paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
0909.4893 | 2 | 0909 | 2010-01-12T19:40:45 | Range Non-Overlapping Indexing | [
"cs.DS"
] | We study the non-overlapping indexing problem: Given a text T, preprocess it so that you can answer queries of the form: given a pattern P, report the maximal set of non-overlapping occurrences of P in T. A generalization of this problem is the range non-overlapping indexing where in addition we are given two indexes i,j to report the maximal set of non-overlapping occurrences between these two indexes. We suggest new solutions for these problems. For the non-overlapping problem our solution uses O(n) space with query time of O(m + occ_{NO}). For the range non-overlapping problem we propose a solution with O(n\log^\epsilon n) space for some 0<\epsilon<1 and O(m + \log\log n + occ_{ij,NO}) query time. | cs.DS | cs |
Range Non-Overlapping Indexing
Hagai Cohen and Ely Porat⋆
Department of Computer Science, Bar-Ilan University, 52900 Ramat-Gan, Israel
{cohenh5,porately}@cs.biu.ac.il
Abstract. We study the non-overlapping indexing problem: Given a
text T , preprocess it in order to answer queries of the form: given a
pattern P , report the maximal set of non-overlapping occurrences of P in
T . A generalization of this problem is the range non-overlapping indexing
where in addition we are given two indexes i, j to report the maximal set
of non-overlapping occurrences between these two indexes. We suggest
new solutions for these problems. For the non-overlapping problem our
solution uses O(n) space with query time of O(m+occN O). For the range
non-overlapping problem we propose a solution with O(n logǫ n) space
for some 0 < ǫ < 1 and O(m + log log n + occij,N O ) query time.
1
Introduction and Related Work
Given a text T of length n over an alphabet Σ, the text indexing problem is to
build an index on T which can answer pattern matching queries efficiently: Given
a pattern P of length m, we want to report all its occurrences in T . There are
some known solutions for this problem. For instance, the suffix tree, proposed by
Weiner [1], which is a compacted trie storing all suffixes of the text. A suffix tree
for text T of length n requires O(n) space and can be built in O(n) preprocessing
time. It has query time of O(m + occ) where occ is the number of occurrences of
P in T .
Range text indexing, also known as position restricted substring searching, is
the problem of finding a pattern P in a substring of the text T between two given
positions i, j. A solution for this problem was presented by Makinen and Navarro
[2]. It uses O(n logǫ n) space and has query time of O(m + log log n + occ). Their
solution is based on another problem - the range searching problem.
The range searching problem is to preprocess a set of points in a d-dimensional
space for answering queries about the set of points which are contained within
a specific range. Alstrup et al [3] proposed a solution for the orthogonal two
dimensional range searching problem when all the points are from n × n grid,
which costs O(n logǫ n) space and has O(log log n + k) query time, where k is
the number of points inside the range. Grossi and Iwona in [4] have shown how
to use Alstrup's data structure to get all the points inside a particular range in
a specific order using some kind of rank function.
⋆ This work was supported by BSF and ISF
In text indexing we are sometimes interested in reporting only the non-
overlapping occurrences of P in T . There is such interest in fields such as pat-
tern recognition, computational linguistics, speech recognition, data compres-
sion, etc. For instance, we might want to compress a text by replacing each
non-overlapping occurrence of a substring of it with a pointer to a single copy
of the substring.
Another problem is the string statistics problem [5, 6] which consists of pre-
processing a text T such that when given a query pattern P , the maximum
number of non-overlapping occurrences of P in T can be reported efficiently.
However, in the string statistics problem we only return the number of non-
overlapping occurrences not the actual occurrences. In this paper, we present
the first non-trivial solution for the non-overlapping indexing problem where we
want to report the maximal sequence of non-overlapping occurrences of P in T .
Keller et al [7] proposed a solution for a generalization of this problem
called the range non-overlapping indexing where we want to report the non-
overlapping occurrences in a substring of T . Their solution has query time of
O(m+occij,N O log log n) and uses O(n log n) space, where occij,N O is the number
of the maximal non-overlapping occurrences in the substring T [i : j].
Crochemore et al [8] suggested another solution for the range non-overlapping
indexing problem. Their solution has optimal query time of O(m + occij,N O) but
requires O(n1+ǫ) space.
In this paper, we present new solutions for the non-overlapping indexing
problem, which use the periodicity of the text and pattern in order to minimize
the query time. Our solution for non-overlapping indexing problem uses O(n)
space with optimal query time of O(m + occN O). For the range non-overlapping
indexing problem we present a solution of O(n logǫ n) space for some 0 < ǫ < 1
with O(m + log log n + occij,N O) query time.
2 Preliminaries
Let n be the length of the text T . And let m be the length of the pattern P . For
two integers i, j (i ≤ j), T [i : j] is the substring of T from i to j.
We will use the Suffix Tree as our main data structure. Each leaf in the Suffix
Tree represents a suffix in the text. In each leaf we save two values: y - the start
location of its suffix in the text and x - the location of the leaf in a left to right
order of all the leaves of the Suffix Tree (lexicographic order, for example). We
have two orders on the leaves and therefore on the suffixes as well: x-order and
y-order. The y-order is the text order and the x-order is the suffix tree leaves
order.
When we search for a pattern P in a Suffix Tree ST of T , we finish searching
at some node v. The subtree rooted by v has all the occurrences of P in T in its
leaves. We denote by l and r the x-value of the leftmost leaf and the x-value of
the rightmost leaf of that subtree respectively. Therefore, the occurrences of P
in T are all the leaves with x-value between l and r.
We denote the number of occurrences of P in T by occ. The number of non-
overlapping occurrences will be denoted as occN O. The number of occurrences of
P in T [i : j] and the non-overlapping occurrences of P in T [i : j] will be denoted
by occij and occij,N O respectively.
3 A Solution for Non-Overlapping Indexing
We use a new approach for solving this problem. Our solution uses the periodicity
of the text and the pattern. We divide patterns for two types: periodic and
aperiodic. A different strategy will be used for each type.
Definition 1. A pattern that can appear more than twice overlapping is called a
periodic pattern. A pattern that can appear at most twice overlapping is called
an aperiodic pattern.
3.1 Aperiodic Pattern
In the aperiodic case we use the periodicity of the pattern to answer a query.
We use the familiar Suffix Tree to get all the leaves that correspond to the
given pattern. After we have all the leaves, we need to remove the overlapping
occurrences. This can be done by sorting the leaves in y-order, going over the
sorted list and filtering the overlapping occurrences. However, sorting occ items
costs O(occ log occ) which is greater than the optimal O(occ). In order to solve
this sorting part we use the following theorem.
Theorem 1. All occurrences of a pattern can be reported and sorted in text
order in O(m + occ) time using O(n) space.
Proof. We use a Suffix Tree to get all the occurrences in O(m + occ) time. For
sorting all the occurrences we will use a renaming method on the Suffix Tree.
Each leaf has its location, i.e., its y-value index, in the whole tree. Saving this
location for a leaf costs log n bits because the whole tree has n leaves. Hence,
the domain for the location is n. Nevertheless, we are interested in the order of
the leaves in a subtree of the occurrences and not in the whole tree. Thus, we
would like to save the location of a leaf for a subtree with less leaves. If we save
for each leaf its location in a subtree with less leaves, for example √n leaves, it
will cost us only log √n. Therefore, for each leaf, aside from keeping its location
in the whole suffix tree, we save its location in a subtree of size √n, its location
i√n for i ≥ 1 until we
in a subtree of size 4√n, and so on for all subtrees of size 2
reach a constant size.
We use Radix Sort which can sort n numbers in a domain of n2 in O(n) time
for sorting the leaves by their locations. Given a subtree whose leaves we wish
to sort in y-order, we can sort them by the locations of the subtree of size at
most O(occ2), this will cost us only O(occ) by using Radix Sort because we sort
occ items in a domain of at most occ2.
In each leaf we save:
log n bits for its location in the ST.
log √n bits for its location in the subtree of size √n.
log 4√n bits for its location in the subtree of size 4√n.
etc.
This sums as following: log n + log √n + log 4√n + log 8√n... = log n + 1
4 log n + 1
1
2 log n +
8 log n + ... ≤ 2 log n.
to n · 2 log n = O(n log n) bits which is O(n) space.
Therefore, we save only 2 log n bits per leaf. We have n leaves summing up
⊓⊔
Theorem 1 provides us a sorted list of all occurrences in O(m + occ) query
time and O(n) space. By filtering the overlapping occurrences which costs O(occ)
time, we are through. Because in an aperiodic pattern, O(occ) = O(occN O), the
query time equals to O(m + occN O).
3.2 Periodic Pattern
The periodic case is more complex. In this case we use the periodicity of the text
in order to answer a query.
Definition 2. A node in the Suffix Tree which represents a suffix which is a
periodic pattern is called period node.
Definition 3. Let s be a string. We define a period of s to be a string p, such
that s = pt ´p, for t ≥ 1 where ´p is prefix of p.
Lemma 1. A period node has only one son which can also be a period node.
Proof. Let a be a period node. Therefore, the string represented by a has a
period p. For a son of a to be a period node too it must continue the period p.
If a ends with a character c than the node which has the next character in p
that is after that c is the period node. There can be only one child of a which
can start with this character. Thus, a period node can have only one son which
⊓⊔
is also period node.
Note that by the period definition a string can have more than one period,
by taking pnew = pp for example. Nevertheless, each period must overlap with
a shorter period of the same string. Thus, there can't be more than one such
character c.
Definition 4. The path that starts from the a period node and goes through all
nodes which continue that period in the Suffix Tree is called a period path. We
denote the number of nodes in a period path to be the period path length.
Lemma 2. Let p2 be a period path on a path P T to the root in the Suffix Tree.
Than p2 must be at least twice as long as the previous period path p1 on P T .
Proof. On P T , between p1 and p2 there must be at least one node which is
not a period node. For p2 to be a period path is must represent a period suffix
which must be started with the period of p1. Moreover, p2 period suffix should
be continued by the character of the next node after p1 which is not a period
node. After it there must be the period of p1 again. Therefore, p2 length is at
⊓⊔
least twice longer than p1.
Lemma 3. The largest number of different period paths contained in the path
from the root to a period node is log n.
Proof. Let P T be the path from the root to some period node in the Suffix Tree.
According to Lemma 2 each period path on P T must be at least twice as long
as the previous period path on P T . Therefore, if we have more than log n period
paths on P T , than the length of the last period path must be greater than n
which is the text length. Hence, the number of period paths on the same path
⊓⊔
can't be more than log n.
Definition 5. A period sequence is the maximum substring in the text of
some period which is repeated more than twice. We will mark it as [s, e], where
s and e are the start index and the end index of the period sequence accordingly.
The period sequences of a period pattern are all the period sequences which start
with that periodic pattern. The period length of a period sequence is the length
of the period inside repeated the sequence.
Example 1. Lets T be the text "abababcabababcabababc". The period sequences
are: For the period "ab", the period sequences are [1,6], [8,13], [15,20] which have
period length 2. For the period "abababc", the period sequence is [1,21] which
has period length 7.
Lemma 4. Given a list L of all the period sequences of some periodic pattern
in the text. All the non-overlapping occurrences of that periodic pattern can be
retrieved from this list in O(occN O) time.
step which can be easily calculated.
Proof. For a period sequence [s, e] with period length pl, the non-overlapping
occurrences of a periodic pattern P with length m are the group: s+i∗stepstep =
⌈ m
pl⌉ ∗ pl, 0 ≤ i ≤ e−s−1
We report all the occurrences in each period sequence in L. The number of
all occurrences we report is O(occN O), so the total time for reporting all the
⊓⊔
non-overlapping occurrences from L is O(occN O).
For answering the non-overlapping indexing we use the following data struc-
ture. We build a data structure for each period path on the Suffix Tree, saving
a list of all period sequences sorted by their length for each period path. Each
period node in the Suffix Tree is on a period path. We save a pointer from each
period node on the Suffix Tree to the period sequence list of its period path. This
pointer will point to the period node appropriate length on the period sequences
list.
Theorem 2. Using the data structure described above, all period sequences of a
period pattern can be retrieved in O(m + occN O) time using O(n log n) space.
Proof. On a query we go to that data structure, get all the period sequences
and by Lemma 4 we calculate all the non-overlapping occurrences. This will
take O(occN O) time, because the number of the period sequences that we will
get is less than the number of the non-overlapping occurrences of the pattern.
If we come up with a long pattern we won't get shorter period sequences which
don't fit the pattern so we won't get unnecessary period sequences.
The space for this data structure is O(n log n). This is because there are n
nodes where each one can be at most in log n period paths. For each node, we
⊓⊔
save all its period paths so it needs O(n log n).
Now, we will show how to reduce the space needed for this data structure.
Definition 6. Let's define a degree of a period sequence to be the maximum
degree of a period sequence included in it plus one. A period sequence without
any period sequences in it will has the degree 0.
Lemma 5. The maximum degree of a period sequence can be at most log n.
Proof. Let ps be the period sequence with the maximum degree in the text. In
ps there is a period sequence with a degree decreased by one. In that period
sequence there is another period sequence with a degree decreased by one. And
so on until we receive a period sequence ps0 with the degree 0. The length of
each period sequence from ps0 to ps is at least twice the length of the period
sequence in it. The maximum length of ps can be at most n, therefore, its degree
⊓⊔
can be at most log n.
Lemma 6. There are at most O(n) period sequences.
Proof. We will count the number of period sequences in each degree:
The number of period sequences of degree 0 can be at most n.
The number of period sequences of degree 1 can be at most n
2 .
The number of period sequences of degree 2 can be at most n
4 .
. . .
The number of period sequences of degree log n can be at most 1.
Summery: n + n
⊓⊔
Theorem 3. The data structure in Theorem 2 can be saved using only O(n)
space.
2 + .. + 1 ≤ 2n = O(n)
Proof. We save all the period paths in a data structure. Each one with its own
period sequences. Each period sequence appears in only one period path. From
Lemma 6 we have O(n) period sequences. Therefore we save at most O(n) space
for all the period sequences. Thus, we need only O(n) space for this data struc-
⊓⊔
ture.
Corollary 1. Using these two different strategies for each type of pattern we can
solve the non-overlapping indexing problem in O(n) space with O(m + occN O)
query time.
4 A solution for Range Non-Overlapping Indexing
We propose a better solution for this problem. Our solution costs O(n logǫ n)
space for some 0 < ǫ < 1 and has query time of O(m + log log n + occij,N O).
4.1 Rank Sensitive Range Searching
We use a data structure for answering the two-dimensional orthogonal range
searching problem. Alstrup et al [3] proposed a data structure for this problem
which requires O(n logǫ n) space with query time of O(log log n + k) where k is
the number of points in the range.
Nevertheless, this range query data structure reports all the points in the
range with no specific order. We want to get those points in a specific order.
Therefore, in addition to this data structure we will use a method suggested by
Grossi et al [4] for a rank sensitive data structure. This gives us a data structure
which uses O(n logǫ n) space with query time of O(log log n) and O(1) per point,
where the points will be reported in rank order. For simplicity we will call this
data structure RSDS from now on.
4.2 Aperiodic Pattern
We use a Rank Sensitive Data Structure to answer aperiodic queries. In the
RSDS we store all the occurrences as points by their x value and y value, where
the rank function of a point will be its y value. Given a pattern P and range
i, j we can get l, r from the Suffix Tree, the leftmost leaf and the rightmost leaf
which are occurrences of P . Then we will do a range query for points within
[i, j]x[l, r] to get all the correct occurrences. Because the rank in the RSDS is by
y value, we will get the points and therefore the occurrences, sorted in the text
order. The only remaining action is filtering the overlapping occurrences.
The RSDS costs O(n logǫ n) space. RSDS query time is O(log log n+k) where
k is the size of the output which is equal to O(occij ). In our case k equals
O(occij,N O) because for an aperiodic pattern O(occij ) = O(occij,N O). Therefore,
aperiodic pattern has query time of O(m) for searching the Suffix Tree plus
O(log log n + occij,N O) for the RSDS query. Concluded in O(m + log log n +
occij,N O).
4.3 Periodic Pattern
The periodic case is more complex. We save all period sequences in the text as
points in two RSDS. For a periodic sequence [s, e] with period length pl, we save
two points (x1, y1), (x2, y2) with the following values:
x1 = Index of the suffix of s in the left to right order of all the ST leaves
y1 = s
x2 = Index of the suffix of s in the left to right order of all the ST leaves
y2 = e − pl + 1
Point (x1, y1) will be saved in the first RSDS with a rank function of x value
in descending order. Point (x2, y2) will be saved in the second RSDS with a rank
function of x value in ascending order. Sometimes there will be multiple period
sequences with the same start index or end index, each with a different degree.
When this happens we save only the one with the highest degree. We can easily
convert a period sequence [s, e] to these two points and vice versa.
Following Lemma 6 the number of points in the two RSDS is O(n). Hence,
each RSDS costs O(n logǫ n) space.
Given a pattern P of length m and range i, j we answer using Algorithm 1.
1
2
3
4
5
6
7
8
9
10
11
12
13
Get the range l, r from the Suffix Tree ;
S ←− Query first RSDS for all points within [i, j]x[l, r] ;
S ←− S ∪ Query second RSDS for all points within [i, j]x[l, r] ;
S ←− S ∪ Query first RSDS for the first point within [1, i]x[l, r] ;
P S ←− convertAllP ointsT oP eriodSequences(S) ;
P S2 ←− ∅ ;
for ps ∈ P S do
x ←− ps ;
while x period length is greater than m do
x ←− the first period sequence inside x ;
end
P S2 ←− P S2 ∪ {x}
end
Algorithm 1: Periodic Pattern Range Query
Getting the first period sequence inside a period sequence can be done by
using another data structure saving for each period sequence its period length,
and a pointer to the first period sequence in it. Thus, given a period sequence it
costs O(1) to find the first period sequence inside it with a degree decreased by
one.
Lemma 7. The number of period sequences we have to go down in order to find
our appropriate period sequence is lower than the number of occurrences that
will be extracted from the period sequences inside the first period sequence we
received.
Proof. Each degree we get down means that there is another occurrence in the
next period sequence. Each step down, adds at least another occurrence. There-
fore, until we get to the appropriate period sequence we work at most O(k) where
k is the number of occurrences we will get from the period sequences inside the
⊓⊔
first period sequence we encounter.
By Lemma 7 it does not cost us more time when we get a period sequence
whose period length is longer than the pattern length.
Theorem 4. All the non-overlapping occurrences can be calculated from the
period sequences got by these three queries.
Proof. First of all we will see that each period sequence we get from these queries
has at least one occurrence of P . Let (x, y) be the point we get. It corresponds
to a period sequence [s, e]. We get only points which have x values between l
and r. The x value of a point is the index of the suffix of s in the left to right
order of all the ST leaves. So if we get a point (x, y) with x between l and r
it means that the corresponding period sequence [s, e] has an occurrence of P .
This happens because all the leaves in the ST between l and r are occurrences
of P .
Now, we need to prove two more things. The first is that all the period
sequences we get from the RSDS are suitable for us and that we haven't got
unnecessary period sequences, which don't fit to P in the range i, j. The second
thing is that we didn't miss any period sequence which can have some suitable
occurrences.
We start by proving that we get all the occurrences of P in the range [i, j]
from the period sequences we get in the three queries. Period sequences of P in
T can be in some cases. Let [s, e] be our period sequence.
The first case is that [s, e] is out of the range [i, j], s < e ≤ i < j or
i < j ≤ s < e. In this case we wouldn't like to get this period sequences at
all. The first two queries will not resolve these period sequences because we do
a query on [i, j]x[l, r] but s is out of range and the points corresponding to
this period sequence have y value equals s. Nevertheless, we can get this period
sequence in the third query. However, it will be at most one period sequence
which can be checked in O(1) time.
The second case, which is the simplest, is that [s, e] is fully inside the range
[i, j], i ≤ s < e ≤ j. In this case we get all the suitable period sequences from
the first query. Nevertheless, we can get the same period sequence twice, first
from the first query and again from the second query. Therefore, we will have to
check any period sequence that we get in order to prevent duplicate occurrence
reporting.
The third case is when only e or s is out of range but not both, s < i < e ≤ j
or i ≤ s < j < e. This time if i ≤ s < j < e we will get the period sequence from
the first query. Otherwise, if s < i < e ≤ j we will get the period sequence from
the second query.
The fourth case is when the range [i, j] is fully inside the period sequence
[s, e], s < i < j < e. In order to solve this case we have the third query which will
resolve the last start of a period sequence which is before index i. This period
⊓⊔
sequence can be checked in O(1) time.
Corollary 2. Using these two different strategies for each type of pattern, the
range non-overlapping text indexing problem can be solved in O(n logǫ n) space
for some 0 < ǫ < 1 and query time of O(m + log log n + occij,N O).
5 Conclusion
We have studied the problem of non-overlapping indexing. In this paper, we
provide the first non-trivial solution for this problem. In addition we proposed
a better solution for a generalization of this problem, the range non-overlapping
problem.
References
1. Weiner, P.: Linear pattern matching algorithms. In: FOCS, IEEE (1973) 1 -- 11
2. Makinen, V., Navarro, G.: Position-restricted substring searching. In Correa, J.R.,
Hevia, A., Kiwi, M.A., eds.: LATIN. Volume 3887 of Lecture Notes in Computer
Science., Springer (2006) 703 -- 714
3. Alstrup, S., Brodal, G.S., Rauhe, T.: New data structures for orthogonal range
searching. In: In 41st Symposium on Foundations of Computer Science (FOCS00.
(2000) 198 -- 207
4. Bialynicka-Birula, I., Grossi, R.: Rank-sensitive data structures. In Consens, M.P.,
Navarro, G., eds.: SPIRE. Volume 3772 of Lecture Notes in Computer Science.,
Springer (2005) 79 -- 90
5. Apostolico, A., Preparata, F.P.: Data structures and algorithms for the string statis-
tics problem. Algorithmica 15(5) (1996) 481 -- 494
6. Brodal, G.S., Lyngsø, R.B., Ostlin, A., Pedersen, C.N.S.: Solving the string statistics
problem in time o(n log n). In: ICALP '02: Proceedings of the 29th International
Colloquium on Automata, Languages and Programming, London, UK, Springer-
Verlag (2002) 728 -- 739
7. Keller, O., Kopelowitz, T., Lewenstein, M.: Range non-overlapping indexing and
In Dehne, F.K.H.A., Sack, J.R., Zeh, N., eds.: WADS.
successive list indexing.
Volume 4619 of Lecture Notes in Computer Science., Springer (2007) 625 -- 636
8. Crochemore, M., Iliopoulos, C.S., Kubica, M., Rahman, M.S., Walen, T.: Improved
algorithms for the range next value problem and applications. In Albers, S., Weil,
P., eds.: STACS. Volume 08001 of Dagstuhl Seminar Proceedings., Internationales
Begegnungs- und Forschungszentrum fuer Informatik (IBFI), Schloss Dagstuhl, Ger-
many (2008) 205 -- 216
|
1802.03587 | 2 | 1802 | 2018-02-14T11:39:54 | Network Flow-Based Refinement for Multilevel Hypergraph Partitioning | [
"cs.DS"
] | We present a refinement framework for multilevel hypergraph partitioning that uses max-flow computations on pairs of blocks to improve the solution quality of a $k$-way partition. The framework generalizes the flow-based improvement algorithm of KaFFPa from graphs to hypergraphs and is integrated into the hypergraph partitioner KaHyPar. By reducing the size of hypergraph flow networks, improving the flow model used in KaFFPa, and developing techniques to improve the running time of our algorithm, we obtain a partitioner that computes the best solutions for a wide range of benchmark hypergraphs from different application areas while still having a running time comparable to that of hMetis. | cs.DS | cs | Network Flow-Based Refinement for Multilevel
Hypergraph Partitioning
Tobias Heuer1, Peter Sanders2, and Sebastian Schlag3
1 Karlsruhe Institute of Technology, Karlsruhe, Germany
2 Karlsruhe Institute of Technology, Karlsruhe, Germany
[email protected]
[email protected]
3 Karlsruhe Institute of Technology, Karlsruhe, Germany
[email protected]
8
1
0
2
b
e
F
4
1
]
S
D
.
s
c
[
2
v
7
8
5
3
0
.
2
0
8
1
:
v
i
X
r
a
Abstract
We present a refinement framework for multilevel hypergraph partitioning that uses max-flow
computations on pairs of blocks to improve the solution quality of a k-way partition. The frame-
work generalizes the flow-based improvement algorithm of KaFFPa from graphs to hypergraphs
and is integrated into the hypergraph partitioner KaHyPar. By reducing the size of hypergraph
flow networks, improving the flow model used in KaFFPa, and developing techniques to improve
the running time of our algorithm, we obtain a partitioner that computes the best solutions
for a wide range of benchmark hypergraphs from different application areas while still having a
running time comparable to that of hMetis.
1998 ACM Subject Classification G.2.2 Graph Theory, G.2.3 Applications
Keywords and phrases multilevel hypergraph partitioning, network flows, refinement
Digital Object Identifier 10.4230/LIPIcs...
1
Introduction
Given an undirected hypergraph H = (V, E), the k-way hypergraph partitioning problem
is to partition the vertex set into k disjoint blocks of bounded size (at most 1 + ε times
the average block size) such that an objective function involving the cut hyperedges is
minimized. Hypergraph partitioning (HGP) has many important applications in practice
such as scientific computing [12] or VLSI design [43]. Particularly VLSI design is a field
where small improvements can lead to significant savings [56].
It is well known that HGP is NP-hard [38], which is why practical applications mostly
use heuristic multilevel algorithms [11, 13, 25, 26]. These algorithms successively contract
the hypergraph to obtain a hierarchy of smaller, structurally similar hypergraphs. After
applying an initial partitioning algorithm to the smallest hypergraph, contraction is undone
and, at each level, a local search method is used to improve the partitioning induced by the
coarser level. All state-of-the-art HGP algorithms [2, 4, 7, 16, 28, 31, 32, 33, 48, 51, 52, 54]
either use variations of the Kernighan-Lin (KL) [34, 49] or the Fiduccia-Mattheyses (FM)
heuristic [19, 46], or simpler greedy algorithms [32, 33] for local search. These heuristics move
vertices between blocks in descending order of improvements in the optimization objective
(gain) and are known to be prone to get stuck in local optima when used directly on the
input hypergraph [33]. The multilevel paradigm helps to some extent, since it allows a more
global view on the problem on the coarse levels and a very fine-grained view on the fine levels
of the multilevel hierarchy. However, the performance of move-based approaches degrades
© Tobias Heuer, Peter Sanders and Sebastian Schlag;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
for hypergraphs with large hyperedges. In these cases, it is difficult to find meaningful vertex
moves that improve the solution quality because large hyperedges are likely to have many
vertices in multiple blocks [53]. Thus the gain of moving a single vertex to another block is
likely to be zero [41].
While finding balanced minimum cuts in hypergraphs is NP-hard, a minimum cut separ-
ating two vertices can be found in polynomial time using network flow algorithms and the
well-known max-flow min-cut theorem [21]. Flow algorithms find an optimal min-cut and do
not suffer the drawbacks of move-based approaches. However, they were long overlooked as
heuristics for balanced partitioning due to their high complexity [40, 57]. In the context of
graph partitioning, Sanders and Schulz [47] recently presented a max-flow-based improvement
algorithm which is integrated into the multilevel partitioner KaFFPa and computes high
quality solutions.
Outline and Contribution. Motivated by the results of Sanders and Schulz [47], we gener-
alize the max-flow min-cut refinement framework of KaFFPa from graphs to hypergraphs.
After introducing basic notation and giving a brief overview of related work and the tech-
niques used in KaFFPa in Section 2, we explain how hypergraphs are transformed into flow
networks and present a technique to reduce the size of the resulting hypergraph flow network
in Section 3.1. In Section 3.2 we then show how this network can be used to construct a flow
problem such that the min-cut induced by a max-flow computation between a pair of blocks
improves the solution quality of a k-way partition. We furthermore identify shortcomings of
the KaFFPa approach that restrict the search space of feasible solutions significantly and
introduce an advanced model that overcomes these limitations by exploiting the structure of
hypergraph flow networks. We implemented our algorithm in the open source HGP framework
KaHyPar and therefore briefly discuss implementation details and techniques to improve
the running time in Section 3.3. Extensive experiments presented in Section 4 demonstrate
that our flow model yields better solutions than the KaFFPa approach for both hypergraphs
and graphs. We furthermore show that using pairwise flow-based refinement significantly
improves partitioning quality. The resulting hypergraph partitioner, KaHyPar-MF, performs
better than all competing algorithms on all instance classes and still has a running time
comparable to that of hMetis. On a large benchmark set consisting of 3222 instances from
various application domains, KaHyPar-MF computes the best partitions in 2427 cases.
Preliminaries
2
2.1 Notation and Definitions
An undirected hypergraph H = (V, E, c, ω) is defined as a set of n vertices V and a set of m
hyperedges/nets E with vertex weights c : V → R>0 and net weights ω : E → R>0, where
each net is a subset of the vertex set V (i.e., e ⊆ V ). The vertices of a net are called pins.
e∈F ω(e). A vertex v is
incident to a net e if v ∈ e. I(v) denotes the set of all incident nets of v. The degree of a
vertex v is d(v) := I(v). The size e of a net e is the number of its pins. Given a subset
V 0 ⊂ V , the subhypergraph HV 0 is defined as HV 0 := (V 0,{e ∩ V 0 e ∈ E : e ∩ V 0 6= ∅}).
We extend c and ω to sets, i.e., c(U) :=P
v∈U c(v) and ω(F) :=P
Π = {V1, . . . , Vk} such that Sk
A k-way partition of a hypergraph H is a partition of its vertex set into k blocks
i=1 Vi = V , Vi 6= ∅ for 1 ≤ i ≤ k, and Vi ∩ Vj = ∅ for
i 6= j. We call a k-way partition Π ε-balanced if each block Vi ∈ Π satisfies the balance
k e for some parameter ε. For each net e, Λ(e) :=
constraint: c(Vi) ≤ Lmax := (1 + ε)d c(V )
{Vi Vi∩ e 6= ∅} denotes the connectivity set of e. The connectivity of a net e is λ(e) := Λ(e).
T. Heuer, P. Sanders and S. Schlag
XX:3
cost functions are the cut-net metric cut(Π) := P
(λ − 1)(Π) := P
A net is called cut net if λ(e) > 1. Given a k-way partition Π of H, the quotient graph
Q := (Π,{(Vi, Vj) ∃e ∈ E : {Vi, Vj} ⊆ Λ(e)}) contains an edge between each pair of
adjacent blocks. The k-way hypergraph partitioning problem is to find an ε-balanced k-way
partition Π of a hypergraph H that minimizes an objective function over the cut nets for
some ε. Several objective functions exist in the literature [5, 38]. The most commonly used
e∈E0 ω(e) and the connectivity metric
e∈E0(λ(e) − 1) ω(e) [1], where E0 is the set of all cut nets [17]. In this
paper, we use the (λ − 1)-metric. Optimizing both objective functions is known to be
NP-hard [38]. Hypergraphs can be represented as bipartite graphs [29]. In the following,
we use nodes and edges when referring to graphs and vertices and nets when referring to
hypergraphs. In the bipartite graph G∗(V ∪E, F) the vertices and nets of H form the node
set and for each net e ∈ I(v), we add an edge (e, v) to G∗. The edge set F is thus defined as
F := {(e, v) e ∈ E, v ∈ e}. Each net in E therefore corresponds to a star in G∗.
Let G = (V, E, c, ω) be a weighted directed graph. We use the same notation as for
hypergraphs to refer to node weights c, edge weights ω, and node degrees d(v). Furthermore
Γ(u) := {v : (u, v) ∈ E} denotes the neighbors of node u. A path P = hv1, . . . , vki is a
sequence of nodes, such that each pair of consecutive nodes is connected by an edge. A
strongly connected component C ⊆ V is a set of nodes such that for each u, v ∈ C there exists
a path from u to v. A topological ordering is a linear ordering ≺ of V such that every directed
edge (u, v) ∈ E implies u ≺ v in the ordering. A set of nodes B ⊆ V is called a closed
set iff there are no outgoing edges leaving B, i.e., if the conditions u ∈ B and (u, v) ∈ E
imply v ∈ B. A subset S ⊂ V is called a node separator if its removal divides G into two
disconnected components.
A flow network N = (V,E, c) is a directed graph with two distinguished nodes s and
t in which each edge e ∈ E has a capacity c(e) ≥ 0. An (s, t)-flow (or flow) is a function
f : V × V → R that satisfies the capacity constraint ∀u, v ∈ V : f(u, v) ≤ c(u, v), the skew
symmetry constraint ∀v ∈ V × V : f(u, v) = −f(v, u), and the flow conservation constraint
v∈V f(s, v) is defined as the
total amount of flow transferred from s to t. The residual capacity is defined as rf(u, v) =
c(u, v) − f(u, v). Given a flow f, Nf = (V,Ef , rf) with Ef = {(u, v) ∈ V × V rf(u, v) > 0}
is the residual network. An (s, t)-cut (or cut) is a bipartition (S,V \ S) of a flow network N
e∈E0 c(e), where
E0 = {(u, v) ∈ E : u ∈ S, v ∈ V \ S}. The max-flow min-cut theorem states that the value f
of a maximum flow is equal to the capacity of a minimum cut separating s and t [21].
∀u ∈ V \ {s, t} :P
with s ∈ S ⊂ V and t ∈ V \ S. The capacity of an (s, t)-cut is defined asP
v∈V f(u, v) = 0. The value of a flow f :=P
2.2 Related Work
Hypergraph Partitioning. Driven by applications in VLSI design and scientific computing,
HGP has evolved into a broad research area since the 1990s. We refer to [5, 8, 43, 50] for an
extensive overview. Well-known multilevel HGP software packages with certain distinguishing
characteristics include PaToH [12] (originating from scientific computing), hMetis [32, 33]
(originating from VLSI design), KaHyPar [2, 28, 48] (general purpose, n-level), Mondriaan [54]
(sparse matrix partitioning), MLPart [4] (circuit partitioning), Zoltan [16], Parkway [51]
and SHP [31] (distributed), UMPa [52] (directed hypergraph model, multi-objective), and
kPaToH (multiple constraints, fixed vertices) [7]. All of these tools either use variations
of the Kernighan-Lin (KL) [34, 49] or the Fiduccia-Mattheyses (FM) heuristic [19, 46], or
algorithms that greedily move vertices [33] or nets [32] to improve solution quality in the
refinement phase.
XX:4
Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Flows on Hypergraphs. While flow-based approaches have not yet been considered as
refinement algorithms for multilevel HGP, several works deal with flow-based hypergraph
min-cut computation. The problem of finding minimum (s, t)-cuts in hypergraphs was first
considered by Lawler [36], who showed that it can be reduced to computing maximum
flows in directed graphs. Hu and Moerder [29] present an augmenting path algorithm to
compute a minimum-weight vertex separator on the star-expansion of the hypergraph. Their
vertex-capacitated network can also be transformed into an edge-capacitated network using a
transformation due to Lawler [37]. Yang and Wong [57] use repeated, incremental max-flow
min-cut computations on the Lawler network [36] to find ε-balanced hypergraph bipartitions.
Solution quality and running time of this algorithm are improved by Lillis and Cheng [39]
by introducing advanced heuristics to select source and sink nodes. Furthermore, they
present a preflow-based [22] min-cut algorithm that implicitly operates on the star-expanded
hypergraph. Pistorius and Minoux [45] generalize the algorithm of Edmonds and Karp [18]
to hypergraphs by labeling both vertices and nets. Liu and Wong [40] simplify Lawler's
hypergraph flow network [36] by explicitly distinguishing between graph edges and hyperedges
with three or more pins. This approach significantly reduces the size of flow networks derived
from VLSI hypergraphs, since most of the nets in a circuit are graph edges. Note that the
above-mentioned approaches to model hypergraphs as flow networks for max-flow min-cut
computations do not contradict the negative results of Ihler et al. [30], who show that, in
general, there does not exist an edge-weighted graph G = (V, E) that correctly represents
the min-cut properties of the corresponding hypergraph H = (V, E).
Flow-Based Graph Partitioning. Flow-based refinement algorithms for graph partitioning
include Improve [6] and MQI [35], which improve expansion or conductance of bipartitions.
MQI also yields as small improvement when used as a post processing technique on hypergraph
bipartitions initially computed by hMetis [35]. FlowCutter [24] uses an approach similar
to Yang and Wong [57] to compute graph bisections that are Pareto-optimal in regard to
cut size and balance. Sanders and Schulz [47] present a flow-based refinement framework
for their direct k-way graph partitioner KaFFPa. The algorithm works on pairs of adjacent
blocks and constructs flow problems such that each min-cut in the flow network is a feasible
solution in regard to the original partitioning problem.
KaHyPar. Since our algorithm is integrated into the KaHyPar framework, we briefly review
its core components. While traditional multilevel HGP algorithms contract matchings or
clusterings and therefore work with a coarsening hierarchy of O(log n) levels, KaHyPar
instantiates the multilevel paradigm in the extreme n-level version, removing only a single
vertex between two levels. After coarsening, a portfolio of simple algorithms is used to create
an initial partition of the coarsest hypergraph. During uncoarsening, strong localized local
search heuristics based on the FM algorithm [19, 46] are used to refine the solution. Our work
builds on KaHyPar-CA [28], which is a direct k-way partitioning algorithm for optimizing the
(λ − 1)-metric. It uses an improved coarsening scheme that incorporates global information
about the community structure of the hypergraph into the coarsening process.
2.3 The Flow-Based Improvement Framework of KaFFPa
We discuss the framework of Sanders and Schulz [47] in greater detail, since our work makes
use of the techniques proposed by the authors. For simplicity, we assume k = 2. The
techniques can be applied on a k-way partition by repeatedly executing the algorithm on
pairs of adjacent blocks. To schedule these refinements, the authors propose an active block
T. Heuer, P. Sanders and S. Schlag
XX:5
scheduling algorithm, which schedules blocks as long as their participation in a pairwise
refinement step results in some changes in the k-way partition.
An ε-balanced bipartition of a graph G = (V, E, c, ω) is improved with flow computations
as follows. The basic idea is to construct a flow network N based on the induced subgraph
G[B], where B ⊆ V is a set of nodes around the cut of G. The size of B is controlled by an
imbalance factor ε0 := αε, where α is a scaling parameter that is chosen adaptively depending
on the result of the min-cut computation. If the heuristic found an ε-balanced partition
using ε0, the cut is accepted and α is increased to min(2α, α0) where α0 is a predefined upper
bound. Otherwise it is decreased to max( α
2 , 1). This scheme continues until a maximal
number of rounds is reached or a feasible partition that did not improve the cut is found.
In each round, the corridor B := B1 ∪ B2 is constructed by performing two breadth-first
searches (BFS). The first BFS is done in the induced subgraph G[V1]. It is initialized with the
2 e− c(V2). The second BFS
boundary nodes of V1 and stops if c(B1) would exceed (1+ 0)d c(V )
constructs B2 in an analogous fashion using G[V2]. Let δB := {u ∈ B ∃(u, v) ∈ E : v /∈ B}
be the border of B. Then N is constructed by connecting all border nodes δB ∩ V1 of G[B]
to the source s and all border nodes δB ∩ V2 to the sink t using directed edges with an
edge weight of ∞. By connecting s and t to the respective border nodes, it is ensured that
edges incident to border nodes, but not contained in G[B], cannot become cut edges. For
α = 1, the size of B thus ensures that the flow network N has the cut property, i.e., each
(s, t)-min-cut in N yields an ε-balanced partition of G with a possibly smaller cut. For larger
values of α, this does not have to be the case.
After computing a max-flow in N , the algorithm tries to find a min-cut with better
balance. This is done by exploiting the fact that one (s, t)-max-flow contains information
about all (s, t)-min-cuts [44]. More precisely, the algorithm uses the 1–1 correspondence
between (s, t)-min-cuts and closed sets containing s in the Picard-Queyranne-DAG Ds,t of
the residual graph Nf [44]. First, Ds,t is constructed by contracting each strongly connected
component of the residual graph. Then the following heuristic (called most balanced minimum
cuts) is repeated several times using different random seeds. Closed node sets containing s
are computed by sweeping through the nodes of DAGs,t in reverse topological order (e.g.
computed using a randomized DFS). Each closed set induces a differently balanced min-cut
and the one with the best balance (with respect to the original balance constraint) is used as
resulting bipartition.
3 Hypergraph Max-Flow Min-Cut Refinement
In the following, we generalize the flow-based refinement algorithm of KaFFPa to hypergraph
partitioning. In Section 3.1 we first show how hypergraph flow networks N are constructed
in general and introduce a technique to reduce their size by removing low-degree hypernodes.
Given a k-way partition Πk = {V1, . . . , Vk} of a hypergraph H = (V, E), a pair of blocks
(Vi, Vj) adjacent in the quotient graph Q, and a corridor B, Section 3.2 then explains how
N is used to build a flow problem F based on a B-induced subhypergraph HB = (VB, EB).
The flow problem F is constructed such that an (s, t)-max-flow computation optimizes the
cut metric of the bipartition Π2 = (Vi, Vj) of HB and thus improves the (λ − 1)-metric in H.
Section 3.3 then discusses the integration into KaHyPar and introduces several techniques to
speed up flow-based refinement. Algorithm 1 gives a pseudocode description of the entire
flow-based refinement framework.
XX:6
Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Algorithm 1: Flow-Based Refinement
Input: Hypergraph H, k-way partition Πk = {V1, . . . , Vk}, imbalance parameter ε.
Algorithm MaxFlowMinCutRefinement(H, Πk)
Q := QuotientGraph(H, Πk)
while ∃ active blocks ∈ Q do
foreach {(Vi, Vj) ∈ Q Vi ∨ Vj is active} do
Πold = Πbest := {Vi, Vj} ⊆ Πk
εold = εbest := imbalance(Πk)
α := α0
do
// in the beginning all blocks are active
// choose a pair of blocks
// extract bipartition to be improved
// imbalance of current k-way partition
// use large B-corridor for first iteration
// adaptive flow iterations
B := computeB-Corridor(H, Πbest, αε)
// as described in Section 2.3
HB := SubHypergraph(H, B)
// create B-induced subhypergraph
NB := FlowNetwork(HB)
// as described in Section 3.1
F := FlowProblem(NB)
// as described in Section 3.2
// compute maximum flow on F
f := maxFlow(F)
Πf := mostBalancedMinCut(f,F)
// as in Section 2.3 & 3.1
εf := imbalance(Πf ∪ Πk \ Πold)
// imbalance of new k-way partition
if (cut(Πf) < cut(Πbest) ∧ εf ≤ ε) ∨ εf < εbest then // found improvement
// update α, Πbest,εbest
// decrease size of B-corridor in next iteration
α := min(2α, α0), Πbest := Πf , εbest := εf
else α := α
2
while α ≥ 1
if Πbest 6= Πold then
Πk := Πbest ∪ Πk \ Πold
activateForNextRound(Vi, Vj)
// improvement found
// replace Πold with Πbest
// reactivate blocks for next round
return Πk
Output: improved ε-balanced k-way partition Πk = {V1, . . . , Vk}
3.1 Hypergraph Flow Networks
The Liu-Wong Network [40]. Given a hypergraph H = (V, E, c, ω) and two distinct nodes
s and t, an (s, t)-min-cut can be computed by finding a minimum-capacity cut in the following
flow network N = (V,E):
V contains all vertices in V .
For each multi-pin net e ∈ E with e ≥ 3, add two bridging nodes e0 and e00 to V and
a bridging edge (e0, e00) with capacity c(e0, e00) = ω(e) to E. For each pin p ∈ e, add two
edges (p, e0) and (e00, p) with capacity ∞ to E.
For each two-pin net e = (u, v) ∈ E, add two bridging edges (u, v) and (v, u) with capacity
ω(e) to E.
The flow network of Lawler [36] does not distinguish between two-pin and multi-pin nets.
This increases the size of the network by two vertices and three edges per two-pin net.
Figure 1 shows an example of the Lawler and Liu-Wong hypergraph flow networks as well as
of our network described in the following paragraph.
Removing Low Degree Hypernodes. We further decrease the size of the network by using
the observation that the problem of finding an (s, t)-min-cut of H can be reduced to finding
a minimum-weight (s, t)-vertex-separator in the star-expansion, where the capacity of each
star-node is the weight of the corresponding net and all other nodes (corresponding to
T. Heuer, P. Sanders and S. Schlag
XX:7
Figure 1 Illustration of hypergraph flow networks. Our approach further sparsifies the flow
network of Liu and Wong [40]. Thin edges have infinite capacity.
vertices in H) have infinite capacity [29]. Since the separator has to be a subset of the
star-nodes, it is possible to replace any infinite-capacity node by adding a clique between
all adjacent star-nodes without affecting the separator. The key observation now is that
an infinite-capacity node v with degree d(v) induces 2d(v) infinite-capacity edges in the
Lawler network [36], while a clique between star-nodes induces d(v)(d(v) − 1) edges. For
hypernodes with d(v) ≤ 3, it therefore holds that d(v)(d(v) − 1) ≤ 2d(v). Thus we can
reduce the number of nodes and edges of the Liu-Wong network as follows. Before applying
the transformation on the star-expansion of H, we remove all infinite-capacity nodes v
corresponding to hypernodes with d(v) ≤ 3 that are not incident to any two-pin nets and add
a clique between all star-nodes adjacent to v. In case v was a source or sink node, we create
a multi-source multi-sink problem by adding all adjacent star-nodes to the set of sources
resp. sinks [20].
Reconstructing Min-Cuts. After computing an (s, t)-max-flow in the Lawler or Liu-Wong
network, an (s, t)-min-cut of H can be computed by a BFS in the residual graph starting
from s. Let S be the set of nodes corresponding to vertices of H reached by the BFS. Then
(S, V \ S) is an (s, t)-min-cut. Since our network does not contain low degree hypernodes,
we use the following lemma to compute an (s, t)-min-cut of H:
(cid:73) Lemma 1. Let f be a maximum (s, t)-flow in the Lawler network N = (V,E) of a
hypergraph H = (V, E) and (S,V \ S) be the corresponding (s, t)-min-cut in N . Then for
each node v ∈ S ∩ V , the residual graph Nf = (Vf ,Ef) contains at least one path hs, . . . , e00i
to a bridging node e00 of a net e ∈ I(v).
Proof. Since v ∈ S, there has to be some path s (cid:32) v in Nf. By definition of the flow
network, this path can either be of the form P1 = hs, . . . , e00, vi or P2 = hs, . . . , e0, vi for some
bridging nodes e0, e00 corresponding to nets e ∈ I(v). In the former case we are done, since
e00 ∈ P1. In the latter case the existence of edge (e0, v) ∈ Ef implies that there is a positive
flow f(v, e0) > 0 over edge (v, e0) ∈ E. Due to flow conservation, there exists at least one
edge (e00, v) ∈ E with f(e00, v) > 0, which implies that (v, e00) ∈ Ef. Thus we can extend the
path P2 to hs, . . . , e0, v, e00i.
(cid:74)
Thus (A, V \ A) is an (s, t)-min-cut of H, where A := {v ∈ e ∃e ∈ E : hs, . . . , e00i in Nf}.
Furthermore this allows us to search for more balanced min-cuts using the Picard-Queyranne-
DAG of Nf as described in Section 2.3. By the definition of closed sets it follows that if a
bridging node e00 is contained in a closed set C, then all nodes v ∈ Γ(e00) (which correspond
to vertices of H) are also contained in C. Thus we can use the respective bridging nodes e00
as representatives of removed low degree hypernodes.
Lawler[36]LiuandWong[40]OurNetworkHypergraphstststst XX:8
Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
B )∩ I(←−
EB to denote the set of border nets.
3.2 Constructing the Hypergraph Flow Problem
Let HB = (VB, EB) be the subhypergraph of H = (V, E) that is induced by a corridor B
computed in the bipartition Π2 = (Vi, Vj). In the following, we distinguish between the set of
internal border nodes −→
B := {v ∈ VB ∃e ∈ E : {u, v} ⊆ e ∧ u /∈ VB} and the set of external
border nodes ←−
B := {u /∈ VB ∃e ∈ E : {u, v} ⊆ e ∧ v ∈ VB}. Similarly, we distinguish
between external nets (e ∩ VB = ∅) with no pins inside HB, internal nets (e ∩ VB = e) with
all pins inside HB, and border nets e ∈ I(−→
B ) with some pins inside HB and some pins
outside of HB. We use ←→
A hypergraph flow problem consists of a flow network NB = (VB,EB) derived from HB
and two additional nodes s and t that are connected to some nodes v ∈ VB. Our approach
works with all flow networks presented in Section 3.1. A flow problem has the cut property if
the resulting min-cut bipartition Πf of HB does not increase the (λ − 1)-metric in H. Thus
it has to hold that cut(Πf) ≤ cut(Π2). While external nets are not affected by a max-flow
computation, the max-flow min-cut theorem [21] ensures the cut property for all internal
nets. Border nets however require special attention. Since a border net e is only partially
contained in HB, it will remain connected to the blocks of its external border nodes in H. In
case external border nodes connect e to both Vi and Vj, it will remain a cut net in H even if
it is removed from the cut-set in Πf. It is therefore necessary to "encode" information about
external border nodes into the flow problem.
The KaFFPa Model and its Limitations.
In KaFFPa, this is done by directly connecting
internal border nodes −→
B to s and t. This approach can also be used for hypergraphs. In
the hypergraph flow problem FG, the source s is connected to all nodes S = −→
B ∩ Vi and
all nodes T = −→
B ∩ Vj are connected to t using directed edges with infinite capacity. While
this ensures that FG has the cut property, applying the graph-based model to hypergraphs
unnecessarily restricts the search space. Since all internal border nodes −→
B are connected
to either s or t, every min-cut (S, VB \ S) will have S ⊆ S and T ⊆ VB \ S. The KaFFPa
model therefore prevents all min-cuts in which any non-cut border net (i.e., e ∈ ←→
EB with
λ(e) = 1) becomes part of the cut-set. This restricts the space of possible solutions, since
corridor B was computed such that even a min-cut along either side of the border would
result in a feasible cut in HB. Thus, ideally, all vertices v ∈ B should be able to change
their block as result of an (s, t)-max-flow computation on FG – not only vertices v ∈ B \ −→
B .
This limitation becomes increasingly relevant for hypergraphs with large nets as well as for
partitioning problems with small imbalance ε, since large nets are likely to be only partially
contained in HB and tight balance constraints enforce small B-corridors. While the former
is a problem only for HGP, the latter also applies to GP.
A more flexible Model. We propose a more general model that allows an (s, t)-max-flow
computation to also cut through border nets by exploiting the structure of hypergraph
flow networks. Instead of directly connecting s and t to internal border nodes −→
B and thus
preventing all min-cuts in which these nodes switch blocks, we conceptually extend HB to
contain all external border nodes ←−
EB. The resulting hypergraph is
←−
HB = (VB ∪←−
B ,{e ∈ E e∩ VB 6= ∅}). The key insight now is that by using the flow network
HB and connecting s resp. t to the external border nodes ←−
of ←−
B ∩ Vj, we get
a flow problem that does not lock any node v ∈ VB in its block, since none of these nodes
is directly connected to either s or t. Due to the max-flow min-cut theorem [21], this flow
problem furthermore has the cut property, since all border nets of HB are now internal nets
B and all border nets ←→
B ∩ Vi resp. ←−
T. Heuer, P. Sanders and S. Schlag
XX:9
Figure 2 Comparison of the KaFFPa flow problem FG and our flow problem FH. For clarity the
zoomed in view is based on the Lawler network.
HB instead of HB to achieve this result. For all vertices v ∈ ←−
and all external border nodes ←−
B are locked inside their block. However, it is not necessary
to use ←−
B the flow network of
←−
HB contains paths hs, v, e0i and he00, v, ti that only involve infinite-capacity edges. Therefore
we can remove all nodes v ∈ ←−
B by directly connecting s and t to the corresponding bridging
nodes e0, e00 via infinite-capacity edges without affecting the maximal flow [20]. More precisely,
in the hypergraph flow problem FH, we connect s to all bridging nodes e0 corresponding to
border nets e ∈ ←→
B ∩ Vi and all bridging nodes e00 corresponding to border nets
e ∈ ←→
Single-Pin Border Nets. Furthermore, we model border nets with e ∩ −→
B = 1 more
efficiently. For such a net e, the flow problem contains paths of the form hs, e0, e00, vi or
hv, e0, e00, ti which can be replaced by paths of the form hs, e0, vi or hv, e00, ti with c(e0, v) = ω(e)
resp. c(v, e00) = ω(e). In both cases we can thus remove one bridging node and two infinite-
capacity edges. A comparison of FG and FH is shown in Figure 2.
B ∩ Vj to t using directed, infinite-capacity edges.
EB : e ⊂ ←−
EB : e ⊂ ←−
3.3 Implementation Details
Since KaHyPar is an n-level partitioner, its FM-based local search algorithms are executed
each time a vertex is uncontracted. To prevent expensive recalculations, it therefore uses
a cache to maintain the gain values of FM moves throughout the n-level hierarchy [2]. In
order to combine our flow-based refinement with FM local search, we not only perform the
moves induced by the max-flow min-cut computation but also update the FM gain cache
accordingly.
Since it is not feasible to execute our algorithm on every level of the n-level hierarchy, we use
an exponentially spaced approach that performs flow-based refinements after uncontracting
i = 2j vertices for j ∈ N+. This way, the algorithm is executed more often on smaller
flow problems than on larger ones. To further improve the running time, we introduce the
following speedup techniques:
S1: We modify active block scheduling such that after the first round the algorithm is
only executed on a pair of blocks if at least one execution using these blocks improved
connectivity or imbalance of the partition on previous levels.
ViVjCutB1B2stFlowProblemFGofSandersandSchulz[38]CutViVjB1B2←−B←−BtsConceptual:Actual:∞∞∞∞ω(e)∞∞∞∞∞ω(e)∞∞OurFlowProblemFH∞∞∞∞ω(e)∞∞e(cid:48)e(cid:48)(cid:48)e(cid:48)e(cid:48)(cid:48)e(cid:48)e(cid:48)(cid:48)−→B−→B−→BFlowNetworkNB:−→BActual:ω(e)∞e(cid:48)(cid:48)∞∞∞ω(e)e(cid:48)e(cid:48)(cid:48)HB←−HB XX:10 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Table 1 Overview about different benchmark sets. Set B and set C are subsets of set A.
Source
[28]
[28]
new
[42]
Set A
Set B
Set C
Set D
# DAC ISPD98 Primal Dual
92
477
30
165
25
5
-
15
18
10
5
-
10
5
-
-
Literal
92
30
5
-
SPM Graphs
-
184
-
60
5
-
15
-
92
30
5
-
S2: For all levels except the finest level: Skip flow-based refinement if the cut between
two adjacent blocks is less than ten.
S3: Stop resizing the corridor B if the current (s, t)-cut did not improve the previously
best solution.
4
Experimental Evaluation
We implemented the max-flow min-cut refinement algorithm in the n-level hypergraph
partitioning framework KaHyPar (Karlsruhe Hypergraph Partitioning). The code is written
in C++ and compiled using g++-5.2 with flags -O3 -march=native. The latest version of
the framework is called KaHyPar-CA [28]. We refer to our new algorithm as KaHyPar-MF.
Both versions use the default configuration for community-aware direct k-way partitioning.1
Instances. All experiments use hypergraphs from the benchmark set of Heuer and Sch-
lag [28]2, which contains 488 hypergraphs derived from four benchmark sets: the ISPD98
VLSI Circuit Benchmark Suite [3], the DAC 2012 Routability-Driven Placement Contest [55],
the University of Florida Sparse Matrix Collection [15], and the international SAT Competi-
tion 2014 [9]. Sparse matrices are translated into hypergraphs using the row-net model [12],
i.e., each row is treated as a net and each column as a vertex. SAT instances are converted
to three different representations: For literal hypergraphs, each boolean literal is mapped to
one vertex and each clause constitutes a net [43], while in the primal model each variable
is represented by a vertex and each clause is represented by a net. In the dual model the
opposite is the case [41]. All hypergraphs have unit vertex and net weights.
Table 1 gives an overview about the different benchmark sets used in the experiments.
The full benchmark set is referred to as set A. We furthermore use the representative subset
of 165 hypergraphs proposed in [28] (set B) and a smaller subset consisting of 25 hypergraphs
(set C), which is used to devise the final configuration of KaHyPar-MF. Basic properties of
set C can be found in Table 10 in Appendix C. Unless mentioned otherwise, all hypergraphs
are partitioned into k ∈ {2, 4, 8, 16, 32, 64, 128} blocks with ε = 0.03. For each value of k, a
k-way partition is considered to be one test instance, resulting in a total of 175 instances
for set C, 1155 instances for set B and 3416 instances for set A. Furthermore we use 15
graphs from [42] to compare our flow model FH to the KaFFPa [47] model FG. Table 11 in
Appendix C summarizes the basic properties of these graphs, which constitute set D.
1 https://github.com/SebastianSchlag/kahypar/blob/master/config/km1_direct_kway_sea17.ini
2 The complete benchmark set along with detailed statistics for each hypergraph is publicly available
from http://algo2.iti.kit.edu/schlag/sea2017/.
T. Heuer, P. Sanders and S. Schlag
XX:11
Table 2 Statistics of benchmark set B. We use x to denote mean andex to denote the median.
Type
DAC
ISPD
Primal
Literal
Dual
SPM
# d(v) gd(v)
5
10
30
30
30
60
3.32
4.20
16.29
8.21
2.63
24.78
3.28
4.24
9.97
4.99
2.38
14.15
e
3.37
3.89
2.63
2.63
16.29
26.58
ee
3.35
3.90
2.39
2.39
9.97
15.01
System and Methodology. All experiments are performed on a single core of a machine
consisting of two Intel Xeon E5-2670 Octa-Core processors (Sandy Bridge) clocked at 2.6
GHz. The machine has 64 GB main memory, 20 MB L3- and 8x256 KB L2-Cache and
is running RHEL 7.2. We compare KaHyPar-MF to KaHyPar-CA, as well as to the k-
way (hMetis-K) and the recursive bisection variant (hMetis-R) of hMetis 2.0 (p1) [32, 33],
and to PaToH 3.2 [12]. These HGP libraries were chosen because they provide the best
solution quality [2, 28]. The partitioning results of these tools are already available from
http://algo2.iti.kit.edu/schlag/sea2017/. For each partitioner except PaToH the
results summarize ten repetitions with different seeds for each test instance and report the
arithmetic mean of the computed cut and running time as well as the best cut found. Since
PaToH ignores the random seed if configured to use the quality preset, the results contain
both the result of single run of the quality preset (PaToH-Q) and the average over ten
repetitions using the default configuration (PaToH-D). Each partitioner had a time limit of
eight hours per test instance. We use the same number of repetitions and the same time
limit for our experiments with KaHyPar-MF.
In the following, we use the geometric mean when averaging over different instances
in order to give every instance a comparable influence on the final result.
In order to
compare the algorithms in terms of solution quality, we perform a more detailed analysis
using improvement plots. For each algorithm, these plots relate the minimum connectivity of
KaHyPar-MF to the minimum connectivity produced by the corresponding algorithm on a
per-instance basis. For each algorithm, these ratios are sorted in decreasing order. The plots
use a cube root scale for the y-axis to reduce right skewness [14] and show the improvement
of KaHyPar-MF in percent (i.e., 1−(KaHyPar-MF/algorithm)) on the y-axis. A value below
zero indicates that the partition of KaHyPar-MF was worse than the partition produced by
the corresponding algorithm, while a value above zero indicates that KaHypar-MF performed
better than the algorithm in question. A value of zero implies that the partitions of both
algorithms had the same solution quality. Values above one correspond to infeasible solutions
that violated the balance constraint. In order to include instances with a cut of zero into the
results, we set the corresponding cut values to one for ratio computations.
4.1 Evaluating Flow Networks, Models, and Algorithms
Flow Networks and Algorithms. To analyze the effects of the different hypergraph flow
networks we compute five bipartitions for each hypergraph of set B with KaHyPar-CA using
different seeds. Statistics of the hypergraphs are shown in Table 2. The bipartitions are then
used to generate hypergraph flow networks for a corridor of size B = 25 000 hypernodes
around the cut. Figure 3 (top) summarizes the sizes of the respective flow networks in terms
of number of nodes V and number of edges E for each instance class. The flow networks of
XX:12 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Figure 3 Top: Size of the flow networks when using the Lawler network NL, the Liu-Wong
network NW and our network NOur. Network N 1
Our additionally models single-pin border nets more
efficiently. The dashed line indicates 25 000 nodes. Bottom: Speedup of BK [10] and IBFS [23]
max-flow algorithms over the execution on NL.
primal and literal SAT instances are the largest in terms of both numbers of nodes and edges.
High average vertex degree combined with low average net sizes leads to subhypergraphs HB
containing many small nets, which then induce many nodes and (infinite-capacity) edges in
NL. Dual instances with low average degree and large average net size on the other hand
lead to smaller flow networks. For VLSI instances (DAC, ISPD) both average degree and
average net sizes are low, while for SPM hypergraphs the opposite is the case. This explains
why SPM flow networks have significantly more edges, despite the number of nodes being
comparable in both classes.
As expected, the Lawler-Network NL induces the biggest flow problems. Looking at the
Liu-Wong network NW , we can see that distinguishing between graph edges and nets with
e ≥ 3 pins has an effect for all hypergraphs with many small nets (i.e., DAC, ISPD, Primal,
Literal). While this technique alone does not improve dual SAT instances, we see that the
combination of the Liu-Wong approach and our removal of low degree hypernodes in NOur
reduces the size of the networks for all instance classes except SPM. Both techniques only
have a limited effect on these instances, since both hypernode degrees and net sizes are large
on average. Since our flow problems are based on B-corridor induced subhypergraphs, N 1
Our
additionally models single-pin border nets more efficiently as described in Section 3.2. This
further reduces the network sizes significantly. As expected, the reduction in numbers of
nodes and edges is most pronounced for hypergraphs with low average net sizes because
these instances are likely to contain many single-pin border nets.
To further see how these reductions in network size translate to improved running times
214215216217218219220221222V216217218219220221222223E12345SpeedupoverNL12345SpeedupoverNLBoykov-KolmogorovIBFSNetworkNLNWNOurN1OurNetworkNLNWNOurN1OurNetworkNWNOurN1OurNetworkNWNOurN1OurDACISPDDualPrimalLiteralSPMDACISPDDualPrimalLiteralSPMDACISPDDualPrimalLiteralSPMInstanceClassDACISPDDualPrimalLiteralSPMInstanceClass214215216217218219220221222V216217218219220221222223E12345SpeedupoverNL12345SpeedupoverNLBoykov-KolmogorovIBFSNetworkNLNWNOurN1OurNetworkNLNWNOurN1OurNetworkNWNOurN1OurNetworkNWNOurN1OurDACISPDDualPrimalLiteralSPMDACISPDDualPrimalLiteralSPMDACISPDDualPrimalLiteralSPMInstanceClassDACISPDDualPrimalLiteralSPMInstanceClass T. Heuer, P. Sanders and S. Schlag
XX:13
Table 3 Comparing the KaFFPa flow model FG with our model FH as described in Section 3.2.
The table shows the average improvement of FH over FG (in [%]) for different imbalance parameters
ε on hypergraphs as well as on graphs. All experiments use configuration (+F,-M,-FM).
Hypergraphs
Graphs
ε = 1% ε = 3% ε = 5% ε = 1% ε = 3% ε = 5%
10.5
7.8
5.4
3.9
3.5
11.3
9.1
7.3
5.3
4.1
7.7
7.9
6.9
5.1
3.4
8.1
6.6
3.9
2.3
1.3
7.6
4.8
2.7
1.5
1.2
11.7
11.0
9.9
8.6
7.0
α0
1
2
4
8
16
of max-flow algorithms, we use these networks to create flow problems using our flow model
FH and compute min-cuts using two highly tuned max-flow algorithms, namely the BK-
algorithm3 [10] and the incremental breadth-first search (IBFS) algorithm4 [23]. These
algorithms were chosen because they performed best in preliminary experiments [27]. We
then compare the speedups of these algorithms when executed on NW , NOur, and N 1
Our
to the execution on the Lawler network NL. As can be seen in Figure 3 (bottom) both
algorithms benefit from improved network models and the speedups directly correlate with the
reductions in network size. While NW significantly reduces the running times for Primal and
Literal instances, NOur additionally leads to a speedup for Dual instances. By additionally
considering single-pin border nets, N 1
Our results in an average speedup between 1.52 and 2.21
(except for SPM instances). Since IBFS outperformed the BK algorithm in [27], we use N 1
Our
and IBFS in all following experiments.
Flow Models. We now compare the flow model FG of KaFFPa to our advanced model FH
described in Section 3.2. The experiments summarized in Table 3 were performed using
sets C and D. To focus on the impact of the models on solution quality, we deactivated
KaHyPar's FM local search algorithms and only use flow-based refinement without the most
balanced minimum cut heuristic. The results confirm our hypothesis that FG restricts the
space of possible solutions. For all flow problem sizes and all imbalances tested, FH yields
better solution quality. As expected, the effects are most pronounced for small flow problems
and small imbalances where many vertices are likely to be border nodes. Since these nodes
are locked inside their respective block in FG, they prevent all non-cut border nets from
becoming part of the cut-set. Our model, on the other hand, allows all min-cuts that yield a
feasible solution for the original partitioning problem. The fact that this effect also occurs
for the graphs of set D indicates that our model can also be effective for traditional graph
partitioning. All following experiments are performed using FH.
4.2 Configuring the Algorithm
We now evaluate different configurations of the max-flow min-cut based refinement framework
on set C. In the following, KaHyPar-CA [28] is used as a reference. Since it neither uses
(F)lows nor the (M)ost balanced minimum cut heuristic and only relies on the (FM) algorithm
3 Available from: https://github.com/gerddie/maxflow
4 Available from: http://www.cs.tau.ac.il/~sagihed/ibfs/code.html
XX:14 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
for local search, it is referred to as (-F,-M,+FM). This basic configuration is then successively
extended with specific components. The results of our experiments are summarized in
Table 4 for increasing scaling parameter α0. The table furthermore includes a configuration
Constant128. In this configuration all components are enabled (+F,+M,+FM) and we
perform flow-based refinements every 128 uncontractions. While this configuration is slow, it
is used as a reference point for the quality achievable using flow-based refinement.
Table 4 Different configurations of our flow-based refinement framework for increasing α0. Column
Avg[%] reports the quality improvement relative to the reference configuration (-F,-M,+FM).
(+F,-M,-FM)
α0 Avg[%]
t[s]
−6.09
12.91
1
−3.19
15.75
2
−1.82
20.37
4
−0.85
28.49
8
−0.19
16
43.32
(-F,-M,+FM)
Ref.
(+F,+M,-FM)
t[s]
Avg[%]
−5.60
13.40
−2.07
16.74
−0.19
21.88
30.67
0.98
46.66
1.75
6373.88
13.73
(+F,-M,+FM)
t[s]
Avg[%]
15.03
0.25
0.59
16.98
20.43
0.90
26.51
1.24
1.60
37.51
(+F,+M,+FM)
Avg[%]
0.23
0.73
1.21
1.71
2.21
t[s]
15.16
17.51
21.53
28.68
41.31
Constant128
t[s]
Avg[%]
67.90
0.54
1.11
140.24
269.63
1.66
512.12
2.20
2.76
973.91
The results indicate that only using flows (+F,-M,-FM) as refinement technique is inferior
to localized FM local search in regard to both running time and solution quality. Although
the quality improves with increasing flow problem size (i.e., increasing α0), the average
connectivity is still worse than the reference configuration. Enabling the most balanced
minimum cut heuristic improves partitioning quality. Configuration (+F,+M,-FM) performs
better than the basic configuration for α0 ≥ 8. By combining flows with the FM algorithm
(+F,-M,+FM) we get a configuration that improves upon the baseline configuration even for
small flow problems. However, comparing this variant with (+F,+M,-FM) for α0 = 16, we see
that using large flow problems together with the most balanced minimum cut heuristic yields
solutions of comparable quality. Enabling all components (+F,+M,+FM) and using large flow
problems performs best. Furthermore we see that enabling FM local search slightly improves
the running time for α0 ≥ 8. This can be explained by the fact that the FM algorithm already
produces good cuts between the blocks such that fewer rounds of pairwise flow refinements
are necessary to further improve the solution. Comparing configuration (+F,+M,+FM) with
Constant128 shows that performing flows more often further improves solution quality at
the cost of slowing down the algorithm by more than an order of magnitude. In all further
experiments, we therefore use configuration (+F,+M,+FM) with α0 = 16 for KaHyPar-MF.
This configuration also performed best in the effectiveness tests presented in Appendix A.
While this configuration performs better than KaHyPar-CA, its running time is still more
than a factor of 3 higher.
We therefore perform additional experiments on set B and successively enable the speedup
heuristics described in Section 3.3. The results are summarized in Table 5. Only executing
pairwise flow refinements on blocks that lead to an improvement on previous levels (S1)
reduces the running time of flow-based refinement by a factor of 1.27, while skipping flows
in case of small cuts (S2) results in a further speedup of 1.19. By additionally stopping
the resizing of the flow problem as early as possible (S3), we decrease the running time of
flow-based improvement by a factor of 2 in total, while still computing solutions of comparable
quality. Thus in the comparisons with other systems, all heuristics are enabled.
T. Heuer, P. Sanders and S. Schlag
XX:15
Table 5 Comparison of quality improvement and running times using speedup heuristics. Column
tflow[s] refers to the running time of flow-based refinement, column t[s] to the total partitioning time.
Configuration
KaHyPar-CA
KaHyPar-MF
KaHyPar-MF(S1)
KaHyPar-MF(S1,S2)
KaHyPar-MF(S1,S2,S3)
Avg [%] Min [%]
6820.17
7077.20
2.12
2.47
2.41
2.06
2.05
2.40
2.41
2.06
tflow[s]
-
43.04
33.89
28.52
21.23
t[s]
29.26
72.30
63.15
57.78
50.49
4.3 Comparison with other Systems
Finally, we compare KaHyPar-MF to different state-of-the-art hypergraph partitioners on the
full benchmark set. We exclude the same 194 out of 3416 instances as in [28] because either
PaToH-Q could not allocate enough memory or other partitioners did not finish in time. The
excluded instances are shown in Table 12 in Appendix D. Note that KaHyPar-MF did not
lead to any further exclusions. The following comparison is therefore based on the remaining
3222 instances. As can be seen in Figure 4, KaHyPar-MF outperforms all other algorithms
on all benchmark sets. Comparing the best solutions of KaHyPar-MF to each partitioner
individually across all instances (top left), KaHyPar-MF produced better partitions than
PaToH-Q, PaToH-D, hMetis-K, KaHyPar-CA, hMetis-R for 92.1%, 91.7%, 85.1%, 83.7%,
and 75.6% of the instances, respectively.
Comparing the best solutions of all systems simultaneously, KaHyPar-MF produced the
best partitions for 2427 of the 3222 instances. It is followed by hMetis-R (678), KaHyPar-CA
(388), hMetis-K (352), PaToH-D (154), and PaToH-Q (146). Note that for some instances
multiple partitioners computed the same best solution and that we disqualified infeasible
solutions that violated the balance constraint.
Figure 5 shows that KaHyPar-MF also performs best for different values of k and that
pairwise flow refinements are an effective strategy to improve k-way partitions. As can be
seen in Table 6, the improvement over KaHyPar-CA is most pronounced for hypergraphs
derived from matrices of web graphs and social networks5 and dual SAT instances. While
the former are difficult to partition due to skewed degree and net size distributions, the latter
are difficult because they contain many large nets.
Finally, Table 9 compares the running times of all partitioners. By using simplified flow
networks, highly tuned flow algorithms and several techniques to speed up the flow-based
refinement framework, KaHyPar-MF is less than a factor of two slower than KaHyPar-CA
and still achieves a running time comparable to that of hMetis.
5 Based on the following matrices: webbase-1M, ca-CondMat, soc-sign-epinions, wb-edu, IMDB,
as-22july06, as-caida, astro-ph, HEP-th, Oregon-1, Reuters911, PGPgiantcompo, NotreDame_www,
NotreDame_actors, p2p-Gnutella25, Stanford, cnr-2000.
XX:16 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Figure 4 Min-Cut improvement plots comparing KaHyPar-MF with KaHyPar-CA and other
partitioners for different instance classes.
infeasiblesolutions-80-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DAll(3222Instances)infeasiblesolutions-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DDac(70Instances)infeasiblesolutions-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DIspd98(126Instances)infeasiblesolutions-80-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DPrimal(559Instances)infeasiblesolutions-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DDual(590Instances)infeasiblesolutions-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DLiteral(620Instances)infeasiblesolutions-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%#Instances(λ−1)Improvement[%]AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DSpm(1257Instances)infeasiblesolutions-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%#InstancesAlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-DWebSocial(115Instances) T. Heuer, P. Sanders and S. Schlag
XX:17
Figure 5 Min-Cut improvement plots comparing KaHyPar-MF with KaHyPar-CA and other
partitioners for different values of k.
infeasiblesolutions-80-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]k=2(467Instances)infeasiblesolutions-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%k=4(466Instances)infeasiblesolutions-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]k=8(468Instances)infeasiblesolutions-60-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%k=16(461Instances)infeasiblesolutions-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%(λ−1)Improvement[%]k=32(457Instances)infeasiblesolutions-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%#Instancesk=64(455Instances)infeasiblesolutions-40-20-10-5-1015102040601000%10%20%30%40%50%60%70%80%90%100%#Instances(λ−1)Improvement[%]k=128(448Instances)AlgorithmKaHyPar-CAhMetis-RhMetis-KPaToH-QPaToH-D XX:18 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Table 6 Comparing the best solutions of KaHyPar-MF with the best results of KaHyPar-CA and
other partitioners for different benchmark sets (top) and different values of k (bottom). All values
correspond to the quality improvement of KaHyPar-MF relative to the respective partitioner (in %).
Algorithm
All
2.22
14.40
12.92
11.48
12.06
k = 2
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
DAC
KaHyPar-MF 7542.8816 828.15
KaHyPar-CA
2.80
4.75
7.77
15.24
15.57
k = 4
KaHyPar-MF 1005.76 2985.22
KaHyPar-CA
2.16
17.62
13.66
12.60
10.41
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
1.71
22.25
21.82
14.92
8.54
Min. (λ − 1)
Literal
SPM WebSoc
7478.06
3.91
41.92
40.45
18.45
23.04
2.46
4.17
6.91
14.98
15.17
k = 32
ISPD98 Primal
Dual
5511.40 15 236.13 15 197.60 2927.42
3.33
31.20
21.51
11.44
13.64
k = 64
6010.05
1.74
1.85
16.37
3.88
16.23
4.78
8.36
14.36
9.45
12.47
k = 128
k = 16
9097.31 14 352.34 21 537.33 31 312.48
2.05
8.01
7.81
8.63
11.89
1.92
2.76
2.17
9.53
10.90
k = 8
5805.19
2.51
15.63
12.76
11.81
13.64
2.51
14.29
13.49
11.66
14.50
2.45
11.94
10.62
10.66
12.70
2.16
9.80
9.18
9.77
12.66
Table 7 Comparing the average running times of KaHyPar-MF with KaHyPar-CA and other
hypergraph partitioners for different benchmark sets (top) and different values of k (bottom).
Algorithm
KaHyPar-MF
KaHyPar-CA
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
KaHyPar-MF
KaHyPar-CA
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
All
55.67
31.05
79.23
57.86
5.89
1.22
k = 2
19.75
12.68
27.87
25.47
1.93
0.43
DAC
504.27
368.97
446.36
240.92
28.34
6.45
k = 4
32.89
17.16
51.59
32.27
3.61
0.77
Running Time t[s]
ISPD98 Primal
61.78
32.91
66.25
44.23
6.90
1.12
k = 16
60.38
31.01
91.09
53.41
7.01
1.42
20.83
12.35
29.03
23.18
1.89
0.35
k = 8
47.52
23.88
74.74
42.50
5.44
1.12
Literal
119.51
64.65
142.12
94.89
9.24
1.58
k = 32
78.51
41.69
109.13
74.00
8.40
1.71
Dual
97.22
68.27
200.36
125.55
10.57
2.87
k = 64
100.34
57.35
128.66
109.12
10.06
2.02
SPM WebSoc
27.40
110.15
67.14
13.91
89.69
41.79
111.95
35.95
3.42
4.71
0.77
0.88
k = 128
119.15
76.61
149.34
152.92
11.44
2.29
T. Heuer, P. Sanders and S. Schlag
XX:19
Conclusion
5
We generalize the flow-based refinement framework of KaFFPa [47] from graph to hypergraph
partitioning. We reduce the size of Liu and Wong's hypergraph flow network [40] by
removing low degree hypernodes and exploiting the fact that our flow problems are built
on subhypergraphs of the input hypergraph. Furthermore we identify shortcomings of the
KaFFPa [47] approach that restrict the search space of feasible solutions significantly and
introduce an advanced model that overcomes these limitations by exploiting the structure
of hypergraph flow networks. Lastly, we present techniques to improve the running time
of the flow-based refinement framework by a factor of 2 without affecting solution quality.
The resulting hypergraph partitioner KaHyPar-MF performs better than all competing
algorithms on all instance classes of a large benchmark set and still has a running time
comparable to that of hMetis.
Since our flow problem formulation yields significantly better solutions for both hyper-
graphs and graphs than the KaFFPa [47] approach, future work includes the integration
of our flow model into KaFFPa and the evaluation in the context of a high quality graph
partitioner. Furthermore an approach similar to Yang and Wong [57] could be used as an
alternative to the most balanced minimum cut heuristic and adaptive B-corridor resizing.
We also plan to extend our framework to optimize other objective functions such as cut or
sum of external degrees.
XX:20 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
1
References
P. Agrawal, B. Narendran, and N. Shivakumar. Multi-way partitioning of VLSI circuits.
In Proceedings of 9th International Conference on VLSI Design, pages 393–399, Jan 1996.
2 Y. Akhremtsev, T. Heuer, P. Sanders, and S. Schlag. Engineering a direct k-way hypergraph
In 19th Workshop on Algorithm Engineering and Experiments,
partitioning algorithm.
(ALENEX), pages 28–42, 2017.
3 C. J. Alpert. The ISPD98 Circuit Benchmark Suite. In Proceedings of the 1998 Interna-
tional Symposium on Physical Design, pages 80–85. ACM, 1998.
4 C. J. Alpert, J.-H. Huang, and A. B. Kahng. Multilevel Circuit Partitioning. IEEE Trans-
actions on Computer-Aided Design of Integrated Circuits and Systems, 17(8):655–667, 1998.
5 C. J. Alpert and A. B. Kahng. Recent Directions in Netlist Partitioning: a Survey. Integ-
ration, the VLSI Journal, 19(1–2):1 – 81, 1995.
6 Reid Andersen and Kevin J Lang. An Algorithm for Improving Graph Partitions.
In
Proceedings of the 19th annual ACM-SIAM Symposium on Discrete Algorithms, pages 651–
660. Society for Industrial and Applied Mathematics, 2008.
7 C. Aykanat, B. B. Cambazoglu, and B. Uçar. Multi-level Direct K-way Hypergraph Parti-
tioning with Multiple Constraints and Fixed Vertices. Journal of Parallel and Distributed
Computing, 68(5):609–625, 2008.
8 D. A. Bader, H. Meyerhenke, P. Sanders, and D. Wagner, editors. Proc. Graph Partitioning
and Graph Clustering - 10th DIMACS Implementation Challenge Workshop, volume 588 of
Contemporary Mathematics. AMS, 2013.
9 A. Belov, D. Diepold, M. Heule, and M. Järvisalo. The SAT Competition 2014. http:
//www.satcompetition.org/2014/, 2014.
10 Yuri Boykov and Vladimir Kolmogorov. An Experimental Comparison of Min-Cut/Max-
Flow Algorithms for Energy Minimization in Vision. IEEE Transactions on Pattern Ana-
lysis and Machine Intelligence, 26(9):1124–1137, 2004.
11 T. N. Bui and C. Jones. A Heuristic for Reducing Fill-In in Sparse Matrix Factorization.
In SIAM Conference on Parallel Processing for Scientific Computing, pages 445–452, 1993.
12 Ü. V. Catalyürek and C. Aykanat. Hypergraph-Partitioning-Based Decomposition for Par-
allel Sparse-Matrix Vector Multiplication. IEEE Transactions on Parallel and Distributed
Systems, 10(7):673–693, Jul 1999.
J. Cong and M. Smith. A Parallel Bottom-up Clustering Algorithm with Applications to
Circuit Partitioning in VLSI Design.
In 30th Conference on Design Automation, pages
755–760, June 1993.
14 N. J. Cox. Stata tip 96: Cube roots. Stata Journal, 11(1):149–154(6), 2011. URL: http:
13
//www.stata-journal.com/article.html?article=st0223.
15 T. A. Davis and Y. Hu. The University of Florida Sparse Matrix Collection. ACM Trans-
actions on Mathematical Software, 38(1):1:1–1:25, 2011.
18
16 K. D. Devine, E. G. Boman, R. T. Heaphy, R. H. Bisseling, and Ü. V. Catalyürek. Parallel
In 20th International Conference on
Hypergraph Partitioning for Scientific Computing.
Parallel and Distributed Processing, IPDPS, pages 124–124. IEEE, 2006.
17 W.E. Donath. Logic partitioning. Physical Design Automation of VLSI Systems, pages
65–86, 1988.
J. Edmonds and R. M. Karp. Theoretical Improvements in Algorithmic Efficiency for
Network Flow Problems. Journal of the ACM, 19(2):248–264, 1972.
19 C. Fiduccia and R. Mattheyses. A Linear Time Heuristic for Improving Network Partitions.
In 19th ACM/IEEE Design Automation Conf., pages 175–181, 1982.
20 D. R. Ford and D. R. Fulkerson. Flows in Networks. Princeton University Press, 1962.
21
Lester R Ford and Delbert R Fulkerson. Maximal Flow through a Network. Canadian
Journal of Mathematics, 8(3):399–404, 1956.
T. Heuer, P. Sanders and S. Schlag
XX:21
22 A. V. Goldberg and R. E. Tarjan. A new approach to the maximum-flow problem. Journal
of the ACM, 35(4):921–940, 1988.
23 Andrew Goldberg, Sagi Hed, Haim Kaplan, Robert Tarjan, and Renato Werneck. Maximum
Flows by Incremental Breadth-First Search. Proceedings of 2011 European Symposium on
Algorithms, pages 457–468, 2011.
24 M. Hamann and B. Strasser. Graph Bisection with Pareto-Optimization. In Proceedings
of the Eighteenth Workshop on Algorithm Engineering and Experiments, ALENEX 2016,
Arlington, Virginia, USA, January 10, 2016, pages 90–102, 2016.
S. Hauck and G. Borriello. An Evaluation of Bipartitioning Techniques. IEEE Transactions
on Computer-Aided Design of Integrated Circuits and Systems, 16(8):849–866, Aug 1997.
26 B. Hendrickson and R. Leland. A Multi-Level Algorithm For Partitioning Graphs. SC
25
Conference, 0:28, 1995.
27 T. Heuer. High Quality Hypergraph Partitioning via Max-Flow-Min-Cut Computations.
Master's thesis, KIT, 2018.
28 T. Heuer and S. Schlag. Improving Coarsening Schemes for Hypergraph Partitioning by
Exploiting Community Structure. In 16th International Symposium on Experimental Al-
gorithms, (SEA), page 21:1–21:19, 2017.
30
29 T. C. Hu and K. Moerder. Multiterminal Flows in a Hypergraph. In T.C. Hu and E.S.
Kuh, editors, VLSI Circuit Layout: Theory and Design, chapter 3, pages 87–93. IEEE
Press, 1985.
E. Ihler, D. Wagner, and F. Wagner. Modeling Hypergraphs by Graphs with the Same
Mincut Properties. Inf. Process. Lett., 45(4):171–175, 1993.
I. Kabiljo, B. Karrer, M. Pundir, S. Pupyrev, A. Shalita, Y. Akhremtsev, and Presta.
A. Social Hash Partitioner: A Scalable Distributed Hypergraph Partitioner. PVLDB,
10(11):1418–1429, 2017.
31
32 G. Karypis, R. Aggarwal, V. Kumar, and S. Shekhar. Multilevel Hypergraph Partitioning:
Applications in VLSI Domain. IEEE Transactions on Very Large Scale Integration VLSI
Systems, 7(1):69–79, 1999.
33 G. Karypis and V. Kumar. Multilevel K-way Hypergraph Partitioning. In Proceedings of
the 36th ACM/IEEE Design Automation Conference, pages 343–348. ACM, 1999.
34 B. W. Kernighan and S. Lin. An Efficient Heuristic Procedure for Partitioning Graphs.
The Bell System Technical Journal, 49(2):291–307, Feb 1970.
35 K. Lang and S. Rao. A Flow-Based Method for Improving the Expansion or Conductance
of Graph Cuts. In Proceedings of 10th International IPCO Conference, volume 4, pages
325–337. Springer, 2004.
E. Lawler. Cutsets and Partitions of Hypergraphs. Networks, 3(3):275–285, 1973.
E. Lawler. Combinatorial Optimization : Networks and Matroids. Holt, Rinehart, and
Whinston, 1976.
36
37
38 T. Lengauer. Combinatorial Algorithms for Integrated Circuit Layout. John Wiley & Sons,
Inc., 1990.
J. Li, J. Lillis, and C. K. Cheng. Linear decomposition algorithm for VLSI design ap-
plications. In Proceedings of IEEE International Conference on Computer Aided Design
(ICCAD), pages 223–228, Nov 1995.
39
40 H. Liu and D. F. Wong. Network-Flow-Based Multiway Partitioning with Area and Pin
IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Constraints.
Systems, 17(1):50–59, Jan 1998.
Z. Mann and P. Papp. Formula partitioning revisited. In Daniel Le Berre, editor, POS-14.
Fifth Pragmatics of SAT workshop, volume 27 of EPiC Series in Computing, pages 41–56.
EasyChair, 2014.
41
XX:22 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
42 H. Meyerhenke, P. Sanders, and C. Schulz. Partitioning Complex Networks via Size-
In 13th International Symposium on Experimental Algorithms,
Constrained Clustering.
(SEA), pages 351–363, 2014.
46
44
45
43 D. A. Papa and I. L. Markov. Hypergraph Partitioning and Clustering. In T. F. Gonza-
lez, editor, Handbook of Approximation Algorithms and Metaheuristics. Chapman and
Hall/CRC, 2007.
Jean-Claude Picard and Maurice Queyranne. On the Structure of all Minimum Cuts in a
Network and Applications. Combinatorial Optimization II, pages 8–16, 1980.
Joachim Pistorius and Michel Minoux. An Improved Direct Labeling Method for the
Max–Flow Min–Cut Computation in Large Hypergraphs and Applications. International
Transactions in Operational Research, 10(1):1–11, 2003.
L. A. Sanchis. Multiple-way Network Partitioning. IEEE Trans. on Computers, 38(1):62–
81, 1989. doi:10.1109/12.8730.
P. Sanders and C. Schulz. Engineering Multilevel Graph Partitioning Algorithms. In 19th
European Symposium on Algorithms, volume 6942 of LNCS, pages 469–480. Springer, 2011.
S. Schlag, V. Henne, T. Heuer, H. Meyerhenke, P. Sanders, and C. Schulz. k-way Hy-
pergraph Partitioning via n-Level Recursive Bisection.
In 18th Workshop on Algorithm
Engineering and Experiments (ALENEX), pages 53–67, 2016.
49 D. G. Schweikert and B. W. Kernighan. A Proper Model for the Partitioning of Electrical
Circuits. In Proceedings of the 9th Design Automation Workshop, DAC, pages 57–62. ACM,
1972.
50 A. Trifunovic. Parallel Algorithms for Hypergraph Partitioning. PhD thesis, University of
47
48
London, 2006.
51 A. Trifunović and W. J. Knottenbelt. Parallel Multilevel Algorithms for Hypergraph Par-
titioning. Journal of Parallel and Distributed Computing, 68(5):563 – 581, 2008.
52 Ü. V. Çatalyürek and M. Deveci and K. Kaya and B. Uçar. UMPa: A multi-objective,
multi-level partitioner for communication minimization. In Bader et al. [8], pages 53–66.
53 B. Uçar and C. Aykanat. Encapsulating Multiple Communication-Cost Metrics in Parti-
tioning Sparse Rectangular Matrices for Parallel Matrix-Vector Multiplies. SIAM Journal
on Scientific Computing, 25(6):1837–1859, 2004.
54 B. Vastenhouw and R. H. Bisseling. A Two-Dimensional Data Distribution Method for
Parallel Sparse Matrix-Vector Multiplication. SIAM Review, 47(1):67–95, 2005.
56
55 N. Viswanathan, C. Alpert, C. Sze, Z. Li, and Y. Wei. The dac 2012 routability-driven
placement contest and benchmark suite. In Proceedings of the 49th Annual Design Auto-
mation Conference, DAC '12, pages 774–782. ACM, 2012.
S. Wichlund. On multilevel circuit partitioning.
Computer-aided Design, ICCAD, pages 505–511. ACM, 1998.
In 1998 International Conference on
57 H. H. Yang and D. F. Wong. Efficient Network Flow Based Min-Cut Balanced Partition-
ing. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems,
15(12):1533–1540, 1996.
T. Heuer, P. Sanders and S. Schlag
XX:23
Effectiveness Tests
A
To evaluate the effectiveness of our configurations presented in Section 4.2 we give each
configuration the same time to compute a partition. For each instance (hypergraph, k),
we execute each configuration once and note the largest running time tH,k. Then each
configuration gets time 3tH,k to compute a partition (i.e., we take the best partition out of
several repeated runs). Whenever a new run of a partition would exceed the largest running
time, we perform the next run with a certain probability such that the expected running time
is 3tH,k. The results of this procedure, which was initially proposed in [47], are presented
in Table 8. We see that the combinations of flow-based refinement and FM local search
perform better than repeated executions of the baseline configuration (-F,-M,+FM). The
most effective configuration is (+F,+M,-FM) with α0 = 16, which was chosen as the default
configuration for KaHyPar-MF.
Table 8 Results of the effectiveness test for different configurations of our flow-based refinement
framework for increasing α0. The quality in column Avg[%] is relative to the baseline configuration
(-F,-M,+FM).
Config
α0
1
2
4
8
16
Ref.
(+F,-M,-FM)
(+F,+M,-FM)
(+F,-M,+FM)
(+F,+M,+FM)
Avg[%]
−6.06
−3.15
−1.89
−0.87
−0.29
(-F,-M,+FM)
Avg[%]
−5.52
−2.06
−0.19
0.96
1.66
6377.15
Avg[%]
0.23
0.55
0.86
1.20
1.52
Avg[%]
0.24
0.73
1.20
1.69
2.17
XX:24 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
B
Average Connectivity Improvement
Table 9 Comparing the average solution quality of KaHyPar-MF with the average results of
KaHyPar-CA and other partitioners for different benchmark sets (top) and different values of
k (bottom). All values correspond to the quality improvement of KaHyPar-MF relative to the
respective partitioner (in %).
Algorithm
All
2.44
13.61
13.23
8.67
14.35
k = 2
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
DAC
KaHyPar-MF 7782.5117 476.91
3.04
KaHyPar-CA
3.53
7.82
11.97
19.21
k = 4
KaHyPar-MF 1057.93 3130.20
2.57
KaHyPar-CA
15.92
15.15
8.36
14.24
hMetis-R
hMetis-K
PaToH-Q
PaToH-D
2.27
21.38
21.63
10.51
13.26
Avg. (λ − 1)
Literal
SPM WebSoc
7783.87
3.84
41.32
40.14
15.12
24.18
2.66
2.72
8.49
11.78
17.72
k = 32
ISPD98 Primal
Dual
5643.99 15 863.75 15 770.06 3038.55
3.39
30.28
22.26
8.08
15.47
k = 64
6143.73
2.07
2.03
16.42
2.03
16.33
3.89
6.33
10.83
11.43
15.45
k = 128
k = 16
9362.55 14 693.96 21 893.59 31 706.57
2.05
7.80
7.83
7.48
12.66
2.16
1.60
1.37
7.35
13.11
k = 8
6032.58
2.80
14.47
13.61
8.35
16.07
2.68
13.63
13.49
9.09
16.59
2.48
11.35
10.52
8.54
13.87
2.24
9.63
9.30
8.27
13.61
T. Heuer, P. Sanders and S. Schlag
XX:25
C
Properties of Benchmark Sets
Table 10 Basic properties of our parameter tuning benchmark set. The number of pins is denoted
with p.
Class
ISPD
Dual
Literal
Primal
SPM
Hypergraph
ibm06
ibm07
ibm08
ibm09
ibm10
6s9
6s133
6s153
dated-10-11-u
dated-10-17-u
6s133
6s153
aaai10-planning-ipc5
dated-10-11-u
atco_enc2_opt1_05_21
6s153
aaai10-planning-ipc5
hwmcc10-timeframe
dated-10-11-u
atco_enc2_opt1_05_21
mult_dcop_01
vibrobox
RFdevice
mixtank_new
laminar_duct3D
n
32 498
45 926
51 309
53 395
69 429
100 384
140 968
245 440
629 461
1 070 757
96 430
171 292
107 838
283 720
112 732
85 646
53 919
163 622
141 860
56 533
25 187
12 328
74 104
29 957
67 173
m
34 826
48 117
50 513
60 902
75 196
34 317
48 215
85 646
141 860
229 544
140 968
245 440
308 235
629 461
526 872
245 440
308 235
488 120
629 461
526 872
25 187
12 328
74 104
29 957
67 173
p
128 182
175 639
204 890
222 088
297 567
234 228
328 924
572 692
1 429 872
2 471 122
328 924
572 692
690 466
1 429 872
2 097 393
572 692
690 466
1 138 944
1 429 872
2 097 393
193 276
342 828
365 580
1 995 041
3 833 077
XX:26 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
Table 11 Basic properties of the graph instances.
Graph
p2p-Gnutella04
wordassociation-2011
PGPgiantcompo
email-EuAll
as-22july06
soc-Slashdot0902
loc-brightkite
enron
loc-gowalla
coAuthorsCiteseer
wiki-Talk
citationCiteseer
coAuthorsDBLP
cnr-2000
web-Google
m
n
29 215
6 405
63 788
10 617
24 316
10 680
60 260
16 805
48 436
22 963
379 445
28 550
212 945
56 739
254 449
69 244
950 327
196 591
227 320
814 134
232 314 ≈1.5M
268 495 ≈1.2M
299 067
977 676
325 557 ≈2.7M
356 648 ≈2.1M
T. Heuer, P. Sanders and S. Schlag
XX:27
D Excluded Instances
Table 12 Instances excluded from the full benchmark set evaluation. Note that using flow-based
refinements did not lead to any further exclusions.
Hypergraph
Primal
10pipe-q0-k
11pipe-k
11pipe-q0-k
9dlx-vliw-at-b-iq3
9vliw-m-9stages-iq3-C1-bug7
9vliw-m-9stages-iq3-C1-bug8
blocks-blocks-37-1.130-
NOTKNOWN
openstacks-p30-3.085-SAT
openstacks-sequencedstrips-nonadl-
nonnegated-os-sequencedstrips-
p30-3.025-NOTKNOWN
openstacks-sequencedstrips-nonadl-
nonnegated-os-sequencedstrips-
p30-3.085-SAT
transport-transport-city-
sequential-25nodes-1000size-
3degree-100mindistance-3trucks-
10packages-2008seed.050-
NOTKNOWN
velev-vliw-uns-2.0-uq5
velev-vliw-uns-4.0-9
Literal
11pipe-k
9vliw-m-9stages-iq3-C1-bug7
9vliw-m-9stages-iq3-C1-bug8
blocks-blocks-37-1.130
Dual
10pipe-q0-k
11pipe-k
11pipe-q0-k
9dlx-vliw-at-b-iq3
9vliw-m-9stages-iq3-C1-bug7
9vliw-m-9stages-iq3-C1-bug8
blocks-blocks-37-1.130-
NOTKNOWN
E02F20
E02F22
q-query-3-L100-coli.sat
q-query-3-L150-coli.sat
q-query-3-L200-coli.sat
q-query-3-L80-coli.sat
transport-transport-city-
sequential-25nodes-1000size-
3degree-100mindistance-3trucks-
10packages-2008seed.030-
NOTKNOWN
velev-vliw-uns-2.0-uq5
velev-vliw-uns-4.0-9
2
4
(cid:3) (cid:3)
(cid:3) (cid:3)
(cid:3) (cid:3)
(cid:3) (cid:3)
4 4
4 4
(cid:3) (cid:3)
(cid:3) (cid:3)
(cid:3) (cid:3)
(cid:3) (cid:3)
(cid:3)
(cid:3) (cid:3)
(cid:3) (cid:3)
4 4
4 4
(cid:3)
8
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
16
(cid:3)
(cid:3)
(cid:3)
(cid:3)
4
4
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
32
(cid:3)
(cid:3)
(cid:3)
(cid:3)
H4
H4
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
64
(cid:3)
(cid:3)
(cid:3)
(cid:3)
H4
H4
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
128
(cid:3)
H(cid:3)
(cid:3)
(cid:3)
H4
H4
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
H
H
H
GH4 GH4 GH4 GH(cid:3)4 GH(cid:3)4
GH4 GH4 GH4 GH(cid:3)4 GH(cid:3)4
(cid:3)
(cid:3)
(cid:3)
(cid:3)
(cid:3)
H
H4
4
H4
4 H4
4
H4
H4
4 GH4 GH4 GH4 GH4 GH4
4 GH4 GH4 GH4 GH4 GH4
4
H4
4
H
GH
GH
GH
GH
H
4
4
4
4
4
4
4
4
4
H4
H4
H4
4
GH4
GH4
GH4
H
H
4
4
4
4
4
4
4
XX:28 Network Flow-Based Refinement for Multilevel Hypergraph Partitioning
SPM
192bit
appu
ESOC
human-gene2
IMDB
kron-g500-logn16
Rucci1
sls
Trec14
(cid:3)
(cid:3) (cid:3)
4
(cid:3) (cid:3)
4
(cid:3)
(cid:3)
4
4
H(cid:3)
(cid:3)
H4
4
4
(cid:3)
H(cid:3)
H
H(cid:3)
H4
4
H4
H(cid:3)
H
(cid:3)
H4
4
H4
H(cid:3)
H
4 : KaHyPar-CA exceeded time limit
G :
H :
(cid:3) :
hMetis-R exceeded time limit
hMetis-K exceeded time limit
PaToH-Q memory allocation error
|
1704.07291 | 1 | 1704 | 2017-04-21T04:26:23 | Minimal Controllability of Conjunctive Boolean Networks is NP-Complete | [
"cs.DS",
"cs.CC",
"eess.SY",
"math.OC",
"q-bio.MN"
] | Given a conjunctive Boolean network (CBN) with $n$ state-variables, we consider the problem of finding a minimal set of state-variables to directly affect with an input so that the resulting conjunctive Boolean control network (CBCN) is controllable. We give a necessary and sufficient condition for controllability of a CBCN; an $O(n^2)$-time algorithm for testing controllability; and prove that nonetheless the minimal controllability problem for CBNs is NP-hard. | cs.DS | cs | Minimal Controllability of Conjunctive Boolean
Networks is NP-Complete
Eyal Weiss, Michael Margaliot and Guy Even
1
7
1
0
2
r
p
A
1
2
]
S
D
.
s
c
[
1
v
1
9
2
7
0
.
4
0
7
1
:
v
i
X
r
a
Abstract
Given a conjunctive Boolean network (CBN) with n state-variables, we consider the problem of finding a
minimal set of state-variables to directly affect with an input so that the resulting conjunctive Boolean control
network (CBCN) is controllable. We give a necessary and sufficient condition for controllability of a CBCN;
an O(n2)-time algorithm for testing controllability; and prove that nonetheless the minimal controllability problem
for CBNs is NP-hard.
Logical systems, controllability, Boolean control networks, computational complexity, minimum dominating
set, systems biology.
Index Terms
I. INTRODUCTION
Many modern networked systems include a large number of nodes (or state-variables). Examples range
from the electric grid to complex biological processes. If the system includes control inputs then a natural
question is whether the system is controllable, that is, whether the control authority is powerful enough to
steer the system from any initial condition to any desired final condition. Controllability is an important
property of control systems, and it plays a crucial role in many control problems, such as stabilization of
unstable systems by feedback, and optimal control [1].
If the system is not controllable (and in particular if there are no control inputs) then an important
problem is what is the minimal number of control inputs that should be added to the network so that
it becomes controllable. This calls for finding the key locations within the system such that controlling
them allows driving the entire system to any desired state. This problem is interesting both theoretically
and for applications, as in many real-world systems it is indeed possible to add control actuators, but
naturally this may be timely and costly, so minimizing the number of added controls is desirable.
Several recent papers studied minimal controllability in networks with a linear and time-invariant (LTI)
dynamics (see, e.g. [2], [3] and the references therein). In particular, Olshevsky [2] considered the following
problem. Given the n-dimensional LTI system xi(t) = Pn
j=1 aijxj(t), i = 1, . . . , n, determine a minimal
set of indices I ⊆ {1, . . . , n}, such that the modified system:
xi(t) =
xi(t) =
n
X
j=1
n
X
j=1
aijxj(t) + ui(t),
aijxj(t),
i ∈ I,
i 6∈ I,
is controllable. Olshevsky [2] showed, using a reduction to the minimum hitting set problem, that this
problem is NP-hard (in the number of state-variables n). For a general survey on the computational
complexity of various problems in systems and control theory, see [4].
Research supported in part by a research grant from the Israel Science Foundation (ISF grant 410/15).
The authors are with the School of Electrical Engineering-Systems, Tel Aviv University, Israel 69978. Corresponding author: Prof. Michael
Margaliot, School of Electrical Engineering-Systems and the Sagol School of Neuroscience, Tel Aviv University, Israel 69978. Homepage:
www.eng.tau.ac.il/michaelm
Email: [email protected]
Boolean control networks (BCNs) are discrete-time dynamical systems with Boolean state-variables
and Boolean control inputs. A BCN without inputs is called a Boolean network (BN). BCNs date back to
the early days of digital switching networks and neural network models with on-off type neurons. They
have also been used to model many other important phenomena such as social networks (see, e.g. [5],
[6]), the spread of epidemics [7], etc.
2
More recently, BCNs have been extensively used to model biological processes and networks where
the possible set of states is assumed to be finite (see, e.g. [8], [9], [10]). For example, in gene regulation
networks one may assume that each gene may be either "on" or "off" (i.e. expressed or not expressed).
Then the state of each gene can be modeled as a state-variable in a BN and the interactions between
the genes (e.g., via the proteins that they encode) determine the Boolean update function for each state-
variable.
A BCN with state-vector X(k) = (cid:2)X1(k)
is said to be controllable if for any pair of
states a, b ∈ {0, 1}n there exists an integer N ≥ 0 and a control sequence u(0), . . . , u(N − 1) steering the
state from X(0) = a to X(N) = b.
. . . Xn(k)(cid:3)′
A natural representation of a BCN is via its graph of states G = (V, E), where the vertices V =
{1, . . . , 2n} represent all the possible 2n states, and a directed edge e = (vi → vj) ∈ E means that there
is a control such that X(k) = vi and X(k + 1) = vj. Then clearly a BCN is controllable if and only
if G is strongly connected [11]. Since testing strong connectivity of a digraph takes linear time in the
number of its vertices and directed edges (see, e.g. [12]), one may expect that verifying controllability of
a general BCN is intractable.
Akutsu et al. [13] showed, using a reduction to the 3SAT problem, that determining if there exists a
control sequence steering a BCN between two given states is NP-hard, and that this holds even for BCNs
with restricted network structures. This implies in particular that verifying controllability is NP-hard. Of
course, it is still possible that the problems of verifying controllability and finding the minimal number
of controls needed to make a BN controllable are tractable in some special classes of Boolean networks.
An important special class of BNs are those comprised of nested canalyzing functions (NCFs) [14]. A
Boolean function is called canalyzing if there exists a certain value, called the canalyzing value, such that
any input with this value uniquely determines the output of the function regardless of the other variables.
For example, 0 is a canalyzing value for the function AND, as AND(0, x1, . . . , xk) = 0 for any xi ∈ {0, 1}.
BNs with nested canalyzing functions are often used to model genetic networks [15], [16], [17].
Here, we consider the subclass consisting of those NCFs that are constructed only with the AND
operator, the conjunctive functions. A BN is called a conjunctive Boolean network (CBN) if every update
function includes only AND operations, i.e., the state-variables satisfy an equation of the form:
Xi(k + 1) =
n
Y
j=1
(Xj(k))ǫji,
i = 1, . . . , n,
(1)
where ǫji ∈ {0, 1} for all i, j.
Recall that a BN is called a disjunctive Boolean network (DBN) if every update function includes only
OR operations. By applying De Morgan Law's, one can reduce DBNs to CBNs, so all the results in this
note hold also for DBNs.
A CBN with n state-variables can be represented by its dependency graph (or wiring diagram) that has n
vertices corresponding to the Boolean state-variables. There is a directed edge (i → j) if Xi(k) appears
in the update function of Xj(k + 1). That is, the dependency graph encodes the variable dependencies
in the update functions. We will assume from here on that none of the update functions is constant,
so every vertex in the dependency graph has a positive in-degree. In this case, there is a one-to-one
correspondence between the CBN and its dependency graph, and this allows a graph-theoretic analysis
of the CBN. For example, the problem of characterizing all the periodic orbits of a CBN with a strongly
connected dependency graph has been solved in [18], and the robustness of these orbits has been studied
in [19].
3
In the context of modeling gene regulation, CBNs encode synergistic regulation of a gene by several
transcription factors [18], and there is increasing evidence that this type of mechanism is common in
regulatory networks [20], [21], [22].
Here, we consider the following problem.
Problem 1. Given a CBN with n state-variables, suppose that for any i ∈ {1, . . . , n} we can replace
the update function of Xi(k + 1) by an independent Boolean control Ui(k). Determine a minimal1 set of
indices I ⊆ {1, . . . , n}, such that the modified system:
Xi(k + 1) = Ui(k),
n
Xi(k + 1) =
Y
(Xj(k))ǫji,
j=1
i ∈ I,
i 6∈ I,
(2)
is controllable.
We refer to a BCN in the form (2) as a conjunctive Boolean control network (CBCN).
Problem 1 is important because it calls for finding a minimal set of key variables in the CBN such that
controlling them makes the system controllable. Of course, an efficient algorithm for solving this problem
must encapsulate an efficient algorithm for testing controllability of a CBCN.
Example 1. Consider Problem 1 for the CBN
Suppose that we replace the update function for X2(k) by a control U2(k) so that the network becomes:
X1(k + 1) = X2(k),
X2(k + 1) = X1(k)X2(k).
X1(k + 1) = X2(k),
X2(k + 1) = U2(k).
This CBCN is clearly controllable. Indeed, given a desired final state s = (cid:2)s1 s2(cid:3)′ ∈ {0, 1}2, the control
sequence U2(0) = s1, U2(1) = s2, steers the CBCN from an arbitrary initial condition X1(0), X2(0)
to (cid:2)X1(2) X2(2)(cid:3)′
= s. Thus, in this case a solution to Problem 1 is to replace the update function
of X2 by a control.
The main contributions of this note are:
1) a necessary and sufficient condition for the controllability of a CBCN;
2) a polynomial-time algorithm for determining whether a CBCN is controllable (more specifically the
time complexity of this algorithm is O(n2), where n is the number of state-variables in the BCN);
3) a proof that Problem 1 is NP-hard.
Together, these results imply that checking the controllability of a given CBCN is "easy", yet there does
not exist a polynomial-time algorithm for solving Problem 1 (unless P=NP).
The next section reviews definitions and notations from graph theory that will be used later on.
Section III describes our main results. Section IV concludes and presents several directions for further
research.
II. PRELIMINARIES
Let G = (V, E) be an undirected graph, where V is the set of vertices, and E is the set of edges. If
two vertices vi, vj are connected by an edge then we denote this edge by eij or by (vi, vj), and say that vi
and vj are neighbors. The set of neighbors of vi is denoted by N (vi), and the degree of vi is N (vi).
1In computer science, this is usually called a minimum cardinality set rather than a minimal set. We follow the terminology used in control
theory.
4
A dominating set for G is a subset D of V such that every vertex in V \D has at least one neighbor
in D.
Problem 2 (Dominating set problem). Given a graph G = (V, E) and a positive integer k ≤ V , does
there exist a dominating set D of V with D ≤ k?
This is known to be an NP-complete decision problem (see, e.g. [23]).
Let G = (V, E) be a directed graph (digraph), with V the set of vertices, and E the set of directed
edges (arcs). Let ei→j (or (vi → vj)) denote the arc from vi to vj. When such an arc exists, we say that vi
is an in-neighbor of vj, and vj as an out-neighbor of vi. The set of in-neighbors and out-neighbors of a
vertex vi is denoted by Nin(vi) and Nout(vi), respectively. The in-degree and out-degree of vi are Nin(vi)
and Nout(vi), respectively.
Let vi and vj be two vertices in V . A walk from vi to vj, denoted by wij, is a sequence: vi0vi1 . . . viq ,
with vi0 = vi, viq = vj, and eik→ik+1 ∈ E for all k ∈ {0, 1, . . . q − 1}. A simple path is a walk with
pairwise distinct vertices. We say that vi is reachable from vj if there exists a simple path from vj to vi.
A closed walk is a walk that starts and finishes at the same vertex. A closed walk is called a cycle if all
the vertices in the walk are distinct, except for the start-vertex and the end-vertex. A strongly connected
digraph is a digraph for which every vertex in the graph is reachable from any other vertex in the graph.
Recall that given the CBN (1), the associated dependency graph is a digraph G = (V, E) with n
vertices, such that ei→j ∈ E if and only if ǫij = 1. A CBN is uniquely determined by its dependency
graph, and for this reason we interchangeably refer to the ith state-variable in the CBN and the ith vertex
in its dependency graph. We extend the definition of a dependency graph to a CBCN in a natural way:
upon replacing the update equation for Xi(k + 1) to Xi(k + 1) = Ui(k) we remove all the arcs pointing
to vi, introduce a new vertex vUi for the new control input, and add an arc eUi→i.
An m-layer graph is a digraph G = (V, E) for which each layer Lk, k = 1, . . . , m, is a subset of V ,
every vertex in V belongs to a single layer, and any arc ei→j ∈ E is such that vi ∈ Lk and vj ∈ Lk+1,
for some k ∈ {1, 2, . . . , m − 1}.
A. Complexity Analysis
III. MAIN RESULTS
We begin by analyzing the computational complexity of Problem 1. We will prove a hardness result
for a CBCN whose dependency graph is a 3-layer graph. Our first result uses the special structure of
this CBCN to provide a simple necessary and sufficient condition for controllability. We will later use
this condition to relate controllability analysis for this CBCN to the dominating set problem.
Lemma 1. Consider a CBCN with a dependency graph that is a 3-layer graph satisfying: every vertex in
layer-1 is a control input to a vertex in layer-2, and every vertex in layer-2 has its own control input (in
layer-1). This CBCN is controllable if and only if every vertex in layer-3 has an in-neighbor (in layer-2)
with out-degree equal to one.
Proof ofLemma1. Assume that the CBCN is controllable. Seeking a contradiction, suppose that there
exists a vertex vi in layer 3 that has no in-neighbor with out-degree one. Let Xi denote the corresponding
state-variable. If vi has zero in-neighbors then Xi(k) is constant, contradicting the assumption of control-
lability. Hence vi must have at least one in-neighbor and each of them with out-degree greater than one.
From the controllability of the CBCN, it follows that for any initial state X(0), there exists a time T ≥ 0
and a control sequence {u(0), . . . , u(T − 1)} steering the CBCN to the final state:
Xi(T ) = 0,
Xj(T ) = 1 for all j 6= i.
(3)
In other words, the state-variables corresponding to all the nodes in layers 2 and 3, except for Xi, are
one at time T . But for (3) to hold at least one of the in-neighbors of vi must be zero at time T − 1. Since
5
zero is the canalyzing value and every in-neighbor of vi has out-degree greater than one, there exists a
state variable Xq, q 6= i, such that Xq(T ) = 0. A contradiction.
To prove the converse implication, suppose that every vertex in layer 3 of the CBCN has an in-neighbor
(in layer 2) with out-degree equal to one. We need to prove that the CBCN is controllable. Denote the
nodes in layer 3 by w1, . . . , wq. For any i ∈ {1, . . . , q} the node wi in layer 3 has an in-neighbor vi in
layer 2 such that vi has out degree one. Let p ≥ q denote the number of nodes in layer 2, so that the
nodes in layer 2 are v1, . . . , vq, vq+1, . . . , vp, and their corresponding controls in layer 1 are u1, . . . , up.
Fix arbitrary a ∈ {0, 1}p and b ∈ {0, 1}q. We will show that for any initial condition there exists a control
sequence that steers the network to the state v(4) = a and w(4) = b, where v = [v1 . . . vp]′, w = [w1 . . . wq]′,
thus proving controllability. In time steps 0 and 1, set all the inputs to one. Then vi(2) = wj(2) = 1
for all i, j. Now let ui(2) = bi for all i ∈ {1, . . . , q}, and ui(2) = 1 for all i > q. Then vi(3) = bi
for all i ∈ {1, . . . , q}, and vi(3) = 1 for all i > q. Finally, let ui(3) = ai for all i ∈ {1, . . . , p}.
Then wi(4) = vi(3) = bi for all i ∈ {1, . . . , q} (because AND(1, z) = z for all z ∈ {0, 1}) and vi(4) = ai
for all i ∈ {1, . . . , p}. This completes the proof of Lemma 1.
We are now ready to present our main complexity result.
Theorem 2. The decision version of Problem 1 is NP-hard.
Proofof Thm.2. Given an undirected graph G = (V, E), and a positive integer k, consider Problem 2.
We will show that we can solve this instance of the dominating set problem by solving a minimal
controllability problem for a 3-layer CBCN with dependency graph G = ( V , E) constructed as follows.
Define L2 := E, L3 := V , and let V := L2 S L3. In other words, G has E + V nodes. The nodes
in L3 are denoted v1, . . . , vs, where s := V and those in L2 are denoted by euv. The set E includes two
sets of directed edges. First, every euv ∈ L2 induces an arc (euv → euv) (i.e., a self loop). Second, every
edge e ∈ E induces two directed edges in E: (euv → v) and (euv → u). Thus, E = 3E (Example 2
below demonstrates this construction). Note that this construction is polynomial in V , E.
Consider the CBN induced by G as a dependency graph, and the minimal controllability problem for
this CBN. A solution for this problem includes adding controls to some of the state-variables. Since every
node in L2 has a self-loop in G, a control input must be added to each node e ∈ L2. Denoting by L1
the set of nodes corresponding to the control inputs added to the CBN, it is evident that we obtained a
3-layer graph. The solution to the controllability problem may also add control inputs to nodes in L3.
Let Y ⊆ L3 denote the set of these nodes. Note that Y is also a set of nodes in the original graph G. We
require the following result.
Lemma 3. The set Y is a minimum dominating set of G.
Proof of Lemma 3. Let ¯G denote the dependency graph of the controllable CBCN obtained by solving
the minimal controllability problem described above. Define a new dependency graph ¯G′ by removing
the nodes in Y , their adjacent directed edges, and the nodes of the control inputs added to them in ¯G.
Then ¯G′ is a 3-layer controllable CBCN: the first layer consists of the controls that were added to the
nodes in L2 (namely the nodes of L1), the second layer is composed of the nodes in L2, and the third is
composed of the nodes in V \ Y . It follows from Lemma 1 that for every vertex v ∈ (V \ Y ) there exists a
vertex e ∈ L2 such that e is an in-neighbor of v and e is not an in-neighbor of any other vertex in V \ Y .
Thus, e must be an in-neighbor of a node in Y . We conclude that Y is a dominating set of G. Since the
construction solves the minimal controllability problem, it is clear that Y is a minimum dominating set
of G.
We now consider two cases. If Y ≤ k then we found a dominating set with cardinality ≤ k and thus
the answer to Problem 2 for the given instance is yes. If Y > k then the argument above implies that
there does not exist a dominating set with cardinality ≤ k and thus the answer to Problem 2 for the given
instance is no. Since Problem 2 is NP-hard, this completes the proof of Theorem 2.
.
v2
e23
v1
e13
e34
v3
v4
6
u1
u2
u3
u4
e13
e23
e34
e13
e23
e34
v1
v2
v3
v4
v1
v2
v3
v4
Fig. 1. An example demonstrating the construction in the proof of Thm. 2. Left: a graph G. The unique solution to the minimal dominating
set problem for G is D = {v3}. Middle: the digraph G described in the proof of Thm. 2. Right: the digraph ¯G describing the unique
solution to the minimal controllability problem for G obtained by adding a control to each node eij and to the node v3 in G.
Example 2. Consider the graph G = (V, E) shown on the left of Fig. 1, with V = {v1, v2, v3, v4}
and E = {e13, e23, e34}. The unique solution to the minimal dominating set for G is D = {v3}. The
construction in the proof above yields the directed graph G = ( V , E) depicted in the middle of Fig. 1.
Note that this has two layers of nodes: L2 and L3 that include the edges and vertices in G, respectively.
The solution of the minimal controllability problem for G is depicted on the right of Fig. 1. Here Y = {v3}.
Our next goal is to present a polynomial-time algorithm for determining if (2) is controllable. To do
this, we first derive graph-theoretic necessary and sufficient conditions for the controllability of a CBCN.
B. Necessary and Sufficient Conditions for Controllability of a CBCN
We begin by introducing several definitions that will be used later on. In a digraph, a source is a node
that has in-degree zero, and a sink is a node that has out-degree zero. A digraph that does not contain
cycles is called a directed acyclic graph (DAG).
We now introduce several definitions referring to the dependency graph of a CBCN. A node that
represents a state-variable (SV) is called a simple node. A node that represents a control input is called
a generator. Note that a generator is always a source, and that its set of out-neighbors always contains
a single simple node (see (2)). A node that represents an SV that has an added control input is called a
directly controlled node. Note that a directly controlled node is also a simple node. A simple node with
out-degree one, and without a self-loop, is called a channel. Note that the only out-neighbor of a channel
is another simple node (since a generator is always a source).
We are now ready to derive necessary conditions for the controllability of a CBCN. We will assume
throughout that no SV in the network has a constant updating function. Indeed, in this case it is clear
that the updating function of such an SV must be replaced by a control input. Under this assumption, a
simple node cannot be a source.
Proposition 4. The dependency graph of a controllable CBCN is acyclic.
Proof of Prop. 4. Consider a CBCN with a cycle in its dependency graph. Every vertex in the cycle
corresponds to an SV (i.e., it is a simple node), as a generator is a source, so it cannot be part of a cycle.
Moreover, any simple node in the cycle is not a directly controlled node, since the only in-neighbor of a
directly controlled node is a generator. This implies that if at time 0 the SVs in the cycle are all zero then
they can never be steered to the a state where they are all one. Hence, the CBCN is not controllable.
It is natural to expect that in a controllable CBCN every SV has the following property. There exists
a path from a control input to the SV that allows to set the SV to zero (the canalyzing value), without
affecting the other SVs. To make this precise, we say that a CBCN has Property P if every simple node
in its dependency graph contains in its set of in-neighbors either a generator or a channel. The next result
provides another necessary condition for controllability of a CBCN.
Proposition 5. A controllable CBCN has Property P.
7
Proof of Prop. 5. Seeking a contradiction, assume that the CBCN is controllable and that there exists
a simple node v in its dependency graph that does not contain a generator nor a channel in its set of
in-neighbors. This means that there does not exist a node w in the graph such that v is the only simple
node in the out-neighbors of w. Hence, the SV that corresponds to v cannot change its value to zero (the
canalyzing value) without at least one other SV changing its value to zero as well. Consider two states: a
corresponding to all SVs being zero, and b corresponding to all SVs being one except for v that is zero.
Since the CBCN is controllable it is possible to steer it from a to b. This implies that v has a self-loop,
as it holds the value zero while the other SVs change their values. Prop. 4 implies that the CBCN is not
controllable.
Props. 4 and 5 provide two necessary conditions for the controllability of a CBCN. The next result
shows that together they are also sufficient.
Theorem 6. A CBCN is controllable if and only if its dependency graph is a DAG and satisfies Property P.
To prove this, we introduce another definition and some auxiliary results. A controlled path is an
ordered non-empty set of nodes in the dependency graph such that: the first element in the ordered set is
a generator, and if the set contains more than one element, then for any i > 1 the ith node is a simple node,
and is the only element in the set of out-neighbors of node i − 1. Controlled paths with non-overlapping
nodes are called disjoint controlled paths.
Proposition 7. Consider a CBCN with a dependency graph G that is a DAG and satisfies Property P.
Then G can be decomposed into disjoint controlled paths, such that every vertex in the graph belongs to
a single controlled path (i.e., the union of the disjoint controlled paths forms a vertex cover of G).
Proof of Prop. 7. The proof is based on Algorithm 1 detailed below that accepts such a graph G and
terminates after each vertex in the graph belongs to exactly one controlled path.
Algorithm 1 Decompose the nodes in G into disjoint controlled paths
1: while there exists a simple node v ∈ G that is not included in any ordered set do
2:
cnode ← v and cset ← {v}
if Nin(cnode) contains a channel then
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
pick a channel u ∈ Nin(cnode)
if u does not belong to any ordered set then
insert u to the ordered set cset just before cnode
cnode ← u
goto 3
else
let H denote the ordered set that contains u
merge cset into H keeping the order between any two adjacent elements
goto 1
else
pick a generator u ∈ Nin(cnode)
if u does not belong to any ordered set then
insert it to cset just before cnode
goto 1
else
goto 10
Note that the assumption that G has Property P is used in line 14 of the algorithm. The proof that
Algorithm 1 terminates in a finite number of steps and divides all the nodes in the graph to a set of
disjoint controlled paths is straightforward and thus omitted. The basic idea is that Property P and the
8
U1
U2
X1
X2
X3
Fig. 2. Dependency graph of the CBCN (4) .
fact that the graph is a DAG implies that we can "go back" from any simple node through a chain of
channels ending with a generator, thus creating a controlled path. This completes the proof of Prop. 7.
Example 3. Consider the CBCN:
X1(k + 1) = U1(k),
X2(k + 1) = U2(k),
X3(k + 1) = X1(k)X2(k).
(4)
It is straightforward to verify that (4) is controllable. Fig. 2 depicts the dependency graph G of this CBCN.
Note that G is a DAG and that it satisfies Property P. A decomposition into two disjoint controlled paths
is C 1 = {U1 → X1 → X3}, C 2 = {U2 → X2}.
We can now prove Thm. 6.
Proof of Thm. 6. Consider a CBCN satisfying the conditions stated in the Theorem. By Prop. 7, there
exists a decomposition of its dependency graph into a set of m ≥ 1 disjoint controlled paths C 1, . . . , C m,
such that every vertex in the graph belongs to a single controlled path. We assume from here on that m = 2
(the proof in the general case is a straightforward generalization).
Pick two states a, b ∈ {0, 1}n. We prove that the CBCN is controllable by providing a control sequence
that steers the CBCN from X(0) = a to X(T ) = b, in time T ≥ 0. Since the paths provide a vertex
cover, the desired state b can be decomposed into b1 and b2 such that when the state is b the simple nodes
in C 1 [C 2] have state b1 [b2].
First, feed a control sequence with all ones to both generators until all the SVs reach the value one
at some time τ ≥ 0. Such a τ exists, because by the properties of the paths there are no arcs between
simple nodes in C1 and C2, except perhaps an arc from the final node in one path, say C1, to the other
path C2 (and since the graph is a DAG there is no arc from a node in C2 to a node in C1).
By adding a chain of dummy control inputs at the beginning of the shorter path, if needed, we may
assume that C 1 = C 2. Now we may view each path as a shift register (with all SVs initiated to one) and
it is straightforward to feed each path with a suitable sequence of controls to obtain the desired states b1
and b2 at some time T ≥ 0. Thus, the CBCN is controllable.
Note that this proof also provides the sequence of controls needed to steer the CBCN from a to b, i.e.
it solves the control synthesis problem (given the decomposition into a set of disjoint controlled paths).
Combining this with Prop. 7 implies the following.
Corollary 8. A CBCN is controllable if and only if its dependency graph can be decomposed into a set
of disjoint controlled paths.
Using the necessary and sufficient condition for controllability it is possible to derive an efficient
algorithm for determining if a CBCN is controllable.
C. An efficient algorithm for determining controllability
Algorithm 2 below tests if a CBCN is controllable. It is based on the condition in Thm. 6.
Algorithm 2 Testing controllability of a CBCN in the form (2) with n SVs and q ≤ n control inputs
9
1: generate the dependency graph G = (V, E)
2: if G is not a DAG then return ("not controllable")
3: create a list L of n bits
4: set all bits in L to 0
5: for all nodes v ∈ V do
6:
if Nout(v) 6= 1 then return ("not controllable")
else
7:
8:
9:
j ← the element in Nout(v)
L(j) ← 1
endfor
10: if all bits in L are 1 then
return ("controllable")
11:
12: else
13:
return ("not controllable")
The input to Algorithm 2 is a CBCN with n SVs and q ≤ n control inputs. The first step is to build
the dependency graph G = (V, E). The complexity of this step is O(n2), as this requires going through n
updating functions, and each function has at most n arguments. The resulting graph satisfies V = n+q ≤
2n, and E ≤ n2.
Checking if G is a DAG in line 2 can be done using a topological sort algorithm. The complexity is
linear in V , E (see, e.g. [24]), i.e. it is O(n2).
Lines 3-13 use the list L to check Property P, that is, to verify that the set of in-neighbors of every SV
contains either a generator or a channel. This part has complexity O(n + q) = O(n).
The total time-complexity of the algorithm is thus O(n2). More precisely, the complexity of the
algorithm is linear in the length of the description of the CBCN, and the latter is O(n2).
IV. CONCLUSIONS
Minimal controllability problems for dynamical systems are important both theoretically and for real-
world applications where actuators can be added to control the SVs. Here, we considered a minimal
controllability problem for an important subclass of BNs, namely, CBNs. Using a graph-theoretic approach
we derived: a necessary and sufficient condition for the CBCN (2) to be controllable, and a polynomial-time
algorithm for testing controllability. We also showed that the minimal controllability problem is NP-hard.
Our approach is based on the new concept of a controlled path and the decomposition of the dependency
graph of the CBCN into disjoint controlled paths. Roughly speaking, this corresponds to decomposing
the CBCN into a set of shift registers that are almost decoupled.
Recall that given a digraph G = (V, E) a path cover is a set of directed paths such that every v ∈ V
belongs to at least one path. A vertex-disjoint path cover is a set of paths such that every v ∈ V belongs
to exactly one path. The minimum path cover problem (MPCP) consists of finding a vertex-disjoint path
cover having the minimal number of paths. This problem has applications in software validation [25].
The MPCP may seem closely related to the problem studied here, but this is not necessarily so. First,
the MPCP for a DAG can be solved in polynomial time (see, e.g. [26]). Second, the solution of the MPCP
does not provide enough information on the controllability of a CBCN. For this, we need the more specific
properties of controlled paths.
We believe that the notions introduced here will find more applications in other control-theoretic
problems for CBNs. An interesting direction for further research is to derive an efficient algorithm that
10
is guaranteed to approximately solve the minimal controllability problem for CBCNs, with a guaranteed
approximation error in the number of needed control inputs.
REFERENCES
[1] E. D. Sontag, Mathematical Control Theory: Deterministic Finite Dimensional Systems, 2nd ed. New York: Springer, 1998.
[2] A. Olshevsky, "Minimal controllability problems," IEEE Trans. Control of Network Systems, vol. 1, no. 3, pp. 249 -- 258, 2014.
[3] Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabasi, "Controllability of complex networks," Nature, vol. 473, pp. 167 -- 173, 2011.
[4] V. D. Blondel and J. N. Tsitsiklis, "A survey of computational complexity results in systems and control," Automatica, vol. 36, pp.
1249 -- 1274, 2000.
[5] D. G. Green, T. G. Leishman, and S. Sadedin, "The emergence of social consensus in Boolean networks," in Proc. 2007 IEEE Symposium
on Artificial Life (IEEE-ALife'07), Honolulu, Hawaii, 2007, pp. 402 -- 408.
[6] S. Kochemazov and A. Semenov, "Using synchronous Boolean networks to model several phenomena of collective behavior," PLoS
ONE, vol. 9, no. 12, p. e115156, 2014.
[7] A. Kasyanov, L. Kirkland, and M. T. Mihaela, "A spatial SIRS Boolean network model for the spread of H5N1 avian influenza virus
among poultry farms," in Proc. 5th Int. Workshop Computational Systems Biology (WCSB), 2008, pp. 73 -- 76.
[8] S. Bornholdt, "Boolean network models of cellular regulation: prospects and limitations," J. R. Soc. Interface, vol. 5, pp. S85 -- S94,
2008.
[9] T. Helikar, N. Kochi, J. Konvalina, and J. Rogers, "Boolean modeling of biochemical networks," The Open Bioinformatics Journal,
vol. 5, pp. 16 -- 25, 2011.
[10] A. Faure, A. Naldi, C. Chaouiya, and D. Thieffry, "Dynamical analysis of a generic Boolean model for the control of the mammalian
cell cycle," Bioinformatics, vol. 22, pp. e124 -- e131, 2006.
[11] D. Laschov and M. Margaliot, "Controllability of Boolean control networks via the Perron-Frobenius theory," Automatica, vol. 48, pp.
1218 -- 1223, 2012.
[12] R. Tarjan, "Depth first search and linear graph algorithms," SIAM J. Computing, vol. 1, no. 2, pp. 146 -- 160, 1972.
[13] T. Akutsu, M. Hayashida, W.-K. Ching, and M. K. Ng, "Control of Boolean networks: Hardness results and algorithms for tree structured
networks," J. Theoretical Biology, vol. 244, pp. 670 -- 679, 2007.
[14] S. A. Kauffman, The Origins of order: Self Organization and Selection in Evolution. Oxford University Press, 1993.
[15] S. E. Harris, B. K. Sawhill, A. Wuensche, and S. Kauffman, "A model of transcriptional regulatory networks based on biases in the
observed regulation rules," Complexity, vol. 7, no. 4, pp. 23 -- 40, 2002.
[16] S. Kauffman, C. Peterson, B. Samuelsson, and C. Troein, "Random Boolean network models and the yeast transcriptional network,"
Proceedings of the National Academy of Sciences, vol. 100, no. 25, pp. 14 796 -- 14 799, 2003.
[17] S. Kauffman, C. Peterson, B. Samuelssonand, and C. Troein, "Genetic networks with canalyzing Boolean rules are always stable,"
Proceedings of the National Academy of Sciences, vol. 101, no. 49, pp. 17 102 -- 17 107, 2004.
[18] A. S. Jarrah, R. Laubenbacher, and A. Veliz-Cuba, "The dynamics of conjunctive and disjunctive Boolean network models," Bull. Math.
Biology, vol. 72, no. 6, pp. 1425 -- 1447, 2010.
[19] Z. Gao, X. Chen,
and T. Basar, "Stability structures of
conjunctive Boolean networks," 2016.
[Online]. Available:
https://arxiv.org/abs/1603.04415#
[20] M. Merika and S. H. Orkin, "Functional synergy and physical interactions of the erythroid transcription factor gata-1 with the Kruppel
family proteins Sp1 and EKLF," Mol. Cell. Bio., vol. 15, no. 5, pp. 2437 -- 2347, 1995.
[21] B. M. Gummow, J. O. Scheys, V. R. Cancelli, and G. D. Hammer, "Reciprocal regulation of a glucocorticoid receptor-steroidogenic
factor-1 transcription complex on the Dax-1 promoter by glucocorticoids and adrenocorticotropic hormone in the adrenal cortex," Mol.
Endocrinol., vol. 20, no. 11, pp. 2711 -- 2723, 2006.
[22] D. H. Nguyen and P. D'haeseleer, "Deciphering principles of transcription regulation in eukaryotic genomes," Molecular Systems
Biology, vol. 2, no. 1, 2006.
[23] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness. San Francisco, CA: W.
H. Freeman, 1979.
[24] A. B. Kahn, "Topological sorting of large networks," Communications of the ACM, vol. 5, no. 11, pp. 558 -- 562, 1962.
[25] S. C. Ntafos and S. L. Hakimi, "On path cover problems in digraphs and applications to program testing," IEEE Trans. Software
Engineering, vol. SE-5, no. 5, pp. 520 -- 529, 1979.
[26] S. Even, Graph Algorithms, 2nd ed., G. Even, Ed. New York, NY: Cambridge University Press, 2011.
|
1503.05638 | 2 | 1503 | 2015-09-21T04:31:50 | Entropy-scaling search of massive biological data | [
"cs.DS",
"q-bio.GN"
] | Many datasets exhibit a well-defined structure that can be exploited to design faster search tools, but it is not always clear when such acceleration is possible. Here, we introduce a framework for similarity search based on characterizing a dataset's entropy and fractal dimension. We prove that searching scales in time with metric entropy (number of covering hyperspheres), if the fractal dimension of the dataset is low, and scales in space with the sum of metric entropy and information-theoretic entropy (randomness of the data). Using these ideas, we present accelerated versions of standard tools, with no loss in specificity and little loss in sensitivity, for use in three domains---high-throughput drug screening (Ammolite, 150x speedup), metagenomics (MICA, 3.5x speedup of DIAMOND [3,700x BLASTX]), and protein structure search (esFragBag, 10x speedup of FragBag). Our framework can be used to achieve "compressive omics," and the general theory can be readily applied to data science problems outside of biology. | cs.DS | cs | Entropy-scaling search of massive biological data
Y. William Yua,b,∗, Noah M. Danielsa,b,∗, David Christian Dankob, Bonnie
Bergera,b,∗∗
aDepartment of Mathematics, Massachusetts Institute of Technology, Cambridge,
bComputer Science and AI Lab, Massachusetts Institute of Technology, Cambridge,
Massachusetts 02139
Massachusetts 02139
Highlights
• We describe entropy-scaling search for finding approximate matches
in a database
• Search complexity is bounded in time and space by the entropy of the
database
• We make tools that enable search of three largely intractable real-world
databases
• The tools dramatically accelerate metagenomic, chemical, and protein
structure search
eTOC Blurb
We describe a general framework for efficiently searching massive datasets
having certain properties common in biology.
Summary
Many datasets exhibit a well-defined structure that can be ex-
ploited to design faster search tools, but it is not always clear when
such acceleration is possible. Here, we introduce a framework for
similarity search based on characterizing a dataset's entropy and
∗These authors contributed equally to this work.
∗∗Corresponding author
Email address: [email protected] (Bonnie Berger)
Cell Systems
Volume 1, Issue 2, pages 130-140
arXiv:1503.05638v2 [cs.DS] 21 Sep 2015
fractal dimension. We prove that searching scales in time with
metric entropy (number of covering hyperspheres), if the fractal
dimension of the dataset is low, and scales in space with the sum
of metric entropy and information-theoretic entropy (randomness
of the data). Using these ideas, we present accelerated versions of
standard tools, with no loss in specificity and little loss in sensi-
tivity, for use in three domains -- high-throughput drug screening
(Ammolite, 150× speedup), metagenomics (MICA, 3.5× speedup
of DIAMOND [3,700× BLASTX]), and protein structure search
(esFragBag, 10× speedup of FragBag). Our framework can be
used to achieve "compressive omics," and the general theory can
be readily applied to data science problems outside of biology.
Source code: http://gems.csail.mit.edu
Introduction
Throughout all areas of data science, researchers are confronted with in-
creasingly large volumes of data. In many fields, this increase is exponential
in nature, outpacing Moore's and Kryder's laws on the respective doublings
of transistors on a chip and long-term data storage density (Kahn, 2011).
As such, the challenges posed by the massive influx of data cannot be solved
by waiting for faster and larger capacity computers but, instead, require
instead the development of data structures and representations that exploit
the complex structure of the dataset.
Here, we focus on similarity search, where the task at hand is to find all
entries in some database that are "similar," or approximate matches, to a
query item. Similarity search is a fundamental operation in data science and
lies at the heart of many other problems, much like how sorting is a primi-
tive operation in computer science. Traditionally, approximate matching has
been studied primarily in the context of strings under edit distance metrics
(Box 1) (e.g., for a spell-checker to suggest the most similar words to a mis-
spelled word) (Ukkonen, 1985). Several approaches, such as the compressed
suffix array and the FM-index (Grossi & Vitter, 2005; Ferragina & Manzini,
2000), have been developed to accelerate approximate matching of strings.
However, it has been demonstrated that similarity search is also important
in problem domains where biological data are not necessarily represented
as strings, including computational screening of chemical graphs (Schaeffer,
2007) and searching protein structures (Budowski-Tal et al., 2010). There-
fore, approaches that apply to more general conditions are needed.
2
As available data grow exponentially (Berger et al., 2013; Yu et al.,
2015) (e.g., genomic data in Figure S1), algorithms that scale linearly (Box
1) with the amount of data no longer suffice. The primary ways in which
the literature addresses this problem -- locality sensitive hashing (Indyk &
Motwani, 1998), vector approximation (Ferhatosmanoglu et al., 2000), and
space partitioning (Weber et al., 1998) -- involve the construction of data
structures that support more efficient search operations. However, we note
that, as biological data increase, not only does the redundancy present in
the data also increase (Loh et al., 2012), but also internal structure (such
as the fact that not all conceivable configurations, e.g. all possible protein
sequences, actually exist) also becomes apparent. Existing general-purpose
methods do not explicitly exploit the particular properties of biological data
to accelerate search (see the Theory section in the Supplemental Methods).
Previously, our group demonstrated how redundancy in genomic data
could be used to accelerate local sequence alignment. Using an approach that
we called "compresive genomics," we accelerated BLAST and BLAT (Kent,
2002) by taking advantage of high redundancy between related genomes us-
ing link pointers and edit scripts to a database of unique sequences (Loh
et al., 2012). We have used similar strategies to obtain equally encouraging
results for local alignment in proteomics (Daniels et al., 2013). Empirically,
this compressive acceleration appears to scale almost linearly in the entropy
of the database, often resulting in orders of magnitude better performance;
however, these previous studies neither proved complexity bounds nor es-
tablished a theory to explain these empirical speedups.
Here, we generalize and formalize this approach by introducing a frame-
work for similarity search of omics data. We prove that search performance
primarily depends on a measure of the novelty of new data, also known as
entropy. This framework, which we call entropy-scaling search, supports the
creation of a data structure that provably scales linearly in both time and
space with the entropy of the database, and thus sublinearly with the entire
database.
We introduce two key concepts for characterizing a dataset: metric en-
tropy and fractal dimension. Intuitively, metric entropy measures how dis-
similar the dataset is from itself, and fractal dimension measures how the
number of spheres needed to cover all points in a database scales with the
radii of those spheres. Both are rigorously defined later, but note that
metric entropy is not to be confused with the notion of a distance metric
(Box 1). Using these two concepts, we show that, if similarity is defined
by a metric-like distance function (e.g., edit or Hamming distance) and
the database exhibits both low metric entropy and fractal dimension, the
3
entropy-scaling search performs much better than nave and even optimized
methods. Through three applications to large databases in chemogenomics,
metagenomics, and protein structure search, we show that this framework
allows for minimal (or even zero) loss in recall, coupled with zero loss in
specificity. The key benefit of formulating entropy-scaling search in terms
of metric entropy and fractal dimension is that this allows us to provide
mathematically rigorous guidance as to how to determine the efficacy of the
approach for any dataset.
• Edit distance: the number of edits (character insertions, deletions, or substi-
tutions) needed to turn one string into another.
• Scale, in time and space: the amount of time or space a task takes as a
function of the amount of data on which it must operate. A task requiring
time directly proportional to the size of the data is said to scale linearly; for
example, searching a database takes twice as long if the database grows by
a factor of two.
• Distance metric: a measure of distance that obeys several mathematical prop-
erties, including the triangle inequality.
• Covering spheres: a set of spheres around existing points so that every point
is contained in at least one sphere and no sphere is empty.
• Metric entropy: a measure of how dissimilar a dataset is from itself. Defined
as the number of covering spheres.
• Fractal dimension: a measure of how the number of points contained within
a sphere scales with the radius of that sphere.
• Information-theoretic entropy: often used in data compression as shorthand
for the number of bits needed to encode a database or a measure of the
randomness of that database.
• Pattern matching: refers to searching for matches that might differ in specific
ways from a query, such as wildcards or gaps, as opposed to searching for
all database entries within a sphere of a specified radius as defined by an
arbitrary distance function.
Box 1: Definitions
4
Figure 1: Entropy-scaling framework for similarity search.
(A-D) As
shown, (A) The naıve approach tests each query against each database entry
to find entries within distance r of the query (inside the small green disc).
(B) By selecting appropriate cluster centers with maximum radius rc to par-
tition the database, we can (C) first do a coarse search to find all cluster
centers within distance r + rc of a query (larger green disc), and then the
(D) triangle inequality guarantees that a fine search over all corresponding
cluster entries (blue polygonal regions) will suffice.
Results
Entropy-scaling similarity search
The basic framework for the entropy-scaling search of a database involves
four steps.
(1) Analyze the database to define a high-dimensional space
and determine how to map database entries onto points in this space (this
mapping may be one-to-one). (2) Use this space and a measure of similarity
between points to group entries in the database into clusters. (3) To search
for a particular query item, perform a coarse-grained search to identify the
clusters that could possibly contain the query. (4) Do a fine-grained search
of the points contained within these clusters to find the closest matches to
the query (Figure 1).
Here, we provide conceptual motivation for this process. In the following
text, we consider entropy to be nearly synonymous with distance between
points in a high-dimensional space; thus, with low entropy, newly added
points do not tend to be far from all existing points. For genomic sequences,
the distance function can be edit distance; for chemical graphs, it can be
Tanimoto distance; and for general vectors, it can be Euclidean or cosine
distance. We are interested in the similarity search problem of finding all
points in a set that are close to (i.e., similar to) the query point.
Let us first consider what it means for a large biological dataset, consid-
ered as points in a high-dimensional space, to be highly redundant. Perhaps
many of the points are exact duplicates; this easy scenario is trivially ex-
ploited by de-duplication and is already standard practice with datasets such
as the NCBI's non-redundant (NR) protein database (Pruitt et al., 2005).
Maybe the points mostly live on a low-dimensional subspace; statistical tools
such as Principal Component Analysis (PCA) exploit this property in data
analysis. Furthermore, if the dimension of the subspace is sufficiently low, it
can be divided into cells, allowing quick similarity searches by looking only
at nearby cells (Weber et al., 1998). However, when the dimensionality of
the subspace increases, cell search time grows exponentially; additionally, in
sparse datasets, most of the cells will be empty, which wastes search time.
More importantly, biological datasets generally do not live in low-dimensional
subspaces. Consider the instructive case of genomes along an evolutionary
"tree of life" (Figure 2). Such a tree has many branches (although admixture
merges branches back together), and looks nearly one-dimensional locally,
but it is globally of higher dimension. Additionally, because of differences
due to mutation, each of the branches is also "thick" (high dimensional)
when looked at closely. Viewing this example as a low-dimensional sub-
space, as in PCA, is incorrect.
6
Figure 2: Cartoon depiction of points in a high-dimensional space. This
cartoon depics points in an arbitrary high-dimensional space that live close
to a one-dimensional tree-like structure, as might arise from genomes gener-
ated by mutation and selection along an evolutionary tree of life. Although
high-dimensional at a fine scale, at the coarser scale of covering spheres, the
data cloud looks nearly one-dimensional, which enables entropy-scaling of
similarity search. The cluster center generation was performed using the
same method we used for protein structure search. The blue circles around
the green query point illustrate low fractal dimension: the larger-radius cir-
cle contains only linearly more points than the smaller one, rather than
exponentially more.
In contrast, the red circles around the orange query
point illustrate higher local fractal dimension.
However, the local low-dimensionality can be exploited by looking on the
right scales: a coarse scale in which the tree looks one-dimensional locally
and a fine scale where the branch width matters. We cover the tree with
spheres (Box 1) of radius rc, where rc is on the order of the branch width;
these spheres determine our clusters, and the number of them is the metric
entropy of the tree (Tao, 2008). Because all the points within a sphere are
close to each other, they are highly redundant and can be encoded in terms
of one another, saving space.
By the triangle inequality, in order to search for all points within distance
r of a query, we need only to look in nearby spheres with centers (i.e.,
representatives) within a distance r + rc of the query (Figure 1D). However,
because each sphere has a radius comparable to branch width, the tree is
locally one-dimensional on the coarse scale; that is, spheres largely tend to
extend along the branches of the tree rather than in all directions. We will
call this property of local scaling the fractal dimension d of the tree at the
scale rc (Falconer, 1990), where rc is essentially our ruler size and d = 1.
Thus, increasing the search radius for coarse search only linearly increases
the number of points that need to be searched in a fine search.
A similar analysis holds in the more general case where d 6= 1. The
entropy-scaling frameworks we introduce can be expected to provide a boost
to approximate search when fractal dimension d of a dataset D is low (i.e.,
close to 1) and metric entropy k is low. Specifically, the ratio D
provides
k
an estimate of the acceleration factor for just the coarse search component
compared to a full linear search of a database D. Local fractal dimension
around a data point can be computed by determining the number of other
data points within two radii r1 and r2 of that point; given those point counts
log(n2/n1)
log(r2/r1)
Sampling this property over a dataset can provide a global average fractal
dimension. When we search a larger radius around a query, the number
of points we encounter grows exponentially with the fractal dimension; low
fractal dimension implies that this growth will not obviate the gains provided
by an entropy-scaling data structure.
(n1 and n2, respectively), fractal dimension d is simply d =
.
More formally, given a database with fractal dimension d and metric
entropy k at the scale rc, we show in the Supplemental Methods that the
time-complexity of similarity search on database D for query q with radius
8
r is
O
+
metric entropy
k{z}
output size
r (cid:19)d
BD(q, r)(cid:18) r + 2rc
z
}
{z
}
scaling factor
!.
{
Thus, for small fractal dimension and output size, similarity search is asymp-
totically linear in metric entropy. Additionally, because the search has to
look at only a small subset of the clusters, the clusters can be stored in
compressed form, and only decompressed as needed, giving space savings
that also scale with entropy. The space-complexity scales with the sum of
metric and information-theoretic entropy, rather than just metric entropy
(Supplemental Methods: Theory).
Practical application of entropy-scaling search
We have presented the simplest such data to analyze for clarity of ex-
position. However, real data is generally messier. Sometimes the distance
function is not a metric, so we lose the triangle inequality guarantee of 100%
sensitivity; sometimes different distance functions can be used for the clus-
tering versus search; and sometimes even what counts as a distinct data
point is not entirely clear without domain knowledge (for example, long
genomic sequences might be better broken into shorter subsequences).
To show that entropy-scaling frameworks are robust to the variations
presented by real data, we explored a diversity of applications from three
major biological "big challenges of big data" -- pharamaceuticals, metage-
nomics, and protein structure (Marx, 2013). We demonstrate that the gen-
eral scheme results in order-of-magnitude improvements in running time in
these different contexts, promising to enable new workflows for practition-
ers (e.g., fast first-pass computational drug screens and local analyses of se-
quencing data in remote field sites for real-time epidemic monitoring). These
applications are enabled by augmenting the framework with domain-specific
distance functions in different stages of the process, as well as preprocessing
to take advantage of domain-specific knowledge. We expect that as long as
the dataset exhibits both low entropy and low fractal dimension -- and this
is empirically true in biological systems -- our entropy-scaling framework has
the potential to achieve massive speedup over more naıve methods and sig-
nificant speedup even over other highly optimized methods.
Source code for the applications discussed here is available at http:
//gems.csail.mit.edu and in the Supplemental Information.
9
Application to high-throughput drug screening
Chemogenomics is the study of drug and target discovery by using chem-
ical compounds to probe and characterize proteomic functions (Bredel &
Jacoby, 2004). Particularly in the field of drug discovery and drug repur-
posing, prediction of biologically active compounds is a critical task. Com-
putational high-throughput screening can eliminate many compounds from
wet-lab consideration, but even this screening can be time-consuming. Pub-
Chem (Bolton et al., 2008), a widely-used repository of molecular compound
structures, has grown greatly since 2008. In July 2007, PubChem contained
10.3 million compounds. In October 2013, PubChem contained roughly 47
million compounds, while in December 2014 it contained 61.3 million com-
pounds.
We designed this compression and search framework around one of the
standard techniques for high-throughput screening of potential drug com-
pounds, the use of maximum common subgraph (MCS) to identify simi-
lar motifs among molecules (Cao et al., 2008; Rahman et al., 2009). We
introduce Ammolite, a method for clustering molecular databases such as
PubChem, and for quickly searching for similar molecular structures in com-
pressed space. Ammolite demonstrates that entropy-scaling methods can be
extended to data types that are not inherently sequence based. Ammolite
is a practical tool that provides approximately a factor of 150 speed-up
with greater than 92% accuracy compared to the popular small molecular
subgraph detector (SMSD) (Rahman et al., 2009).
An MCS-based search of molecule databases typically matches pairs of
molecules by Tanimoto distance (Rahman et al., 2009). Tanimoto distance
obeys the triangle inequality and is more useful in the domain of molecular
graphs than other distance metrics such as graph distance (Bunke & Shearer,
1998).
To compress a molecule database, we project the space of small molecules
onto a subspace by removing nodes and edges that do not participate in sim-
ple cycles (Figure S2); note that a molecule without cycles will collapse to
a single node. Clusters are exactly pre-images of this projection operator
(i.e., all molecules that are isomorphic after simplification form a cluster).
Coarse search is performed by finding the MCS on this much smaller pro-
jection subspace. This step increases speed by reducing both the required
number of MCS operations and the time required for each MCS operation,
which scales with the size of the molecule. Further reduction in search
time is accomplished by grouping clusters according to size of the molecules
within; because Tanimoto distance relies on molecule size, clusters contain-
10
ing molecules significantly larger or smaller then the query need not be
searched at all.
The time required to cluster a large database such as PubChem is,
nonetheless, significant; clustering the 306-GB PubChem required approxi-
mately 400 hours on a 12-core Xeon X5690 running at 3.47GHz, and required
128 GB RAM. However, this database can easily be appended to as new
molecules become available, and the clustering time can be amortized over
future queries. It is worth noting that the preprocessing of molecular graphs
can cause the triangle inequality to be violated; while the distance function
is a metric, the clustering does not respect that metric. Ammolite can
be readily incorporated into existing analysis pipelines for high-throughput
drug screening.
Our entropy-scaling framework can be applied to PubChem because it
has both low fractal dimension and low metric entropy. In particular, we
determined the mean local fractal dimension of PubChem to be approx-
imately 0.2 in the neighborhood between 0.2 and 0.4 Tanimoto distance,
and approximately 1.9 in the neighborhood between 0.4 and 0.5. The ex-
pected speedup is measured by the ratio of database size to metric entropy,
which, for PubChem is approximately 11:1. This is not taking into account
the clustering according to molecule size, which further reduces the search
space.
Because SMSD is not computationally tractable on the entire PubChem
database, we benchmarked Ammolite against SMSD on a subset of 1 mil-
lion molecules from PubChem. Since SMSD's running time should scale
linearly with the size of the database, we extrapolated the running time
of SMSD to the entire PubChem database. Benchmarking Ammolite and
SMSD required 60GB RAM and used 12 threads, although Ammolite's
search, used normally, requires < 20 GB RAM. For these benchmarks, we
used five randomly-chosen query molecules with at least two rings (Pub-
Chem IDs 1504670, 19170294, 28250541, 4559889, and 55484477), as well
as five medically-interesting molecules chosen by hand (adenosine triphos-
phate [atp], clindamycin, erythromycin, teixobactin, and thalidomide). We
also used SMSD as a gold standard against which we measured Ammolite's
recall.
Ammolite achieves an average of 92.5% recall with respect to SMSD
(Table 1a). This recall is brought down by one poorly-performing compound,
PubChem ID 1504670, with only 62.5% recall, but is otherwise over 80%.
Furthermore, Ammolite's speed gains with respect to SMSD grow as the
database grows (Table 1b).
11
Table 1: Benchmarks of Ammolite vs. SMSD on databases of (a) 1 million
molecules and (b) 47 million molecules (all of PubChem)
(a) Ammolite benchmark on database of 1 million molecules
PubChem ID
5957 (atp)
446598 (clindamycin)
12560 (erythromycin)
86341926 (teixobactin)
5426 (thalidomide)
1504670
19170294
28250541
4559889
55484477
SMSD (hours) Ammolite (hours) Speedup Recall (%)
4.4
18.7
849.6
618.5
48.9
8.1
31.3
43.3
108.8
23.3
0.14
1.5
3.0
2.3
0.81
0.8
0.8
4.8
2.7
2.5
31
11.7
279.2
265.5
60.4
10.3
39.7
9.0
41.0
9.1
81%
90%
91%
100%
100%
62.5%
100%
100%
100%
100%
(b) Ammolite benchmark on entire PubChem database as of October 2013. See
also Figure S2.
PubChem ID
5957 (atp)
446598 (clindamycin)
12560 (erythromycin)
86341926 (teixobactin)
5426 (thalidomide)
1504670
19170294
28250541
4559889
55484477
Ammolite (hours) Speedup
4.1
28.4
79.1
96.5
29.2
4.6
6.0
38.9
57.3
35.5
51.3
14.5
512.9
305.9
80.0
84.4
247.4
53.2
90.7
31.4
Application to metagenomics
Metagenomics is the study of genomic data sequenced directly from envi-
ronmental samples. It has led to improved understanding of how ecosystems
recover from environmental damage (Tyson et al., 2004) and how the hu-
man gut responds to diet and infection (David et al., 2014). Metagenomics
has even provided some surprising insights into disorders such as Autism
Spectrum Disorder (MacFabe, 2012).
BLASTX (Altschul et al., 1990) is widely used in metagenomics to
map reads to protein databases such as KEGG (Kanehisa & Goto, 2000)
and NCBI's NR (Sayers et al., 2011). This mapping is additionally used
as a primitive in pipelines such as MetaPhlAn (Segata et al., 2012), PI-
CRUSt (Langille et al., 2013), and MEGAN (Huson et al., 2011) to de-
termine the microbial composition of a sequenced sample. Unfortunately,
BLASTX's runtime requirements scale linearly with the product of the size
of the full read dataset and the targeted protein database, and thus each
year require exponentially more runtime to process the exponentially grow-
ing read data. These computational challenges are at present a barrier
to widespread use of metagenomic data throughout biotechnology, which
constrains genomic medicine and environmental genomics (Frank & Pace,
2008). For example, Mackelprang et al. (2011) reported that using BLASTX
to map 246 million reads against KEGG required 800,000 CPU hours at a
supercomputing center.
Although this is already a problem for major research centers, it is es-
pecially limiting for on-site analyses in more remote locations. In surveying
the 2014 Ebola outbreak, scientists physically shipped samples on dry ice to
Harvard for sequencing and analysis (Gire et al., 2014). Even as sequencers
become more mobile and can thus be brought on site, lack of fast Internet
connections in remote areas can make it impossible to centralize and ex-
pedite processing (viz., the cloud); local processing on resource-constrained
machines remains essential. Thus, a better-scaling and accurate version
of BLASTX raises the possibility of not only faster computing for large re-
search centers, but also of performing entirely on-site sequencing and desktop
metagenomic analyses.
Recently, approaches such as RapSearch2 (Zhao et al., 2012) and Dia-
mond (Buchfink et al., 2015) have provided faster alternatives to BLASTX.
We have applied our entropy-scaling framework to the problem of metage-
nomic search and demonstrate MICA, a method whose software implementa-
tion provides an acceleration of DIAMOND by a factor of 3.5, and BLASTX
by a factor of up to 3700. This application illustrates the potential of
entropy-scaling frameworks, while providing a useful tool for metagenomic
13
research. It can be readily incorporated into existing analysis pipelines (e.g.,
for microbial composition analysis using MEGAN). MICA clustering of the
September 17, 2014 NCBI NR database (containing 49.3 million sequences)
required 39 hr on a 12-core Xeon X5690 running at 3.47GHz; it used ap-
proximately 84 GB of resident memory.
Our entropy-scaling framework can be applied to the NCBI's NR database
because it, like PubChem, exhibits low fractal dimension and metric entropy.
We determined the mean local fractal dimension of the NCBI's NR database,
using sequence identity of alignment as a distance function, to be approx-
imately 1.6 in the neighborhood between 70% and 80% protein sequence
identity. The ratio of database size to metric entropy, which gives an indi-
cator of expected speedup, is approximately 30:1. Indeed, the notion that
protein sequence space exhibits structure, and lends itself to clustering, has
precedent (Linial et al., 1997).
To evaluate the runtime performance of MICA, we tested it against
BLASTX, RapSearch2 (Zhao et al., 2012) and Diamond (Buchfink et al.,
2015). On five read sets (ERR335622, ERR335625, ERR335631, ERR335635,
ERR335636) totalling 207,623 151-nucleotide (nt) reads from the American
Gut Microbiome project, we found that MICA provides measurable runtime
improvements over DIAMOND with no further loss in accuracy (Table 2a),
and substantial runtime improvements over BLASTX. Notably, the mean
running time for BLASTX was 58,215 minutes, while MICA took an av-
erage of 15.6 minutes, a speedup of 3,724x. MICA uses DIAMOND for its
coarse search, and can use either DIAMOND or BLASTX for its fine search.
We also evaluated MICA using BLASTX for both the coarse and the fine
search; this approach performed slightly slower than DIAMOND, requiring
an average of 89 min, though it was somewhat more accurate, at 95.9%
recall compared to DIAMOND's 90.4% recall. MICA using BLASTX for
both coarse and fine searches relied on a query-side clustering (discussed
in Supplemental Methods); we note that the time spent performing query-
side clustering is included here; without query-side clustering, this variant
of MICA takes 2,278 min, a speedup of 25x over BLASTX.
MICA accelerates DIAMOND with no further loss in accuracy: 90.4%
compared to unaccelerated BLASTX Table 2b. Experiments validating ac-
curacy treated BLASTX as a gold standard. Since MICA accelerates DI-
AMOND using entropy-scaling techniques, false positives with respect to
DIAMOND are not possible, but false negatives are. We report as accuracy
the fraction of BLASTX hits that are also returned by MICA.
DIAMOND's clever indexing and alphabet reduction provide excellent
runtime performance already, though its running time still scales linearly
14
Table 2:
(a) Running time and (b) accuracy of BLASTX, RapSearch2,
DIAMOND, and MICA. Data set is the American gut microbiome project
read sets ERR335622, ERR335625, ERR335631, ERR335635, ERR335636
(a) Running time in minutes (standard deviation)
BLASTX
58215 (1561.8) 206 (5.4)
RapSearch2 DIAMOND MICA-DIAMOND MICA-BLASTX
54 (1.1)
15.6 (0.5)
21.9 (1.7)
(b) Accuracy against BLASTX (standard deviation)
RapSearch2 DIAMOND MICA-DIAMOND MICA-BLASTX
79.5% (1.63) 90.4% (3.10) 90.4% (3.10)
90.4% (3.10)
with database size.
In contrast, as an entropy-scaling search, MICA will
demonstrate greater acceleration as database sizes grow (Daniels et al.,
2013). Moreover, MICA can use standard BLASTX for its fine search, which
allows the user to pass arbitrary parameters to the underlying BLASTX call
but which also comes at a small runtime penalty (40% in our testing). This
option allows for additional BLAST arguments that DIAMOND does not
support, such as XML output, which may be useful in some pipelines. Thus,
MICA with BLASTX may be suitable for a wider variety of existing analysis
pipelines.
Application to protein structure search
The relationship between protein structure and function has been a sub-
ject of intense study for decades, and this strong link has been used for the
prediction of function from structure (Hegyi & Gerstein, 1999). Specifically,
given a protein of solved (or predicted) structure but unknown function, the
efficient identification of structurally similar proteins in the Protein Data
Bank (PDB) is critical to function prediction. Finding structural neigh-
bors can also give insight into the evolutionary origins of proteins of interest
(Yona et al., 1999; Nepomnyachiy et al., 2014).
One approach to finding structural neighbors is to attempt to align the
query protein to all the entries in the PDB using a structural aligner, such
as STRUCTAL (Subbiah et al., 1993), ICE (Shindyalov & Bourne, 1998), or
Matt (Menke et al., 2008). However, performing a full alignment against ev-
ery entry in the PDB is prohibitively expensive, especially as the database
grows. To mitigate this, (Budowski-Tal et al., 2010) introduced the tool
FragBag, which avoids performing full alignments but rather describes each
protein as a "bag of fragments," where each fragment is a small structural
motif. FragBag has been reported as comparable to structural aligners such
15
as STRUCTAL or ICE, and its bag-of-fragments approach allows it to per-
form comparisons much faster than standard aligners. Importantly for us,
the bag of fragments is just a frequency vector, making FragBag amenable
to acceleration through entropy-scaling.
By first verifying that the local fractal dimension of PDB FragBag fre-
quency vectors is low in most regimes (d ≈ 2 − 3), Figure S3), we are given
reason to think that this problem is amenable to entropy-scaling search. As
an estimate of potential speedup, the ratio of PDB database size to metric
entropy at for the chosen cluster radii is on average, ∼10:1. We directly ap-
plied our entropy-scaling framework without any additional augmentation:
esFragBag (entropy-scaling FragBag) is able to achieve an average factor
of 10 speedup of the highly-optimized FragBag with less than 0.2% loss in
sensitivity and no loss in specificity.
For this last example, we intentionally approach the application of entropy-
scaling frameworks to FragBag in a blind manner, without using any domain-
specific knowledge. Instead, we use the very same representation (bag of
fragments) and distance functions (Euclidean and cosine distances) as Frag-
Bag, coupled with a greedy k-centers algorithm to generate the clustered
representation. Note that this is in contrast to MICA and Ammolite, which
both exploit domain knowledge to further improve performance. Thus, es-
FragBag only involves extending an existing codebase with new database
generation and similarity search functions.
We investigate the increases in speed resulting from directly applying
the entropy-scaling framework for both Euclidean and cosine distances and
found the acceleration is highly dependent on both the search radius and
cluster radius (Figure 3). For cosine distance, we generated databases with
maximum cluster radii of 0.1, 0.2, 0.3, 0.4, and 0.5. Then, for each query
protein from the set {4rhv, 1ake, 1bmf, 1rbp} (identified by PDB IDs), we
ran both naıve and accelerated similarity searches with radii of 0.02i,∀i ∈
{0, . . . , 49}. This test was repeated 5 times for each measurement, and the
ratio of average accelerated vs naıve times is shown in Figure 3a.
For Euclidean distance, we generated databases with maximum cluster
radii of 10, 20, 25, 50, and 100. Again, for each query protein drawn from
the same set, we compared the average over five runs of the ratio of average
accelerated versus naıve times (Figure 3b). The cluster generation required
anywhere from 65 to 23,714 seconds, depending on the choice of radii (See
table 3) and no more than a small constant (< 3) times as much memory
as it takes to simply load the PDB database (no more than 2 GB RAM).
Clustering used 20 threads on a 12-core Xeon X5690, while search used only
one thread.
16
Table 3: Cluster generation time for esFragBag
(a) Cosine distance:
radius
time (s)
0.1
0.2
21,037
11,088
0.3
7,409
0.4
5,288
0.5
3,921
(b) Euclidean distance:
radius
time (s)
10
20
23,714
3,062
30
483
40
144
50
65
Not only is the acceleration highly dependent on both the search radius
r and the maximum cluster radius rc, but the choice of query protein also
affects the results. We suspect that this effect is due to the geometry of pro-
tein fragment frequency space being very "spiky" and "star-like". Proteins
that are near the core (and thus similar to many other proteins) show very
little acceleration when our framework is used because the majority of the
database is nearby, whereas proteins in the periphery have fewer neighbors
and are thus found much more quickly. Changing the maximum cluster ra-
dius effectively makes more proteins peripheral proteins, but at the cost of
overall acceleration.
Naturally, as the search radius expands, it quickly becomes necessary
to compare against nearly the entire database, destroying any acceleration.
For the cosine space in particular, note that the maximum distance between
any two points is 1, so once the coarse search radius of r + rc ≥ 1.0, there
cannot ever be any acceleration as the fine search encompasses the entire
database. Similarly, once the coarse search encompasses all (or nearly all)
the clusters in Euclidean space, the acceleration diminishes to a factor 1,
and the overhead costs make the entropy-scaling framework perform worse
than a naıve search. However, as we are most interested in proteins that
are very similar to the query, the low-radius behavior is of primary interest.
In the low-radius regime, esFragBag demonstrates varying though substan-
tial acceleration (2-30x, averaging >10x for both distance functions for the
proteins chosen) over FragBag.
It is instructive to note that because of the very different geometries of
Euclidean vs cosine space, acceleration varies tremendously for some pro-
teins, such as 4rhv and 1bmf, which display nearly opposite behaviors.
Whereas there is nearly 30x acceleration for 4rhv in cosine space for low
radius, and the same for 1bmf in Euclidean space, neither achieves better
than ∼ 2.5x acceleration in the other space.
17
(a) Cosine distance
(b) Euclidean distance
Figure 3: Scaling behavior of esFragBag. EsFragBag benchmarking data
with parameters varied until the acceleration advantage of esFragBag dis-
appears. As search radius increases, the fraction of the database returned
by the coarse search increases, ultimately returning the whole database.
Unsurprisingly, when returning the whole database in the coarse search re-
sults, there are no benefits to using entropy-scaling frameworks. (a) Cosine
distance gives on the whole better acceleration, but results in > 99.8% sen-
sitivity, whereas (b) Euclidean distance as a metric is guaranteed by the
Triangle Inequality to get 100% sensitivity.
Table 4: Average sensitivity of esFragBag compared to FragBag when using
cosine distance for the trials described in Figure 3a. This table averages
the sensitivities for each choice of search radii {0, 0.01, . . . , 0.49}. (NB: no
analogous table is given for Euclidean distance as the Triangle Inequality
ensures perfect recall).
Cluster radii
Query protein
4rhv
1ake
1bmf
1rbp
0.10
0.20
0.30
0.40
0.50
1
1
1
1
1
0.999840
0.999918
0.999926
0.999974
0.999984
0.998490
0.999001
0.999649
0.999796
0.999934
0.999950
0.999978
1
1
1
Finally, while Euclidean distance is a metric -- for which the triangle
inequality guarantees 100% sensitivity -- cosine distance is not. Empirically,
however, for all of the queries we performed, we achieve > 99.8% sensitivity
(Table 4).
Application to other domains
We anticipate that our entropy-scaling approach will be useful to other
kinds of biological data sets; applying it to new data sets will require sev-
eral steps. Here we provide a "cookbook" for applying our entropy-scaling
framework to a new data set. Given a new data set, we first define what
the high-dimensional space is. For metagenomic sequence data, it is the
set of enumerable protein sequences up to some maximum length, while for
small-molecule data, it is the set of connected chemical graphs up to some
maximum size, and for protein structure data (using the FragBag model) it
is the set of "bag-of-words" frequency vectors of length 400.
Given the high dimensional space, we determine how database entries
map onto points (for example, in the case of MICA, they are greedily broken
into subsequences with a minimum length). Next, clustering can be imple-
mented; a simple greedy clustering may suffice (as for esFragBag) but clus-
tering of sequence data may be dramatically accelerated by using BLAST-
style seed-and-extend matching (as used in MICA). Finally, coarse and fine
search can be implemented; in many cases, existing tools may be used "out of
the box," as with esFragBag and MICA. With MICA, we note that coarse
search by default uses DIAMOND, while fine search provides a choice of
19
DIAMOND or BLASTX. With Ammolite, we used the SMSD library, but
incorporated it into our own search tool.
Discussion
We have introduced an entropy-scaling framework for accelerating ap-
proximate search, allowing search on large omics datasets to scale, even as
those datasets grow exponentially. The primary advance of this framework
is that it bounds both time and space as functions of the dataset entropy
(albeit using two different notions of entropy: metric entropy bounds time,
while information-theoretic entropy bounds space). We proved that run-
time scales linearly with the entropy of the database, but we also show
(Supplemental Methods: Theory) that under certain additional constraints,
this entropy-scaling framework permits a compressed representation on disk.
This compression is particularly applicable in the case of metagenomic anal-
ysis, where the collection of read data presents a major problem for storage
and transfer. Although we did not optimize for on-disk compression in any
of our applications, choosing instead to focus on search speed, implementing
this compression is feasible using existing software tools and libraries such
as Blocked GZip (BZGF); each cluster would be compressed separately on
disk.
Furthermore, we have justified and demonstrated the effectiveness of
this framework in three distinct areas of computational molecular biology,
providing the following open-source software: Ammolite for small-molecule
structure search, MICA for metagenomic analysis, and esFragBag for protein
structure search. All of our software is available under the GNU General
Public License, and not only can the tools we are releasing be readily plugged
into existing pipelines, but the code and underlying methods can also be
easily incorporated into the original software that we are accelerating.
The reason for the speedup is the combination of low fractal dimension
and low metric entropy. Low fractal dimension ensures that runtime is
dominated by metric entropy. The size of the coarse database provides an
estimate of metric entropy. Furthermore, we can directly measure the local
fractal dimension of the database by sampling points from the database and
looking at the scaling behavior of the number of points contained in spheres
of increasing radii centered on those sampled points. We have shown that
for three domains within biological data science, metric entropy, and fractal
dimension are both low.
As discussed in the theoretical results, although the data live locally on
a low dimension subspace, the data are truly high-dimensional globally. At
20
small scales, biological data often lives on a low-dimensional polytope (Hart
et al., 2015). However, omics data are, by nature, comprehensive and include
not just one but many such polytopes. Although each polytope can be
individually projected onto a subspace using techniques such as PCA, the
same projection cannot be used for all the polytopes at once because they
live on different low-dimensional subspaces. Furthermore, as is the case
with genomes, the low-dimensional polytopes are also often connected (e.g.,
through evolutionary history). Thus, collections of local projections become
unwieldy. By using our clustering approach, we are able to take advantage
of the existence of these low-dimensional polytopes for accelerated search
without having to explicitly characterize each one.
A hierarchical clustering approach, rather than our flat clustering, has
the potential to produce further gains (Loh et al., 2012). We have taken the
first steps in exploring this idea here; the molecule size clustering in Am-
molite can be thought of as an initial version of a multi-level or hierarchical
clustering.
Entropy-scaling search is related to succinct, compressed, and oppor-
tunistic data structures, such as the compressed suffix array, the FM-index,
and the sarray (Grossi & Vitter, 2005; Ferragina & Manzini, 2000; Conway
& Bromage, 2011). However, these solve the problem of theoretically fast
and scalable pattern matching (Box 1), whereas we solve, theoretically and
practically, the much more general similarity search problem. An entropy-
scaling search tree is also related to a metric ball tree (Uhlmann, 1991),
although with different time complexity. Querying a metric ball tree re-
quires O(log n) time, assuming the relatively uniform distribution of data
points in a metric space. This distribution differs from the non-uniform dis-
tribution under which entropy-scaling search behaves well. In future work,
we will investigate further acceleration of coarse search by applying a metric
ball tree to the cluster representatives themselves; this approach may reduce
the coarse search time to O(log k). This step, too, can be thought of as an
additional level of clustering.
Other metric search trees can also be found in the database literature
(Zezula et al., 2006), although, to our knowledge, they have not been ex-
plicitly applied to biological data science. The closest analogue to entropy-
scaling search trees is the M-tree (Ciaccia et al., 1997, 1998), which resembles
a multi-level variation of our entropy-scaling search trees. However, the M-
tree time-complexity analysis (Ciaccia et al., 1998) does not have a nice
closed form and is more explicitly dependent on the exact distribution of
points in the database. By using and combining the concepts of metric en-
tropy and fractal dimension for our analysis, we are able to give an easier to
21
understand and more intuitive, if somewhat looser, bound on entropy-scaling
search tree complexity.
Entropy-scaling frameworks have the advantage of becoming proportion-
ately faster and space efficient with the size of the available data. Although
the component pieces (e.g., the clustering method chosen) of the framework
can be either standard (as in esFragBag) or novel (as in Ammolite), the
key point is that these pieces are used in a larger framework to exploit
the underlying complex structure of biological systems, enabling massive
acceleration by scaling with entropy. We have demonstrated this scaling
behavior for common problems drawn from metagenomics, cheminformat-
ics, and protein structure search, but the general strategy can be applied
directly or with simple domain knowledge to a vast array of other problems
faced in data science. We anticipate that entropy-scaling frameworks should
be applicable beyond the life sciences, wherever physical or empirical laws
have constrained data to a subspace of low entropy and fractal dimension.
Methods
Ammolite small molecule search
Ammolite's clustering approach relies on structural similarity. We aug-
mented the entropy-scaling data structure by using a clustering scheme
based on molecular structural motifs instead of a distance function. Each
molecule is "simplified" by removing nodes and edges that do not participate
in simple cycles. Clusters are formed of molecules that are isomorphic after
this simplification step. Each cluster can then be represented by a single
molecular structure, along with pointers to "difference sets" between that
structure and each of the full molecules in the cluster it represents. For both
coarse and fine search, we use the Tanimoto distance metric, defined as
d(G1, G2) = 1 −
mcs(G1, G2)
G1 + G2 − mcs(G1, G2)
,
where mcs refers to the maximum common subgraph of two chemical graphs.
The coarse search is performed in compressed space, by searching the
coarse database with the goal of identifying possible hits. The query molecule
is simplified in exactly the same manner as the molecular database during
clustering, and this transformed query graph is matched against the coarse
database. To preserve sensitivity, this coarse search is performed with a
permissive similarity score. Any possible hits -- molecular graphs from the
coarse database whose MCS to the transformed query molecule was within
22
the similarity score threshold -- are then reconstructed by following point-
ers to the removed atom and bond information and recreating the original
molecules. Since the Tanimoto distance is used, we can bound the size of
candidate molecules based on the size of the query molecule and the desired
Tanimoto cutoff. Thus, a second level of clustering, at query time, based
on molecule size, allows further gains in runtime performance. Finally, the
fine search is performed against these decompressed possible hits that are
within the appropriate size range based on the Tanimoto distance cutoff.
MICA metagenomic search
CaBLASTX's clustering approach relies on sequence similarity. We aug-
mented the entropy-scaling data structure by using different distance func-
tions for clustering and search. For clustering, we rely on sequence identity,
while for search, we use the E-value measure that is standard for BLAST.
All benchmarks were performed with an E-value of 10−7. For coarse search,
MICA uses the DIAMOND argument --top 60 in order to return all queries
with a score within 60% of the top hit. When MICA was tested using
BLASTX for coarse search, it used an E-value of 1000. This seemingly sur-
prisingly large coarse E-value is used because E-values are poorly behaved
for short sequences; in sensitivity analysis, coarse E-values of 1 and 10 ex-
hibited recall below 10%, and an E-value of 100 exhibited recall below 60%.
Furthermore, during clustering (compression), we apply a preprocessing step
that identifies subsequences to be treated as distinct points in the database.
We apply a reversible alphabet reduction to the protein sequences, which
projects them into a subspace (Supplemental Methods).
When applied to high-coverage next-generation sequencing queries, ca-
BLASTX can also perform clustering on the reads (Supplemental Methods).
In this instance, coarse search is performed by matching each representa-
tive query with a set of representative database entries. Fine search then
matches the original queries within each cluster with the candidate database
entries resulting from the coarse search.
esFragBag protein structure search
In FragBag, the bag of fragments is essentially a term frequency vector
representing the number of occurrences of each structural motif within the
protein. FragBag turns out to be amenable to acceleration using an entropy-
scaling data structure because much of the computation is spent in doing a
similarity search on that frequency vector.
For the cluster generation, we trivially used a naıve randomized greedy
2-pass approach. First, all proteins in the Protein Data Bank were randomly
23
ordered. Then in the first pass, proteins were selected as cluster centers if
and only if they were not within a user-specified Euclidean distance rc from
an existing center (i.e., the first protein is always selected, and the second
if further away than rc from the first, etc.). Recall that this generation of
cluster centers is the same as the one used to generate covering spheres in
Figure 2; the covering spheres were overlapping, but we assign every protein
uniquely to a single cluster by assigning to the nearest cluster center in the
second pass.
Similarity search here was performed exactly as described in the sec-
tion "Entropy-Scaling Similarity Search", with no modification. For a given
search query q and search radius r, a coarse search was used to find all clus-
ter centers within distance r + rc of q. Then, all corresponding clusters were
unioned into a set F . Finally, a fine search was performed over the set F to
find all proteins within distance r of q.
Author Contributions
Y.W.Y., N.M.D., and B.B. conceived the project. Y.W.Y., N.M.D.,
and B.B. developed the theoretical analyses. N.M.D. and D.C.D. imple-
mented and benchmarked MICA. D.C.D. implemented and benchmarked
Ammolite, with help from N.M.D. and Y.W.Y. Y.W.Y. implemented and
benchmarked esFragBag, with help from N.M.D. B.B. guided all research
and provided critical advice on the study. Y.W.Y., N.M.D. and B.B. wrote
the manuscript.
Acknowledgments
Y.W.Y. is supported by a Hertz Foundation fellowship. N.M.D. and
B.B. are supported by NIH GM108348. We thank Andrew Gallant for his
implementation of Fragbag. We thank Joseph V. Barrile for graphic design.
We thank Jian Peng for suggesting high-throughput drug screening as an
application.
References
Altschul, S. F., Gish, W., Miller, W., Myers, E. W., & Lipman, D. J. (1990).
Basic local alignment search tool. Journal of molecular biology, 215, 403 --
410.
24
Berger, B., Peng, J., & Singh, M. (2013). Computational solutions for omics
data. Nature Reviews Genetics, 14, 333 -- 346.
Bolton, E. E., Wang, Y., Thiessen, P. A., & Bryant, S. H. (2008). Pubchem:
integrated platform of small molecules and biological activities. Annual
reports in computational chemistry, 4, 217 -- 241.
Bredel, M., & Jacoby, E. (2004). Chemogenomics: an emerging strategy for
rapid target and drug discovery. Nature Reviews Genetics, 5, 262 -- 275.
Buchfink, B., Xie, C., & Huson, D. H. (2015). Fast and sensitive protein
alignment using DIAMOND. Nature methods, 12, 59 -- 60.
Budowski-Tal, I., Nov, Y., & Kolodny, R. (2010). FragBag, an accurate
representation of protein structure, retrieves structural neighbors from the
entire PDB quickly and accurately. Proceedings of the National Academy
of Sciences, 107, 3481 -- 3486.
Bunke, H., & Shearer, K. (1998). A graph distance metric based on the
maximal common subgraph. Pattern recognition letters, 19, 255 -- 259.
Cao, Y., Jiang, T., & Girke, T. (2008). A maximum common substructure-
based algorithm for searching and predicting drug-like compounds. Bioin-
formatics, 24, i366 -- i374.
Ciaccia, P., Patella, M., & Zezula, P. (1997). Deis-csite-cnr. In Proceedings
of the... International Conference on Very Large Data Bases (p. 426).
Morgan Kaufmann Pub volume 23.
Ciaccia, P., Patella, M., & Zezula, P. (1998). A cost model for similarity
queries in metric spaces. In Proceedings of the seventeenth ACM SIGACT-
SIGMOD-SIGART symposium on Principles of database systems (pp. 59 --
68). ACM.
Conway, T. C., & Bromage, A. J. (2011). Succinct data structures for
assembling large genomes. Bioinformatics, 27, 479 -- 486.
Daniels, N. M., Gallant, A., Peng, J., Cowen, L. J., Baym, M., & Berger,
B. (2013). Compressive genomics for protein databases. Bioinformatics,
29, i283 -- i290.
David, L. A., Materna, A. C., Friedman, J., Campos-Baptista, M. I., Black-
burn, M. C., Perrotta, A., Erdman, S. E., & Alm, E. J. (2014). Host
lifestyle affects human microbiota on daily timescales. Genome Biol, 15,
R8.
25
Falconer, K. (1990). Fractal geometry: mathematical foundations and appli-
cations. John Wiley & Sons.
Ferhatosmanoglu, H., Tuncel, E., Agrawal, D., & El Abbadi, A. (2000). Vec-
tor approximation based indexing for non-uniform high dimensional data
sets. In Proceedings of the ninth international conference on Information
and knowledge management (pp. 202 -- 209). ACM.
Ferragina, P., & Manzini, G. (2000). Opportunistic data structures with
In Foundations of Computer Science, 2000. Proceedings.
applications.
41st Annual Symposium on (pp. 390 -- 398). IEEE.
Frank, D. N., & Pace, N. R. (2008). Gastrointestinal microbiology enters
the metagenomics era. Current opinion in gastroenterology, 24, 4 -- 10.
Gire, S. K., Goba, A., Andersen, K. G., Sealfon, R. S., Park, D. J., Kanneh,
L., Jalloh, S., Momoh, M., Fullah, M., Dudas, G. et al. (2014). Genomic
surveillance elucidates ebola virus origin and transmission during the 2014
outbreak. Science, 345, 1369 -- 1372.
Grossi, R., & Vitter, J. S. (2005). Compressed suffix arrays and suffix trees
with applications to text indexing and string matching. SIAM Journal on
Computing, 35, 378 -- 407.
Hart, Y., Sheftel, H., Hausser, J., Szekely, P., Ben-Moshe, N. B., Korem,
Y., Tendler, A., Mayo, A. E., & Alon, U. (2015).
Inferring biological
tasks using pareto analysis of high-dimensional data. Nature methods,
12, 233 -- 235.
Hegyi, H., & Gerstein, M. (1999). The relationship between protein struc-
ture and function: a comprehensive survey with application to the yeast
genome. Journal of molecular biology, 288, 147 -- 164.
Huson, D. H., Mitra, S., Ruscheweyh, H.-J., Weber, N., & Schuster, S. C.
Integrative analysis of environmental sequences using megan4.
(2011).
Genome research, 21, 1552 -- 1560.
Indyk, P., & Motwani, R. (1998). Approximate nearest neighbors: towards
removing the curse of dimensionality.
In Proceedings of the thirtieth
annual ACM symposium on Theory of computing (pp. 604 -- 613). ACM.
Kahn, S. D. (2011). On the future of genomic data. Science(Washington),
331, 728 -- 729.
26
Kanehisa, M., & Goto, S. (2000). KEGG: kyoto encyclopedia of genes and
genomes. Nucleic acids research, 28, 27 -- 30.
Kent, W. J. (2002). BLAT-the BLAST-like alignment tool. Genome re-
search, 12, 656 -- 664.
Langille, M. G., Zaneveld, J., Caporaso, J. G., McDonald, D., Knights,
D., Reyes, J. A., Clemente, J. C., Burkepile, D. E., Thurber, R. L. V.,
Knight, R., & Huttenhower, C. (2013). Predictive functional profiling of
microbial communities using 16S rRNA marker gene sequences. Nature
biotechnology, 31, 814 -- 821.
Linial, M., Linial, N., Tishby, N., & Yona, G. (1997). Global self-
organization of all known protein sequences reveals inherent biological
signatures. Journal of molecular biology, 268, 539 -- 556.
Loh, P.-R., Baym, M., & Berger, B. (2012). Compressive genomics. Nature
biotechnology, 30, 627 -- 630.
MacFabe, D. F. (2012). Short-chain fatty acid fermentation products of the
implications in autism spectrum disorders. Microbial
gut microbiome:
ecology in health and disease, 23.
Mackelprang, R., Waldrop, M. P., DeAngelis, K. M., David, M. M., Chavar-
ria, K. L., Blazewicz, S. J., Rubin, E. M., & Jansson, J. K. (2011). Metage-
nomic analysis of a permafrost microbial community reveals a rapid re-
sponse to thaw. Nature, 480, 368 -- 371.
Marx, V. (2013). Biology: The big challenges of big data. Nature, 498,
255 -- 260.
Menke, M., Berger, B., & Cowen, L. (2008). Matt:
local flexibility aids
protein multiple structure alignment. PLoS computational biology, 4,
e10.
Nepomnyachiy, S., Ben-Tal, N., & Kolodny, R. (2014). Global view of the
protein universe. Proceedings of the National Academy of Sciences, 111,
11691 -- 11696.
Pruitt, K. D., Tatusova, T., & Maglott, D. R. (2005). NCBI reference se-
quence (RefSeq): a curated non-redundant sequence database of genomes,
transcripts and proteins. Nucleic acids research, 33, D501 -- D504.
27
Rahman, S. A., Bashton, M., Holliday, G. L., Schrader, R., & Thornton,
J. M. (2009). Small molecule subgraph detector (SMSD) toolkit. Journal
of Cheminformatics, 1, 1 -- 13.
Sayers, E. W., Barrett, T., Benson, D. A., Bolton, E., Bryant, S. H., Canese,
K., Chetvernin, V., Church, D. M., DiCuccio, M., Federhen, S. et al.
(2011). Database resources of the national center for biotechnology infor-
mation. Nucleic acids research, 39, D38 -- D51.
Schaeffer, S. E. (2007). Graph clustering. Computer Science Review, 1,
27 -- 64.
Segata, N., Waldron, L., Ballarini, A., Narasimhan, V., Jousson, O., & Hut-
tenhower, C. (2012). Metagenomic microbial community profiling using
unique clade-specific marker genes. Nature methods, 9, 811 -- 814.
Shindyalov, I. N., & Bourne, P. E. (1998). Protein structure alignment by
incremental combinatorial extension (CE) of the optimal path. Protein
engineering, 11, 739 -- 747.
Subbiah, S., Laurents, D., & Levitt, M. (1993). Structural similarity of
DNA-binding domains of bacteriophage repressors and the globin core.
Current Biology, 3, 141 -- 148.
Tao, T. (2008). Product set estimates for non-commutative groups. Com-
binatorica, 28, 547 -- 594.
Tyson, G. W., Chapman, J., Hugenholtz, P., Allen, E. E., Ram, R. J.,
Richardson, P. M., Solovyev, V. V., Rubin, E. M., Rokhsar, D. S., &
Banfield, J. F. (2004). Community structure and metabolism through
reconstruction of microbial genomes from the environment. Nature, 428,
37 -- 43.
Uhlmann, J. K. (1991). Satisfying general proximity/similarity queries with
metric trees. Information processing letters, 40, 175 -- 179.
Ukkonen, E. (1985). Algorithms for approximate string matching. Informa-
tion and control, 64, 100 -- 118.
Weber, R., Schek, H.-J., & Blott, S. (1998). A quantitative analysis and per-
formance study for similarity-search methods in high-dimensional spaces.
In VLDB (pp. 194 -- 205). volume 98.
28
Yona, G., Linial, N., & Linial, M. (1999). Protomap: automatic classification
of protein sequences, a hierarchy of protein families, and local maps of
the protein space. Proteins: Structure, Function, and Bioinformatics, 37,
360 -- 378.
Yu, Y. W., Yorukoglu, D., Peng, J., & Berger, B. (2015). Quality score
compression improves genotyping accuracy. Nature Biotechnology, 33,
240 -- 243.
Zezula, P., Amato, G., Dohnal, V., & Batko, M. (2006). Similarity search:
the metric space approach volume 32. Springer Science & Business Media.
Zhao, Y., Tang, H., & Ye, Y. (2012). RAPSearch2: a fast and memory-
efficient protein similarity search tool for next-generation sequencing data.
Bioinformatics, 28, 125 -- 126.
29
Supplemental Methods
Theory
Time-complexity
We introduced the definition of the entropy-scaling similarity search data
structure in Figure 1. For ease of analysis, we will work in a high-dimensional
metric space and consider the database as a set D of n unique points in that
metric space. Define BS(q, r) = {p ∈ S : q − p < r}. The similarity
search problem is thus to compute BD(q, r) for a query q and radius r. Note
however that the metricity requirement is needed only for a 100% sensitivity
guarantee; other distance functions can be used, but result in some loss in
sensitivity. However, regardless of the distance function chosen, there cannot
be a loss of specificity; false positives will never be introduced because the
fine search is just the original search function on a smaller subset of the
database.
A set C of k cluster centers are chosen such that no cluster has radius
greater than a user-specified parameter rc and no two cluster centers are
within distance rc of one another. The data structure then clusters the points
in the set by assigning them to their nearest cluster center. Overloading
notation a bit, we will identify each cluster with its center, so C is also
the set of clusters. For a given similarity search query for all items within
distance r of a query q, this data structure breaks the query into coarse
and fine search stages. The coarse search is over the list of cluster centers,
returning BC(q, r + rc). Let
F = [c∈BC (q,r+rc)
c,
the union of all the returned clusters. By the Triangle Inequality, BD(q, r) ⊆
F , which combined with F ⊆ D implies that BF (q, r) = BD(q, r). Thus, a
fine search over the set F will return all items within radius r of q.
Note that we require the metricity requirement only for the Triangle
Inequality.
It turns out that many interesting distance functions are not
metrics, but still almost satisfy the Triangle Inequality, which is nearly suf-
ficient. More precisely, if a fraction α of the triples in S do not satisfy the
Triangle Inequality, then in expectation, we will have sensitivity 1 − α. As
shown in the results, empirically, this loss in sensitivity appears to be low
and can likely be ameliorated by increasing the coarse search radius.
Provided the fractal dimension of the database is low, and given a few
technical assumptions on the distribution of points, this data structure al-
S1
arXiv:1503.05638v2 [cs.DS] 21 Sep 2015
lows for similarity search queries in time roughly linear in the metric en-
tropy of the database. Additionally, without increasing the asymptotic
time-complexity, this data structure can also be stored in an information
theoretic entropy-compressed form.
Note that entropy-scaling data structures are distinct from both succinct
data structures and compressed data structures. Succinct data structures
are ones that use space close to the information-theoretic limit in the worst
case while permitting efficient queries; i.e. succinct data structures do not
depend on the actual entropy of the underlying data set, but have size-
dependence on the potential worst-case entropy of the data set (Jacobson,
1988). Compressed (and opportunistic) data structures, on the other hand,
bound the amount of the space used by the entropy of the data set while per-
mitting efficient queries (Grossi & Vitter, 2005; Ferragina & Manzini, 2000).
Entropy-scaling data structures are compressed data structures, but are dis-
tinct, as unlike entropy-scaling data structures, compressed data structures
do not measure time-complexity in terms of metric entropy. Additionally,
existing compressed data structures such as the compressed suffix array and
the FM-index are designed for the problem of pattern matching (Grossi &
Vitter, 2005; Ferragina & Manzini, 2000). While related to similarity search,
pattern matching does not admit as general of a notion of distance as the
similarity search problem. While compressed sensing has also been applied
to the problem of finding a representative set of genes for a collection of
expression samples (Prat et al., 2011), compressed sensing is distinct from
entropy-scaling data structures.
The primary advance of entropy-scaling data structures is that
they bound both space and time as functions of the data set en-
tropy (albeit using two different notions of entropy).
Unifying different notions of fractal dimension
In this paper, we make use of two different notions of fractal dimension,
including the more intuitive formulation in the main paper and a more
classical definition in this following section.
In order to both unify these
different notions of fractal dimension and to allow us to prove our complexity
bounds, we make a technical assumption about the self-similarity of the
data set. Roughly speaking, we require that the density of points be similar
throughout the data set.
More precisely, for a given radius, consider the number of neighbors
about any point for that radius, which we will call the density around a
point. We assume that for any given radius, the densities around all points
are bounded within a constant multiplicative factor γ of one another. This
S2
particular technical assumption is likely stronger than we need, but makes
the following arguments much more convenient; we conjecture but do not
prove here that so long as the distribution of densities has low variance,
all our statements below also hold with high probability. We give special
thanks to one of our commenters for having brought this subtlety to our
attention.
Complexity bounds
We first define the concept of metric entropy and entropy dimension in
the classical manner:
Definition 1 ((Tao, 2008) Definition 6.1). Let X be a metric space, let D
be a subset of X, and let ρ > 0 be a radius.
• The metric entropy Nρ(D) is the fewest number of points x1, . . . , xn ∈
D such that the balls B(x1, ρ), . . . B(xn, ρ) cover D.
Definition 2 ((Falconer, 1990)). The Hausdorff dimension of a set D is
given by
dimHausdorff(D) := lim
ρ→0
log Nρ(D)
log 1/ρ
Unfortunately, as D is a finite, discrete, set, the given definision always
gives dimHausdorff(D) = 0. However, we are only interested in scaling behav-
iors around large radii, so instead we use:
Definition 3. Define fractal dimension d of a set D at a scale [ρ1, ρ2] by
log Nρ(D)
Nρ1 (D)
log ρ1
ρ
Intuitively, this means that when we double the radii, the metric entropy,
or number of covering spheres needed, decreases by a multiplicative factor of
2d. This definition of fractal dimension is classical, but the reader will note
also different in formulation from the intuitive one given in the main text.
However, the two notions are related given the bounded density assumption
we made, from which immediately follows a bound on the number of points
within each cluster. On average, when we double the radius of a sphere
around a point, the number of points in the larger sphere is roughly the
number of points in the smaller sphere multiplied by 2d, because otherwise
the spheres could not cover the space. This latter behavior is what we
measure when we talk about local fractal dimension around a point in the
d = max
ρ∈[ρ1,ρ2]
S3
main paper. However, because the number of points within each cluster must
remain within a multiplicative constant of each other by the bounded density
assumption, not only is this true on average, but the increase in number of
points within a cluster has to be roughly uniform across all clusters. Thus,
this scaling behavior of points in a doubled sphere must be true everywhere
in the data set, and thus we can connect the global average local fractal
dimension alluded to in the main paper and measured for our data sets of
interest to this more classical notion of fractal dimension based on covering
spheres.
Recall that k entries are selected as cluster centers for partitioning the
database to result in clusters with maximum radius rc. From the definition
above, when setting ρ = rc, it is trivial to verify k ≤ Nrc(D). This upper
bound is guaranteed by our requirement that the cluster centers not be
within distance rc.
Given any query q, the coarse search over the cluster centers always
requires k comparisons. Additionally, the fine search is over the set F ,
defined to be the union of clusters with centers within distance r + rc from
q. As the time-complexity of similarity search is just the total of the coarse
and fine searches, this implies that the total search time is O(k + F).
By the triangle inequality, F ⊂ BD(q, r + 2rc), so we can bound F ≤
BD(q, r + 2rc). Let the fractal dimension D at the scale between rc and
2rc + r be d. Recall that the local fractal dimension determines how many
more points we hit when we scale the radius of a sphere. Also, we use the
density bound to give us that
BD(p, ρ) ≤ Eq BD(q, ρ) ≤ γ BD(p, ρ)
1 γ
for any choice of point p and radius ρ. Then
Eq [BD(q, r + 2rc)] ≤ γ BD(p, r + 2rc)
≤ γ BD(p, r)(cid:18) r + 2rc
r (cid:19)d
≤ γ2Eq [BD(q, r)](cid:18) r + 2rc
r (cid:19)d
,
(1)
(2)
(3)
because we are measuring the relative number of points found in spheres of
radius r vs radius r + 2rc respectively for some particular point p. But γ is
a constant, and disappears in asymptotic notation. Thus, total search time
S4
is in expectation over points q
O k + BD(q, r)(cid:18) r + 2rc
r (cid:19)d! .
However, note that k is linear in metric entropy and BD(q, r) is the output
size, so similarity search can be performed in time linear to metric entropy
and a polynomial factor of output size. Provided that the fractal dimension
d is small and k is large, the search time will be dominated by the metric
entropy component, which turns out to be the regime of greatest interest
for us. We have thus proven bounds for the time-complexity of similarity
search, given a self-similarity condition on the density of the dataset.
Space-complexity
Here we relate the space-complexity of our entropy-scaling similarity
search data structure to information-theoretic entropy. Traditionally, information-
theoretic entropy is a measure of the uncertainty of a distribution or random
variable and is not well-defined for a finite database. However, the notion of
information-theoretic entropy is often used in data compression as a short-
hand for the number of bits needed to encode the database, or a measure
of the randomness of that database. We use entropy in the former sense;
precisely, we define the entropy of a database as the number of bits needed
to encode that database, a standard practice in the field. Thus, we consider
entropy-compressed forms of the original database, such as that obtained
by Prediction by Partial Matching (PPM), Lempel-Ziv compression (e.g.
Gzip), or a Burrows-Wheeler Transform (as in Bzip2), and use their size as
an estimate of the entropy Sorig of the database.
For all commonly used compression techniques, decompression time is
linear in the size of the uncompressed data. Obviously, even with linear
decompression, decompressing the entire database for each similarity search
would squander the entropy-scaling benefits of our approach. However, note
that the fine search detailed above only needs access to a subset of clusters
and furthermore needs full access to that set of clusters.
It is therefore
always asymptotically 'free' to decompress an entire cluster at once, if any
member of that cluster needs to accessed. Thus, one ready solution is to
simply store entropy-compressed forms of each cluster separately.
Compressing each cluster separately preserves runtime bounds, but makes
it difficult to compare the compressed clustered database size to the original
compressed database size. This results from the possibility that redundancy
across clusters that would originally have been exploited by the compressor
S5
can no longer be exploited once the database is partitioned. Intuitively, for
any fixed-window or block compressor, grouping together similar items into
clusters should increase the performance of the compressor, but it is unclear
a priori if that balances out the loss of redundancy across clusters.
A somewhat more sophisticated solution is to reorder the entries of the
database by cluster, compress the entire database, and then store indexes
into the starting offset of each cluster. For popular tools such as Gzip
or Bzip2, this is possible with constant overhead κ per index. Because
the entire database is still being compressed, redundancy across clusters
can be exploited to reduce compressed size, while still taking advantage of
similar items being grouped together. Thus, in expectation over uniformly-
randomly chosen orderings of the database entries (obviously, there is some
optimal ordering, but computing that is computationally infeasible), the
compressed clustered database size Sclust ≤ Sorig. Then, total expected
space-complexity of our data structure is O(κk + Sorig); recall here that k is
the number of clusters and is bounded by the metric entropy of the database.
Thus, space complexity is linear in metric entropy plus information-theoretic
entropy.
Additionally, given that our distance function measures marginal information-
theoretic entropies, we can also give a bound on the total information-
theoretic entropy of the database by using metric entropy and the cluster
radius. Let l be the maximum distance of two points in the space. The
naıve upper bound on total entropy is then O(nl), where n is the total
number of points in the database, because distance and entropy are re-
lated. Recall that we chose k points as cluster centers, where k is bounded
by metric entropy, for a maximum cluster radius rc. Encoding each non-
center point p as a function of the nearest cluster center requires O(nrc)
bits. Specifying the privileged points again requires O(kl) bits, so together
the total information-theoretic entropy is O(kl + nrc). In other words, not
only is space complexity linear in metric entropy plus information-theoretic
entropy, but information-theoretic entropy itself is also bounded by the low-
dimensional coarse structure of the database.
Clustering time complexity
Although clustering the database is a one-time cost that can further be
amortized over future queries, we still require that cluster generation be
tractable. Here we present a trivial O(kn) algorithm for cluster generation
with clusters of maximum radius rc that appears to work sufficiently well in
practice (and is the algorithm used in esFragBag):
S6
• Initialize an empty set of cluster centers C. Let δ(x, C) be the distance
from a point x to C, defined to be ∞ if C = ∅.
• Randomly order the n entries of the database D = {d1, . . . , dn}
• For i = 1, . . . n,
-- If δ(di, C) > rc, append di to C.
• For i = 1, . . . n,
-- Assign di to the cluster represented by the nearest item in C.
Because we need to compare each of n items against up to k items in C in
each of the for loops, this trivial algorithm takes O(kn) time.
Additionally, insertions can be performed in O(k) time. For a new entry
dn+1, if δ(dn+1, C) ≤ rc, assign dn+1 to the cluster represented by the nearest
item in C. Otherwise, append dn+1 to C as a new cluster center. This clearly
requires exactly k comparisons to do. Note that with insertions of this kind,
items are no longer guaranteed to be assigned to the nearest cluster center;
however, they are still guaranteed to be assigned to some cluster center
within distance rc, which is all that is needed for entropy-scaling to work.
Deletions are slightly more complicated. If the entry to be deleted is not
a cluster center, then removing it takes constant time. However, if it is a
cluster center, we effectively have to remove the entire cluster and reinsert
all the non-center elements, which will take O(k · [size of cluster]). Thus, in
expectation over a uniform random choice of item to be deleted, deletions
can also be performed in O(k) time.
Ammolite
Simplification and compression
Given a molecular graph, any vertex or edge that is not part of a simple
cycle or a tree is removed, and any edge that is part of a tree is removed.
This preserves the node count, but not the topology, of tree-like structures,
and preserves simple cycles, which represent rings in chemical compounds.
For example, as shown in Figure S2, both caffeine and adenine would be
reduced to a purine-like graph.
After this transformation is applied to each molecule in a database to be
compressed, we identify all clusters of fully-isomorphic transformed molec-
ular graphs. Isomorphism detection is performed using the VF2 (Cordella
et al., 2001) algorithm; a simple hash computed from the number of ver-
tices and edges in each transformed molecular graph is first used to filter
S7
molecular graphs that cannot possibly be isomorphic. A representative from
each such cluster is stored in SDF format; collectively, these representatives
form a "coarse" database. Along with each representative, we preserve the
information necessary to reconstruct each original molecule, as a pointer to
a set of vertices and edges that have been removed or unlabeled.
Ammolite is implemented in Java, and its source code is available on
Github.
MICA
Alphabet Reduction
Alphabet reduction -- reducing the 20-letter standard amino acid alpha-
bet to a smaller set, in order to accelerate search or improve homology
detection -- has been proposed and implemented several times (Bacardit et al.,
2007; Peterson et al., 2009). In particular, Murphy et al. (2000) considered
reducing the amino-acid alphabet to 17, 10, or even 4 letters. More recently,
Zhao et al. (2012) and Huson & Xie (2013) applied a reduction to a 4-letter
alphabet, termed a "pseudoDNA" alphabet, in sequence alignment.
When using BLASTX for coarse search (which we call caBLASTX), we
extend the compression approach of Daniels et al. (2013) using a reversible
alphabet reduction. We use the alphabet reduction of Murphy et al. (2000)
to map the standard amino acid alphabet (along with the four common
ambiguous letters ) onto a 4-letter alphabet. Specifically, we map F, W,
and Y into one cluster; C, I, L, M, V, and J into a second cluster, A, G,
P, S, and T into a third cluster, and D, E, N, Q, K, R, H, B, and Z into
a fourth cluster. By storing the offset of the original letter within each
cluster, the original sequence can be reconstructed, making this a reversible
reduction. This alphabet reduction is not used when using DIAMOND for
coarse search, as DIAMOND already relies on its own alphabet reduction.
Database Compression
Given a protein sequence database to be compressed, we proceed as
follows:
1. First, initialize a table of all possible k-mer seeds of our (possibly 4-
letter reduced) alphabet, as well as a coarse database of sequences,
initially containing the (possibly reduced-alphabet) first sequence in
the input database.
2. For each k-mer of the first sequence, store its position in the corre-
sponding entry in the seed table.
S8
4.
3. For each subsequent sequence s in the input, reduce its alphabet and
slide a window of length k along the sequence, skipping single-letter
repeats of length greater than 10.
(a) Look up these k residues in the seed table. For every entry match-
ing to that k-mer in the seed table, follow it to a corresponding
subsequence in the coarse database and attempt extension (de-
fined below).
If no subsequences from this window can be ex-
tended, move the window by m positions, where m defaults to
20.
(b) If a match was found via extension, move the k-mer window to
the first k-mer in s after the match, and repeat the extension
process.
Given a k-mer in common between sequence s and a subsequence s0
pointed to by the seed table, first attempt ungapped extension:
1. Within each window of length m beginning with a k-mer match, if
there are at least 60% matches between s and s0, then there is an
ungapped match.
2. Continue ungapped matching using m-mer windows until no more m-
mers of at least 60% sequence identity are found.
3. The result of ungapped extension is that there is an alignment between
s and s0 where the only differences are substitutions, at least 60% of
the positions contain exact matches.
When ungapped extension terminates, attempt gapped extension. From
the end of the aligned regions thus far, align 25-mer windows of both s and s0
using the Needleman-Wunsch (Needleman & Wunsch, 1970) algorithm using
an identity matrix. Note that the original caBLASTP (Daniels et al., 2013)
used BLOSUM62 as it was operating in amino acid space; as we are now
operating in a reduced-alphabet space, an identity matrix is appropriate,
just as it is for nucleotide space. After gapped extension on a window length
of 25, attempt ungapped extension again.
If neither gapped nor ungapped extension can continue, end the exten-
sion phase. If the resulting alignment has less than 70% sequence identity
(in the reduced-alphabet space), or is shorter than 40 residues, discard it,
and attempt extension on the next entry in the seed table for the original
k-mer, continuing on to the next k-mer if there are no more entries.
If the resulting alignment does have at least 70% sequence identity in
the reduced-alphabet space, and is at least 40 residues long, then create
a link from the entry for s0 in the coarse database to the subsequence of s
S9
corresponding to the alignment. If there are unaligned ends of s shorter than
30 residues, append them to the match. Longer unaligned ends that did not
match any subsequences reachable from the seed table are added into the
coarse database themselves, following the same k-mer indexing procedure as
the first sequence.
Finally, in order to be able to recover the original sequence with its
original amino acid identities, a difference script is associated with each link.
This difference script is a representation of the insertions, deletions, and
substitutions resulting from the Needleman-Wunsch alignment, along with
(if alphabet reduction is used) the offset in each reduced-alphabet cluster
needed to recover the original alphabet. Thus, for example, a valine (V) is
in the cluster containing C, I, L, M, V, and J. Since it is the 4th entry in
that 5-entry cluster, we can represent it with the offset 4. Since the largest
cluster contains 9 elements, only four bits are needed to store one entry
in the difference script. More balanced clusters would have allowed 3-bit
storage, but at the expense of clusters that less faithfully represented the
BLOSUM62 matrix and the physicochemical properties of the constituent
amino acids.
Because of the seed table, compression is memory-intensive and CPU-
intensive. Compressing the September, 2014 NCBI NR database required
approximately 39 hours on a 12-core Xeon with 128GB RAM.
Query Clustering
Metagenomic reads are themselves nucleotide sequences, so no alphabet
reduction is performed on them directly. When BLASTX is used for coarse
search (caBLASTX), MICA relies on query-side clustering. Metagenomic
reads are compressed using the same approach as the protein database,
without the alphabet reduction step and with a number of different parame-
ters. The difference scripts for metagenomic reads do not rely on the cluster
offsets, but simply store the substituted nucleotides.
Furthermore, unlike protein databases, where most typical sequences
range in length from 100 to over 1000 amino acids, next-generation sequenc-
ing reads are typically short and usually of fixed length, which is known
in advance. Thus, the minimum alignment length required for a match,
and the maximum length unaligned fragment to append to a match, require
different values based on the read length.
An additional complication is that insertions and deletions from one read
to another will change the reading frame, potentially resulting in different
amino acid sequences. For this reason, query clustering requires long, un-
gapped windows of high sequence identity. Specifically, for 202-nucleotide
S10
reads, for two sequences to cluster together, we require a 150-nucleotide
ungapped region of at least 80% sequence identity.
We note that unlike the compression of the database, which can be amor-
tized over future queries, the time spent clustering and compressing the
queries cannot be amortized. Thus, we would not refer to the query clus-
tering as entropy-scaling, but it still provides a constant speed-up. For this
reason, we include the time spent clustering and compressing queries in the
search time for MICA. When using DIAMOND for coarse search, MICA
does not perform query-side clustering, and instead relies on DIAMOND's
indexing of the queries.
Search
Given a compressed protein database and a compressed query read set,
search comprises two phases. The first, coarse search, considers only the
coarse sequences -- the representatives -- resulting from compression of the
protein database and the query set. When BLASTX is used for the coarse
search (caBLASTX), each coarse nucleotide read is transformed into each of
the six possible amino acid sequences that could result from it (three reading
frames for both the sequence and its reverse complement). Then, each of
these amino acid sequences is then reduced back to a four-letter alphabet
using the same mapping as for protein database compression. For conve-
nience, the four-letter alphabet is represented using the standard nucleotide
bases, though this has no particular biological significance. This is done so
that the coarse search can rely on BLASTN (nucleotide BLAST) to search
these sequences against the compressed protein database.
For each coarse query representative (identified using a a coarse E-
value of 1000, along with the BLASTN arguments -task blastn-short
-penalty -1; these arguments are recommended by the NCBI BLAST+
manual when queries are short), the set of coarse hits is used to reconstruct
all corresponding sequences from the original database by following links
to original sequence matches and applying their difference scripts. The re-
sulting candidates are thus original sequences from the protein database, in
their original amino acid alphabet. The query representative is also used to
reconstruct all corresponding sequences from the original read set. Thus, for
each coarse query representative, there is now a subset of the metagenomic
read set (the reads represented by that coarse query) and also a subset of
the protein database (the candidates).
When DIAMOND is used for coarse search, instead of an E-value thresh-
old, the argument --top 60 is used; this causes DIAMOND to return all
coarse hits whose score is within 60% of the top-scoring hit. Without this
S11
argument, DIAMOND defaults to returning at most 25 hits for each query
sequence, which would result in significant loss of recall.
The second phase, fine search, uses standard DIAMOND or BLASTX to
translate each of these reads associated with a coarse query representative
and search for hits only in the subset of the database comprising the candi-
dates. This fine search phase relies on a user-specified E-value threshold (or
other user-specified parameters to DIAMOND or BLASTX) to filter hits.
To ensure that E-value calculation is correct, the call to BLASTX uses a cor-
rected database size which is the size of the original, uncompressed protein
database.
Benchmarking
Although our primary result is the direct acceleration of DIAMOND us-
ing our entropy-scaling data structures, we also compared MICA to RapSearch2 (Zhao
et al., 2012) version 2.22 and the November 29, 2014 version of DIAMOND (Buchfink
et al., 2015). All tests were performed on a 12-core Intel Xeon X5690 running
at 3.47GHz with 88GB RAM and hyperthreading; 24 threads were allowed
for all programs. Diamond was run with the --sensitive option. In all
cases, an E-value threshold of 1e-7 was used.
For the raw-read dataset, we filtered out reads starting or ending with
10 or more no-calls ('N').
MICA is implemented in Go, and its source code is available on Github.
esFragBag
We took the existing FragBag method as a black box and by design
did not do anything clever in esFragBag except apply the entropy-scaling
similarity search data structure. We used a Go language implementation
of FragBag, written by Andrew Gallant. Additionally, we removed the
sorting-by-distance feature of Andrew Gallant's FragBag search implemen-
tation, which does not improve the all-matching results we were interested in
here -- it lowers k-nearest neighbor search memory requirements while dom-
inating the running time of ρ-nearest neighbor, the problem at hand. This
was done for both the FragBag and the esFragBag benchmarks, to ensure
comparability. All code was written in Go, and is available on Github.
The entire 2014 Oct 31 version of the Protein Data Bank was down-
loaded and the database was composed of fragment frequency vectors gen-
erated from all of the relevant PDB files using the 400-11.json fragment list
(Budowski-Tal et al., 2010). For this paper, we implemented the benchmark-
ing in Go, and have provided the source code for the benchmarking routine
on Github. This allowed us to benchmark just the search time, excluding
S12
the time to load the database from disk. Note that the prototype imple-
mentation of esFragBag available only supports the all ρ-nearest neighbor
search query found in FragBag.
Supplemental References
Bacardit, J., Stout, M., Hirst, J. D., Sastry, K., Llor`a, X., & Krasnogor, N.
(2007). Automated alphabet reduction method with evolutionary algo-
rithms for protein structure prediction. In Proceedings of the 9th annual
conference on Genetic and evolutionary computation (pp. 346 -- 353). ACM.
Buchfink, B., Xie, C., & Huson, D. H. (2015). Fast and sensitive protein
alignment using DIAMOND. Nature methods, 12, 59 -- 60.
Budowski-Tal, I., Nov, Y., & Kolodny, R. (2010). FragBag, an accurate
representation of protein structure, retrieves structural neighbors from the
entire PDB quickly and accurately. Proceedings of the National Academy
of Sciences, 107, 3481 -- 3486.
Cordella, L. P., Foggia, P., Sansone, C., & Vento, M. (2001). An improved
In 3rd IAPR-TC15 workshop on
algorithm for matching large graphs.
graph-based representations in pattern recognition (pp. 149 -- 159).
Daniels, N. M., Gallant, A., Peng, J., Cowen, L. J., Baym, M., & Berger,
B. (2013). Compressive genomics for protein databases. Bioinformatics,
29, i283 -- i290.
Falconer, K. (1990). Fractal geometry: mathematical foundations and appli-
cations. John Wiley & Sons.
Ferragina, P., & Manzini, G. (2000). Opportunistic data structures with
In Foundations of Computer Science, 2000. Proceedings.
applications.
41st Annual Symposium on (pp. 390 -- 398). IEEE.
Grossi, R., & Vitter, J. S. (2005). Compressed suffix arrays and suffix trees
with applications to text indexing and string matching. SIAM Journal on
Computing, 35, 378 -- 407.
Huson, D. H., & Xie, C. (2013). A poor man's BLASTX-high-throughput
metagenomic protein database search using PAUDA. Bioinformatics, (p.
btt254).
S13
Jacobson, G. J. (1988). Succinct static data structures. Ph.D. thesis
Carnegie Mellon University.
Murphy, L. R., Wallqvist, A., & Levy, R. M. (2000). Simplified amino acid
alphabets for protein fold recognition and implications for folding. Protein
Engineering, 13, 149 -- 152.
Needleman, S. B., & Wunsch, C. D. (1970). A general method applicable
to the search for similarities in the amino acid sequence of two proteins.
Journal of molecular biology, 48, 443 -- 453.
Peterson, E. L., Kondev, J., Theriot, J. A., & Phillips, R. (2009). Reduced
amino acid alphabets exhibit an improved sensitivity and selectivity in
fold assignment. Bioinformatics, 25, 1356 -- 1362.
Prat, Y., Fromer, M., Linial, N., & Linial, M. (2011). Recovering key bi-
ological constituents through sparse representation of gene expression.
Bioinformatics, 27, 655 -- 661.
Tao, T. (2008). Product set estimates for non-commutative groups. Com-
binatorica, 28, 547 -- 594.
Zhao, Y., Tang, H., & Ye, Y. (2012). RAPSearch2: a fast and memory-
efficient protein similarity search tool for next-generation sequencing data.
Bioinformatics, 28, 125 -- 126.
S14
Supplemental Figures
S1 Genomic data available has grown at a faster exponential
rate than computer processing power and disk storage. These
plots represent, on a log scale, the daily growth in sequence
data from GenBank along with (a) the combined computing
power (in TeraFLOPs) of the Top 500 Supercomputer list,
and (b) the largest commercially-available hard disk drives.
S2
S3
(Related to Table 1b) Ammolite's preprocessing during the
clustering phase. Ammolite removes nodes and edges that do
not participate in simple cycles, and treats all edges as simple,
unlabeled edges. In this example, both caffeine and adenine
become a purine-like graph structure. Note that the resulting
graph has no implicit hydrogens.
(Related to Figure 3) Local fractal dimension at different
scales for the space of PDB FragBag frequency vectors. Each
data point is defined by dimension d = log(n2/n1)
log(r2/r1) , where n1, n2
are the number of similarity search hits within radius respec-
tively r1, r2, and r2−r1 is the increment size of 0.01 for cosine
distance and 1 for euclidean distance. In most regimes, local
fractal dimension is consistently low, except for a large spike
when radius expands to include the central cluster of proteins.
esFragBag achieves the most acceleration when both output
size is small and we remain in a low fractal dimension regime.
15
(a)
(b)
Figure S1: Genomic data available has grown at a faster exponential rate
than computer processing power and disk storage. These plots represent,
on a log scale, the daily growth in sequence data from GenBank along with
(a) the combined computing power (in TeraFLOPs) of the Top 500 Super-
computer list, and (b) the largest commercially-available hard disk drives.
S16
Figure S2: (Related to Table 1b) Ammolite's preprocessing during the
clustering phase. Ammolite removes nodes and edges that do not participate
in simple cycles, and treats all edges as simple, unlabeled edges.
In this
example, both caffeine and adenine become a purine-like graph structure.
Note that the resulting graph has no implicit hydrogens.
S17
N
N
C
C C
C
N
N C
(a) Cosine distance
(b) Euclidean distance
Figure S3: (Related to Figure 3) Local fractal dimension at different
scales for the space of PDB FragBag frequency vectors. Each data point is
defined by dimension d = log(n2/n1)
log(r2/r1) , where n1, n2 are the number of similar-
ity search hits within radius respectively r1, r2, and r2 − r1 is the increment
size of 0.01 for cosine distance and 1 for euclidean distance. In most regimes,
local fractal dimension is consistently low, except for a large spike when ra-
dius expands to include the central cluster of proteins. esFragBag achieves
the most acceleration when both output size is small and we remain in a
low fractal dimension regime.
|
1804.06601 | 1 | 1804 | 2018-04-18T08:30:59 | Independent Distributions on a Multi-Branching AND-OR Tree of Height 2 | [
"cs.DS"
] | We investigate an AND-OR tree T and a probability distribution d on the truth assignments to the leaves. Tarsi (1983) showed that if d is an independent and identical distribution (IID) such that probability of a leaf having value 0 is neither 0 nor 1 then, under a certain assumptions, there exists an optimal algorithm that is depth-first. We investigate the case where d is an independent distribution (ID) and probability depends on each leaf. It is known that in this general case, if height is greater than or equal to 3, Tarsi-type result does not hold. It is also known that for a complete binary tree of height 2, Tarsi-type result certainly holds. In this paper, we ask whether Tarsi-type result holds for an AND-OR tree of height 2. Here, a child node of the root is either an OR-gate or a leaf: The number of child nodes of an internal node is arbitrary, and depends on an internal node. We give an affirmative answer. Our strategy of the proof is to reduce the problem to the case of directional algorithms. We perform induction on the number of leaves, and modify Tarsi's method to suite height 2 trees. We discuss why our proof does not apply to height 3 trees. | cs.DS | cs |
Independent Distributions on a Multi-Branching
AND-OR Tree of Height 2
Mika Shigemizu1, Toshio Suzuki2 ∗and Koki Usami3
Department of Mathematical Sciences, Tokyo Metropolitan University,
Minami-Ohsawa, Hachioji, Tokyo 192-0397, Japan
1: mksgmn [email protected]
2: [email protected]
3: [email protected]
August 6, 2018
Abstract
We investigate an AND-OR tree T and a probability distribution d
on the truth assignments to the leaves. Tarsi (1983) showed that if d is
an independent and identical distribution (IID) such that probability of a
leaf having value 0 is neither 0 nor 1 then, under a certain assumptions,
there exists an optimal algorithm that is depth-first. We investigate the
case where d is an independent distribution (ID) and probability depends
on each leaf. It is known that in this general case, if height is greater than
or equal to 3, Tarsi-type result does not hold. It is also known that for
a complete binary tree of height 2, Tarsi-type result certainly holds. In
this paper, we ask whether Tarsi-type result holds for an AND-OR tree
of height 2. Here, a child node of the root is either an OR-gate or a leaf:
The number of child nodes of an internal node is arbitrary, and depends
on an internal node. We give an affirmative answer. Our strategy of the
proof is to reduce the problem to the case of directional algorithms. We
perform induction on the number of leaves, and modify Tarsi's method to
suite height 2 trees. We discuss why our proof does not apply to height 3
trees.
Keywords: Depth-first algorithm; Independent distribution; Multi-
branching tree; Computational complexity; Analysis of algorithms
MSC[2010] 68T20; 68W40
1 Introduction
Depth-first algorithm is a well-known type of tree search algorithm. Algorithm
A on a tree T is depth-first if the following holds for each internal node x of T :
∗Corresponding author. This work was partially supported by Japan Society for the Pro-
motion of Science (JSPS) KAKENHI (C) 16K05255.
1
Once A probes a leaf that is a descendant of x, A does not probe leaves that
are not descendant of x until A finds value of x.
With respect to analysis of algorithm, the concept of depth-first algorithm
has an advantage that it is well-suited for induction on subtrees. An example
of such type of induction may be found in our former paper [12].
Thus, given a problem on a tree, it is theoretically interesting question to
ask whether there exists an optimal algorithm that is depth-first. Here, cost
denotes (expected value of) the number of leaves probed during computation,
and an algorithm is optimal if it achieves the minimum cost among algorithms
considered in the question.
If an associated evaluation function of a mini-max tree is bi-valued and the
label of the root is MIN (MAX, respectively), the tree is equivalent to an AND-
OR tree (an OR-AND tree). In other words, the root is labeled by AND (OR),
and AND layers and OR layers alternate. Each leaf has a truth value 0 or 1,
where we identify 0 with false, and 1 with true.
A fundamental result on optimal algorithms on an AND-OR trees is given by
Tarsi. A tree is balanced (in the sense of Tarsi) if (1) any two internal nodes of
the same depth (distance from the root) have the same number of child nodes,
and (2) all of the leaves have the same depth. A probability distribution d
on the truth assignments to the leaves is an independent distribution (ID) if
the probability of each leaf having value 0 depends on the leaf, and values of
leaves are determined independently. If, in addition, all the probabilities of the
leaves are the same (say, p), d is an independent and identical distribution (IID).
Algorithm A is directional [3] if there is a fixed linear order of the leaves and for
any truth assignment, the order of probing by A is consistent with the order.
The result of Tarsi is as follows. Suppose that AND-OR tree T is balanced
and d is an IID such that p 6= 0, 1. Then there exists an optimal algorithm that
is depth-first and directional [13].
This was shown by an elegant induction. For an integer n such that 0 ≤
n ≤ h = (the height of the tree), an algorithm A is called n-straight if for each
node x whose distance from the leaf is at most n, once A probes a leaf that is
a descendant of x, A does not probe leaves that are not descendant of x until
A finds value of x. The proof by Tarsi is induction on n. Under an induction
hypothesis, we take an optimal algorithm A that is (n − 1)-straight but not
n-straight. Then we modify A and get two new algorithms. By the assumption,
the expected cost by A is not greater than the costs by the modified algorithms,
thus we get inequalities. By means of the inequalities, we can eliminate a non-
n-straight move from A without increasing cost.
In the above mentioned proof by Tarsi, derivation of the inequalities heavily
depends on the hypothesis that the distribution is an IID. In the same paper,
Tarsi gave an example of an ID on an AND-OR tree of the following properties:
The tree is of height 4, not balanced, and no optimal algorithm is depth-first.
Later, we gave another example of such an ID where a tree is balanced, binary
and height 3 [10].
On the other hand, as is observed in [10], in the case of a balanced binary
AND-OR tree of height 2, Tarsi-type result holds for IDs in place of IIDs.
2
height 2, binary
height 2, general
height ≥ 3
ID
Yes. S. [10]
IID
No. S. [10]
(see also Tarsi [13])
Yes. Tarsi [13]
Table 1: Existence of an optimal algorithm that is depth-first
Table 1 summarizes whether Tarsi-type result holds or not. In the table, we
assume that an AND-OR tree is balanced, and that the probability of each leaf
having value 0 is neither 0 nor 1.
In this paper, we ask whether Tarsi-type result holds for the case where a tree
is height 2 and the number of child nodes is arbitrary. We give an affirmative
answer.
We show a slightly stronger result. We are going to investigate a tree of
the following properties. The root is an AND-gate, and a child node of the
root is either an OR-gate or a leaf. The number of child nodes of an internal
node is arbitrary, and depends on an internal node. Figure 1 is an example of
such a tree. Now, suppose that an ID on the tree is given and that at each
leaf, the probability of having value 0 is neither 0 nor 1. Under these mild
assumptions, we show that there exists an optimal algorithm that is depth-first
and directional.
xÕ
x
0
x1
x2
x3
x
4
x20 x21 x30 x31 x32
x40 x41 x42
Figure 1: Example of a height 2 AND-OR tree that is not balanced
Our strategy of the proof is to reduce the problem to the case of directional
algorithms. We perform induction on the number of leaves, and modify Tarsi's
method to go along with properties particular to hight 2 trees. The first author
and the third author showed, in their talk [8], a restricted version of the present
3
result in which only directional (possibly non-depth-first) algorithms are taken
into consideration. In the talk, the core of the strategy is suggested by the first
author. The second author reduced the general case, in which non-directional
algorithms are taken into consideration, to the case of directional algorithms.
The paper [9] gives an exposition of the background. It is a short survey on
the work of Liu and Tanaka [2] and its subsequent developments [11, 12, 6]. The
paper [5] is also in this line. Some classical important results by 1980's may be
found in the papers [1, 3, 4, 13] and [7].
We introduce notation in section 2. We show our result in section 3.
In
section 4, we discuss why our proof does not apply to the case of height 3, and
discuss future directions.
2 Preliminaries
If T is a balanced tree (in the sense of Tarsi, see Introduction) and there exists a
positive integer n such that all of the internal nodes have exactly n child nodes,
we say T is a complete n-ary tree.
For algorithm A and distribution d, we denote (expected value of) the cost
by cost(A, d).
We are interested in a multi-branching AND-OR tree of height 2, where a
child node of the root is either a leaf or an OR-gate, and the number of leaves
depend on each OR-gate. For simplicity of notation, we investigate a slightly
larger class of trees.
Hereafter, by "a multi-branching AND-OR tree of height at most 2", we
denote an AND-OR tree T of the following properties.
• The root is an AND-gate.
• We allow an internal node to have only one child, provided that the tree
has at least two leaves.
• All of the child nodes of the root are OR-gates.
The concept of "multi-branching AND-OR tree of height at most 2" include
the multi-branching AND-OR trees of height 2 in the original sense (because an
OR-gate of one leaf is equivalent to a leaf), the AND-trees of height 1 (this case
is achieved when all of the OR-gates have one leaf) and the OR-trees of height
1 (this case is achieved when the root has one child). The simplest case is a tree
of just two leaves, and this case is achieved exactly in either of the following
two. (1) The tree is equivalent to a binary AND-tree of height 1: (2) The tree
is equivalent to a binary OR-tree of height 1.
Suppose that T is a multi-branching AND-OR tree of height at most 2.
• We let xλ denote the root.
• By r we denote the number of child nodes of the root. x0, . . . , xr−1 are
the child nodes of the root.
4
• For each i (0 ≤ i < r), we let a(i) denote the number of child leaves of xi.
xi,0, . . . , xi,a(i)−1 are the child leaves of xi.
Figure 2 is an example of such a tree, where r = 5, a(0) = 1 and a(4) = 3.
The tree is equivalent to the tree in Figure 1.
xÕ
x
4
x40 x41 x42
x
0
x00
Figure 2: Example of an AND-OR tree of height at most 2
Suppose that d is an ID on T . For each i (0 ≤ i < r) and j (0 ≤ j < a(i)),
we use the following symbols.
• p(i, j) is the probability of xi,j having value 0.
• p(i) is the probability of xi having value 0.
• Since d is an ID, its restriction to the child nodes of xi is an ID on the
subtree whose root is xi. Here, we denote it by the same symbol d. Then
we define c(i) as follows.
c(i) = min
A
cost(A, d),
where A runs over algorithms finding value of xi, and cost(A, d) is expected
cost.
Thus, p(i) is the product p(i, 0) · · · p(i, a(i) − 1). If a(i) = 1 then c(i) = 1.
If a(i) ≥ 2 and we have p(i, 0) ≤ · · · ≤ p(i, a(i) − 1), then c(i) = 1 + p(i, 0) +
p(i, 0)p(i, 1) + · · · + p(i, 0) · · · p(i, a(i) − 1).
Tarsi [13] investigated a depth-first algorithm that probes the leaves from
left to right, skipping a leaf whenever there is sufficient information to evaluate
one of its ancestors, and he called it SOLVE. We investigate a similar algorithm
depending on a given independent distribution.
5
Definition 1. Suppose that T is a multi-branching AND-OR tree of height at
most 2.
1. Suppose that d is an ID on T and for each i (0 ≤ i < r) and j (0 ≤ j <
a(i)), we have p(i, j) 6= 0, 1.
By SOLVEd, we denote the unique depth-first directional algorithm such
that the following hold for all i, s, j, k (0 ≤ i < r, 0 ≤ s < r, 0 ≤ j < a(i),
0 ≤ k < a(i)).
(a) If c(i)/p(i) < c(s)/p(s) then priority of (probing the leaves under) xi
is higher than that of xs.
(b) If c(i)/p(i) = c(s)/p(s) and i < s then priority of xi is higher than
that of xs.
(c) If p(i, j) < p(i, k) then priority of (probing) xi,j is higher than xi,k.
(d) If p(i, j) = p(i, k) and j < k then priority of xi,j is higher than xi,k.
2. Suppose that we remove some nodes (except for the root of T ) from T ,
and if a removed node has descendants, we remove them too. Let T ′ be
the resulting tree. Suppose that δ is an ID on T ′ and for each i (0 ≤ i < r)
and j (0 ≤ j < a(i)) such that x(i, j) is a leaf of T ′, we have p(i, j) 6= 0, 1.
By SOLVET
δ , we denote the unique depth-first directional algorithm of the
following properties. For all i, s, j, k (0 ≤ i < r, 0 ≤ s < r, 0 ≤ j < a(i),
0 ≤ k < a(i)), if xi and xs are nodes of T ′, the above-mentioned assertions
(a) and (b) hold; and if xi,j and xi,k are leaves of T ′, the above-mentioned
assertions (c) and (d) hold.
Lemma 1. Suppose that T is a multi-branching AND-OR tree of height at most
2. Suppose that d is an ID on T and for each i (0 ≤ i < r) and j (0 ≤ j < a(i)),
we have p(i, j) 6= 0, 1. Then SOLVEd achieves the minimum cost among depth-
first directional algorithms. To be more precise, if A is a depth-first directional
algorithm then cost(SOLVEd, d) ≤ cost(A, d).
Proof. It is straightforward.
3 Result
Theorem 2. Suppose that T is a multi-branching AND-OR tree of height at
most 2. Suppose that d is an ID on T and for each i (0 ≤ i < r) and j
(0 ≤ j < a(i)), we have p(i, j) 6= 0, 1. Then SOLVEd achieves the minimum
cost among all of the algorithms (depth-first or non-depth-first, directional or
non-directional). Therefore, there exists a depth-first directional algorithm that
is optimal among all of the algorithms.
Proof. We perform induction on the number of leaves. The base cases are the
binary AND-trees of height 1 and the binary OR-trees of height 1. In general, if
T is equivalent to a tree of height 1, the assertion of the theorem clearly holds.
6
To investigate induction step, we assume that T has at least three leaves.
Our induction hypothesis is that for any multi-branching AND-OR tree T ′ of
height at most 2, if the number of leaves of T ′ is less than that of T then the
assertion of theorem holds for T ′.
We fix an algorithm A that minimizes cost(A, d) among all of the algorithms
(depth-first or non-depth-first, directional or non-directional).
Case 1: At the first move of A, A makes a query to a leaf xi,0 such that
a(i) = 1. In this case, if A finds that xi,0 has value 0 then A returns 0 and
finish. Otherwise, A calls a certain algorithm (say, A′) on T − xi, that is, the
tree given by removing xi and xi,0 from T . The probability distribution given
by restricting d to T − xi is an ID, and a probability of any leaf is neither 0 nor
1.
Therefore, by induction hypothesis, without loss of generality, we may as-
sume that A′ is a depth-first directional algorithm on T − xi. Therefore, A is
a depth-first directional algorithm. Hence, by Lemma 1, the same cost as A is
achieved by SOLVEd.
Case 2: Otherwise. At the first move of A, A makes a query to a leaf xi,j
such that a(i) ≥ 2.
Let T0 := T − xi,j, the tree given by removing xi,j from T . In addition, let
T1 := T − xi, the tree given by removing xi and all of the leaves under xi from
T . Here, T0 and T1 inherit all of the indices (for example, "3, 1" of x3,1) from
T .
If T1 is empty then T is equivalent to a tree of height 1, and this case reduces
to our observation in the base case. Thus, throughout rest of Case 2, we assume
that T1 is non-empty.
If A finds that xi,j has value 0 then A calls a certain algorithm (say, A0) on
T0.
T1.
If A finds that xi,j has value 1 then A calls a certain algorithm (say, A1) on
Case 1, without loss of generality, we may assume that As is SOLVET
For each s = 0, 1, let d[s] be the restriction of d to Ts. In the same way as
d[s] on Ts.
Hence, there is a permutation X = hxi,s(0), . . . , xi,s(a(i)−2)i of the leaves
under xi except xi,j , and there are possibly empty sequences of leaves, Y =
hy0, . . . , yk−1i and Z = hz0, . . . , zm−1i, with the following properties.
• The three sets {xi}, Y ∗ = {y : y is a parent of a leaf in Y }, and Z ∗ = {z : z
is a parent of a leaf in Z} are mutually disjoint, and their union equals
{x0, . . . , xr−1}, the set of all child nodes of xλ.
• The search priority of A0 is in accordance with Y XZ (thus, y0 is the first).
• The search priority of A1 is in accordance with Y Z.
Case 2.1: Z is non-empty.
We are going to show that Y is empty. Assume not. Let B be the depth-first
directional algorithm on T whose search priority is Y xi,j XZ.
7
Let AY (AX , AZ, respectively) denote the depth-first directional algorithm
on the subtree above by Y (X, Z), where search priority is in accordance with
Y (X, Z). Thus, we may write A0 as "AY ; AX ; AZ ", A1 as "AY ; AZ", and B
as "AY ; Probe xi,j ; AX ; AZ".
We look at the following events. Recall that Y ∗ is the set of all parent nodes
of leaves in Y .
EY : "At least one element of Y ∗ has value 0."
EX : "All of the elements of X have value 0."
Since the tree is height 2 and Z is non-empty, in each of A0, A1 and B, the
following holds: "AY finds value of xλ if and only if EY happens."
In the same way, in A0 and in B, under assumption that AX is called, AX
finds value of the root if and only if EX happens.
Thus, flowcharts of A and B as Boolean decision trees are as described in
Figure 3 and Figure 4, respectively.
0
x ij
1
AY
EY
end
AY
EY
end
AZ
AX
EX
end
AZ
A0
A1
Figure 3: Flowchart of A (Case 2.1, in the presence of Y )
AY
EY
end
0
x ij
1
AZ
AX
EX
end
AZ
Figure 4: Flowchart of B (in the presence of Y )
Therefore, letting pY = prob[¬EY ] (that is, probability of the negation) and
pX = prob[¬EX ], the cost of A and B are as follows. In the following formulas,
CY denotes cost(AY , dY ), where dY denotes the probability distribution given
by restricting d to Y . CX and CZ are similarly defined.
8
cost(A, d)
= 1 + p(i, j){CY + pY (CX + pXCZ )} + (1 − p(i, j))(CY + pY CZ)
= 1 + CY + p(i, j)pY (CX + pX CZ ) + (1 − p(i, j))pY CZ
(1)
cost(B, d)
= CY + pY [1 + {p(i, j)(CX + pX CZ) + (1 − p(i, j))CZ }]
= CY + pY + pY p(i, j)(CX + pX CZ ) + pY (1 − p(i, j))CZ
(2)
Therefore, cost(A, d) − cost(B, d) = 1 − pY . However, by our assumption
on d that probability of each leaf (having value 0) is neither 0 nor 1, EY has
positive probability, thus pY < 1. Thus cost(A, d) − cost(B, d) is positive, and
this contradicts to the assumption that A achieves the minimum cost.
Hence, we have shown that Y is empty. Therefore, A is the following algo-
rithm (Figure 5): "Probe xi,j . If xi,j = 0 then perform depth-first directional
search on T0 = T − xi,j , where search priority is in accordance with XZ. Oth-
erwise (that is, xi,j = 1), perform depth-first directional search on T1 = T − xi,
where search priority is in accordance with Z."
0
x ij
1
AX
EX
end
AZ
AZ
A0
A1
Figure 5: Flowchart of A (in the absence of Y )
Thus, A is a depth-first directional algorithm. Hence, by Lemma 1, the same
cost as A is achieved by SOLVEd.
Case 2.2: Otherwise. In this case, Z is empty. The proof is similar to Case
2.1.
4 Concluding remarks
4.1 Difference between height 2 case and heihgt 3 case
As is mentioned in Introduction, the counterpart to Theorem 2 does not hold
for the case of height 3. We are going to discuss why the proof of Theorem 2
9
does not work for the case of height 3.
Figure 6 is a complete binary OR-AND tree of height 3.
x0
x
000
x
001
X
xÕ
x1
x10
x
100
x
101
Y
Figure 6: A binary OR-AND tree of height 3
Suppose that algorithm A is as follows. Let Y = hx100, x101i, X = hx001i,
and Z = hx010, x011, x110, x111i. At the first move, A probes x000. If x000 = 0
then A probes in accordance with order Y XZ. If x000 = 1 then A probes in
accordance with order Y Z.
Let A′
than x101.
Y be the algorithm on Y such that x100 has higher priority of probing
Suppose that an ID d on the tree is given, and that, at each leaf, probability
of having value 0 is neither 0 nor 1. We investigate the following event.
Y : "x10 has value 0."
E ′
On the one hand, by our assumption on ID d, E ′
On the other hand, whether E ′
In other words, probability of "A′
equivalent to the assertion "A′
observation in Case 2.1 of Theorem 2 does not work for the present setting.
Y has positive probability.
Y happens or not, AY does not find value of xλ.
Y is not
Y finds value of xλ". Hence, counterpart to our
Y finds value of xλ" is 0. Therefore, E ′
4.2 Summary and future directions
Given a tree T , let IID+
T , respectively) denote the set of all IIDs on T (IDs
on T ) such that, at each leaf, probability having value 0 is neither 0 nor 1. Now
we know the following.
T (ID+
1. (Tarsi [13]) Suppose that T is a balanced AND-OR tree of any height, and
T . Then there exists an optimal algorithm that is depth-first
that d ∈ IID+
and directional.
2. (S. [10]) Suppose that T is a complete binary OR-AND tree of height 3.
T such that no optimal algorithm is depth-first.
Then there exists d ∈ ID+
10
3. (Theorem 2) Suppose that T is an AND-OR tree of height 2, and that
T . Then there exists an optimal algorithm that is depth-first and
d ∈ ID+
directional.
Suppose that T is a complete binary AND-OR tree of height h ≥ 3. There
is yet some hope to find a subset D of ID+
T of the following properties.
• IID+
T ( D ( ID+
T
• For each d ∈ D, there exists an optimal algorithm that is depth-first and
directional.
References
[1] Donald E. Knuth and Ronald W. Moore. An analysis of alpha-beta pruning.
Artif. Intell., 6 (1975) 293 -- 326.
[2] ChenGuang Liu and Kazuyuki Tanaka. Eigen-distribution on random as-
signments for game trees. Inform. Process. Lett., 104 (2007) 73 -- 77.
[3] Judea Pearl. Asymptotic properties of minimax trees and game-searching
procedures. Artif. Intell., 14 (1980) 113 -- 138.
[4] Judea Pearl. The solution for the branching factor of the alpha-beta pruning
algorithm and its optimality. Communications of the ACM, 25 (1982) 559 --
564.
[5] Weiguang Peng, NingNing Peng, KengMeng Ng, Kazuyuki Tanaka and
Yue Yang. Optimal depth-first algorithms and equilibria of independent
distributions on multi-branching trees. Inform. Process. Lett., 125 (2017)
41 -- 45.
[6] Weiguang Peng, Shohei Okisaka, Wenjuan Li and Kazuyuki Tanaka. The
Uniqueness of eigen-distribution under non-directional algorithms. IAENG
International Journal of Computer Science, 43 (2016) 318 -- 325.
http://www.iaeng.org/IJCS/issues_v43/issue_3/IJCS_43_3_07.pdf
[7] Michael Saks and Avi Wigderson. Probabilistic Boolean decision trees and
the complexity of evaluating game trees. In: Proc. 27th IEEE FOCS, 1986,
29 -- 38.
[8] Mika Shigemizu and Koki Usami. Search algorithms on a multi-branching
AND-OR tree of height 2. (in Japanese) In:
S¯urikaisekikenky¯usho
K¯oky¯uroku, Research Institute for Mathematical Sciences, Kyoto Univer-
sity, to appear.
[9] Toshio Suzuki. Kazuyuki Tanaka's work on AND-OR trees and subsequent
developments. Annals of the Japan Association for Philosophy of Science,
25 (2017) 79 -- 88.
11
[10] Toshio Suzuki. Non-Depth-First Search against Independent Distributions
on an AND-OR Tree. preprint, arXiv:1709.07358v1[cs.DS] (2017).
[11] Toshio Suzuki and Ryota Nakamura. The eigen distribution of an AND-OR
tree under directional algorithms. IAENG Int. J. Appl. Math., 42 (2012)
122 -- 128.
http://www.iaeng.org/IJAM/issues_v42/issue_2/IJAM_42_2_07.pdf
[12] Toshio Suzuki and Yoshinao Niida. Equilibrium points of an AND-OR tree:
under constraints on probability. Ann. Pure Appl. Logic , 166 (2015) 1150 --
1164.
[13] Michael Tarsi. Optimal search on some game trees. J. ACM, 30 (1983)
389 -- 396.
12
|
1910.06185 | 1 | 1910 | 2019-10-14T14:51:57 | An Improved FPT Algorithm for the Flip Distance Problem | [
"cs.DS"
] | Given a set $\cal P$ of points in the Euclidean plane and two triangulations of $\cal P$, the flip distance between these two triangulations is the minimum number of flips required to transform one triangulation into the other. Parameterized Flip Distance problem is to decide if the flip distance between two given triangulations is equal to a given integer $k$. The previous best FPT algorithm runs in time $O^{*}(k\cdot c^{k})$ ($c\leq 2\times 14^{11}$), where each step has fourteen possible choices, and the length of the action sequence is bounded by $11k$. By applying the backtracking strategy and analyzing the underlying property of the flip sequence, each step of our algorithm has only five possible choices. Based on an auxiliary graph $G$, we prove that the length of the action sequence for our algorithm is bounded by $2|G|$. As a result, we present an FPT algorithm running in time $O^{*}(k\cdot 32^{k})$. | cs.DS | cs | An Improved FPT Algorithm for the Flip Distance Problem∗†
Qilong Feng‡
Shaohua Li§
Xiangzhong Meng¶
Jianxin Wang(cid:107)
9
1
0
2
t
c
O
4
1
]
S
D
.
s
c
[
1
v
5
8
1
6
0
.
0
1
9
1
:
v
i
X
r
a
Abstract
Given a set P of points in the Euclidean plane and two triangulations of P, the flip distance between
these two triangulations is the minimum number of flips required to transform one triangulation into
the other. Parameterized Flip Distance problem is to decide if the flip distance between two given
triangulations is equal to a given integer k. The previous best FPT algorithm runs in time O∗(k · ck)
(c ≤ 2 × 1411), where each step has fourteen possible choices, and the length of the action sequence is
bounded by 11k. By applying the backtracking strategy and analyzing the underlying property of the
flip sequence, each step of our algorithm has only five possible choices. Based on an auxiliary graph G,
we prove that the length of the action sequence for our algorithm is bounded by 2G. As a result, we
present an FPT algorithm running in time O∗(k · 32k).
Introduction
1
Given a set P of n points in the Euclidean plane, a triangulation of P is a maximal planar subdivision whose
vertex set is P [6]. A flip operation to one diagonal e of a convex quadrilateral in a triangulation is to remove
e and insert the other diagonal into this quadrilateral. Note that if the quadrilateral associated with e is
not convex, the flip operation is not allowed. The flip distance between two triangulations is the minimum
number of flips required to transform one triangulation into the other.
Triangulations play an important role in computational geometry, which are applied in areas such as
computer-aided geometric design and numerical analysis [7, 8, 16].
Given a point set P in the Euclidean plane, we can construct a graph GT (P) in which every triangulation
of P is represented by a vertex, and two vertices are adjacent if their corresponding triangulations can be
transformed into each other through one flip operation. GT (P) is called the triangulations graph of P.
[1] showed that the
Properties of the triangulations graph are studied in the literature. Aichholzer et al.
lower bound of the number of vertices of GT (P) is Ω(2.33n). Lawson and Charles [12] showed that the
diameter of GT (P) is O(n2). Hurtado et al. [9] proved that the bound is tight. Since GT (P) is connected
[12], any two triangulations of P can be transformed into each other through a certain number of flips.
Flip Distance problem consists in computing the flip distance between two triangulations of P, which
was proved to be NP-complete by Lubiw and Pathak [13]. Pilz showed that the Flip Distance is APX-hard
[2] proved that Parameterized Flip Distance is NP-complete on triangulations
[15]. Aichholzer et al.
of simple polygons. However, the complexity of Flip Distance on triangulations of convex polygons has
been open for many years, which is equivalent to the problem of computing the rotation distance between
two rooted binary trees [17].
dations of Computer Science, MFCS 2017.
∗A preliminary version of this paper appeared in the proceedings of 42nd International Symposium on Mathematical Foun-
†This work is supported by the National Natural Science Foundation of China under Grants (61672536, 61420106009,
61872450, 61828205,), Hunan Provincial Science and Technology Program (2018WK4001) and the European Research Council
(ERC) under the European Unions Horizon 2020 research and the innovation programme (grant Nr: 714704).
‡School of Computer Science and Engineering, Central South University, Changsha, Hunan, P.R.China.
§Institute of Informatics, University of Warsaw, Poland.
¶School of Computer Science and Engineering, Central South University, Changsha, Hunan, P.R.China.
(cid:107)School
of Computer Science
and Engineering, Central South University, Changsha, Hunan, P.R.China,
[email protected]
1
Parameterized Flip Distance problem is: given two triangulations of a set of points in the plane and
an integer k, deciding if the flip distance between these two triangulations is equal to k. For Parameterized
Flip Distance on triangulations of a convex polygon, Lucas [14] gave a kernel of size 2k and an O∗(kk)-time
algorithm. Kanj and Xia [11] studied Parameterized Flip Distance on triangulations of a set of points
in the plane, and presented an O∗(k · ck)-time algorithm (c ≤ 2 · 1411), which applies to triangulations of
general polygonal regions (even with holes or points inside it).
In this paper, we exploit Parameterized Flip Distance further. At first, we give a nondeterministic
construction process to illustrate our idea. The nondeterministic construction process contains only two types
of actions, which are the moving action as well as the flipping and backing action. Given two triangulations
and a parameter k, we prove that either there exists a sequence of actions of length at most 2k, following
which we can transform one triangulation into the other, or we can conclude that no valid sequence of length
k exists. Thus we get an improved O∗(k · 32k)-time FPT algorithm, which also applies to triangulations of
general polygonal regions (even with holes or points inside it).
2 Preliminaries
In a triangulation T , a flip operation f to an edge e that is the diagonal of a convex quadrilateral Q is
to delete e and insert the other diagonal e(cid:48) into Q. We define e as the underlying edge of f , denoted by
ε(f ), and e(cid:48) as the resulting edge of f , denoted by ϕ(f ). (For consistency and clarity, we continue to use
some symbols and definitions from [11]). Note that if e is not a diagonal of any convex quadrilateral in the
triangulation, flipping e is not allowed. Suppose that we perform a flip operation f on a triangulation T1
and get a new triangulation T2. We say f transforms T1 into T2. T1 is called an underlying triangulation
of f , and T2 is called a resulting triangulation of f . Given a set P of n points in the Euclidean plane,
let Tstart and Tend be two triangulations of P, in which Tstart is the initial triangulation and Tend is the
objective triangulation. Let F = (cid:104)f1, f2, ..., fr(cid:105) be a sequence of flips, and (cid:104)T0, T1, ..., Tr(cid:105) be a sequence of
triangulations of P in which T0 = Tstart and Tr = Tend. If Ti−1 is an underlying triangulation of fi, and Ti
is a resulting triangulation of fi for each i = 1, 2, ..., r, we say F transforms Tstart into Tend, or F is a valid
F−→ Tend. The flip distance between Tstart and Tend is the length of a shortest
sequence, denoted by Tstart
valid flip sequence.
Now we give the formal definition of Parameterized Flip Distance problem.
Parameterized Flip Distance
Input: Two triangulations Tstart and Tend of P and an integer k.
Question: Decide if the flip distance between Tstart and Tend is equal to k.
The triangulation on which we are performing a flip operation is called the current triangulation. An
edge e which belongs to the current triangulation but does not belong to Tend is called a necessary edge in
the current triangulation. It is easy to see that for any necessary edge e, there must exist a flip operation f
in a valid sequence such that e = ε(f ). Otherwise, we cannot get the objective triangulation Tend.
For a directed graph D, a maximal connected component of its underlying graph is called a weakly
connected component of D. We define the size of an undirected tree as the number of its vertices. A node in
D is called a source node if the indegree of this node is 0.
A parameterized problem is a decision problem for which every instance is of the form (x, k), where x is
the input instance and k ∈ N is the parameter. A parameterized problem is fixed-parameter tractable (FPT )
if it can be solved by an algorithm (FPT algorithm) in O(f (k)xO(1)) time, where f (k) is a computable
function of k. For a further introduction to parameterized algorithms, readers could refer to [3, 5].
3 The Improved Algorithm for Parameterized Flip Distance
Given Tstart and Tend, let F = (cid:104)f1, f2, ..., fr(cid:105) be a valid sequence, that is, Tstart
the adjacency of two flips in F .
F−→ Tend. Definition 1 defines
2
Definition 1. [11] Let fi and fj be two flips in F (1 ≤ i < j ≤ r). We define that flip fj is adjacent to flip
fi, denoted by fi → fj, if the following two conditions are satisfied:
(1) either ϕ(fi) = ε(fj), or ϕ(fi) and ε(fj) share a triangle in triangulation Tj−1;
(2) ϕ(fi) is not flipped between fi and fj, that is, there does not exist a flip fp in F , where
i < p < j, such that ϕ(fi) = ε(fp).
By Definition 1, we can construct a directed acyclic graph (DAG), denoted by DF . Every node in DF
represents a flip operation of F , and there is an arc from fi to fj if fj is adjacent to fi. For convenience, we
label the nodes in DF using labels of the corresponding flip operations. In other words, we can see a node
in DF as a flip operation and vice versa.
The intuition of Definition 1 is that if there is an arc from fi to fj, then fj cannot be flipped before
fi because the quadrilateral corresponding to the flip fj is formed after fi or the underlying edge of fj,
namely ε(fj) is the resulting edge of fi, namely ϕ(fi). The following lemma gives a stronger statement: any
topological sorting of DF is a valid sequence.
Lemma 1. [11] Let T0 and Tr be two triangulations and F = (cid:104)f1, f2, ..., fr(cid:105) be a sequence of flips such that
F−→ Tr. Let π(F ) be a permutation of the flips in F such that π(F ) is a topological sorting of DF . Then
T0
π(F ) is a valid sequence of flips such that T0
π(F )−−−→ Tr.
Lemma 1 ensures that if we repeatedly remove a source node from DF and flip the underlying edge of
this node until DF becomes empty, we can get a valid sequence and the objective triangulation Tend. On
the basis of Lemma 1, the essential task of our algorithm is to find an edge which is the underlying edge of
a source node. Thus we introduce the definition of a walk, which describes the "track" to find such an edge.
Definition 2. [10] A walk in a triangulation T (starting from an edge e ∈ T ) is a sequence of edges of T
beginning with e in which any two consecutive edges share a triangle in T .
According to Lemma 1, if there is a valid sequence F for the input instance, any topological sorting of
DF is also a valid sequence for the given instance. The difficulty is that F is unknown. In order to find
the topological sorting of DF , the algorithm of Kanj and Xia [11] takes a nondeterministic walk to find an
edge e which is the underlying edge of a source node, flips this edge (removing the corresponding node from
DF ), nondeterministically walks to an edge which shares a triangle with e and recursively searches for an
edge corresponding to a source node. Their algorithm deals with weakly connected components of DF one
after another (refer to Corollary 4 in [11]), that is, the algorithm tries to find a solution F in which all
flips belonging to the same weakly connected component of DF appear consecutively. In order to keep this
procedure within the current weakly connected component, the algorithm uses a stack to preserve the nodes
(defined as connecting point in [11]) whose removal separates the current weakly connected component into
small weakly connected components. When removing all nodes of a small component, their algorithm jumps
to the connecting point at the top of the stack in order to find another small component.
We observe that it is not necessary to remove all nodes of a weakly connected component before dealing
with other weakly connected components, that is, our algorithm may find a solution F in which the nodes
belonging to the same weakly connected components appear dispersedly. Thus our algorithm leaves out
the stack which is used to preserve connecting points. We show that it suffices to use two types of actions
(Section 3.2) instead of the five types in [11]. Moreover, every time our algorithm finds a source node,
it removes the node, flips the underlying edge and backtracks to the previous edge in the walk instead of
searching for the next node, thus reducing the number of choices for the actions. In a word, our algorithm
traverses DF in a reverse way. However, two adjacent edges in a walk may not correspond to two adjacent
vertices in DF . In order to emphasize this fact and make it more convenient for the proof, we construct
an auxiliary graph G and prove that G is a forest. Actually G can be seen as the track of the reverse
traversal of DF . However, G is not the underlying graph of DF (Fig. 3). Since there is a bijection between
nondeterministic actions and nodes as well as edges of G, we prove that there exists a sequence of actions of
length at most 2DF, which is smaller than 11DF in [11]. In addition, we make some optimization on the
strategy of finding the objective sequence. As a result, we improve the running time of the algorithm from
O∗(k · ck) where c ≤ 2 · 1411 [11] to O∗(k · 32k).
3
Tstart
f1
f6
f2
f7
f3
f8
f5
f4
f9
Tend
Figure 1: An example of Flip Distance. F = (cid:104)f1, ..., f9(cid:105) is a shortest valid sequence for Tstart and Tend.
The red edge is the current edge to be flipped.
3.1 Nondeterministic construction process
Now we give a description of our nondeterministic construction process NDTRV (see Fig. 2). The construc-
tion is nondeterministic, that is, we suppose it always guesses the optimal choice correctly when running. The
actual deterministic algorithm enumerates all possible choices to simulate the nondeterministic actions (see
Fig. 4). Readers could refer to [4] as an example of nondeterministic algorithm. We present this construction
process in order to depict the idea behind our deterministic algorithm clearly and vividly.
Let Tstart be the initial triangulation, and Tend be the objective triangulation. Suppose that F is a
shortest valid sequence, that is, F has the shortest length among all valid sequences. Let DF be the DAG
constructed after F according to Definition 1. NDTRV traverses DF reversely, removes the vertices of
DF in a topologically-sorted order and transforms Tstart into Tend. Although DF is unknown, for further
analysis, we assume that NDTRV can remove and copy nodes in DF so that it can construct an auxiliary
undirected graph G and a list L during the traversal. In later analysis we show that G is a forest. Moreover,
there is a bijection between flipping actions of NDTRV and nodes of G while there is a bijection between
moving actions of NDTRV and edges of G. Obviously G and L are unknown as well. We just show that
if a shortest valid sequence F exists, then DF exists. So do G and L. We can see DF and G as conceptual
or dummy graphs. We construct G instead of analysing a subgraph of DF because one moving action (see
Section 3.2) of NDTRV may correspond to one or more edges in DF (see Fig. 3), while there is a one-to-one
correspondence between moving actions and edges in G .
At the beginning of an iteration, NDTRV picks a necessary edge e = ε(fh) arbitrarily and nondeter-
ministically guesses a walk W to find the underlying edge of a source node fs. Lemma 2 shows that there
exists such a walk W whose length is bounded by the length of a directed path B from fs to fh, and every
edge e(cid:48) in W is the underlying edge of some flip f(cid:48) on B. NDTRV uses L to preserve a sequence of nodes
Γ = (cid:104)fs = v1, ..., fh = v(cid:96)(cid:105) on B, whose underlying edges are in W . Simultaneously NDTRV constructs
a path S by copying all nodes in Γ as well as adding an undirected edge between the copy of vi and vi+1
for i = 1, ..., (cid:96). S is defined as a searching path. The node fh is called a starting node. If a starting node
is precisely a source node in DF , the searching path consists only of the copy of this starting node. When
finding ε(fs), NDTRV removes fs from DF , flips ε(fs) and moves back(backtracks) to the previous edge
ε(v2) of ε(fs) in W . If v2 becomes a source node of DF , NDTRV removes v2 from DF , flips ε(v2) and
moves back to the previous edge ε(v3). NDTRV repeats the above operations until finding a node vi in Γ
which is not a source node in DF . Then NDTRV uses vi as a new starting node, and recursively guesses
a walk nondeterministically from ε(vi) to find another edge which is the underlying edge of a source node
as above. NDTRV performs these operations until the initial starting node fh becomes a source node in
DF . Finally NDTRV removes fh and flips ε(fh), terminating this iteration. If Tcurrent is not equal to Tend,
4
NDTRV picks a new necessary edge and starts a new iteration as above until Tstart is transformed into
Tend. We give the formal presentation of NDTRV in Fig. 2 and an example in Fig. 1 and Fig. 3.
3.2 Actions of the construction
Our construction process contains two types of actions operating on triangulations. The edge which the
algorithm is operating on is called the current edge. The current triangulation is denoted by Tcurrent.
(i) Move to one edge that shares a triangle with the current edge in Tcurrent. We formalize it as
(move, e1 (cid:55)→ e2), where e1 is the current edge and e2 shares a triangle with e1.
(ii) Flip the current edge and move back to the previous edge of the current edge in W . We
formalize it as (f, e4 (cid:55)→ e3), where f is the flip performed on the current edge, e4 equals ϕ(f )
and e3 is the previous edge of ε(f ) in the current walk W .
Since there are four edges that share a triangle with the current edge, there are at most four directions
for an action of type (i). However, there is only one choice for an action of type (ii).
3.3 The sequence of actions
The following theorem is the main theorem for the deterministic algorithm FLIPDT, which bounds the
length of the sequence of actions by 2V (DF ).
Theorem 1. There exists a sequence of actions of length at most 2V (DF ) following which we can perform
a sequence of flips F (cid:48) of length V (DF ), starting from a necessary edge in Tstart, such that F (cid:48) is a topological
sorting of DF .
In order to prove theorem 1, we need to introduce some lemmas. We will give the proof for Theorem 1
at the end of Section 3.3.
Lemma 2 is one of the main structural results of Kanj and Xia [10]. It is also the structural basis of our
algorithm.
Lemma 2. [10] Suppose that a sequence of flips F − is performed such that every time we flip an edge, we
delete the corresponding source node in the DAG resulting from preceding deleting operations. Let fh be a
node in the remaining DAG such that ε(fh) is an edge in the triangulation T resulting from performing the
sequence of flips F −. There is a source node fs in the remaining DAG satisfying:
(1) There is a walk W in T from ε(fh) to ε(fs).
(2) There is a directed path B from fs to fh in the remaining DAG that we refer to as the
backbone of the DAG.
(3) The length of W is at most that of B.
(4) Any edge in W is the underlying edge of a flip in B, that is, W = (cid:104)ε(v1), ..., ε(v(cid:96))(cid:105), where
v1 = fs, ..., v(cid:96) = fh are nodes in B and there is a directed path Bi from vi to vi+1 for
i = 1, ..., (cid:96) − 1 such that Bi ⊂ B.
Lemma 3. NDTRV transforms Tstart into Tend with the minimum number of flips and stops in polynomial
time if it correctly guesses every moving and flipping action.
Proof. Suppose that F is a shortest valid sequence. According to Lemma 2, every edge flipped in NDTRV
is the underlying edge of a source node in the remaining graph of DF , and every node removed from the
remaining graph of DF in NDTRV is a source node. If Tcurrent is equal to Tend but DF is not empty, then
there exists a valid sequence F (cid:48) which is shorter than F , contradicting that F is a shortest valid sequence.
Thus NDTRV traverses DF , removes all nodes of DF in a topologically-sorted order and transforms Tstart
into Tend with the minimum number of flips by Lemma 1. Since the diameter of a transformations graph
GT (P) is O(n2) [12], NDTRV stops in polynomial time.
5
NDTRV(Tstart, Tend; DF )
Input: the initial triangulation Tstart and objective triangulation Tend.
/*Assume that F is a shortest sequence, DF is the corresponding DAG by Definition 1*/
/*G is an auxiliary undirected graph */
/*L is a list keeping track of searching paths for backtracking */
/*Q is a list preserving the sequence of nondeterministic actions */
/* Tcurrent is the current triangulation*/
a. Let V (G) and E(G) be empty sets, L and Q be empty lists;
b. Tcurrent = Tstart;
c. While Tcurrent (cid:54)= Tend do
c.1. Pick a necessary edge e = ε(fh) in Tcurrent arbitrarily;
c.2. Add a copy of fh to G;
c.3. Append fh to L;
c.4. TrackTree(Tcurrent, e, DF , G, L, Q);
TrackTree(Tcurrent, ε(fh), DF , G, L, Q)
1. Nondeterministically guess a walk in Tcurrent from ε(fh) to find ε(fs) according to
Lemma 2, let Γ = (cid:104)fs = v1, ..., fh = v(cid:96)(cid:105), where fs is a source node in DF , be a sequence
of nodes on the backbone B whose underlying edges are in the walk W such that ε(vi)
and ε(vi+1) are consecutive in W for i = 1, ..., (cid:96) − 1;
/*record actions*/
2. Add a copy of v1,...,v(cid:96)−1 to G respectively;
3. Connect the copies of v1,...,v(cid:96) in G into a path;
4. Append v(cid:96)−1,...,v1 to L;
5. Append (move, ε(v(cid:96)) (cid:55)→ ε(v(cid:96)−1)),...,(move, ε(v2) (cid:55)→ ε(v1)) to Q;
6. Remove fs = v1 from L;
7. Remove fs = v1 from DF ;
8. Flip ε(fs) in Tcurrent and move back to ε(v2);
9. Append (fs, ϕ(v1) (cid:55)→ ε(v2)) to Q;
10. For i = 2 to (cid:96) do
10.1 Nondeterministically guess if vi is a source node in DF ;
10.2
10.2.1 Remove vi from L;
10.2.2 Remove vi from DF ;
10.2.3
10.2.4 Append (vi, ϕ(vi) (cid:55)→ ε(vi+1)) to Q;
10.3 Else
10.3.1 TrackTree(Tcurrent, ε(vi), DF , G, L, Q);
If vi is a source node of DF then /*flip and move back*/
Flip ε(vi) in Tcurrent and move back to ε(vi+1);
/*record actions*/
Figure 2: Nondeterministic construction NDTRV
6
v1
v2
v9
v7
v8
v4
v3
v5
v6
v1
v2
v4
v5
v3
v7
v8
v6
v9
Figure 3: Constructing G according to the example in Fig. 1. The graph on the left is DF after F in Fig. 1.
The graph on the right is the auxiliary graph G constructed by NDTRV. ε(v6) is the necessary edge chosen
at the beginning of the first iteration and ε(v9) is the one chosen at the beginning of the second iteration.
Lemma 4. The auxiliary graph G constructed during NDTRV is a forest and V (G) = V (DF ). Morover,
G consists of a set of vertex-disjoint trees called track trees. Each track tree is created during an iteration
of NDTRV.
Proof. Since NDTRV makes a topological sorting during execution and the copy of a vertex of DF is added
to G only when it is removed from DF , it follows that V (G) = V (DF ). Suppose that there is a cycle in G.
According to lemma 2 and NDTRV, we can find a directed cycle in DF , contradicting that DF is a directed
acyclic graph. Thus G is a forest. From the execution of NDTRV, we get that it creates a connected
subgraph in G during every iteration and the subgraphs created during each iteration are vertex-disjoint.
Thus the subgraph created during each iteration is a tree. This concludes the proof.
We give the proof of Theorem 1 below.
Proof. (Theorem 1) During the procedure of NDTRV(Fig. 2), it constructs a list Q consisting of actions
of type (i) and (ii). We claim that Q is exactly the sequence satisfying the requirement of this theorem.
NDTRV appends an action of type (ii) to Q if and only if it adds a vertex to G. Meanwhile, NDTRV
appends an action of type (i) to Q if and only if it adds an an edge to G. It follows that there is a one-to-one
correspondence between actions of type (i) in Q and E(G) and there is a one-to-one correspondence between
actions of type (ii) in Q and V (G). According to Lemma 4, G is a forest. As a result, E(G) ≤ V (G), and
the length of Q is bounded by E(G) + V (G) ≤ 2V (G) = 2V (DF ).
3.4 The deterministic algorithm
Now we are ready to give the deterministic algorithm FLIPDT for Parameterized Flip Distance. The
specific algorithm is presented in Fig. 4. As mentioned above, we assume that NDTRV is always able to
guess the optimal choice correctly. In fact, FLIPDT achieves this by trying all possible sequences of actions
and partitions of k. At the top level, FLIPDT branches into all partitions of k, namely (k1, ..., kt) satisfying
k1 + ... + kt = k and k1, ..., kt ≥ 1, in which ki (i = 1, ..., t) equals the size of the track tree Ai constructed
during the i-th iteration.
Suppose that FLIPDT is under some partition (k1, ..., kt). Let T 0
iteration = Tstart. FLIPDT permutates
all necessary edges in Tstart in the lexicographical order, and the ordering is denoted by Olex. Here we
number the given points of P in the Euclidean plane from 1 to n arbitrarily and label one edge by a tuple
consisting of two numbers of its endpoints. Thus we can order the edges lexicographically. FLIPDT
performs t iterations. At the beginning of the i-th iteration, i = 1, ..., t, we denote the current triangulation
7
iteration. For i = 1, ..., t, T i
by T i−1
iteration is also the triangulation resulting from the execution of the first i
iterations. At the beginning of the i-th iteration (i = 1, ..., t), FLIPDT repeatedly picks the next edge
in Olex until finding a necessary edge e belonging to T i−1
iteration. Note that one edge in Olex may not be a
necessary edge anymore with respect to T i−1
iteration. Moreover, if FLIPDT reaches the end of Olex but does
not find a necessary edge belonging to T i−1
iteration, it needs to update Olex by clearing Olex and permutating
all necessary edges in T i−1
iteration lexicographically, and choose the first edge in the updated ordering Olex.
Then FLIPDT branches into every possible sequence of actions seqi of length 2ki − 1.
Under each enumeration of seqi, FLIPDT branches into every possible sequence of actions seqi+1 of
length 2ki+1 − 1. FLIPDT proceeds as above. When FLIPDT finishes the last iteration, it judges if the
resulting triangulation T t
iteration is equal to Tend. If they are equal, the input instance is a yes-instance.
Otherwise, FLIPDT rejects this branch and proceeds.
Now we analyse how to enumerate all possible sequences of length 2ki − 1. According to Lemma 4 and
Theorem 1, in every iteration NDTRV constructs a track tree in which a node corresponds to an action of
type (ii) while an edge corresponds to an action of type (i). It follows that the number of actions of type (ii)
is ki, and the number of actions of type (i) is ki − 1. According to NDTRV, the last action γ in seqi must
be of type (ii), and in any prefix of seqi − γ the number of actions of type (i) must not be less than that
of type (ii). Thus FLIPDT only needs to enumerate all sequences of length 2ki − 1 satisfying the above
constraints.
FLIPDT(Tstart, Tend, k)
Input: two triangulations Tstart and Tend of a point set P in the Euclidean plane and an
integer k.
Output: return YES if there exists a sequence of flips of length k that transforms Tstart
into Tend; otherwise return NO.
1. For each partition (k1, ..., kt) of k satisfying k1 + k2 + ... + kt = k and k1, ..., kt ≥ 1 do
1.1
1.2
2. Return NO;
Order all necessary edges in Tstart lexicographically and denote this ordering by Olex;
FDSearch(Tstart,1,(k1, ..., kt)); /*iteration 1 distributed with k1*/
FDSearch(T ,i,(k1, ..., kt))
/*the concrete branching procedure*/
Input: a triangulation T , an integer i denoting that the algorithm is at the i-th iteration and
a partition (k1, ..., kt) of k.
Output: return YES if the instance is accepted.
1. Repeatedly pick the next edge in Olex until finding a necessary edge e with respect to T
and Tend;
iteration in lexicographical order,
and pick the first edge e in Olex;
T (cid:48) = Transform(T ,seqi,e);
If i < t then /*continue to the next iteration distributed with ki+1*/
2. If it reaches the end of Olex but finds no necessary edge in T then
2.1 Update Olex by permutating all necessary edges in T i−1
3. For each possible sequence of actions seqi of length 2ki − 1 do
3.1
3.2
3.2.1
3.3 Else if i = t and T (cid:48) = Tend then
3.3.1
/*compare T (cid:48) with Tend*/
FDSearch(T (cid:48),i + 1,(k1, ..., kt));
Return YES;
Transform(T ,s,e)
/*subprocess for transforming triangulations*/
Input: a triangulation T , a sequence of actions s and a starting edges e.
Output: a new triangulation T (cid:48).
1. Perform a sequence of actions s starting from e in T , getting a new triangulation T (cid:48);
2. Return T (cid:48);
Figure 4: The deterministic algorithm for Parameterized Flip Distance
8
The following theorem proves the correctness of the algorithm FLIPDT.
Theorem 2. Let (Tstart, Tend, k) be an input instance. FLIPDT is correct and runs in time O∗(k · 32k).
Proof. Suppose that (Tstart, Tend, k) is a yes-instance. There must exist a sequence of flips F of length k such
F−→ Tend. Thus DF exists according to Definition 1. By NDTRV and Lemma 4, there exists an
that Tstart
undirected graph G consisting of a set of vertex-disjoint track trees A1, ..., At. Moreover, Theorem 1 shows
that there exists a sequence of actions Q following which we can perform all flips of DF in a topologically-
sorted order. Due to NDTRV, Q consists of several subsequences seq1, ..., seqt, in which seqi is constructed
in the i-th iteration and corresponds to the track tree Ai for i = 1, ..., t. Supposing the size of Ai is λi for
i = 1, ..., t satisfying λ1 + ... + λt = k, seqi contains λi actions of type (ii) corresponding to the nodes of Ai as
well as λi − 1 actions of type (i) corresponding to the edges of Ai. FLIPDT guesses the size of every track
tree by enumerating all possible partitions of k into (k1, ..., kt) such that k1 + ... + kt = k and k1, ..., kt ≥ 1.
We say that ki is distributed to the i-th iteration or the distribution for the i-th iteration is ki for i = 1, ..., t.
We claim that FLIPDT is able to perform a sequence Σ of actions which correctly guesses every subse-
quence seq1, ..., seqt of the objective sequence Q, that is, Σ is a concatenation of seq1, ..., seqt. Suppose that
FLIPDT has completed i iterations. We prove this claim by induction on i. At the first iteration, FLIPDT
starts by picking the first necessary edge e1 in list Olex. In the first iteration of constructing Q, NDTRV
starts by picking an arbitrary necessary edge. Without loss of generality, it chooses e1 and construct seq1
starting from e1. The length of seq1 is 2λ1 − 1. Since FLIPDT tries every distribution in {1, ..., k} for
the first iteration and 1 ≤ λ1 ≤ k, there is a correct guess of the distribution equal to λ1 for this iteration.
Under this correct guess, FLIPDT tries all possible sequences of actions of length 2λ1 − 1 starting from e1.
It follows that FLIPDT is able to perform a sequence that is equals to seq1 in the first iteration resulting
in a triangulation T1.
Suppose that the claim is true for any first i iterations (1 ≤ i < t). That is, under some guess for
the partition of k, λ1, ..., λi are distributed to the first i iterations respectively. Moreover, FLIPDT has
completed i iterations and performed a sequence of actions seqconcat,i, which is equal to the concatenation
of seq1, ..., seqi, resulting in a triangulation Ti. Based on Ti and seqconcat,i, FLIPDT is ready to perform
the (i + 1)-th iteration. Suppose that FLIPDT picks ei+1 from Olex. Let us see the construction of Q in
NDTRV. Suppose NDTRV has constructed the first i track trees A1, ..., Ai, and it is ready to begin a
new iteration by arbitrarily picking a necessary edge in the current triangulation. Since FLIPDT correctly
guessed and performed the first i subsequences of Q, Ti is exactly equal to the current triangulation in
NDTRV. Thus ei+1 is a candidate edge belonging to the set of all selectable necessary edges for NDTRV
in this iteration. Without loss of generality, it chooses ei+1 and constructs seqi+1 of length 2λi+1− 1 starting
from ei+1. Since the sizes of A1, ..., Ai are λ1, ..., λi respectively, we get that 1 ≤ λi+1 ≤ k−(λ1 +...+λi). We
argue that FLIPDT is able to perform a sequence that is equal to the concatenation of seq1, ..., seqi+1. Since
the edges in Olex are ordered lexicographically and FLIPDT chooses necessary edges in a fixed manner,
FLIPDT is sure to choose ei+1 to begin the (i + 1)-th iteration for every guessed sequence in which the
first i subsequences are equal to seq1,...,seqi respectively. Thus FLIPDT actually tries every distribution in
{1, ..., k − (λ1 + ... + λi}) for the (i + 1)-th iteration starting from ei+1 based on Ti and seqconcat,i. It follows
that there is a correct guess of distribution for the (i + 1)-th iteration which is equal to λi+1. Under this
correct guess of distribution, FLIPDT tries all possible sequences of length 2λi+1 − 1 starting from ei+1 on
Ti based on seqconcat,i, ensuring that one of them is equal to seqi+1. It follows that the claim is true for the
first i + 1 iterations. This completes the inductive proof for the claim.
If (Tstart, Tend, k) is a yes-instance, the action sequence Q of length at most 2k exists and the deterministic
algorithm can find such a sequence. Otherwise, there is no valid sequence F of length k. Thus there is no
such action sequence Q. As a result, FLIPDT returns NO. It is proved that FLIPDT decides the given
instance (Tstart, Tend, k) correctly.
Finding and ordering all necessary edges in Tstart takes O(n+k log k) time, and FLIPDT may update the
ordering Olex at the beginning of each iteration. The number of partitions of k is known as the composition
number of k, which is 2k−1. Under each partition (k1, ..., kt) of k and for each ki, i = 1, ..., t, we enumerate
all possible subsequences of actions in which there are ki actions of type (ii) and ki − 1 actions of type (i).
9
there are four choices for action (i) and one choice for action (ii). Here we use Stirling's approximation
2πn(n/e)n and get that (cid:0)2(ki−1)
It follows that the number of all possible subsequences is bounded by (cid:0)2(ki−1)
n! ≈ √
O∗(16kt) = O∗(16k) cases under each partition. Since for each case we can perform the sequence of actions
in O(k) time, and the resulting triangulation can be compared to Tend in O(k) time, the running time of the
whole algorithm is bounded by O∗(k · 2k−1 · (n + k log k) + k · 2k−1 · 16k) = O∗(k · 32k).
(cid:1) × 4ki−1 = O∗(16ki) since
(cid:1) = O∗(4ki). It follows that there are O∗(16k1) × O∗(16k2) × ... ×
on each instance (Tstart, Tend, k(cid:48)) for k(cid:48) = 0, ..., k. The running time is bounded by (cid:80)k
According to the definition of Parameterized Flip Distance, we need to check if we can find a
shorter valid sequence for the given triangulations Tstart and Tend. This is achieved by calling FLIPDT
) =
O∗(k · 32k).
k(cid:48)=0 O∗(k(cid:48) · 32k(cid:48)
ki−1
ki−1
4 Conclusion
In this paper we presented an FPT algorithm running in time O∗(k · 32k) for Parameterized Flip Dis-
tance, improving the previous O∗(k · ck)-time (c ≤ 2 × 1411) FPT algorithm by Kanj and Xia [11]. An
important related problem is computing the flip distance between triangulations of a convex polygon, whose
traditional complexity is still unknown. Although our algorithm can be applied to the case of convex polygon,
it seems that an O(ck) algorithm with smaller c for this case probably exists due to its more restrictive geo-
metric property. In addition, whether there exists a polynomial kernel for Parameterized Flip Distance
is also an interesting problem.
References
[1] Oswin Aichholzer, Ferran Hurtado, and Marc Noy. A lower bound on the number of triangulations of
planar point sets. Computational Geometry, 29(2):135 -- 145, 2004. doi:10.1016/j.comgeo.2004.02.
003.
[2] Oswin Aichholzer, Wolfgang Mulzer, and Alexander Pilz. Flip distance between triangulations of a
simple polygon is NP-complete. Discrete & Computational Geometry, 54(2):368 -- 389, 2015. doi:10.
1007/s00454-015-9709-7.
[3] Jianer Chen. Parameterized computation and complexity: A new approach dealing with np-hardness.
J. Comput. Sci. Technol., 20(1):18 -- 37, 2005. doi:10.1007/s11390-005-0003-7.
[4] Jianer Chen, Donald K. Friesen, Weijia Jia, and Iyad A. Kanj. Using nondeterminism to design efficient
deterministic algorithms. Algorithmica, 40(2):83 -- 97, 2004. doi:10.1007/s00453-004-1096-z.
[5] Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, D´aniel Marx, Marcin Pilipczuk,
Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer, 2015. doi:10.1007/
978-3-319-21275-3.
[6] Mark de Berg, Otfried Cheong, Marc J. van Kreveld, and Mark H. Overmars. Computational geometry:
algorithms and applications, 3rd Edition. Springer, 2008.
[7] Gerald E. Farin. Curves and surfaces for computer-aided geometric design - a practical guide (4. ed.).
Computer science and scientific computing. Academic Press, 1997.
[8] Bernd Hamann. Modeling contours of trivariatc data. Mathematical Modeling and Numerical Analysis,
26:51 -- 75, 1992.
[9] Ferran Hurtado, Marc Noy, and Jorge Urrutia. Flipping edges in triangulations. Discrete & Computa-
tional Geometry, 22(3):333 -- 346, 1999. doi:10.1007/PL00009464.
10
[10] Iyad A. Kanj, Eric Sedgwick, and Ge Xia. Computing the flip distance between triangulations. Discrete
& Computational Geometry, 58(2):313 -- 344, 2017. doi:10.1007/s00454-017-9867-x.
[11] Iyad A. Kanj and Ge Xia. Flip Distance is in FPT time O(n + k · ck).
In proceedings of the 32nd
International Symposium on Theoretical Aspects of Computer Science, STACS, Garching, Germany,
pages 500 -- 512, 2015. doi:10.4230/LIPIcs.STACS.2015.500.
[12] Charles L. Lawson. Transforming triangulations. Discrete Mathematics, 3(4):365 -- 372, 1972. doi:
10.1016/0012-365X(72)90093-3.
[13] Anna Lubiw and Vinayak Pathak. Flip distance between two triangulations of a point set is NP-
complete. Computational Geometry, 49:17 -- 23, 2015. doi:10.1016/j.comgeo.2014.11.001.
[14] Joan M. Lucas. An improved kernel size for rotation distance in binary trees. Information Processing
Letters, 110(12-13):481 -- 484, 2010. doi:10.1016/j.ipl.2010.04.022.
[15] Alexander Pilz. Flip distance between triangulations of a planar point set is APX-hard. Computational
Geometry, 47(5):589 -- 604, 2014. doi:10.1016/j.comgeo.2014.01.001.
[16] Larry L. Schumaker. Triangulations in CAGD. IEEE Computer Graphics and Applications, 13(1):47 -- 52,
1993. doi:10.1109/38.180117.
[17] Daniel D. Sleator, Robert E. Tarjan, and William P. Thurston. Rotation distance, triangulations, and
hyperbolic geometry. In Proceedings of the 18th Annual ACM Symposium on Theory of Computing,
STOC, Berkeley, USA, pages 122 -- 135, 1986. doi:10.1145/12130.12143.
11
|
1310.5786 | 1 | 1310 | 2013-10-22T03:01:41 | Contracting Graphs to Split Graphs and Threshold Graphs | [
"cs.DS"
] | We study the parameterized complexity of Split Contraction and Threshold Contraction. In these problems we are given a graph G and an integer k and asked whether G can be modified into a split graph or a threshold graph, respectively, by contracting at most k edges. We present an FPT algorithm for Split Contraction, and prove that Threshold Contraction on split graphs, i.e., contracting an input split graph to a threshold graph, is FPT when parameterized by the number of contractions. To give a complete picture, we show that these two problems admit no polynomial kernels unless NP\subseteq coNP/poly. | cs.DS | cs |
Contracting Graphs to Split Graphs and
Threshold Graphs
Leizhen Cai and Chengwei Guo
Department of Computer Science and Engineering, The Chinese University of Hong Kong,
Hong Kong S.A.R., China
{lcai,cwguo}@cse.cuhk.edu.hk
Abstract. We study the parameterized complexity of Split Contraction and Thresh-
old Contraction. In these problems we are given a graph G and an integer k and
asked whether G can be modified into a split graph or a threshold graph, respectively,
by contracting at most k edges. We present an FPT algorithm for Split Contrac-
tion, and prove that Threshold Contraction on split graphs, i.e., contracting an
input split graph to a threshold graph, is FPT when parameterized by the number of
contractions. To give a complete picture, we show that these two problems admit no
polynomial kernels unless N P ⊆ coN P/poly.
1
Introduction
Graph modification problems constitute a fundamental and well-studied family of problems
in algorithmic graph theory, many famous graph problems can be formulated as graph
modification problems such as Clique, Feedback Vertex Set, and Minimum Fill-
in. A graph modification problem takes as input a graph G and an integer k, and the
question is whether G can be modified to belong to a specified graph class, using at most k
operations of a certain specified type such as vertex deletion or edge deletion. The number k
of operations measures how close a graph is to such a specified class of graphs. Recently the
study of modifying a graph by using operations of edge contraction has been initiated from
the parameterized point of view, yielding several results for the Π-Contraction problem.
Π-Contraction
Instance:
Question:
Parameter:
Graph G = (V, E), positive integer k.
Can we obtain a Π-graph (i.e. a graph belonging to class Π)
from G by contracting at most k edges?
k.
A problem (with a particular parameter k) is fixed-parameter tractable (FPT) if it can
be solved in time f (k)nO(1) where f (k) is a computable function depending only on k.
Considering parameterization by the number of edge contractions used to modify graphs,
the Π-Contraction problem has been proved to be FPT when Π is the class of bipartite
graphs (Heggernes et al. [14]), the class of trees or paths (Heggernes et al. [13]), the class
of planar graphs (Golovach et al. [11]), the class of cliques (Cai et al. [5] and Lokshtanov
et al. [16]). On the negative side, very recently two groups of authors (Cai et al. [5] and
1
Lokshtanov et al. [16]) showed that Chordal Contraction is fixed-parameter intractable.
It is then natural to ask whether the Π-contraction problems are FPT for two well-known
subclasses Π of chordal graphs: split graphs and threshold graphs.
In this paper, we study the parameterized complexity of Π-Contraction when Π is
the class of split graphs and when Π is the class of threshold graphs. Their edge deletion
versions known as Split Deletion and Threshold Deletion, asking whether an input
graph can be modified into a split graph or a threshold graph, respectively, by deleting at
most k edges, are FPT and have polynomial kernels [4, 9, 12]. The combination of split
graphs, threshold graphs, and edge contractions has been studied in a closely related set-
ting. Belmonte et al. [1] showed that given a split graph G and a threshold graph H, the
problem of determining whether G is contractible to H is NP-complete, and it can be solved
in polynomial time for fixed V (H). Inspired by this, we consider the case that H is an
arbitrary threshold graph that is not a part of the input, and take as parameter the number
of edge contractions instead of the size of the target graph H.
Our Contribution: We show that Split Contraction can be solved in 2O(k2)nO(1) time.
This result complements the FPT results of two other graph modification problems related
to split graphs: Split Deletion and Split Vertex Deletion. Our algorithm starts
by finding a large split subgraph and further is partitioned into two parts in terms of the
clique size of this split subgraph. If the clique is large, we use a branch-and-search algorithm
to enumerate edge contractions and reduce to the Clique Contraction problem that is
known to be FPT. Otherwise, there will be a large independent set in the input graph. We
partition all vertices into bounded number of independent sets such that vertices in each set
have the same neighbors, and then use reduction rules to reduce to a smaller graph. The
ideas of reduction rules are applicable to obtain kernelization algorithms for other contraction
problems such as Clique Contraction and Biclique Contraction.
We also obtain an 2O(k2)nO(1) time algorithm for Threshold Contraction when the
input is a split graph. One motivation to study this problem is for considering an aspect
of modification problems: contracting few edges to make all vertices in a graph obey some
ordering (in this problem the ordering by neighborhood inclusion). We feel that it is of more
interest to study the Threshold Contraction problem for split graphs than for general
graphs, since in the latter case the part of contracting input graphs to split graphs and the
part of ordering vertices by edge contractions have lack of connection.
Furthermore, we prove that the above problems admit no polynomial kernels unless
N P ⊆ coN P/poly by polynomial-time reductions from Red-Blue Dominating Set and
One-Sided Dominating Set, respectively, with polynomial bounds on the new parameter
values. The incompressibility of Red-Blue Dominating Set is provided by Dom et al. [7].
We introduce One-Sided Dominating Set and One-Sided Domatic Number that were
defined by Feige et al. [8], and show that these problems are incompressible. Such results
might have application in obtaining kernelization lower bounds for further problems.
2 Preliminaries
Graphs: We consider simple and undirected graphs G = (V, E), where V is the vertex set
and E is the edge set. Two vertices u, v ∈ V are adjacent iff uv ∈ E. A vertex v is incident
with an edge e iff v ∈ e, i.e., v is an endpoint of e. The neighbor set NG(v) of a vertex v ∈ V
in graph G is the set of vertices that are adjacent to v in G. We use NG[v] to denote the
closed neighbor set of v in G where NG[v] = NG(v) ∪ {v}. For a set X of vertices or edges
in G, we use G \ S or G − X to denote the graph obtained by deleting X from G. For a set
2
of vertices V ′ ⊆ V , we denote by E[V ′] the set of edges whose both endpoints are in V ′. An
l-path (or Pl) is a path on l vertices.
We study two hereditary graph classes. A graph G is a split graph if its vertex set can be
partitioned into a clique K and an independent set I, where (K; I) is called a split partition
of G. The class of split graphs is characterized by a set of forbidden induced subgraphs:
{2K2, C4, C5}. A graph G is a threshold graph if there exist non-negative reals r(v) for each
v ∈ V (G) and "threshold" t such that for every vertex set X ⊆ V (G), X is an independent
set iff Σv∈X r(v) ≤ t. The class of threshold graphs is characterized by a set of forbidden
induced subgraphs: {2K2, C4, P4}.
Edge Contraction: The contraction of edge e = uv in G removes u and v from G, and
replaces them by a new vertex adjacent to precisely those vertices which were adjacent to
at least one of u or v. The resulting graph is denoted by G/e or G · e. For a set of edges
F ⊆ E(G), we write G/F to denote the graph obtained from G by sequentially contracting
all edges from F .
For a graph H, if H can be obtained from G by a sequence of edge contractions, then G
is contractible to H, or called H-contractible. Let V (H) = {h1, · · · , hl}. G is H-contractible
if G has a so-called H-witness structure: a partition of V (G) into l sets W (h1), · · · , W (hl),
called witness sets, such that each W (hi) induces a connected subgraph of G and for any
two hi, hj ∈ V (H), there is an edge between W (hi) and W (hj) in G iff hihj ∈ E(H). We
obtain H from G by contracting vertices in each W (hi) into a single vertex.
Parameterized Complexity: A paramerized problem Q is a subset of Σ∗ × N for some
finite alphabet Σ. The second component is called the parameter. The problem Q is fixed-
parameter tractable if it admits an algorithm deciding whether (I, k) ∈ Q in time f (k)IO(1),
where f is a computable function depending only on k.
A kernelization of Q is a polynomial-time computable function that maps instance (I, k)
to another instance (I ′, k′) such that:
• (I, k) ∈ Q ⇔ (I ′, k′) ∈ Q;
• I ′, k′ ≤ g(k) for some computable function g.
If g is a polynomial function then we say that Q admits a polynomial kernel. A problem Q
is incompressible if it admits no polynomial kernel unless N P ⊆ coN P/poly.
3 Contracting graphs to split graphs
In this section, we consider the Split Contraction problem: Given a graph G and an
integer k, can we obtain a split graph from G by contracting at most k edges?
We point out that this problem is NP-complete by reducing from another NP-complete
edge contraction problem Clique Contraction [5]: Given a graph G and an integer k,
can we modify G into a clique by contracting at most k edges? We construct a graph G′
from G by adding an independent set of k + 2 new vertices and making new vertices adjacent
to all vertices in G. Observe that at least two of the new vertices are not involved in any
edge contraction, then at least one of them belongs to the independent set of the resulting
split graph, which implies that all the old vertices belong to the clique of the resulting graph.
Thus G′ can be modified into a split graph using k edge contractions iff there exists a set of
k edges in G whose contraction makes G into a clique.
Theorem 3.1. Split Contraction is NP-complete.
3
We now present an FPT algorithm for Split Contraction based on an kO(k) + O(m)
time algorithm for Clique Contraction [5].
Note that an n-vertex graph G must contain an induced split subgraph of (n − 2k)
vertices if (G, k) is yes-instance of Split Contraction, because k edge contractions can
affect at most 2k vertices. We start by finding an (n − 2k)-vertex induced split subgraph H
in 22knO(1) time using a known algorithm for Split Vertex Deletion (Ghosh et al. [9]).
Let Vk = V (G) − V (H), and let (KH ; IH ) be a split partition of H where KH is a maximal
clique and IH is an independent set. Here we first assume that KH > 2k implying that at
least one vertex in KH is not involved in any edge contraction, and will discuss the case for
KH ≤ 2k in the last part of the algorithm.
We branch out by contracting every possible set E′ ⊆ E[Vk] of at most k edges and
obtain the resulting instance (G′, k′) where G′ = G/E′ and k′ = k − E′. For each resulting
instance (G′, k′), suppose that vertices Vk are contracted to vertices V ′
k in G′.
Proposition 3.2. (G, k) has a solution S iff there exists a resulting instance (G′, k′) such
that (G′, k′) has a solution F = S − E′ satisfying F ∩ E[V ′
k] = ∅.
Suppose that (G′, k′) is a yes-instance and has a solution F . Thus the graph G′/F is
a split graph and has a split partition (KF ; IF ). We further branch on at most 3V ′
k ways
to find a partition V ′
k = R ∪ Kp ∪ Ip such that R consists of exactly those vertices in V ′
k
that are incident with some edges in F , Kp ⊆ KF induces a clique, and Ip ⊆ IF induces an
independent set. See Fig. 1 for an illustration. It is easy to see that R ≤ k′.
IH
KH
H
Ip
Kp
R
G′[V ′
k]
Figure 1: An illustration of the structure of G′
It is clear that those vertices in IH that are adjacent to some vertices in Ip must be in
the clique KF of the target split graph, and other vertices in IH could be in the independent
set IF after contractions. The following proposition states that almost all vertices in KF
are finally in the clique of some target graph.
Proposition 3.3. If (G′, k′) is a yes-instance, then it has a solution F such that there is
at most one vertex in KH that is finally in the independent set IF .
Proof. For an arbitrary solution F of (G′, k′), if there are at least two vertices a, b ∈ KH
that are finally in the independent set IF , they must be contained in a same witness set WF
because they are adjacent originally. Thus the number of such vertices is bounded by k + 1,
implying that there is a vertex u ∈ KH that is finally in the clique KF since KH > 2k.
By the definition of witness set we see that the induced subgraph G′[WF ] has a spanning
tree whose edges are entirely in F . Thus we can remove one edge from F to separate the
vertex b from the witness set WF , and add an edge ub into F which implies that b is adjacent
to all vertices in KF after contracting F . It is easy to see that the resulting set obtained
from F is also a solution of (G′, k′) and b is no longer in the independent set.
4
Our algorithm further considers the following two cases. It outputs "YES" if either case
outputs "YES".
Case 1. There is no vertex in KH that is finally in IF . Let T1 be a subset of IH
containing exactly those vertices that are adjacent to some vertices in Ip. It is clear that
all vertices in T1 are finally in KF after contractions. Therefore every vertex in T1 must be
involved in at least one edge contraction, and we have T1 ≤ k′.
Remember that R consists of exactly those vertices in V ′
k that are incident with some
edges in F , which implies that vertices in R are merged into KH ∪ IH after contracting some
edges in F . It is easy to see that in order to contract G′ into a split graph, it is better to
merge vertices R into KH ∪ T1 than into IH − T1. Thus we may assume that all vertices in R
are finally in KF after contractions. Our goal becomes to check whether T1 ∪ KH ∪ R ∪ Kp
induces a clique after contracting at most k′ edges in G′.
Since KH > 2k, there exists a vertex u ∈ KH that is not involved in any edge contrac-
tion. Obviously u is adjacent to all vertices in KF . We can obtain an edge set F1 from F by
removing every edge of F whose endpoints are both outside KH , and replacing every edge
ab ∈ F that a ∈ IH − T1, b ∈ KH by an edge ub. Since contracting such edge ab only affects
the vertex b which can be merged into the clique KF by contracting ub, it is clear that F1
is also a solution set of G′ consisting of edges whose one endpoint is in KH and another
endpoint is in T1 ∪ KH ∪ R ∪ Kp.
Proposition 3.4. (G′, k′) has a solution set that is entirely contained in G′[T1∪KH ∪R∪Kp]
if it is a yes-instance for Case 1.
By the above result, we first find the set T1 in polynomial time, and then apply the FPT
algorithm for Clique Contraction [5] to determine whether G′[T1 ∪ KH ∪ R ∪ Kp] can
be made into a clique by at most k′ edge contractions. If it outputs "YES", then (G′, k′) is
a yes-instance. The running time is bounded by k′O(k′) + O(m).
Case 2. There is exactly one vertex in KH that is finally in IF . We check whether
(G′, k′) is a yes-instance with the assumption KH ∩ IF = {w} for every w ∈ KH that is
not adjacent to any vertex in Ip.
Let T2 be a subset of IH containing exactly those vertices that are adjacent to some
vertices in Ip ∪ {w}. Since each vertex in T2 is finally in KF and thus is involved in some
edge contraction, we have T2 ≤ k′. Our goal is to check whether T2 ∪ (KH − {w}) ∪ R ∪ Kp
induces a clique after contracting at most k′ edges in G′. Similar to Case 1, we have the
following proposition.
Proposition 3.5. (G′, k′) has a solution set that is entirely contained in G′[T2 ∪ (KH −
{w}) ∪ R ∪ Kp] if it is a yes-instance for Case 2.
We also apply the k′O(k′) + O(m) time algorithm for Clique Contraction [5] to
determine whether G′[T2 ∪ (KH − {w}) ∪ R ∪ Kp] can be made into a clique by contracting
at most k′ edges. If it outputs "YES", then (G′, k′) is a yes-instance.
Combing Case 1 and Case 2, we can decide whether a resulting instance (G′, k′) is a
yes-instance in kO(k) + O(m) time when KH > 2k.
Furthermore, we deal with the remaining case: KH ≤ 2k. We partition the vertex set
V (G) into disjoint sets X1, · · · , Xd such that each Xi induces a maximal independent set
satisfying that the vertices in Xi have the same sets of neighbors in G. This procedure is
equivalent to partitioning the complement graph G into critical cliques, which can be done
5
in linear time [6, 15]. A critical clique K in a graph is a clique such that all vertices in K
have the same closed neighbor sets, and K is maximal under this property. It has been
proved that all vertices in a graph can be uniquely partitioned into groups such that each
group induces a critical clique [6, 15]. We now use the following reduction rules to obtain a
smaller instance:
Rule 1.
If d > 24k + 4k, then output "NO".
Rule 2.
2k + 5 vertices among them and remove others in Xi from G.
If there are more than 2k + 5 vertices in Xi for some i, then arbitrarily retain
It is clear that applying these reduction rules requires linear time. We now show the
correctness of Rule 1 and Rule 2.
Lemma 3.6. Rule 1 and Rule 2 are correct.
Proof. Since KH ≤ 2k, we have Vk ∪ KH ≤ 4k and IH ≥ n − 4k. Note that vertices
in IH have at most 24k different connection configurations to vertices Vk ∪ KH . Thus IH
can be partitioned into at most 24k maximal independent sets such that vertices in each set
have the same neighbors in G. Including vertices of Vk ∪ KH , the number d is bounded by
24k + 4k if (G, k) is a yes-instance. Thus Rule 1 is correct.
Moreover, we prove that the input graph G has a k-solution iff the graph G∗ obtained
after one application of Rule 2 has a k-solution. Let Yi be the set of remaining vertices in
Xi for i = 1, · · · , d.
Suppose that G has a solution S ⊆ E(G). For every vertex a that is incident with edges
in S and is removed after applying Rule 2, there exists 1 ≤ j ≤ d such that a ∈ Xj − Yj.
Note that Yj = 2k + 5 > 2k, there is another vertex b ∈ Yj that is not incident with any
edge in S. We replace all edges {ac ∈ S : c ∈ V (G)} by {bc : ac ∈ S} in S for every such
a, and obtain a set S′ ⊆ E(G∗). It is clear that G∗/S′ is an induced subgraph of G/S′
because S′ ⊆ E(G∗). Since NG(a) = NG(b) for every such a, it is easy to see that G/S′ is
isomorphic to G/S that is a split graph. Therefore G∗/S′ is a split graph.
Conversely, suppose that G∗ has a solution S∗, i.e., G∗/S∗ is a split graph. We claim
that G/S∗ is also a split graph. Towards a contradiction, we assume that G/S∗ contains an
induced subgraph D isomorphic to one graph of {2K2, C4, C5}. For every vertex a that is
contained in V (D) and is removed after applying Rule 2, a is in Xj − Yj for some j. Since
S∗ ⊆ E(G∗), a is not incident with any edge in S∗. Note that Yj = 2k + 5 ≥ 2k + V (D),
there is a vertex b ∈ Yj that is not contained in V (D) and is not incident with any edge in
S∗. We replace a by b in V (D) for every such a, and obtain an induced subgraph D′ whose
vertices are all included in V (G∗/S∗). Since NG(a) = NG(b) for every such a, it is easy to
see that D′ is isomorphic to D, contradicting to the fact that G∗/S∗ is a split graph that is
D-free. Thus G/S∗ is {2K2, C4, C5}-free, and S∗ is a solution of G.
After applying Rule 1 and Rule 2, we reduce G into a graph of O(24kk) vertices, implying
that the problem can be solved in 2O(k2) + O(m) time using exhaustive search if K1 ≤ 2k.
Our FPT algorithm for Split Contraction is summarized in the following table:
6
Algorithm Split Contraction (G, k)
• Find an induced split subgraph H = (KH ; IH ) of size (n − 2k) in G. If it does not exist,
output "NO". Let Vk = V (G) − V (H).
• If KH > 2k, then
-- Branch into instances (G′, k′) by contracting edges E ′ ⊆ E[Vk]. Enumerate all
partitions V ′
k = (R, Kp, Ip).
-- Case 1: Find T1 = {v ∈ IH ∃x ∈ Ip, vx ∈ E(G′)}. Determine whether (G′[T1 ∪
KH ∪ R ∪ Kp], k′) is a yes-instance of Clique Contraction. If yes, output "YES".
-- Case 2: For every w ∈ KH , find T2 = {v ∈ IH ∃x ∈ Ip ∪ {w}, vx ∈ E(G′)}.
Determine whether (G′[T2 ∪ (KH − {w}) ∪ R ∪ Kp], k′) is a yes-instance of Clique
Contraction. If yes, output "YES".
-- If for every (G′, k′) neither Case 1 nor Case 2 output "YES", then output "NO".
• Else
-- Partition V (G) into disjoint sets X1, · · · , Xd such that each Xi induces a maximal
independent set satisfying that vertices in Xi have the same neighbors.
-- Reduction Rule 1: If d > 24k + 4k, then output "NO".
-- Reduction Rule 2: If there are more than 2k + 5 vertices in Xi for some i, then
arbitrarily retain 2k + 5 vertices among them and remove others in Xi from G.
-- Use exhaustive search to find a solution in the reduced graph G∗.
Theorem 3.7. Split Contraction can be solved in 2O(k2)nO(1) time.
Proof. In the above algorithm, we use an 22knO(1) time algorithm to find an (n − 2k)-vertex
induced split subgraph H. If KH > 2k, we branch into at most E[Vk]k = kO(k) instances
and enumerate at most 3V ′
k = 3O(k) partitions, and for each resulting instance it costs
kO(k) + O(m) time to determine whether it is a yes-instance. Thus the running time of
this step is bounded by kO(k)nO(1).
If KH ≤ 2k, the running time of the algorithm is
2O(k2) + O(m). Therefore the total running time is 2O(k2)nO(1).
4 Contracting split graphs to threshold graphs
After obtaining an FPT algorithm for contracting general graphs to split graphs, we consider
the problem of contracting split graphs to threshold graphs, which is formally defined as
Threshold Contraction on split graphs: Given a split graph G and an integer k, can
we obtain a threshold graph from G by contracting at most k edges?
Threshold graphs constitute a subclass of both split graphs and interval graphs in graph
theory. A graph G is a threshold graph if it is a split graph and for any split partition
(K; I) of G, there is an ordering (u1, u2, · · · , up) of the vertices in K such that N [u1] ⊆
N [u2] ⊆ · · · ⊆ N [up] and there is an ordering (v1, v2, · · · , vq) of the vertices in I such that
N (v1) ⊇ N (v2) ⊇ · · · ⊇ N (vq), i.e., a split graph containing no induced P4.
Note that the class of split graphs is closed under edge contractions, and contracting edges
in a split graph never generates a new induced P4 because split graphs contain no induced
Pl for l > 4. Thus contracting a split graph to a threshold graph is essentially removing all
induced P4 from the split graph by edge contractions. Based on this observation, we obtain
the following FPT algorithm.
7
Given an input split graph G with n vertices and an integer k, we find a split partition
(K; I) for G where K is a maximal clique and I is an independent set. We construct a
bounded search tree whose root is labelled by (G, k, T ), where T represents a set of vertices
in K that are incident with edges in a solution set and is initially an empty set ∅.
Observation 4.1. For any v ∈ I and any x, y ∈ K such that vx, vy ∈ E(G), these two
graphs, G/vx and G/vy, are isomorphic.
For any induced 4-path v1v2v3v4 in G where v1, v4 ∈ I and v2, v3 ∈ K, there are five
ways to destroy this path by edge contractions: contracting v1x ∈ E(G) for any x ∈ K,
contracting v4y ∈ E(G) for any y ∈ K, contracting v2v3, making v2 adjacent to v4, or
making v3 adjacent to v1. The first two cases are based on Observation 4.1. The last two
cases imply that v2 or v3 is involved in at least one edge contraction. So we branch into
5 instances: (G/v1v2, k − 1, T ), (G/v3v4, k − 1, T ), (G/v2v3, k − 1, T ), (G, k, T ∪ {v2}), and
(G, k, T ∪ {v3}). For the fourth case, we mark every induced P4 containing both v2 and v4
in the current graph. For the fifth case, we mark every induced P4 containing both v3 and
v1 in the current graph. We continue branching whenever k > 0, T < 2k, and the graph
contains an induced P4 that is not marked. Clearly the search tree has height at most 2k,
and its size is bounded by O(52k).
Proposition 4.2. (G, k) is a yes-instance iff there exists at least one leaf (G′, k′, T ′) in the
search tree such that (G′, k′) is a yes-instance, T ′ ≤ 2k′, and every induced P4 in G′ is
marked.
Next, we determine whether a leaf (G′, k′, T ′) is a yes-instance in FPT time. Let (K ′; I ′)
be the split partition of G′ corresponding to (K; I). Since every induced P4 in G′ is marked
and thus intersects T ′, then G′ − T ′ contains no induced P4. Therefore (K ′ − T ′; I ′) con-
stitutes a threshold graph and there exists a vertex u ∈ K ′ − T ′ such that N [u] ⊇ N [w] for
every w ∈ K ′ − T ′. For each w ∈ K ′ − T ′, let P (w) be a set of exactly those vertices v ∈ T ′
such that there exists an induced P4 containing the edge wv in G′. We group the vertices
{w ∈ K ′ − T ′ : P (w) > 0} into different sets R1, · · · , Rd by P (w) (i.e., any w1 and w2 are
grouped into a same set iff P (w1) = P (w2) 6= ∅), and guarantee that Ri ≤ 2k′ + 1 for each
i. If there are more than 2k′ + 1 vertices w that have the same P (w) values, then choose
2k + 1 of them with largest degrees to form Ri. Obviously d < 2T ′ ≤ 22k′
.
Suppose that (G′, k′) is a yes-instance and has a minimal solution set S. If there is a
vertex w ∈ I ′ that is incident with some edge wv in S, then removing wv from S also yields
a solution set because every induced P4 in G′ is marked, contradicting to the fact that S is
minimal. Thus we have S ⊆ E[K ′]. We now use S to obtain a solution set that is entirely
contained in a set whose size is bounded by a function of k′.
We first replace every xw ∈ S that x ∈ K ′, w ∈ K ′ − T ′, and w is not contained in any
induced P4 (i.e., P (w) = ∅) by an edge xu, and then obtain a set S1. Note that contracting
xw only affects such induced 4-path a1xa2a3 that a2 ∈ K ′ and a1, a3 ∈ I ′. To destroy this
path, w must be adjacent to a3 in G′. Thus u is adjacent to a3 because N [w] ⊆ N [u],
implying that contracting xu instead of xw also destroys the path a1xa2a3 (see Fig. 2(a)).
Lemma 4.3. S1 is a solution set of (G′, k′) if S is.
8
a3
a2
K ′
a1
x
u
w
(a)
b4
b1
b3 b5
w
w′
(b)
K ′
b2
Figure 2: (a) a1xa2a3 can be destroyed by contracting xu instead of xw; (b) b1wb2b3 can be
destroyed by contracting S1 − {wy}
Furthermore for any w ∈ K ′ − T ′ such that P (w) 6= ∅ and w is not contained in any Ri,
there exits a vertex w′ ∈ Rj for some j such that P (w′) = P (w) and w′ is not involved in
any edge contraction since Rj = 2k′ + 1 > 2k′. We replace the edges {wy ∈ S1 : y ∈ K ′}
incident with w by {w′y : wy ∈ S1} for every such w and then obtain a set S2. We show
that S2 is also a solution set.
Lemma 4.4. S2 is a solution set of (G′, k′) if S1 is.
Proof. Note that contraction of each wy ∈ S1 only destroys the induced 4-paths containing
either w or y, then other induced 4-paths in G′ are destroyed by contracting S1 − {wy}.
Since w′ ∈ Rj, w /∈ Rj, and P (w′) = P (w), we have N [w] ⊆ N [w′] by the construction rule
of Rj. Similar to the above proof for S1, contracting w′y instead of wy will destroy the
paths that contain y but not w.
Moreover for any induced path b1wb2b3 containing w, there exists another induced path
b4w′b2b5 containing w′ and b2 in G′ because b2 ∈ P (w) = P (w′). Clearly {b1, w′, b2, b5}
constitute an induced P4 in G′ since b1 and b2 are not adjacent and w′ and b5 are not
adjacent either. Note that b1w′b2b5 is destroyed by contracting S1 − {wy} and w′ is not
involved in any edge contraction. Therefore b2 must be adjacent to b1 after contractions,
implying that b1wb2b3 can be destroyed by contracting S1 − {wy} (see Fig. 2(b)).
Thus contracting (S1 − {wy}) ∪ {w′y} also removes all induced 4-paths in G′, implying
that S2 is a solution.
Let R = R1 ∪ · · · ∪ Rd, we have the following result.
Corollary 4.5. (G′, k′) has a solution set that is entirely contained in E[T ′ ∪ R ∪ {u}] if it
is a yes-instance.
Thus to determine whether (G′, k′) is a yes-instance, we use exhaustive search to check
k′
whether there is a k′-size solution set in E[T ′ ∪ R ∪ {u}], which costs O((cid:0)T ′+R+1
(cid:1)
2O(k′2) time since R ≤ (2k′ + 1)22k′
Contraction on split graphs.
) =
. We have obtained an FPT algorithm for Threshold
2
Theorem 4.6. Threshold Contraction can be solved in 2O(k2)nO(1) time for split
graphs.
5
Incompressibility
In this section, we show that Split Contraction and Threshold Contraction on
split graphs are very unlikely to have polynomial kernels. To this end, we give polynomial
9
parameter transformations from two domination problems: Red-Blue Dominating Set
and One-Sided Dominating Set. A polynomial parameter transformation from a problem
Q to another problem Q′ is a polynomial-time computable function that maps (I, k) to
(I ′, k′) such that: (a) (I, k) ∈ Q ⇔ (I ′, k′) ∈ Q′; (b) k′ ≤ h(k) for some computable
function h. Bodlaender et al. [3] proved that if Q is incompressible and there is a polynomial
parameter transformation from Q to Q′, then Q′ is incompressible as well. Red-Blue
Dominating Set and One-Sided Dominating Set are defined as follows:
Red-Blue Dominating Set
Instance:
Question:
Parameter:
Bipartite graph G = (X, Y ; E) and an integer t.
Does Y have a subset of at most t vertices that dominates X?
X, t.
One-Sided Dominating Set
Instance:
Question:
Parameter:
Bipartite graph G = (X, Y ; E) and an integer t.
Does X have a subset of at most t vertices that dominates Y ?
X.
These two problems of similar definitions have different meanings. If we consider the set
X of the input bipartite graph as the "small side" since the cardinality of X is the parameter
of the problem, and the set Y as the "large side" since the cardinality of Y could be very
large, the Red-Blue Dominating Set problem is to find a set in "large side" to dominate
"small side" while the One-Sided Dominating Set problem is to find a set in "small side"
to dominate "large side".
Dom et al. [7] proved that Red-Blue Dominating Set is incompressible using the
Colors and IDs technique. This result is applicable to several other incompressible problems
such as Steiner Tree, Connected Vertex Cover [7], and Tree Contraction [13].
Theorem 5.1 (Dom et al. [7]). Red-Blue Dominating Set admits no polynomial kernel
unless N P ⊆ coN P/poly.
We point out that One-Sided Dominating Set can be solved in O(2XIO(1)) time
by enumerating all subsets of X, and is incompressible. The proof will appear in full version
of the paper [5] by the same authors.
Theorem 5.2. One-Sided Dominating Set admits no polynomial kernel unless N P ⊆
coN P/poly.
We now give a polynomial parameter transformation from Red-Blue Dominating Set
to Split Contraction. Without loss of generality, we may assume that every vertex in
X has at least one neighbor in Y and then t ≤ X. Our result is inspired by the reduction
for Tree Contraction by Heggernes et al. [13].
Theorem 5.3. Split Contraction admits no polynomial kernel unless N P ⊆ coN P/poly.
Proof. Given a bipartite graph G = (X, Y ; E) and a positive integers t, we construct a graph
G′ from G by creating a clique C of size X + t + 3 where a designated vertex u ∈ C is made
adjacent to all vertices of Y , and for every v ∈ X adding X + t + 1 new leaves appending
to v. We claim that Y has a subset of at most t vertices that dominates X iff G′ can be
made into a split graph by contracting at most B + t edges.
Suppose that Y ′ is a t-subset of Y that dominates X. Since vertices of {u} ∪ Y ′ ∪ X
induce a connected graph, we can make these X + t + 1 vertices into a single vertex by
10
using X + t edge contractions. The subgraph of G′ induced by {u} ∪ Y ∪ X is contracted
to a star, and G′ is modified into a split graph after contractions.
Conversely, suppose that G′ contains at most X + t edges F whose contraction results
in a split graph. Note that at least two vertices a and b other than u in C survive after
contracting F , and for each vertex x ∈ X there exists a leaf x′ appending to x that is
not involved in any edge contraction. If any x ∈ X and u are in the different witness set,
then {x, x′, a, b} form an induced 2K2, contradicting to the fact that G′/F is a split graph.
Therefore all vertices in X ∪{u} are in the same witness set. Observe that each path starting
from one vertex in X to u must go through some vertices in Y , which implies that X is
dominated by a subset I of Y containing exactly those vertices that are in the same witness
set with u. Thus we obtain a solution of G with I ≤ F − X ≤ t vertices.
In order to show the incompressibility of Threshold Contraction in a clear sketch, we
first give a polynomial parameter transformation from One-Sided Dominating Set to the
following One-Sided Domatic Number problem, and next give another transformation
from One-Sided Domatic Number to Threshold Contraction on split graphs.
One-Sided Domatic Number
Instance:
Question:
Bipartite graph G = (X, Y ; E) and integer t.
Can we partition X into t disjoint sets such that vertices in each
set dominate Y ?
X.
Parameter:
One-Sided Domatic Number can be solved in O(2XlogXIO(1)) time by enumerat-
ing all partitions for X, and is incompressible.
Theorem 5.4. One-Sided Domatic Number admits no polynomial kernel unless N P ⊆
coN P/poly.
Proof. Given a bipartite graph G = (X, Y ; E) and an integer t, we construct a graph G′ by
creating a new vertex w that is made adjacent to every vertex of X, and a set Z of X − t
vertices that are made adjacent to every vertex of Y .
X
Y
Z
w
Figure 3: An example of construction of G′ with X = 4, t = 2.
We claim that X has a subset of at most t vertices that dominates Y iff X ∪ Z can be
partitioned into X − t + 1 disjoint sets such that vertices in each set dominate Y ∪ {w}.
Suppose that X has a subset S of t vertices that dominates Y . Obviously S dominates
Y ∪ {w}. Note that every vertex in X \ S is adjacent to w and every vertex in Z dominates
Y . Thus vertices in (X \ S) ∪ Z can be partitioned into X − t pairs such that each pair of
11
vertices dominates Y ∪ {w}. Included the set S, we obtain X − t + 1 disjoint dominating
sets for Y ∪ {w} in X ∪ Z.
Conversely, suppose that X ∪Z has a disjoint partition (S1, · · · , SX−t+1) such that each
Si dominates Y ∪ {w}. Note that each Si that contains some vertex of Z must contain at
least one vertex in X because w is only adjacent to vertices in X. Let r be the number of
sets Si that intersect Z where r ≤ Z = X − t. It is easy to see that there are at most
X − r vertices in X that constitute X − t + 1 − r disjoint dominating sets for Y ∪ {w},
X−t+1−r ≤ t
and thus there exists one dominating set containing at most
vertices which is clearly a dominating set for Y .
X−t+1−r = 1 +
X−r
t−1
Since X ∪ Z = 2X − t, the reduction is parameter preserving.
Theorem 5.5. Threshold Contraction on split graphs admits no polynomial kernel
unless N P ⊆ coN P/poly.
Proof. Given a bipartite graph G = (X, Y ; E) where X = {x1, · · · , xk} and Y = {y1, · · · , yn},
and a positive integer t, we construct a graph G′ as following:
• Create a clique K of X vertices {u1, · · · , uk}, and a clique A of size 2X + 1. Make
K ∪ A into a large clique.
• Create an independent set B of X + 1 vertices that are made adjacent to every vertex
1 , · · · , v(l)
of K, and create X + 1 disjoint independent sets of size Y : Il = {v(l)
n }
(l = 1, · · · , X + 1). All vertices in I1 ∪ · · · ∪ IX+1 are made adjacent to every vertex
of A.
• For each 1 ≤ l ≤ X + 1, make ui adjacent to v(l)
j
iff xiyj ∈ E.
• G′ = (K ∪ A; B ∪ I1 ∪ · · · ∪ IX+1) is a split graph.
X
Y
K
A
B
· · ·
I1
Ik+1
Figure 4: An example of the reduction from Red-Blue Domatic Number to Threshold
Contraction.
We claim that X can be partitioned into t disjoint dominating sets for Y iff G′ can be
modified into a threshold graph by contracting at most X − t edges.
Suppose that the set B has a disjoint partition X = X1 ∪· · · ∪Xt such that each Xi dom-
inates Y . We partition the set {u1, · · · , uk} into t disjoint sets {R1, · · · , Rt} corresponding
to {X1, · · · , Xt}, and contract vertices in each Ri to a single vertex ri. The total number
of edge contractions we use is Σi(Ri − 1) = (ΣiRi) − t = X − t, and ri is adjacent to
12
every vertex of I1 ∪ · · · ∪ IX+1 for each 1 ≤ i ≤ t because Xi dominates Y in G. The closed
neighbor sets of vertices in K ∪A satisfy that N [r1] = · · · = N [rt] ⊇ N [A] after contractions,
thus G′ is made into a threshold graph.
Conversely, suppose that G′ contains at most X − t edges F whose contraction results
in a threshold graph. K is partitioned into r witness sets W1, · · · , Wr, and vertices in
each Wj are contracted to a single vertex wj . Note that Σj(Wj − 1) ≤ F , we have
r ≥ ΣjWj − F ≥ t. Since F ≤ X, there exists some 1 ≤ p ≤ X + 1 such that none
vertex in Ip is incident with any edge in F . Moreover, there exist vertices a ∈ A and b ∈ B
that are not incident with any edge in F . By the construction, each vertex in K is adjacent
to b which is non-adjacent to a, and a is adjacent to all vertices in Ip. Thus to modify
G′ into a threshold graph, each vertex in {w1, · · · , wr} must be adjacent to all vertices in
Ip = {vp,1, · · · , vp,n} after contractions, implying that Wj dominates {yi, · · · , yn} in G for
each 1 ≤ j ≤ r. Therefore (G, t) is a yes-instance of One-Sided Domatic Number.
The above reduction implies the NP-hardness of Threshold Contraction.
Theorem 5.6. Threshold Contraction is NP-complete even for split graphs.
6 Concluding Remarks
In this paper we have proved that the following two edge contraction problems are FPT
but admit no polynomial kernels unless N P ⊆ coN P/poly: Split Contraction, and
Threshold Contraction on split graphs. We note that Threshold Contraction
on split graphs is essentially the same as Cograph Contraction on split graphs. The
Cograph Contraction problem is claimed to be FPT for general graphs by a recent paper
(Lokshtanov et al. [16]) using the notion of rankwidth. We have provided a conceptually
simpler and faster algorithm for the problem when the input is a split graph. Our result in
Theorem 5.5 also implies that Cograph Contraction admits no polynomial kernel unless
N P ⊆ coN P/poly.
Although we have proved that contracting general graphs to split graphs and contracting
split graphs to threshold graphs are both F P T , our results do not imply the fixed-parameter
tractability for Threshold Contraction on general graphs since our algorithm for Split
Contraction does not find all split graphs obtained by contractions.
It remains open
whether Threshold Contraction is FPT for general graphs.
References
[1] Belmonte, R., Heggernes, P., van't Hof, P.: Edge contractions in subclasses of chordal
graphs. Discrete Applied Mathematics 160, 999 -- 1010 (2012)
[2] Bodlaender, H.L., Thomass´e, S., Yeo, A.: Kernel bounds for disjoint cycles and disjoint
paths. In: ESA 2009. LNCS, vol. 5757, pp. 635 -- 646. Springer, Heidelberg (2009)
[3] Bodlaender, H.L., Downey, R.G., Fellows, M.R., Hermelin, D.: On problems without
polynomial kernels (extended abstract). In: ICALP 2008, Part I. LNCS, vol. 5125, pp.
563 -- 574. Springer, Heidelberg (2008)
[4] Cai, L.: Fixed-parameter tractability of graph modification problems for hereditary
properties. Information Processing Letters 58, 171 -- 176 (1996)
13
[5] Cai, L., Guo, C.: Contracting few edges to remove forbidden induced subgraphs.
Manuscript. (extended abstract) In: IPEC 2013. To appear.
[6] Chen, J., Meng, J.: A 2k kernel for the cluster editing problem. In: COCOON 2010.
LNCS, vol. 6196, pp. 459 -- 468. Springer, Heidelberg (2010)
[7] Dom, M., Lokshtanov, D., Saurabh, S.: Incompressibility through colors and IDs. In:
ICALP 2009. LNCS, vol. 5555, pp. 378 -- 389. Springer, Heidelberg (2009)
[8] Feige, U., Halldorsson, M.M., Kortsarz, G., Srinivasan, A.: Approximating the Domatic
Number. SIAM Journal on computing 32(1), 172 -- 195 (2002)
[9] Ghosh, E., Kolay, S., Kumar, M., Misra, P., Panolan, F., Rai, A., Ramanujan, M.:
Faster parameterized algorithms for deletion to split graphs. In: SWAT 2012. LNCS,
vol. 7357, pp. 107 -- 118. Springer, Heidelberg (2012)
[10] Golovach, P.A., Kami´nski, M., Paulusma, D., Thilikos, D.M.: Increasing the minimum
degree of a graph by contractions. Theoretical Computer Science 481, 74 -- 84 (2013)
[11] Golovach, P.A., van't Hof, P., Paulusma, D.: Obtaining planarity by contracting few
edges. Theoretical Computer Science 476, 38 -- 46 (2013)
[12] Guo, J.: Problem kernels for NP-complete edge deletion problems: split and related
graphs. In: ISAAC 2007. LNCS, vol. 4835, pp. 915 -- 926. Springer, Heidelberg (2007)
[13] Heggernes, P., van't Hof, P., L´eveque, B., Lokshtanov, D., Paul, C.: Contracting graphs
to paths and trees. In: IPEC 2011. LNCS, vol. 7112, pp. 55 -- 66. Springer, Heidelberg
(2011)
[14] Heggernes, P., van't Hof, P., Lokshtanov, D., Paul, C.: Obtaining a bipartite graph
In: FSTTCS 2011. LIPIcs, vol. 13, pp. 217 -- 228. Schloss
by contracting few edges.
Dagstuhl, Leibniz-Zentrum fur Informatik (2011)
[15] Lin, G.-H., Kearney, P.E., Jiang, T.: Phylogenetic k-root and Steiner k-root.
In:
ISAAC 2000. LNCS, vol. 1969, pp. 539 -- 551. Springer, Heidelberg (2000)
[16] Lokshtanov, D., Misra, N., Saurabh, S.: On the hardness of eliminating small induced
subgraphs by contracting edges. In: IPEC 2013. To appear.
14
|
1511.07412 | 1 | 1511 | 2015-11-23T20:48:29 | Approximation Algorithms for Route Planning with Nonlinear Objectives | [
"cs.DS"
] | We consider optimal route planning when the objective function is a general nonlinear and non-monotonic function. Such an objective models user behavior more accurately, for example, when a user is risk-averse, or the utility function needs to capture a penalty for early arrival. It is known that as nonlinearity arises, the problem becomes NP-hard and little is known about computing optimal solutions when in addition there is no monotonicity guarantee. We show that an approximately optimal non-simple path can be efficiently computed under some natural constraints. In particular, we provide a fully polynomial approximation scheme under hop constraints. Our approximation algorithm can extend to run in pseudo-polynomial time under a more general linear constraint that sometimes is useful. As a by-product, we show that our algorithm can be applied to the problem of finding a path that is most likely to be on time for a given deadline. | cs.DS | cs | Approximation Algorithms for Route Planning with Nonlinear Objectives
Ger Yang
Electrical and Computer Engineering
University of Texas at Austin
[email protected]
Evdokia Nikolova
Electrical and Computer Engineering
University of Texas at Austin
[email protected]
5
1
0
2
v
o
N
3
2
]
S
D
.
s
c
[
1
v
2
1
4
7
0
.
1
1
5
1
:
v
i
X
r
a
Abstract
We consider optimal route planning when the objective func-
tion is a general nonlinear and non-monotonic function. Such
an objective models user behavior more accurately, for exam-
ple, when a user is risk-averse, or the utility function needs
to capture a penalty for early arrival. It is known that as
nonlinearity arises, the problem becomes NP-hard and little
is known about computing optimal solutions when in addi-
tion there is no monotonicity guarantee. We show that an ap-
proximately optimal non-simple path can be efficiently com-
puted under some natural constraints. In particular, we pro-
vide a fully polynomial approximation scheme under hop
constraints. Our approximation algorithm can extend to run
in pseudo-polynomial time under a more general linear con-
straint that sometimes is useful. As a by-product, we show
that our algorithm can be applied to the problem of finding a
path that is most likely to be on time for a given deadline.
Introduction
In this paper, we present approximation algorithms for route
planning with a nonlinear objective function. Tradition-
ally, route planning problems are modeled as linear shortest
path problems and there are numerous algorithms, such as
Bellman-Ford or Dijkstra's algorithm, that are able to find
the optimal paths efficiently. However, nonlinearity arises
when we would like to make more complicated decision-
making rules. For example, in road networks, the objective
might be a nonlinear function that represents the trade-off
between traveling time and cost. Unfortunately, dynamic
programming techniques used in linear shortest path algo-
rithms no longer apply as when nonlinearity appears, sub-
paths of optimal paths may no longer be optimal.
Applications include risk-averse routing or
routing
with a nonlinear objective when we are uncertain about
the traffic conditions (Nikolova, Brand, and Karger 2006;
Nikolova et al. 2006). A risk-averse user tends to choose a
route based on both the speediness and reliability. Such risk-
averse attitude can often be captured by a nonlinear ob-
jective, e.g. a mean-risk objective (Nikolova 2010). On the
other hand, the user might seek a path that maximizes the
probability for him to be on time given a deadline. This ob-
jective was considered by Nikolova (2010) to capture the
risk-averse behavior. For Gaussian travel times, the latter
paper only considered the case where at least one path has
if
literature
mean travel time less than the deadline, and consequently the
objective function can be modeled as a monotonic function.
The problem of finding a path that minimizes a
monotonic increasing objective has been studied in
(Goyal and Ravi 2013; Nikolova 2010;
the
Tsaggouris and Zaroliagis 2009). What
there is no
monotonicity guarantee? For example, consider the dead-
line problem when we know that all paths have mean
travel
time longer than the deadline, or consider the
well-known cost-to-time ratio objective (Megiddo 1979;
Gu´erin and Orda 1999) in various combinatorial optimiza-
tion problems. Without
the monotonicity assumption,
the problem usually becomes very hard. It has been
shown (Nikolova, Brand, and Karger 2006)
a
general objective function, finding the optimal simple
path is NP-hard. This is because such a problem often
involves finding the longest path for general graphs,
which is known to be a strongly NP-hard problem
(Karger, Motwani, and Ramkumar 1997), namely finding a
constant-factor approximation is also NP-hard. This sug-
gests that for a very general class of functions, including the
cost-to-time ratio objective (Ahuja, Batra, and Gupta 1983),
not only is the problem itself NP-hard but so is also its
approximation counterpart.
that
for
Therefore, in this paper, we focus on the problem that ac-
cepts non-simple paths, in which it is allowed to visit nodes
more than once. To make this problem well-defined, we con-
sider the problem under either a hop constraint or an addi-
tional linear constraint, where we say a path is of γ hops if it
contains γ edges. We design a fully polynomial approxima-
tion scheme under hop constraints, and show that the algo-
rithm can extend to run in pseudo-polynomial time under an
additional linear constraint. Further, we show how our algo-
rithm can be applied to the cost-to-time ratio objective and
the deadline problem.
Related Work
The route planning problem with nonlinear objective we
consider here is related to multi-criteria optimization (e.g.,
Papadimitriou and Yannakakis 2000). The solutions to these
problems typically look for approximate Pareto sets, namely
a short list of feasible solutions that provide a good approx-
imation to any non-dominated or optimal solution. A fully
polynomial approximation scheme (FPTAS) for route plan-
ning with general monotonic and smoothed objective func-
tions was presented by Tsaggouris and Zaroliagis (2009),
based on finding an approximate Pareto set for the feasi-
ble paths. Later, Mittal and Schulz (2013) provided a general
approach for monotonic and smoothed combinatorial opti-
mization problems and showed how to reduce these prob-
lems to multi-criteria optimization.
A special class of nonlinear optimization is quasi-
concave minimization. This class of problems has the
property that
the optimal solution is an extreme point
of the feasible set (Horst, Pardalos, and Van Thoai 2000).
For quasi-concave combinatorial minimization problems,
Kelner and Nikolova (2007) proved (log n)-hardness of ap-
proximation and they also gave an additive approximation
algorithm for low-rank quasi-concave functions based on
smoothed analysis. Later, Nikolova (2010) presented an FP-
TAS for risk-averse quasi-concave objective functions. In re-
cent work, Goyal and Ravi (2013) gave an FPTAS for gen-
eral monotone and smooth quasi-concave functions.
Without assuming monotonicity, the problem was studied
by Nikolova, Brand, and Karger (2006), who proved hard-
ness results and gave a pseudo-polynomial approximation
algorithm with integer edge weights for some specific ob-
jective functions.
Problem Statement and Preliminaries
Consider a directed graph G = (V, E). Denote the number
of vertices as V = n, and the number of edges as E = m.
Suppose we are given a source node s and a destination node
t. In the nonlinear objective shortest path problem, we seek
an s − t path p, possibly containing loops, that minimizes
some nonlinear objective function f (p).
More precisely, denote the feasible set of all s − t paths
by Pt, and let P = Sv∈V Pv be the set of paths that start
from s. Here we have not yet imposed assumptions that P
contains only simple paths. Let f : P → R+ be a given
objective function. Then we can write the problem as finding
the path p∗ that minimizes this objective function:
p∗ = arg min
p∈Pt
f (p)
(1)
We assume that f is a low-rank nonlinear function with
some smoothness condition. Specifically, we say that f is
of rank d if there exists a function g : Rd
+ → R+ and
an m × d weight matrix W = (we,k)m×d, with pos-
itive entries, such that for any path p ∈ P, f (p) =
g(Pe∈p we,1,Pe∈p we,2, . . . ,Pe∈p we,d).
We interpret this as having d criteria, such as for example
length in miles, cost, travel time, etc. so that each edge is
associated with d additive weights, where the k-th weight
for edge e is denoted by we,k. Then, the k-th weights of
all edges can be represented by the column vector wk =
(w1,k, w2,k, . . . , wm,k)T . For simplicity, we denote the sum
of the k-th weights of the edges along path p as lk(p) =
Pe∈p we,k, and we call it the k-th criterion of path p. Hence,
the objective function f can be rewritten as a function of all
criteria:
f (p) = g(l1(p), l2(p), . . . , ld(p))
(2)
Figure 1: Illustration of a path polytope on a doubly
weighted graph with a hop constraint. Each blue cross and
red dot correspond to an s − t path projected onto W. We
can use polynomially many red dots to approximate the path
polytope and consequently to approximate the optimal path.
Consider the linear space spanned by the d criteria W =
span(w1, w2, . . . , wd). Then, each feasible s − t path p ∈
Pt has an associated point in this d-dimensional space W
given by l(p) = (l1(p), l2(p), . . . , ld(p)), providing the val-
ues of the d criteria for that path. We shall call the con-
vex hull of such points the path polytope.1 We assume that
the weights matrix W contains only positive entries, and
thus the path polytope is bounded from below: for any path
p ∈ Pt and criterion k ∈ {1, . . . , d}, we have lk(p) ≥
minp′∈Pt lk(p′). This lower-bound corresponds to the short-
est path with respect to the k-th criterion. For upper bounds
of the polytope, we defer the discussion to the later sections.
The size of the path polytope is in worst case exponential
in the problem size. The main idea of this paper is to ap-
proximate the path polytope with polynomially many paths,
as illustrated in Figure 1. As we will show in this paper, the
optimal path can be approximated if there are some smooth-
ness conditions on the objective function. The smoothness
assumption we adapt here is as follows: g is assumed to be β-
Lipschitz on the log-log scale with respect to the L1-norm:
that is, for any y, y′ in the path polytope,
log g(y) − log g(y′) ≤ βk log y − log y′k1
(3)
where log y is defined as a column vector:
(log y1, log y2, . . . , log yd)T .
log y =
In this paper, we design approximation algorithms for this
problem, formally defined as follows:
Definition 1. An α-approximation algorithm for a mini-
mization problem with optimal solution value OP T and
α > 1 is a polynomial-time algorithm that returns a solu-
tion of value at most αOP T .
1Strictly speaking, it is the projection of the feasible path poly-
tope in Rm onto this d-dimensional subspace but for simplicity we
will just call it the path polytope in this paper.
The best possible approximation algorithms provide solu-
tions that are arbitrarily close to the optimal solution:
Definition 2. A fully-polynomial approximation scheme
(FPTAS) is an algorithm for an optimization problem that,
given desired accuracy ǫ > 0, finds in time polynomial in
1
ǫ and the input size, a solution of value OP T ′ that satisfies
OP T − OP T ′ ≤ ǫOP T , for any input, where OP T is the
optimal solution value.
Hardness Results
We can categorize the problem on general directed graphs
into two types: one is the problem that accepts only sim-
ple paths, by which we mean each vertex can only be vis-
ited once in any feasible path; the other is the problem
that accepts non-simple paths. Unfortunately, both types
of problems turn out to be hard. In fact, the problem that
only accepts simple paths in general involves finding the
longest simple path, which is strongly NP-hard, namely it
is NP-hard to approximate to within any constant factor
(Karger, Motwani, and Ramkumar 1997).
First, we present a natural objective function that has been
considered in the literature, that is hard to approximate.
Minimum Cost-to-Time Ratio Path Problem
The cost-to-time ratio objective function is a rank 2 function,
defined as follows:
f (p) = g(cid:18)Xe∈p
we,1,Xe∈p
we,2(cid:19) = Pe∈p we,1
Pe∈p we,2
strictly
has
function
different
previously
optimization
simplicity, we
entries. This
assume W has
objective
for
problems
For
pos-
itive
been
considered
combina-
torial
(Megiddo 1979;
Correa, Fernandes, and Wakabayashi 2010). However, find-
ing the minimum cost-to-time ratio simple path is known
to be an NP-hard problem (Ahuja, Batra, and Gupta 1983).
Here, we show its hardness of approximation:
Theorem 1. There is no polynomial-time α-approximation
algorithm for the minimum cost-to-time ratio simple path
problem for any constant factor α, unless P = N P .
Proof. To facilitate the presentation of our hardness
of approximation proof, we first show that
this prob-
lem is NP-hard. We reduce a longest path problem
instance to a minimum cost-to-time ratio simple path
problem instance. The longest path problem is known
to be NP-hard to approximate to within any constant
factor (Karger, Motwani, and Ramkumar 1997). First, we
choose some λ > 0. For an instance of the longest path
problem, which is some graph G with unweighted edges and
a designated source node s and sink node t, we create an in-
stance of the minimum cost-to-time ratio path problem with
the same graph G, source s and sink t and the following
(cost,time) edge weights:
1. For each edge e incident to s, set we = (λn + 1, 1).
2. For the remaining edges e, set we = (1, 1).
We can see that a path is optimal under the ratio objective
if and only if it is the longest path with respect to the un-
weighted edges (or, equivalently, with respect to the sec-
ond weight). This shows that the minimum cost-to-time ratio
simple path problem is NP-hard.
To show the hardness of approximation, similarly, we
use the fact that the longest path problem cannot be ap-
proximated to within any constant factor unless P = N P
(Karger, Motwani, and Ramkumar 1997). Suppose the con-
trary, namely that we can find a (1 + ǫ)-approximation algo-
rithm for the minimum cost-to-time ratio simple path prob-
lem, for some ǫ ∈ (0, λ). To be precise, since we can choose
λ while ǫ comes from the assumed approximation algorithm,
we can set λ > ǫ. For a given instance of longest path, con-
sider the corresponding instance of minimum cost-to-time
ratio path problem above. Let p∗ be the optimal min cost-to-
time ratio path. Then, the (1 + ǫ)-approximation algorithm
will return a simple path p such that
λn + l2(p)
l2(p)
≤ (1 + ǫ)
λn + l2(p∗)
l2(p∗)
Rearranging this, we get the inequality:
(λn − ǫl2(p))l2(p∗) ≤ λ(1 + ǫ)nl2(p)
Since l2(p) ≤ n and ǫ < λ, we get the following:
l2(p∗) ≤
λ(1 + ǫ)
λ − ǫ
l2(p)
This shows that we can approximate the longest path prob-
lem to within a constant factor of λ(1+ǫ)
λ−ǫ , a contradiction.
Therefore, the minimum cost-to-time ratio simple path prob-
lem cannot be approximated to within any constant factor
unless P = N P .
General Objective Function
For general objective functions, the following hardness re-
sults were proven by Nikolova, Brand, and Karger (2006):
Fact 1. Suppose function g attains a global minimum at y >
0. Then problem (1) of finding an optimal s − t simple path
is NP-hard.
Fact 2. Consider problem (1) of finding an optimal s − t
simple path with f being a rank-1 function. Let p∗ be the
optimal path, and l∗ = l(p∗). If g is any strictly decreasing
and positive function with slope of absolute value at least
λ > 0 on [0, l∗], then there does not exist a polynomial-
time constant factor approximation algorithm for finding an
optimal simple path, unless P = N P .
Although these facts are proven for rank-1 functions, they
show that for a large class of objective functions, the prob-
lem of finding the optimal simple path is not only NP-hard,
but also NP-hard to approximate within any constant factor.
As a result, we turn our attention to the problem that
accepts non-simple paths. While this problem is also NP-
hard (Nikolova, Brand, and Karger 2006), we are able to de-
sign a polynomial-time algorithm to approximate the opti-
mal solution arbitrarily well.
Approximation Algorithms
Consider problem (1) that accepts non-simple paths. For
some cases, it is possible that there does not exist an opti-
mal path of finite length. Consider an example of a rank 1
monotonic decreasing function g(x) = 1/x. Suppose there
is a positive weight loop in any of the s − t paths. Then,
we can always find a better solution by visiting such a loop
arbitrarily many times: the optimal path here is ill-defined.
Naturally, for the problem that accepts non-simple paths,
the path polytope is infinite. Even when the optimal path is
finite, for a large class of problems, including the minimum
cost-to-time ratio path problem, finding a finite subset of the
path polytope that contains the optimal path is difficult, as
shown in our Hardness section above. In this section, we
will show that, under some additional constraints, this prob-
lem becomes well-defined and well-approximable. In what
follows, we identify the additional constraints and present
the algorithms to solve the corresponding problems.
Hop Constraint
Suppose the path contains at most γ hops, then the set of
feasible paths Pt becomes finite and the problem is well-
defined. Denote hop(p) to be the number of hops in path p,
or equivalently, the number of edges in p. Note that an edge
can be counted multiple times if path p contains loops.
k
k
k
k
/cmin
, γcmax
, γcmax
k=1[cmin
], for each criterion k ∈ {1, . . . , d}.
To approximate the optimal path, we approximate the path
polytope with polynomially many paths, and then search for
the best one among them. To approximate the path poly-
tope, we first identify a region of weights space W that con-
tains all paths starting from s, which are within γ edges.
We can upper- and lower-bound this region as follows: For
each criterion k = 1, . . . , d, define cmin
k = mine∈E we,k
k = maxe∈E we,k. Then for any path p that starts
and cmax
from the source s and contains at most γ edges, we have
lk(p) ∈ [cmin
Next, we partition the region Qd
] of the
weights space W into polynomially many hypercubes.
Specifically, for each criterion k ∈ {1, 2, . . . , d}, let ǫ >
0 be a given desired approximation factor, and Ck =
. Define Γk = {0, 1, . . . ,⌊log(1+ǫ)(γCk)⌋} as the
cmax
k
set of points that evenly partitions the segment [cmin
, γcmax
]
on the log scale. Then, we generate a d-dimensional lattice
Γ = Γ1×Γ2×···×Γd according to these sets. After that, we
define a hash function h = (h1, h2, . . . , hd) : P → Γ that
maps a path to a point on the lattice Γ. Each hk : P → Γk is
defined as hk(p) = ⌊log(1+ǫ)
The main part of the algorithm iterates over i ∈
{1, 2, . . . , γ}, sequentially constructing s− v sub-paths con-
taining at most i edges, for each vertex v ∈ V . For each
iteration i and each vertex v ∈ V , we maintain at most
Γ configurations as a d-dimensional table Πi
v, indexed by
v(y) can be either an s − v
y ∈ Γ. Each configuration Πi
path p of at most i edges and h(p) = y, given there is one,
or null otherwise. The table can be implemented in various
ways. 2 Here we assume that given index y ∈ Γ, both check-
2In practice, we can either allocate the memory of the entire
table in the beginning, or use another data structure that only stores
lk(p)
cmin
k
⌋.
k
k
k
v(y), and inserting/updating a path in Πi
v(y) is null, retrieving the
v(y) can
ing whether the configuration Πi
path in Πi
be done in TΠ = TΠ(n,Γ) time.
At iteration i, for each vertex u, we will construct a set
of s − u sub-paths containing i edges based on (i − 1)-edge
sub-paths we previously constructed. We loop over each ver-
tex v ∈ V and each edge e = (v, u) incident to v, and we
append the edge to the end of each (i− 1)-edge s− v path in
. However, not all generated sub-paths will be stored.
Πi−1
Consider any s − u path p generated with this procedure, it
will be stored in table Πi
u(h(p)) is empty. With
this technique, we can ensure that the number of sub-paths
we record in each iteration is at most polynomial, because it
is bounded by the size of the lattice Γ times the number of
vertices. This procedure is outlined in Algorithm 1.
u only if Πi
v
Set Πi
v(y) = null
s = {s}.
Algorithm 1 Nonlinear Objective Shortest Path with Hop-
constraint
1: for each y ∈ Γ, v ∈ V , and i = 1, 2, . . . , γ do
2:
3: Set Π0
4: for i = 1 to γ do
v = Πi−1
5:
6:
7:
8:
9:
Set Πi
for each vertex v ∈ V do
for each path p ∈ Πi−1
for each edge e = (v, u) ∈ E do
v with i − 1 edges do
for each vertex v ∈ V .
v
Set p′ be the path that appends edge e to the
tail of p.
Compute hash y = h(p′).
if Πi
Πi
u(y) = null then
u(y) = p′.
10:
11:
12:
13: return maxp∈Πγ
t
f (p)
To prove the correctness of this algorithm, we show the
following lemmas:
Lemma 1. Let pu and p′u be two s − u paths such that for
each criterion k ∈ {1, 2, . . . , d},
log lk(pu) − log lk(p′u) ≤ log(1 + ǫ).
(4)
Suppose there is a vertex v and an edge e such that e =
(u, v) ∈ E. Let pv and p′v be the s − v paths that append
(u, v) to the tails of pu and p′u, respectively. Then, for each
criterion k ∈ {1, 2, . . . , d},
log lk(pv) − log lk(p′v) ≤ log(1 + ǫ)
Proof. By (4), it directly follows that
lk(pv) = lk(pu) + lk(e) ≤ (1 + ǫ)lk(p′u) + lk(e)
≤ (1 + ǫ)(lk(p′u) + lk(e)) = (1 + ǫ)lk(p′v)
lk(p′v) = lk(p′u) + lk(e) ≤ (1 + ǫ)lk(pu) + lk(e)
≤ (1 + ǫ)(lk(pu) + lk(e)) = (1 + ǫ)lk(pv)
Combining the above two inequalities yields the proof.
the used entries of the table. According to our experimental results,
only a small portion of the table Πi
v will be used.
Lemma 2. For any v ∈ V , and for any path s − v that is of
at most i edges, there exists an s − v path p′ ∈ Πi
v such that
for each criterion k ∈ {1, 2, . . . , d},
log lk(p) − log lk(p′) ≤ i log(1 + ǫ)
Proof. We prove this lemma by induction on the number of
edges in a path. The base case is a direct result from the def-
inition of the hash function h. For the induction step, con-
sider vertex v ∈ V , and we would like to prove the statement
is correct for any s − v sub-paths of at most i∗ edges, given
the statement is true for any i < i∗ and any vertices. Let pv
be any s − v path no more than i∗ edges. If pv is of i′ < i∗
edges, then by the induction hypothesis, there exists an s− v
path p′v in Πi′
v such that log(lk(pv)/lk(p′v)) ≤ i′ log(1+ǫ).
The lemma is proved for this case since p′v is in Πi∗
v as well.
Next consider pv to be an s− v path with i∗ edges, and sup-
pose pv is formed by appending the edge (u, v) to an s − u
path pu. By the induction hypothesis, we can see that there
is an s − u path p′u ∈ Πi∗−1
such that for each criterion
k ∈ {1, 2, . . . , d},
u
log lk(pu) − log lk(p′u) ≤ (i∗ − 1) log(1 + ǫ)
Then, with Lemma 1, as we form s − v sub-paths pv and
p′v by appending (u, v) to the end of pu and p′u respectively,
we have the following relationship for each criterion k ∈
{1, 2, . . . , d}:
log lk(pv) − log lk(p′v) ≤ (i∗ − 1) log(1 + ǫ)
(5)
Consider any s − v path p′′v such that h(p′′v ) = h(p′v). We
must have log lk(p′v) − log lk(p′′v ) ≤ log(1 + ǫ) for each
criterion k ∈ {1, . . . , d}, according to the definition of the
hash function h. Combining this with (5), we can see that
even for the case that path p′v is not stored in the table Πi∗
v ,
the path p′′v = Πi∗
v (h(p′v)) must satisfy
log lk(pv) − log lk(p′′v ) ≤ i∗ log(1 + ǫ)
for each criterion k ∈ {1, . . . , d}. This indicates that for any
s − v path pv within i∗ edges, there exists a path p′′v in Πi∗
such that their ratio is bounded by (1 + ǫ)i∗. Q.E.D.
v
Now, we are ready to state and prove the main result:
Theorem 2. Suppose g defined in (2) is β-Lipschitz on
the log-log scale. Then Algorithm 1 is a (1 + ǫ)βdγ-
approximation algorithm for the γ-hop-constrained nonlin-
ear objective shortest path problem. Besides, Algorithm 1
has time complexity O(γm( log(γC)
log(1+ǫ) )dTΠ), where C =
maxk∈{1,2,...,d} Ck.
Proof. By Lemma 2, we can see that for any s − t path p of
at most γ edges, there exists an s − t path p′ ∈ Πγ
t such that
log lk(p) − log lk(p′) ≤ γ log(1 + ǫ),
∀k = 1, 2, . . . , d
It then follows from (3) that our algorithm finds a path p′
such that log f (p′) − log f (p∗) ≤ βdγ log(1 + ǫ). Then,
we have
f (p′) ≤ (1 + ǫ)βdγf (p∗)
This shows that p′ is an (1 + ǫ)βdγ-approximation to p∗.
Next, we show the time complexity. It is easy to see that
at each iteration, the algorithm generates at most mΓ sub-
paths. Each sub-path requires at most O(TΠ) time to oper-
ate. Since Γ = ( log(γC)
log(1+ǫ) )d, we can conclude that Algo-
rithm 1 runs in time O(γm( log(γC)
log(1+ǫ) )dTΠ).
Note that we can obtain a (1+δ)-approximation algorithm
by setting δ = (1 + ǫ)βdγ − 1. With this setting, then Algo-
rithm 1 has time complexity O(γm( βdγ log(γC)
)dTΠ). If we
assume γ = O(n), then Algorithm 1 is indeed an FPTAS.
δ
Rm
Linear Constraint
The hop constraint we considered can be relaxed: we now
consider the more general problem that any returned path p
must satisfy an additional linear constraint Pe∈p we,d+1 ≤
b, for some budget b > 0 and some weight vector wd+1 ∈
++ with strictly positive entries. We define the (d + 1)-th
criterion ld+1(p) = Pe∈p we,d+1, which can be interpreted
as having a budget b for a path. To solve this problem, we can
find an upper bound of the number of edges for any feasible
s − t path. A simple upper bound can be given by
+ hopmin(cid:25)
γ = (cid:24) b −Pe∈p∗d+1
we,d+1
(6)
mine∈E we,d+1
where p∗d+1 = arg minp∈P Pe∈p we,d+1 is the shortest path
with respect to the weight wd+1, and hopmin is the mini-
mum number of edges for any s − t path. Note that either
hopmin and p∗d+1 can be solved by Dijkstra's or Bellman-
Ford's shortest path algorithms efficiently.
Then, we can adapt the techniques discussed for the hop-
constrained problem but with some modifications. Instead of
assigning the least hop path to a configuration, we now as-
sign the path with the least (d+ 1)-th criterion. Formally, the
modification is to change line 11 and line 12 of Algorithm 1
u(y) if p′ satisfies both of the following:
to : Assign p′ to Πi
• ld+1(p′) < b
• Πi
u(y) is null or ld+1(p′) < ld+1(Πi
As a corollary of Theorem 2, we have the following result:
Corollary 1. Algorithm 1 with the modifications stated
above is a (1 + ǫ)βdγ-approximation algorithm for the
linear-constrained nonlinear objective shortest path prob-
lem, where γ is defined as in Eq. (6).
u(y))
Application: Deadline Problem
With the techniques mentioned in the above section, we are
able to solve various problems, including the minimum cost-
to-time ratio path problem with either a hop or a linear bud-
get constraint. In this section, we discuss a related problem,
which is the deadline problem in a stochastic setting.
The deadline problem is to find a path that is the most
likely to be on time, which is defined as follows: Suppose
each edge is associated with an independent random vari-
able Xe, which we call the travel time of edge e. Given
a deadline D > 0, we are going to find an s − t path
such that the probability of being on time is maximized:
Table 1: Deadline Problem with Hop Constraint
Network
Hops
5 × 5 Grid
10 × 10 Grid
Anaheim
12
12
15
15
15
Worst-case
Theoretical
Accuracy (α)
(Exhaustive)
2.2923
2.2926
2.2891
3.0436
Memory
Usage
(106 paths)
1.63
0.67
4.30
Run
time (s)
3.441
2.057
20.276
18.05
83.430
9.60
37.671
maxp∈Pt Pr(Pe∈p Xe < D). We assume that the random
variables {Xe} are Gaussian. Based on this, we denote the
mean travel time for edge e as µe and variance as σ2
e . For
simplicity, we assume for each edge e ∈ E, its mean µe
e are bounded by µe ∈ [µmin, µmax], and
and variance σ2
max]. The problem can hence be written as:
min, σ2
σ2
e ∈ [σ2
max
p∈Pt
Φ(cid:18) D −Pe∈p µe
qPe∈p σ2
e
(cid:19)
(7)
where Φ(·) is the cumulative distribution function for
the standard normal distribution. The objective function
in (7) is a rank 2 function. This problem is studied in
(Nikolova et al. 2006) for a risk-averse shortest path prob-
lem, where they considered the case that the optimal path
has mean travel time less than the deadline: for this case (7)
is a monotonic decreasing function. In this paper, we con-
sider the case that all feasible s − t paths have mean travel
time strictly greater than the deadline. This means that we
assume each path has poor performance. Then, the objec-
tive function in (7) becomes monotonic decreasing with re-
spect to the mean but monotonic increasing with respect to
the variance, for which known algorithms no longer apply.
However, we are able to use techniques presented in this pa-
per to solve this problem under either a hop or an additional
linear constraint.
Theorem 3. Suppose each s − t path p ∈ Pt satisfies
Pe∈p µe > D and Pe∈p σ2
e > S. For the deadline prob-
lem with a hop constraint γ, if Algorithm 1 returns some
path p′ such that f (p′) > Φ(−3) ≈ 0.0013, then p′ is
an α-approximation for the deadline problem, where α =
min{384.62, (1 + ǫ)6.568(3+ D
)γ)}. Similarly, if p′ is such
that f (p′) > Φ(−2) ≈ 0.023, then α = min{21.93, (1 +
ǫ)4.745(2+ D
√S
√S
Note that this theorem does not guarantee Algorithm 1
admits FPTAS for the deadline problem, since given α, the
running time is polynomial in the factor D/√S.
)γ}.
Experimental Evaluations
We tested our algorithm on a grid graph and a real trans-
portation network. The algorithm is implemented in C++.
Experiments are run on an Intel Core i7 1.7 GHz (I7-4650U)
processor. Due to memory constraints, the table {Πi
vv ∈ V }
is implemented as a binary search tree, where query, in-
sertion, and update operations can be done in time TΠ =
Figure 2: Memory usage for running our algorithm on a 5×5
grid graph to the number of iterations.
vv ∈ V }.
O(log(nΓ)). Memory turns out to be the limiting factor for
the problem size and accuracy available to our experiments.
In this section, memory usage refers to the total number of
sub-paths maintained in the table {Πi
We tested the deadline problem with a hop constraint on
a 5 × 5 grid graph, 10 × 10 grid graph, and the Anaheim
network (416 vertices, 914 edges) from the real world data
set (Bar-Gera 2002). The grid graphs are bi-directional, and
the mean and the variance for each edge are randomly gener-
ated from [0.1, 5]. We routed from node (0, 0) to node (4, 4)
on the 5 × 5 grid, and from (1, 1) to (8, 8) on the 10 × 10
grid. The Anaheim dataset provides the mean travel time.
The variance for each edge is randomly generated from 0 to
the mean. The source and destination are randomly chosen.
Table 1 summarizes the results on the datasets, which are
the average of 20 experiments. Even though theoretically the
algorithm can be designed for arbitrary desired accuracy, it is
limited by the memory of the environment.3 However, com-
paring with exhaustive search, we found that the practical
accuracy is 100% on 5 × 5 grid graphs.4 This means that
our algorithm is far more accurate than its theoretical worst
case bound. We explain this phenomenon through Figure 2,
which shows the number of sub-paths generated up to each
iteration. We see that the number of sub-paths generated in
the initial iterations grows exponentially, just as in exhaus-
tive search, but levels off eventually. Recall Lemma 2, which
tells us that an error is induced when a sub-path is eliminated
due to similarity to a saved sub-path, and Figure 2 suggests
that this rarely happens in the initial iterations.
Conclusion
We investigated the shortest path problem with a general
nonlinear and non-monotonic objective function. We proved
3It takes about 3GB physical memory usage for the 18M paths
in the 10 × 10 grid network.
4Exhaustive search becomes prohibitive for paths with more
than 12 hops on 5 × 5 grid graphs.
that this problem is hard to approximate for simple paths and
consequently, we designed and analyzed a fully polynomial
approximation scheme for finding the approximately opti-
mal non-simple path under either a hop constraint or an ad-
ditional linear constraint. We showed that this algorithm can
be applied to the cost-to-time ratio problem and the dead-
line problem. Our experiments show that our algorithm is
capable of finding good approximations efficiently.
Acknowledgement
This work was supported in part by NSF grant numbers
CCF-1216103, CCF-1350823, CCF-1331863, and a Google
Faculty Research Award.
References
[Ahuja, Batra, and Gupta 1983] Ahuja, R.; Batra, J.; and
Gupta, S. 1983. Combinatorial optimization with rational
objective functions: A communication. Mathematics of Op-
erations Research 8(2):314 -- 314.
[Bar-Gera 2002] Bar-Gera, H. 2002. Transportation network
test problems.
[Correa, Fernandes, and Wakabayashi 2010] Correa, J.; Fer-
nandes, C. G.; and Wakabayashi, Y. 2010. Approximat-
ing a class of combinatorial problems with rational objective
function. Mathematical programming 124(1-2):255 -- 269.
[Goyal and Ravi 2013] Goyal, V., and Ravi, R. 2013. An
fptas for minimizing a class of low-rank quasi-concave
functions over a convex set. Operations Research Letters
41(2):191 -- 196.
[Gu´erin and Orda 1999] Gu´erin, R. A., and Orda, A. 1999.
Qos routing in networks with inaccurate information: the-
ory and algorithms. IEEE/ACM Transactions on Networking
(TON) 7(3):350 -- 364.
[Horst, Pardalos, and Van Thoai 2000] Horst, R.; Pardalos,
P. M.; and Van Thoai, N. 2000. Introduction to global opti-
mization. Springer Science & Business Media.
[Karger, Motwani, and Ramkumar 1997] Karger, D.; Mot-
wani, R.; and Ramkumar, G. 1997. On approximating the
longest path in a graph. Algorithmica 18(1):82 -- 98.
[Kelner and Nikolova 2007] Kelner, J. A., and Nikolova, E.
2007. On the hardness and smoothed complexity of quasi-
concave minimization. In Foundations of Computer Science,
2007. FOCS'07. 48th Annual IEEE Symposium on, 472 --
482. IEEE.
[Megiddo 1979] Megiddo, N. 1979. Combinatorial opti-
mization with rational objective functions. Mathematics of
Operations Research 4(4):414 -- 424.
[Mittal and Schulz 2013] Mittal, S., and Schulz, A. S. 2013.
An fptas for optimizing a class of low-rank functions over a
polytope. Mathematical Programming 141(1-2):103 -- 120.
[Nikolova et al. 2006] Nikolova, E.; Kelner, J. A.; Brand, M.;
and Mitzenmacher, M. 2006. Stochastic shortest paths via
quasi-convex maximization. In Algorithms -- ESA 2006, 552 --
563. Springer.
[Nikolova, Brand, and Karger 2006] Nikolova, E.; Brand,
M.; and Karger, D. R. 2006. Optimal route planning un-
der uncertainty. In ICAPS, volume 6, 131 -- 141.
[Nikolova 2010] Nikolova, E. 2010. Approximation algo-
rithms for offline risk-averse combinatorial optimization.
[Papadimitriou and Yannakakis 2000] Papadimitriou, C. H.,
and Yannakakis, M. 2000. On the approximability of trade-
offs and optimal access of web sources. In Foundations of
Computer Science, 2000. Proceedings. 41st Annual Sympo-
sium on, 86 -- 92. IEEE.
[Tsaggouris and Zaroliagis 2009] Tsaggouris, G., and Zaro-
liagis, C. 2009. Multiobjective optimization: Improved fp-
tas for shortest paths and non-linear objectives with applica-
tions. Theory of Computing Systems 45(1):162 -- 186.
Appendix: Missing Proofs
Proof of Corollary 1
We need the following modified version of Lemma 2 to
prove Corollary 1:
Lemma 3. For any v ∈ V , and for any path s − v that is of
at most i edges, there exists an s − v path p′ ∈ Πi
v such that
1
(1 + ǫ)i ≤
lk(p)
lk(p′) ≤ (1 + ǫ)i,
∀k = 1, 2, . . . , d
and
ld+1(p′) ≤ ld+1(p)
Proof. We prove this lemma by induction on the number
of edges in a path as what we have done in Lemma 2. The
base case is a direct result from the definition of the hash
function h. For the induction step, consider vertex v ∈ V ,
and we would like to prove the statement is correct for any
s − v sub-paths of at most i∗ edges, given the statement is
true for any i < i∗ and any vertices.
Let pv be any s − v path of no more than i∗ edges. First
consider the case that the path pv is of i′ < i∗ edges, then
by induction hypothesis, there exists an s − v path p′v in Πi′
such that lk(pv)/lk(p′v) ≤ (1+ǫ)i′. Then consider the s−v
path p′′v ∈ Πi∗
v such that h(p′′v ) = h(p′v). By the definition
of the hash function h, it must hold that
v
1
1 + ǫ ≤
Then we have
lk(p′v)
lk(p′′v ) ≤ 1 + ǫ,
∀k = 1, 2, . . . , d
1
1
lk(pv)
lk(p′′v ) ≤ (1 + ǫ)i′+1,
(1 + ǫ)i′+1 ≤
∀k = 1, 2, . . . , d
By the induction hypothesis and the replacement rule in
the algorithm, we have ld+1(p′′v ) ≤ ld+1(p′v) ≤ ld+1(pv).
Therefore the lemma is proved for this case.
Next consider the case that pv is s− v path with i∗ edges,
and suppose pv is formed by appending the edge (u, v) to an
s − u path pu. By the induction hypothesis, we can see that
there is an s − u path p′u ∈ Πi∗−1
(1 + ǫ)i∗−1 ≤
lk(pu)
lk(p′u) ≤ (1 + ǫ)i∗−1,
∀k = 1, 2, . . . , d
such that
u
and ld+1(p′u) ≤ ld+1(pu). Then, with Lemma 1, as we form
the s− v sub-paths pv and p′v by appending (u, v) to the end
of pu and p′u respectively, we have the following relation-
ship:
1
lk(pv)
lk(p′v) ≤ (1 + ǫ)i∗−1,
(1 + ǫ)i∗−1 ≤
∀k = 1, 2, . . . , d
(8)
Consider any s − v path p′′v such that h(p′′v ) = h(p′v). The
following is always satisfied according to the definition of
the hash function h:
1
1 + ǫ ≤
lk(p′v)
lk(p′′v ) ≤ 1 + ǫ,
∀k = 1, 2, . . . , d
Combining this with (8), we can see that the path p′′v =
Πi∗
v (h(p′v)) must satisfy
lk(pv)
lk(p′′v ) ≤ (1 + ǫ)i∗ ,
(1 + ǫ)i∗ ≤
1
∀k = 1, 2, . . . , d
and ld+1(p′′v ) ≤ ld+1(pv). This indicates that for any s − v
path pv which is within i∗ edges, there exists a path p′′v in Πi∗
such that their ratio is bounded by (1 + ǫ)i∗, and p′′v always
has the smaller (d + 1)-th weight than pv. Q.E.D.
v
With this lemma, we are able to prove Corollary 1:
Proof of Corollary 1. By Lemma 3, we can see that for any
s − t path p that is within γ edges, there exists an s − t path
p′ ∈ Πγ
≤ γ log(1 + ǫ) ∀k = 1, 2, . . . , d
and ld+1(p′) ≤ ld+1(p). It then follows from (3) that our
algorithm finds a path p′ such that
t such that
lk(p)
log
(cid:12)(cid:12)(cid:12)(cid:12)
lk(p′)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
log
Then, we have
f (p′)
f (p∗)(cid:12)(cid:12)(cid:12)(cid:12)
≤ βdγ log(1 + ǫ)
f (p′) ≤ (1 + ǫ)βdγf (p∗)
Clearly p′ satisfies the constraint ld+1(p′) ≤ b since
ld+1(p′) ≤ ld+1(p∗) ≤ b. This shows that p′ is an (1+ǫ)βdγ-
approximation to p∗.
Proof of Theorem 3
Before we prove Theorem 3, we need the following lemma:
Lemma 4. The function Φ( D−x√y ) is 3.284(3 + D√S
)-
Lipschitz on log-log scale with respect to L1 norm on the
region C = {(x, y)x > D, y > S, x − D ≤ 3√y}.
Proof. We want to show that for any (u1, v1), (u2, v2) ∈
C′ = {(u, v)eu > D, ev > S, eu − D ≤ 3√ev},
√ev1 (cid:19) − log Φ(cid:18) D − eu2
√ev2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
log Φ(cid:18) D − eu1
≤ 3.284(cid:18)3 +
√S(cid:19)(u1 − u2 + v1 − v2)
(cid:12)(cid:12)(cid:12)(cid:12)
(9)
D
We take the derivatives from the function log Φ( D−eu
√ev ):
(cid:12)(cid:12)(cid:12)(cid:12)
√ev(cid:19) (10)
D
(11)
(12)
(cid:12)(cid:12)(cid:12)(cid:12)
(13)
√ev (cid:19)
Φ′(cid:18) D−eu
Φ(cid:18) D−eu
√ev (cid:19)
∂
∂u
(cid:12)(cid:12)(cid:12)(cid:12)
log Φ(cid:18) D − eu
√ev (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
∂
∂v
(cid:12)(cid:12)(cid:12)(cid:12)
log Φ(cid:18) D − eu
√ev (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
+
eu
√ev ·
Φ′(cid:18) D−eu
√ev (cid:19)
= (cid:12)(cid:12)(cid:12)(cid:12)
Φ(cid:18) D−eu
√ev (cid:19)
≤ 3.284(cid:18)eu − D
√ev
≤ 3.284(cid:18)3 +
√ev(cid:19)
D
√S(cid:19)
≤ 3.284(cid:18)3 +
= (cid:12)(cid:12)(cid:12)(cid:12)
D − eu
3 ·
2√ev
ev ·
D
≤
3
2 × 3.284
comes
Inequality (10)
where
that
Φ′(x)/Φ(x) ≤ 3.284 for x ∈ [−3, 0), Inequality (11)
that eu − D < 3√ev. Since
comes from the fact
log Φ( D−eu
√ev ) is differentiable in C′, we obtain (9).
from the
fact
Proof of Theorem 3. Let l1(p) denote the mean traveling
time for path p, and l2(p) denote the variance for path p.
Further denote l(p) = (l1(p), l2(p)). By Lemma 2, we can
see that for the optimal s − t path p∗, there exists an s − t
path p′′ ∈ Πγ
t such that
log
(cid:12)(cid:12)(cid:12)(cid:12)
lk(p∗)
lk(p′′)(cid:12)(cid:12)(cid:12)(cid:12)
≤ γ log(1 + ǫ)
Denote the region C = {(x, y)x > D, y > S, x − D ≤
3√y}. We can see that (x, y) ∈ C if and only if Φ( D−x√y ) ∈
[Φ(−3), 0.5) and y > S. The optimal path must satisfy
l(p∗) ∈ C since f (p∗) ≥ f (p′) > Φ(−3).
First consider the case that l(p′′) is in the region C. Then
we can adapt the same technique as in the proof of Theo-
rem 2 and the smoothness condition we showed in Lemma 4
to claim that the path p′ returned by the algorithm satisfies
f (p∗) ≤ (1 + ǫ)2×3.284×(3+ D
≤ (1 + ǫ)2×3.284×(3+ D
√S
√S
)γf (p′′)
)γf (p′)
x0
log
Now consider the case that l(p′′) is not in the region C,
then there must exist a point (x0, y0) ∈ C on the line L =
{x − D = 3√y} such that
l1(p∗)(cid:12)(cid:12)(cid:12)(cid:12)
l2(p∗)(cid:12)(cid:12)(cid:12)(cid:12)
≤ 2 × 3.284 ×(cid:18)3 +
√S(cid:19)γ log(1 + ǫ)
It then follows from Lemma 4 that
≤ γ log(1 + ǫ)
≤ γ log(1 + ǫ)
Φ( D−x0√y0
f (p∗)
log
log
y0
D
)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
Since the algorithm returns some other path p′ such that
f (p′) > Φ(−3) = Φ( D−x0√y0
), we must have
f (p∗) ≤ (1 + ǫ)6.568(3+ D
< (1 + ǫ)6.568(3+ D
√S
√S
)γΦ(−3)
)γf (p′)
However, the approximation ratio is upper-bounded by
0.5/Φ(−3) ≈ 384.62, so this shows that p′
is an
min{384.62, (1 + ǫ)6.568(3+ D
)γ}-approximation to p∗. We
can apply the same technique for the case that f (p′) >
Φ(−2) or any other values.
√S
|
1109.5931 | 1 | 1109 | 2011-09-27T15:19:08 | Online Primal-Dual For Non-linear Optimization with Applications to Speed Scaling | [
"cs.DS"
] | We reinterpret some online greedy algorithms for a class of nonlinear "load-balancing" problems as solving a mathematical program online. For example, we consider the problem of assigning jobs to (unrelated) machines to minimize the sum of the alpha^{th}-powers of the loads plus assignment costs (the online Generalized Assignment Problem); or choosing paths to connect terminal pairs to minimize the alpha^{th}-powers of the edge loads (online routing with speed-scalable routers). We give analyses of these online algorithms using the dual of the primal program as a lower bound for the optimal algorithm, much in the spirit of online primal-dual results for linear problems.
We then observe that a wide class of uni-processor speed scaling problems (with essentially arbitrary scheduling objectives) can be viewed as such load balancing problems with linear assignment costs. This connection gives new algorithms for problems that had resisted solutions using the dominant potential function approaches used in the speed scaling literature, as well as alternate, cleaner proofs for other known results. | cs.DS | cs | Online Primal-Dual For Non-linear Optimization
with Applications to Speed Scaling
Anupam Gupta∗
Ravishankar Krishnaswamy∗
Kirk Pruhs†
Abstract
We reinterpret some online greedy algorithms for a class of nonlinear "load-balancing" problems
as solving a mathematical program online. For example, we consider the problem of assigning jobs to
(unrelated) machines to minimize the sum of the αth-powers of the loads plus assignment costs (the
online Generalized Assignment Problem); or choosing paths to connect terminal pairs to minimize
the αth-powers of the edge loads (i.e., online routing with speed-scalable routers). We give analyses
of these online algorithms using the dual of the primal program as a lower bound for the optimal
algorithm, much in the spirit of online primal-dual results for linear problems.
We then observe that a wide class of uni-processor speed scaling problems (with essentially arbi-
trary scheduling objectives) can be viewed as such load balancing problems with linear assignment
costs. This connection gives new algorithms for problems that had resisted solutions using the dom-
inant potential function approaches used in the speed scaling literature, as well as alternate, cleaner
proofs for other known results.
1
1
0
2
p
e
S
7
2
]
S
D
.
s
c
[
1
v
1
3
9
5
.
9
0
1
1
:
v
i
X
r
a
∗Computer Science Department, Carnegie Mellon University, Pittsburgh, PA 15213, USA. Supported in part by NSF
awards CCF-0964474 and CCF-1016799. R.K. supported in part by an IBM Graduate Fellowship.
†Computer Science Department, University of Pittsburgh, Pittsburgh, PA 15260, USA. [email protected]. Supported
in part by NSF grant CCF-0830558 and an IBM Faculty Award.
1
Introduction
In this paper, we consider two online problems related to load balancing. We call the first problem
Online Generalized Assignment Problem (OnGAP):
Definition of OnGAP: Jobs arrive one by one in an online manner, and the algorithm must fractionally
assign these jobs to one of m machines. When a job j arrives, the online algorithm learns ℓje, the amount
by which the load of machine e would increase for each unit of work of job j that is assigned to machine
e, and cje, the assignment cost incurred for each unit of work of job j that is assigned to machine e.
The goal is to minimize the sum of the αth powers of the machine loads, plus the total assignment cost.
The version of OnGAP without assignment costs was studied by [AAF+97, AAG+95]. Our original
motivation for studying OnGAP is that it models a well-studied class of speed scaling problems with
sum cost scheduling objectives. In these problems, jobs arrive over time and must be scheduled on a
speed scalable processor -- i.e., a processor that can run at any non-negative speed, and uses power sα
when run at speed s. The objective is the sum of the energy used by the processor plus a fractional
scheduling objective that is the sum over jobs of the "scheduling cost" of the individual jobs. These
speed scaling problems are a special case of OnGAP where the machines model the times that jobs can
be scheduled, the assignment cost cje models the scheduling cost for scheduling a unit of job j at time
e. For example, one such scheduling objective is the sum of the fractional flow/response times squared.
For this objective, cje is (e − rj)2 for all times e that are at least the release time rj of job j, and infinite
otherwise. Another example is the problem of minimizing energy usage subject to deadline constraints,
introduced by [YDS95] and considered in followup papers [BKP07, BBCP11, BCPK09]. This problem
can be viewed as a special case of OnGAP, where each job j has an associated deadline dj, and cje is
zero if e ∈ [rj, dj] and infinite otherwise.
The second problem that we consider is a variation/generalization of OnGAP involving online routing
with speed scalable routers to minimize energy, which was previously considered in [AAF+97, AAZ11].
Definition of Online Routing with Speed Scalable Routers Problem: A sequence of requests
arrive one by one over time. Each request j has an associated source-sink pair (sj, tj) in a network of
speed scalable routers, and the online algorithm must route flow between the source-sink pair, with an
objective of minimizing the total energy used by the network, where the energy incurred by an edge e
is the αth power of the load flowing through it.
For load balancing and online routing, it was known that natural online greedy algorithms, which assign
jobs to the machine(s) that minimize the increase in cost, can be shown to be Oα(1)-competitive via an
exchange argument, and directly bounding the cost compared to the optimal cost [AAF+97, AAG+95].
(In fact, basically the same argument shows that the online greedy algorithm is Oα(1)-competitive for
integer assignments.) Once we observe that speed scaling problems with sum scheduling objectives can
be reduced to OnGAP, it is not too difficult to see that the analysis technique in [AAG+95] can be used
to show that natural greedy speed scaling algorithms are Op(1)-competitive.
Our Contribution. In this paper, we first interpret these online problems as solving a mathematical
program online, where the constraints arrive one-by-one, and in response to the arrival of a new con-
straint, the online algorithm has to raise some of the primal variables so that the new constraint will be
satisfied. The online algorithms that we consider raise the primal variables greedily. Our competitive
analysis will use the dual function of the primal program as a lower bound for optimal. For analysis
purposes, we assign a value to the dual variable corresponding to a constraint after the online algorithm
has satisfied that constraint. Our goal is to set the dual variables so that the resulting dual solution
is closely related to the online solution. (In our analyses, the settings of the dual variables naturally
correspond to (some approximation for) the increase in the objective function.) We first show how to
obtain fractional solutions to these problems, and subsequently show how similar ideas can be used for
integer assignments.
1
Our analyses are very much in the spirit of the online primal dual technique for linear programs [BN07].
The main difference is that in the nonlinear setting, the dual is more complicated than in the linear
setting (where the dual is just another linear program). Indeed, in the nonlinear setting, one can not
disentangle the objective and the constraints, since the dual itself contains a version of the primal
objective, and hence copies of the primal variables, within it. Consequently, the arguments for the
dual function in the nonlinear setting have a somewhat different feel to them than in the linear setting.
In particular, we need to set the dual variables λ, and then find minimizing settings for the copies of
the primal variables to get a good lower bound. For the load-balancing and speed scaling problems,
this proceeds relatively naturally. But for the routing problem the dual minimization problem is itself
non-trivial:
in this case we first show how to write a "relaxed/decoupled" dual, which is potentially
weaker than the original dual, but easier to argue about, and then set the variables of this relaxed dual
to achieve a good lower bound. We hope this analytical technique will be useful for other problems.
We then show how a wide class of uni-processor speed-scaling problems (with essentially arbitrary
scheduling objectives) can be viewed as load balancing problems with linear assignment costs. This
connection gives new algorithms for speed-scaling problems that had resisted solutions using the dom-
inant potential function approaches used in the speed scaling literature, as well as alternate, cleaner
analyses for some known results. For speed scaling problems, our analysis using duality is often cleaner
(compare for example, the analysis of OA in [BKP07] to the analysis given here) and more widely
applicable (for example, to nonlinear scheduling objectives) than the potential function-based analy-
ses. Furthermore, we believe that much like the online primal-dual approach for linear problems, the
techniques presented here have potential for wide applicability in the design and analysis of online
algorithms for other non-linear optimization problems.
In Section 1.1 we discuss related work. In Section 2 we consider OnGAP. In Section 3, we
Roadmap:
make some comments about the application of these results to speed scaling problems. In Section 4 we
consider the online routing problem. In Section 5 we show how to alter the water-filling algorithm to
obtain integer assignments with a similar competitive ratio, as well a simple randomized rounding with
a slightly worse performance.
1.1 Related Work
An O(α)-competitive online greedy algorithm for the unrelated machines load-balancing problem in
the Lα-norm was given by [AAF+97, AAG+95]; Caragiannis [Car08] gave better analyses and im-
provements using randomization. An offline O(1)-approximation for this problem was given by [AE05]
and [AKMPS09], via solving the convex program and then rounding the solution in a correlated fashion.
For the routing problem, the O(α)α-algorithm can be inferred from the ideas of [AAF+97, AAG+95].
Followup work in a setting of a network consisting of routers with static and dynamic power components
can be found in [AAZ10, AAZ11].
There are two main lines of speed scaling research that fit within the framework that we consider here.
This first is the problem of minimizing energy used with deadline feasibility constraints. [YDS95] pro-
posed two online algorithms OA and AVR, and showed that AVR is Oα(1)-competitive by reasoning
directly about the optimal schedule.
[BKP07] introduced the use of potential functions for analyzing
online scheduling problems, and showed that OA and another algorithm BKP are Oα(1)-competitive.
[BBCP11] gave a potential function analysis to show that AVR is Oα(1)-competitive. [BCPK09] intro-
duced the algorithm qOA, and gave a potential function analysis to show that it has a better competitive
ratio than OA or AVR for smallish α.
The second main line of speed scaling research is when the scheduling objective is total flow, or more
[PUW08, AF07] gave offline algorithms for unit-weight unit-work jobs.
generally total weighted flow.
All of the work on online algorithms consider some variation of the "natural" algorithm, which uses
the "right" job selection algorithm from the literature on scheduling fixed speed processors, and sets
2
the power of the processor equal to the weight of the outstanding jobs. This speed scaling policy is
"natural" in that it balances the energy and scheduling costs. By reasoning directly about the energy
optimal schedule, [AF07] showed that a batched version of the natural algorithm is Oα(1)-competitive
for unit-work unit-weight jobs. Using a potential function analysis, [BPS09] showed that a variation of
the natural algorithm is Oα(1)-competitive for arbitrary-weight arbitrary-work jobs. For the objective
of total flow plus energy, the bound on the competitive ratio was improved in [LLTW08] by use of
potential function tailored to integer flow instead of fractional flow. Using a potential function analysis,
[BCP09] showed a variation on the natural algorithm is O(1)-competitive for total flow plus energy for
an arbitrary power function, and a variation on the natural algorithm is scalable, for fractional weighted
flow plus energy for an arbitrary power function. [ALW10] improved the bound on the competitive ratio
for total flow plus energy. Nonclairvoyant algorithms are analyzed in [CEL+09, CLL10]. A relatively
recent survey of the algorithmic power management literature in general, and the speed scaling literature
in particular, can be found in [Alb10].
An extensive survey/tutorial on the online primal dual technique for linear problems, as well the history
of the development of this technique, can be found in [BN07].
2 The Online Generalized Assignment Problem
In this section we consider the problem of Online Generalized Assignment Problem (OnGAP). If xje
denotes the extent to which job j is assigned on machine e, then this problem can be expressed by the
following mathematical program:
min Xe (cid:18)Xj
ℓjexje(cid:19)α
+Xe Xj
cjexje
xje ≥ 1
j = 1, . . . , n
The dual function of the primal relaxation is then
subject to Xe
x(cid:23)0(cid:18)Xj
λj +Xe (cid:18)Xj
g(λ) = min
ℓjexje(cid:19)α
+Xj,e
cjexje −Xj,e
λj xje(cid:19)
(2.1)
One can think of the dual problem as having the same instance as the primal, but where jobs are allowed
to be assigned to extent zero. In the objective, in addition to the same load costPe(cid:0)Pj ℓjexje(cid:1)α as in
the primal, a fixed cost of λj is paid for each job j, and a payment of λj − cje is obtained for each unit
of job j assigned. It is well known that each feasible value of the dual function is a lower bound to the
optimal primal solution; this is weak duality [BV04].
αα−1 . To
Online Greedy Algorithm Description: Let δ be a constant that we will later set to
schedule job j, the load is increased on the machines for which the increase the cost will be the least,
assuming that energy costs are discounted by a factor of δ, until a unit of job j is scheduled. More
formally, the value of all the primal variables xje for all the machines e that minimize
1
δ · α · ℓje(cid:18)Xi≤j
ℓiexie(cid:19)α−1
+ cje
(2.2)
are increased until all the work from job j is scheduled, i.e.,Pe xje = 1. Notice that α · ℓje(cid:0)Pi≤j ℓiexie(cid:1)α−1
is the rate at which the load cost is increasing for machine e, and cje is the rate that assignment costs
are increasing for machine e. In other words, our algorithm fractionally assigns the job to the machines
on which the overall objective function increases at the least rate. Furthermore, observe that if the
3
algorithm begins assigning the job to some machine e, it does not stop raising the primal variable xje
until the job is fully assigned1. By this monotonicity property, it is clear that all machines e for which
xje > 0 have the same value of the above derivative when j is fully assigned. Now, for the purpose of
analysis, we set the valuebλj to be the rate of increase of the objective value when we assigned the last
infinitesimal portion of job j. More formally, if e is any machine on which job j is run, i.e., if xje > 0,
then
+ cje
(2.3)
bλj := δ · α · ℓje(cid:18)Xi≤j
ℓiexie(cid:19)α−1
variables for the online algorithm.
assignment of job j (we assign a total of 1 unit of job j, and λj is set to be the rate at which objective
value increases).
Intuitively, bλj is a surrogate for the total increase in objective function value due to our fractional
We now move on to the analysis of our algorithm. To this end, let ex denote the final value of the xje
(by weak duality) to show that g(bλ) is at least
end, let bx be the value of the minimizing x variables in g(bλ), namely
x(cid:23)0(cid:18)Xj bλj +Xe (cid:18)Xj
Algorithm Analysis. To establish that the online algorithm is αα-competitive, note that it is sufficient
1
αα times the cost of the online solution. Towards this
−Xj,e (cid:18)bλj − cje(cid:19) xje(cid:19)
ℓjexje(cid:19)α
bx = arg min
Then by the usual argument of either increasing or decreasing these variables along the line that keeps
choice for the job ϕ(e) is the latest arriving job for which the online algorithm scheduled some bit of
work on machine e; Let us denote this latest job by ψ(e).
Observe that the values bx could be very different from the values ex, and indeed the next few Lemmas
try to characterize these values. Lemma 2.1 notes that bx only has one job ϕ(e) on each machine e,
and Lemma 2.2 shows how to determine ϕ(e) andbxϕ(e)e. Then, in Lemma 2.3, we show that a feasible
Lemma 2.1 There is a minimizing solution bx such that if bxje > 0, then bxie = 0 for all i 6= j.
Proof. Suppose for some machine e, there exist distinct jobs i and k such that bxie > 0 and bxke > 0.
their sum constant, we can keep the convex term (Pj ℓjebxje)α term fixed and not increase the linear
termPj(bλj − cje)bxje. This allows us to either setbxie orbxke to zero without increasing the objective.
αℓϕ(e)e (cid:19)1/(α−1)
and bxje = 0
for j 6= ϕ(e). Moreover, the contribution of machine e towards g(bλ) is exactly (1−α)(cid:18) bλϕ(e)−cϕ(e)e
αℓϕ(e)e (cid:19)α/(α−1)
Proof. By Lemma 2.1 we know that in bx there is at most one job (say j, if any) run on machine e.
Then the contribution of this machine to the value of g(bλ) is
(ℓjebxje)α − (bλj − cje)bxje
Sincebx is a minimizer for g(bλ), we know that the partial derivative of the above term evaluates to zero.
This gives αℓje · (ℓjebxje)α−1 −(cid:16)bλj − cje(cid:17) = 0, or equivalently, bxje = 1
. Then bxϕ(e)e = 1
je′ at different rates so as to balance the derivatives where e and e′ are both machines
ℓϕ(e)e(cid:18) bλϕ(e)−cϕ(e)e
Lemma 2.2 Define ϕ(e) = arg maxj
ℓje(cid:16) bλj −cje
αℓje (cid:17)1/(α−1)
1It may however increase xje and x
. Substituting
(bλj −cje)
ℓje
.
(2.4)
which minimize equation 2.2
4
into this value of bxje into equation (2.4), the contribution of machine e towards the dual g(bλ) is
αℓje !α/(α−1)
= (1 − α) bλj − cje
bλj − cje
αℓje !α/(α−1)
bλj − cje
αℓje !1/(α−1)
(bλj − cje)
ℓje
−
Hence, for each machine e, we want to choose that the job j that minimizes this expression, which is
also the job j that maximizes the expression (bλj − cje)/ℓje since α > 1. This is precisely the job ϕ(e)
and the proof is hence complete.
Lemma 2.3 For all machines e, job ψ(e) is feasible choice for ϕ(e).
Proof. The line of reasoning is the following:
ϕ(e) = arg max
j
(cid:0)bλj − cje(cid:1)
ℓje
= arg max
j (cid:18)δ · α ·(cid:18)Xi≤j
ℓjexie(cid:19)α−1(cid:19) = arg max
j (cid:18)(cid:18)Xi≤j
ℓiexie(cid:19)α−1(cid:19) = ψ(e) .
The first equality is the definition of ϕ(e). For the second equality, observe that for any job k,
bλk ≤ δ · α · ℓek(Xi≤k
ℓiexie)α−1 + cke =⇒ bλk − cke
ℓke
≤ δ α (Xi≤k
ℓiexie)α−1 .
The expression on the right is monotone increasing in Pi≤k ℓiexie, the load due to jobs up to (and
including k). Moreover, it is maximized by the last job to assign fractionally to e (since the inequality
is strict for all other jobs). Since this last job is ψ(e), the last equality follows.
Theorem 2.4 The online greedy algorithm is αα-competitive.
Proof. By weak duality it is sufficient to show that g(bλ) ≥ ON/αα. Applying Lemma 2.2 to the
expression for g(bλ) (equation (2.1)) and substituting the contribution of each machine towards the
dual, we get that
(1 − α)(cid:18)bλψ(e) − cψ(e)e
αℓψ(e)e
(cid:19)α/(α−1)(cid:19)
g(bλ) =(cid:18)Xj bλj +Xe
Xj
λj =Xj,e bλjexje
=Xe exje(cid:18)Xj
= (δ · α)Xe Xj
≥ δXe (cid:18)Xj
Now we consider only the first termPjbλj and evaluate it.
δ · α · ℓje(cid:18)Xi≤j
ℓjeexje(cid:18)Xi≤j
ℓjeexje(cid:19)α
ℓieexie(cid:19)α−1
ℓieexie(cid:19)α−1
+Xj,e exjecje
+ cje(cid:19)
+Xj,e exjecje
(2.5)
(2.6)
(2.7)
(2.8)
(2.9)
Now consider the second term of equation (2.5). Note that if we substitute the value ofbλψ(e), it evaluates
to (1 − α)δα/(α−1)Pe(cid:0)Pj ℓjeexje(cid:1)α Putting the above two estimates together, we get
+Xj,e exjecje + (1 − α)δα/(α−1)Xe (cid:18)Xj
ℓjeexje(cid:19)α
g(bλ) ≥ δXe (cid:18)Xj
ℓjeexje(cid:19)α
(2.10)
5
=(cid:18)δ + (1 − α)δα/(α−1)(cid:19)Xe (cid:18)Xj exjeℓje(cid:19)α
≥ ON/αα
+Xj,e exjecje
(2.11)
(2.12)
The final inequality is due to the choice of δ = 1/αα−1 which maximizes(cid:0)δ + (1 − α)δα/(α−1)(cid:1).
As observed, e.g., in [AAG+95], this O(α)α result is the best possible, even for the (fractional) OnGAP
In Section 5, we show how to obtain an O(α)α-competitive
problem without any assignment costs.
algorithm for integer assignments by a very similar greedy algorithm, and dual-fitting, albeit with a
more careful analysis.
3 Application to Speed Scaling
We now discuss the application of our results for OnGAP to some well-studied speed scaling problems.
In these problems a collection of jobs arrive over time. The jth job arrives at time rj, and has size/work
pj. These jobs must be scheduled on a speed scalable processor that can run at any non-negative
speed. There is a convex function P (s) = sα specifying the dynamic power used by the processor as
a function of speed s. The value of α is typically around 3 for CMOS based processors. Commonly,
one considers objectives G of the form S + βE, where S is a scheduling objective, and E is the energy
used by the system. Moreover, the scheduling objective S is a fractional sum objective of the form
Cjt, where Cjt is the cost of completing job a unit of work of job j at time t, and xjt is
PjPt
the amount of work completed at time t, or the corresponding integer sum objective PjPt yjt Cjt,
where yjt indicates whether or not job j was completed at time t. Fractional scheduling objectives
are interesting in their own right (for example, in situations where the client gains some benefit from
the early partial completion of a job), and are often used in an intermediate step in the analysis of
algorithms for integer scheduling objectives.
xjt
pj
Normally one thinks of the online scheduling algorithm as having two components: a job selection policy
to determine the job to run, and a speed scaling policy to determine the processor speed. However, one
gets a different view when one thinks of the online scheduler as solving online the following mathematical
programming formulation of the problem (which is an instance of the OnGAP problem):
min Xt (cid:18)Xj
xjt(cid:19)α
+Xj Xt
Cjt xjt
xjt ≥ 1
j = 1, . . . , n
subject to Xt
Here the variables xjt specify how much work from job j is run at time t. Because we are initially
concerned with fractional scheduling objectives, we can assume without loss of generality that all jobs
have unit length. The arrival of a job j corresponds to the arrival of a constraint specifying that job j
must be completed. Greedily raising the primal variables corresponds to committing to complete the
work of job j in the cheapest possible way, given the previous commitments.
Two Special Cases. A well-studied speed-scaling problem in this class is when the scheduling ob-
jective is energy minimization subject to deadline feasibility [YDS95, BKP07, BBCP11, BCPK09]: for
each job j there is a deadline dj, and cjt = 0 for t ∈ [rj, dj] and is infinite otherwise. Our algorithm for
OnGAP is essentially equivalent to the algorithm Optimal Available (OA), introduced in [YDS95] and
shown to be αα-competitive in [BKP07] -- specifically, the speeds set by both OA and our algorithm are
the same at all times, but the jobs that are run may be different, since OA uses Earliest Deadline First
for scheduling. Our analysis of the online greedy algorithm for OnGAP is an alternate, and simpler,
analysis of OA than the potential function analysis in [BKP07].
In this instance, our duality based
analysis is tight, as OA is no better than αα-competitive [YDS95, BKP07].
6
Another well-studied special case is when the objective is total flow [AF07, BPS09, BCP09, LLTW08,
ALW10, CEL+09, CLL10]. That is, cjt equals (t − rj) for t ≥ rj and infinite otherwise. All prior
algorithms for this objective assume some variation of the balancing speed scaling algorithm that sets
the power equal to the (fractional) number/weight of unfinished jobs. Unlike the earlier example above,
OnGAP behaves differently than these balancing algorithms for the following reason. When a job arrives,
OnGAP only focuses on choosing assignments which minimize the rate of increase of the objective, and
this rule determines both the scheduling policy (in fact, the entire schedule of job j is decided upon
arrival) and the power usage over time. However, in the balancing algorithms the speed profile (and
hence power usage) is entirely determined by the scheduling policy, and the scheduling policies used
are typically the ones optimal for fractional flow like SJF. Hence it is likely that our algorithms will
actually have a different work profile from balancing algorithms.
A Note about our Approximation Guarantees. A closer examination of our analysis (especially
equation (2.11)) shows that our algorithm has a Lagrangian multiplier preserving property: we get that
our convex cost + αα times the linear term is at most αα times the dual. This separation between the
linear and non-linear terms in the objective happens because the constraints are linear, and when we
compute the dual, the dual variables are involved in the linear terms whereas the convex terms in the
minimizer are identical to their expressions in the primal. In order to argue about the dual minimizer,
this somehow forces us to be exact on the linear terms. This can perhaps explain why our analysis
is tight for the deadline feasibility and load balancing problems [BKP07, AAF+97, AAG+95] (where
there are no linear terms), but is an O(αα)-factor worse than the algorithm of [BCP09] for fractional
flow+energy (which has non-trivial linear terms).
In most potential function-based analyses of speed scaling
Comparison to Previous Techniques.
problems in the literature, the potential function is defined to be the future cost for the online algorithm
to finish the jobs if the remaining sizes of the jobs were the lags of the jobs, which is how far the online
algorithm is behind on the job [IMP11]. A seemingly necessarily condition to apply this potential
function technique is that there must be is a relatively simple algebraic expression for the future cost
for the online scheduling algorithm starting from an arbitrary state. As it is not clear how to obtain such
an algebraic expression for the most obvious candidate algorithms for nonlinear scheduling objectives, 2
this to date has limited the application of this potential function method to speed scaling problems
with linear scheduling objectives. However, our dual-based analysis for OnGAP yields an online greedy
speed scaling algorithm that is Oα(1)-competitive for any sum scheduling objective.
Our algorithm for OnGAP has the advantage that, at the release time of a job, it can commit to the
client exactly the times that each portion of the job will be run. One can certainly imagine situations
when this information would be useful to the client. Also the OnGAP analysis applies to a wider class
of machine environments than does the previous potential function based analyses in the literature. For
example, our analysis of the OnGAP algorithm can handle the case that the processor is unavailable at
certain times, without any modification to the analysis. Although this generality has the disadvantage
that it gives sub-optimal bounds for some problems, such as when the scheduling objective is total flow.
By speeding up the processor by a (1+ǫ) factor, one can obtain an online speed scaling algorithm that one
can show, using known techniques [BLMSP06, BPS09], has competitive ratios at most min(αα(1+ǫ)α, 1
ǫ )
for the corresponding integer scheduling objective.
2 For Lα norms of flow on a single fixed speed processor, no potential function is required to prove scalability of natural
online algorithms [BP10]. For more complicated non-work conserving machine environments, there are analyses that use a
potential function that is a rough approximation of future costs [GIK+10, IM10]. But despite some effort, it has not been
clear how to extend these potential functions to apply to Lα norms of flow in the speed scaling setting.
7
4 Routing with Speed Scalable Routers
In this section we consider the following online routing problem. Routing requests in a graph/network
arriving over time. The jth request consists of a source sj, a sink tj, and a flow requirement fj. In
the unsplittable flow version of the problem the online algorithm must route fj units of flow along a
single (sj, tj)-path. In the splittable flow version of this problem, the online algorithm may partition the
fj units of flow among a collection of (sj, tj)-paths. In either case, we assume speed scalable network
elements (routers, or links, or both) that use power ℓα when they have load ℓ, where the load is the sum
of the flows through the element. We consider the objective of minimizing the aggregate power. We show
that an intuitive online greedy algorithm is αα-competitive using the dual function of a mathematical
programming formulation as a lower bound to optimal.
High Level Idea: The proof will follow the same general approach as for OnGAP: we define dual
just job assignments) is not so straight-forward: the different edges on a path p might want to set f (p)
to different values. So we do something seemingly bad: we relax the dual to decouple the variables, and
allow each (edge, path) pair to choose its own "flow" value f (p, e). And which of these should we use as
variablesbλj for the demand pairs, but now the minimization problem (which is over flow paths, and not
our surrogate for f (p)? We use a convex combination Pe∈p he f (p, e) -- where the multipliers h(e) are
chosen based on the primal loads(!), hence capturing which edges are important and which are not.
4.1 The Algorithm and Analysis
We first consider the splittable flow version of the problem. Therefore, we can assume without loss of
generality that all flow requirements are unit, and all sources and sinks are distinct (so we can associate
a unique request j(p) with each path p). This will also allow us to order paths by the order in which flow
was sent along the paths. We now model the problem by the following primal optimization formulation:
min Xe (cid:18)Xj Xp∋e:p∈Pj
f (p)(cid:19)α
f (p) ≥ 1
j = 1, . . . , n
subject to Xp∈Pj
where Pj is the set of all (sj, tj) paths, and f (p) is a non-negative real variable denoting the amount of
flow routed on the path p. In this case, the dual function is:
g(λ) = min
f (p)(cid:18)Xj
λj +Xe (cid:18)Xj Xp∋e:p∈Pj
f (p)(cid:19)α
− Xj,p∈Pj
λjf (p)(cid:19)
One can think of the dual function as a routing problem with the same instance, but without the
constraints that at least a unit of flow must be routed for each request. In the objective, in addition
to energy costs, a fixed cost of λj is payed for each request j, and a payment of λj is received for each
unit of flow routed from sj to tj.
Description of the Online Greedy Algorithm: To route flow for request j, flow is continuously
routed along the paths that will increase costs the least until enough flow is routed to satisfy the request.
That is, flow is routed along all (sj, tj) paths p that minimizePe∈p α ·(cid:16)Pq≤p:q∋e f (q)(cid:17)α−1
purposes, after the flow for request j is routed, we define
. For analysis
where p is any path along which flow for request j was routed, and δ is a constant (later set to
1
αα−1 ).
bλj = αδ(cid:18)Xe∈p Xq≤p:q∋e
f (q)(cid:19)α−1
8
The Analysis: Unfortunately, unlike the previous section for load balancing, it is not so clear how to
there (per machine). In order to circumvent this difficulty, we consider the following relaxed function
compute the dual function g(bλ) or its minimizer since the variables cannot be nicely decoupled as we did
bg(bλ, h), which does not enforce the constraint that flow must be routed along paths. This enables us
to decouple variables and then argue about the objective value. Indeed, let ef (p) be the final flow on
path p for the routing produced by the online algorithm. Let h(e) = αPp∋e ef (p)α−1 be the incremental
cost of routing additional flow along edge e, and h(p) =Pe∈p h(e) be the incremental cost of routing
additional flow along path p. We then define:
f (p,e)(cid:18)Xj bλj +Xe (cid:18)Xj Xp∋e:p∈Pj
bg(bλ, h) = min
f (p, e)(cid:19)α
−Xj bλj Xp∈PjXe∈p
h(e)
h(p)
f (p, e)(cid:19)
in terms of the final online primal solution.
has the option of increasing the load on individual edges e ∈ p ∈ Pj, but the income from edge e will
be a factor of h(e)
Conceptually, f (p, e) can be viewed as the load placed on edge e by request j(p). Inbg(bλ, h), the scheduler
h(p) less than the income achieved in g(bλ). In Lemma 4.1 we prove thatbg(bλ, h) is a lower
bound for g(bλ). We then proceed as in the analysis of OnGAP. Lemma 4.2 shows how the minimizer
and value ofbg(bλ, h) can be computed, and Lemma 4.3 shows how to bound some of the dual variables
Lemma 4.1 For the above setting of h(·), bg(bλ, h) ≤ g(bλ).
Proof. We show that there is a feasible value ofbg(bλ, h) that is less than g(bλ). Let the value of f (p, e)
inbg(bλ, h) be the same as the value of f (p) in g(bλ). Plugging these values for f (p, e) into the expression
forbg(bλ, h), and simplifying, we get:
The first equality holds by the definitions of h(e) and h(p), and the second equality holds by the
optimality of f (p).
.
that p(e) = arg maxp∋e
α·h(p(e)) (cid:19)1/(α−1)
follows the same reasoning as in the proof of Lemma 2.1. Once we know only one request sends flow on
any edge we can use calculus to identify the minimizer and the value which achieves it. Indeed, we get
Lemma 4.2 There is a minimizer bf ofbg(bλ, h) in which for each edge e, there is a single path p(e) such
that bf (p, e) is positive, and bf (p(e), e) =(cid:18) bλj(p(e))h(e)
Proof. The argument that bg(bλ, h) has a minimizer where each edge only has flow from one request
. and the value of bf (p(e), e) is set so that the incremental energy cost
would just offset the incremental income from routing the flow, that is αbf (p(e), e)α−1 =bλj(p(e))
Solving for bf (p(e), e), the result follows.
Lemma 4.3 bλj(p(e)) ≤ δ · h(p(e))
Proof. bλj(p(e)) is δ times the rate at which the energy cost was increasing for the online algorithm when
it routed the last bit of flow for request j(p(e)). h(p(e)) is the rate at which the energy cost would
h(e)
h(p(e)) .
bλj(p)h(e)
h(p)
9
bg(bλ, h) ≤Xj bλj +Xe (cid:18)Xj Xp∋e:p∈Pj
=Xj bλj +Xe (cid:18)Xj Xp∋e:p∈Pj
= g(bλ)
f (p)(cid:19)α
f (p)(cid:19)α
−Xj bλj Xp∈Pj
−Xj bλj Xp∈Pj
f (p)Xe∈p
h(e)
h(p)
f (p)
increase for the online algorithm if additional flow was pushed along p(e) after the last request was
satisfied. If p(e) was a path on which the online algorithm routed flow, then the result follows from
the fact the online algorithm never decreases the flow on any edge. If p(e) was not a path on which
the online algorithm routed flow, then the result follows from the fact that at the time that the online
algorithm was routing flow for request j(p(e)), p(e) was more costly than the selected paths, and the
cost can't decrease subsequently due to monotonicity of the flows in the online solution.
Theorem 4.4 The online greedy algorithm is αα competitive.
Proof. We will show thatbg(bλ, h) is at least the online cost ON divided by αα, which is sufficient since
bg(bλ, h) is a lower bound to g(bλ) by Lemma 4.1, and since g(bλ) is a lower bound to optimal.
f (p, e)(cid:19)
h(e)
h(p)
(4.13)
−Xj bλj Xp∈PjXe∈p
bg(bλ, h) = min
f (p, e)(cid:19)α
f (p,e)(cid:18)Xj bλj +Xe (cid:18)Xj Xp∋e:p∈Pj
=Xj bλj − (α − 1)Xe (cid:18)bλj(p(e))h(e)
α · h(p(e))(cid:19)α/(α−1)
≥Xj bλj − (α − 1)Xe (cid:18) δ · h(e)
α (cid:19)α/(α−1)
=Xj bλj − (α − 1)δα/(α−1)Xe (cid:18)Xp∋e ef (p)(cid:19)α
=Xj bλj Xp∈Pj ef (p) − (α − 1)δα/(α−1)Xe (cid:18)Xp∋e ef (p)(cid:19)α
= δαXj Xp∈Pj ef (p)(cid:18)Xe∈p Xq≤p:q∋eef (q)(cid:19)α−1
− (α − 1)δα/(α−1)Xe (cid:18)Xp∋e ef (p)(cid:19)α
≥ δXe (cid:18)Xp∋e ef (p)(cid:19)α
ααXe (cid:18)Xp∋e ef (p)(cid:19)α
=
≥ ON/αα
1
− (α − 1)δα/(α−1)Xe (cid:18)Xp∋e ef (p)(cid:19)α
(4.14)
(4.15)
(4.16)
(4.17)
(4.18)
(4.19)
(4.20)
(4.21)
4.2. The inequality in line (4.15) follows from Lemma 4.3. The equality in line (4.16) follows from the
The equality in line (4.13) is the definition ofbg(bλ, h). The equality in line (4.14) follows from Lemma
definition of h(e). The equality in line (4.17) follows from the feasibility of ef . The equality in line (4.18)
follows from the definition ofbλ. The equality in line (4.19) follows from the definition of δ.
While the above algorithm only gives a splittable routing, i.e., a fractional routing, we note that the
ideas of the next section, Section 5, can be used to obtain an Oα(1)-competitive algorithm for integer
flow as well by using a slightly modified primal program (we have to handle non-uniform demands, and
also strengthen the basic convex program to prevent some trivial integrality gaps. The next section
described how we can handle these issues for the load balancing problem.
5 Online Load Balancing: Integral Assignments
For simplicity, let us consider online integer load balancing without assignment costs; it is easy to see
the extension to the other problems that we consider. In this problem each job has the values ℓje, and
the goal is to integrally assign it to a single machine so as to minimize the sumPe(Pj Xjeℓje)α where
10
Xje is the indicator variable for whether job j is assigned to machine e. The most natural reduction
to our general model OnGAP is to set all cje's to 0. However, the convex relaxation for this setting
has a large integrality gap with respect to integral solutions. For example, consider the case of just a
single job which splits into m equal parts (where m is the number of machines) -- the integer primal
optimal pays a factor of mα−1 times the fractional primal optimal. To handle this case, we add a fixed
assignment cost of cje = ℓα
je for assigning job j to machine e. It is easy to see that the cost of an optimal
integral solution at most doubles in this relaxation. This is the convex program that we use for the rest
of this section.
5.1 Approach I: Integer Assignment
In this section, we show that an algorithm which gives an O(α)α-competitive ratio. Consider the
following greedy algorithm: when job j arrives, it picks the machine e that minimizes
δ · α · ℓje(Xi<j
ℓiexie)α−1 + ℓα
je,
settings of the primal variables.
and set xje = 1. Moreover, set bλj for job j to be precisely the quantity above. Let exie be the final
For the analysis, we again need to set thebxj's. For machine e, again consider ϕ(e) and ψ(e) as defined
in the previous proofs. We can no longer claim that ψ(e) is a feasible setting for ϕ(e). However, we can
claim that the load on machine e that is seen by ϕ(e) (and indeed, by any job k) is at most the load
seen by ψ(e) when it arrived, plus the length ℓψ(e)e. Hence
bλϕ(e) ≤ δ α ℓϕ(e)e( Xi<ψ(e)
But since ψ(e) is the last job on machine e, this last expression is exactly δ (Pi ℓieexie)α−1. Now recall
the dual from (2.5). Observing that that α > 1, we can use the calculations we just did to get
ℓieexie + ℓψ(e)e)α−1 + ℓα
ϕ(e)e =⇒ bλϕ(e) − ℓα
αℓϕ(e)e
ℓieexie)α−1.
≤ δ ( Xi≤ψ(e)
(5.22)
ϕ(e)e
g(bλ) ≥Xj bλj + (1 − α)δα/(α−1) Xe (cid:18)Xi
ℓieexie(cid:19)α
ℓjeexje(cid:18)Xi:i<j
ℓieexie(cid:19)α−1
jeexje + (1 − α)δα/(α−1)(cid:18)Xj,e
=αδXe,j
+Xj,e
αδ Xj∈Se
je + (1 − α)δα/(α−1)(cid:18)Xj∈Se
ℓie(cid:19)α−1
ℓje(cid:18) Xi∈Se:i<j
=Xe
+Xj∈Se
ℓje(cid:19)α
e(e(α + 1))α ×Xe (cid:18)Xj∈Se
ℓα
ℓα
≥
1
ℓjeexje(cid:19)α
ℓje(cid:19)α
where the last inequality is obtained by applying Lemma 5.1 to the expression for each machine e. This
implies the O(α)α competitive ratio for the integral assignment algorithm.
Lemma 5.1 Given non-negative numbers a0, a1, a2, . . . , aT and δ = (e(α + 1))α−1, we get
αδ Xj∈[T ]
aj(cid:18)Xi<j
ai(cid:19)α−1
+ Xj∈[T ]
aα
j + (1 − α)δα/(α−1)(cid:18) Xj∈[T ]
aj(cid:19)α
≥
1
e(e(α + 1))α ×(cid:18) Xj∈[T ]
aj(cid:19)α
.
(5.23)
11
Proof. First, consider the case when α ≥ 2:
in this case, we bound the LHS of (5.23) when the
sequence of numbers is non-decreasing, and then we show the non-decreasing sequence makes this LHS
the smallest. We then consider the (easier) case of α ∈ [1, 2].
(lower) bound the following term
Suppose a0 ≤ a1 ≤ · · · ≤ aT , then PT
ai(cid:19)α−1
aj(cid:18)Xi:i≤j
T −1Xj=0
αδ
j=0 aj(Pi<j ai)α−1 ≥ PT −1
j=0 aj(Pi≤j ai)α−1, and it suffices to
+
TXj=1
aα
j − (α − 1)δα/(α−1)(cid:18) TXj=1
aj(cid:19)α
(5.24)
There are two cases, depending on the last term: whether aT ≤ 1
j=0 aj, or not.
j=0 aj. Now, consider the first term in (5.24):
• If aT ≤ 1
j=0 aj, we getPT
αPT −1
ai(cid:19)α−1
aj(cid:18)Xi:i≤j
T −1Xj=0
j=0 aj ≤ (1 + 1/α)PT −1
≥ δ(cid:18) T −1Xj=0
aj(cid:19)α
αδ
≥
αPT −1
(1 + 1/α)α(cid:18) TXj=0
δ
aj(cid:19)α
≥
δ
e(cid:18) TXj=0
aj(cid:19)α
.
(The last inequality used the fact that (1 + 1/α)α approaches e from below.) Finally, plugging
this back into (5.24), ignoring the second sum, and using δ = (e(α + 1))1−α, we can lower bound
the expression of (5.24) by (PT
− (α − 1)δα/(α−1) =
δ
e
j=0 aj)α times
1
e(e(α + 1))α−1 −
α − 1
(e(α + 1))α =
(α + 1) − (α − 1)
(eα)α(1 + 1/α)α ≥
2
e(eα)α
• In case aT ≥ 1
term aα
j=0 aj, we get Pj aj =Pj<T aj + aT ≤ (1 + α)aT . Now using just the single
αPT −1
T from the first two summations in (5.24), we can lower bound it by
aα
T − (α − 1)δα/(α−1)(1 + α)αaα
T = aα
T(cid:18)1 −
(α − 1)(1 + α)α
(e(α + 1))α (cid:19) = aα
T(cid:18)1 −
α − 1
eα (cid:19) .
This is at least aα
T /2 ≥
1
2(1+α)α (Pj aj)α ≥ 1
2eαα .
So in either case the inequality of the statement of Lemma 5.1 is satisfied.
Now to show that the non-decreasing sequence makes the LHS smallest for α ≥ 2. Only the first
summation depends on the order, so focus on Pj∈[T ] aj(Pi<j ai)α−1. Suppose ak > ak+1, then let
ak = l, ak+1 = s; moreover, we can scale the numbers so thatPi<k ai = 1. Now swapping ak and ak+1
causes a decrease of
(ak − ak+1) · 1α−1 + ak+1(1 + ak)α−1 − ak(1 + ak+1)α−1 = (l − s) + s(1 + l)α−1 − l(1 + s)α−1.
And for α ≥ 2 this quantity is non-negative.
Finally, for the case α ∈ [1, 2). Note that αδ ≤ 1 for our choice of δ, so the LHS of (5.23) is at least
αδ Xj∈[T ]
≥ αδ Xj∈[T ]
aj(cid:18)Xi<j
ai(cid:19)α−1
ai(cid:19)α−1
aj(cid:18)Xi≤j
+ aα−1
j + (1 − α)δα/(α−1)(cid:18) Xj∈[T ]
+ (1 − α)δα/(α−1)(cid:18) Xj∈[T ]
aj(cid:19)α
aj(cid:19)α
The second inequality used the fact that aβ + bβ ≥ (a + b)β for β ∈ (0, 1). Now the proof proceeds as
usual and gives us the desired O(eα)α bound.
12
5.2 Approach II: Randomized Rounding
We now explain a different way of obtaining integer solutions: by rounding (in an online fashion) the
fractional solutions obtained for the problems that we consider into integral solutions. While this has a
weaker result, it is a simple strategy that may be useful in some contexts.
Suppose we have a fractional solution w.r.t the above parameters (after including the assignment
cost). While it is known that the convex programming formulation we use has an integrality gap
of 2 [AE05, AKMPS09], these proofs use correlated rounding procedures which we currently are not
able to implement online. Instead we analyze the simple online rounding procedure that independently
and randomly rounds the fractional assignment. Indeed, suppose we independently assign each job j to
a machine e with probability exje. Denote the integer assignment induced by this random experiment
by Yje. Let Le = Pj ℓjeYje denote the random load on machine e after the independent random-
ized rounding. Note that Le is a sum of non-negative and independent random variables. We can
now use the following inequality (for bounding higher moments of sums of random variables) due to
Rosenthal [Ros70, JSZ85] to get that
E [Lα
e ]1/α ≤ Kα max(cid:18)Xj
E [ℓjeYje] ,(cid:18)Xj
jeY α
je(cid:3)(cid:19)1/α(cid:19),
E(cid:2)ℓα
where Kα = O(α/ log α). However we know that Pj
Pj E[ℓα
jeexje(cid:19)1/α(cid:19)(cid:19)α
E [ℓjeYje] =Pj ℓjeexje, and that Pj E[ℓα
jeexje. Substituting this back in and using (a + b)α ≤ 2α−1(aα + bα), we get
cjexje(cid:19).
jeYje] =Pj ℓα
e ] ≤(cid:18)Kα max(cid:18)Xj
ℓjeexje(cid:1)α +Xj
α(cid:18)(cid:0)Xj
E[Lα
ℓjeexje,(cid:18)Xj
Summing over all e, we infer that E[Pe Lα
ℓα
e ] is at most (2Kα)α times the value of the online frac-
tional solution objective, and hence at most (2αKα)α = O(α2/ log α)α times the integer optimum by
[AAG+95, Car08] give
Theorem 2.4.
O(α)α-competitive online algorithms for the integer case.)
(Note that the results of the previous section, and those of
≤ 2α−1K α
jeY α
je] =
6 Conclusion
The online primal-dual dual technique (surveyed in [BN07]) has proven to be a widely-systematically-
applicable method to analyze online algorithms for problems expressible by linear programs. This paper
develops an analogous technique to analyze online algorithms for problems expressible by nonlinear
programs. The main difference is that in the nonlinear setting one can not disentangle the objective
and the constraints in the dual, and hence the arguments for the dual have a somewhat different feel
to them than in the linear setting. We apply this technique to several natural nonlinear covering
problems, most notably obtaining competitive analysis for greedy algorithms for uniprocessor speed
scaling problems with essentially arbitrary scheduling objectives that researchers were not previously
able to analyze using the prevailing potential function based analysis techniques.
Independently and concurrently with this work, Anand, Garg and Kumar [AGK12] obtained results that
are in the same spirit as the results obtained here. Mostly notably, they showed how to use nonlinear-
duality to analyze a greedy algorithm for a multiprocessor speed-scaling problem involving minimizing
flow plus energy on unrelated machines. More generally, [AGK12] showed how duality based analyses
could be given for several scheduling algorithms that were analyzed in the literature using potential
functions.
13
References
[AAF+97]
James Aspnes, Yossi Azar, Amos Fiat, Serge Plotkin, and Orli Waarts. On-line routing of virtual
circuits with applications to load balancing and machine scheduling. J. ACM, 44(3):486 -- 504, 1997.
[AAG+95] Baruch Awerbuch, Yossi Azar, Edward F. Grove, Ming-Yang Kao, P. Krishnan, and Jeffrey Scott
Vitter. Load balancing in the lp norm. In FOCS, pages 383 -- 391, 1995.
[AAZ10]
[AAZ11]
[AE05]
[AF07]
[AGK12]
Matthew Andrews, Spyridon Antonakopoulos, and Lisa Zhang. Minimum-cost network design with
(dis)economies of scale. In FOCS, pages 585 -- 592, 2010.
Matthew Andrews, Spyridon Antonakopoulos, and Lisa Zhang. Energy-aware scheduling algorithms
for network stability. In INFOCOM, pages 1359 -- 1367, 2011.
Yossi Azar and Amir Epstein. Convex programming for scheduling unrelated parallel machines.
In STOC'05: Proceedings of the 37th Annual ACM Symposium on Theory of Computing, pages
331 -- 337. ACM, New York, 2005.
Susanne Albers and Hiroshi Fujiwara. Energy-efficient algorithms for flow time minimization. ACM
Transactions on Algorithms, 3(4):49, 2007.
S. Anand, Naveen Garg, and Amit Kumar. Resource augmentation for weighted flow-time explained
by dual fitting. In SODA, 2012.
[AKMPS09] V. S. Anil Kumar, Madhav V. Marathe, Srinivasan Parthasarathy, and Aravind Srinivasan. A unified
approach to scheduling on unrelated parallel machines. J. ACM, 56(5):Art. 28, 31, 2009.
[Alb10]
Susanne Albers. Energy-efficient algorithms. Commun. ACM, 53(5):86 -- 96, 2010.
[ALW10]
Lachlan L. H. Andrew, Minghong Lin, and Adam Wierman. Optimality, fairness, and robustness in
speed scaling designs. In SIGMETRICS, pages 37 -- 48, 2010.
[BBCP11] Nikhil Bansal, David P. Bunde, Ho-Leung Chan, and Kirk Pruhs. Average rate speed scaling.
Algorithmica, 60(4):877 -- 889, 2011.
[BCP09]
Nikhil Bansal, Ho-Leung Chan, and Kirk Pruhs. Speed scaling with an arbitrary power function.
In SODA, pages 693 -- 701, 2009.
[BCPK09] Nikhil Bansal, Ho-Leung Chan, Kirk Pruhs, and Dmitriy Katz. Improved bounds for speed scaling
in devices obeying the cube-root rule. In ICALP (1), pages 144 -- 155, 2009.
[BKP07]
N. Bansal, T. Kimbrel, and K. Pruhs. Speed scaling to manage energy and temperature. JACM,
54(1), 2007.
[BLMSP06] Luca Becchetti, Stefano Leonardi, Alberto Marchetti-Spaccamela, and Kirk Pruhs. Online weighted
flow time and deadline scheduling. Journal of Discrete Algorithms, 4(3):339 -- 352, 2006.
[BN07]
[BP10]
[BPS09]
[BV04]
[Car08]
[CEL+09]
[CLL10]
Niv Buchbinder and Joseph Naor. The design of competitive online algorithms via a primal-dual
approach. Found. Trends Theor. Comput. Sci., 3(2-3):front matter, 93 -- 263 (2009), 2007.
Nikhil Bansal and Kirk Pruhs. Server scheduling to balance priorities, fairness, and average quality
of service. SIAM J. Comput., 39(7):3311 -- 3335, 2010.
Nikhil Bansal, Kirk Pruhs, and Clifford Stein. Speed scaling for weighted flow time. SIAM J.
Comput., 39(4):1294 -- 1308, 2009.
Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, New
York, NY, USA, 2004.
Ioannis Caragiannis. Better bounds for online load balancing on unrelated machines. In Proceedings
of the Nineteenth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 972 -- 981, New
York, 2008. ACM.
Ho-Leung Chan, Jeff Edmonds, Tak Wah Lam, Lap-Kei Lee, Alberto Marchetti-Spaccamela, and
Kirk Pruhs. Nonclairvoyant speed scaling for flow and energy. In STACS, pages 255 -- 264, 2009.
Sze-Hang Chan, Tak Wah Lam, and Lap-Kei Lee. Non-clairvoyant speed scaling for weighted flow
time. In ESA (1), pages 23 -- 35, 2010.
14
[GIK+10]
Anupam Gupta, Sungjin Im, Ravishankar Krishnaswamy, Benjamin Moseley, and Kirk Pruhs.
Scheduling jobs with varying parallelizability to reduce variance. In SPAA, pages 11 -- 20, 2010.
[IM10]
[IMP11]
[JSZ85]
Sungjin Im and Benjamin Moseley. An online scalable algorithm for average flow time in broadcast
scheduling. In SODA, pages 1322 -- 1333, 2010.
Sungjin Im, Benjamin Moseley, and Kirk Pruhs. A tutorial on amortized local competitiveness in
online scheduling. SIGACT News, 42(2):83 -- 97, 2011.
W. B. Johnson, G. Schechtman, and J. Zinn. Best constants in moment inequalities for linear
combinations of independent and exchangeable random variables. Ann. Probab., 13(1):234 -- 253,
1985.
[LLTW08] T.W. Lam, L.K. Lee, Isaac To, and P. Wong. Speed scaling functions based for flow time scheduling
based on active job count. In European Symposium on Algorithms, pages 647 -- 659, 2008.
[PUW08]
Kirk Pruhs, Patchrawat Uthaisombut, and Gerhard J. Woeginger. Getting the best response for
your erg. ACM Transactions on Algorithms, 4(3), 2008.
[Ros70]
[YDS95]
Haskell P. Rosenthal. On the subspaces of Lp (p > 2) spanned by sequences of independent random
variables. Israel J. Math., 8:273 -- 303, 1970.
F. Yao, A. Demers, and S. Shenker. A scheduling model for reduced CPU energy. In Proc. IEEE
Symp. Foundations of Computer Science, pages 374 -- 382, 1995.
15
|
1906.11327 | 1 | 1906 | 2019-06-26T20:15:54 | The Adversarial Robustness of Sampling | [
"cs.DS",
"cs.CG",
"cs.CR",
"cs.DB",
"cs.DC"
] | Random sampling is a fundamental primitive in modern algorithms, statistics, and machine learning, used as a generic method to obtain a small yet "representative" subset of the data. In this work, we investigate the robustness of sampling against adaptive adversarial attacks in a streaming setting: An adversary sends a stream of elements from a universe $U$ to a sampling algorithm (e.g., Bernoulli sampling or reservoir sampling), with the goal of making the sample "very unrepresentative" of the underlying data stream. The adversary is fully adaptive in the sense that it knows the exact content of the sample at any given point along the stream, and can choose which element to send next accordingly, in an online manner.
Well-known results in the static setting indicate that if the full stream is chosen in advance (non-adaptively), then a random sample of size $\Omega(d / \varepsilon^2)$ is an $\varepsilon$-approximation of the full data with good probability, where $d$ is the VC-dimension of the underlying set system $(U,R)$. Does this sample size suffice for robustness against an adaptive adversary? The simplistic answer is \emph{negative}: We demonstrate a set system where a constant sample size (corresponding to VC-dimension $1$) suffices in the static setting, yet an adaptive adversary can make the sample very unrepresentative, as long as the sample size is (strongly) sublinear in the stream length, using a simple and easy-to-implement attack.
However, this attack is "theoretical only", requiring the set system size to (essentially) be exponential in the stream length. This is not a coincidence: We show that to make Bernoulli or reservoir sampling robust against adaptive adversaries, the modification required is solely to replace the VC-dimension term $d$ in the sample size with the cardinality term $\log |R|$. This nearly matches the bound imposed by the attack. | cs.DS | cs | The Adversarial Robustness of Sampling
Omri Ben-Eliezer∗
Eylon Yogev†
Abstract
Random sampling is a fundamental primitive in modern algorithms, statistics, and ma-
chine learning, used as a generic method to obtain a small yet "representative" subset of the
data.
In this work, we investigate the robustness of sampling against adaptive adversarial
attacks in a streaming setting: An adversary sends a stream of elements from a universe U to
a sampling algorithm (e.g., Bernoulli sampling or reservoir sampling), with the goal of mak-
ing the sample "very unrepresentative" of the underlying data stream. The adversary is fully
adaptive in the sense that it knows the exact content of the sample at any given point along
the stream, and can choose which element to send next accordingly, in an online manner.
Well-known results in the static setting indicate that if the full stream is chosen in advance
(non-adaptively), then a random sample of size Ω(d/ε2) is an ε-approximation of the full
data with good probability, where d is the VC-dimension of the underlying set system (U, R).
Does this sample size suffice for robustness against an adaptive adversary? The simplistic
answer is negative: We demonstrate a set system where a constant sample size (corresponding
to a VC-dimension of 1) suffices in the static setting, yet an adaptive adversary can make the
sample very unrepresentative, as long as the sample size is (strongly) sublinear in the stream
length, using a simple and easy-to-implement attack.
However, this attack is "theoretical only", requiring the set system size to (essentially) be
exponential in the stream length. This is not a coincidence: We show that in order to make the
sampling algorithm robust against adaptive adversaries, the modification required is solely to
replace the VC-dimension term d in the sample size with the cardinality term log R. That
is, the Bernoulli and reservoir sampling algorithms with sample size Ω(log R/ε2) output a
representative sample of the stream with good probability, even in the presence of an adaptive
adversary. This nearly matches the bound imposed by the attack.
9
1
0
2
n
u
J
6
2
]
S
D
.
s
c
[
1
v
7
2
3
1
1
.
6
0
9
1
:
v
i
X
r
a
∗Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel.
†Department of Computer Science, Technion, Haifa, Israel. Supported by the European Union's Horizon 2020
research and innovation program under grant agreement no. 742754, and by a grant from the Israel Science Foundation
(no. 950/16).
1 Introduction
Random sampling is a simple, generic, and universal method to deal with massive amounts
of data across all scientific disciplines. It has wide-ranging applications in statistics, databases,
networking, data mining, approximation algorithms, randomized algorithms, machine learning,
and other fields (see e.g., [CJSS03, JMR05, JPA04, CDK+11, CG05, CMY11] and [Cha01, Chapter
4]). Perhaps the central reason for its wide applicability is the fact that it (provably, and with
high probability) suffices to take only a small number of random samples from a large dataset
in order to "represent" the dataset truthfully (the precise geometric meaning is explained later).
Thus, instead of performing costly and sometimes infeasible computations on the full dataset,
one can sample a small yet "representative" subset of a data, perform the required analysis on
this small subset, and extrapolate (approximate) conclusions from the small subset to the entire
dataset.
The analysis of sampling algorithms has mostly been studied in the non-adaptive (or static)
setting, where the data is fixed in advance, and then the sampling procedure runs on the fixed
data. However, it is not always realistic to assume that the data does not change during the
sampling procedure, as described in [MNS11, GHR+12, GHS+12, HW13, NY15]. In this work,
we study the robustness of sampling in an adaptive adversarial environment.
In high-level, the model is a two-player game between a ran-
The adversarial environment.
domized streaming algorithm, called Sampler, and an adaptive player, Adversary. In each round,
1. Adversary first submits an element to Sampler. The choice of the element can depend, pos-
sibly in a probabilistic manner, on all elements submitted by Adversary up to this point, as
well as all information that Adversary observed from Sampler up to this point.
2. Next, Sampler probabilistically updates its internal state, i.e., the sample that it currently
maintains. An update step usually involves an insertion of the newly received element to
the sample with some probability, and sometimes deletion of old elements from the sample.
3. Finally, Adversary is allowed to observe the current (updated) state of Sampler, before pro-
ceeding to the next round.
Adversary's goal is to make the sample as unrepresentative as possible, causing Sampler to come
with false conclusions about the data stream. The game is formally described in Section 2.
Adversarial scenarios are common and arise in different settings. An adversary uses adver-
sarial examples to fool a trained machine learning model [SZS+14, MHS19]; In the field of online
learning [Haz16], adversaries are typically adaptive [SS17, LMPL18]. An online store suggests
recommended items based on a sample of previous purchases, which in turn influences future
sales [Sha12, GHR+12]. A network device routes traffic according to statistics pulled from a sam-
pled substream of packets [DLT05], and an adversary that observes the network's traffic learns the
device's routing choices might cause a denial-of-service attack by generating a small amount of
adversarial traffic [NY15]. A high-frequency stock trading algorithm monitors a stream of stock
orders places buy/sell requires based on statistics drawn from samples; A competitor might fool
the sampling algorithm by observing its requests and modifying future stock orders accordingly.
An autonomous vehicle receives physical signals from its immediate environment (which might
be adversarial [SBM+18]) and has to decide on a suitable course of action.
1
Even when there is no apparent adversary, the adaptive perspective is sometimes natural
and required. For instance, adaptive data analysis [DFH+15, WFRS18] aims to understand the
challenges arising when data arrives online, such as data reuse, the implicit bias "collected"
over time in scientific discovery, and the evolution of statistical hypotheses over time. In graph
algorithms, [CGP+18] observed that an adversarial analysis of dynamic spanners would yield a
simpler (and quantitively better) alternative to their work.
In view of the importance of robustness against adaptive adversaries, and the fact that random
sampling is very widely used in practice (including in streaming settings), we ask the following.
Are sampling algorithms robust against adaptive adversaries?
Bernoulli and reservoir sampling. We mainly focus on two of the most basic and well-known
sampling algorithms: Bernoulli sampling and reservoir sampling. The Bernoulli sampling al-
gorithm with parameter p ∈ [0, 1] runs as follows: whenever it receives a stream element xi,
the algorithm stores the element with probability p. For a stream of length n the sample size is
expected to be np; and furthermore, it is well-concentrated around this value. We denote this
algorithm by BernoulliSample.
The classical reservoir sampling algorithm [Vit85] (see also [Knu97, Section 3.4.2] and a formal
description in Section 2) with parameter k ∈ [n] maintains a uniform sample of fixed size k,
acting as follows. The first k elements it receives, x1, . . . , xk, are simply added to the memory
with probability one. When the algorithm receives its ith element xi, where i > k, it stores it with
probability k/i, by overriding a uniformly random element from the memory (so the memory
size is kept fixed to k). We henceforth denote this algorithm by ReservoirSample.
Attacking sampling algorithms. To answer the question above of whether sampling algorithms
are robust against adversarially chosen streams, we must first define a notion of a representative
sample, as several notions might be appropriate. However, we begin the discussion with an
example showing how to attack the Bernoulli (and reservoir) sampling algorithm with respect to
merely any definition of "representative".
Consider a setting where the stream consists of n points x1, . . . , xn in the one-dimensional
range of real numbers [0, 1]. BernoulliSample receives these points and samples each one indepen-
dently with probability p < 1. One can observe that, in the static setting and for sufficiently large
p, the sampled set will be a good representation of the entire n points for various definitions of
the term "representation". For example, the median of the stream will be ε-close1 to the median
of the sampled elements with high probability, as long as p = c
ε2n for some constant c > 0 (this
also holds for any other quantile).
Consider the following adaptive adversary which will demonstrate the difference of the adap-
tive setting. Adversary keeps a "working range" at any point during the game, starting with the
full range [0, 1]. In the first round, Adversary chooses the number x1 = 0.5 as the first element
in the stream. If x1 is sampled, then Adversary moves to the range [0.5, 1], and otherwise, to the
range [0, 0.5]. Next, Adversary submits x2 as the middle of the current range. This continues for n
steps; Formally, Adversary's strategy is as follows. Set a1 = 0 and b1 = 1. In round i, where i runs
1The term "close" here means that the median of the sampled set will be an element whose order among the
elements of the full stream, when the elements are sorted by value from smallest to largest, is within the range
(1 ± ǫ)n/2, with high probability where the parameter ǫ depends on the probability p.
2
from 1 to n, Adversary submits xi = ai+bi
2
ai+1 = xi, bi+1 = bi, and otherwise, it sets ai+1 = ai, bi+1 = xi. The final stream is x1, . . . , xn.
to BernoulliSample; If xi is sampled then Adversary sets
Note that at any point throughout the process, Adversary always submits an element that is
larger than all elements in the current sampled set, and also smaller than all the non-sampled ele-
ments of the stream. Therefore, the end result is that after this process is over, with probability 1,
the k sampled elements are precisely the smallest k elements in the stream. Of course, the median
of the sampled set is far from the median of the stream as such a subset is very unrepresentative
of the data. Actually, one might consider it as the "most unrepresentative" subset of the data.
The exact same attack on BernoulliSample works almost as effectively against ReservoirSample.
In this case, the attack will cause all of the k sampled elements at the end of the process to lie
among the first O(k ln n) elements with high probability. For more details, see Section 5.
The good news. This attack joins a line of attacks in the adversarial model. Lipton and
Naughton [LN93] showed that an adversary that can measure the time of operations in a dic-
tionary can use this information to increase the probability of a collision and as a result, signif-
icantly decrease the performance of the hashtable. Hardt and Woodruff [HW13] showed that
linear sketches are inherently non-robust and cannot be used to compute the Euclidean norm
of its input (where in the static setting they are used mainly for this reason). Naor and Yogev
[NY15] showed that Bloom filters are susceptible to attacks by an adaptive stream of queries if the
adversary is computationally unbounded and they also constructed a robust Bloom filter against
computationally bounded adversaries.
In our case, we note that the given attack might categorize it as "theoretical" only. In practice,
it is unrealistic to assume that the universe from which Adversary can pick elements is an infinite
set; how would the attack look, then, if the universe is the discrete set [N] = {1, . . . , N}? Adversary
splits the range [0, 1] to half for n times, meaning that the precision of the elements required is
exponential; The analogous attack in the discrete setting requires N to be exponentially large
with respect to the stream size n. Such a universe size is large and "unrealistic": for Sampler to
memorize even a single element requires memory size that is linear in n, whilst sampling and
streaming algorithms usually aim to use an amount sublinear in n of memory.
Thus, the question remains whether there exist attacks that can be performed on elements us-
ing substantially less precision, that is, on a significantly smaller size of discrete universe. In this
work, we bring good news to both the Bernoulli and reservoir sampling algorithms by answer-
ing this question negatively. We show that both sampling algorithms, with the right parameters,
will output a representative sample with good probability regardless of Adversary's strategy, thus
exhibiting robustness for these algorithms in adversarial settings.
We note that any deterministic algorithm that works in the static setting is inherently robust in
the adversarial adaptive setting as well. However, in many cases, deterministic algorithms with
small memory simply do not exist, or they are complicated and tailored for a specific task. Here,
we enjoy the simplicity of a generic randomized sampling algorithm combined with the robust
guarantees of our framework.
What is a representative sample? Perhaps the most standard and well-known notion of being
representative is that of an ε-approximation, first suggested by Vapnik and Chervonenkis [VC71]
(see also [MV17]), which originated as a natural notion of discrepancy [Cha01] in the geometric
literature. It is closely related to the celebrated notion of VC-dimension [VC71, Sau72, She72], and
3
captures many quantitative properties that are desired in a random subset. Let X = (x1, . . . , xn)
be a sequence of elements from the universe U (repetitions are allowed) and let R ⊆ U. The
density of R in X is the fraction of elements in X that are also in R (i.e., dR(X) = Pri∈[n][xi ∈ R]).
A set system is simply a pair (U, R) where R ⊆ 2U is a collection of subsets. A non-empty
subsequence S of X is an ε-approximation of X with respect to the set system (U, R) if it preserves
densities (up to an ε factor) for all subsets R ∈ R.
Definition 1.1 (ε-approximation). We say that a (non-empty) sample S is an ε-approximation of X
with respect to R if for any subset R ∈ R it holds that dR(X) − dR(S) ≤ ε.
If the universe U is well-ordered, it is natural to take R as the collection of all consecutive
intervals in U, that is, R = {[a, b] : a ≤ b ∈ U} (including all singletons [a, a]). With this set
system in hand, ε-approximation is a natural form of "good representation" in the streaming
setting, pointed out by its deep connection to multiple classical problems in the streaming litera-
ture, like approximate median, and more generally, quantile estimation [MRL99, GK01, WLYC13,
GK16, KLL16] and range searching [BCEG07]. In particular, if S is an ε-approximation of X w.r.t.
(U, R), then any q-quantile of S is ε-close to the q-quantile of X; this holds simultaneously for all
quantiles (see Section 1.2).
1.1 Our Results
Fix a set system (U, R) over the universe U. A sampling algorithm is called (ε, δ)-robust if
for any (even computationally unbounded) strategy of Adversary, the output sample S is an ε-
approximation of the whole stream X with respect to (U, R), with probability at least 1 − δ.
Our main result is an upper bound ("good news") on the (ε, δ)-robustness of Bernoulli and
reservoir sampling, later to be complemented them with near-matching lower bounds.
Theorem 1.2. For any 0 < ε, δ < 1, set system (U, R), and stream length n, the following holds.
• BernoulliSample with parameter p ≥ 10 · ln R+ln(4/δ)
ε2n
is (ε, δ)-robust.
• ReservoirSample with parameter k ≥ 2 · ln R+ln(2/δ)
ε2
is (ε, δ)-robust.
The proof appears in Section 4. As the total number of elements sampled by BernoulliSample
is well-concentrated around np, the above theorem implies that a sample of total size (at least)
Θ( ln R+ln 1
approximation with probability 1 − δ.
), obtained by any of the algorithms, BernoulliSample or ReservoirSample, is an ε-
ε2
δ
This should be compared with the static setting, where the same result is known as long
as p ≥ c · d+ln 1
ε2n
dimension of (U, R) and c > 0 is a constant [VC71, Tal94, LLS01] (see also [MV17]).
for BernoulliSample, and k ≥ c · d+ln 1
ε2
for ReservoirSample, where d is the VC-
δ
δ
As you can see, to make the static sampling algorithm robust in the adaptive setting one
solely needs to modify the sample size by replacing the VC-dimension term d with the cardinality
dimension ln R (and update the multiplicative constant). Below, in our lower bounds, we show
that this increase in the sample size is inherent, and not a byproduct of our analysis.
4
Lower Bounds. We next show that being adaptively robust comes at a price. That is, the de-
pendence on the cardinality dimension, as opposed to the VC dimension, is necessary. By an
improved version of the attack described in the introduction, we show the following:
Theorem 1.3. There exists a constant c > 0 and a set system (U, R) with VC-dimension 1, where such
that for any 0 < ε, δ < 1/2:
1. The BernoulliSample algorithm with parameter p < c · ln R
n ln n is not (ε, δ)-robust.
2. The ReservoirSample algorithm with parameter k < c · ln R
ln n is not (ε, δ)-robust.
Moreover, for any n6 ln n ≤ N ≤ 2n/2, there exists (U, R) as above where R = U = N.
The proof can be found in Section 5.
Continuous robustness. The condition of (ε, δ)-robustness requires that the sample will be
ε-representative of the stream in the end of the process. What if we wish the sample to be repre-
sentative of the stream at any point throughout the stream? Formally, we say that a sampling
algorithm is (ε, δ)-continuously robust if, with probability at least 1 − δ, at any point i ∈ [n] the
sampled set Si is an ε-approximation of the first i elements of the stream, i.e., of Xi = (x1, . . . , xi).
The next theorem shows that continuous robustness of ReservoirSample can be obtained with just
a small overhead compared to "standard" robustness. (For BernoulliSample one cannot hope for
such a result to be true, at least for the above definition of continuous robustness.)
Theorem 1.4. There exists c > 0, such that for any 0 < ε, δ < 1/2, set system (U, R), and stream
length n, ReservoirSample with parameter k ≥ c · ln R+ln 1/δ+ln 1/ε+ln ln n
is (ε, δ)-continuously robust.
ε2
Moreover, if only continuous robustness against a static adversary is desired, then the ln R term can
be replaced with the VC-dimension of (U, R).
We are not aware of a previous analysis of continuous robustness, even in the static setting.
The proof, appearing in Section 6, follows by applying Theorem 1.2 (or its static analogue) in
carefully picked "checkpoints" k = i1 ≤ i2 ≤ . . . ≤ it = n along the stream, where t = O(ε−1 ln n).
It shows that if the sample Si is representative of the stream Xi in any of the points i = i1, . . . , it−1,
then with high probability, the sample is also representative in any other point along the stream.
(We remark that a similar statement with weaker dependence on n can be obtained from Theorem
1.2 by a straightforward union bound.) The proof can be found in Section 6.
Comparison to deterministic sampling algorithms. Our results show that sampling algorithms
provide an ε-approximation in the adversarial model. One advantage of using the notion of ε-
approximation is its wide array of applications, where for each such task we get a streaming
algorithm in the adversarial model as described in the following subsection. We stress that for
any specific task a deterministic algorithm that works in the static setting will also automatically
be robust in the adversarial setting. However, deterministic algorithms tend to be more compli-
cated, and in some cases they require larger memory. Here, we focus on showing that the most
simple and generic sampling algorithms "as is" are robust in our adaptive model and yield a
representative sample of the data that can be used for many different applications.
The best known deterministic algorithm for computing an ε-approximating sample in the
streaming model is that of Bagchi et al. [BCEG07]. The sample size they obtain is O(ε−2 ln 1/ε);
5
the working space of their algorithm and the processing time per element are of the form
ε−2d−O(1)(ln n)O(d), where d is the scaffold dimension2 of the set system. The exact bounds are
rather intricate, see Corollary 4.2 in [BCEG07]. While the space requirement of their approach
does not have a dependence on ln R, its dependence on ε and ln n is generally worse than ours,
making their bounds somewhat incomparable to ours. Finally, we note that there exist more
efficient methods to generate an ε-approximation in some special cases, e.g., when the set system
constitutes of rectangles or halfspaces [STZ04].
1.2 Applications of Our Results
We next describe several representative applications and usages of ε-approximations (see also
[BCEG07] for more applications in the area of robust statistics). For some of these applications,
there exist deterministic algorithms known to require less memory than the simple random sam-
pling models discuss in this paper. However, one area where our generic random sampling
approach shines compared to deterministic approaches is the query complexity or running time
(under a suitable computational model). Indeed, while deterministic algorithms must inherently
query all elements in the stream in order to run correctly, our random sampling methods query
just a small sublinear portion of the elements in the stream.
Consequently, to the best of our knowledge, Bernoulli and reservoir sampling are the first two
methods known to compute an ε-approximation (and as a byproduct, solve the tasks described in
this subsection) in adversarial situations where it is unrealistic or too costly to query all elements
in the stream. The last part of this subsection exhibits an example of one such situation.
Quantile approximation. As was previously mentioned, ε-approximations have a deep connec-
tion to approximate median (and more generally, quantile estimation). Assume the universe U
is well-ordered. We say that a streaming algorithm is an (ε, δ)-robust quantile sketch if, in our
adversarial model, it provides a sample that allows to approximate the rank3 of any element in
the stream up to additive error εn with probability at least 1 − δ. Observe that this is achieved
with an ε-approximation with respect to the set system (U, R) where R = {[1, b] : b ∈ U}. For
example, set b to be the median of the stream. Since the density of the range [1, b] is preserved in
the sample, we know that the median of the sample will be ε-close to the median of the stream.
This works for any other quantile simultaneously. The sample size is Θ( ln U+ln(1/δ)
).
ε2
Corollary 1.5. For any 0 < ε, δ < 1, well-ordered universe U, and stream length n, BernoulliSample with
parameter p ≥ 10 · ln U+ln(4/δ)
is an (ε, δ)-robust quantile sketch. The same holds for the ReservoirSample
algorithm with parameter k ≥ 2 · ln U+ln(2/δ)
ε2n
.
ε2
A corollary in the same spirit regarding continuously robust quantile sketches can be derived
from Theorem 1.4.
Range queries. Suppose that the universe is of the form U = [m]d for some parameters m and
d. One basic problem is that of range queries: one is given a set of ranges R and each query
consists of a range R ∈ R where the desired answer is the number of points in the stream that
2The scaffold dimension is a variant of the VC-dimension equal to ⌈ln R/ ln U⌉.
3The rank of an element xi in a stream x1, . . . , xn is the total amount of elements xj in the stream so that xj ≤ xi.
6
are in this range. Popular choices of such ranges are axis-aligned or rotated boxes, spherical
ranges and simplicial ranges. An ε-approximation allows us to answer such range queries up
to an additive error of εn. Suppose the sampled set is S, then an answer is given by computing
dR(S) · n/S. For example, when R consists of all axis-parallel boxes, ln R = O(d ln m) and
thus the sample size required to answer range queries that are robust against adversarial streams
[BCEG07] for more details on the connection between ε-approximations and range queries.
(cid:17); for rotated boxes, one should replace d with d2 in this expression. See
is S = O(cid:16) d ln m+ln(1/δ)
ε2
Center points. Our result is also useful for computing β-center points. A point c in the stream
is a β-center point if every closed halfspace containing c in fact contains at least βn points of
the stream. In [CEM+96, Lemma 6.1] it has been shown that an ε-approximation (with respect
to half-spaces) can be used to get a β-center point for suitable choices of the parameters. For
example, setting ε = β/5 we get that a 6β/5-center of the sample S is a β-center of the stream X.
Thus, we can compute a β-center of a stream in the adversarial model. See also [BCEG07].
Heavy hitters. Finding those elements that appear many times in a stream is a fundamental
problem in data mining, with a myriad of practical applications. In the heavy hitters problem,
there is a threshold α and an error parameter ε. The goal is to output a list of elements such that
if an element x appears more than αn times in the stream (i.e., dx(X) ≥ α) it must be included
in the list, and if an element appears less than (α − ε)n times in the stream (i.e., dx(X) ≤ α − ε it
cannot be included in the list.
Our results yield a simple and efficient heavy hitters streaming algorithm in the adversarial
model. For any universe U let R = {{a} : a ∈ U} be the set of all singletons. Now, pick ε′ = ε/3
and use either Bernoulli or reservoir sampling to compute an ε′-approximation S of the stream
X, outputting all elements x ∈ S with d{x}(S) ≥ α − ε′. Indeed, if da(X) ≥ α then dx(S) ≥ α − ε′.
On the other hand, if dx(X) ≤ α − ε then dx(S) ≤ α − ε + ε′ < α − ε′.
Corollary 1.6. There exists c > 0 such that for any 0 < ε, δ < 1/2, universe U, and stream length n,
BernoulliSample with parameter p ≥ c · ln U+ln(1/δ)
solves the heavy hitters problem with error ε in the
adversarial model. The same holds for ReservoirSample with parameter k ≥ c · ln U+ln(1/δ)
ε2n
.
ε2
Clustering. The task of partitioning data elements into separate groups, where the elements
in each group are "similar" and elements in different groups are "dissimilar" is fundamental
and useful for numerous applications across computer science. There has been lots of interest
on clustering in a streaming setting, see e.g. [GLA16] for a survey on recent results. Our results
suggest a generic framework to accelerate clustering algorithms in the adversarial model: Instead
of running clustering on the full data, one can simply sample the data to obtain (with high
probability, even against an adversary) an ε-approximation of it, run the clustering algorithm on
the sample, and then extrapolate the results to the full dataset.
It is very common to use random sampling
Sampling in modern data-processing systems.
(sometimes "in disguise") in modern data-intensive systems that operate on streaming data, ar-
riving in an online manner. As an illustrative example, consider the following distributed database
[OV11] setting. Suppose that a database system must receive and process a huge amount of
7
It is unrealistic for a single server to handle all the queries, and hence,
queries per second.
for load balancing purposes, each incoming query is randomly assigned to one of K query-
processing servers. Seeing that the set of queries that each such server receives is essentially a
Bernoulli random sample (with parameter p = 1/K) of the full stream, one hopes that the por-
tion of the stream sampled by each of these servers would truthfully represent the whole data
stream (e.g., for query optimization purposes), even if the stream changes with time (either unin-
tentionally or by a malicious adversary). Such "representation guarantees" are also desirable in
distributed machine learning systems [GDG+17, SKYL17], where each processing unit learns a
model according to the portion of the data it received, and the models are then aggregated, with
the hope that each of the units processed "similar" data.
In general, modern data-intensive systems like those described above become more and more
complicated with time, consisting of a large number of different components. Making these
systems robust against environmental changes in the data, let alone adversarial changes, is one
of the greatest challenges in modern computer science. From our perspective, the following
question naturally emerges:
Is random sampling a risk in modern data processing systems?
Fortunately, our results indicate that the answer to this question is largely negative. Our up-
per bounds, Theorems 1.2 and 1.4, show that a sufficiently large sample suffices to circumvent
adversarial changes of the environment.
1.3 Related Work
Online learning. One related field to our work is online learning, which was introduced for
settings where the data is given in a sequential online manner or where it is necessary for the
learning algorithm to adapt to changes in the data. Examples include stock price predictions, ad
click prediction, and more (see [Sha12] for an overview and more examples).
Similar to our model, online learning is viewed as a repeated game between a learning algo-
rithm (or a predictor) and the environment (i.e., the adversary). It considers n rounds where in
each round the environment submits an instance xi, the learning algorithm then makes a predic-
tion for xi, the environment, in turn, chooses a loss for this prediction and sends it as feedback
to the algorithm. The goal in this model is usually to minimize regret (the sum of losses) com-
pared to the best fixed prediction in hindsight. This is the typical setting (e.g., [HAK07, SST10]),
however, many different variants exist (e.g., [DGS15, ZLZ18]).
In the PAC-learning framework [Val84], the learner algorithm receives samples
PAC learning.
generated from an unknown distribution and must choose a hypothesis function from a family
of hypotheses that best predicts the data with respect to the given distribution.
It is known
that the number of samples required for a class to be learnable in this model depends on the
VC-dimension of the class.
A recent work of Cullina et al. [CBM18] investigates the effect of evasion adversaries on the
PAC-learning framework, coining the term of adversarial VC-dimension for the parameter govern-
ing the sample complexity. Despite the name similarity, their context is seemingly unrelated to
ours (in particular, it is not a streaming setting), and correspondingly, their notion of adversarial
VC-dimension does not seem to relate to our work.
8
Adversarial examples in deep learning. A very popular line of research in modern deep learn-
ing proposes methods to attack neural networks, and countermeasures to these attacks. In such a
setting, an adversary performs adaptive queries to the learned model in order to fool the model
via a malicious input. The learning algorithms usually have an underlying assumption that the
training and test data are generated from the same statistical distribution. However, in practice,
the presence of an adaptive adversary violates this assumption. There are many devastating ex-
amples of attacks on learning models [SZS+14, BCM+13, PMG+17, BR18, MHS19] and we stress
that currently, the understanding of techniques to defend against such adversaries is rather lim-
ited [GMP18, MW18, MM19, MHS19].
Maintaining random samples. Reservoir sampling is a simple and elegant algorithm for main-
taining a random sample of a stream [Vit85], and since its proposal, many flavors have been
introduced. Chung, Tirthapura, Woodruff [CTW16] generalized reservoir sampling to the setting
of multiple distributed streams, which need to coordinate in order to continuously respond to
queries over the union of all streams observed so far (see also Cormode et al. [CMYZ12]). An-
other variant is weighted reservoir sampling where the probability of sampling an element is
proportional to a weight associated with the element in the stream [ES06, BOV15]. A distributed
version as above was recently considered for the weighted case as well [JSTW19].
1.4 Paper Organization
Section 2 contains an overview of our adversarial model and a more precise and detailed def-
inition than the one given in the introduction. In Section 3 we mention several concentration
In Section 4 we present and prove our main technical
inequalities required for our analysis.
Lemma, from which we derive Theorem 1.2. This includes analysis of both BernoulliSample and
ReservoirSample. In Section 5 we present our "attack", i.e., our lower bound showing the tightness
of our result. Finally, in Section 6, we prove our upper bounds in the continuous setting.
2 The Adversarial Model for Sampling
In this section, we formally define the online adversarial model discussed in this paper. Roughly
speaking, we say that Sampler is an (ε, δ)-robust sampling algorithm for a set system (U, R) if
for any adversary choosing an adaptive stream of elements X = (x1, . . . , xn), the final state of
the sampling algorithm σn is an ε-approximation of the stream with probability 1 − δ. This is
formulated using a game, AdaptiveGame, between two players, Sampler and Adversary.
Rules of the game:
1. Sampler is a streaming algorithm, which gets a sequence of n elements one by one x1, . . . , xn
in an online manner (the sampling algorithms we discuss in this paper do not need to know
n in advance). Upon receiving an element xi, Sampler can perform an arbitrary computation
(the running time can be unbounded) and update a local state σ. We denote the local state
after i steps by σi, and write σi ← Sampler(σi−1, xi).
2. The stream is chosen adaptively by Adversary: a probabilistic (unbounded) player that, given
all previously sent elements x1, . . . , xi−1 and the current state σi−1, chooses the next element
9
xi to submit. The strategy that Adversary employs along the way, that is, the probability
distribution over the choice of xi given any possible set of values x1, . . . , xi−1 and σi−1, is
fixed in advance. The underlying (finite or infinite) set from which Adversary is allowed to
choose elements during the game is called the universe, and denoted by U. We assume that
U does not change along the game.
3. Once all n rounds of the game have ended, Sampler outputs σn. For the sampling algorithms
discussed in this paper, S := σn is a subsequence of the stream X = (x1, . . . , xn). S is usually
called the sample obtained by Sampler in the game.
For an illustration on the rules of the game see Figure 1.
The game AdaptiveGame
Parameters: n, ε, (U, R).
1. Set σ0 = ⊥.
2. For i = 1 . . . n do:
(a) Adversary(σi−1, x1, . . . , xi−1) submits the query xi.
(b) Set σi ← Sampler(σi−1, xi).
3. Let S = σn, and output 1 if S is an ε-representative sample of X = x1, . . . , xn with
respect to (U, R), and 0 otherwise.
Figure 1: The definition of the game AdaptiveGameε between a streaming algorithm Sampler
and Adversary. Here the adversary chooses the next element to the stream while given the state
(memory) of the streaming algorithm thus-far. In the beginning of the game, Adversary receives
the parameters n, ε, (U, R) and knows exactly which sampling algorithm is employed by Sampler.
Using the game defined above, we now describe what it means for a sampling algorithm to
be (adversarially) robust.
Definition 2.1 (Robust sampling algorithm). We say that a sampling algorithm Sampler is (ε, δ)-
robust with respect to the set system (U, R) and the stream length n if for and any (even unbounded)
strategy of Adversary, it holds that
Pr[AdaptiveGame(Sampler, Adversary) = 1] ≥ 1 − δ
The memory size used by Sampler is defined to be the maximal size of σ throughout the process of
AdaptiveGame.
A stronger requirement that one can impose on the sampling algorithm is to hold an ε-
approximation of the stream at any step during the game. To handle this, we define a continuous
variant of AdaptiveGame which we denote ContinuousAdaptiveGame, presented in Figure 2.
For the sampling algorithms that we consider, the state at any time σi is essentially equal to
the sample Si. In any case, the definition of the framework given in Figure 2 generally allows σi
10
The game ContinuousAdaptiveGame
Parameters: n, ε, (U, R).
1. Set σ0 = ⊥.
2. For i = 1 . . . n do:
(a) Adversary((σi−1, Si−1), x1, . . . , xi−1) submits the query xi.
(b) Set (σi, Si) ← Sampler(σi−1, xi).
(c) If Si is not an ε-approximation of Xi = x1, . . . , xi with respect to (U, R) then
output 0 and halt.
3. Output 1.
Figure 2:
The game corresponding to the continuous variant, ContinuousAdaptiveGame be-
tween a streaming algorithm Sampler and an Adversary. Here, Sampler is required to hold an
ε-approximating sampled set Si after each step.
to contain additional information, if needed. A sampling algorithm is called (ε, δ)-continuously
robust if the following holds with probability at least 1 − δ: for any strategy of Adversary, and all
i ∈ [n], the sample Si is an ε-approximation of the stream at time i.
Definition 2.2 (Continuously robust sampling algorithm). We say that a sampling algorithm Sampler
is (ε, δ)-continuously robust with respect to the set system (U, R) and the stream length n if for and any
(even unbounded) strategy of Adversary, it holds that
Pr[ContinuousAdaptiveGame(Sampler, Adversary) = 1] ≥ 1 − δ
The memory size used by Sampler is defined to be the maximal size of σ throughout the process of
ContinuousAdaptiveGame.
Reservoir sampling. For completeness, we provide the pseudocode of the reservoir sampling
algorithm [Vit85, Knu97]. Here, k denotes the (fixed) memory size of the algorithm, i denotes the
current round number, and xi is the currently received element.
ReservoirSample(k, i, σi−1, xi):
1. If i < k then parse σi−1 = x1, . . . , xi−1 and output σi = x1, . . . , xi.
2. Otherwise, parse σi−1 = s1, . . . , sk.
3. With probability k/i do:
choose j ∈ [k] uniformly at random and output σi = s1, . . . , sj−1, xi, sj+1, . . . , sk.
4. Otherwise, output σi = σi−1.
11
3 Technical Preliminaries
The logarithms in this paper are usually of base e, and denoted by ln. The exponential function
exp (x) is ex. For an integer n ∈ N we denote by [n] the set {1, . . . , n}. We state some con-
centration inequalities, useful for our analysis in later sections. We start with the well-known
Chernoff's inequality for sums of independent random variables.
Theorem 3.1 (Chernoff Bound [Che52]; see Theorem 3.2 in [CL06]). Let X1, . . . , Xm be independent
random variables that take the value 1 with probability pi and 0 otherwise, X = ∑m
i=1 Xi, and µ = E[X].
Then for any 0 < δ < 1,
and
δ2µ
Pr[X ≤ (1 − δ)µ] ≤ exp(cid:18)−
Pr[X ≥ (1 + δ)µ] ≤ exp(cid:18)−
2 (cid:19)
2 + 2δ/3(cid:19) .
δ2µ
Our analysis of adversarial strategies crucially makes use of martingale inequalities. We thus
provide the definition of a martingale.
Definition 3.2. A martingale is a sequence X = (X0, . . . , Xm) of random variables with finite means, so
that for 0 ≤ i < m, it holds that E[Xi+1 X0, . . . , Xi] = Xi.
The most basic and well-known martingale inequality, Azuma's (or Hoeffding's) inequality,
asserts that martingales with bounded differences Xi+1 − Xi are well-concentrated around their
mean. For our purposes, this inequality does not suffice, and we need a generalized variant of
it, due to McDiarmid [McD98, Theorem 3.15]; see also Theorem 4.1 in [Fre75]. The formulation
that we shall use is given as Theorem 6.1 in the survey of Chung and Lu [CL06].
Lemma 3.3 (See [CL06], Theorem 6.1). Let X = (X0, X1, . . . , Xn) be a martingale. Suppose further
that for any 1 ≤ i ≤ n, the variance satisfies Var(XiX0, . . . , Xi−1) ≤ σ2
for some values σ1, . . . , σn ≥ 0,
i
and there exists some M ≥ 0 so that Xi − Xi−1 ≤ M always holds. Then, for any λ ≥ 0, we have
Pr(X − X0 ≥ λ) ≤ exp(cid:18)−
λ2
i=1(σ2
i ) + Mλ/3(cid:19) .
2 ∑n
In particular,
Pr(X − X0 ≥ λ) ≤ 2 exp(cid:18)−
λ2
i=1(σ2
i ) + Mλ/3(cid:19) .
2 ∑n
Unlike Azuma's inequality, Lemma 3.3 is well-suited to deal with martingales where the max-
imum value M of Xi+1 − Xi is large, but the maximum is rarely attained (making the variance
much smaller than M2). The martingales we investigate in this paper depict this behavior.
4 Adaptive Robustness of Sampling: Main Technical Result
In this section, we prove the main technical lemma underlying our upper bounds for Bernoulli
sampling and reservoir sampling. The lemma asserts that for both sampling methods, and any
given subset R of the universe U, the fraction of elements from R within the sample typically
does not differ by much from the corresponding fraction among the whole stream.
12
Lemma 4.1. Fix ε, δ > 0, a universe U and a subset R ⊆ U, and let X = (x1, x2, . . . , xn) be the sequence
chosen by Adversary in AdaptiveGameε against either BernoulliSample or ReservoirSample.
1. For BernoulliSample with parameter p ≥ 10 · ln(4/δ)
ε2n
, we have Pr(dR(X) − dR(S) ≥ ε) ≤ δ.
2. For ReservoirSample with memory size k ≥ 2 · ln(2/δ)
ε2
, it holds that Pr(dR(X) − dR(S) ≥ ε) ≤ δ.
Both of these bounds are tight up to an absolute multiplicative constant, even for a static
adversary (that has to submit all elements in advance); see Section 6 for more details.
The proof of Theorem 1.2 follows immediately from Lemma 4.1, and is given below. The
proof of Theorem 1.4 requires slightly more effort, and is given in Section 6.
Proof of Theorem 1.2. Let (U, R), ε, δ, n be as in the statement of the theorem, and let X and
S denote the stream and sample, respectively. We start with the Bernoulli sampling case, and
assume that p ≥ 10 · ln(4/δ)+ln R
. For each R ∈ R, we apply the first part of
Lemma 4.1 with parameters ε and δ/R, concluding that
= 10 · ln(4R/δ)
ε2n
ε2n
Pr(dR(X) − dR(S) ≥ ε) ≤ δ/R.
In the event that dR(X) − dR(S) ≤ ε for any R, by definition S is an ε-approximation of X.
Taking a union bound over all RR, we conclude that the probability of this event not to hold is
bounded by R · (δ/R) = δ, meaning that BernoulliSample with p as above is (ε, δ)-robust.
The proof for ReservoirSample is identical, except that we replace the condition on p with the
condition that k ≥ 2 · ln(2/δ)+ln R
ε2
, and apply the second part of Lemma 4.1.
It is important to note that the typical proofs given for statements of this type in the static
setting (i.e., when Adversary submits all elements in advance, and cannot act adaptively) do
not apply for our adaptive setting. Indeed, the usual proof of the static analogue of the above
lemma goes along the following lines: Adversary chooses which elements to submit in advance,
and in particular, determines the number of elements from A sent, call it nA. Then, the number
of sampled elements from A is distributed according to the binomial distribution Bin(nA, p) for
Bernoulli sampling, and Bin(nA, k/n) for reservoir sampling. One can then employ Chernoff
bound to conclude the proof. This kind of analysis crucially relies on the adversary being static.
Here, we need to deal with an adaptive adversary. Recall that Adversary at any given point
is modeled as a probabilistic process, that given the sequence Xi−1 = (x1, . . . , xi−1) of elements
sent until now, and the current state σi−1 of Sampler, probabilistically decides which element xi
to submit next. Importantly, this makes for a well-defined probability space, and allows us to
analyze Adversary's behavior with probabilistic tools, specifically with concentration inequalities.
Chernoff bound cannot be used here, as it requires the choices made by the adversary along
the process to be independent of each other, which is clearly not the case. In contrast, martingale
inequalities are suitable for this setting. We shall thus employ these, specifically Lemma 3.3, to
prove both parts of our main result in this section.
4.1 The Bernoulli Sampling Case
We start by proving the Bernoulli sampling case (first statement of Lemma 4.1). Recall that here
each element is sampled, independently, with probability p. At any given point 0 ≤ i ≤ n along
13
the process, let Xi = (x1, . . . , xi) denote the sequence of elements submitted by the adversary
until round i, and let Si ⊆ Xi denote the subsequence of sampled elements from Xi. Note that
Xn = X and Sn = S, and hence, to prove the lemma, we need to show that dR(Xn) − dR(Sn) ≤ ε.
As a first attempt, it might make sense to try applying a martingale concentration inequal-
ity on the sequence of random variables (Y0, Y1, . . . , Yn), where we define Yi = dR(Xi) − dR(Si).
Indeed, our end-goal is to bound the probability that Yn significantly deviates from zero. How-
ever, a straightforward calculation shows that this is not a martingale, since the condition that
E[YiY0, . . . , Yi−1] = 0 does not hold in general. To overcome this, we show that a slightly different
formulation of the random variables at hand does yield a martingale. Given the above R ⊆ U,
for any 0 ≤ i ≤ n we define the random variables
AR
i =
i
n
· dR(Xi) =
R ∩ Xi
n
;
BR
i =
R ∩ Si
np
;
i = BR
ZR
i − AR
i ,
(1)
where, as before, the intersection between a set R and a sequence Xi is the subsequence of Xi
consisting of all elements that also belong to R.
Importantly, as is described in the next claim, the sequence of random variables ZR =
0 , . . . , ZR
n ) defined above forms a martingale. The claim also demonstrates several useful prop-
(ZR
erties of these random variables, to be used later in combination with Lemma 3.3.
Claim 4.2. The sequence (ZR
on ZR
0 , . . . , ZR
0 , ZR
1 , . . . , ZR
n ) is a martingale. Furthermore, the variance of ZR
i conditioned
i−1 is bounded by 1/n2 p, and it always holds that ZR
i − ZR
i−1 ≤ 1/np.
We shall prove Claim 4.2 later on; first we use it to complete the proof of the main result.
Proof of Lemma 4.1, Bernoulli sampling case. It suffices to prove the following two inequalities for
any p satisfying the conditions of the lemma for the Bernoulli sampling case:
Pr(AR
n − BR
n ≥ ε/2) ≤ δ/2
;
Pr(BR
n − dR(Sn) ≥ ε/2) ≤ δ/2.
(2)
Indeed, taking a union bound over these two inequalities, applying the triangle inequality, and
observing that AR
n = dR(Xn), we conclude that Pr(dR(Xn) − dR(Sn) ≥ ε) ≤ δ, as desired.
The first inequality follows from Claim 4.2 and Lemma 3.3. Indeed, in view of Claim 4.2, we
i = 1/n2 p, and M = 1/np. As
n ) with parameters λ = ε/2, σ2
0 , . . . , ZR
can apply Lemma 3.3 on (ZR
ZR
0 = 0, we have AR
n − BR
n = ZR
n − ZR
0 , and so
Pr(AR
n − BR
n ≥ ε/2) ≤ 2 exp −
(ε/2)2
n2 p + ε
6np! < 2 exp(cid:18)−
2n · 1
ε2np
9 (cid:19) .
The right hand side is bounded by δ/2 when np ≥ 9
We next prove the second inequality of (2). Observe that BR
ε2 ln(δ/4), settling the first inequality of (2).
np . Since each
element is added to the sample with probability p, independently of other elements, the size of
Sn is distributed according to the binomial distribution Bin(n, p), regardless of the adversary's
strategy. Applying Chernoff inequality with δ = ε/2, we get that
n = dR(Sn) · Sn
Pr((cid:12)(cid:12)Sn − np(cid:12)(cid:12) ≥ εnp/2) ≤ 2 exp −
14
(ε/2)2 np
2 + ε/3 ! < 2 exp(cid:18)−
ε2np
10 (cid:19) .
This probability is bounded by δ/2 provided that np ≥ 10 ln(4/δ)
occurring, we have that
ε2
. Conditioning on this event not
n(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)dR(Sn) − BR
1 −
Sn
np (cid:12)(cid:12)(cid:12)(cid:12)
· dR(Sn) ≤ (cid:12)(cid:12)(cid:12)(cid:12)
1 −
Sn
np (cid:12)(cid:12)(cid:12)(cid:12)
≤
ε
2
,
where the first inequality follows from the fact that densities (in this case, dR(Sn)) are always
bounded from above by one, and the second inequality follows from our conditioning. This
completes the proof of the second inequality in (2).
The proof of Claim 4.2 is given next.
Proof of Claim 4.2. We first show that (ZR
that the first i − 1 rounds of AdaptiveGameε have just ended (so the values of ZR
0 , . . . , ZR
already fixed), and that Adversary now picks an element xi to submit in round i of the game.
n ) is a martingale. Fix 1 ≤ i ≤ n, and suppose
i−1 are
1 , . . . , ZR
0 , ZR
If xi /∈ R then AR
i = AR
i−1 and BR
i = BR
i−1 and so ZR
i = ZR
i−1, which trivially means that
E(cid:2)ZR
ZR
0 , . . . , ZR
i
When xi ∈ R, we have
i−1 ; xi /∈ R(cid:3) = ZR
i−1 as desired.
AR
i = AR
i−1 +
1
n
;
BR
i = (
i = (cid:26)
i−1
BR
i−1 + 1
BR
np
if xi is not sampled.
if xi is sampled.
ZR
i−1 − 1/n
if xi is not sampled.
ZR
i−1 + 1/np − 1/n if xi is sampled.
⇒ ZR
Recall that Sampler uses Bernoulli sampling with probability p, that is, xi is sampled with proba-
bility p (regardless of the outcome of the previous rounds). Therefore, we have that
EhZR
i
ZR
0 , . . . , ZR
i−1 ; xi ∈ Ri = ZR
i−1 + p · (
1
np
−
1
n
) + (1 − p) · (−
1
n
) = ZR
i−1.
The analysis of both cases xi /∈ R and xi ∈ R implies that E[ZR
i−1, as desired.
We now turn to prove the other two statements of Claim 4.2. The maximum of the expression
i − ZR
0 , . . . , ZR
ZR
i−1
is zero given the additional assumption that xi /∈ R; assuming that xi ∈ R, the variance satisfies
np , obtained when xi ∈ R. The variance of ZR
i−1 is max{ 1
i given ZR
i−1] = ZR
0 , . . . , ZR
n , 1
np − 1
n } ≤ 1
i ZR
Var(ZR
i
ZR
0 , . . . , ZR
n(cid:19)2
i−1 ; xi ∈ R) = (1 − p) ·(cid:18) 1
+ p ·(cid:18) 1
=
1
n2 (cid:18) 1
p
− 1(cid:19) ≤
1
n2 p
.
ZR
0 , . . . , ZR
n2 p , completing the proof.
1
n(cid:19)2
−
np
i−1) ≤ 1
Combining both cases, we conclude that Var(ZR
i
4.2 The Reservoir Sampling Case
We continue to the proof of the second statement of Lemma 4.1, which considers reservoir sam-
pling. In high level, the proof goes along the same lines, except that we work with a different
martingale. Specifically, for k < i ≤ n we define
AR
i = i · dR(Xi) = R ∩ Xi,
BR
i = i · dR(Si) =
i = BR
ZR
i − AR
i ,
i
k
· R ∩ Si,
15
whereas for i ≤ k we simply define AR
i = BR
definition for i > k; specifically, in view of the definition of BR
k elements appear in the stream, the reservoir simply keeps all of the stream's elements.)
The following claim is the analogue of Claim 4.2 for the setting of reservoir sampling.
i = R ∩ Xi. (This is a natural extension of the
i , note that as long as no more than
Claim 4.3. The sequence (ZR
on ZR
0 , . . . , ZR
i−1 is bounded by i/k, and it always holds that ZR
i − ZR
i−1 ≤ i/k.
0 , ZR
1 , . . . , ZR
n ) is a martingale. Furthermore, the variance of ZR
i conditioned
Proof. We follow the same kind of analysis as in Claim 4.2. Fix i > k (for i ≤ k the claim holds
trivially), and suppose that the first i − 1 rounds have ended, so ZR
i−1 are already fixed.
Denote the next element that the adversary submits by xi. First, it is easy to verify that
0 , . . . , ZR
AR
i = (cid:26) AR
AR
i−1
xi /∈ R
i−1 + 1 xi ∈ R
is determined by three factors:
i requires a more subtle case analysis. Given BR
The calculation of BR
i−1 and xi, the value
of BR
(i) is xi ∈ R or not? (ii) is xi sampled or not? and
i
(iii) conditioning on xi being sampled, does it replace an element from R in the sample, or an
element not in R? We separate the analysis into several cases; in cases where xi is sampled, we
denote the element removed from the sample to make room for xi by ri.
0 , . . . , BR
Case 1: xi /∈ R.
from R are neither added nor removed from the sample. That is, R ∩ Si = R ∩ Si−1. Hence,
In the cases where xi is either not sampled, or sampled but with ri /∈ R, elements
BR
i =
i
k
· R ∩ Si =
i − 1
k
· R ∩ Si−1 +
1
k
· R ∩ Si−1 = BR
i−1 + dR(Si−1),
where the first equality is by definition, and the third equality follows again by definition and
since Si−1 = k for i > k.
It remains to consider the event where xi is sampled and ri ∈ R. The probability that xi is
sampled equals k/i, and conditioning on this occurring, the probability that ri belongs to R is
dR(Si−1), so the above event holds with probability (k/i) · dR(Si−1). In this case, one element
from R is removed from the sample, that is, R ∩ Si = R ∩ Si−1 − 1, and therefore
BR
i =
i
k
· R ∩ Si =
i
k
· R ∩ Si−1 −
i
k
= BR
i−1 + dR(Si−1) −
i
k
.
Thus, conditioned on xi /∈ R, the expectation of BR
i
is
(cid:18)1 −
k
i
· dR(Si−1)(cid:19) ·(cid:16)BR
i−1 + dR(Si−1)(cid:17) +
k
i
· dR(Si−1) ·(cid:18)BR
i−1 + dR(Si−1) −
i
k(cid:19) = BR
i−1.
Since AR
i = AR
i−1 when xi /∈ R, we deduce that
E[ZR
i ZR
0 , . . . , ZR
i−1 ; xi /∈ R] = ZR
i−1.
16
Case 2: xi ∈ R. Similarly, whenever Si = Si−1 we have that BR
i−1 + dR(Si−1). The only case
where this does not hold is when xi is sampled and ri /∈ R, which has probability (k/i) · (1 −
dR(Si−1)). In this case, R ∩ Si = R ∩ Si−1 + 1, implying that
i = BR
BR
i =
i
k
· R ∩ Si =
i
k
· R ∩ Si−1 +
i
k
= BR
i−1 + dR(Si−1) +
i
k
.
Combining these two we get, conditioned on xi ∈ R, that the expectation of BR
i
is
BR
i−1 + dR(Si−1) +(cid:18) k
i
· (1 − dR(Si−1))(cid:19) ·
i
k
= BR
i−1 + 1.
Finally, since AR
i = AR
i−1 + 1 when xi ∈ R, we have that
E[ZR
i ZR
0 , . . . , ZR
i−1 ; xi ∈ R] = ZR
i−1.
The analysis of these two cases implies that (ZR
0 , . . . , ZR
n ) is indeed a martingale.
It remains to obtain the bounds on the difference ZR
i given
ZR
0 , . . . , ZR
i−1. This follows rather easily as a byproduct of the above analysis (and the fact that the
density dR is always bounded between zero and one). When xi /∈ R, we know from the analysis
that AR
i−1 + 1 and
i−1 ≤ BR
BR
i−1 + 1, whereas if xi ∈ R, we have AR
i−1 ≤ i/k.
i−1 + 1 + i/k. In both cases, we conclude that ZR
i−1 and the variance of ZR
i−1 − i/k ≤ BR
i−1 and BR
i = AR
i = AR
i ≤ BR
i ≤ BR
i − ZR
i − ZR
We next bound the variance of ZR
i conditioned on the values of ZR
i−1 (the analysis
also implicitly conditions on the value dR(Si−1); the bound we shall eventually derive holds
regardless of this value). We start with the case that xi /∈ R, and revisit Case 1 above: with
probability (k/i) · dR(Si−1), the value of ZR
is smaller than its expectation by i/k − dR(Si−1); and
i
otherwise (with probability 1 − (k/i) · dR(Si−1)), the value of ZR
is larger than its expectation by
i
dR(Si−1). Thus, we have that
0 , . . . , ZR
Var(ZR
i
=
=
k
i
i
k
ZR
0 , . . . , ZR
i−1, xi /∈ R, dR(Si−1))
· dR(Si−1) ·(cid:18) i
k
− dR(Si−1)(cid:19)2
· dR(Si−1) − (dR(Si−1))2 ≤
i
k
+(cid:18)1 −
k
i
· dR(Si−1)(cid:19) · (dR(Si−1))2
.
We next address the case where xi ∈ R, which correspond to Case 2 above. Here, with probability
(k/i) · (1 − dR (Si−1)), the value of ZR
is larger than its conditional expectation by i/k + dR (Si−1) −
i
1; otherwise, ZR
i
is smaller than the expectation by 1 − dR(Si−1). Thus,
Var(ZR
i
=
=
k
i
i
k
ZR
0 , . . . , ZR
i−1, xi ∈ R, dR(Si−1))
· (1 − dR(Si−1)) ·(cid:18) i
k
+ dR(Si−1) − 1(cid:19)2
· (1 − dR(Si−1)) − (1 − dR(Si−1))2 ≤
i
k
+(cid:18)1 −
k
i
· (1 − dR(Si−1))(cid:19) · (1 − dR(Si−1))2
.
As the conditional variance is always bounded by i/k, the bound remains intact if we remove the
conditioning on the value of dR(Si−1) and the predicate assessing whether xi ∈ R or not. In other
words, Var(ZR
i−1) ≤ i/k, completing the proof.
0 , . . . , ZR
i ZR
17
The proof of the second part of Lemma 4.1 now follows from the last claim.
Proof of Lemma 4.1, reservoir sampling case. Observe that
Pr(dR(X) − dR(S) ≥ ε) = Pr(BR
= Pr(ZR
n − AR
n − ZR
n ≥ εn)
0 ≥ εn).
In view of Claim 4.3, we apply Lemma 3.3 on the martingale ZR = (ZR
i = i/k for any i ≥ k (for i ≤ k, we can set σ2
σ2
i = 0), and M = n/k. We get that
0 , . . . , ZR
n ) with λ = εn,
Pr(ZR
n − ZR
0 ≥ λ) ≤ 2 exp(cid:18)−
= 2 exp(cid:18)−
= 2 exp(cid:18)−
≤ 2 exp(cid:18)−
λ2
i=1 σ2
i + Mλ/3(cid:19)
2 ∑n
ε2n2
2 ∑n
ε2kn2
i=1(i/k) + (n/k) · εn/3(cid:19)
n(n + 1) + εn2/3(cid:19)
2n2 (cid:19) = 2 exp(cid:18)−
2 (cid:19) ,
ε2kn2
ε2k
where the second inequality holds for n ≥ 2. Therefore, it suffices to require k ≥ 2
the bound Pr(dR(X) − dR(S) ≥ ε) ≤ δ.
ε2 ln(cid:0) 2
δ(cid:1) to get
5 An Adaptive Attack on Sampling
In this section, we present our lower bounds. Specifically, we show that the sample size cannot
depend solely on the VC-dimension, but rather that the dependency on the cardinality is neces-
sary. This is done by describing a set system (U, R) with large U and VC-dimension of one,
together with a strategy for the adversary that will make the sampled set unrepresentative with
respect to (U, R). That is, the sampled set will not be an ε-approximation of (U, R) with high
probability. This is in contrast to the static setting where the same sample size suffices to an
ε-approximation with high probability. Moreover, in the case of the BernoulliSample algorithm,
the sampled set under attack is extremely unrepresentative, consisting precisely of the k smallest
elements in the stream (where k is the total sample size at the end of the stream).
Proof of Theorem 1.3. Set the universe to be the well-ordered set U = {1, 2, . . . , N} for an arbitrary
n6 ln n ≤ N ≤ 2n/2 and let R = {[1, b] : b ∈ U}. Clearly, (U, R) has VC-dimension 1. Adversary's
strategy (for both sampling algorithms BernoulliSample and ReservoirSample) is described in Fig-
ure 3.
Let S denote the subsequence of elements sampled by the algorithm BernoulliSample along
the stream. The expected size of S is np ≤ np′, and it follows from the well-known Markov
inequality (see e.g. [AS16], Appendix A) that Pr(S ≥ 2np′) < 1/2 (in fact the probability is
much smaller, by Chernoff inequality, but we will not need the stronger bound). From here on,
we condition on the complementary event: we assume that S < 2np′. The next claim asserts
that for S of this size, Adversary's strategy does not fail, in the sense that it never runs out of
elements (i.e., ai < bi for all i ∈ [n]).
18
The adversarial strategy
1. Set a1 = 1 and b1 = N.
2. Let p′ = max{p, ln n/n}.
3. For i = 1 . . . n do:
(a) Set xi = ⌊ai + (1 − p′)(bi − ai)⌋.
(b) If xi is sampled then set ai+1 = xi and bi+1 = bi.
(c) Otherwise set ai+1 = ai and bi+1 = xi.
4. The final stream is X = x1, . . . , xn.
Figure 3: The description of Adversary's strategy for making the sample unrepresentative.
Claim 5.1. If S < 2np′ then bi − ai ≥ n for any i ∈ [n].
Proof. For any i ∈ [n], set ℓi = bi − ai. We prove by induction that ℓi ≥ n. If xi is sampled, then
we have that ℓi+1 ≥ p′ℓi and otherwise we have that ℓi+1 = (1 − p′)ℓi − 2 ≥ (1 − 2p′)ℓi, where
the inequality follows from the induction assumption. Since S < 2np′, we get that
ℓi ≥ p′S(1 − 2p′)n−S · N
≥ p′S(1 − 2p′)n · N
−(S ln 1
p′ +n ln(
1
1−2p′ ))
· N
−(S ln 1
p′ +3np′)
· N
= e
> e
−(2np′ ln 1
p′ +3np′)
· N
> e
≥ eln n−ln N · N = n ,
where the third inequality holds since 3p ≥ ln(
inequality follows since p′ ≤ ln N
6n ln n and p′ ≥ ln n/n, which means that
ln N ≥ 6np′ ln n ≥ 2np′ ln(1/p′) + 3np′ + ln n.
1
1−2p ) for small enough p > 0, and the last
This proves the induction step, and completes the proof of the claim.
The last claim means that if S < 2np′, then the attack in Figure 3 successfully generates a
stream of n elements. We now show that the sampled set is not an ε-approximation. We begin
by analyzing the BernoulliSample algorithm.
Claim 5.2. Consider Adversary's attack on BernoulliSample described in Figure 3. At round i of the game,
• All elements that were previously submitted by Adversary and sampled are no bigger than ai.
• All elements that were previously submitted but not sampled are no smaller than bi.
19
• The element submitted during round i is between ai and bi.
Proof. By induction, where the base case i = 1 is trivial. Suppose that the claim holds for the
first i − 1 rounds; we now prove it for round i. By definition of the attack, and from Claim 5.1
it holds that ai−1 ≤ ai < bi ≤ bi−1 and so any of the elements xj for j < i − 1 satisfies the
desired condition, by the induction assumption. It remains to address the case where j = i − 1.
If xi−1 was sampled, then the attack sets ai = xi−1, that is, xi−1 is a sampled element and satisfies
xi−1 ≤ ai. Otherwise, the attack sets bi = xi−1 and so xi−1 is a non-sampled element and satisfies
xi−1 ≥ bi. Finally, ai < xi < bi always holds. Thus, the three desired conditions are retained.
As the last claim depicts, all sampled elements are smaller than all non-sampled ones at any
point along the stream. This, of course, suffices for the sampled set to not be an ε-approximation
of (U, R). Denote the sampled set by S, and let s be the maximal element in S (if S is empty,
we are done). Consider now the range [1, s] ∈ R: its density in the sampled set is 1, namely,
d[1,s](S) = 1, while its density in the stream is d[1,s](X) = S/n. To summarize,
d[1,s](S) − d[1,s](X) ≥ 1 − S/n ≥ 1 − 2p′ > 1/2 ≥ ε .
Altogether, the attack does not fail provided that S < 2np′, which holds with probability at least
1/2. Thus, BernoulliSample with parameter p as in the theorem's statement is not (ε, δ)-robust.
The analysis of the ReservoirSample algorithm is very similar. Recall that k denotes the sample
size, and let k′ be the total number of elements that were sampled during the reservoir sampling
process. That is, k′ counts sampled elements that were evicted at a future iteration. We bound k′
as follows. E[k′] = k + ∑n
i=1 k/n ≤ 2k ln n. Again, Markov inequality shows that with probability
at least 1/2, we will have k′ ≤ 4k ln n. Using the previous analysis, we know that all k′ elements
are the smallest elements in the stream. The sample set S consists of some k elements among
these k′ elements (in other words, the sample set is not necessarily the set of k smallest element,
but it is still a subset of the k′ smallest elements). Thus, taking the interval [1, s] where s is
the maximal element among the k′ elements, we have that the density of [1, s] in the sample is
d[1,s](S) = k
k = 1. On the other hand, the density of [1, s] is the stream is
Together, we entail that
d[1,s](X) =
k′
n
≤
4k ln n
n
≤
ln N
n
≤ 1/2.
d[1,s](S) − d[1,s](X) > 1 − 1/2 ≥ ε ,
meaning that ReservoirSample with k as in the statement of the theorem is not (ε, δ)-robust.
6 Continuous Robustness
In this section, we prove that the ReservoirSample algorithm is (ε, δ)-continuous robust against
static and adaptive adversaries. Recall that a sampling algorithm is (ε, δ)-continuously robust
if the following holds with probability at least 1 − δ: at any point throughout the stream, the
current sample held by Sampler is an ε-approximation of the current stream (i.e., of the set of all
elements submitted by Adversary until now).
20
With this definition in hand, BernoulliSample cannot possibly be continuously robust in gen-
eral (even in the static setting)4. We thus restrict our discussion to ReservoirSample from here on,
and turn to the proof of Theorem 1.4. The proof examines O(ε−1 ln n) carefully picked points
along the stream, applying Theorem 1.2 on each of the points. It then shows that if the sam-
ple is a good approximation of the stream at all of these points, then continuous robustness is
guaranteed with high probability.
Proof of Theorem 1.4. We provide the proof for the setting of an adaptive adversary. The proof for
the static setting is essentially identical, with the only difference being that, instead of making
black-box applications of Theorem 1.2, we apply the static analogue of it; Recall that the bound in
the static analogue is of the form Θ(cid:16) d+ln 1/δ
in the statement of Theorem 1.2.
ε2
(cid:17), compared to the Θ(cid:16) ln R+ln 1/δ
ε2
(cid:17) bound appearing
Let (U, R), n, ε, δ be as in the statement of the theorem. As a warmup, let us analyze a
simple yet non-optimal proof based on a naıve union bound. Denote the stream and sample
after i rounds by Xi and Si, respectively. Consider for a moment the first i rounds of the game as
a "standalone" game where the stream length is i. Applying the second part of Theorem 1.2 with
parameters (U, R), ε, δ′, i, where δ′ = δ/n, we get that if the memory size k of ReservoirSample
satisfies
ln R + ln(2n/δ)
ln R + ln(2/δ) + ln n
ε2
then regardless of Adversary's strategy,
k ≥ 2 ·
= 2 ·
ε2
,
(3)
Pr(Si is not an ε-approximation of Xi) ≤ δ/n.
Taking a union bound, the probability that Si is an ε-approximation of Xi for all i ∈ [n] is at
least 1 − n · (δ/n) = 1 − δ. Thus, it follows that ReservoirSample whose parameter k satisfies the
condition of (3) is (ε, δ)-continuously robust.
We now continue to the proof of the improved bound, appearing in the statement of the
theorem. The proof is also based, at its core, on a union bound argument, albeit a more efficient
one. The key idea is to take a sparse set of "checkpoints" i1, . . . , it along the stream, where ij+1 =
(1 + Θ(ε))ij, apply Theorem 1.2 at any of the times i1, . . . , it to make sure the sample is an (ε/2)-
approximation of the stream in any of these times. Finally, we show that with high probability, for
any j ∈ [t − 1], the approximation is preserved (the approximation factor might become slightly
worse, but no worse than ε) in the "gaps" between any couple of such neighboring points.
For this, we first need the following simple claims.
Claim 6.1. Let T, T′ be two sequences of length k over U, which differ in up to v values. Then dR(T) −
dR(T′) ≤ v/k for any R ⊆ U. In particular, if T is an α-approximation of some sequence X ⊇ T, T′,
then T′ is an (α + v/k)-approximation of X.
Proof. For any subset R ⊆ U we have −v ≤ R ∩ T − R ∩ T′ ≤ v. Dividing by k = T = T′,
and recalling that dR(T) = R ∩ T/T and dR(T′) = R ∩ T′/T′, we conclude that −v/k ≤
4To see this, consider any set system (U, R) where R contains a singleton {u} for some u ∈ U, which is the first
element of the stream. With probability 1 − p this element is not sampled and the density of {u} in the sample at
the current point is 0, while its density in the stream is 1. This violates the ε-approximation requirement (unless
p ≥ 1 − δ).
21
dR(T) − dR(T′) ≤ v/k, that is, dR(T) − dR(T′) ≤ v/k. To prove the second part, note that
dR(T′) − dR(X) ≤ dR(T′) − dR(T) + dR(T) − dR(X) ≤ v/k + α
for any R ⊆ U.
Claim 6.2. Suppose that T ⊆ X ⊆ X′ are three sequences over U, where T is an α-approximation of X,
and X′ ≤ (1 + β)X. Then T is an (α + β)-approximation of X′.
Proof. For any subset R ⊆ U, we have that R ∩ X ≤ R ∩ X′ ≤ R ∩ X + βX. We also know
that dR(T) − dR(X) ≤ α, since T is an α-approximation of X. On the one hand, it follows that
On the other hand,
dR(T) ≥ dR(X) − α =
R ∩ X
X
− α ≥
R ∩ X′ − βX
X
− α
≥
R ∩ X′
X′
− β − α = dR(X′) − (α + β).
dR(T) ≤ dR(X) + α =
R ∩ X
X
+ α ≤
R ∩ X′
X′/(1 + β)
+ α
= (1 + β)dR(X′) + α ≤ dR(X′) + (α + β).
As these inequalities hold for any R ⊆ U, the claim follows.
As a consequence of the above two claims, we get the following useful claim. (Recall that for
any i ∈ [n], the sample and stream after i rounds are denoted by Si and Xi, respectively.)
Claim 6.3. Consider ReservoirSample with memory size k, and suppose that exactly v elements were
sampled in rounds l + 1, l + 2, . . . , m of the game, where k ≤ l < m ≤ (1 + β)l.
If Sl is an α-
approximation of Xl, then Sm is an (α + β + v/k)-approximation of Xm.
Proof. By Claim 6.2, Sl is an (α + β)-approximation of Xm. As Sm differs from Sl by at most v
elements, we conclude from Claim 6.1 that Sm is an (α + β + v/k)-approximation of Xm.
The last claim equips us with an approach to ensure continuous robustness, which is more
efficient compared to the simple union bound approach. Suppose that there exists a set of
integers k = i1 < i2 < . . . < it = n satisfying the following for any j ∈ [t − 1].
1. Sij is an α-approximation of Xij, where α = ε/4.
2. ij+1 ≤ (1 + β)ij, where β = ε/4.
3. The number of elements sampled in rounds ij + 1, ij + 2, . . . , ij+1 is bounded by v = εk/2.
We claim that the above three conditions suffice to ensure that Si is an ε-approximation of Xi for
any i ∈ [n]. Indeed, for i ≤ k, Si = Xi is trivially an ε-approximation. When i > k, consider the
maximum j < t for which ij ≤ i, and apply Claim 6.3 with l = ij, m = i, and α, β, v as dictated
above. Since α + β + v/k = ε, the claim implies that Si is an ε-approximation of Xi, as desired.
Specifically, given k satisfying the assumption of Theorem 1.4, we pick i1, i2, . . . , it recursively
as follows: we start with i1 = k; and given ij we set ij+1 ≤ n as the largest integer satisfying that
22
ij+1 ≤ (1 + β)ij = (1 + ε/4)ij. It is not hard to verify that ij = k · (1 + θ(ε))j−1 (this implicitly relies
on the fact that k ≥ 4/ε, ensured by the assumption of the theorem). Note that t = O(ln1+ε n) =
O(ε−1 ln n). We next show that for this choice of i1, . . . , it, the above three conditions are satisfied
simultaneously for all j ∈ [t − 1] with probability at least 1 − δ. This shall conclude the proof.
For the first condition, apply Theorem 1.2 for any j ∈ [t − 1] with parameters (U, R), ε/42, δ′, ij
where δ′ = δ/2t, concluding that if the memory size k satisfies
k ≥ 2 ·
ln R + ln(4t/δ)
(ε/4)2
= Θ(cid:18) ln R + ln(1/δ) + ln(1/ε) + ln ln n
ε2
(cid:19)
then for any j ∈ [t − 1],
Pr(Sij is an ε/2-approximation of Xij) ≥ 1 − δ/2t.
Taking a union bound, with probability at least 1 − δ/2 the first condition holds for all j ∈ [t − 1].
The second condition, regarding the boundedness of ij+1 as a function of ij, holds trivially
(and deterministically) for our choice of i1 ≤ i2 ≤ . . . ≤ it.
Finally, it remains to address the third condition. For any j ∈ [t − 1], let Aj denote the total
number of sampled elements in rounds ij + 1, ij + 2, . . . , ij+1 of the game. Note that each such Aj
is a random variable. We wish to show that
Pr(Aj > εk/2) ≤ δ/2t.
(4)
Indeed, if (4) is true for any j ∈ [t − 1], then the probability that the third condition holds for
any j is at least 1 − δ/2, which (in combination with our analysis of the other two conditions)
completes the proof. Thus, it remains to prove (4).
Recall that the probability of an element to be sampled in round i is exactly k/i, and that
ij+1 ≤ (1 + ε/4)ij. Hence, Aj is a sum of up to ⌊εij/4⌋ independent random variables, each of
which has probability less than k/ij to be sampled. In particular, the mean of Aj is less than
(εij/4) · (k/ij) = εk/4. From Chernoff bound (Theorem 3.1), we get the desired bound:
Pr(Aj ≥ εk/2) < exp(cid:18)−
22 · εk/4
2 + 2 · 2/3(cid:19) ≤ exp(cid:18)−
εk
4 (cid:19) ≤
δ
2t
,
where the last inequality holds for k ≥ c · ε−1(ln 1/δ + ln 1/ε + ln ln n), for a sufficiently large
constant c > 0; note that k in the theorem's statement indeed satisfies this inequality.
Acknowledgments
We are grateful to Moni Naor for suggesting the study of streaming algorithms in the adversarial
setting and for helpful and informative discussions about it. We additionally thank Noga Alon,
Nati Linial, and Ohad Shamir for invaluable comments and suggestions for the paper.
References
[AS16]
Noga Alon and Joel H. Spencer. The Probabilistic Method. Wiley Publishing, 4th edition,
2016.
23
[BCEG07] Amitabha Bagchi, Amitabh Chaudhary, David Eppstein, and Michael T. Goodrich.
Deterministic sampling and range counting in geometric data streams. ACM Transac-
tions on Algorithms, 3(2):16, 2007.
[BCM+13] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndic, Pavel
Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learn-
ing at test time. In Machine Learning and Knowledge Discovery in Databases - European
Conference, ECML PKDD, pages 387 -- 402, 2013.
[BOV15] Vladimir Braverman, Rafail Ostrovsky, and Gregory Vorsanger. Weighted sampling
Information Processing Letters, 115(12):923 --
without replacement from data streams.
926, 2015.
[BR18]
Battista Biggio and Fabio Roli. Wild patterns: Ten years after the rise of adversarial
machine learning. Pattern Recognition, 84:317 -- 331, 2018.
[CBM18] Daniel Cullina, Arjun Nitin Bhagoji, and Prateek Mittal. PAC-learning in the presence
In Proceedings of the 32Nd International Conference on Neural
of evasion adversaries.
Information Processing Systems, NIPS, pages 228 -- 239, 2018.
[CDK+11] Edith Cohen, Nick G. Duffield, Haim Kaplan, Carsten Lund, and Mikkel Thorup. Ef-
ficient stream sampling for variance-optimal estimation of subset sums. SIAM Journal
on Computing, 40(5):1402 -- 1431, 2011.
[CEM+96] Kenneth L. Clarkson, David Eppstein, Gary L. Miller, Carl Sturtivant, and Shang-Hua
Teng. Approximating center points with iterative radon points. International Journal
of Computational Geometry and Applications, 6(3):357 -- 377, 1996.
[CG05]
Graham Cormode and Minos N. Garofalakis. Sketching streams through the net: Dis-
tributed approximate query tracking. In Proceedings of the 31st International Conference
on Very Large Data Bases, VLDB, pages 13 -- 24, 2005.
[CGP+18] Timothy Chu, Yu Gao, Richard Peng, Sushant Sachdeva, Saurabh Sawlani, and Junx-
ing Wang. Graph sparsification, spectral sketches, and faster resistance computation,
In 59th IEEE Annual Symposium on Foundations of
via short cycle decompositions.
Computer Science, FOCS, pages 361 -- 372, 2018.
[Cha01]
[Che52]
Bernard Chazelle. The discrepancy method - randomness and complexity. Cambridge
University Press, 2001.
Herman Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based
on the sum of observations. The Annals of Mathematical Statistics, 23(4):493 -- 507, 1952.
[CJSS03] Charles D. Cranor, Theodore Johnson, Oliver Spatscheck, and Vladislav Shkapenyuk.
The gigascope stream database. IEEE Data Engineering Bulletin, 26(1):27 -- 32, 2003.
[CL06]
Fan Chung and Linyuan Lu. Concentration inequalities and martingale inequalities:
a survey. Internet Mathematics, 3(1):79 -- 127, 2006.
[CMY11] Graham Cormode, S. Muthukrishnan, and Ke Yi. Algorithms for distributed func-
tional monitoring. ACM Transactions on Algorithms, 7(2):21:1 -- 21:20, 2011.
24
[CMYZ12] Graham Cormode, S. Muthukrishnan, Ke Yi, and Qin Zhang. Continuous sampling
from distributed streams. Journal of the ACM, 59:10:1 -- 10:25, 2012.
[CTW16] Yung-Yu Chung, Srikanta Tirthapura, and David P. Woodruff. A simple message-
optimal algorithm for random sampling from a distributed stream. IEEE Transactions
on Knowledge and Data Engineering, 28:1356 -- 1368, 2016.
[DFH+15] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and
Aaron Roth. The reusable holdout: Preserving validity in adaptive data analysis.
Science, 349(6248):636 -- 638, 2015.
[DGS15] Amit Daniely, Alon Gonen, and Shai Shalev-Shwartz. Strongly adaptive online learn-
ing. In Proceedings of the 32nd International Conference on Machine Learning, ICML, pages
1405 -- 1411, 2015.
[DLT05] Nick G. Duffield, Carsten Lund, and Mikkel Thorup. Estimating flow distributions
IEEE/ACM Transactions on Networking, 13(5):933 -- 946,
from sampled flow statistics.
2005.
[ES06]
[Fre75]
Pavlos S. Efraimidis and Paul G. Spirakis. Weighted random sampling with a reser-
voir. Information Processing Letters, 97(5):181 -- 185, 2006.
David A. Freedman. On tail probabilities for martingales. The Annals of Probability,
3(1):100 -- 118, 1975.
[GDG+17] Priya Goyal, Piotr Doll´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo
Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch
sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.
[GHR+12] Anna C. Gilbert, Brett Hemenway, Atri Rudra, Martin J. Strauss, and Mary Wootters.
Recovering simple signals. In Information Theory and Applications Workshop, ITA, pages
382 -- 391, 2012.
[GHS+12] Anna C. Gilbert, Brett Hemenway, Martin J. Strauss, David P. Woodruff, and Mary
In
Wootters. Reusable low-error compressive sampling schemes through privacy.
IEEE Statistical Signal Processing Workshop, SSP, pages 536 -- 539, 2012.
[GK01]
[GK16]
Michael Greenwald and Sanjeev Khanna. Space-efficient online computation of quan-
tile summaries. SIGMOD Record, 30(2):58 -- 66, 2001.
Michael B. Greenwald and Sanjeev Khanna. Quantiles and equi-depth histograms
over streams. In Minos Garofalakis, Johannes Gehrke, and Rajeev Rastogi, editors,
Data Stream Management: Processing High-Speed Data Streams, pages 45 -- 86. Springer
Berlin Heidelberg, 2016.
[GLA16] Mohammed Ghesmoune, Mustapha Lebbah, and Hanene Azzag. State-of-the-art on
clustering data streams. Big Data Analytics, 1(1):13, 2016.
[GMP18]
Ian J. Goodfellow, Patrick D. McDaniel, and Nicolas Papernot. Making machine
learning robust against adversarial inputs. Communications of the ACM, 61(7):56 -- 66,
2018.
25
[HAK07] Elad Hazan, Amit Agarwal, and Satyen Kale. Logarithmic regret algorithms for on-
line convex optimization. Machine Learning, 69(2-3):169 -- 192, 2007.
[Haz16]
Elad Hazan. Introduction to online convex optimization. Foundations and Trends in
Optimization, 2(3-4):157 -- 325, 2016.
[HW13] Moritz Hardt and David P. Woodruff. How robust are linear sketches to adaptive
inputs? In Symposium on Theory of Computing Conference, pages 121 -- 130, 2013.
[JMR05]
[JPA04]
Theodore Johnson, S. Muthukrishnan, and Irina Rozenbaum. Sampling algorithms
in a stream operator. In Proceedings of the ACM SIGMOD International Conference on
Management of Data, pages 1 -- 12, 2005.
Chris Jermaine, Abhijit Pol, and Subramanian Arumugam. Online maintenance of
very large random samples. In Proceedings of the ACM SIGMOD International Confer-
ence on Management of Data, pages 299 -- 310, 2004.
[JSTW19] Rajesh Jayaram, Gokarna Sharma, Srikanta Tirthapura, and David P. Woodruff.
Weighted reservoir sampling from distributed streams. In Proceedings of the 38th ACM
SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, PODS, pages
218 -- 235, 2019.
[KLL16]
Zohar Karnin, Kevin Lang, and Edo Liberty. Optimal quantile approximation in
In 2016 IEEE 57th Annual Symposium on Foundations of Computer Science
streams.
(FOCS), pages 71 -- 78, 2016.
[Knu97]
Donald E. Knuth. The Art of Computer Programming, Volume 2 (3rd Ed.): Seminumerical
Algorithms. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1997.
[LLS01]
Yi Li, Philip M. Long, and Aravind Srinivasan.
complexity of learning. Journal of Computer and System Sciences, 62(3):516 -- 527, 2001.
Improved bounds on the sample
[LMPL18] Thodoris Lykouris, Vahab Mirrokni, and Renato Paes Leme. Stochastic bandits robust
to adversarial corruptions. In Proceedings of the 50th Annual ACM SIGACT Symposium
on Theory of Computing, STOC 2018, pages 114 -- 122, 2018.
[LN93]
Richard J. Lipton and Jeffrey F. Naughton. Clocked adversaries for hashing. Algorith-
mica, 9(3):239 -- 252, 1993.
[McD98] Colin McDiarmid. Concentration. In Michel Habib, Colin McDiarmid, Jorge Ramirez-
Alfonsin, and Bruce Reed, editors, Probabilistic Methods for Algorithmic Discrete Mathe-
matics, pages 195 -- 248. Springer Berlin Heidelberg, 1998.
[MHS19] Omar Montasser, Steve Hanneke, and Nathan Srebro. Vc classes are adversarially
robustly learnable, but only improperly. In Alina Beygelzimer and Daniel Hsu, edi-
tors, Proceedings of the Thirty-Second Conference on Learning Theory, COLT, volume 99 of
Proceedings of Machine Learning Research, pages 2512 -- 2530, 2019.
[MM19]
Saeed Mahloujifar and Mohammad Mahmoody. Can adversarially robust learning
In Algorithmic Learning Theory, ALT, pages 581 --
leverage computational hardness?
609, 2019.
26
[MNS11]
Ilya Mironov, Moni Naor, and Gil Segev. Sketching in adversarial environments.
SIAM Journal on Computing, 40(6):1845 -- 1870, 2011.
[MRL99] Gurmeet Singh Manku, Sridhar Rajagopalan, and Bruce G. Lindsay. Random sam-
pling techniques for space efficient online computation of order statistics of large
datasets. SIGMOD Record, 28(2):251 -- 262, 1999.
[MV17]
Nabil H. Mustafa and Kasturi R. Varadarajan. Epsilon-approximations and epsilon-
nets. In Csaba D. Toth, Joseph O'Rourke, and Jacob E. Goodman, editors, Handbook
of Discrete and Computational Geometry, 3rd Edition, chapter 47, page 27. Chapman and
Hall/CRC, New York, 2017.
[MW18] Michael McCoyd and David A. Wagner. Background class defense against adversarial
examples. In 2018 IEEE Security and Privacy Workshops, SP Workshops, pages 96 -- 102,
2018.
[NY15]
[OV11]
Moni Naor and Eylon Yogev. Bloom filters in adversarial environments. In Advances
in Cryptology - CRYPTO 2015 - 35th Annual Cryptology Conference, pages 565 -- 584, 2015.
M. Tamer Ozsu and Patrick Valduriez. Principles of Distributed Database Systems.
Springer Publishing Company, Incorporated, 3rd edition, 2011.
[PMG+17] Nicolas Papernot, Patrick D. McDaniel, Ian J. Goodfellow, Somesh Jha, Z. Berkay
Celik, and Ananthram Swami. Practical black-box attacks against machine learning.
In ACM Asia Conference on Computer and Communications Security, AsiaCCS, pages 506 --
519, 2017.
[Sau72]
Norbert Sauer. On the density of families of sets. Journal of Combinatorial Theory, Series
A, 13(1):145 -- 147, 1972.
[SBM+18] Chawin Sitawarin, Arjun Nitin Bhagoji, Arsalan Mosenia, Mung Chiang, and Prateek
Mittal. DARTS: deceiving autonomous cars with toxic signs. CoRR, abs/1802.06430,
2018.
[Sha12]
[She72]
Shai Shalev-Shwartz. Online learning and online convex optimization. Foundations
and Trends in Machine Learning, 4(2):107 -- 194, 2012.
Saharon Shelah. A combinatorial problem; stability and order for models and theories
in infinitary languages. Pacific Journal of Mathematics, 41(1):247 -- 261, 1972.
[SKYL17] Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, and Quoc V. Le. Don't decay
the learning rate, increase the batch size, 2017. Published as a conference paper at
ICLR 2018.
[SS17]
[SST10]
Ohad Shamir and Liran Szlak. Online learning with local permutations and delayed
feedback. In Proceedings of the 34th International Conference on Machine Learning, ICML,
pages 3086 -- 3094, 2017.
Nathan Srebro, Karthik Sridharan, and Ambuj Tewari. Smoothness, low noise and
fast rates. In Advances in Neural Information Processing Systems 23: Annual Conference
on Neural Information Processing Systems, NIPS, pages 2199 -- 2207, 2010.
27
[STZ04]
Subhash Suri, Csaba D. T ´oth, and Yunhong Zhou. Range counting over multidimen-
sional data streams. In Proceedings of the Twentieth Annual Symposium on Computational
Geometry, SCG '04, pages 160 -- 169, 2004.
[SZS+14] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan,
Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd
International Conference on Learning Representations, ICLR, 2014.
[Tal94]
[Val84]
[VC71]
Michel Talagrand. Sharper bounds for gaussian and empirical processes. The Annals
of Probability, 22(1):28 -- 76, 1994.
L. G. Valiant. A theory of the learnable. Communications of the ACM, 27(11):1134 -- 1142,
1984.
V. Vapnik and A. Chervonenkis. On the uniform convergence of relative frequencies
of events to their probabilities. Theory of Probability & Its Applications, 16(2):264 -- 280,
1971.
[Vit85]
Jeffrey Scott Vitter. Random sampling with a reservoir. ACM Transactions on Mathe-
matical Software, 11(1):37 -- 57, 1985.
[WFRS18] Blake E. Woodworth, Vitaly Feldman, Saharon Rosset, and Nati Srebro. The ever-
In Advances in Neural Informa-
lasting database: Statistical validity at a fair price.
tion Processing Systems 31: Annual Conference on Neural Information Processing Systems,
NeurIPS, pages 6532 -- 6541, 2018.
[WLYC13] Lu Wang, Ge Luo, Ke Yi, and Graham Cormode. Quantiles over data streams: An
experimental study. In Proceedings of the 2013 ACM SIGMOD International Conference
on Management of Data, pages 737 -- 748, 2013.
[ZLZ18]
Lijun Zhang, Shiyin Lu, and Zhi-Hua Zhou. Adaptive online learning in dynamic en-
vironments. In Advances in Neural Information Processing Systems 31: Annual Conference
on Neural Information Processing Systems, NeurIPS, pages 1330 -- 1340, 2018.
28
|
1511.00648 | 2 | 1511 | 2016-10-19T23:01:30 | Parameterized Algorithms for Constraint Satisfaction Problems Above Average with Global Cardinality Constraints | [
"cs.DS"
] | Given a constraint satisfaction problem (CSP) on $n$ variables, $x_1, x_2, \dots, x_n \in \{\pm 1\}$, and $m$ constraints, a global cardinality constraint has the form of $\sum_{i = 1}^{n} x_i = (1-2p)n$, where $p \in (\Omega(1), 1 - \Omega(1))$ and $pn$ is an integer. Let $AVG$ be the expected number of constraints satisfied by randomly choosing an assignment to $x_1, x_2, \dots, x_n$, complying with the global cardinality constraint. The CSP above average with the global cardinality constraint problem asks whether there is an assignment (complying with the cardinality constraint) that satisfies more than $(AVG+t)$ constraints, where $t$ is an input parameter.
In this paper, we present an algorithm that finds a valid assignment satisfying more than $(AVG+t)$ constraints (if there exists one) in time $(2^{O(t^2)} + n^{O(d)})$. Therefore, the CSP above average with the global cardinality constraint problem is fixed-parameter tractable. | cs.DS | cs | Parameterized Algorithms for Constraint Satisfaction Problems
Above Average with Global Cardinality Constraints
Xue Chen∗
Computer Science Department
University of Texas at Austin
[email protected]
Yuan Zhou
Computer Science Department
Indiana University at Bloomington
[email protected]
August 15, 2018
Abstract
Given a constraint satisfaction problem (CSP) on n variables, x1, x2, . . . , xn ∈ {±1}, and m con-
straints, a global cardinality constraint has the form ofPn
i=1 xi = (1−2p)n, where p ∈ (Ω(1), 1−Ω(1))
and pn is an integer. Let AV G be the expected number of constraints satisfied by randomly choosing an
assignment to x1, x2, . . . , xn, complying with the global cardinality constraint. The CSP above average
with the global cardinality constraint problem asks whether there is an assignment (complying with the
cardinality constraint) that satisfies more than (AV G + t) constraints, where t is an input parameter.
In this paper, we present an algorithm that finds a valid assignment satisfying more than (AV G + t)
constraints (if there exists one) in time (2O(t2) + nO(d)). Therefore, the CSP above average with the
global cardinality constraint problem is fixed-parameter tractable.
6
1
0
2
t
c
O
9
1
]
S
D
.
s
c
[
2
v
8
4
6
0
0
.
1
1
5
1
:
v
i
X
r
a
∗Supported by NSF Grant CCF-1526952.
1 Introduction
In a d-ary constraint satisfaction problem (CSP), we are given a set of boolean variables {x1, x2,··· , xn}
over {±1} and m constraints C1,··· , Cm, where each constraint Ci consists of a predicate on at most d
variables. A constraint is satisfied if and only if the assignment of the related variables is in the predicate of
the constraint. The task is to find an assignment to {x1,··· , xn} so that the greatest (or the least) number
of constraints in {C1,··· , Cm} are satisfied. Simple examples of binary CSPs are the MAXCUT prob-
lem and the MINCUT problem, where each constraint includes 2 variables, and the predicate is always
{(−1, +1), (+1,−1)}. The MAX3SAT problem is an example of a ternary CSP. In a MAX3SAT problem,
each constraint includes 3 variables, and the predicate includes 7 out 8 possible assignments to the 3 vari-
ables. Another classical example of a ternary CSP is called the MAX3XOR problem, where each predicate
is either {(−1,−1,−1), (−1, +1, +1), (+1,−1, +1), (+1, +1,−1)} or
{(+1, +1, +1), (+1,−1,−1), (−1, +1, −1), (−1, −1, +1)}.
Constraint satisfaction problem above average. For each CSP problem, there is a trivial randomized
algorithm which chooses an assignment uniformly at random from all possible assignments. In a seminal
work [21], Hastad showed that for MAX3SAT and MAX3XOR, it is NP-hard to find an assignment sat-
isfying ǫm more constraints than the trivial randomized algorithm (in expectation) for an arbitrarily small
constant ǫ > 0. In other words, there is no non-trivial approximation algorithm (i.e. better approximation
than the trivial randomization) for MAX3SAT or MAX3XOR assuming P6=NP. Recently, this result was
extended to many more CSPs by Chan [7]. Assuming the Unique Games Conjecture [23], Austrin and Mos-
sel [3] provided a sufficient condition for CSPs to admit non-trivial approximation algorithms. Later, Khot,
Tulsiani and Worah [24] gave a complete characterization for these CSPs. Guruswami et al. [15] showed
that all ordering CSPs (a variant of CSP) do not admit non-trivial approximation algorithm under the Unique
Games Conjecture.
Therefore, it is natural to set the expected number of constraints satisfied by the trivial randomized
algorithm as a baseline, namely AV G, and ask for better algorithms. In the constraint satisfaction problem
above average, we are given a CSP instance I and a parameter t, and the goal is to decide whether there is
an assignment satisfying t constraints more than the baseline AV G.
The CSP above average problem has been extensively studied in the parameterized algorithms designing
research community. Gutin et al. [18] showed that the MAX3XOR above average (indeed MAXdXOR for
arbitrary d) is fixed-parameter tractable (FPT), i.e., there is an algorithm that makes the correct decision in
time f (t)poly(n). Later, Alon et al. [1] showed that every CSP above average admits an algorithm with
runtime (cid:16)2O(t2) + O(m)(cid:17), and therefore is fixed-parameter tractable. Later, Crowston et al. [8] improved
the running time of MAXXOR to 2O(t log t) · poly(nm). Recently, Makarychev, Makarychev and Zhou [26]
studied a variant, and showed that the ordering CSP above average is fixed-parameter tractable.
Constraint satisfaction problem with a global cardinality constraint. Given a boolean CSP instance I,
we can impose a global cardinality constraintPn
i=1 xi = (1− 2p)n (we assume that pn is an integer). Such
a constraint is called the bisection constraint if p = 1/2. For example, the MAXBISECTION problem is the
MAXCUT problem with the bisection constraint. Constraint satisfaction problems with global cardinality
constraints are natural generalizations of boolean CSPs. Researchers have been studying approximation
algorithms for CSPs with global cardinality constraints for decades, where the MAXBISECTION problem
[2, 10, 12, 16, 20, 34, 35] and the SMALLSET EXPANSION problem [31 -- 33] are two prominent examples.
1
Adding a global cardinality constraint could strictly enhance the hardness of the problem. The SMALL-
SET EXPANSION problem can be viewed as the MINCUT problem with the cardinality of the selected
subset to be ρV (ρ ∈ (0, 1)). While MINCUT admits a polynomial-time algorithm to find the optimal
solution, we do not know a good approximation algorithm for SMALLSET EXPANSION. Raghavendra and
Steurer [31] suggested that the SMALLSET EXPANSION problem is where the hardness of the notorious
UNIQUEGAMES problem [23] stems from.
For many boolean CSPs, via a simple reduction described in [16,34], if such a CSP does not admit a non-
trivial approximation algorithm, neither does the CSP with the bisection constraint. Therefore, it is natural
to set the performance of the trivial randomized algorithm as the baseline, and ask for better algorithms for
CSPs with a global cardinality constraint. Specifically, given a CSP instance I and the cardinality constraint
Pn
i=1 xi = (1 − 2p)n, we define our baseline AV G to be the expected number of constraints satisfied by
uniform randomly choosing an assignment complying with the cardinality constraint. The task is to decide
whether there is an assignment (again complying with the cardinality constraint), satisfying at least AV G+t
constraints. We call this problem CSP above average with a global cardinality constraint, and our goal is to
design an FPT algorithm for the problem.
Recently, Gutin and Yeo [19] showed that it is possible to decide whether there is an assignment sat-
Zenklusen [27]. However, observe that in the MAXBISECTION problem, the trivial randomized algorithm
isfying more than ⌈m/2 + t⌉ constraints in time(cid:16)2O(t2) + O(m)(cid:17) for the MAXBISECTION problem with
m constraints and n variables. The running time was later improved to (cid:0)2O(t) + O(m)(cid:1) by Mnich and
2(n−1)(cid:17) m constraints in expectation. Therefore, when m ≫ n, our problem
satisfies AV G = (cid:16) 1
MAXBISECTION above average asks more than what was proved in [19, 27]. For the MAXCUT problem
without any global cardinality constraint, Crowston et al. [9] showed that optimizing above the Edwards-
Erdos bound is fixed-parameter tractable, which is comparable to the bound in our work, while our algorithm
2 + 1
outputs a solution strictly satisfying the global cardinality constraintPn
i=1 xi = (1 − 2p)n.
1.1 Our results
In this paper, we show FPT algorithms for boolean CSPs above average with a global cardinality constraint
problem. Our main theorem is stated as follows.
Theorem 1.1 (Informal version of Theorem 4.1 and Theorem 6.1) For any integer constant d and real
constant p0 ∈ (0, 1/2], given a d-ary CSP with n variables and m = nO(1) constraints, a global cardinality
constraintPn
i=1 xi = (1− 2p)n such that p ∈ [p0, 1− p0], and an integer parameter t, there is an algorithm
that runs in time (nO(1) + 2O(t2)) and decides whether there is an assignment complying with the cardinality
constraint to satisfy at least (AV G + t) constraints or not. Here AV G is the expected number of constraints
satisfied by uniform randomly choosing an assignment complying with the cardinality constraint.
One important ingredient in the proof of our main theorem is the 2 → 4 hypercontractivity of low-
degree multilinear polynomials in a correlated probability space. Let Dp be the uniform distribution on all
assignments to the n variables complying with the cardinality constraint Pn
i=1 xi = (1 − 2p)n. We show
the following inequality.
Theorem 1.2 (Informal version of Corollary 4.8 and Corollary 5.2) For any degree d multilinear poly-
nomial f on variables x1, x2, . . . , xn, we have
EDp[f 4] ≤ poly(d) · C d
p · EDp[f 2]2,
2
where the constant Cp = poly( 1
1−p , 1
p ).
The ordinary 2 → 4 hypercontractive inequality (see Section 2.1 for details of the inequality) has wide
applications in computer science, e.g., invariance principles [29], a lower bound on the influence of variables
on Boolean cube [22], and an upper bound on the fourth moment of low degree functions [1,26] (see [30] for
a complete introduction and more applications with the reference therein). The inequality admits an elegant
induction proof, which was first introduced in [28]; and the proof was later extended to different settings (e.g.
to the low-level sum-of-squares proof system [4], and to more general product distributions [26]). All the
previous induction proofs, to the best of our knowledge, rely on the local independence of the variables (i.e.
the independence among every constant-sized subset of random variables). In the 2 → 4 hypercontractive
inequality we prove, however, every pair of the random variables is correlated.
Because of the lack of pair-wise independence, our induction proof (as well as the proof to the main
theorem (Theorem 1.1)) crucially relies on the analysis of the eigenvalues of several nO(d) × nO(d) set-
symmetric matrices. We will introduce more details about this analysis in the next subsection.
Independent work on the 2 → 4 hypercontractive inequality.
Independently, Filmus and Mossel [11]
provided a hypercontractive inequality over Dp based on the log-Sobolev inequality due to Lee and Yau [25].
They utilized the property that harmonic polynomials constitute an orthogonal basis in Dp. In this work,
we use parity functions and their Fourier coefficients to analyze the eigenspaces of VarDp and prove the
hypercontractivity in Dp. Parity functions do not constitute an orthogonal basis in Dp, e.g., the n variables
i=1 xi = (1− 2p)n. However, there is another
important component in the proof of our main theorem -- we need to prove the variance of a random solution
is high if the optimal solution is much above average. Parity functions also play an important role in this
component. More specifically, our analysis relies on the fact that the null space of VarDp is equivalent to
the subspace spanned by the global cardinality constraint and all non-zero eigenvalues of VarDp are Θ(1) in
terms of their Fourier coefficients (respect to parity functions).
are not independent under any global cardinality constraintPn
1.2 Techniques and proof overview
Our high level idea is similar to the framework introduced by Gutin et al. [17]. However, we employ very
different techniques to deal with the fixed-parameter tractability above average under a global cardinality
constraint. Specifically, we extensively use the analysis of the eigenspaces of the association schemes. For
simplicity, we will first use the bisection constraint Pn
i=1 xi = 0 as the global cardinality constraint to
illustrate our main ideas. Then we discuss how to generalize it to a global cardinality constraint specified by
p ∈ (0, 1).
We first review the basic ideas underlying the work [1,26]. Let fI be the degree d multilinear polynomial
counting the number of satisfied constraints in a d-ary CSP instance I and fI(S) be the coefficient of
Qi∈S xi, i.e., the Fourier coefficient of χS =Qi∈S xi. We will also view the multilinear polynomial f as a
≤d(cid:1)o.
vector in the linear space spannχSS ∈(cid:0)[n]
We consider the variance of fI in the uniform distribution of {±1}n, namely Var(fI), and discuss the
following 2 cases.
only on t (i.e. a small kernel of variables).
1. Var(fI) = O(t2). In this case, we would like to reduce the problem to a problem whose size depends
In [1], this was done because the coefficients in fI
fI (S)2 in the uniform distribution. Then
are multiples of 2−d and the variance Var(fI ) = PS6=∅
3
there are at most O(t222d) terms with non-zero coefficients in fI, so the size of the kernel is at most
O(d · t222d). One can simply enumerate all possible assignments to the kernel.
2. Var(fI) = Ω(t2). We will claim that the optimal value is at least AV G+t. This is done via the 2 → 4
hypercontractive inequality (Theorem 1.2), which shows that the low-degree polynomial fI is smooth
under the uniform distribution over all valid assignments. Therefore fI is greater than its expectation
plus its standard deviation with positive probability, via the fourth moment method [1, 5, 30].
example -- the value of a bisection in the complete graph with n vertices is always n
Now we take a detailed look at the first case. For the uniform distribution D on all assignments comply-
fI(S)2. Indeed, VarD(fI) might
ing with the bisection constraint, we no longer have VarD(fI) = PS6=∅
fI(S)2. Let us consider a complete graph in the MAXBISECTION problem for
be very different fromPS6=∅
2 in D, which indicates
fI(S)2 is Ω(n2). Another example for the MAXBISECTION problem
that VarD(fI) = 0. However,PS6=∅
is a star graph, i.e., a graph with (n− 1) edges connecting an arbitrary vertex i to the rest of (n− 1) vertices.
The variance for this instance is also 0. However, one can check thatPS6=∅
To solve the problem above, we first observe that in the star graph instance, the degree 2 polynomial
fI = n
2 in the support of D.
This is becausePi xi = 0 is a requirement for all valid assignments in D. In general, we have (Pi xi)h = 0
for all polynomials h under the bisection constraint. Therefore,
2 (say the center of the star is at Vertex 1), which always equals n
fI(S)2 = Ω(n).
2 · n
2 − (Pi xi) x1
S =((Xi
xi)h + c : h a degree-at-most (d − 1) multilinear polynomial, c constant)
constitutes a linear subspace where all the functions in the subspace are equivalent to a constant function
≤d(cid:1) ×(cid:0)[n]
multilinear polynomials f (this can be done because VarD(f ) is just a quadratic polynomial on the Fourier
≤d(cid:1) matrix so that f T Bf = VarD(f ) for all degree-at-most-d
≤d(cid:1)o). From the discussion above, we know that S is in the nullspace of B.
in the support of D. Let B be a (cid:0)[n]
coefficientsn f (S)S ∈(cid:0)[n]
Projection onto the orthogonal complement of S, and the spectral analysis of B via association schemes.
Somewhat surprisingly, we show that S indeed coincides with the nullspace of B, and the eigenvalues of
B of the orthogonal complement of S are Θ(1) (when we treat d as a constant). Once we have this, we
let (Pi xi)hf be the projection of (fI − fI(∅)) onto S = spann1, (Pi xi)χSS ∈(cid:0) [n]
≤d−1(cid:1)o, and have
that the 2-norm2 of the projection onto the orthogonal complement of S, kfI − fI(∅) − (Pi xi)hfk2
Θ(VarD(fI)).
the value of BS,T (where S, T ∈ (cid:0) n
≤d(cid:1)) only depends on S, T, and S ∩ T. If we only consider the
homogeneous submatrices of B (i.e. the submatrix of all rows S and columns T so that S = T = d′ for
some 0 ≤ d′ ≤ d), their eigenvalues and eigenspaces are well understood and easy to characterize as they
correspond to a special association shceme called the Johnson scheme [13]. Then we follow the approach
of [14] to provide a self-contained description of the eigenspaces and eigenvalues of B, which extensively
To analyze the eigenvalues and eigenspaces of B, we crucially use the fact that B is set-symmetric, i.e.
2 =
use the property Pi xi = 0 in the support of D. We will finally show that all eigenvalues of B in the
orthogonal complement of S are between .5 and d.
and eigenvectors of A in a similar fashion, for the purpose of our proof.
We will also let A be the matrix corresponding to the quadratic form ED[f 2] and analyze the eigenvalues
4
follow the approach of [1].
Rounding procedure for the projection vector hf . We would like to use the argument that since kfI −
fI(∅)− (Pi xi)hfk2
2 = Θ(VarD(fI)) = O(t2) is small, there are not so many non-zero Fourier coefficients
for the function (fI − (Pi xi)hf ) (which is equivalent to fI on all valid assignments). Therefore we could
reduce the whole problem to a small kernel. However, since hf might not have any integrality property,
we cannot directly say that the Fourier coefficients of (fI − fI(∅) − (Pi xi)hf ) are multiples of 2−d, and
To solve this problem, we need an extra rounding process. We "round" the projection vector hf to
a new vector h so that the coefficients of h are multiples of 1/Γ (for some big constant Γ which only
depends on d). At the same time, the rounding process guarantees that kfI − fI(∅) − (Pi xi)hk2
2 ≤
O(kfI − fI(∅)− (Pi xi)hfk2
2), so that we can use the argument of [1] with (fI − (Pi xi)h) (which is also
The rounding process proceeds in a iterative manner. At each iteration, we only round the degree-d′
homogeneous part of hf (starting from d′ = d − 1, and decrease d′ after each iteration). We would like
2 is bounded by O(kfI − fI(∅) − (Pi xi)hfk2
to prove the rounding error k(Pi xi)h − (Pi xi)hfk2
2),
so that kfI − fI(∅) − (Pi xi)hk2
2 + k(Pi xi)h − (Pi xi)hfk2
2) ≤
O(kfI − fI(∅) − (Pi xi)hfk2
2). We will first claim that the coefficients of hf are "quite" close to those
of h, where, however, such closeness does not directly guarantee that the rounding error is well bounded.
Then, using this closeness, we will bound the rounding error via a different argument.
2 ≤ 2(kfI − fI(∅) − (Pi xi)hfk2
equivalent to f on all valid assignments) instead.
The 2 → 4 hypercontractive inequality for low-degree multilinear polynomials under D. For the
second case, we need to prove the 2 → 4 hypercontractive inequality for low-degree multilinear polynomials
in D, i.e. ED[f 4] = O(1) ED[f 2]2 for any low-degree multilinear polynomial f . We will use the special
property of the bisection constraint and reduce the task to the ordinary 2 → 4 hypercontractive inequality
(on the uniform production distribution).
Specifically, we view the sampling process in D as first uniformly sampling a perfect matching M
among the n variables then assigning +1 and −1 to the two vertices in each pair of matched variables
independently. Observe that once we have the perfect matching, the second sampling step is a product
distribution over n/2 unbiased coins. Let fM be the function on n/2 variables so that each pair of matched
variables always take opposite values, we have ED[f 4] = EM E[f 4
M ], where the second expectation is over
the uniform distribution.
Now we can directly apply the ordinary hypercontractive inequality to E[f 4
M ])2, it remains to show that EM E[f 2
O(1) EM E[f 2
This final step, can be viewed as proving the "1 → 2 hypercontractivity of E[f 2
lyzing the Fourier coefficients of f .
M ]2. Since ED[f 2]2 = (EM E[f 2
M ]". We prove this by ana-
M ] and have ED[f 4] =
M ]2 = O(1)(EM E[f 2
M ])2.
Generalization to the general cardinality constraint Pi xi = (1 − 2p)n via random restriction. We
now discuss how to generalize the parameterized algorithm to the general cardinality constraint Pi xi =
(1−2p)n for any p ∈ [p0, 1−p0] (where p0 > 0 is a constant). Let Dp be the uniform distribution on all valid
assignments complying with the global cardinality constraint. We first focus on Case 1: VarDp(fI ) = O(t2).
There are two natural approaches to generalize our previous argument to Dp. The first idea is to define
φi =pp/(1 − p) when xi = 1 and φi = −p(1 − p)/p when xi = −1 so that E φi = 0 and E φ2
i = 1, and
do the analysis for g(φ1, . . . , φn) = fI(x1, . . . , xn). The second idea is to work with xi = ±1, and try to
generalize the analysis. However, we adopt neither of the two approaches because of the following reasons.
For the first idea, indeed one can prove that there exists some polynomial hg, so that kg − g(∅) −
2 = Θ(VarDp(fI)). However, since the Fourier coefficients of g have factors such as p1/2, p, p3/2,
(Pi φi)hgk2
5
and fix these variables to 1. For any valid assignment, we see that the remaining .98n variables satisfy the
when we round hg to h. For the second idea, the super-constant on the right-hand-side of the constraint
Our final approach for the general cardinality constraint is via a reduction to the existing algorithm for
the bisection constraint, using random restriction. Let us assume p = .49 to illustrate the high-level idea
etc., it is not clear why the Fourier coefficients of (g − g(∅)− (Pi φi)h) are multiples of some number even
Pi xi = (1 − 2p)n imposes a technical difficulty for bounding the rounding error.
(and it works for any p). Given the constraintPi xi = .02n, we randomly choose a set Q of .02n variables
bisection constraintPi∈Q xi = 0. Let fQ be the function on the remaining .98n variables derived by fixing
the variables in Q to 1 for fI, i.e. let fQ(xQ) = fI(xQ, xQ = ~1), where xQ means {xi : i ∈ Q} and xQ
means {xi : i ∈ Q}.
Since sampling x from Dp is equivalent to sampling Q first then sampling xQ ∼ D and setting xQ = ~1,
we have EQ VarD[fQ] ≤ VarDp[fI] ≤ O(t2). Therefore, for most Q's, we have VarD[fQ] ≤ O(t2). Using
our parameterized algorithm for the bisection constraint, we know that for most Q's, fQ depends on merely
O(t2) variables (on all assignments complying with the bisection constraint). Then we manage to show that
to make this happen, a low-degree polynomial f itself has to depend on only O(t2) variables (on all valid
assignments in Dp).
The 2 → 4 hypercontractive inequality under distribution Dp. To deal with Case 2, we need the 2 → 4
hypercontractive inequality for low-degree multilinear polynomials in Dp. We let φi = pp/(1 − p) when
xi = 1 and φi = −p(1 − p)/p when xi = −1 so that E φi = 0, E φ2
i = 1, andPi φi = 0 in the support
of Dp then prove the inequality for multilinear polynomials on φ1, φ2, . . . , φn. We follow the paradigm
introduced in [28] and apply induction on the number of variables and the degree of the polynomial. For
any multilinear polynomial f , we write it in the form of f = φ1h0 + h1 (say the first variable f depends on
is φ1), expand EDp[f 4] = EDp[(φ1h0 + h1)4], and control each term using induction hypothesis separately.
0h1 = 0] because of the lack
of independence. While the latter one is easy to deal with, the main technical difficulty comes from the
first term, namely upper-bounding EDp[φ1h0h3
1] in terms of kh0k2 and kh1k2. We reduce this problem to
upper-bounding the spectral norm of a carefully designed set-symmetric matrix L, which again utilizes the
analysis developed in Section 3.
Unlike the proof in [28], we no longer have EDp[φ1h0h3
1] = 0 or EDp[φ3
1h3
1.3 Organization
This paper is organized as follows. We introduce some preliminaries in Section 2. In Section 3, we provide a
self-contained description of the eigenspaces and eigenvalues of the matrices corresponding to EDp[f 2] and
VarDp(f ). In Section 4, we prove the fixed-parameter tractability of CSPs above average with the bisection
constraint, and the 2 → 4 hypercontractive inequality under the distribution conditioned on the bisection
constraint. In Section 5, we prove the more general 2 → 4 hypercontractive inequality under distribution Dp.
At last, we show the fixed-parameter tractability of any d-ary CSP above average with a global cardinality
constraint in Section 6.
2 Preliminaries
Let [n] = {1, 2,··· , n}. For convenience, we always use (cid:0)[n]
in [n] and (cid:0)[n]
d(cid:1) to denote the set of all subsets of size d
≤d(cid:1) to denote the set of all subsets of size at most d in [n] (including ∅). For two subsets S
6
indicator variable of an event E, i.e. 1E = 1 when E is true, and 1E = 0 otherwise.
and T , we use S∆T to denote the symmetric difference of S and T . Let n! denote the productQn
i=1 i and
[ n−1
2 ]
i=0 (n − 2i). We use ~0 (~1 resp.) to denote the all 0 (1 resp.) vector. We also use 1E to denote the
n!! = Q
In this work, we only consider f : {±1}n → R. Let U denote the uniform distribution on {±1}n and
Up denote the biased product distribution on {±1}n such that each bit equals to −1 with probability p and
equals to 1 with probability 1 − p. For a distribution V , we use supp(V ) to denote the support of V .
For a random variable X with standard deviation σ, it is known that the fourth moment is necessary and
sufficient to guarantee that there exists x ∈ supp(X) greater than E[X] + Ω(σ) from [1,5,30]. We state this
result as follows.
Lemma 2.1 Let X be a real random variable. Suppose that E[X] = 0, E[X 2] = σ2 > 0, and E[X 4] < bσ4
for some b > 0. Then Pr[X ≥ σ/(2√b)] > 0.
A global cardinality constraint defined by a parameter 0 < p < 1 is Pi∈[n] xi = (1 − 2p)n, which
indicates that (1 − p) fraction of xi's are 1 and the rest p fraction are −1. For convenience, we call the
global cardinality constraint with p = 1/2 as the bisection constraint. In this paper, we always use D to
denote the uniform distribution on all assignments to the n variables complying with the bisection constraint
i=1 xi = 0 and Dp to denote the uniform distribution on all assignments complying with the cardinality
Pn
constraintPn
i=1 xi = (1 − 2p)n.
2.1 Basics of Fourier Analysis of Boolean functions
We state several basic properties of the Fourier transform for Boolean functions those will be useful in this
instead of 1 − p.
work. We follow the notations in [30] except for q = 2p−1√p(1−p)
We first introduce the standard Fourier transform in {±1}n. We will also use the p-biased Fourier
transform in several proofs especially for the 2 → 4 hypercontractive inequality under Dp. More specifically,
in Section 2.2, Section 3, and Section 5, we will use the Fourier transform with the p-biased basis {φS}. In
Section 4 and Section 6, we will use the standard Fourier transform in the basis {χS}.
For the uniform distribution U, we define the inner-product on a pair of functions f, g : {±1}n → R
by hf, gi = Ex∼U [f (x)g(x)]. Hence χS(x) = Qi∈S xi over all subsets S ⊆ [n] constitute an orthonormal
basis for the functions from {±1}n to R. For simplicity, we abuse the notation by writing χS instead of
f (S)χS,
χS(x). Hence every Boolean function has a unique multilinear polynomial expression f =PS⊆[n]
where f (S) = hf, χSi is the coefficient of χS in f . In particular, f (∅) = Ex∼U [f (x)]. An important fact
f (S)2 = Ex∼U [f (x)2], which indicates VarU (f ) =
about Fourier coefficients is Parseval's identity, i.e.,PS
PS6=∅
Given any Boolean function f , we define its degree to be the largest size of S with non-zero Fourier
coefficient f (S). In this work, we focus on the multilinear polynomials f with degree-at-most d. We use
the Fourier coefficients of weight i to denote all Fourier coefficients { f (S)S ∈ (cid:0)[n]
i (cid:1)} of size i character
space span{χSS ∈(cid:0)[n]
≤d(cid:1)}, where each coordinate corresponds to a character function χS of a subset S.
We state the standard Bonami Lemma for Bernoulli ±1 random variables [6, 30], which is also known
as the 2 → 4 hypercontractivity for low-degree multilinear polynomials.
Lemma 2.2 Let f : {−1, 1}n → R be a degree-at-most d multilinear polynomial. Let X1,··· , Xn be
independent unbiased ±1-Bernoulli variables. Then
functions. For a degree-at-most d polynomial f , we abuse the notation f to denote a vector in the linear
f (S)2.
E[f (X1,··· , Xn)4] ≤ 9d · E[f (X1,··· , Xn)2]2.
7
1−p 1xi=1−q 1−p
f (S)2. We state two facts of φi that will be useful in the later section.
For the p-biased distribution Up, we define the inner product on pairs of function f, g : {±1}n → R by
p 1xi=−1 and φS(x) =Qi∈S φi(x).
We abuse the notation by writing φS instead of φS(x). It is straightforward to verify EUp[φi] = 0 and
EUp[φ2
i ] = 1. Notice that φSφT 6= φS∆T unlike χSχS = χS∆T for all x. However, hφS , φTi = 0 for
f (S)φS (x), where
f (S)2 = EUp[f 2], which
hf, gi = Ex∼Up[f (x)g(x)]. Then we define φi(x) =q p
different S and T under Up. Thus we have the biased Fourier expansion f (x) =PS⊆[n]
f (S) = hf, φSi in Up. We also have f (∅) = EUp[f ] and Parseval's identity PS
demonstrates VarUp(f ) =PS6=∅
1. xi = 2pp(1 − p) · φi + 1 − 2p. HencePi φi(x) = 0 for any x satisfyingPi xi = (1 − 2p)n.
i = q · φi + 1 for q = 2p−1√p(1−p)
Observe that the largest size of T with non-zero Fourier coefficient f (T ) in the basis {φSS ∈ (cid:0)[n]
≤d(cid:1)}
is equivalent to the degree of f defined in {χSS ∈ (cid:0)[n]
≤d(cid:1)}. Hence we still define the degree of f to be
maxS: f (S)6=0 S. We abuse the notation f to denote a vector in the linear space span{φSS ∈(cid:0)[n]
≤d(cid:1)}.
1−p + (1−p)2
i ] = p2
p ≥ 1. Therefore we state the 2 → 4
hypercontractivity in the biased distribution Up as follows.
Lemma 2.3 Let f : {−1, 1}n → R be a degree-at-most d multilinear polynomial of φ1,··· , φn. Then
. Thus we always write f as a multilinear polynomial of φi.
For the biased distribution Up, we know EUp[φ4
2. φ2
EUp[f (X1,··· , Xn)4] ≤(cid:18)9 ·
p2
1 − p
(1 − p)2
p
+ 9 ·
(cid:19)d
· EUp[f (X1,··· , Xn)2]2.
At last, we notice that the definition of φS is consistent with the definition of χS when p = 1/2. When
the distribution Up is fixed and clear, we use kfk2 = Ex∼Up[f (x)2]1/2 to denote the L2 norm of a Boolean
f (S)2)1/2. From the Cauchy-Schwarz inequality,
function f . From Parseval's identity, kfk2 is also (PS
one useful property is kf gk2 ≤ kf 2k1/2
2.2 Distributions conditioned on global cardinality constraints
2 kg2k1/2
2
.
as a constant and hide it in the big-Oh notation.
We will study the expectation and the variance of a low-degree multilinear polynomial f in Dp. Because φS
is consistent with χS when p = 1/2, we fix the basis to be φS of the p-biased Fourier transform. Because
p ∈ [p0, 1 − p0], we treat q = 2p−1√p(1−p)
We first discuss the expectation of f under Dp. Because EDp[φS ] is not necessary 0 for any non-
empty subset S, EDp[f ] 6= f (∅). Let δS = EDp[φS]. From symmetry, δS = δS′ for any S and S′ with
the same size. For convenience, we use δk = δS for any S ∈ (cid:0)[n]
k(cid:1). From the definition of δ, we have
EDp[f ] =PS
For p = 1/2 and D, δk = 0 for all odd k and δk = (−1)k/2
(n−1)·(n−3)···(n−k+1) for even k. We
calculate it this way: pick any T ∈(cid:0)[n]
k(cid:1) and consider ED[(Pi xi)χT ] = 0. This indicates
k · δk−1 + (n − k)δk+1 = 0.
f (S) · δS.
(k−1)!!
From δ0 = 1 and δ1 = 0, we could obtain δk for every k > 1.
8
Because δ2i−2 = (−1)i−1Θ(n−i+1) and δ2i−1 = (−1)iO(n−i), the major term of δ2i is determined by
δ2i−2. We choose k = 2i − 1 in the equation (1) to obtain
+ O(
(2i − 1)!!
δ2i = (−1)i
(n − 1)(n − 3)··· (n − 2i + 1)
1
ni+1 ) = (−1)i (2i − 1)!!
ni
+ O(
1
ni+1 ).
For p 6= 1/2 and Dp under the global cardinality constraintPi∈n xi = (1− 2p)n, we consider EDp[φS],
becausePi∈n xi = (1 − 2p)n indicatesPi φi = 0. Thus we use δS = EDp[φS] and calculate it as follows:
pick any T ∈(cid:0)[n]
k(cid:1) and consider EDp[(Pi φi)φT ] = 0. φiφT = φT∪i for i /∈ T ; and φiφT = q · φT + φT\i
for i ∈ T from the fact φ2
i = q · φi + 1. We have
k · δk−1 + k · q · δk + (n − k)δk+1 = 0
(1)
Remark 2.4 For p = 1/2 and the bisection constraint, q = 0 and the recurrence relation becomes k ·
δk−1 + (n − k)δk+1 = 0, which is consistent with the above characterization. Thus we abuse the notation
δk when Up is fixed and clear.
From δ0 = 1, δ1 = 0, and the relation above, we can determine δk for every k. For example, δ2 = − 1
n−1
2q
n−2 · 2 · q =
and δ3 = − δ2
Claim 2.5 For any i ≥ 1, δ2i−1 = (−1)iO(n−i) and δ2i = (−1)i (2i−1)!!
Proof. We use induction on i. Base Case: δ0 = 1 and δ1 = 0.
(n−1)(n−2) . We bound δi as follows:
ni + O(n−i−1).
At the same time, from δ2i and δ2i−1,
δ2i+1 = (−1)i+1q ·
(2i)(2i − 1)!! + (2i)(2i − 2)(2i − 3)!! + ··· + (2i)!!
1
ni+2 ).
⊓⊔
Now we turn to EDp[f 2] and VarDp[f ] for a degree-at-most-d multilinear polynomial f . From the
1
ni+2 ) = (−1)i+1O(
+ O(
ni+1
f (S)φS,
f (S) f (T )δS∆T , VarDp(f ) = EDp[f 2] − EDp[f ]2 =XS,T
≤d(cid:1) ×(cid:0) n
definition and the Fourier transform f =PS
f (S) f (T )(δS∆T − δSδT ).
EDp[f 2] =XS,T
We associate a (cid:0) n
≤d(cid:1) matrix A with EDp[f 2] that A(S, T ) = δS∆T . Hence EDp[f 2] = f T · A · f
from the definition when we think f is a vector in the linear space of span{φSS ∈(cid:0)[n]
≤d(cid:1)}.
Similarly, we associate a(cid:0) n
≤d(cid:1) matrix B with VarDp(f ) that B(S, T ) = δS∆T − δS · δT . Hence
VarDp(f ) = f T · B · f . Notice that an entry (S, T ) in A and B only depends on the size of S, T, and S ∩ T .
Remark 2.6 Because B(∅, S) = B(S,∅) = 0 for any S and VarDp(f ) is independent with f (∅), we could
neglect f (∅) in B such that B is a ((cid:0)[n]
1(cid:1)) matrix. f (∅) is the only difference
between the analysis of eigenvalues in A and B. Actually, the difference δS · δT between A(S, T ) and
B(S, T ) will not effect the analysis of their eigenvalues except the eigenvalue induced by f (∅).
In Section 3, we study the eigenvalues of EDp[f 2] and VarDp(f ) in the linear space span{φSS ∈ (cid:0)[n]
≤d(cid:1)},
d(cid:1) + ···(cid:0)[n]
d(cid:1) + ···(cid:0)[n]
i.e., the eigenvalues of A and B.
1(cid:1)) × ((cid:0)[n]
≤d(cid:1) ×(cid:0) n
9
2.3 Eigenspaces in the Johnson Schemes
We shall use a few characterizations about the eigenspaces of the Johnson scheme to analyze the eigenspaces
and eigenvalues of A and B in Section 3 (please see [13] for a complete introduction).
r )×([n]
We divide A into (d + 1) × (d + 1) submatrices where Ai,j is the matrix of A(S, T ) over all S ∈ (cid:0)[n]
i (cid:1)
and T ∈ (cid:0)[n]
j (cid:1). For each diagonal matrix Ai,i, observe that Ai,i(S, T ) only depends on S ∩ T because of
S = T = i, which indicates Ai,i is in the association schemes, in particular, Johnson scheme.
Definition 2.7 A matrix M ∈ R([n]
r ) is set-symmetric if for every S, T ∈(cid:0)[n]
r(cid:1), M (S, T ) depends only
on the size of S ∩ T.
For n, r ≤ n/2, let Jr ⊆ R([n]
r ) be the subspace of all set-symmetric matrices. Jr is called the
Johnson scheme.
Let M ∈ R([n]
r ) as a homogeneous
f (T )φT , where each coordinate corresponds to a r-subset. Although
degree r polynomial f = PT∈([n]
the eigenvalues of M depend on the entries of M, the eigenspaces of M are independent with M as long as
M is in the Johnson scheme.
r ) be a matrix in the Johnson scheme Jr. We treat a vector in R([n]
r )×([n]
r )×([n]
r )
V0).
f (S).
properties:
i(cid:1) −(cid:0) n
i−1(cid:1);
i (cid:1), although M and f only
i (cid:1)} with the following two
i (cid:1)} satisfies thatPj /∈T ′ f (T ′ ∪ j) = 0 (neglect this property for
Fact 2.8 There are r + 1 eigenspaces V0, V1,··· , Vr in M. For i ∈ [r], the dimension of Vi is(cid:0)n
and the dimension of V0 is 1. We define Vi through f (S) over all S ∈ (cid:0)[n]
r(cid:1)}. Vi is the linear space spanned by { f (S)φSS ∈(cid:0)[n]
depend on { f (T )T ∈(cid:0)[n]
i−1(cid:1), { f (S)S ∈ (cid:0)[n]
1. For any T ′ ∈ (cid:0) [n]
r(cid:1), f (T ) =PS∈(T
2. For any T ∈(cid:0)[n]
It is straightforward to verify that the dimension of Vi is(cid:0)n
that the homogeneous degree i polynomialPS∈([n]
Claim 2.9 For any j ≤ r and any S ∈(cid:0)[n]
<j(cid:1),PT∈([n]
Base Case S = j − 1: from the definition of f ,PT :S⊂T
SupposePT :S⊂T
j ):S⊂T
Proof. We use induction on the size of S to show it is true.
To show the orthogonality between Vi and Vj, it is enough to prove that
f (T ) = 0 for any S ∈(cid:0) [n]
i−1(cid:1) and Vi is an eigenspace in M. Notice
f (S)φS is an eigenvector of matrices in Ji.
i(cid:1) −(cid:0) n
f (T ) = 0 for any f ∈ Vj.
f (T ) = 0.
i )
i )
k+1(cid:1). We prove it is true for any S ∈(cid:0)[n]
k(cid:1):
j − SXi /∈S XT :(S∪i)⊂T
f (T ) = 0.
1
f (T ) =
XT :S⊂T
⊓⊔
10
2.4 CSPs with a global cardinality constraint
In this work, we consider the constraint satisfaction problem on {−1, 1}n with a global cardinality con-
straint. We allow different constraints using different predicates. Because we can add dummy variables in
each constraint, we assume the number of variables in each constraint is d for simplicity.
Definition 2.10 An instance I of a constraint satisfaction problem of arity d consists of a set of variables
V = {x1,··· , xn} and a set of m constraints C1,··· , Cm. Each constraint Ci consists of d variables
xi1,··· , xid and a predicate Pi ⊆ {−1, 1}d. An assignment on xi1,··· , xid satisfies Ci if and only if
(xi1,··· , xid) ∈ Pi. The value valI (α) of an assignment α is the number of constraints in C1,··· , Cm that
are satisfied by α. The goal of the problem is to find an assignment with maximum possible value.
An instance I of a constraint satisfaction problem with a global cardinality constraint consists of an
instance I of a CSP and a global cardinality constraint Pi∈[n] xi = (1 − 2p)n specified by a parameter
p. The goal of the problem is to find an assignment of maximum possible value complying with the global
cardinality constraintPi∈[n] xi = (1 − 2p)n. We denote the value of the optimal assignment by
OP T =
max
α:Pi αi=(1−2p)n
valI(α).
The average value AV G of I is the expected value of an assignment chosen uniformly at random among all
assignments complying the global cardinality constraint
AV G = Eα:Pi αi=(1−2p)n[valI (α)].
Given an instance I of a constraint satisfaction problem of arity d, we associate a degree-at-most d multi-
linear polynomial fI with I such that fI(α) = valI (α) for any α ∈ {±1}n as in [1].
fI(x) = Xi∈[m] Xσ∈Pi Qj∈[d](1 + σj · xi,j)
2d
.
Remark 2.11 The degree of fI is at most d; and the coefficients of fI in the standard basis {χSS ∈(cid:0)[n]
≤d(cid:1)}
are always multiples of 2−d.
Thus we focus on the study of degree-d polynomial f with coefficients of multiples of 2−d instead of the m
constraints C1,··· , Cm. From the discussion above, given an instance I and a global cardinality constraint
Pi∈n xi = (1 − 2p)n, the expectation of I under the global cardinality constraint is different than its
expectation in the uniform distribution, even for CSPs of arity 2 in the bisection constraint:
AV G = ED[fI] =XS
f (S) ED[χS] = f (∅) +XS6=∅
f (S)δS .
Definition 2.12 In the satisfiability above Average Problem, we are given an instance of a CSP of arity
d, a global cardinality constraint Pi∈n xi = (1 − 2p)n, and a parameter t. We need to decide whether
OP T ≥ AV G + t or not.
In this work, we show that it is fixed-parameter tractable. Namely, given a parameter t and an instance of a
CSP problem of arity d under a global cardinality constraintPi∈n xi = (1 − 2p)n, we design an algorithm
that either finds a kernel on O(t2) variables or certifies that OP T ≥ AV G + t.
11
3 Eigenspaces and Eigenvalues of EDp[f 2] and VarDp(f )
In this section we analyze the eigenvalues and eigenspaces of A and B, following the approach of Grigoriev
[14].
(d + 1) submatrices where we know the eigenspaces of the diagonal submatrices from the Johnson scheme,
eigenspaces of these diagonal matrices characterized in Section 2.3, which is motivated by Grigoriev [14].
We will focus on the analysis of A in most time and discuss about B in the end of this section.
We fix any p ∈ (0, 1) with the global cardinality constraint Pi xi = (1 − 2p)n and use the p-biased
Fourier transform in this section, i.e., {φSS ∈ (cid:0)[n]
≤d(cid:1)}. Because χS is consistent with φS for p = 1/2, it is
enough to study the eigenspaces of A and B in span{φSS ∈(cid:0)[n]
≤d(cid:1)}. Since A can be divided into (d + 1)×
we study the eigenspaces of A through the global cardinality constraintPi φi = 0 and the relations between
We first show the eigenspace V ′null with an eigenvalue 0 in A, i.e., the null space of A. BecausePi xi =
(1 − 2p)n in the support of Dp,Pi φi(x) = 0 for any x in the support of Dp. Thus (Pi φi)h = 0 for all
polynomial h of degree-at-most d − 1, which is in the linear subspace span{(Pi φi)φSS ∈(cid:0) [n]
≤d−1(cid:1)}. This
≤d−1(cid:1) = (cid:0) [n]
d−1(cid:1) +(cid:0) [n]
linear space is the eigenspace of A with an eigenvalue 0; and its dimension is(cid:0) n
d−2(cid:1) +
···(cid:0)[n]
0(cid:1). By the same reason, V ′null is the eigenspace in B with an eigenvalue 0.
Let Vd be the largest eigenspace in Ad,d on(cid:0)[n]
d(cid:1) ×(cid:0)[n]
d(cid:1). We demonstrate how to find an eigenspace of
fd(T ∪ j) = 0 for any
A based on Vd. From the definition of Vd, for any fd ∈ Vd, fd satisfies that Pj /∈T
T ∈ (cid:0) [n]
d−1(cid:1) from the property of the Johnson scheme. Thus, from Claim 2.9 and the fact that A(S, T ) only
i (cid:1) and T ∈ (cid:0)[n]
depends on S ∩ T given S ∈ (cid:0)[n]
d(cid:1), we know Ai,dfd = ~0 for all i ≤ d − 1. We construct
an eigenvector f in A from fd as follows: f (S) = 0 for all S ∈ (cid:0)[n]
<d(cid:1) and f (T ) = fd(T ) for all T ∈ (cid:0)[n]
d(cid:1),
Then we move to Vd−1 in Ad,d and illustrate how to use an eigenvector in Vd−1 to construct an eigenvec-
fd−1(S)φS be the homogeneous degree d − 1 polyno-
tor of A. For any fd ∈ Vd−1, let fd−1 = PS∈( [n]
fd−1(S)(cid:17) φT . From Claim 2.9, Ai,dfd = 0 for all i < d − 1 and
mial such that fd =PT∈([n]
Ai,d−1fd−1 = 0 for all i < d − 2. Observe that fd−1 is an eigenvector of Ad−1,d−1, although it is possible
that the eigenvalue of fd−1 in Ad−1,d−1 is different than the eigenvalue of fd in Ad,d. At the same time, from
the symmetry of A and the relationship between fd and fd−1, Ad−1,dfd = β0fd−1 and Ad,d−1fd−1 = β1fd
for some constants β0 and β1 only depending on δ and d. Thus we can find a constant αd−1,d such that
(~0, fd−1, αd−1,dfd) becomes an eigenvector of A.
More directly, we determine the constant αd−1,d from the orthogonality between (~0, fd−1, αd−1,d · fd)
and the null space span{(Pi φi)φSS ∈ (cid:0) [n]
d−1(cid:1) and rewrite (Pi φi)φT =
Pj∈T φT\j + q · T · φT +Pj /∈T φT∪j. From the orthogonality,
qT · fd−1(T ) +Xj /∈T
i.e., f = (~0, fd). It is straightforward to verify that A(~0, fd) = λd(~0, fd), where the eigenvalue λd is the
eigenvalue of Vd in Ad,d.
≤d−1(cid:1)}. We pick any T ∈ (cid:0) [n]
fd−1(T ′)(cid:1) = 0
d−1)
d )(cid:16)PS∈( T
d−1)
αd−1,d(cid:0) XT ′∈(T ∪j
d−1)
⇒(cid:18)qT + (n − T)αd−1,d(cid:19) fd−1(T ) + αd−1,d(cid:18) XT ′′∈( T
d−2)Xj /∈T
fd−1(T ′′ ∪ j)(cid:19) = 0.
12
From the property of fd−1 thatPj /∈T ′′ fd−1(T ′′ ∪ j) = 0 for all T ′′ ∈(cid:0) [n]
(cid:0)qT + (n − 2T)αd−1,d(cid:1) fd−1(T ) = 0,
d−2(cid:1), we simplify it to
n−2d+2 directly.
which determines αd−1,d = −(d−1)q
Following this approach, we figure out all eigenspaces of A from the eigenspaces V0, V1,··· , Vd in Ad,d.
For convenience, we use V ′k for k ≤ d to denote the kth eigenspace in A extended by Vk in Ad,d. We first
choose the coefficients in the combination. Let αk,i = 0 for all i < k, αk,k = 1, and αk,k+1,··· , αk,d
satisfy the recurrence relation (we will show the choices of α later):
i · αk,k+i−1 + (k + i) · q · αk,k+i + (n − 2k − i)αk,k+i+1 = 0.
(2)
f (S);
k)
k )
k)
following three properties:
k(cid:1)} satisfy the
f (T ∪ j) = 0 (neglect this property for V ′0);
≤d(cid:1) spanned by { f (S)S ∈(cid:0)[n]
Now we show the recurrence relation of αk,k+i from the fact that f is orthogonal to the null space of A.
k−1(cid:1),Pj /∈T
>k(cid:1), f (T ) = αk,T ·PS∈(T
<k(cid:1), f (T ) = 0.
Then for every f ∈ Vk, the coefficients of f (T ) over all T ∈(cid:0)[n]
1. ∀T ∈(cid:0) [n]
2. ∀T ∈(cid:0)[n]
3. ∀T ∈(cid:0)[n]
We consider (Pi φi)φT in the null space for a subset T of size k + i < d and simplify (Pi φi)φT to
Pj∈T φT\j + q · T · φT +Pj /∈T φT∪j. We have
Xj∈T
αk,k+i−1 XS∈(T \j
k)(cid:0)i· αk,k+i−1 + (k + i)q· αk,k+i + (n− k− i)αk,k+i+1(cid:1) f (S) + XT ′∈( T
XS∈(T
Using the first property ∀T ′ ∈(cid:0) [n]
k−1(cid:1),Pj /∈T ′ f (T ′ ∪ j) = 0 to eliminate all S′ not in T , We obtain
f (S) + (k + i) · q · αk,k+i XS∈(T
αk,k+i+1 XS∈(T ∪j
αk,k+i+1Xj /∈T
f (S) +Xj /∈T
k)
k)
f (S) = 0.
relation in (2).
(cid:0)i · αk,k+i−1 + (k + i) · q · αk,k+i + (n − 2k − i)αk,k+i+1(cid:1) XS∈(T
f (S) is not necessary equal to 0 to satisfy the first property (actually PS∈(T
Because PS∈(T
for all T ∈ (cid:0) [n]
k(cid:1)), the coefficient is 0, which provides the recurrence
The dimension of V ′k is(cid:0)[n]
k(cid:1)−(cid:0) [n]
k−1(cid:1) from the first property (It is straightforward to verifyPd
dim(V ′null) = Pd
k(cid:1) −(cid:0) [n]
k−1(cid:1)) +(cid:0) [n]
d−2(cid:1) + ···(cid:0)[n]
≤d(cid:1)). The orthogonality between V ′i
and V ′j follows from Claim 2.9 and the orthogonality of Vi and Vj.
Remark 3.1 V ′1,··· , V ′d are the non-zero eigenspaces in B except for V ′0. For f ∈ V ′0, observe that f (T )
only depends on the size of T and f (∅). Hence for any polynomial f ∈ V ′0, f is a constant function over the
support of Dp, i.e., VarDp(V ′0) = 0. Therefore V ′0 is in the null space of B.
k+i(cid:1) indicates f (S) = 0 for all S ∈ (cid:0)[n]
d−1(cid:1) +(cid:0) [n]
0(cid:1) = (cid:0)[n]
k=0((cid:0)[n]
f (S) = 0
k)
f (S) = 0 ⇒
k )
f (T ′∪ j) = 0.
k−1)
k=0 dim(V ′k)+
13
We use induction on i to bound αk,k+i. From αk,k = 1 and the recurrence relation (2), the first few
n−2k−1
n−2k and αk,k+2 = − 1+(k+1)q·αk,k+1
ni + O(n−i−1) and αk,k+2i+1 = (−1)i+1O(n−i−1).
terms would be αk,k+1 = − kq
Claim 3.2 αk,k+2i = (−1)i (2i−1)!!
Proof. We use induction on i again. Base Case: αk,k = 1 and αk,k+1 = − kq
n−2k .
From the induction hypothesis αk,k+2i−2 = (−1)i−1Θ(n−i+1) and αk,k+2i−1 = (−1)in−i, the ma-
jor term of αk,k+2i is determined αk,k+2i−2 such that αk,k+2i = (−1)i (2i−1)!!
ni + O(n−i−1). Similarly,
αk,k+2i+1 = (−1)i+1O(n−i−1).
⊓⊔
n−2k−1 + O(n−2).
= − 1
Now we bound the eigenvalue of V ′k. For convenience, we think 0! = 1 and (−1)!! = 1.
(i−1)!!(i−1)!!
Theorem 3.3 For any k ∈ {0,··· , d}, the eigenvalue of V ′k in A isPd−k
any k ∈ {1,··· , d}, the eigenvalue of V ′k in B is the samePd−k
≤d) A(S, T ) f (T ) for the eigenvalue of
Proof. We fix a polynomial f ∈ V ′k and S ∈(cid:0)[n]
k(cid:1) to calculatePT∈([n]
f (S′), we expand PT A(S, T ) f (T ) into the summation
V ′k in A. From the fact f (T ) = αk,T ·PS′∈(T
of f (S′) over all S′ ∈ (cid:0)[n]
k(cid:1) with coefficients. From the symmetry of A, the coefficients of f (S′) in the
expansion only depends on the size of S ∩ S′ (the sizes of S and S′ are k). Hence we use τi to denote the
coefficients of f (S′) given S′∆S = i. ThusPT A(S, T ) f (T ) =PS′∈([n]
We calculate τ0,··· , τ2d as follows. Because S′ = S = k, S∆S′ is always even. For τ0, we only
± O(n−1). For
even i=0
(i−1)!!(i−1)!!
± O(n−1).
k ) τS′△S
consider T containing S and use k + i to denote the size of T .
f (S′).
even i=0
k)
i!
i!
τ0 =
d−k
Xi=0(cid:18)n − k
i (cid:19) · αk,k+i · δi.
(3)
i
d−k
τ2l =
αk,k+i
Xi=0
i − t (cid:19)δ2l+i−2t.
For τ2l, we fix a subset S′ with S∆S′ = 2l and only consider T containing S′. We use k + i to denote the
size of T and t to denote the size of the intersection of T and S \ S′.
t(cid:19)(cid:18)n − k − 2l
Xt=0(cid:18)l
We will prove that τ0 = Θ(1) and τ2l = O(n−l) for all l ≥ 1 then eliminate all S′ 6= S inPT A(S, T ) f (T ) =
k ) τS′△S f (S′) to obtain the eigenvalue of V ′k.
PS′∈([n]
From Claim 2.5 and Claim 3.2, we separate the summation of τ0 = Pd−k
i (cid:1) · αk,k+i · δi to
Peven i(cid:0)n−k
i (cid:1) · αk,k+i · δi +Podd i(cid:0)n−k
i (cid:1) · αk,k+i · δi. We replace δi and αk,k+i by the bound in Claim 2.5
Xeven i(cid:18)n − k
i (cid:19)(−1)
2 (cid:18) (i − 1)!!
(i − 1)!!
i=0 (cid:0)n−k
and Claim 3.2:
2 + i
(4)
n
n
i
2
i
2
i
i+1
2 + i+1
2
· O(n− i+1
2 ) · O(n− i+1
2 ).
+ O(n−i−1)(cid:19)
+ Xodd i(cid:18)n − k
i (cid:19)(−1)
14
even i=0
It shows τ0 = Pd−k
the fact that for any T ′ ∈(cid:0) [n]
(k − i)
i) XS1∈([n]\S
k−i )
XS0∈(S
(i−1)!!(i−1)!!
where O(n−l) comes from the fact that αk,k+i = O(n− i
+ O(n−1). For τ2l, we bound it by O(n−l) through similar method,
).
At last, we show the eigenvalue of V ′k is O(1/n) close to τ0, which is enough to finish the proof. From
i−t (cid:1) < ni−t, and δ2l+i−2t = O(n− 2l+i−2t
2 ),(cid:0)n−k−2l
i!
2
k−1(cid:1),Pj /∈T ′ f (T ′ ∪ j) = 0, we have (recall thatS = k)
f (S0 ∪ S1)
+ (i + 1)
XS0∈( S
i+1) XS1∈( [n]\S
= XS0∈(S
i) XS1∈( [n]\S
k−i−1)
k−i−1) Xj /∈S0∪S1
f (S0 ∪ S1)
f (S0 ∪ S1 ∪ j) = 0.
Thus we apply it onPi τ2i(cid:0)PS′∈([n]
k ):S∩S′=k−i
f (S′)(cid:1) to remove all S′ 6= S. Let τ′2k = τ2k and
Using the above rule, it is straightforward to verify
τ′2k−2i−2 = τ2k−2i−2 −
i + 1
k − i · τ′2k−2i.
k
Xi=j
τ2i(cid:18)
XS′∈([n]
k ):S∩S′=k−i
f (S′)(cid:19) = τ′2j(cid:18)
i=0 τ2i(cid:0)PS′∈([n]
XS′∈([n]
k ):S∩S′=k−j
f (S′)(cid:19)
f (S′)(cid:1) = τ′0
i=0 τ2i · (−1)i(cid:0)k
i(cid:1).)
even i=0
from j = k to j = 0 by induction. Therefore Pd
τ2i = O(n−i), we have τ′0 = τ0 ± O(1/n). (Remark: actually, τ′0 =Pk
From all discussion above, the eigenvalue of V ′k in A is τ′0, which isPd−k
± O(n−1).
For V ′k in B of k ≥ 1, observe that the difference δS · δT between A(S, T ) and B(S, T ) will not
f (T )(cid:1) = 0 from the fact f is
i ) δSδT f (T ) = δSδi(cid:0)PT∈([n]
change the calculation of τ , because PT∈([n]
⊓⊔
orthogonal to V ′0.
k ):S∩S′=k−i
f (S). Because
(i−1)!!(i−1)!!
i )
i!
i!
Because (i−1)!!(i−1)!!
.5 and [ d
2 ] + 1 ≤ d.
between .5 and [ d+1
≤ 1 for any even integer i ≥ 0, we have the following two corollaries.
Corollary 3.4 All non-zero eigenvalues of EDp[f 2] in the linear space of span{φSS ∈(cid:0)[n]
≤d(cid:1)} are between
Corollary 3.5 All non-zero eigenvalues of VarDp[f ] in the linear space of span{φSS ∈ (cid:0) [n]
1,··· ,d(cid:1)} are
Because f + (Pi φi)h ≡ f over supp(Dp) for any h of degree-at-most d − 1, we define the projection of
f onto V ′null to compare kfk2
Definition 3.6 Fix the global cardinality constraint Pi xi = (1 − 2p)n and the Fourier transform φS, let
hf denote the projection of a degree d multilinear polynomial f onto the null space span{(Pi φi)φSS ∈
(cid:0) [n]
≤d−1(cid:1)} of EDp[f 2] and VarDp(f ), i.e., f − (Pi φi)hf is orthogonal to the eigenspace of an eigenvalue 0
in EDp[f 2] and VarDp(f ).
2 and EDp[f 2].
2 ] ≤ d.
15
f (S)2.
From the above two corollaries and the definition of hf , we bound EDp[f 2] by kfk2
VarDp(f ), we exclude f (∅) because VarDp(f ) is independent with f (∅). Recall that kfk2
PS
Corollary 3.7 For any degree d multilinear polynomial f and a global cardinality constraint Pi xi =
(1 − 2p)n, EDp[f 2] ≤ dkfk2
Corollary 3.8 For any degree d multilinear polynomial f and a global cardinality constraint Pi xi =
(1 − 2p)n, VarDp(f ) ≤ dkf − f (∅)k2
2 and EDp[f 2] ≥ 0.5kf − (Pi φi)hfk2
2 as follows. For
2 = EUp[f 2] =
2 and VarDp(f ) ≥ 0.5kf − f (∅) − (Pi φi)hf− f (∅)k2
4 Parameterized algorithm for CSPs above average with the bisection con-
2.
2.
straint
We prove that CSPs above average with the bisection constraint are fixed-parameter tractable. Given an
f (S)2 and
bisection constraint.
instance I from d-ary CSPs and the bisection constraint Pi xi = 0, we use the standard basis {χSS ∈
(cid:0)[n]
≤d(cid:1)} of the Fourier transform in U and abbreviate fI to f . Recall that kfk2
2 = EU [f 2] = PS
D is the uniform distribution on all assignments in {±1}n complying with the bisection constraint.
For f with a small variance in D, we use hf− f (∅) to denote the projection of f − f (∅) onto the null space
≤d−1(cid:1)}. We know kf − f (∅)− (Pi xi)hf− f (∅)k2
span{(Pi xi)χSS ∈(cid:0) [n]
2 ≤ 2VarD(f ) from Corollary 3.8,
i.e., the lower bound of the non-zero eigenvalues in VarD(f ). Then we show how to round hf− f (∅) in
Section 4.1 to a degree d − 1 polynomial h with integral coefficients such that kf − f (∅) − (Pi xi)hk2
2 =
2), which indicates that f − f (∅) − (Pi xi)h has a small kernel under the
O(kf − f (∅) − (Pi xi)hf− f (∅)k2
Otherwise, for f with a large variance in D, we show the hypercontractivity in D that ED[(f −
ED[f ])4] = O(ED[(f − ED[f ])2]2) in Section 4.2. From the fourth moment method, we know there
exists α in the support of D satisfying f (α) ≥ ED[f ] + Ω(pVarD[f ]2). At last, we prove the main theorem
in Section 4.3.
Theorem 4.1 Given an instance I of a CSP problem of arity d and a parameter t, there is an algorithm with
running time O(n3d) that either finds a kernel on at most Cdt2 variables or certifies that OP T ≥ AV G + t
under the bisection constraint for a constant Cd = 24d2 · 7d · 9d · 22d ·(cid:0)d!(d − 1)!··· 2!(cid:1)2.
4.1 Rounding
2).
In this section, we show that for any polynomial f of degree d with integral coefficients, there exists an
2 =
efficient algorithm to round hf into an integral-coefficient polynomial h while it keeps kf − (Pi xi)hk2
O(kf − (Pi xi)hfk2
Theorem 4.2 For any constants γ and d, given a degree d multilinear polynomial f with kf−(Pi xi)hfk2
2 ≤
√n whose Fourier coefficient f (S) is a multiple of γ for all S ∈(cid:0)[n]
≤d(cid:1), there exists an efficient algorithm to
find a degree-at-most d − 1 polynomial h such that
d!(d−1)!···2! , which demonstrates that the Fourier coeffi-
1. The Fourier coefficients of h are multiples of
γ
γ
cients of f − (Pi xi)h are multiples of
2. kf − (Pi xi)hk2
2 ≤ 7d · kf − (Pi xi)hfk2
2.
d!(d−1)!···2! .
16
2) in the same order.
The high level idea of the algorithm is to round hf (S) to h(S) from the coefficients of weight d − 1 to the
coefficient of weight 0. At the same time, we guarantee that for any k < d, the rounding on the coefficients
2 = O(kf − (Pi xi)hfk2
of weight k will keep kf − (Pi xi)hk2
Because hf contains non-zero coefficients up to weight d − 1, we first prove that we could round
{hf (S)S ∈(cid:0) [n]
d−1(cid:1)} to multiples of γ/d!. Observe that for T ∈(cid:0)[n]
d(cid:1), the coefficient of χT in f − (Pi xi)hf
hf (T \ j). BecausePT∈([n]
is f (T )−Pj∈T
hf (T \ j)
hf (T \ j))2 = o(n), f (T )−Pj∈T
d )( f (T )−Pj∈T
hf (T \ j) mod γ is close to 0 for most T . Our start point is
is close to 0 for most T in(cid:0)[n]
d(cid:1). HencePj∈T
d−1(cid:1), h(S) is close to a multiple of γ/d! from the above discussion.
to prove that for any S ∈(cid:0) [n]
Lemma 4.3 If f (T ) is a multiple of γ and f (T ) −PS∈( T
multiple of γ/d! for all S ∈(cid:0) [n]
d−1(cid:1).
hf (S) = 0 for all T ∈ (cid:0)[n]
d(cid:1), then hf (S) is a
Proof. From the two conditions, we know
d−1)
hf (S) ≡ 0 mod γ
XS∈( T
d−1)
for any T ∈(cid:0)[n]
d(cid:1). We prove that
(d − 1)! · hf (S1) + (−1)d(d − 1)! · hf (S2) ≡ 0 mod γ
for any S1 ∈(cid:0) [n]
d−1(cid:1) and S2 ∈(cid:0)[n]\S
d−1(cid:1). Thus
0 ≡ (d − 1)! · XS2∈( T
d−1)
hf (S2) ≡ d! · hf (S1) mod γ,
d
d−1)
(cid:1), becausePS∈( T
hf (T \ j), we use (T ) to denote the equation
for any T with S1 ∩ T = ∅, which indicates hf (S1) is a multiple of γ/d!.
Without loss of generality, we assume S1 = {1, 2,··· , d − 1} and S2 = {k1, k2,··· , kd−1}. For a
subset T ∈(cid:0)S1∪S2
hf (S) =Pj∈T
hf (T \ j) ≡ 0 mod γ
Xj∈T
Let βd−1,1 = (d − 2)! and βd−i−1,i+1 = −i
d−i−1 · βd−i,i for any i ∈ {1,··· , d − 2} (we choose βd−1,1 to
guarantee that all coefficients are integers). Consider the following linear combination of equations over
T ∈(cid:0)S1∪S2
hf (T1∪ T2\ j)(cid:1) ≡ 0 mod γ.
Xi=1
i(cid:1), the coefficient of hf (S ∪ S′) is
Observe that for any i ∈ {1,··· , d − 2}, S ∈ (cid:0) S1
i · βd−i,i + (d − i − 1) · βd−i−1,i+1 = 0 in equation (5), where i comes from the number of choices of T1 is
d − 1 − S = i and d − i − 1 comes from the number of choices of T2 is (d − 1) − S′.
i )(cid:0) Xj∈T1∪T2
d−i−1(cid:1), and S′ ∈ (cid:0)S2
(cid:1) with coefficients βd−i,i:
(T1∪ T2) ⇒
βd−i,i XT1∈( S1
βd−i,i XT1∈( S1
d−i),T2∈(S2
i )
d−i),T2∈(S2
d−1
Xi=1
d
d−1
(T)
(5)
17
Hence equation (5) indicates that (d − 1)βd−1,1hf (S1) + (d − 1)β1,d−1hf (S2) ≡ 0 mod γ. Setting
into βd−1,1 = (d − 2)! and β1,d−1 = (−1)d−2(d − 2)!, we obtain
(d − 1)! · hf (S1) + (−1)d(d − 1)! · hf (S2) ≡ 0 mod γ.
⊓⊔
d−1,(cid:1), hf (S) is
0.1
d−1)
d−1)
d(cid:1), PS∈( T
d )( f (T ) −PS∈( T
Following the proof in Lemma 4.3, we obtain that hf (S) is (2d)!(d!)2
hf (S))2 = k = o(n0.6), then for all S ∈ (cid:0) [n]
Corollary 4.4 If PT∈([n]
d · γ/d! close to a multiple of γ/d!.
hf (S) is n−.1
Proof. From the condition, we know that except for n.8 choices of T ∈ (cid:0)[n]
close to a multiple of γ because of n.8 · (n−.1)2 > k. Observe that the above proof depends on the Fourier
coefficients in at most 2d + 1 variables of S1 ∪ T . Because n0.8 = o(n), for any subset S1 ∈(cid:0) [n]
d−1(cid:1), there is
a subset T ∈(cid:0)[n]\S1
d (cid:1) such that for any T ′ ∈(cid:0)S1∪T
d · γ/d! close to a multiple of
γ/d! for any S ∈(cid:0) [n]
d−1(cid:1).
⊓⊔
We consider a natural method to round hf , which is to round hf (S) to the closet multiple of γ/d! for every
S ∈(cid:0) [n]
d−1(cid:1).
Claim 4.5 Let hd−1 be the rounding polynomial of hf such that hd−1(S) = hf (S) for any S 6= d − 1 and
hd−1(S) is the closest multiple of γ/d! to hf (S) for any S ∈(cid:0) [n]
If ǫ(S) < .1/d · γ/d! and α(T ) is a multiple of γ for any T , then
d )(cid:0)α(T ) − XS∈( T
d−1(cid:1). Let ǫ(S) = hd−1(S) − hf (S).
hf (S) is n−.1 close to a multiple of γ.
ǫ(S)(cid:1)2 ≤ XT∈([n]
d (cid:1),PS∈( T ′
n.1 < 0.1
d−1)
d−1)
d−1)
d−1)
d )(cid:0)α(T ) −PS∈( T
d−1) ǫ(S)(cid:1)2 ≤PT∈([n]
hf (S)(cid:1)2.
d−1) ǫ(S) < 0.1 · γ/d!, then PS∈( T
d )(cid:0)PS∈( T
d )(cid:0) XS∈( T
XT∈([n]
Proof. For each T ∈ (cid:0)[n]
d(cid:1), Because PS∈( T
d−1) ǫ(S) < α(T ) −
hf (S)(cid:1)2.
hf (S). Hence we knowPT∈([n]
PS∈( T
⊓⊔
From now on, we use hd−1 to denote the degree d − 1 polynomial of hf after the above rounding process
on the Fourier coefficients of weight d − 1. Now we bound the summation of the square of the Fourier
2. Observe that rounding hf (S) only affect the
coefficients in f − (Pi xi)hd−1, i.e., kf − (Pi xi)hd−1k2
hf (S)χS\i +
terms of T ∈(cid:0)[n]
hf (S)χS∪i.
Pi /∈S
Lemma 4.6 kf − (Pi xi)hd−1k2
2 ≤ 7kf − (Pi xi)hfk2
Proof. Let ǫ(S) = hd−1(S) − hf (S). It is sufficient to prove
ǫ(S)
d−2(cid:1) inside S, because (Pi xi)hf (S)χS =Pi∈S
d(cid:1) containing S and T ′ ∈(cid:0) [n]
d )(cid:0) f (T ) − XS∈( T
hf (S) − XS∈( T
f (T ) − XS∈( T
≤ 4 XT∈([n]
hf (S)(cid:1)2,
XT∈([n]
d−1)
d−1)
d−1)
d−1)
2.
2
18
d )
(6)
and
Claim 4.5. From the inequality of arithmetic and geometric means, we know the cross terms:
2
hf (T ′ ∪ {j}) − Xj /∈T ′
d )(cid:0)PS∈( T
hf (S)(cid:12)(cid:12) · XS∈( T
d−1) ǫ(S)(cid:1)2 ≤PT∈([n]
ǫ(S) ≤ 2 XT∈([n]
ǫ(T ′ ∪ {j})
≤ 2kf−(Xi
d )(cid:0) f (T ) −PS∈( T
d )(cid:0) f (T ) − XS∈( T
d−1)
d−1)
hf (S)(cid:1)2.
xi)hfk2
2. (7)
d−1)
hf (S)(cid:1)2 by
d−2)
f (T ′) − XS∈( T ′
hf (S) − Xj /∈T ′
XT ′∈( [n]
Equation (6) follows the fact thatPT∈([n]
d−3)
XT∈([n]
d )
f (T ) − XS∈( T
2 ·(cid:12)(cid:12)
d−1)
For (7), observe that
XT ′∈( [n]
d−2)(cid:0)Xj /∈T ′
Hence we have
ǫ(T ∪ {j})(cid:1)2 = (d − 1) XS∈( [n]
≤ XT∈([n]
ǫ(S)2 + XS,S′:S∩S′=d−2
ǫ(S)(cid:1)2 ≤ XT∈([n]
d )(cid:0) XS∈( T
d−1)
d−1)
2ǫ(S)ǫ(S′)
d )(cid:0) f (T ) − XS∈( T
d−1)
hf (S)(cid:1)2.
XT ′∈( [n]
d−2)(cid:0) f (T ′) − XS∈( T ′
≤ XT ′∈( [n]
d−2)(cid:0) f (T ′) − XS∈( T ′
d−3)
d−3)
hf (S) − Xj /∈T ′
hf (T ′ ∪ {j})(cid:1) + XT ′∈( [n]
d−2)(cid:0)Xj /∈T ′
hf (T ′ ∪ {j})(cid:1) + XT∈([n]
ǫ(T ∪ {j})(cid:1)2
d )(cid:0) f (T ) − XS∈( T
hf (S) − Xj /∈T ′
d−1)
hf (S)(cid:1)2
xi)hfk2
2.
≤ kf − (Xi
We use the inequality of arithmetic and geometric means again to obtain inequality (7).
⊓⊔
Proof of Theorem 4.2. We apply Claim 4.5 and Lemma 4.6 for d times on the Fourier coefficients of hf
d−1(cid:1)},{hf (S)S ∈ (cid:0) [n]
from {hf (S)S ∈ (cid:0) [n]
d−2(cid:1)},··· to {hf (S)S ∈ (cid:0)[n]
specific, let hi be the polynomial after rounding the coefficients on(cid:0)[n]
Claim 4.5 to round coefficients of {hi(S)S ∈ (cid:0)[n]
parameters of γ in different rounds: γ in the rounding of hd−1, γ/d! in the rounding of hd−2,
rounding of hd−3 and so on. After d rounds, all coefficients in h0 are multiples of
Because kf − (Pi xi)hik2
2 ≤ 7kf − (Pi xi)hi+1k2
7d · kf − (Pi xi)hfk2
4.2 2 → 4 hypercontractive inequality under distribution D
We prove the 2 → 4 hypercontractivity for a degree d polynomial g in this section.
0(cid:1)} by choosing γ properly. More
≥i(cid:1) and hd = hf . Every time, we use
i (cid:1)} from hi+1 for i = d − 1,··· , 0. We use different
d!·(d−1)! in the
2 from Lemma 4.6. Eventually, kf − (Pi xi)h0k2
2 ≤
⊓⊔
d!(d−1)!(d−2)!···2! .
2.
γ
γ
19
Theorem 4.7 For any degree-at-most d multilinear polynomial g, ED[g4] ≤ 3d · 92d · kgk4
2.
Recall that kgk2 = EU [g2]1/2 = (PS g(S)2)1/2 and g − (Pi xi)hg ≡ g in the support of D. Because
2 ≤ 2 Ex∼D[g2] from the lower bound of non-zero eigenvalues in ED[g2] in Corollary 3.4,
kg − (Pi xi)hgk2
without loss of generality, we assume g is orthogonal to the null space span{(Pi xi)χSS ∈(cid:0) [n]
Corollary 4.8 For any degree-at-most d multilinear polynomial g, ED[g4] ≤ 12d · 92d · ED[g2]2.
≤d−1(cid:1)}.
Before proving the above Theorem, we observe that uniform sampling a bisection (S, ¯S) is as same as
first choosing a random perfect matching M and independently assigning each pair of M to the two subsets.
For convenience, we use P (M ) to denote the product distribution on M and EM to denote the expectation
over a uniform random sampling of perfect matching M. Let M (i) denote the vertex matched with i in M
and M (S) = {M (i)i ∈ S}. From the 2 → 4 hypercontractive inequality on product distribution P (M ),
we have the following claim:
Claim 4.9 EM [EP (M )[g4]] ≤ 9d EM [EP (M )[g2]2].
Now we prove the main technical lemma of the 2 → 4 hypercontractivity under the bisection constraint
to finish the proof.
Lemma 4.10 EM [EP (M )[g2]2] ≤ 3d · 9d · kgk4
2.
Theorem 4.7 follows from Claim 4.9 and Lemma 4.10.
Now we proceed to the proof of Lemma 4.10.
Proof of Lemma 4.10. Using g(x) =PS∈([n]
EM(cid:20) EP (M )(cid:2)( XS∈([n]
g(S)χS)2(cid:3)2(cid:21)
≤d)
≤d) g(S)χS, we rewrite EM [EP (M )[g2]2] as
≤d),S′6=S
g(S)2 +
≤d)
XS∈([n]
≤d),S′∈([n]
= EM(cid:20) EP (M )(cid:2) XS∈([n]
g(S)g(S′)χS△S′(cid:3)2(cid:21).
Notice that EP (M )[χS∆S′] = (−1)S∆S′/2 if and only if M (S∆S′) = S∆S′; otherwise it is 0. We expand
it to
EM(cid:20)(cid:16)kgk2
= kgk4
g(S)g(S′) · 1S△S′=M (S△S′) · (−1)S△S′/2(cid:17)2(cid:21)
2 · EM(cid:20)
XS∈([n]
+ EM
(cid:16)
XS∈([n]
g(S)g(S′) · 1S△S′=M (S△S′) · (−1)S△S′/2(cid:21)
g(S)g(S′) · 1S△S′=M (S△S′) · (−1)S△S′/2(cid:17)2
XS∈([n]
2 + 2kgk2
≤d),S′∈([n]
≤d),S′∈([n]
≤d),S′6=S
≤d),S′∈([n]
≤d),S′6=S
2 +
.
≤d),S′6=S
20
≤d),S′6=S g(S)g(S′) · 1S△S′=M (S△S′) · (−1)S△S′/2 in
We first bound the expectation of PS∈([n]
≤d),S′∈([n]
the uniform distribution over all perfect matchings, then bound the expectation of its square. Observe that
(m−1)(m−3)···(m−U+1) such that EM [1U =M (U ) ·
for a subset U ⊆ [n] with even size, EM [1U =M (U )] =
(−1)U/2] = δU, i.e., the expectation ED[χU ] of χU in D. Hence
EM(cid:2)
g(S)g(S′)1S△S′=M (S△S′) · (−1)S△S′/2(cid:3) ≤XS,S′
g(S)g(S′) · δS△S′ = ED[g2].
XS∈([n]
(U−1)(U−3)···1
≤d),S′∈([n]
≤d),S′6=S
From Corollary 3.4, the largest non-zero eigenvalue of the matrix constituted by δS∆S′ is at most d. Thus
the expectation is upper bounded by d · kgk2
2.
We define g′ to be a degree 2d polynomialPT∈( [n]
≤2d) g′(T )χT with coefficients
X
g(S)g(S′)
≤d):S∆S′=T
g′(T ) =
S∈([n]
≤d),S′∈([n]
for all T ∈(cid:0) [n]
≤2d(cid:1). Hence we rewrite
XS∈([n]
EM
(cid:18)
= EM(cid:20)(cid:0) XT∈( [n]
= EM(cid:20)XT,T ′
≤2d)
g(S)g(S′) · 1S△S′=M (S△S′) · (−1)S△S′/2(cid:19)2
≤d),S′∈([n]
≤d),S′6=S
g′(T ) · 1T =M (T )(−1)T/2(cid:1)2(cid:21)
g′(T )g′(T ′) · 1T =M (T )1T ′=M (T ′)(−1)T/2+T ′/2(cid:21).
2 from the discussion above. However, we still need to bound the contribution from the
Intuitively, because T ≤ 2d and T ′ ≤ 2d, most of pairs T and T ′ are disjoint such that EM [1T =M (T )1T ′=M (T ′)] =
EM [1T =M (T )]·EM [1T ′=M (T ′)]. The summation is approximately EM [PT g′(T )1T =M (T )(−1)T/2]2, which
is bounded by d2kgk4
correlated paris of T and T ′.
Notice that kg′k2
2 from the standard 2 → 4
Instead of bounding it by kgk4
2 through the analysis
2 = EU [g4], which can be upper bounded by ≤ 9dkgk4
on its eigenvalues and eigenspaces to this end. For convenience, we rewrite it to
2 ≤ 2d· 9dkgk4
hypercontractivity.
2 directly, we will bound it by 2d·kg′k2
g′(T )g′(T )1T =M (T )1T ′=M (T ′)(−1)T/2+T ′/2(cid:21) = XT,T ′
EM(cid:20)XT,T ′
g′(T )g′(T ′)∆(T, T ′),
where ∆(T, T ′) = 0 if T or T ′ is odd, otherwise
∆(T, T ′) = EM(cid:20)1T∩T ′=M (T∩T ′)1T =M (T )1T ′=M (T ′)(−1)T/2+T ′/2(cid:21)
= T ∩ T ′ − 1!! · T \ T ′ − 1!! · T ′ \ T − 1!!
(n − 1)(n − 3)··· (n − T ∪ T ′ + 1)
(−1)T ∆T ′/2.
21
2d(cid:1) ×(cid:0) n
Let A′ be the(cid:0) n
2d(cid:1) matrix whose entry (T, T ′) is ∆(T, T ′). We prove that the eigenspace of A′ with
eigenvalue 0 is still span{(Pi xi)χTT ∈(cid:0) [n]
≤2d−1(cid:1)}. Because ∆T,T ′ 6= 0 if and only if T,T ′,and T ∩T ′
are even, it is sufficient to showPi A′(S, T ∆i) = 0 for all odd sized T and even sized S.
1. S ∩ T is odd: ∆(S, T △ i) 6= 0 if and only if i ∈ S. We separate the calculation into i ∈ S ∩ T or
∆(S, T ∪ i).
not:
Xi
A′(S, T △ i) = Xi∈S∩T
∆(S, T \ i) + Xi∈S\T
Plugging in the definition of ∆, we obtain
S ∩ T · S ∩ T − 2!! · S \ T!! · T \ S!!
(n − 1)(n − 3)··· (n − S ∪ T + 1)
(−1)S/2+T−1/2
(−1)S/2+T +1/2 = 0.
2. S ∩ T is even: ∆(S, T △ i) 6= 0 if and only if i /∈ S. We separate the calculation into i ∈ T or not:
(n − 1)(n − 3)··· (n − S ∪ T + 1)
+ S \ T · S ∩ T!! · S \ T − 2!! · T \ S!!
Xi
A′(S, T △ i) = Xi∈T\S
∆(S, T \ i) + Xi /∈S∪T
∆(S, T ∪ i).
Plugging in the definition of ∆, we obtain
T \ S · S ∩ T − 1!! · S \ T − 1!! · T \ S − 2!!
(n − 1)(n − 3)··· (n − S ∪ T + 1)
(−1)S/2+T−1/2
+
(n − S ∪ T) · S ∩ T − 1!! · S \ T − 1!! · T \ S!!
(n − 1)(n − 3)··· (n − S ∪ T)
(−1)S/2+T +1/2 = 0.
From the same analysis in Section 3, the eigenspaces of A′ are as same as the eigenspaces of A with
degree 2d except the eigenvalues, whose differences are the differences between ∆S△T and δS△T . We
can compute the eigenvalues of A′ by the same calculation of eigenvalues in A. However, we bound the
eigenvalues of A′ by 0 (cid:22) A′ (cid:22) A as follows.
Observe that for any S and T , A′(S, T ) and A(S, T ) always has the same sign. At the same time,
A′(S, T ) = O(A(S,T )
nS∩T ). For a eigenspace V ′k in A, we focus on τ0 because the eigenvalue is O(1/n)-
close to τ0 from the proof of Theorem 3.3. We replace δi by any ∆(S, T ) of S = k,T = k + i and
i (cid:1) · αk,k+i · δi to obtain τ′0 for A′. Thus αk,k+i · ∆(S, T ) = Θ( αk,k+iδi
S ∩ T = i in τ0 = Pd−k
)
indicates τ′0 = Θ(1) < τ0 from the contribution of i = 0. Repeat this calculation to τ2l, we can show
τ′2l = O(τ2l) for all l. Hence we know the eigenvalue of A′ in V ′k is upper bounded by the eigenvalue of A
from the cancellation rule of τ in the proof of Theorem 3.3. On the other hand, A′ (cid:23) 0 from the definition
that it is the expectation of a square term in M.
i=0 (cid:0)n−k
ni
From Corollary 3.4 and all discussion above, we bound the largest eigenvalue of A′ by 2d. Therefore
EM(cid:20)XT,T ′
g′(T )g′(T )1T =M (T )1T ′=M (T ′)(−1)T/2+T ′/2(cid:21)
= XT,T ′
Over all discussion above, EM(cid:2) EP (M )[g2]2(cid:3) ≤ kgk4
22
g′(T )g′(T ′)∆(T, T ′) ≤ 2dkg′k2
2 ≤ 2d · 9d · kgk4
2.
2 + 2dkgk4
2 + 2d · 9d · kgk4
2 ≤ 3d · 9d · kgk4
2.
⊓⊔
4.3 Proof of Theorem 4.1
In this section, we prove Theorem 4.1. Let f = fI be the degree d multilinear polynomial associated with
the instance I and g = f − ED[f ] for convenience. We discuss VarD[f ] in two cases.
If VarD[f ] = ED[g2] ≥ 12d · 92d · t2, we have ED[g4] ≤ 12d · 92d · ED[g2]2 from the 2 → 4
√ED[g2]
2√12d·92d ] > 0. Thus PrD[g ≥ t] >
hypercontractivity of Theorem 4.7. By Lemma 2.1, we know PrD[g ≥
0, which demonstrates that PrD[f ≥ ED[f ] + t] > 0.
Otherwise we know VarD[f ] < 12d · 92d · t2. We consider f − f (∅) now. Let hf− f (∅) be the projection
of f − f (∅) onto the linear space such that kf − f (∅) − (Pi xi)hf− f (∅)k2
2 ≤ 2Var(f ) from Corollary 3.8
and γ = 2−d. From Theorem 4.2, we could round hf− f (∅) to h for f − f (∅) in time nO(d) such that
1. the coefficients of f − f (∅) − (Pi xi)h are multiples of
d!(d−1)!···2! ;
2. kf − f (∅) − (Pi xi)hk2
We first observe that f (α) = f (α)− (Pi αi)h(α) for any α in the support of D. Then we argue f − f (∅)−
(Pi xi)h has a small kernel, which indicates that f has a small kernel. From the above two properties, we
know there are at most kf − f (∅)−(Pi xi)hk2
d!(d−1)!···2! )2 non-zero coefficients in f − f (∅)−(Pi xi)h.
Because each of the nonzero coefficients contains at most d variables, the instance I has a kernel of at most
24d2 · 7d · 9d · 22d ·(cid:0)d!(d − 1)!··· 2!(cid:1)2t2 variables.
The running time of this algorithm is the running time to find hf and the rounding time O(nd). Therefore
2 ≤ 7dkf − f (∅) − (Pi xi)hf− f (∅)k2
2 ≤ 7d · 2 · 12d · 9d · t2.
2/(
this algorithm runs in time O(n3d).
γ
γ
5 2 → 4 hypercontractive inequality under distribution Dp
In this section, we prove the 2 → 4 hypercontractivity of low-degree multilinear polynomials in the distri-
bution Dp conditioned on the global cardinality constraintPi xi = (1 − 2p)n.
We assume p is in (0, 1) such that p · n is a integer. Then we fix the Fourier transform to be the p-biased
Fourier transform in this section, whose basis is {φSS ∈ (cid:0)[n]
≤d(cid:1)}. Hence we use φ1,··· , φn instead of
x1,··· , xn and say that a function only depends on a subset of characters {φii ∈ S} if this function only
f (S)φS, we
takes input from variables {xii ∈ S}. For a degree d multilinear polynomial f = PS∈([n]
use kfk2 = EUp[f 2]1/2 = (PS
We rewrite the global cardinality constraint as Pi φi = 0. For convenience, we use n+ = (1 − p)n
to denote the number of q p
1−p's in the global cardinality constraint of φi and n− = pn to denote the
number of −q 1−p
p 's. If n+
= 1−p
for some integers p1 and p2, we could follow the approach in
n−
Section 4.2 that first partition [n+ + n−] into tuples of size p1 + p2 then consider the production distribution
over tuples. However, this approach will introduce a dependence on p1 + p2 to the bound, which may be
superconstant. Instead of partitioning, we use induction on the number of characters and degree to prove the
2 → 4 hypercontractivity of low-degree multilinear polynomials in Dp.
Theorem 5.1 For any degree-at-most d multilinear polynomial f on φ1,··· , φn,
f (S)2)1/2 in this section.
p = p1
p2
≤d)
EDp[f (φ1, . . . , φn)4] ≤ 3 · d3/2 ·(cid:18)256 ·(cid:0)(
1 − p
p
)2 + (
p
1 − p
)2(cid:1)2(cid:19)d
· kfk4
2.
23
Recall that hf is the projection of f onto the null space span{(Pi φi)φSS ∈ (cid:0) [n]
≤d−1(cid:1)}. We know f −
(Pi φi)hf ≡ f in supp(Dp), which indicates EDp[f k] = EDp[(f−(Pi φi)hf )k] for any integer k. Without
loss of generality, we assume f is orthogonal to span{(Pi φi)φSS ∈ (cid:0) [n]
≤d−1(cid:1)}. From the lower bound of
eigenvalues in EDp[f 2] by Corollary 3.4, 0.5kfk2
Corollary 5.2 For any degree-at-most d multilinear polynomial f on φ1,··· , φn,
2 ≤ EDp[f 2]. We have a direct corollary as follows.
EDp[f (φ1, . . . , φn)4] ≤ 12 · d3/2 ·(cid:18)256 ·(cid:0)(
1 − p
p
)2 + (
p
1 − p
)2(cid:1)2(cid:19)d
EDp[f (φ1, . . . , φn)2]2.
Note that since xi can be written as a linear function of φi and linear transformation does not change the
degree of the multilinear polynomial, we also have for any degree-at-most d multilinear polynomial g on
x1,··· , xn,
EDp[g(x1, . . . , xn)4] ≤ 12 · d3/2 ·(cid:18)256 ·(cid:0)(
1 − p
p
)2 + (
p
1 − p
)2(cid:1)2(cid:19)d
EDp[g(x1, . . . , xn)2]2.
Proof of Theorem 5.1. We assume the inequality holds for any degree < d polynomials and use induction
on the number of characters in a degree d multilinear polynomial f to prove that if the multilinear polynomial
f of φ1,··· , φn depends on at most k characters of φ1,··· , φn, then EDp[f 4] ≤ d3/2 · C d · βk · kfk4
2 for
C = 256 ·(cid:16)( 1−p
1−p )2(cid:17)2
and β = 1 + 1/n.
p )2 + ( p
Base case.
f is a constant function that is independent from φ1,··· , φn, EDp[f 4] = f (∅)4 = kfk4
2.
Induction step. Suppose there are k ≥ 1 characters of φ1,··· , φn in f . Without loss of generality, we
assume φ1 is one of the characters in f and rewrite f = φ1h0 + h1 for a degree d − 1 polynomial h0 with at
most k− 1 characters and a degree d polynomial h1 with at most k− 1 characters. Because f is a multilinear
polynomial, kfk2
EDp[φ4
2. We expand EDp[f 4] = EDp[(φ1h0 + h1)4] to
0 · h1] + 6 EDp[φ2
2 = kh0k2
1 · h4
2 + kh1k2
1 · h3
1] + 4 EDp[φ1 · h0 · h3
0] + 4 EDp[φ3
1] + EDp[h4
1].
1h4
0] + EDp[φ2
(EDp[φ4
Finally, we bound EDp[φ1 · h0 · h3
0h2
1h2
1])/2 ≤ d3/2(cid:0)C d−0.5βk−1kh0k4
1]. However, we cannot apply the Cauchy-Schwarz inequality or the
inequality of arithmetic and geometric means, because we cannot afford a term like d3/2 · C dβk−1kh1k4
2
2 + C d−1/4 · βk−1 · kh0k2
2kh1k2
2(cid:1)/2.
24
0 · h2
2 and
1 · h2
1] ≤ C dβk−1kh1k4
)2, (
From the induction hypothesis, EDp[h4
1 − p
p
EDp[φ4
1 · h4
0 · h2
0] ≤ max{(
1] ≤ (EDp[φ4
1 · h2
Hence EDp[φ2
p
1 − p
0])1/2(EDp[h4
1h4
above discussion, this is at most
d3/4 · C d/2 · β(k−1)/2 · kh1k2
Applying the inequality of arithmetic and geometric means on EDp[φ3
2 · d3/4 · C (d−0.5)/2β(k−1)/2kh0k2
0] ≤ d3/2 · C d−0.5βk−1kh0k4
2.
)2} EDp[h4
1])1/2 from the Cauchy-Schwarz inequality. From the
(8)
2 ≤ d3/2 · C d−1/4 · βk−1 · kh0k2
2kh1k2
2. (9)
0 · h1], we know it is at most
(10)
1 · h3
any more. We use Dφ1>0 (Dφ1<0 resp.) to denote the conditional distribution of Dp on fixing φ1 =q p
(−q 1−p
resp.) and rewrite
p
1−p
EDp[φ1 · h0 · h3
1] =pp(1 − p) EDφ1>0[h0 · h3
1] −pp(1 − p) EDφi<0[h0 · h3
1].
Let L be the matrix corresponding to the quadratic form EDφ1>0[f g]−EDφ1<0[f g] for low-degree multilinear
polynomials f and g (i.e. let L be a matrix such that f T Lg = EDφ1>0[f g]−EDφ1<0[f g]). The main technical
lemma of this section is a upper bound on the spectral norm of L.
(11)
Lemma 5.3 Let g be a degree d(d ≥ 1) multilinear polynomial on characters φ2,··· , φn, we have
EDφ1>0[g2] − EDφ1<0[g2] ≤
3d3/2
p(1 − p) · kgk2
2√n
.
Therefore, the spectral norm of L is upper bounded by 3d3/2
p(1−p) · kgk2
2√n .
From the above lemma, we rewrite the equation (11) from the upper bound of its eigenvalues:
EDp[φ1 · h0 · h3
1] =pp(1 − p)(cid:16)EDφ1>0[h0 · h3
=pp(1 − p)(h0h1)T L · (h2
1](cid:17)
1] − EDφ1<0[h0 · h3
1) ≤pp(1 − p) ·
p(1 − p) ·
3(2d)3/2
Then we use the inequality of arithmetic and geometric means on it:
1
√n · kh0h1k2 · kh2
1k2.
EDp[φ1 · h0 · h3
1] ≤pp(1 − p) ·
Next, we use the 2 → 4 hypercontractivity kh2k2
Cauchy-Schwarz inequality to further simplify it to:
EDp[φ1 · h0 · h3
1] ≤
5d3/2
pp(1 − p) · (kh2
0k2 · kh2
1k2 +
≤
5d3/2
pp(1 − p) · 9d ·(cid:18) p2
1k2
2 + kh2
2/n
2
.
p (cid:17)d
1−p + (1−p)2
10d3/2
p(1 − p) · kh0h1k2
2 = EUp[h4] ≤ 9d ·(cid:16) p2
p (cid:17)d
1−p + (1−p)2
(cid:19)d
(kh0k2
9d ·(cid:16) p2
(1 − p)2
1 − p
+
n
p
kh1k4
2)
khk4
2 in Up and the
2 · kh1k2
2 +
1
nkh1k4
2).
(12)
25
From all discussion, we bound E[f 4] by the upper bound of each inequalities in (8),(10),(9),(12):
EDp[(φ1h0 + h1)4]
+ 4
5d3/2
1 · h4
(1 − p)2
0 · h1] + 6EDp[φ2
1] + 4EDp[φ1 · h0 · h3
=EDp[φ4
0] + 4EDp[φ3
≤d3/2C d−0.5βk−1kh0k4
1 · h3
1 · h2
0 · h2
2 + 4d3/2(cid:16)C d−0.5βk−1kh0k4
2 + C d−1/4βk−1kh0k2
(cid:19)d(cid:18)kh0k2
1
nkh1k4
2 + 8d3/2 · C d−1/4 · βk−1 +
·
pp(1 − p) · 9d ·(cid:18) p2
1 − p
(cid:19)d
+ d3/2 · C dβk−1!kh1k4
p
2 + d3/2 · (C d/n + C dβk−1)kh1k4
2kh1k2
≤3d3/2 · C d−0.5βk−1kh0k4
pp(1 − p) · 9d ·(cid:18) p2
+ 20d3/2
pp(1 − p) · 9d ·(cid:18) p2
≤d3/2 · C dβkkh0k4
2 + d3/2 · C dβk · 2kh0k2
2kh1k2
20d3/2
(1 − p)2
+
1 − p
1 − p
2 +
+
p
1
n
2
2 ≤ d3/2 · C dβkkfk4
2.
1] + EDp[h4
1]
2(cid:17) /2 + 6d3/2C d−1/4βk−1kh0k2
2kh1k2
2(cid:19) + d3/2 · C dβk−1kh1k4
2
2kh1k2
2
+
(1 − p)2
p
(cid:19)d! · kh0k2
2kh1k2
2
We prove Lemma 5.3 to finish the proof.
⊓⊔
Intuitively, both EDφ1>0[g2] and EDφ1<0[g2] are close to
EDp[g2] (we add the dummy character φ1 back in Dp) for a low-degree multilinear polynomial g; therefore
their gap should be small compared to EDp[g2] = O(kgk2
2). Recall that Dp is a uniform distribution on the
constraintPi φi = 0, i.e., there are always n+ characters of φi withq p
p .
For convenience, we abuse the notation φ to denote a vector of characters (φ1,··· , φn).
Proof of Lemma 5.3.
1−p and n− characters with −q 1−p
Let F be the p-biased distribution on n − 2 characters with the global cardinality
1−p +q 1−p
p = −q, i.e., n+ − 1 of the characters are alwaysq p
constraintPn−1
1−p and n− − 1
of the characters are always −q 1−p
p . Let ~φ−i denote the vector (φ2,··· , φi−1, φi+1,··· , φn) of n − 2
characters such that we could sample ~φ−i from F . Hence φ ∼ Dφ1<0 is equivalent to the distribution
that first samples i from 2,··· , n then fixes φi = q p
1−p and samples ~φ−i ∼ F . Similarly, φ ∼ Dφ1>0
is equivalent to the distribution that first samples i from 2,··· , n then fixes φi = −q 1−p
and samples
~φ−i ∼ F .
i=2 φi = −q p
p
For a multilinear polynomial g depending on characters φ2,··· , φn, we rewrite
EDφ1>0[g2] − EDφ1<0[g2]
= Ei Eφ∼F (cid:20)g(φi =r p
1 − p
, ~φ−i = φ)2 − g(φi = −r 1 − p
p
, ~φ−i = φ)2(cid:21) .
(13)
26
Eφ∼F (cid:20)(cid:18)g(φi =r p
1 − p
≤ Eφ∼F "(cid:18)g(φi =r p
1 − p
, ~φ−i = φ)(cid:19)
p
, ~φ−i = φ) − g(φi = −r 1 − p
·(cid:18)g(φi =r p
1 − p
, ~φ−i = φ) − g(φi = −r 1 − p
· Eφ∼F "(cid:18)g(φi =r p
1 − p
, ~φ−i = φ) + g(φi = −r 1 − p
, ~φ−i = φ)(cid:19)2#1/2
p
p
, ~φ−i = φ)(cid:19)(cid:21)
We will show its eigenvalue is upper bounded by 3d3/2
p(1−p)·p1/n. We first use the Cauchy-Schwarz inequality:
, ~φ−i = φ) + g(φi = −r 1 − p
From the inequality of arithmetic and geometric means and the fact that EDp[g2] ≤ dkgk2
second term is bounded by
p
, ~φ−i = φ)(cid:19)2#1/2
.
2, observe that the
Eφ∼F "(cid:18)g(φi =r p
1 − p
≤2 EDφ1>0(cid:20)g(φi =r p
1 − p
3d
p(1 − p) · kgk2
≤
p(1 − p)
EDp[g2] ≤
3
2.
, ~φ−i = φ)(cid:19)2#
, ~φ−i = φ) + g(φi = −r 1 − p
, ~φ−i = φ)2(cid:21) + 2 EDφ1<0(cid:20)g(φi = −r1 − p
p
p
, ~φ−i = φ)2(cid:21)
1/2
. We use
p , ~φ−i = φ)(cid:17)2(cid:21)1/2
p , ~φ−i = φ):
1−p , ~φ−i = φ) − g(φi = −q 1−p
Then we turn to Eφ∼F (cid:20)(cid:16)g(φi =q p
f (S)φS\i to replace g(φi =q p
gi =PS:i∈S
Eφ∼F
XS:i∈S
+r 1 − p
g(S)(cid:18)r p
1 − p
1−p , ~φ−i = φ) − g(φi = −q 1−p
p (cid:19) φS\i!2
p (cid:19) Eφ∼F (cid:2)gi(φ)2(cid:3)1/2
Eventually we bound Eφ∼F [gi(φ)2] by its eigenvalue. Observe that F is the distribution on n − 2
1−p's and (n−−1) −q 1−p
characters with (n+−1)q p
p 's, which indicatesPj φj +q = 0 in F . However, the
small difference betweenPj φj + q = 0 andPj φj = 0 will not change the major term in the eigenvalues
of Dp. From the same analysis, the largest eigenvalue of EF [g2
a calculation in Section 5.1.
Claim 5.4 For any degree-at-most d multilinear polynomial gi, Eφ∼F [gi(φ)2] ≤ dkgik2
2.
i ] is at most d. For completeness, we provide
Therefore we have Ei[Eφ∼F [gi(φ)2]1/2] ≤ √d · Ei[kgi(φ)k2] and simplify the right hand side of in-
=(cid:18)r p
1 − p
+r 1 − p
.
27
equality (13) further:
Ei(cid:26) Eφ∼F "(cid:18)g(φi =r p
1 − p
· Eφ∼F "(cid:18)g(φi =r p
1 − p
≤ Ei"(cid:18)r p
+r 1 − p
p (cid:19) ·
1 − p
≤
p(1 − p) · kgk2 · Ei[kgik2]
3d
√d · kgik2 ·s 3d
, ~φ−i = φ) − g(φi = −r1 − p
p
, ~φ−i = φ)(cid:19)2#1/2
, ~φ−i = φ) + g(φi = −r 1 − p
p(1 − p) · kgk2#
p
, ~φ−i = φ)(cid:19)2#1/2
(cid:27)
Using the fact thatPi kgik2
to obtain the desired upper bound on the absolute value of eigenvalues of EDφ1=1[g2] − EDφ1=−1[g2] :
2 and Cauchy-Schwartz again, we further simplify the expression above
2 ≤ dkgk2
3d
p(1 − p) · kgk2 · Ei[kgik2] ≤
3d
p(1 − p) · kgk2 ·
5.1 Proof of Claim 5.4
2)1/2
(Pi kgik2
√n
p(1 − p) · kgk2 ·r d
≤
3d
nkgk2 ≤
3d3/2
p(1 − p) · kgk2
2√n
.
⊓⊔
We follow the approach in Section 3 to determine the eigenvalues of EF [f 2] for a polynomial f =PS∈([n−2]
Recall thatPi φi + q = 0 over all support of F for q = 2p−1√p(1−p)
We abuse δk = EF [φS] for a subset S with size k. We start with δ0 = 1 and δ1 = −q/(n − 2). From
(Pi φi + q)φS ≡ 0 for k ≤ d − 1 and any S ∈(cid:0)[n]
k(cid:1), we have
as defined before.
≤d )
φS∪j] = 0 ⇒ k · δk−1 + (k + 1)q · δk + (n − 2 − k) · δk+1 = 0
E[Xj∈S
φjφS + qχS +Xj /∈S
Now we determine the eigenspaces of EF [f 2]. The eigenspace of an eigenvalue 0 is span{(Pi φ+q)φSS ∈
(cid:0)[n−2]
≤d−1(cid:1)}. There are d + 1 non-zero eigenspaces V0, V1,··· , Vd. The eigenspace Vi of EF [f 2] is spanned by
{ f (S)φSS ∈(cid:0)[n−2]
1. ∀T ′ ∈(cid:0)[n−2]
2. ∀T ∈(cid:0)[n]
3. ∀T ∈(cid:0)[n−2]
i (cid:1)}. For each f ∈ Vi, f satisfies the following properties:
i−1 (cid:1),Pj /∈T ′ f (S ∪ j) = 0 (neglect this contraint for V0).
<i(cid:1), f (T ) = 0.
>i (cid:1), f (T ) = αk,TPS∈T
f (S) where αk,k = 1 and αk,k+i satisfies
i · αk,k+i−1 + (k + i + 1) · q · αk,k+i + (n − 2 − 2k − i)αk,k+i+1 = 0.
28
f (S)φS.
We show the calculation of αk,k+i as follows: fix a subset T of size k + i and consider the orthogonality
between (Pi φi + q)]φT and f ∈ Vk:
Xj∈T
⇒ XS∈(T
k )
f (S) + (k + i + 1) · q · αk,k+i XS∈(T
αk,k+i−1 XS∈(T \j
k)(cid:16)(k + i + 1) · q · αk,k+i + (n − 2 − k − i)αk,k+i+1 + i · αk,k+i−1(cid:17) f (S)
αk,k+i+1Xj /∈T
f (S) +Xj /∈T
+ XT ′∈( T
k−1)
k)
αk,k+i+1 XS∈(T ∪j
k )
f (S) = 0
f (T ′ ∪ j) = 0.
Using the first property ∀T ′ ∈(cid:0)[n−2]
i−1 (cid:1),Pj /∈T ′ f (S ∪ j) = 0 to remove all S′ /∈ T , we have
k)(cid:0)i · αk,k+i−1 + (k + i + 1) · q · αk,k+i + (n − 2 − 2k − i)αk,k+i+1(cid:1) f (S) = 0
XS∈(T
We calculate the eigenvalues of Vk following the approach in Section 3. Fix S and S′ with i = S △ S′, we
still use τi to denote the coefficients of f (S′) in the expansion ofPT (δS△T − δS δT ) f (T ). Observe that τi
is as same as the definition in Section 3 in terms of δ and αk :
τ0 =
d−k
Xi=0(cid:18)n − 2 − k
i
(cid:19) · αk,k+i · δi,
τ2l =
d−k
Xi=0
αk,k+i
i
Xt=0(cid:18)l
t(cid:19)(cid:18)n − 2 − k − 2l
i − t
(cid:19)δ2l+i−2t.
Observe that the small difference between (Pi φi + q) ≡ 0 andPi φi ≡ 0 only changes a little in the
recurrence formulas of δ and α. For δ2i and αk,k+2i of an integer i, the major term is still determined by
δ2i−2 and αk,k+2i−2. For δ2i+1 and αk,k+2i+1, they are still in the same order (the constant before n−i−1
will not change the order). Using the same induction on δ and α, we have
ni + O(n−i−1);
1. δ2i = (−1)i (2i−1)!!
2. δ2i+1 = O(n−i−1);
3. αk,k+2i = (−1)i (2i−1)!!
4. αk,k+2i+1 = O(n−i−1).
ni + O(n−i−1);
even i=0
(i−1)!!(i−1)!!
Hence τ0 = Pd−k
we know the eigenvalue of Vk is τ0 ± O(τ2) =Pd−k
(i−1)!!(i−1)!!
From all discussion above, the eigenvalues of EF [f 2] is at most [ d
even i=0
i!
i!
+ O(n−1).
2 ] + 1 ≤ d.
+ O(n−1) and τ2l = O(n−l). Follow the same analysis in Section 3,
6 Parameterized algorithm for CSPs above average with global cardinality
constraints
We show that CSPs above average with the global cardinality constraint Pi xi = (1 − 2p)n are fixed-
parameter tractable for any p ∈ [p0, 1 − p0] with an integer pn. We still use Dp to denote the uniform
distribution on all assignments in {±1}n complying withPi xi = (1 − 2p)n.
29
Without loss of generality, we assume p < 1/2 and (1 − 2p)n is an integer. We choose the standard
basis {χS} in this section instead of {φS}, because the Fourier coefficients in {φS} can be arbitrary small
for some p ∈ (0, 1).
Theorem 6.1 For any constant p0 ∈ (0, 1) and d, given an instance of d-ary CSP, a parameter t, and a
parameter p ∈ [p0, 1 − p0], there exists an algorithm with running time nO(d) that either finds a kernel
on at most C · t2 variables or certifies that OP T ≥ AV G + t under the global cardinality constraint
Pi xi = (1 − 2p)n for a constant C = 160d2 · 30d(cid:16)( 1−p0
Let f be a degree d polynomial whose coefficients are multiples of γ in the standard basis {χSS ∈(cid:0)[n]
≤d(cid:1)}.
We show how to find an integral-coefficients polynomial h such that f −(cid:0)Pi xi−(1−2p)n(cid:1)h only depends
polynomial h such that f −(cid:0)Pi xi(cid:1)h only depends on O(VarD(f )) variables (where D is the distribution
conditioned on the bisection constraint). Without loss of generality, we assume f (∅) = 0 because VarDp(f )
is independent with f (∅).
Before proving that f depends on at most O(VarDp(f )) variables, we first define the inactivity of a
variable xi in f .
on O(VarDp(f )) variables. We use the rounding algorithm in Section 4.1 as a black box, which provides a
· (d!)3d2 · (1/2p0)4d.
6.1 Rounding
)2 + ( p0
1−p0
)2(cid:17)d
p0
equations.
≤d)
f (S)χS if f (S) = 0 for all S containing xi.
In general, there are multiple ways to choose h to turn a variable into inactive. However, if we know a
Definition 6.2 A variable xi for i ∈ [n] is inactive in f =PS
A variable xi is inactive in f under the global cardinality constraint Pi xi = (1 − 2p)n if there exists a
polynomial h such that xi is inactive in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h.
subset S of d variables and the existence of some h to turn S into inactive in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h,
we show that h is uniquely determined by S. Intuitively, for any subset S1 with d − 1 variables, there are
d (cid:1) =(cid:0)2d−1
(cid:0)S1∪S
d (cid:1) ways to choose a subset of size d. Any d-subset T in S1 ∪ S contains at least one inactive
h(T \ j) = 0 from the assumption. At the same time, there are at most
variable such that f (T ) −Pj∈T
d−1(cid:1) coefficients of h in S1 ∪ S. So there is only one solution of coefficients in h to satisfy these(cid:0)2d−1
(cid:0)2d−1
d (cid:1)
f (T )χT and p ∈ [p0, 1 − p0], let S be a subset with at least d variables
Claim 6.3 Given f = PT∈([n]
such that there exists a degree ≤ d − 1 multilinear polynomial h turning S into inactive in f −(cid:0)Pi xi −
(1 − 2p)n(cid:1)h. Then h is uniquely determined by any d elements in S and f .
Proof. Without lose of generality, we assume S = {1,··· , d} and determine h(S1) for S1 = {i1,··· , id−1}.
For simplicity, we first consider the case S ∩ S1 = ∅. From the definition, we know that for any
T ∈ (cid:0)S∪S1
h(T \ j) = 0. Hence
we can repeat the argument in Lemma 4.3 to determine h(S1) from f (T ) over all T ∈(cid:0)S∪S1
d (cid:1).
Let βd−1,1 = (d − 2)! and βd−i−1,i+1 = −i
d−i−1 βd−i,i for any i ∈ {1,··· , d − 2} be the parameters
define in Lemma 4.3. For any S2 ∈(cid:0) S
βd−i,i XT1∈( S1
d (cid:1), T contains at least one inactive variable, which indicates f (T ) −Pj∈T
d−1(cid:1), by the same calculation,
f (T1 ∪ T2) − Xj∈T1∪T2
h(T1 ∪ T2 \ j)
= 0
d−i),T2∈(S2
i )
d−1
Xi=1
30
indicates that (all h(S) not S1 or S2 cancel with each other)
(d − 1)! · h(S1) + (−1)d(d − 1)! · h(S2) =
d−1
Xi=1
βd−i,i XT1∈( S1
d−i),T2∈(S2
i )
f (T1 ∪ T2).
d−i),T2∈(S2
i )
≤d)
the definition, and vice versa.
i=1
βd−i,i
(d−1)!(cid:0)PT1∈( S1
d−1)
h(S2) = 0, we obtain h(S1) in terms of f and S.
Remark 6.4 The coefficients of h are multiples of γ/d! if the coefficients of f are multiples of γ.
d−1(cid:1), we repeat this argument for S1 ∈ (cid:0) [n]
≤d−1(cid:1) from S and the coefficients in f .
Hence h(S2) = (−1)d−1h(S1) + (−1)dPd−1
f (T1 ∪ T2)(cid:1) for any S2 ∈(cid:0) S
d−1(cid:1).
Replacing all h(S2) in the equation f (S) −PS2∈( S
If S ∩ S1 6= ∅, then we set S′ = S \ S1. Next we add arbitrary S ∩ S1 more variables into S′ such
that S′ = d and S′ ∩ S1 = ∅. Observe that S′ ∪ S1 contains at least d inactive variables. Repeat the above
argument, we could determine h(S1).
After determining h(S1) for all S1 ∈ (cid:0) [n]
d−2(cid:1) and so on. There-
fore we could determine h(S1) for all S1 ∈(cid:0) [n]
⊓⊔
Let h1 and h2 be two polynomials such that at least d variables are inactive in both f−(cid:0)Pi xi−(1−2p)n(cid:1)h1
and f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h2. We know that h1 = h2 from the above claim. Furthermore, it implies that
any variable that is inactive in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h1 is inactive in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h2 from
Based on this observation, we show how to find a degree d − 1 function h such that there are fews
active variables left in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h. The high level is to random sample a subset Q of
(1 − 2p)n variables and restrict all variables in Q to 1. Thus the rest variables constitutes the bisection
constraint on 2pn variables such that we could use the rounding process in Section 4.1. Let k be a large
number, Q1,··· , Qk be k random subsets and h1,··· , hk be the k functions after rounding in Section 4.1.
Intuitively, the number of active variables in f − (Pi /∈Q1
xk)hk are small with high
probability such that h1,··· , hk share at least d inactive variables. We can use one function h to represent
xi)hj over all
h1,··· , hk from the above claim such that the union of inactive variables in f − (Pi /∈Qj
j ∈ [k] are inactive in f − (Pi xi)h from the definition. Therefore there are a few active variables in
f − (Pi xi)h.
Let us move to f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h. Because h is a degree-at-most d− 1 function, (1 − 2p)n · h
is a degree ≤ d − 1 function. Thus we know that the number of active variables among degree d terms in
f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h is upper bounded by the number of active variables in f − (Pi xi)h. For the
degree < d terms left in f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h, we repeat the above process again.
Theorem 6.5 Given a global cardinality constraint Pi xi = (1 − 2p)n and a degree d function f =
PS∈([n]
running in time O(dn2d) to find a polynomial h such that there are at most C′
f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h for C′p,d = 20d27d·(d!)2d2
(1−2p)n(cid:1), we consider the assignments conditioned on xQ = ~1 and use fQ to
Proof. For any subset Q ∈ (cid:0)
rest variables isPi /∈Q xi = 0. We use DQ denote the distribution on assignments of {xii /∈ Q} satisfying
Pi /∈Q xi = 0, i.e., the distribution of {xii /∈ ¯Q} under the bisection constraint.
denote the restricted function f on xQ = ~1. Conditioned on xQ = ~1, the global cardinality constraint on the
f (S) with VarDp(f ) < n0.5 and coefficients of multiples of γ , there is an efficient algorithm
xi)h1,··· , f − (Pi /∈Qi
p,d·VarDp (f )
γ2
active variables in
(2p)4d
.
[n]
31
Let XQ(i) ∈ {0, 1} denote whether xi is active in fQ under the bisection constraint of ¯Q or not after
the bisection rounding in Theorem 4.2. From Theorem 4.2, we get an upper bound on the number of active
variables in fQ, i.e.,
for C′d = 7d(d!(d−1)!···2!)2
γ2
We claim that
XQ(i) ≤ 2C′d · VarDQ(fQ)
Xi
and any Q with VarDQ(fQ) = O(n0.6).
EQ[VarDQ(fQ)] ≤ VarDp(f ).
From the definition, EQ[VarDQ(fQ)] = EQ Ey∼DQ[fQ(y)2]− EQ(cid:2) Ey∼DQ[fQ(y)](cid:3)2. At the same time, we
observe that EQ Ey∼DQ[fQ(y)2] = ED[f 2] and EQ[Ey∼DQ fQ(y)]2 ≥ EDp[f ]2. Therefore EQ[VarDQ(fQ)] ≤
VarDp[f ]. One observation is that PrQ[VarDQ ≥ n0.6] < n−0.1 from the assumption VarDp(f ) < n0.5,
which is very small such that we can neglect it in the rest of proof. From the discussion above, we have
EQ[Pi XQ(i)] ≤ 2C′d · VarDp(f ) with high probability.
Now we consider the number of i's with EQ[XQ(i)] ≤ (2p)2d
denote the number of i's with EQ[XQ(i)] ≤ (2p)2d
convenience. Hence for any i > m, EQ[XQ(i)] > (2p)2d
is at least 1 − (2p)2d
5d . Without loss of generality, we use m to
and further assume these variables are {1, 2,··· , m} for
5d . We know the probability VarDQ(f ) ≤
, which implies
2VarDp (f )
(2p)2d
5d
2
n − m ≤
2C′d · VarDQ(f )
(2p)2d
5d
≤
20d · C′d
VarDp(f )
.
(2p)4d
5d − (2p)2d
We are going to show that EQ[XQ(i)] is either 0 or at least (2p)2d
5d , which means that only xm+1,··· , xn are
active in f under Pi xi = 0. Then we discuss how to find out a polynomial hd such that x1,··· , xm are
inactive in the degree d terms of f −(cid:0)Pi xi − (1 − 2p)n(cid:1)hd.
We fix d variables x1,··· , xd and pick d arbitrary variables xj1,··· , xjd from {d + 1,··· , n}. We
focus on {x1, x2,··· , xd, xj1,··· , xjd} now. With probability at least (2p)2d − o(1) ≥ 0.99(2p)2d over
random sampling Q, none of these 2d variables is in Q. At the same time, with probability at least 1 −
2d · (2p)2d
5d , all variables in the intersction {x1, . . . , xm} ∩ {x1, x2,··· , xd, xj1,··· , xjd} are inactive in fQ
under the bisection constraint on ¯Q (2d is for xj1,··· , xjd if necessary). Therefore, with probability at least
0.99(2p)2d − 2d · (2p)2d
2 ≥ 0.09(2p)2d, x1, x2,··· , xd are inactive in fQ underPi /∈Q xi = 0 and
such that the variables in {x1, . . . , xm} ∩
n − m is small. Namely there exists a polynomial hxj1 ,··· ,xjd
{x1, x2,··· , xd, xj1,··· , xjd} are inactive in fQ − (Pi /∈Q xi)hxj1 ,··· ,xjd
Now we apply Claim 6.3 on S = {1,··· , d} in f to obtain the unique polynomial hd, which is the
combination of hxj1 ,··· ,xjd
over all choices of j1,··· , jd, and consider f − (Pi xi)hd. Because of the
arbitrary choices of xj1,··· , xjd, it implies that x1,··· , xd are inactive in f − (Pi xi)hd. For example,
we fix any j1,··· , jd and T = {1, j1,··· , jd−1}. we know f (T ) −Pj∈T
(T \ j) = 0 from the
on the Fourier coefficients from Claim 6.3, we
definition of hxj1 ,··· ,xjd
have f (T ) −Pj∈T
Furthermore, it implies that xd+1,··· , xm are also inactive in f − (Pi xi)hd. For example, we fix
j1 ∈ {d + 1,··· , m} and choose j2,··· , jd arbitrarily. Then x1,··· , xd, and xj1 are inactive in fQ −
for some Q from the discussion above, which indicates that xj1 are inactive in f −
(Pi /∈Q xi)hxj1 ,··· ,xjd
(Pi xi)hd by Claim 6.3.
. Because hd agrees with hxj1 ,··· ,xjd
hd(T \ j) = 0.
hxj1 ,··· ,xjd
.
32
To find hd in time O(n2d), we enumerate all possible choices of d variables in [n] as S. Then we apply
Claim 6.3 to find the polynomial hS corresponding to S and check f − (Pi xi)hS. If there are more than
m inactive variables in f − (Pi xi)hS, then we set hd = hS. Therefore the running time of this process is
(cid:0)n
d(cid:1) · O(nd) = O(n2d).
Hence, we can find a polynomial hd efficiently such that at least m variables are inactive in f −
(Pi xi)hd. Let us return to the original global cardinality constraintPi xi = (1 − 2p)n. Let
fd = f −(cid:0)Xi
xi − (1 − 2p)n(cid:1)hd.
x1,··· , xm are no longer inactive in fd because of the extra term (1 − 2p)n · h. However, x1,··· , xm are
at least independent with the degree d terms in fd. Let Ad denote the set for active variables in the degree d
terms of fd, which is less than 20d·C′
VarDp(f )
from the upper bound of n − m.
d
(2p)4d
For fd, observe that VarDp(fd) = VarDp(f ) and all coefficients of fd are multiples of γd−1 = γ/d!
from Claim 6.3. For fd, we neglect its degree d terms in Ad and treat it as a degree d − 1 function from now
on. Then we could repeat the above process again for the degree d − 1 terms in fd to obtain a degree d − 2
polynomial hd−1 such that the active set Ad−1 in the degree d − 1 terms of fd−1 = fd −(cid:0)Pi xi − (1 −
2p)n(cid:1)hd−1 contains at most 20d·C′
that the degree of(cid:0)Pi xi − (1 − 2p)n(cid:1)hd−1 is at most d − 1 such that it will not introduce degree d terms
to fd−1. Then we repeat it again for terms of degree d − 2, d − 3, and so on.
To summarize, we can find a polynomial h such that Ad∪ Ad−1 ···∪ A1 is the active set in f −(cid:0)Pi xi−
(1 − 2p)n(cid:1)h. At the same time, Ad ∪ Ad−1 ··· ∪ A1 ≤Pi Ai ≤
⊓⊔
variables for C′d−1 = ((d−1)2···2!)2
20d27d·VarDp (f )·(d!)2d2
. At the same time, observe
γ2·(2p)4d
d−1
(2p)4d
VarDp(f )
p )2 + ( p
p
1 − p
· t2, we have
6.2 Proof of Theoreom 6.1
In this section, we prove Theorem 6.1. Let f = fI be the degree d function associated with the instance I
and g = f − EDp[f ] for convenience. We discuss VarDp[f ] in two cases.
If VarDp[f ] = EDp[g2] ≥ 8(cid:16)16 ·(cid:0)( 1−p
EDp[g4] ≤ 12 ·(cid:18)256 ·(cid:0)(
1−p )2(cid:1) · d3(cid:17)d
)2(cid:1)2 · d6(cid:19)d
qEDp[g2]
p )2 + ( p
from the 2 → 4 hypercontractivity in Theorem 5.2. By Lemma 2.1, we know
2r12 ·(cid:16)256 ·(cid:0)( 1−p
1−p )2(cid:1)2 · d6(cid:17)d
Thus PrDp[g ≥ t] > 0, which demonstrates that PrDp[f ≥ EDp[g] + t] > 0.
Otherwise we know VarDp[f ] ≤ 8(cid:16)16 ·(cid:0)( 1−p
· t2. We set γ = 2−d. From Theorem
6.5, we could find a degree d − 1 function h in time O(n2d) such that f −(cid:0)Pi xi − (1 − 2p)n(cid:1)h contains
variables. We further observe that f (α) = f (α) −(cid:0)Pi αi − (1 − 2p)n(cid:1)h(α) for any
1−p )2(cid:1) · d3(cid:17)d
p )2 + ( p
1 − p
p
at most C′
p,d·VarDp (f )
Pr
Dp
g ≥
EDp[g2]2
)2 + (
> 0.
γ2
d−1
.
γ2
33
α in the support of Dp. Then we know the kernel of f and I is at most
8(cid:18)16 ·(cid:0)(
)2(cid:1) · d3(cid:19)d
p
1 − p
1 − p
p
)2 + (
C′p,d
γ2
< 8(cid:18)16 ·(cid:0)(
1 − p
p
)2 + (
)2(cid:1) · d3(cid:19)d
· t2 ·
The running time of this algorithm is O(dn2d).
· t2 ·
p
1 − p
20d27d · VarDp(f ) · (d!)2d2 · 22d
γ2 · (2p)4d
< C · t2.
Acknowledgement
We would like to thank Ryan O'Donnell and Yu Zhao for useful discussion on the hypercontractive inequal-
ities. The first author is grateful to David Zuckerman for his constant support and encouragement, as well
as for many fruitful discussions.
References
[1] Noga Alon, Gregory Gutin, EunJung Kim, Stefan Szeider, and Anders Yeo. Solving MAX-r-SAT
Above a Tight Lower Bound. Algorithmica, 61(3):638 -- 655, 2011. 1, 3, 4, 5, 7, 11
[2] Per Austrin, Siavosh Benabbas, and Konstantinos Georgiou. Better Balance by Being Biased: A
0.8776-Approximation for Max Bisection. In Proceedings of the 24th Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA '13, pages 277 -- 294, 2013. 1
[3] Per Austrin and Elchanan Mossel. Approximation Resistant Predicates from Pairwise Independence.
Computational Complexity, 18(2):249 -- 271, 2009. 1
[4] Boaz Barak, Fernando G.S.L. Brandao, Aram W. Harrow, Jonathan Kelner, David Steurer, and Yuan
Zhou. Hypercontractivity, Sum-of-squares Proofs, and Their Applications. In Proceedings of the 44th
Annual ACM Symposium on Theory of Computing, STOC '14, pages 307 -- 326, 2014. 3
[5] Bonnie Berger. The Fourth Moment Method. SIAM J. Comput., 26(4):1188 -- 1207, August 1997. 4, 7
[6] Aline Bonami. ´Etude des coefficients Fourier des fonctions de Lp(G). Annales de l'Institut Fourier,
20(2):335 -- 402, 1970. 7
[7] Siu On Chan. Approximation Resistance from Pairwise Independent Subgroups. In Proceedings of the
45th Annual ACM Symposium on Theory of Computing, STOC '13, pages 447 -- 456, New York, NY,
USA, 2013. ACM. 1
[8] R. Crowston, M. Fellows, G. Gutin, M. Jones, E.J. Kim, F. Rosamond, I.Z. Ruzsa, S. Thomass´e, and
A. Yeo. Satisfying more than half of a system of linear equations over gf(2). J. Comput. Syst. Sci.,
80(4):687 -- 696, June 2014. 1
[9] Robert Crowston, Mark Jones, and Matthias Mnich. Max-cut parameterized above the edwards-erdos
bound. Algorithmica, 72(3):734 -- 757, 2015. 2
34
[10] Uriel Feige and Michael Langberg. The RPR2 Rounding Technique for Semidefinite Programs. J.
Algorithms, 60(1):1 -- 23, July 2006. 1
[11] Yuval Filmus and Elchanan Mossel. Harmonicity and Invariance on Slices of the Boolean Cube. In
Computational Complexity Conference 2016, CCC 2016, 2016. 3
[12] Alan Frieze and Mark Jerrum.
Improved approximation algorithms for max k-cut and max bisec-
tion. In Proceedings of the 4th International Conference on Integer Programming and Combinatorial
Optimization, pages 1 -- 13, 1995. 1
[13] Chris Godsil. Association Schemes. http://www.math.uwaterloo.ca/cgodsil/pdfs/assoc2.pdf,
2010. Last accessed: November 6, 2015. 4, 10
[14] Dmitry Grigoriev. Complexity of Positivstellensatz proofs for the knapsack. Computational Complex-
ity, 10(2):139 -- 154, 2001. 4, 12
[15] Venkatesan Guruswami, Johan Hastad, Rajsekar Manokaran, Prasad Raghavendra, and Moses
Charikar. Beating the Random Ordering Is Hard: Every Ordering CSP Is Approximation Resistant.
SIAM J. Comput., 40(3):878 -- 914, June 2011. 1
[16] Venkatesan Guruswami, Yury Makarychev, Prasad Raghavendra, David Steurer, and Yuan Zhou. Find-
ing Almost-Perfect Graph Bisections. In Proceedings of the 2nd Symposium on Innovations in Com-
puter Science, ICS '11, pages 321 -- 337, 2011. 1, 2
[17] Gregory Gutin, Eun Jung Kim, Stefan Szeider, and Anders Yeo. A probabilistic approach to problems
parameterized above or below tight bounds.
In Parameterized and Exact Computation, 4th Inter-
national Workshop, IWPEC 2009, Copenhagen, Denmark, September 10-11, 2009, Revised Selected
Papers, pages 234 -- 245, 2009. 3
[18] Gregory Gutin, Eun Jung Kim, Stefan Szeider, and Anders Yeo. A Probabilistic Approach to Problems
Parameterized Above or Below Tight Bounds. J. Comput. Syst. Sci., 77(2):422 -- 429, March 2011. 1
[19] Gregory Gutin and Anders Yeo. Note on maximal bisection above tight lower bound. Inf. Process.
Lett., 110(21):966 -- 969, 2010. 2
[20] Eran Halperin and Uri Zwick. A Unified Framework for Obtaining Improved Approximation Algo-
rithms for Maximum Graph Bisection Problems. Random Struct. Algorithms, 20(3):382 -- 402, May
2002. 1
[21] Johan Hastad. Some optimal inapproximability results. J. ACM, 48(4):798 -- 859, 2001. 1
[22] J. Kahn, G. Kalai, and N. Linial. The Influence of Variables on Boolean Functions. In Proceedings of
the 29th Annual Symposium on Foundations of Computer Science, FOCS '88, pages 68 -- 80, 1988. 3
[23] Subhash Khot. On the Power of Unique 2-prover 1-round Games. In Proceedings of the 34th Annual
ACM Symposium on Theory of Computing, STOC '02, pages 767 -- 775, New York, NY, USA, 2002.
ACM. 1, 2
[24] Subhash Khot, Madhur Tulsiani, and Pratik Worah. A characterization of strong approximation resis-
tance. In Proceedings of the 46th ACM STOC, pages 634 -- 643, 2014. 1
35
[25] Tzong-Yau Lee and Horng-Tzer Yau. Logarithmic Sobolev inequality for some models of random
walks. Ann. Prob., 26(4):1855 -- 1873, 1998. 3
[26] Konstantin Makarychev, Yury Makarychev, and Yuan Zhou. Satisfiability of Ordering CSPs Above
Average Is Fixed-Parameter Tractable. In Proceedings of the 56th Annual IEEE Symposium on Foun-
dations of Computer Science, FOCS '15, pages 975 -- 993, 2015. 1, 3
[27] Matthias Mnich and Rico Zenklusen. Bisections above tight lower bounds. In the 38th International
Workshop on Graph-Theoretic Concepts in Computer Science, WG '12, pages 184 -- 193, 2012. 2
[28] Elchanan Mossel, Ryan O'Donnell, and Krzysztof Oleszkiewicz. Noise stability of functions with
In Proceedings of the 46th Annual IEEE Symposium on
low influences: invariance and optimality.
Foundations of Computer Science, FOCS '05, pages 21 -- 30, 2005. 3, 6
[29] Elchanan Mossel, Ryan O'Donnell, and Krzysztof Oleszkiewicz. Noise stability of functions with low
influences: invariance and optimality. Annals of Mathematics, 171(1):295 -- 341, 2010. 3
[30] Ryan O'Donnell. Analysis of Boolean Functions. Cambridge University Press, 2014. 3, 4, 7
[31] Prasad Raghavendra and David Steurer. Graph expansion and the unique games conjecture. In Pro-
ceedings of the 42nd ACM STOC, pages 755 -- 764, 2010. 1, 2
[32] Prasad Raghavendra, David Steurer, and Prasad Tetali. Approximations for the isoperimetric and
spectral profile of graphs and related parameters. In Proceedings of the 42nd ACM STOC, pages 631 --
640, 2010. 1
[33] Prasad Raghavendra, David Steurer, and Madhur Tulsiani. Reductions between expansion problems.
In Proceedings of the 27th Conference on Computational Complexity, pages 64 -- 73, 2012. 1
[34] Prasad Raghavendra and Ning Tan. Approximating CSPs with global cardinality constraints using
SDP hierarchies. In Proceedings of the 23rd Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA '12, pages 373 -- 387, 2012. 1, 2
[35] Yinyu Ye. A .699-approximation algorithm for Max-Bisection. Math. Program., 90(1):101 -- 111, 2001.
1
36
|
1606.04978 | 1 | 1606 | 2016-06-15T20:54:04 | Distance geometry approach for special graph coloring problems | [
"cs.DS"
] | One of the most important combinatorial optimization problems is graph coloring. There are several variations of this problem involving additional constraints either on vertices or edges. They constitute models for real applications, such as channel assignment in mobile wireless networks. In this work, we consider some coloring problems involving distance constraints as weighted edges, modeling them as distance geometry problems. Thus, the vertices of the graph are considered as embedded on the real line and the coloring is treated as an assignment of positive integers to the vertices, while the distances correspond to line segments, where the goal is to find a feasible intersection of them. We formulate different such coloring problems and show feasibility conditions for some problems. We also propose implicit enumeration methods for some of the optimization problems based on branch-and-prune methods proposed for distance geometry problems in the literature. An empirical analysis was undertaken, considering equality and inequality constraints, uniform and arbitrary set of distances, and the performance of each variant of the method considering the handling and propagation of the set of distances involved. | cs.DS | cs |
Distance geometry approach for special graph coloring problems
Rosiane de Freitasa, Bruno Diasa, Nelson Maculanb, and Jayme Szwarcfiterb
aInstituto de Computa¸cao, Universidade Federal do Amazonas, Av. Rodrigo Ot´avio 3000, 69000-000, Manaus,
bCOPPE, Universidade Federal do Rio de Janeiro, C.P. 68530, 21945-970, Rio de Janeiro, Brazil
E-mail: [email protected] [deFreitas]; [email protected] [Dias]; [email protected]
[Maculan]; [email protected] [Szwarcfiter]
Brazil
Abstract
One of the most important combinatorial optimization problems is graph coloring. There are
several variations of this problem involving additional constraints either on vertices or edges. They
constitute models for real applications, such as channel assignment in mobile wireless networks. In
this work, we consider some coloring problems involving distance constraints as weighted edges,
modeling them as distance geometry problems. Thus, the vertices of the graph are considered
as embedded on the real line and the coloring is treated as an assignment of positive integers to
the vertices, while the distances correspond to line segments, where the goal is to find a feasible
intersection of them. We formulate different such coloring problems and show feasibility conditions
for some problems. We also propose implicit enumeration methods for some of the optimization
problems based on branch-and-prune methods proposed for distance geometry problems in the
literature. An empirical analysis was undertaken, considering equality and inequality constraints,
uniform and arbitrary set of distances, and the performance of each variant of the method consid-
ering the handling and propagation of the set of distances involved.
Keywords: branch-and-prune; channel assignment; constraint propagation; graph theory; T-coloring.
1 Introduction
Let G = (V, E) be an undirected graph. A k-coloring of G is an assignment of colors {1, 2, . . . , k}
to the vertices of G so that no two adjacent vertices share the same color. The chromatic number χG
of a graph is the minimum value of k for which G is k-colorable. The classic graph coloring problem,
which consists in finding the chromatic number of a graph, is one of the most important combinatorial
optimization problems and it is known to be NP-hard (Karp, 1972).
There are several versions of this classic vertex coloring problem, involving additional constraints,
in both edges as vertices of the graph, with a number of practical applications as well as theoretical
challenges. One of the main applications of such problems consists of assigning channels to transmitters
in a mobile wireless network. Each transmitter is responsible for the calls made in the area it covers
and the communication among devices is made through a channel consisting of a discrete slice of the
electromagnetic spectrum. However, the channels cannot be assigned to calls in an arbitrary way,
since there is the problem of interference among devices located near each other using approximate
channels. There are three main types of interferences: co-channel, among calls of two transmitters
using the same channels; adjacent channel, among calls of two transmitters using adjacent channels
and co-site, among calls on the same cell that do not respect a minimal separation. It is necessary
to assign channels to the calls such that interference is avoided and the spectrum usage is minimized
(Audhya et al., 2011; Koster and Munhoz, 2010; Koster, 1999).
Thus, the channel assignment scenario is modeled as a graph coloring problem by considering each
transmitter as a vertex in a undirected graph and the channels to be assigned as the colors that the
vertices will receive. Some more general graph coloring problems were proposed in the literature in
order to take the separation among channels into account, such as the T-coloring problem, also known
as the Generalized Coloring Problem (GCP) where, for each edge, the absolute difference between
1
Figure 1: Example of channel assignment with distance constraints, where the separation is given by the weight
in each edge. The image on the right shows the network as an undirected graph and the projection of vertices
on the real number line, but considering only natural numbers.
colors assigned to each vertex must not be in a given forbidden set (Hale, 1980). The Bandwidth
Coloring Problem, a special case of T-coloring where the absolute difference between colors assigned
to each vertex must be greater or equal a certain value (Malaguti and Toth, 2010), and the coloring
problem with restrictions of adjacent colors (COLRAC), where there is a restriction graph for which
adjacent colors in it cannot be assigned to adjacent vertices (Akihiro et al., 2002).
The separation among channels is a type of distance constraint, so we can see the channel assign-
ment as a type of distance geometry (DG) problem (Liberti et al., 2014) since we have to place the
channels in the transmitters respecting some distances imposed in the edges, as can be seen in Figure
1. One method to solve DG problems is the branch-and-prune approach proposed by Lavor et al.
(2012a,b), where a solution is built and if at some point a distance constraint is violated, then we stop
this construction (prune) and try another option for the current solution in the search space. See also:
Mucherino et al. (2013); Lavor et al. (2012a); Freitas et al. (2014a,b); Dias (2014); Dias et al. (2013,
2012).
For graph theoretic concepts and terminology, see the book by Bondy and Murty (2008).
The main contribution of this paper consists of a distance geometry approach for special cases of
T-coloring problems with distance constraints, involving a study of graph classes for which some of
these distance coloring problems are unfeasible, and branch-prune-and-bound algorithms, combining
concepts from the branch-and-bound method and constraint propagation, for the considered problems.
The remainder of this paper is organized as follows. Section 2 defines the distance geometry models
for some special graph coloring problems. Section 3 shows some properties regarding the structure
of those distance geometry graph coloring problems, including the determination of feasibility for
some graphs classes. Section 4 formulates the branch-prune-and-bound (BPB) algorithms proposed
for the problems and shows properties regarding optimality results. Section 5 shows results of some
experiments done with the BPB algorithms using randomly generated graphs for each proposed model.
Finally, Section 6 concludes the paper and states the next steps for ongoing research.
2 Distance geometry and graph colorings
We propose an approach in distance geometry for special vertex coloring problems with distance
constraints, based on the Discretizable Molecular Distance Geometry Problem (DMDGP), which is a
special case of the Molecular Distance Geometry Problem, where the set V of vertices from the input
graph G are ordered such that the set E of edges contain all cliques on quadruplets of consecutive
vertices, that is, any four consecutive vertices induce a complete graph (∀i ∈ {4, . . . , n} ∀j, k ∈
{i − 3, . . . , i} ({j, k} ∈ E)) (Lavor et al., 2012a). Furthermore, a strict triangular inequality holds
on weights of edges between consecutive vertices in such ordering (∀i ∈ {2, . . . , n − 1} di−1,i+1 <
di−1,i + di,i+1). All coordinates are given in R3 space. The position for a point i (where i ≥ 4) can be
2
123544132512431Channel ofstation 1 = 5Channel ofstation 2 = 1Channel ofstation 3 = 6Channel ofstation 4 = 2Channel ofstation 5 = 314325Color forvertex 1 = 5Color forvertex 3 = 6Color forvertex 4 = 24541113322Color forvertex 2 = 1Color forvertex 5 = 312345(ℕ ⊂ ℝ)ℝ16 Figure 2: Some types of n-spheres. A (n − 1)-sphere is a projection of a n-sphere on a lower dimension.
Figure 3: Example from Figure 1 using 0-spheres (line segments).
determined using the positions of the previous three points i − 1, i − 2 and i − 3 by intersecting three
spheres with radii di−3,i, di−2,i and di−1,i, obtaining two possible points that are checked for feasibility.
A similar reasoning can be used in vertex coloring problems with distance constraints, where the
distances that must be respected involve the absolute difference between two values x(i) and x(j)
(respectively, the color points attributed to i and j), but for these problems the space considered is
actually unidimensional. The positioning of a vertex i can be determined by using a neighbor j that
is already positioned. Thus, we have a 0-sphere, consisting of a projection of a 1-sphere (a circle),
which itself is a projection of a 2-sphere (the three-dimensional sphere), as shown in Figure 2. The
0-sphere is a line segment with a radius di,j, and feasible colorings consist of treating the intersections
of these 0-spheres. Figure 3 exemplifies the correlations between these types of spheres and shows the
example from Figure 1 as the positioning of these line segments.
In this work we focus on problems with exact distances between colors, and also in the analysis of
different types of BPB algorithms and integer programming models.
Based on DMDGP, which is a decision problem involving equality distance constraints, the basic
distance graph coloring model we consider also involves equality constraints between colors of two
neighbor vertices i and j. That is, the absolute difference between them must be exactly equal to
an arbitrary weight imposed on the edge (i, j), and the solution candidate must satisfy all given
constraints. We can formally define as follows.
Given a graph G = (V, E), we define di,j as a positive integer weight associated to an edge
(i, j) ∈ E(G).
In distance coloring, for each vertex i, a color must be determined for it (denoted
by x(i)) such that the constraints imposed on the edges between i and its neighbors are satisfied.
A variation of the classic graph coloring problem consists in finding the minimum span of G, that
is, in determining that the maximum x(i), or color used, be the minimum possible. Based on these
preliminary definitions, we describe the following distance geometry vertex coloring problems.
3
2-sphere (ℝ3)1-sphere (ℝ2)0-sphere (ℝ1)14325Color forvertex 1 = 5Color forvertex 3 = 6Colorforvertex4=24541113322Color forvertex 2 = 1Color forvertex 5 = 312345(ℕ ⊂ ℝ)ℝ1612345(ℕ ⊂ ℝ)ℝ16(2,5) = 2(5,3) = 3(2,4) = 1(4,1) = 3(4,5) = 1(2,3) = 5(5,1) = 2(1,3) = 1(2,1) = 4(4,3) = 4 Figure 4: Specific order of 0-spheres that leads to the optimal solution for Figure 1.
Definition 1. Coloring Distance Geometry Problem (CDGP): Given a simple weighted undi-
rected graph G = (V, E), where, for each (i, j) ∈ E, there is a weight di,j ∈ N, find an embedding
x : V → N (that is, an embedding of G on the real number line, but considering only the natural
number points) such that x(i) − x(j) = di,j for each (i, j) ∈ E.
CDGP involves equality constraints, and thus is named as Equal Coloring Distance Geometry
Problem and labeled as EQ-CDGP. A solution for this problem consists of a tree, whose vertices are
colored with colors that respect the equality constraints involving the weighted edges (see Figure 4).
Since CDGP (or EQ-CDGP) is a decision problem, only a feasible solution is required. This problem
is NP-complete, as shown below.
Theorem 1. EQ-CDGP is NP-complete.
Proof. To prove that EQ-CDGP ∈ NP-complete, we must show that EQ-CDGP ∈ NP and EQ-CDGP
∈ NP-hard.
1. EQ-CDGP ∈ NP.
Given, for a graph G = (V, E), an embedding x : V → N, its feasibility can be checked by taking each
edge (i, j) ∈ E and examining if its endpoints do not violate the corresponding distance constraint,
that is, if x(i) − x(j) = di,j. If all distance constraints are valid, then x is a certificate for a positive
answer to the EQ-CDGP instance, meaning that a certificate for a YES answer can be verified in
O(E) time, which is linear. Thus, EQ-CDGP ∈ NP.
2. EQ-CDGP ∈ NP-hard.
Since EQ-CDGP is equivalent to 1-Embeddability with integer weights, which is NP-hard (Saxe,
1979), we can use the same proof for the latter problem to show that EQ-CDGP is also NP-hard. The
proof is made by reducing the Partition problem, which is known to be NP-complete (Garey and
Johnson, 1979) to EQ-CDGP.
Consider a Partition instance, consisting of a set I of r integers, that is, M = {m1, m2, . . . , mr}. Let
G be a weighted graph G = (V, E), where G is a cycle such that V = E = r and, for each edge (i, j),
its weight is a natural number denoted by di,j. This graph is constructed from M by considering:
• V = {i0, i1, . . . , ir−1}.
• E = {(ib, ib+1 mod r) 0 ≤ b ≤ r}.
• dib,ib+1 mod r = mb (∀0 ≤ b ≤ r).
Now, let x : V → N be an embedding of the vertices on the number line. If it is a valid embedding,
then we can define two sets:
• S1 = {mb x(ib) < x(ib+1 mod r)}.
4
14325Color forvertex 1 = 5Color forvertex 3 = 6Color forvertex 4 = 251133Color forvertex 2 = 1Color forvertex 5 = 3(2,4) = 1(4,5) = 1(4,1) = 3(2,3) = 5Vertex order: 2 → 4 → 5 → 1 → 3 (a) Instance with a YES solution.
(b) Instance with a NO solution.
Figure 5: Partition instances and corresponding transformations to EQ-CDGP.
• S2 = {mb x(ib) > x(ib+1 mod r)}.
We have that S1 and S2 are disjoint subsets of M (that is, they form a partition of M ) where the
sum of all S1 elements is equal to the sum of all S2 elements, that is, if the cyclic graph constructed
from G admits an embedding on the line (which means that its solution to EQ-CDGP is YES), then
M has a YES solution for Partition and vice-versa. This reduction can be made in O(r) time, thus,
EQ-CDGP ∈ NP-hard.
To illustrate the reasoning from Theorem 1, let M be an instance of Partition such that M =
{1, 4, 5, 6, 7, 9}. Figure 5 shows its corresponding EQ-CDGP solution.
Since most graph coloring problems in the literature and in real world applications are optimization
problems, we define an optimization version of this basic distance geometry graph coloring problem,
as shown below.
Definition 2. Minimum Equal Coloring Distance Geometry Problem (MinEQ-CDGP):
Given a simple weighted undirected graph G = (V, E), where, for each (i, j) ∈ E, there is a weight
di,j ∈ N, find an embedding x : V → N such that x(i) − x(j) = di,j for each (i, j) ∈ E whose span S,
defined as S = maxi∈V x(i), that is, the maximum used color, is the minimum possible.
Figure 6 shows an example of this model and its corresponding 0-sphere visualization.
5
x(5) = 1034x(3)=2125= 1x(0) = 10x(1) = 11x(2) = 7x(4) = 8M = {1, 4, 5, 6, 7, 9}S1 = {1, 6, 9} S2 = {4, 5, 7}Sum(S2) = Sum(S1) = 16x(0) < x(1)m1 → S1x(1) > x(2)m2 → S2x(2) > x(3)m3 → S2x(3) < x(4)m4 → S1x(4) > x(5)m5 → S2x(5) < x(0)m6→S1= 4= 5= 6= 7= 9034x(3) = 4125= 2x(0) = 10x(1) = 12x(2) = 9x(4) = 10M = {2, 3, 5, 6, 7, 10}S1 = {2, 4} S2 = {3, 5}7 and 10 → undefined Infeasible for Partitionx(0) < x(1)m1 → S1x(1) > x(2)m2 → S2x(2) > x(3)m3 → S2x(3) < x(4)m4 → S1??Cannot beembeddedx(5) = ?= 3= 5= 6= 7= 10 Figure 6: MinEQ-CDGP instance with solution and its 0-sphere representation.
On the other hand, instead of equalities, we can consider inequalities, such that the weight di,j on
an edge (i, j) is actually a lower bound for the distance to be respected between the color points x(i)
and x(j), that is, x(i) − x(j) ≥ dij. Thus, we can modify MinEQ-CDGP to deal with this scenario,
which becomes the following model.
Definition 3. Minimum Greater than or Equal Coloring Distance Geometry Problem
(MinGEQ-CDGP): Given a simple weighted undirected graph G = (V, E), where, for each (i, j) ∈ E,
there is a weight di,j ∈ N, find an embedding x : V → N such that x(i)−x(j) ≥ di,j for each (i, j) ∈ E
whose span (max
i∈V
x(i)) is the minimum possible.
MinGEQ-CDGP is equivalent to the bandwidth coloring problem (BCP) (Malaguti and Toth,
2010), which itself is equivalent to the minimum span frequency assignment problem (MS-FAP)
(Koster, 1999; Audhya et al., 2011).
Figure 6 In Figure 7, this model, along with its 0-sphere representation, is exemplified.
Figure 7: MinGEQ-CDGP instance with solution and its 0-sphere representation.
2.1 Special cases
For the models previously stated, we can identify some specific scenarios for which additional
properties can be identified. The first special case is for EQ-CDGP, the decision distance coloring
problem, where all distances are the same, that is, the input is a graph with uniform edge weights, as
stated below.
Definition 4. Coloring Distance Geometry Problem with Uniform Distances (EQ-CDGP-
Unif ): Given a simple weighted undirected graph G = (V, E), and a nonnegative integer ϕ, find an
embedding x : V → N such that x(i) − x(j) = ϕ for each (i, j) ∈ E.
For the optimization version, we can also define this special case, as shown below.
Definition 5. Minimum Equal Coloring Distance Geometry Problem with Uniform Dis-
tances (MinEQ-CDGP-Unif ): Given a simple weighted undirected graph G = (V, E), and a
6
x(1) = 1135= 3= 1x(3) = 2x(5) = 3x(6) = 1= 2x(2) = 32= 1= 2= 34x(4) = 461234(ℕ ⊂ ℝ)ℝ1(1,2) = 2(1,4) = 3(3,4) = 2(6,4) = 3(3,2) = 1(5,4) = 1x(1) = 114x(4) = 43x(2) = 425≥ 1≥ 3≥ 2≥ 2≥ 3≥ 2x(3) = 1x(5) = 2x(6) = 261234(ℕ ⊂ ℝ)ℝ1(3,2) = 3(1,4) = 3(5,4) = 2(1,2) = 3(3,4) = 3(6,4) = 2 nonnegative integer ϕ, find an embedding x : V → N such that x(i) − x(j) = ϕ for each (i, j) ∈ E
whose span (maxi∈V x(i)) is the minimum possible.
In this model, an input graph can be defined by its sets of vertices and edges and the ϕ value,
instead of a set of weights for each edge. A similar special case exists for MinGEQ-CDGP, as stated
in the following definition.
Definition 6. Minimum Greater than or Equal Coloring Distance Geometry Problem
with Uniform Distances (MinGEQ-CDGP-Unif ): Given a simple weighted undirected graph
G = (V, E), and a nonnegative integer ϕ, find an embedding x : V → N such that x(i)− x(j) ≥ ϕ for
each (i, j) ∈ E whose span (maxi∈V x(i)) is the minimum possible.
When ϕ = 1, MinGEQ-CDGP-Unif is equivalent to the classic graph coloring problem (Figure 8).
A summary of all distance coloring models, including special cases, is given in Table 1.
Table 1: Summary of distance coloring models.
Problem
Constraint type
EQ-CDGP and MinEQ-CDGP
∀(i, j) ∈ E, x(i)−x(j) = di,j
∀(i, j) ∈ E, x(i)−x(j) ≥ di,j
EQ-CDGP-Unif and MinEQ-CDGP-Unif ∀(i, j) ∈ E, x(i)−x(j) = di,j
MinGEQ-CDGP
MinGEQ-CDGP-Unif
∀(i, j) ∈ E, x(i)−x(j) ≥ di,j
Distance type
∀(i, j) ∈ E, di,j ∈ N
∀(i, j) ∈ E, di,j ∈ N
∀(i, j) ∈ E, di,j = ϕ
∀(i, j) ∈ E, di,j = ϕ
(ϕ ∈ N)
(ϕ ∈ N)
(a) MinEQ-CDGP-Unif.
(b) MinGEQ-CDGP-Unif.
Figure 8: Examples of instances for the special cases of distance coloring models with constant edge weights
and feasible solutions for them.
3 Feasibility conditions of distance graph coloring problems
In this section, we discuss feasibility conditions related to our proposed EQ-CDGP problems.
Clearly, the problems involving inequality constraints are always feasible. This is the case for the
MinGEQ-CDGP and MinGEQ-CDGP problems (and the special cases with uniform distances, MinGEQ-
CDGP-Unif and MinGEQ-CDGP-Unif). However, this is not so for versions that involve only equality
constraints, EQ-CDGP and its special case with uniform distances, the EQ-CDGP-Unif problem.
3.1 Feasibility conditions for EQ-CDGP-Unif
Graphs that admit a solution for the EQ-CDGP-Unif problem are characterized by the following
theorem.
Theorem 2. A graph G has solution YES for EQ-CDGP-Unif problem if and only if G is bipartite.
7
x(1) = 1143x(2) = 3256= 2= 2= 2= 2= 2= 2x(6) = 1x(4) = 3x(3) = 1x(5) = 1x(1) = 114x(4) = 23x(2) = 2256≥ 1≥ 1≥ 1≥ 1≥ 1≥ 1x(3) = 1x(5) = 1x(6) = 1 Proof. Let G be a graph, input to a EQ-CDGP-Unif problem, where for each edge vivj of G, the
distance required is dij = ϕ, ϕ ∈ N, constant. Suppose G has a YES solution for the problem such
that x : V → N is a certificate for that solution. Let x(i) be the color assigned to vi ∈ V . Choose an
arbitrary path v1, v2, ..., vk of G, not necessarily simple. Then x(i) − x(j) = ϕ, for i − j = 1. The
latter implies x(i) = x(i + 2), i = 1, 2, ..., k − 2. Consequently, if the path contains the same vector vi
twice, their corresponding indices are the same. That is, all edges of G are necessarily even, and G is
bipartite.
Conversely, if G is bipartite, its vertices admit a proper coloring with two distinct colors. Assign the
value x(i) to the vertices of the first color, and the value ϕ+1 to the second one. Then x(i)−x(j) = ϕ,
for each edge vivj of G, and EQ-CDGP-Unif has a YES solution. As an alternative way of proving that
if a graph is bipartite then it has a YES solution for EQ-CDGP, observe that, since the input graph
is bipartite, it is also 2-colorable (considering the classic graph coloring problem), that is, the entire
graph can be colored using only two different colors, which can be determined by considering a single
edge from the graph. All edges (i, j) have the same distance constraint, that is, x(i) − x(j) = ϕ, so
the two colors that will be used are {1, 1+ϕ}, which form the solution for the EQ-CDGP-Unif instance.
In order to prove the converse statement, that is, if a graph has a YES solution for EQ-CDGP, it is
bipartite, we will use a proof by contrapositive, which states that if a graph is not bipartite, then it
has a NO solution for EQ-CDGP. This will be done by mathematical induction on odd cycles, since
a graph is not bipartite if, and only if, it contains an odd cycle. Let V = 2z + 1. The proof will be
by induction on z (the number of vertices).
Base case: z = 1. We have the cycle C3, with three vertices (V = {1, 2, 3}) and three edges
({(1, 2), (1, 3), (2, 3)}), with x(i) − x(j) = ϕ for all of them. Without loss of generality, let x(1) = 1
and x(2) = 1 + ϕ. Then we have that:
• Since (1, 3) ∈ E and x(1) = 1, then x(3) − 1 = ϕ. All colors must be positive integers, so
x(3) = 1 + ϕ.
• Since (2, 3) ∈ E and x(2) = 1 + ϕ, then x(3) − (1 + ϕ) = ϕ ⇒ x(3) − 1 − ϕ = ϕ. By this
inequation, x(3) = 1 or x(3) = 1 + 2ϕ.
From this result, we have that x(3) = 1 + ϕ and (x(3) = 1 or x(3) = 1 + 2ϕ) at the same time, which
is impossible. Then C3 has a NO solution for EQ-CDGP, as seen in Figure 9.
Induction hypothesis: The cycle C2z+1 has a NO solution for EQ-CDGP.
Inductive step: By the inductive hypothesis, the cycle C2z+1 is infeasible for EQ-CDGP. If we
consider the cycle C2(z+1)+1 = C2z+3, we have that the size of the cycle increases by two vertices, but
it will still be an odd cycle. If we add only one vertex, that is, we consider the cycle C2z+1+1 = C2z+2,
we will have an even cycle. Since all even cycles are bipartite, they are feasible in EQ-CDGP according
to Theorem 2. Now, consider that another vertex is added to C2z+2, becoming C2z+3. Without loss
of generality, consider that the new vertex 2z + 3 is adjacent to vertices 2z + 2 and 1, that is, we have
{(2z + 2, 2z + 3), (2z + 3, 1)} ⊆ E, and x(2z + 2) = 1 + ϕ and x(1) = 1 (these colors can be seen as
having been assigned when we added only one vertex, generating an even cycle). Then we have that:
• Since (2z + 2, 2z + 3) ∈ E and x(2z + 2) = 1 + ϕ, then x(2z + 3) − (1 + ϕ) = ϕ ⇒
x(2z + 3) − 1 − ϕ = ϕ. By this inequation, x(2z + 3) = 1 or x(2z + 3) = 1 + 2ϕ.
• Since (2z + 3, 1) ∈ E and x(1) = 1, then x(2z + 3)− 1 = ϕ. All colors must be positive integers,
so x(2z + 3) = 1 + ϕ.
From this result, we have that x(2z + 3) = 1 + ϕ and (x(2z + 3) = 1 or x(2z + 3) = 1 + 2ϕ) at the
same time, which is impossible. Therefore C2z+3 has a NO solution for EQ-CDGP, as can be seen in
Figure 10.
8
Figure 9: C3 graph that has a NO solution for EQ-CDGP-Const when all distances are the same.
Figure 10: Odd cycle C2z+3 that has a NO solution for EQ-CDGP-Const when all distances are the same.
As a complementary result, graphs which have odd-length cycles as induced subgraphs will always
have a NO solution for EQ-CDGP-Unif, because a graph is bipartite if, and only if, it contains no
odd-length cycles. Since the recognition of bipartite graphs can be done in linear time using a graph
search algorithm such as depth-first search (DFS), the EQ-CDGP-Unif problem can be solved in linear
time.
3.2 Feasibility conditions for EQ-CDGP
Clearly, Theorem 2 does not apply when the distances are arbitrarily defined. Depending on the
edge weights, bipartite graphs may have NO solutions for EQ-CDGP, and graphs which include odd-
length cycles may have YES solutions. Figure 11 shows examples of instances considering each case.
However, this decision problem can be easily solved for trees, as shown below.
Theorem 3. Let G = (V, E, d), be a tree, where ∀(i, j) ∈ E di,j is an arbitrary positive integer. Then
G always has a YES solution for EQ-CDGP.
Proof. We describe a simple algorithm for assigning colors that satisfy the EQ-CDGP problem.
Initially unmark all vertices. Choose an arbitrary vertex vi, assign any positive integer value x(vi) to
vi, and mark vi. Iteratively, choose an unmarked vertex vj, adjacent to some marked vertex vk. Assign
the value x(vj) = djk + x(vj) and mark vj. Repeat the iteration until all vertices become marked.
The algorithm described in Theorem 3 has linear time complexity. It is important to note that
when this procedure is applied to a MinEQ-CDGP instance, that is, the optimization problem with
equality constraints, it only guarantees that a feasible solution is found for a tree, not the optimal
one.
4 Algorithmic techniques and methods to solve EQ-CDGP models
In this section, we show some algorithmic strategies to solve our distance geometry graph coloring
models, and discuss some algorithmic strategies considering the EQ-CDGP models proposed in the
previous section.
4.1 Branch-prune-and-bound methods
For solving the three distance geometry graph coloring models shown in Section 2, we developed
three algorithms that combine concepts from constraint propagation and optimization techniques.
9
= φ111+φ2= φ= φ3Cannot becolored111+φ22z+3131+φ42z+2=φ=φ=φ=φTo becolored (a) Bipartite graphs.
(b) Graphs with odd-length cycles.
Figure 11: Examples of instances for the special cases of distance coloring models with constant edge weights
and feasible solutions for them.
A branch-and-prune (BP) algorithm was proposed by Lavor et al. (2012a) for the Discretizable
Molecular Distance Geometry Problem (DMDGP), based on a previous version for the MDGP by
Liberti et al. (2008). The algorithm proceeds by enumerating possible positions for the vertices that
must be located in three-dimensional space (R3), by manipulating the set of available distances. The
position for a vertex i, where i ∈ [4, n] and n is the number of vertices that must be placed in R3,
is determined with respect to the last three vertices that have already been positioned, following
the ordering and sphere intersection cited in Section 2. However, a distance between the currently
positioned vertex and a previous one that was placed before the last three can be violated, which
requires feasibility tests to guarantee that the solution being built is valid. The authors applied the
Direct Distance Feasibility (DDF) pruning test, where ∀(i, j) ∈ E x(i)− x(j)− di,j < , and where
is a given tolerance.
In this work, we adapted these concepts to study and solve our proposed distance geometry coloring
models. One of the first reflections that can be made is that for the distance geometry coloring models,
there are no initial assumptions to be respected, and thus, there is no explicit vertex ordering to be
considered, so we build the ordering by an implicit enumerating process. We mix concepts from branch-
and-prune for DMDGP and branch-and-bound procedures to obtain partial solutions (sequences of
vertices that have already been colored) that cannot improve on the current best solution.
Our branch-prune-and-bound (BPB) method works as follows. First, a vertex i that has not been
colored yet is selected as a starting point. This vertex receives the color 1, which is the lowest available
(since all colors are positive integers). Then a neighbor j of i that has not been colored yet is selected.
A color selection algorithm is used for setting a color to j and the process is repeated recursively for
neighbors of j that have not been colored yet. When an uncolored neighbor of the current vertex
cannot be found, a uncolored vertex of the graph is used. Pseudocode for this general procedure is
given in Algorithm 1.
We propose different strategies for selecting a color for a vertex and illustrate how the feasibility
checking can be done in different levels of the procedure. Each of these cases are discussed below.
Color selection for a vertex
There are two possibilities for determining which colors a vertex can use, determined by the call
to SelectColors()), which returns a set of possible colors for a vertex.
10
x(1) = 113x(3) = 2= 2x(2) = 32= 1= 2= 34x(4) = 4x(1) = 113x(3) = 3= 3x(2) = 42= 1= 3= 24Cannot becoloredx(1) = 113x(2) = 22Cannot becolored= 1= 3= 1x(3) = 113x(1) = 22= 1= 1x(2) = 3= 2 Algorithm 1 Branch-prune-and-bound general algorithm.
Require: graph G (with set V of vertices and set E of edges), function d : E → N of distances for each edge,
previous vertex i, current vertex j to be colored, current partial coloring x, best complete coloring found
xbest, upper bound ub, array pred of predecessors from each vertex (initially all set to -1) and enumeration
tree depth dpt.
predec[k] ← i
for each neighbor k of j do
if predec[k] = −1 then
1: function Branch-Prune-And-Bound(G = (V, E), d, i, j, x, xbest, ub, pred, dpt)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
if i = −1 then
i ← predec[j]
colorsAvail ← SelectColors(G, d, i, j, x, ub)
while colorsAvail (cid:54)= ∅ do
color ← element of colorsAvail
colorsAvail ← colorsAvail − {color}
x(j) ← color
max
if
x(v) ≥ ub then
(cid:46) Set current vertex as predecessor of neighbors
(cid:46) If this call did not come from a neighbor, use predecessor information
v∈V v is colored
Remove color from i
continue
(cid:46) Discard this possible partial solution by bounding
if FeasibilityTest(G, d, f, x, i) = false then
Remove color from i
return
if dpt = V then
if max
v∈V
xbest ← x
ub ← max
v∈V
x(v)
x(v) < max
v∈V
xbest(v) then
(cid:46) Distance violation, discard partial solution by pruning
(cid:46) If true, then all vertices are colored
else
hasN eighbor ← false
for each neighbor k of j do
if k is not colored then
hasN eighbor ← true
Branch-Prune-And-Bound(G, d, f, j, k, x, xbest, ub, dpt + 1)
if hasN eighbor = false then
for each vertex k of G such that predec[k] (cid:54)= −1 do (cid:46) Only from vertices with predecessors
if k is not colored then
Branch-Prune-And-Bound(G, d, f,−1, k, x, xbest, ub, dpt + 1)
Remove color from i
return xbest
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
The first one, denoted by BPB-Prev, is based on the original BP algorithm by Lavor et al. (2012a).
When a vertex i has to be colored, the single previously colored vertex j is taken into account. If j
is an invalid vertex, which means that i is not an uncolored neighbor of j, then the only color that i
can receive is 1. Otherwise, the function returns a set of cardinality at most 2, whose elements are:
1. x(j) + di,j.
2. x(j) − di,j (returned only if x(j) > di,j).
This means that this criterion uses only information from the previous vertex to determine colors,
which makes the BPB that uses it an inexact algorithm, something that the original BP for DMDGP
also is (Lavor et al., 2012a). However, to counter this in our BPB, when a vertex is colored, its
neighbors are marked so that they can use the current vertex as a predecessor in case the search
restarts from one of such neighbors. Since we assume the input graph is connected and the algorithm
essentially walks through the graph, this information helps to find the true optimal solution. This
procedure is done in O(1) time, since only two arithmetic operations are made to determine the colors.
An example of this color selection possibility is given in Figure 12.
When using this criterion, we apply the feasibility checking at each colored vertex. However, an
alternative is to prune only infeasible solutions where all vertices have colors, that is, we apply the
11
Figure 12: Partial enumeration of solutions starting from vertex 2 for the MinEQ-CDGP instance defined by
Figures 1 and 2 using BPB-Prev, with color selection based only on the previous vertex and feasibility checking
at each partial solution.
feasibility test only at the last level of the enumeration tree. An example of this alternative is shown
in Figure 13, where it is possible to see that this strategy makes the tree grow very large.
The second selection criterion is undertaken using information from all colored neighbors to deter-
mine the color for the current vertex i. This is done by solving a system of absolute value inequalities
(or equalities, in the case of MinEQ). Those inequalities arise from the distance constraints for the
edges. Let i be the vertex that must be colored. The color x(i) must be the solution of a system of
absolute value (in)equalities where there is one for each colored neighbor j and each one is as follows:
x(j) − x(i) OP di,j
Where OP is either "=" (for MinEQ-CDGP type problems) or "≥" (for MinGEQ-CDGP type prob-
lems). The color that will be assigned to j is the smallest value that satisfies all (in)equalities. We note
that this procedure always returns a set of cardinality 1, that is, only one color (since only the lowest
index is returned) which is also feasible for the partial solution and eventually leads to the optimal
solution, although it requires more work per vertex. This selection strategy runs in O(ub) time, where
ub is an upper bound for the span, since, to solve the system, we have to mark each possible solution
in the interval [1, ub] and select the smallest value. Figure 14 shows an example of an enumeration
tree using this color selection strategy.
Feasibility checking
When building a partial solution we must verify if it is feasible when not all distances are taken
into account at the same time, especially on BPB-Prev. We used a similar feasibility test to the Direct
Distance Feasibility (DDF) used on the BP algorithm for the DMDGP.
Let i be the vertex that has just been colored. Then we must check, for each neighbor j that has
already been colored, if the condition x(i) − x(j) ≥ di,j (if f ((i, j)) = 0) or x(i) − x(j) = di,j (if
f ((i, j)) = 1). This test can be seen as a variation of DDF setting to zero and allowing inequalities in
the test. We denote this procedure as Direct Distance Coloring Feasibility (DDCF) and its pseudocode
12
Vertex 2Color 1Vertex 1Color 5Vertex 1Color -34+-Vertex 3Color 6Vertex 4Color 101+4+Pruned x(4) - x(1) = 3 10 - 5 = 35 = 3 --> FalseVertex 4Color 2-Vertex 5Color 31+Vertex 5Color 1-x(1) = 5 / x(2) = 1x(3) = 6 / x(4) = 2x(5) = 3Feas. solutionUpper bound = 6BoundedCurrent span = 6Upper bound = 66 < 6 --> FalseVertex 5Color 9+Vertex 5Color 3-3BoundedCurrent span = 9Upper bound = 69 < 6 --> FalseBoundedCurrent span = 6Upper bound = 66 < 6 --> FalseVertex 3Color 4Pruned x(3) - x(2) = 5 4 - 1 = 53 = 5--> FalseStartVertex 1Color 1Vertex 3Color 1Vertex 5Color 1Vertex 4Color 1-. . .. . .. . .. . .. . .Vertex 2Color -34+-Invalid colorVertex 2Color 5Vertex 3Color 0Invalid colorVertex 3Color 105+-Pruned x(3) - x(1) = 1 10 - 5 = 19 = 1 --> FalseInvalid color Figure 13: Partial enumeration of solutions starting from vertex 2 for the MinEQ-CDGP instance defined by
Figures 1 and 2 using BPB-Prev-FeasCheckFull with color selection based only on the previous vertex and
feasibility checking only when all vertices are colored. The backtracking points are indicated when the solution
is pruned.
Figure 14: Partial enumeration of solutions starting from vertex 2 for the MinEQ-CDGP instance defined by
Figures 1 and 2 using BPB-Select, where a color is determined using a system of absolute value expressions
(equalities or inequalities).
13
Vertex 2Color 1Vertex 1Color 5Vertex 1Color -34+-Vertex 3Color 6Vertex 4Color 101+4+Pruned x(4) - x(1) = 3 10 - 5 = 35 = 3 --> FalseVertex 4Color 2-Vertex 5Color 31+Vertex 5Color 1-x(1) = 5 / x(2) = 1x(3) = 6 / x(4) = 2x(5) = 3Feas. solutionUpper bound = 6BoundedCurrent span = 6Upper bound = 66 < 6 --> FalseVertex 5Color 9+Vertex 5Color 3-3BoundedCurrent span = 9Upper bound = 69 < 6 --> FalseBoundedCurrent span = 6Upper bound = 66 < 6 --> FalseVertex 3Color 4StartVertex 1Color 1Vertex 3Color 1Vertex 5Color 1Vertex 4Color 1-. . .. . .. . .. . .Vertex 5Color 31+Vertex 5Color 1-Pruned x(4) - x(1) = 3 10 - 5 = 35 = 3 --> False. . .Vertex 2Color -34+-Invalid colorVertex 2Color 5Vertex 3Color 0Invalid colorVertex 3Color 105+-Vertex 4Color6Vertex 4Color 14+-4Vertex 5Color 13Vertex 5Color 15+-1Pruned x(3) - x(1) = 1 10 - 5 = 19 = 1 --> FalsePruned x(3) - x(1) = 1 10 - 5 = 19 = 1 --> FalseVertex5Color 5Vertex 5Color 7+-1Pruned x(3) - x(1) = 1 10 - 5 = 19 = 1 --> FalsePruned x(3) - x(1) = 1 10 - 5 = 19 = 1 --> False. . .Vertex 4Color 8Vertex 4Color 0BoundedCurrent span = 8Upper bound = 68 < 6 --> False4-+Invalid color. . .Invalid colorVertex 2Color 1Vertex 1Color 5Vertex 3Color 6Vertex 4Color 2Vertex 5Color 3x(1) = 5 / x(2) = 1x(3) = 6 / x(4) = 2x(5) = 3Feas. solutionUpper bound = 6BoundedCurrent span = 6Upper bound = 66 < 6 --> FalseStartVertex 1Color 1Vertex 3Color 1Vertex 5Color 1Vertex 4Color 1. . .. . .. . .. . .. . .Vertex 2Color 5Vertex 3Color -1PrunedNo valid coloravailableVertex 5Color 3. . . is given in Algorithm 2.
Algorithm 2 Direct Distance Coloring Feasibility (DDCF) check
Require: graph G (with set V of vertices and set E of edges), problem type t (MinGEQ-CDGP or
MinEQ-CDGP), matrix d of distances for each edge, current coloring x and vertex i.
for each neighbor k of i do
if k is colored then
1: function DDCF-Check(G, d, f, x, i)
2:
3:
4:
5:
6:
7:
8:
9:
10:
if t = MinGEQ-CDGP then
if x(k) − x(i) > di,j then
if x(k) − x(i) (cid:54)= di,j then
return false
else
return false
return true
(cid:46) Inequality constraint
(cid:46) Equality constraint
We note that when selecting a color using the first criterion (only taking into account the previously
colored vertex) the feasibility check can be made at each colored vertex or only when all vertices have
been colored (which will require that the function DDF-Check() is called for each vertex). Each
check (for only one vertex) runs in O(V ) time, and if the entire coloring is checked (that is, for all
vertices), it runs in O(V 2) time. We also note that, when using the second criterion (using a system
of absolute value (in)equalities), the feasibility check can be skipped, since the color that it returns is
always feasible.
The combination of these selection criteria and the corresponding feasibility checks result in three
possible BPB algorithms, which are summarized in Table 2.
Table 2: Summary of branch-prune-and-bound methods.
Algorithm
Color
set size
BPB-Prev - Previous
neighbor
BPB-Prev-CheckFull -
Alternate previous
neighbor
BPB-Select - System of
all neighbors
2
2
1
Color selection of a vertex
Feasibility checking
Strategy
x(i) = x(j) + di,j or
x(i) = x(j) + di,j
(if x(i) > x(j) + di,j)
x(i) = x(j) + di,j or
x(i) = x(j) + di,j
(if x(i) > x(j) + di,j)
O(1)
O(1)
Time
complexity
When
Time
complexity
O(V ) for each
vertex
At each colored
vertex
Only when all
vertices are
colored
O(V 2) for entire
coloring
x(i) = min{k ∈ [1, U B] : ∀(i, j) ∈
E x(j) − k = (or ≥) di,j},
O(ub)
Not needed
-
5 Computational experiments
In order to analyze the behavior of the proposed distance geometry coloring problems and the
branch-prune-and-bound algorithms, we made two main sets of experiments: the first one involved
generating many random graphs with different numbers of vertices according to some configurations
and counting how many include even or odd cycles (while the rest are trees), since some of the
properties of distance geometry coloring are related to these types of graphs.
All algorithms used in these experiments were implemented in C language (compiled with gcc 4.8.4
using options -Ofast -march=native -mtune=native) and executed on a computer equipped with
an Intel Core i7-3770 (3,4GHz), 8GB of memory and Linux Mint 17 operating system. We describe
each set of experiments below.
14
Table 3: Number of random graphs with even, odd or no cycles (trees) and bipartite graphs generated for each
number of vertices. For each size, 1,000,000 graphs were generated.
V Average
E
637.56
2523.52
5656.64
10059.56
15675.94
22586.52
30688.21
40120.76
50678.60
62628.32
Average
Density
# Graphs with
# Graphs with
Odd Cycles
Even Cycles
# Trees
0.5205
0.5098
0.5062
0.5055
0.5036
0.5036
0.5025
0.5028
0.5016
0.5020
998309
999553
999832
999910
999926
999958
999975
999975
999971
999988
854
238
74
45
41
18
13
14
15
6
837
209
94
45
33
24
12
11
14
6
# Bipartite
CPU Time
Graphs
1691
447
168
90
74
42
25
25
29
12
(sec)
257.07
753.25
1808.25
3403.02
4553.07
6764.28
10042.43
11886.32
14415.33
23332.64
50
100
150
200
250
300
350
400
450
500
5.1 Counting members of graph classes in random instances
(cid:104)V − 1,
(cid:105)
Using Theorems 2 and 3, we have information about some types of graphs which always have
feasible embeddings for EQ-CDGP and EQ-CDGP-Unif. Based on this, we generated a large amount
of random graphs with different number of vertices and counted how many were cyclic (and based on
that, how many there were for each possibility of having even or odd cycles) and how many were trees.
Each random graph always starts as a random spanning tree, that is, a connected undirected
graph G = (V, E), where E = V − 1. To generate this initial tree, we used a random walk algorithm
proposed independently by Broder (1989) and Aldous (1990). The procedure works by using a set V ∗
of the vertices outside the tree and a set W of edges of the spanning tree. Then, whenever the random
walk reaches a vertex j outside the tree, the edge (i, j) is added to E and j is removed from V ∗. This
continues until V ∗ = ∅. We note that this amounts to making a random walk in a complete graph
of V vertices and it generates trees in a uniform manner, that is, for all possible spanning trees of a
given complete graph, each one has the same probability of being generated by the algorithm.
V (V −1)
After the initial tree is generated, we add random new edges to it until the graph has the desired
number of edges. This parameter is also randomly set, sampled from interval
.
This interval ensures that the generated graph is always connected and is, at least, a tree and, at
most, a complete graph.
In Table 3, we outline statistics obtained from using the described procedure to generate 1,000,000
(one million) random graphs for each V ∈ {50, 100, 150, 200, 250, 300, 350, 400, 450, 500}. As we can
observe, most of the graphs (more than 99%) generated have odd cycles, which translates into a
very small set of possible EQ-CDGP-Unif instances with feasible embeddings for this configuration of
random graphs. By increasing the number of vertices, more possibilities for generating edges appear,
but the number of possible connections which will lead to trees or graphs with even cycles is very small.
In fact, we can deduce that this configuration generated very few bipartite graphs. For EQ-CDGP
(with arbitrary distances), the space of instances with guaranteed feasible embeddings is even smaller,
since only trees are certain to have them. However, as shown in Section 3, odd and even cycles can
have embeddings depending on how the edges are weighted.
2
In Figure 15, we can observe the growth of the average number of edges between all generated
graphs for each number of vertices. Since the number of edges in a graph is proportional to the square
∈ O(V 2), the curve follows a similar pattern, being a half
of the number of edges (since
parabola.
V (V −1)
2
5.2 Results for branch-prune-and-bound algorithms
In order to use some of the random graphs in experiments with the BPB algorithms, we selected
four graphs of each type (with even cycle, with odd cycle and trees) for each number of vertices and
15
Figure 15: Growth of the average number of edges generated when the number of vertices increases.
(a) V ∈ [50, 500]
(b) V ∈ [200, 500] (for scale)
Figure 16: Total number of bipartite graphs generated from 1,000,000 random graphs of each number of vertices.
16
0 10000 20000 30000 40000 50000 60000 70000 50 100 150 200 250 300 350 400 450 500Average number of edgesNumber of vertices 200 400 600 800 1000 1200 1400 1600 1800 100 200 300 400 500Number of bipartitesNumber of vertices 10 20 30 40 50 60 70 80 90 200 300 400 500Number of bipartitesNumber of vertices Table 4: Results for BP algorithm (decision/search) applied to EQ-CDGP instances - 4, 5, 6, 7 and 8 vertices.
V
Type
Inst
Span # Prunes # Nodes CPU Time (s)
Span # Prunes # Nodes CPU Time (s)
Span # Prunes # Nodes CPU Time (s)
BPB-Prev
BPB-Prev-FeasCheckFull
BPB-Select
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.001
0.001
0.000
0.001
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.003
0.007
0.006
0.006
0.001
0.000
0.001
0.001
0.000
0.000
0.000
0.000
0.041
0.002
0.004
0.018
0.000
0.007
0.005
0.000
0.000
0.000
0.000
0.000
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
24
19
14
14
20
21
38
Infeasible
20
Infeasible
32
24
Infeasible
Infeasible
17
19
20
20
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
21
Infeasible
Infeasible
13
19
27
39
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
23
Infeasible
Infeasible
19
14
27
35
Infeasible
Infeasible
Infeasible
Infeasible
28
Infeasible
Infeasible
18
34
29
22
20
16
12
12
12
8
8
0
0
0
0
0
0
1
20
0
30
1
0
20
20
0
0
0
1
45
76
48
43
30
0
60
30
0
0
5
0
66
148
138
144
91
26
70
100
0
0
2
1
220
145
99
311
1
230
352
4
0
0
0
1
34
30
30
30
28
28
4
4
4
4
4
4
7
54
5
62
7
5
60
60
5
5
5
7
105
128
105
95
97
6
160
100
6
6
17
6
139
229
219
227
246
105
178
259
7
7
11
10
336
438
251
563
10
511
883
19
8
8
8
10
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.001
0.000
0.000
0.001
0.000
0.000
0.000
0.000
0.000
4
5
6
7
8
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
24
19
21
34
29
41
38
Infeasible
20
Infeasible
44
36
Infeasible
Infeasible
20
19
23
39
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
46
Infeasible
Infeasible
16
54
34
39
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
36
Infeasible
Infeasible
34
33
27
35
Infeasible
Infeasible
Infeasible
28
21
21
19
20
20
0
0
0
0
0
0
0
38
1
61
0
0
54
51
0
0
0
0
149
176
167
126
100
0
155
84
0
0
0
0
191
360
383
378
263
151
250
293
0
0
0
0
650
896
696
Infeasible
1092
51
Infeasible
Infeasible
0
1005
1474
40
45
47
43
71
0
0
0
0
0
36
32
32
32
32
32
4
4
4
4
4
4
5
58
5
69
5
5
77
76
5
5
5
5
161
168
161
135
131
6
223
137
6
6
6
6
196
307
323
310
361
193
283
382
7
7
7
7
505
1202
652
1075
8
1072
1934
8
8
8
8
8
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.001
0.001
0.001
0.001
0.000
0.001
0.002
0.000
0.000
0.000
0.000
0.000
54
42
41
41
32
32
4
4
4
4
4
4
5
88
6
175
5
5
97
96
5
5
5
5
533
1114
465
624
187
6
381
203
6
6
6
6
2880
6875
6506
6470
686
243
1422
745
7
7
7
7
41815
2191
4052
18078
8
6544
4369
8
8
8
8
8
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
24
19
21
34
29
41
38
Infeasible
20
Infeasible
44
36
Infeasible
Infeasible
20
19
23
39
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
46
Infeasible
Infeasible
16
54
34
39
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
36
Infeasible
Infeasible
34
33
27
35
36
25
22
21
20
20
0
0
0
0
0
0
0
56
1
106
0
0
66
62
0
0
0
0
378
887
374
504
122
0
242
120
0
0
0
0
2234
5398
5360
5391
429
192
1093
511
0
0
0
0
Infeasible
34796
Infeasible
Infeasible
1750
3192
Infeasible
13940
51
Infeasible
Infeasible
0
5410
3346
40
45
47
43
71
0
0
0
0
0
17
Table 5: Results for BP algorithm (decision/search) applied to EQ-CDGP instances - 9, 10, 12, 14 and 16
vertices.
Type
Inst
Span # Prunes # Nodes CPU Time (s)
Span
# Prunes
# Nodes CPU Time (s)
Span # Prunes # Nodes CPU Time (s)
BPB-Prev
BPB-Prev-FeasCheckFull
BPB-Select
V
9
10
12
14
16
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
Infeasible
Infeasible
Infeasible
Infeasible
61
26
Infeasible
Infeasible
55
68
88
45
1638
1633
1165
1417
0
0
2011
946
0
0
0
0
1537
1219
923
1801
9
9
2542
1142
9
9
9
9
Infeasible
33183
42018
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
72
59
40
49
46
2619
3529
4379
3500
46618
70914
1
0
0
0
0
Infeasible
108359
Infeasible
24433
Infeasible
Infeasible
49
7440
8467
0
Infeasible
13801
41
0
Infeasible
18820
38
50
48
52
0
0
0
0
1974
2963
3333
3114
48740
86532
10
10
10
10
10
103428
23227
5294
5648
12
16577
12
19828
12
12
12
12
Infeasible
2725363
2744206
Infeasible
Infeasible
Infeasible
Infeasible
34217
25749
17520
22438
Infeasible
8815580
Infeasible
8022290
Infeasible
4979521
55
63
67
59
0
0
0
0
23821
16751
11848
20427
9620137
8013781
5359288
14
14
14
14
Infeasible
9674081
8970218
Infeasible
Infeasible
Infeasible
79
40827
64340
79312
0
26130
40739
71850
16
Infeasible
41169779
41680959
72
0
16
Infeasible
61277800
71043803
65
86
74
79
0
0
0
0
16
16
16
16
0.002
0.001
0.001
0.002
0.000
0.000
0.003
0.001
0.000
0.000
0.000
0.000
0.042
0.002
0.003
0.004
0.003
0.045
0.086
0.000
0.000
0.000
0.000
0.000
0.115
0.026
0.006
0.006
0.000
0.017
0.000
0.021
0.000
0.000
0.000
0.000
2.820
0.019
0.018
0.013
0.022
9.640
7.269
4.921
0.000
0.000
0.000
0.000
8.843
0.019
0.031
0.056
0.000
35.448
0.000
55.371
0.000
0.000
0.000
0.000
Infeasible
Infeasible
Infeasible
Infeasible
61
26
Infeasible
Infeasible
55
68
88
45
46903
328822
385150
2479
0
0
3414
1347
0
0
0
0
Infeasible
155000
Infeasible
2108251
Infeasible
316389
Infeasible
1481970
Infeasible
Infeasible
Infeasible
157665
160032
254208
72
59
40
49
46
1
0
0
0
0
60000
400892
467221
3574
9
9
4392
1785
9
9
9
9
173304
2797473
400670
1818952
199713
193409
304707
10
10
10
10
10
Infeasible
965368
Infeasible
1559180
1140595
1807872
0.065
0.397
0.456
0.003
0.000
0.000
0.005
0.002
0.000
0.000
0.000
0.000
0.168
2.814
0.367
1.868
0.174
0.224
0.249
0.000
0.000
0.000
0.000
0.000
1.284
1.927
Infeasible
36724085
44121832
Infeasible
544947529
645413373
41.883
390.399
49
0
Infeasible
53860
41
0
12
62497
12
Infeasible
381806
495804
38
50
48
52
0
0
0
0
12
12
12
12
0.000
0.073
0.000
0.393
0.000
0.000
0.000
0.000
Infeasible
104517592
118562122
Infeasible
3407171273
4097250986
101.605
3314.219
Infeasible
11200102605
14120166774
10800.000
Infeasible
4872771100
5921495589
5148.447
Infeasible
1726508
Infeasible
38618944
Infeasible
32873088
Infeasible
16804920
1934194
42925314
37749326
21169856
55
63
67
59
0
0
0
0
14
14
14
14
1.889
27.755
29.182
17.424
0.000
0.000
0.000
0.000
Infeasible
Infeasible
Infeasible
Infeasible
34
20
Infeasible
Infeasible
38
27
29
21
434
456
452
297
0
7
226
88
10
0
0
0
777
650
620
801
9
24
788
357
30
9
9
9
Infeasible
7939
15821
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
38
27
31
33
29
Infeasible
Infeasible
Infeasible
Infeasible
22
Infeasible
35
792
710
1030
520
4718
4662
0
4
1
4
3
9947
1578
1426
2293
42
954
16
1127
1184
1441
1036
9984
10887
10
18
12
18
16
18283
3718
2058
2876
146
2799
51
Infeasible
6572
13539
27
22
34
26
58
0
440
2
139
12
1016
16
Infeasible
119828
218219
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
29
37
31
39
5217
4890
2844
286
240185
146774
160477
54
164
60
647
6949
6169
3883
955
480564
329350
354148
138
385
166
1432
Infeasible
270950088
312103257
218.788
Infeasible
229364
464467
0.001
0.001
0.000
0.001
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.011
0.001
0.001
0.001
0.001
0.006
0.007
0.000
0.000
0.000
0.000
0.000
0.012
0.003
0.002
0.002
0.000
0.002
0.000
0.009
0.000
0.000
0.001
0.000
0.146
0.005
0.005
0.003
0,001
0.306
0.205
0.225
0.000
0.000
0.000
0.001
0.295
0.010
0.012
0.010
0.001
Infeasible
Infeasible
Infeasible
35
10334
13417
7508
356
12037
15677
14908
799
Infeasible
10410162
19341115
12.607
29
1390
3171
Infeasible
787682
1851608
30
30
36
23
222
15
329
9
502
50
756
42
0.002
1.155
0.000
0.000
0.001
0.000
Infeasible
9212314467
13585438967
10800.000
Infeasible
12857613205
16238297924
10800.000
Infeasible
135939959
153954534
79
0
16
Infeasible
881601120
1035374397
72
0
16
Infeasible
250169040
273392839
65
86
74
79
0
0
0
0
16
16
16
16
104.928
0.000
712.612
0.000
196.395
0.000
0.000
0.000
0.000
18
V
18
20
Table 6: Results for BP algorithm (decision/search) applied to EQ-CDGP instances - 18 and 20 vertices.
Type
Inst
Span
# Prunes
# Nodes CPU Time (s)
Span
# Prunes
# Nodes CPU Time (s)
Span # Prunes # Nodes CPU Time (s)
BPB-Prev
BPB-Prev-FeasCheckFull
BPB-Select
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
OddCycle
OddCycle
OddCycle
OddCycle
EvenCycle
EvenCycle
EvenCycle
EvenCycle
Tree
Tree
Tree
Tree
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4
Infeasible
508400
Infeasible
1151865
Infeasible
Infeasible
164231
117988
411189
809413
137272
72425
Infeasible
11859360
10893632
Infeasible
152740084
167194942
Infeasible
466856235
439993043
Infeasible
54784145
53791828
88
61
91
71
0
0
0
0
18
18
18
18
0.324
0.630
0.093
0.050
8.610
127.611
298.974
36.624
0.000
0.000
0.000
0.000
Infeasible
14249119873
14806963853
10800.000
Infeasible
6511737586
7859547065
10800.000
Infeasible
2830789695
3168256401
3927.393
Infeasible
8764779022
9900318687
10800.000
Infeasible
1336188416
1468424277
Infeasible
611753448
711245102
979.763
598.378
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
18527
29654
1411
25336
92766
Infeasible
996127
Infeasible
6868230379
7523523165
10800.000
Infeasible
2312073
Infeasible
285915264
360281393
505.807
Infeasible
2527172
88
61
91
71
0
0
0
0
18
18
18
18
0.000
0.000
0.000
0.000
31
25
28
30
6170
1171
57
1634
33062
46174
3456
28579
208614
2340311
4373462
5361369
12737
2731
157
3433
Infeasible
273165199
217944054
169.144
Infeasible
6147895146
9044193773
10800.000
Infeasible
1786324
3216901
Infeasible
2231047
1503222
Infeasible
1049440
Infeasible
414762
744982
249339
1.467
0.542
0.170
Infeasible
6371122973
8069928711
Infeasible
7751043598
8898429750
10800.000
10800.000
Infeasible
13950045871
14504063426
10800.000
Infeasible
Infeasible
Infeasible
63034
31642
56488
89385
48916
64593
Infeasible
7618735112
8591239945
7112.081
141
0
20
Infeasible
20754606
17374721
0.000
14.777
$ -
141
-
0
-
20
Infeasible
7684808926
7634511236
10800.000
Infeasible
167148
Running
Infeasible
20719998
47209517
0.000
36
13158
25609
351029
0.023
0.032
0.002
0.022
0.135
1.429
2.787
3.354
0.008
0.002
0.000
0.002
2.101
0.066
0.035
0.051
29.050
0.017
0.225
Infeasible
15355600960
14057177765
10800.000
Infeasible
5426010704
7820985965
10800.000
Infeasible
335043320
686615646
428.222
64
103
96
115
0
0
0
0
20
20
20
20
0.000
0.000
0.000
0.000
64
103
96
115
0
0
0
0
20
20
20
20
0.000
0.000
0.000
0.000
24
29
36
27
107586
201000
560
2908
57601
1443
6536
151814
0.132
0.001
0.004
0.095
randomly weighted the edges with a uniform distribution in the interval [1, 30]. We made two subsets
of experiments: the first one involved only the EQ-CDGP and EQ-CDGP-Const models, in order to
find a feasible solution for each of its instances that were generated (that is, the algorithms use the
pruning procedure, but not bounding - equivalently, stopping the search as soon as the first feasible
solution is found), and the second one involved the optimization models for each discussed model
(MinEQ-CDGP, MinEQ-CDGP-Const, MinGEQ-CDGP and MinGEQ-CDGP-Const).
Tables 4, 5 and 6 give detailed results with 4 to 8, 9 to 16 and 18 to 20 vertices, respectively,
considering each BPB algorithm applied to the decision versions. We can see that BPB-Prev reaches
a feasible solution faster than the other methods (that is, it solves the decision problem in less time),
but it also returns the first feasible solution with a worst span than BPB-Select (noting that BPB-
Prev-FeasCheckFull always returns the same span from BPB-Prev because only the pruning point is
changed). However, it is much slower to prove that an instance is infeasible (that is, the answer to
the decision problem is NO). This is explained by the fact that the enumeration tree is smaller in
BPB-Select, since instead of two color possibilities for each vertex, there is only one. Although the
time complexity of determining the color for a vertex in BPB-Select is higher (as shown in Table 2),
this is compensated by generating a smaller tree and that the feasibility check is not explicitly needed,
since it is guaranteed by the color selection algorithm. We also note that BPB-Prev-FeasCheckFull has
the worst CPU times for infeasible instances, because the method will keep branching the enumeration
tree to find a feasible solution, but since feasibility checking is only done at the leaf nodes, the tree
will tend to become the full enumeration tree.
In the same manner, Table 7 shows the results for the BPB algorithms considering the optimization
versions and applied only to feasible MinEQ-CDGP instances. For almost all of these instances, BPB-
Prev is the best method, BPB-Prev-FeasCheckFull shows similar performance and BPB-Select has
worse CPU times. The ties between BPB-Prev and BPB-Prev-FeasCheckFull are explained by the
fact that although, in the latter method, the feasibility checking at the leaf nodes increases the time
required to find a feasible solution, we keep using the bounding procedures at each node, which reduces
the amount of work needed to find the optimal solution. We also note that, for the 4th Tree instance
with 20 vertices, BPB-Select does not find the true optimum for the problem. This happens because
the method is recursively applied only to vertices which have recorded neighbors, in the same manner
as the other two BPBs, but the system of absolute value expressions returns only one color, which
may not be the one for the optimal solution when applying the recursion only on some vertices. On
some experiments, we detected that, if we generate all vertice orders and apply the color selection of
BPB-Select, on them, the optimal solution is found, but the CPU times become very high, since this
procedure does not take advantage of the 0-sphere intersection characteristic.
19
The last experiments were made by applying the BPB algorithms on all instances considered for
the algorithms, but by transforming them to MinGEQ-CDGP (changing the = in constraints to ≥).
Since they are always feasible for MinGEQ-CDGP because of its equivalence to BCP, we only have
to consider optimization problems, as was already done in Section 2. The same pattern of previous
experiments occur here, with BPB-Prev being the best method of the three, however, the CPU time
difference between it and BPB-Prev-FeasCheckFull becomes much more apparent here, since there are
many more feasible solutions for MinGEQ-CDGP than MinEQ-CDGP.
6 Concluding remarks
In this paper, we proposed some distance geometry models for graph coloring problems with
distance constraints that can be applied to important, modern real world applications, such as in
telecommunications for channel assignment in wireless networks. In these proposed coloring distance
geometry problems (CDGPs), the vertices of the graph are embedded on the real line and the coloring
is treated as an assignment of natural numbers to the vertices, while the distances correspond to line
segments, whose objective is to determine a feasible intersection of them.
We tackle such problems under the graph theory approach, to establish conditions of feasibility
through behavior analysis of the problems in certain classes of graphs, identifying prohibited structures
for which the occurrence indicates that it can not admit a valid solution, as well as identifying classes
graphs that always admit valid solution.
We also developed exact and approximate enumeration algorithms, based on the Branch-and-Prune
(BP) algorithm proposed for the Discretizable Molecular Distance Geometry Problem (DMDGP),
combining concepts from constraint propagation and optimization techniques, resulting in Branch-
Prune-and-Bound algorithms (BPB), that handle the set of distances in different ways in order to
get feasible and optimal solutions. The computational experiments involved equality and inequality
constraints and both uniform and arbitrary sets of distances, where we measure the number of prunes
and bounds and the CPU time needed to reach the best solution.
The main contribution of this paper consists of a distance geometry approach for special cases of
T-coloring problems with distance constraints, involving a study of graph classes for which some of
these distance |
1301.0995 | 1 | 1301 | 2013-01-06T11:44:35 | On the Complexity of Connectivity in Cognitive Radio Networks Through Spectrum Assignment | [
"cs.DS"
] | Cognitive Radio Networks (CRNs) are considered as a promising solution to the spectrum shortage problem in wireless communication. In this paper, we initiate the first systematic study on the algorithmic complexity of the connectivity problem in CRNs through spectrum assignments. We model the network of secondary users (SUs) as a potential graph, where two nodes having an edge between them are connected as long as they choose a common available channel. In the general case, where the potential graph is arbitrary and the SUs may have different number of antennae, we prove that it is NP-complete to determine whether the network is connectable even if there are only two channels. For the special case where the number of channels is constant and all the SUs have the same number of antennae, which is more than one but less than the number of channels, the problem is also NP-complete. For the special cases in which the potential graph is complete, a tree, or a graph with bounded treewidth, we prove the problem is NP-complete and fixed-parameter tractable (FPT) when parameterized by the number of channels. Exact algorithms are also derived to determine the connectability of a given cognitive radio network. | cs.DS | cs | Noname manuscript No.
(will be inserted by the editor)
On the Complexity of Connectivity in Cognitive Radio
Networks Through Spectrum Assignment
Hongyu Liang · Tiancheng Lou ·
Haisheng Tan · Yuexuan Wang ·
Dongxiao Yu
3
1
0
2
n
a
J
6
]
S
D
.
s
c
[
1
v
5
9
9
0
.
1
0
3
1
:
v
i
X
r
a
the date of receipt and acceptance should be inserted later
Abstract Cognitive Radio Networks (CRNs) are considered as a promising
solution to the spectrum shortage problem in wireless communication. In this
paper, we initiate the first systematic study on the algorithmic complexity of
the connectivity problem in CRNs through spectrum assignments. We model
the network of secondary users (SUs) as a potential graph, where two nodes
having an edge between them are connected as long as they choose a common
available channel. In the general case, where the potential graph is arbitrary
and the SUs may have different number of antennae, we prove that it is NP-
complete to determine whether the network is connectable even if there are
only two channels. For the special case where the number of channels is con-
stant and all the SUs have the same number of antennae, which is more than
one but less than the number of channels, the problem is also NP-complete.
For the special cases in which the potential graph is complete, a tree, or a
graph with bounded treewidth, we prove the problem is NP-complete and
fixed-parameter tractable (FPT) when parameterized by the number of chan-
nels. Exact algorithms are also derived to determine the connectability of a
given cognitive radio network.
A preliminary version of this paper appears in ALGOSENSORS 2012 [13]. The main exten-
sions are that we give more details of our results and investigate the special case where the
potential graphs are of bounded treewidth.
Hongyu Liang, Haisheng Tan(Corresponding Author), Yuexuan Wang
Institute for Theoretical Computer Science, IIIS, Tsinghua University, Beijing, China
E-mail: [email protected], {tan,wangyuexuan}@mail.tsinghua.edu.cn
Tiancheng Lou
Google Inc., California, USA
E-mail: [email protected]
Dongxiao Yu
Department of Computer Science, the University of Hong Kong, Hong Kong, China
2
1 Introduction
Hongyu Liang et al.
Cognitive Radio is a promising technology to alleviate the spectrum shortage
in wireless communication. It allows the unlicensed secondary users to utilize
the temporarily unused licensed spectrums, referred to as white spaces, without
interfering with the licensed primary users. Cognitive Radio Networks (CRNs)
is considered as the next generation of communication networks and attracts
numerous research from both academia and industry recently.
In CRNs, each secondary user (SU) can be equipped with one or multiple
antennae for communication. With multiple antennae, a SU can communicate
on multiple channels simultaneously (in this paper, channel and spectrum are
used interchangeably.). Through spectrum sensing, each SU has the capacity
to measure current available channels at its site, i.e. the channels are not used
by the primary users (PUs). Due to the appearance of PUs, the available
channels of SUs have the following characteristics [1]:
-- Spatial Variation: SUs at different positions may have different available
channels;
-- Spectrum Fragmentation: the available channels of a SU may not be con-
tinuous; and
-- Temporal Variation: the available channels of a SU may change over time.
Spectrum assignment is to allocate available channels to SUs to improve
system performance such as spectrum utilization, network throughput and
fairness. Spectrum assignment is one of the most challenging problems in CRNs
and has been extensively studied such as in [12,19,20,22,23].
Connectivity is a fundamental problem in wireless communication. Con-
nection between two nodes in CRNs is not only determined by their distance
and their transmission powers, but also related to whether the two nodes has
chosen a common channel. Due to the spectrum dynamics, communication in
CRNs is more difficult than in the traditional multi-channel radio networks
studied in [4]. Authors in [14,15,16] investigated the impact of different pa-
rameters on connectivity in large-scale CRNs, such as the number of channels,
the activity of PUs, the number of neighbors of SUs and the transmission
power.
In this paper, we initiate the first systematic study on the complexity of
connectivity in CRNs through spectrum assignment. We model the network as
a potential graph and a realized graph before and after spectrum assignment
respectively (refer to Section 2). We start from the most general case, where
the network is composed of heterogenous SUs1, SUs may be equipped with
different number of antennae and the potential graph can be arbitrary (Fig-
ure 1). Then, we proceed to study the special case when all the SUs have the
same number of antennae. If all the SUs are homogenous with transmission
ranges large enough, the potential graph will be a complete graph. For some
hierarchically organized networks, e.g. a set of SUs are connected to an access
1 We assume two heterogenous SUs cannot communicate even when they work on a com-
mon channel and their distance is within their transmission ranges.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
3
Fig. 1 the general case. a) the potential graph: the set besides each SU is its available
channels, and β is its number of antennae. u2 and u4 are not connected directly because they
are a pair of heterogenous nodes or their distance exceeds at least one of their transmission
ranges. b) the realization graph which is connected: the set beside each SU is the channels
assigned to it.
point, the potential graph can be a tree. Therefore, we also study these special
cases. Exact algorithms are also derived to determine connectivity for different
cases. Our results are listed below. To the best of knowledge, this is the first
work that systematically studies the algorithmic complexity of connectivity in
CRNs with multiple antennae.
Our Contributions: In this paper we study the algorithmic complexity of the
connectivity problem through spectrum assignment under different models.
Our main results are as follows.
-- When the potential graph is a general graph, we prove that the problem is
NP-complete even if there are only two channels. This result is sharp as the
problem is polynomial-time solvable when there is only one channel. We
also design exact algorithms for the problem. For the special case when all
SUs have the same number of antennae, we prove that the problem is NP-
complete when k > β ≥ 2, where k and β are the total amount of channels
in the white spaces and the number of antennae on an SU respectively.
-- When the potential graph is complete,2 the problem is shown to be NP-
complete even if each node can open at most two channels. However, in
contrast to the general case, the problem is shown to be polynomial-time
solvable if the number of channels is fixed. In fact, we prove a stronger
result saying that the problem is fixed parameter tractable when parame-
terized by the number of channels. (See [5] for notations in parameterized
complexity.)
-- When the potential graph is a tree, we prove that the problem is NP-
complete even if the tree has depth one. Similar to the complete graph
case, we show that the problem is fixed parameter tractable when parame-
terized by the number of channels. We then generalize this result, showing
that the problem remains fixed parameter tractable when parameterized
2 The complete graph is a special case of disk graphs, which are commonly used to model
wireless networks such as in [11, 21].
4
Hongyu Liang et al.
by the number of channels if the underlying potential graph has bounded
treewidth.
Paper Organization: In Section 2 we formally define our model and problems
studied in this paper. We study the problem with arbitrary potential graphs
in Section 3. The special cases where the potential graph is complete or a tree
are investigated in Sections 4 and 5. The paper is concluded in Section 6 with
possible future works.
2 Preliminaries
2.1 System Model and Problem Definition
We first describe the model used throughout this paper. A cognitive radio
network is comprised of the following ingredients:
-- U is a collection of secondary users (SUs) and C is the set of channels in
the white spaces.
-- Each SU u ∈ U has a spectrum map, denoted by SpecMap(u), which is a
subset of C representing the available channels that u can open.
-- The potential graph PG = (U, E), where each edge of E is also called a
potential edge. If two nodes are connected by a potential edge, they can
communicate as long as they choose a common available channel.
-- Each SU u ∈ U is equipped with a number of antennas, denoted as antenna
budget β(u), which is the maximum number of channels that u can open
simultaneously.
For a set S, let 2S denote the power set of S, i.e., the collection of all
subsets of S. A spectrum assignment is a function SA : U → 2C satisfying
that
SA(u) ⊆ SpecMap(u) and SA(u) ≤ β(u) for all u ∈ U.
Equivalently, a spectrum assignment is a way of SUs opening channels such
that each SU opens at most β channels and can only open those in its spectrum
map.
Given a spectrum assignment SA, a potential edge {u, v} ∈ E is called re-
alized if SA(u) ∩ SA(v) 6= ∅, i.e., there exists a channel opened by both u and
v. The realization graph under a spectrum assignment is a graph RG = (U, E′),
where E′ is the set of realized edges in E. Note that RG is a spanning subgraph
of the potential graph PG. A cognitive radio network is called connectable if
there exists a spectrum assignment under which the realization graph is con-
nected, in which case we also say that the cognitive radio network is connected
under this spectrum assignment. Now we can formalize the problems studied
in this paper.
The Spectrum Connectivity Problem. The Spectrum Connectivity
problem is to decide whether a given cognitive radio network is connectable.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
5
We are also interested in the special case where the number of possible
channels is small3 and SUs have the same antenna budget. Therefore, we
define the following subproblem of the Spectrum Connectivity problem:
The Spectrum (k, β)-Connectivity Problem. For two constants k, β ≥ 1,
the Spectrum (k, β)-Connectivity problem is to decide whether a given
cognitive radio network with k channels in which all SUs have the same budget
β is connectable. For convenience we write SpecCon(k, β) to represent this
problem.
Finally, we also consider the problem with special kinds of potential graphs,
i.e. the potential graph is complete or a tree.
In the sequel, unless otherwise stated, we always use n := U and k := C
to denote respectively the number of secondary users and channels.
2.2 Tree Decomposition
In this subsection we give some basic notions regarding the tree decomposition
of a graph, which will be used later. The concept of treewidth was introduced
by Robertson and Seymour in their seminal work on graph minors [17]. A tree
decomposition of a graph G = (V, E) is given by a tuple (T = (I, F ), {Xi i ∈
I}), where T is a tree and each Xi is a subset of V called a bag satisfying that
-- Si∈I Xi = V ;
-- For each edge {u, v} ∈ E, there exists a tree node i with {u, v} ⊆ Xi;
-- For each vertex u ∈ V , the set of tree nodes {i ∈ I u ∈ Xi} forms a
connected subtree of T . Equivalently, for any three vertices t1, t2, t3 ∈ I
such that t2 lies in the path from t1 to t3, it holds that Xt1 ∩ Xt3 ⊆ Xt2.
The width of the tree decomposition is maxi∈I {Xi − 1}, and the treewidth
of a graph G is the minimum width of a tree decomposition of G. For each
fixed integer d, there is a polynomial time algorithm that decides whether a
given graph has treewidth at most d, and if so, constructs a tree decomposition
of width d [2]. Such a decomposition can easily be transformed to a nice tree
decomposition (T, {Xi}) of G with the same width, in which T is a rooted
binary tree with at most O(V ) nodes (see e.g. [10]).
3 The Spectrum Connectivity Problem
In this section, we study the the Spectrum Connectivity problem from
both complexity and algorithmic points of view.
3 Commonly, the white spaces include spectrums from channel 21 (512Mhz) to 51
(698Mhz) excluding channel 37, which is totally 29 channels [1].
6
Hongyu Liang et al.
3.1 NP-completeness Results
We show that the Spectrum Connectivity problem is NP-complete even if
the number of channels is fixed. In fact we give a complete characterization of
the complexity of SpecCon(k, β) by proving the following dichotomy result:
Theorem 1 SpecCon(k, β) is NP-complete for any integers k > β ≥ 2, and
is in P if β = 1 or k ≤ β.
The second part of the statement is easy: When β = 1, each SU can
only open one channel, and thus all SUs should be connected through the
same channel. Therefore, the network is connectable if and only if there there
exists a channel that belongs to every SU's spectrum map (and of course the
potential graph must be connected), which is easy to check. When k ≤ β, each
SU can open all channels in its spectrum map, and the problem degenerates
to checking the connectivity of the potential graph.
In the sequel we prove the NP-completeness of SpecCon(k, β) when k >
β ≥ 2. First consider the case k = β + 1. We will reduce a special case of the
Boolean Satisfiability (SAT) problem, which will be shown to be NP-complete,
to SpecCon(β + 1, β), thus showing the NP-completeness of the latter.
A clause is called positive if it only contains positive literals, and is called
negative if it only contains negative literals. For example, x1 ∨ x3 ∨ x5 is
positive and x2 ∨ x4 is negative. A clause is called uniform if it is positive
or negative. A uniform CNF formula is the conjunction of uniform clauses.
Define Uniform-SAT as the problem of deciding whether a given uniform
CNF formula is satisfiable.
Lemma 1 Uniform-SAT is NP-complete.
Proof Let F be a CNF formula with variable set {x1, x2, . . . , xn}. For each i
such that xi appears in F , we create a new variable yi, and do the following:
-- substitute yi for all occurrences of xi;
-- add two clauses xi ∨ yi and xi ∨ yi to F . More formally, let F ← F ∧ (xi ∨
yi) ∧ (xi ∨ yi). This ensures yi = xi in any satisfying assignment of F .
Call the new formula F ′. For example, if F = (x1 ∨ x2) ∧ (x1 ∨ x3), then
F ′ = (x1 ∨ y2) ∧ (y1 ∨ x3) ∧ (x1 ∨ y1) ∧ (x1 ∨ y1) ∧ (x2 ∨ y2) ∧ (x2 ∨ y2).
It is easy to see that F ′ is a uniform CNF formula, and that F is satisfi-
able if and only if F ′ is satisfiable. This constitutes a reduction from SAT to
Uniform-SAT, which concludes the proof. ⊓⊔
Theorem 2 SpecCon(β + 1, β) is NP-complete for any integer β ≥ 2.
Proof The membership of SpecCon(β + 1, β) in NP is clear. In what follows
we reduce Uniform-SAT to SpecCon(β +1, β), which by Lemma 1 will prove
the NP-completeness of the latter.
Let c1 ∧ c2 ∧ . . . ∧ cm be an input to Uniform-SAT where cj, 1 ≤ j ≤ m,
is a uniform clause. Assume the variable set is {x1, x2, . . . , xn}. We construct
an instance of SpecCon(β + 1, β) as follows.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
7
-- Channels: There are β + 1 channels {0, 1, 2, . . . , β}.
-- SUs:
-- For each variable xi, there is a corresponding SU Xi with spectrum map
SpecMap(Xi) = {0, 1, 2, . . . , β} (which contains all possible channels);
-- for each clause cj, 1 ≤ j ≤ m, there is a corresponding SU Cj with
SpecMap(Cj) = {pj}, where pj = 1 if cj is positive and pj = 0 if cj is
negative;
-- there is an SU Y2 with SpecMap(Y2) = {2}. For every 1 ≤ i ≤ n and
2 ≤ k ≤ β, there is an SU Yi,k with SpecMap(Yi,k) = {k}; and
-- all SUs have the same antenna budget β.
-- Potential Graph: For each clause cj and each variable xi that appears
in cj (either as xi or xi), there is a potential edge between Xi and Cj . For
each 1 ≤ i ≤ n and 3 ≤ k ≤ β, there is a potential edge between Xi and
Yi,k. Finally, there is a potential edge between Y2 and every Xi, 1 ≤ i ≤ n.
Denote the above cognitive radio network by I, which is also an instance
of SpecCon(β + 1, β). We now prove that c1 ∧ c2 ∧ . . . ∧ cm is satisfiable if
and only if I is connectable.
First consider the "only if" direction. Let A : {x1, . . . , xn} → {0, 1} be a
satisfying assignment of c1 ∧ c2 ∧ . . . ∧ cm, where 0 stands for FALSE and 1 for
TRUE. Define a spectrum assignment as follows. For each 1 ≤ i ≤ n, let user
Xi open the channels {2, 3, . . . , β} ∪ {A(i)}. Every other SU opens the only
channel in its spectrum map.
We verify that I is connected under the above spectrum assignment. For
each 1 ≤ i ≤ n, Xi is connected to Y2 through channel 2. Then, for every
2 ≤ l ≤ β, Yi,l is connected to Xi through channel l. Now consider SU Cj
where 1 ≤ j ≤ m. Since A satisfies the clause cj, there exists 1 ≤ i ≤ n
such that: 1) xi or xi occurs in cj; and 2) A(xi) = 1 if cj is positive, and
A(xi) = 0 if cj is negative. Thus Xi and Cj are connected through channel
A(xi). Therefore the realization graph is connected, completing the proof of
the "only if" direction.
We next consider the "if" direction. Suppose there is a spectrum assign-
ment that makes I connected. For every 1 ≤ i ≤ n and 2 ≤ l ≤ β, Xi must
open channel l, otherwise Yi,l will become an isolated vertex in the realization
graph. Since Xi can open at most β channels in total, it can open at most
one of the two remaining channels {0, 1}. We assume w.l.o.g. that Xi opens
exactly one of them, which we denote by ai.
Now, for the formula c1 ∧ c2 ∧ . . . ∧ cm, we define a truth assignment
A : {x1, . . . , xn} → {0, 1} as A(xi) = ai for all 1 ≤ i ≤ n. We show that
A satisfies the formula. Fix 1 ≤ j ≤ m and assume that cj is negative (the
case where cj is positive is totally similar). Since the spectrum map of SU Cj
only contains channel 0, some of its neighbors must open channel 0. Hence,
there exists 1 ≤ i ≤ n such that xi appears in cj and the corresponding
SU Xi opens channel 0. By our construction of A, we have A(xi) = 0, and
thus the clause cj is satisfied by A. Since j is chosen arbitrarily, the formula
8
Hongyu Liang et al.
c1 ∧c2 ∧. . .∧cm is satisfied by A. This completes the reduction from Uniform-
SAT to SpecCon(β, β + 1), and the theorem follows. ⊓⊔
Corollary 1 SpecCon(k, β) is NP-complete for any integers k > β ≥ 2.
Proof By a simple reduction from SpecCon(β + 1, β): Given an instance of
SpecCon(β+1, β), create k−β−1 new channels and add them to the spectrum
map of an (arbitrary) SU. This gives a instance of SpecCon(k, β). Since the
new channels are only contained in one SU, they should not be opened, and
thus the two instances are equivalent. Hence the theorem follows. ⊓⊔
Theorem 2 indicates that the Spectrum Connectivity problem is NP-
complete even if the cognitive radio network only has three channels. We
further strengthen this result by proving the following theorem:
Theorem 3 The Spectrum Connectivity problem is NP-complete even if
there are only two channels.
Proof We present a reduction from Uniform-SAT similar as in the proof of
Theorem 2. Let c1 ∧ c2 ∧ . . . ∧ cm be a uniform CNF clause with variable
set {x1, x2, . . . , xn}. Construct a cognitive radio network as follows: There are
two channels {0,1}. For each variable xi there is a corresponding SU Xi with
spectrum map SpecMap(Xi) = {0, 1} and antenna budget β(Xi) = 1. For
each clause cj there is a corresponding SU Cj with SpecMap(Cj ) = {pj} and
β(Cj ) = 1, where pj = 1 if cj is positive and pj = 0 if cj is negative. There
is an SU Y with SpecMap(Y ) = {0, 1} and β(Y ) = 2. Note that, unlike in
the case of SpecCon(k, β), SUs can have different antenna budgets. Finally,
the edges of the potential graph include: {Xi, Cj } for all i, j such that xi or
xi appears in cj, and {Y, Xi} for all i. This completes the construction of the
cognitive radio network, which is denoted by I. By an analogous argument as
in the proof of Theorem 2, c1 ∧ c2 ∧ . . . ∧ cm is satisfiable if and only if I is
connectable, concluding the proof of Theorem 3. ⊓⊔
Theorem 3 is sharp in that, as noted before, the problem is polynomial-time
solvable when there is only one channel.
3.2 Exact Algorithms
In this subsection we design algorithms for deciding whether a given cognitive
radio network is connectable. Since the problem is NP-complete, we cannot
expect a polynomial time algorithm.
Let n, k, t denote the number of SUs, the number of channels, and the
maximum size of any SU's spectrum map, respectively (t ≤ k). The simplest
idea is to exhaustively examine all possible spectrum assignments to see if
there exists one that makes the network connected. Since each SU can have
at most 2t possible ways of opening channels, the number of assignments is at
most 2tn. Checking each assignment takes poly(n, k) time. Thus the running
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
9
time of this approach is bounded by 2tn(nk)O(1), which is reasonable when t
is small. However, since in general t can be as large as k, this only gives a
2O(kn) bound, which is unsatisfactory if k is large. In the following we present
another algorithm for the problem that runs faster than the above approach
when k is large.
Theorem 4 There is an algorithm that decides whether a given cognitive radio
network is connectable in time 2O(k+n log n).
Proof Let I be a given cognitive radio network with potential graph PG.
Let n be the number of SUs and k the number of channels. Assume that I
is connected under some spectrum assignment. Clearly the realization graph
contains a spanning tree of PG, say T , as a subgraph. If we change the poten-
tial graph to T while keeping all other parameters unchanged, the resulting
network will still be connected under the same spectrum assignment. Thus,
it suffices to check whether there exists a spanning tree T of G such that I
is connectable when substituting T for PG as its potential graph. Using the
algorithm of [7], we can list all spanning trees of PG in time O(N n) where N
is the number of spanning trees of PG. By Cayley's formula [3,18] we have
N ≤ nn−2. Finally, for each spanning tree T , we can use the algorithm in
Theorem 9 (which will appear in Section 5) to decide whether the network
is connectable in time 2O(k)nO(1). The total running time of the algorithm is
O(nn−2)2O(k)nO(1) = 2O(k+n log n). ⊓⊔
Combining Theorem 4 with the brute-force approach, we obtain:
Corollary 2 The Spectrum Connectivity problem is solvable can be solved
in time 2O(min{kn,k+n log n}).
4 Spectrum Connectivity with Complete Potential Graphs
In this section we consider the special case of the Spectrum Connectivity
problem, in which the potential graph of the cognitive radio network is com-
plete. We first show that this restriction does not make the problem tractable
in polynomial time.
Theorem 5 The Spectrum Connectivity problem is NP-complete even
when the potential graph is complete and all SUs have the same antenna budget
β = 2.
Proof The membership in NP is trivial. The hardness proof is by a reduc-
tion from the Hamiltonian Path problem, which is to decide whether a
given graph contains a Hamiltonian path, i.e., a simple path that passes ev-
ery vertex exactly once. The Hamiltonian Path problem is well-known to
be NP-complete [8]. Let G = (V, E) be an input graph of the Hamiltonian
Path problem. Construct an instance of the Spectrum Connectivity prob-
lem as follows: The collection of channels is E and the set of SUs is V ; that
10
Hongyu Liang et al.
is, we identify a vertex in V as an SU and an edge in E as a channel. For
every v ∈ V , the spectrum map of v is the set of edges incident to v. All SUs
have antenna budget β = 2. Denote this cognitive radio network by I. We will
prove that G contains a Hamiltonian path if and only if I is connectable.
First suppose G contains a Hamiltonian path P = v1v2 . . . vn, where n =
V . Consider the following spectrum assignment of I: for each 1 ≤ i ≤ n,
let SU vi open the channels corresponding to the edges incident to vi in the
path P . Thus all SUs open two channels except for v1 and vn each of whom
opens only one. For every 1 ≤ i ≤ n− 1, vi and vi+1 are connected through the
channel (edge) {vi, vi+1}. Hence the realization graph of I under this spectrum
assignment is connected.
Now we prove the other direction. Assume that I is connectable. Fix a
spectrum assignment under which the realization graph of I is connected,
and consider this particular realization graph RG = (V, E′). Let {vi, vj} be
an arbitrary edge in E′. By the definition of the realization graph, there is a
channel opened by both vi and vj. Thus there is an edge in E incident to both
vi and vj, which can only be {vi, vj}. Therefore {vi, vj} ∈ E. This indicates
E′ ⊆ E, and hence RG is a connected spanning subgraph of G. Since each
SU can open at most two channels, the maximum degree of RG is at most 2.
Therefore RG is either a Hamiltonian path of G, or a Hamiltonian cycle which
contains a Hamiltonian path of G. Thus, G contains a Hamiltonian path.
The reduction is complete and the theorem follows. ⊓⊔
Notice that the reduction used in the proof of Theorem 5 creates a cognitive
radio network with an unbounded number of channels. Thus Theorem 5 is not
stronger than Theorem 1 or 3. Recall that Theorem 3 says the Spectrum
Connectivity problem is NP-complete even if there are only two channels.
In contrast we will show that, with complete potential graphs, the problem is
polynomial-time tractable when the number of channels is small.
Theorem 6 The Spectrum Connectivity problem with complete potential
graphs can be solved in 22k+O(k)nO(1) time.
Proof Consider a cognitive radio network I with SU set U , channel set C
and a complete potential graph, i.e., there is a potential edge between every
pair of distinct SUs. Recall that n = U and k = C. For each spectrum
assignment SA, we construct a corresponding spectrum graph GSA = (V, E)
where V = {C′ ⊆ C ∃u ∈ U s.t. SA(u) = C′} and E = {{C1, C2} C1, C2 ∈
V ; C1 ∩ C2 6= ∅}. Thus, V is the collection of subsets of C that is opened
by some SU, and E reflexes the connectivity between pairs of SUs that open
the corresponding channels. Since each vertex in V is a subset of C, we have
V ≤ 2k, and the number of different spectrum graphs is at most 22k
.
We now present a relation between GSA and the realization graph of I
under SA. If we label each vertex u in the realization graph with SA(u),
and contract all edges between vertices with the same label, then we obtain
precisely the spectrum graph GSA = (V, E). Therefore, in the language of
graph theory, GSA = (V, E) is a minor of the realization graph under SA.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
11
Since graph minor preserves connectivity, I is connectable if and only if there
exists a connected spectrum graph. Hence we can focus on the problem of
deciding whether a connected spectrum graph exists.
Consider all possible graphs G = (V, E) such that V ⊆ 2C, and E =
{{C1, C2} C1, C2 ∈ V ; C1 ∩ C2 6= ∅}. There are 22k
such graphs each of which
has size 2O(k). Thus we can list all such graphs in 22k+O(k) time. For each
graph G, we need to check whether it is the spectrum graph of some spectrum
assignment of I. We create a bipartite graph in which nodes on the left side
are the SUs in I, and nodes on the right side all the vertices of G. We add an
edge between an SU u and a vertex C′ of G if and only if C′ ⊆ SpecMap(u)
and C′ ≤ β(u), that is, u can open C′ in a spectrum assignment. The size
of H is poly(n, 2k) and its construction can be finished in poly(n, 2k) time.
Now, if G is the spectrum graph of some spectrum assignment SA, then we
can identify SA with a subgraph of H consisting of all edges (u, SA(u)) where
u is an SU. In addition, in this subgraph we have
-- every SU u has degree exactly one; and
-- every node C′ on the right side of H has degree at least one.
Conversely, a subgraph of H satisfying the above two conditions clearly
induces a spectrum assignment whose spectrum graph is exactly G. Therefore
it suffices to examine whether H contains such a subgraph. Furthermore, the
above conditions are easily seen to be equivalent to:
-- every SU u has degree at least one in G; and
-- G contains a matching that includes all nodes on the right side.
The first condition can be checked in time linear in the size of H, and the
second one can be examined by any polynomial time algorithm for bipartite
matching (e.g., [9]). Therefore, we can decide whether such subgraph exists
(and find one if so) in time poly(n, 2k). By our previous analyses, this solves
the Spectrum Connectivity problem with complete potential graphs. The
total running time of our algorithm is 22k+O(k)poly(n, 2k) = 22k+O(k)nO(1).
⊓⊔
Theorem 7 The Spectrum Connectivity problem with complete potential
graphs is fixed parameter tractable (FPT) when parameterized by the number
of channels.
5 Spectrum Connectivity on Trees and Bounded Treewidth Graphs
In this section, we study another special case of the Spectrum Connectiv-
ity problem where the potential graph of the cognitive radio network is a tree.
We will also investigate the problem on the class of bounded-treewidth graphs.
Many NP-hard combinatorial problems become easy on trees, e.g., the domi-
nating set problem and the vertex cover problem. Nonetheless, as indicated by
the following theorem, the Spectrum Connectivity problem remains hard
on trees.
12
5.1 Trees
Hongyu Liang et al.
We state the complexity of the spectrum connectivity problem with trees as
the potential graph in the following theorem.
Theorem 8 The Spectrum Connectivity problem is NP-complete even if
the potential graph is a tree of depth one.
Proof We give a reduction from the Vertex Cover problem which is well
known to be NP-complete [8]. Given a graph G = (V, E) and an integer r, the
Vertex Cover problem is to decide whether there exists r vertices in V that
cover all the edges in E. Construct a cognitive radio network I as follows. The
set of channels is C = {cv v ∈ V }. For each edge e = {u, v} ∈ E there is an
SU Ue with SpecMap(Ue) = {cu, cv} and antenna budget 2. There is another
SU M with SpecMap(M ) = C and antenna budget r. The potential graph is
a star centered at M , that is, there is a potential edge between M and Ue for
every e ∈ E. This finishes the construction of I.
We prove that G has a vertex cover of size r if and only if I is connectable.
First assume G has a vertex cover S ⊆ V with S ≤ r. Define a spectrum
assignment A(S) as follows: M opens the channels {cv v ∈ S}, and Ue opens
both channels in its spectrum map for all e ∈ E. Since S is a vertex cover,
we have u ∈ S or v ∈ S for each e = {u, v} ∈ E. Thus at least one of cu
and cv is opened by M , which makes it connected to Ue. Hence the realization
graph is connected. On the other hand, assume that the realization graph is
connected under some spectrum assignment. For each e = {u, v} ∈ E, since
the potential edge {M, Ue} is realized, M opens at least one of cu and cv. Now
define S = {v ∈ V cv is opened by M }. It is clear that S is a vertex cover of
G of size at most β(M ) = r. This completes the reduction, and the theorem
follows. ⊓⊔
We next show that, in contrast to Theorems 2 and 3, this special case of
the problem is polynomial-time solvable when the number of channels is small.
Theorem 9 Given a cognitive radio network whose potential graph is a tree,
we can check whether it is connectable in 2O(t)(kn)O(1) time, where t is the
maximum size of any SU's spectrum map. In particular, this running time is
at most 2O(k)nO(1).
Proof Let I be a given cognitive radio network whose potential graph PG =
(V, E) is a tree. Root PG at an arbitrary node, say r. For each v ∈ V let
PGv denote the subtree rooted at v, and let Iv denote the cognitive radio
network obtained by restricting I on PGv. For every subset S ⊆ SpecMap(v),
define f (v, S) to be 1 if there exists a spectrum assignment that makes Iv
connected in which the set of channels opened by v is exactly S; let f (v, S) = 0
otherwise. For each channel c ∈ C, define g(v, c) to be 1 if there exists S,
{c} ⊆ S ⊆ SpecMap(v), for which f (v, S) = 1; define g(v, c) = 0 otherwise.
Clearly I is connectable if and only if there exists S ⊆ SpecMap(r) such that
f (r, S) = 1.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
13
We compute all f (v, S) and g(v, c) by dynamic programming in a bottom-
up manner. Initially all values to set to 0. The values for leaf nodes are easy to
obtain. Assume we want to compute f (v, S), given that the values of f (v′, S′)
and g(v′, c) are all known if v′ is a child of v. Then f (v, S) = 1 if and only if
for every child v′ of v, there exists c ∈ S such that g(v′, c) = 1 (in which case
v and v′ are connected through channel c). If f (v, S) turns out to be 1, we set
g(v, c) to 1 for all c ∈ S. It is easy to see that g(v, c) will be correctly computed
after the values of f (v, S) are obtained for all possible S. After all values have
been computed, we check whether f (r, S) = 1 for some S ⊆ SpecMap(r).
Recall that n = V , k = C, and denote t = maxv∈V SpecMap(v).
There are at most n(2t + k) terms to be computed, each of which takes time
poly(n, k) by our previous analysis. The final checking step takes 2tpoly(n, k)
time. Hence the total running time is 2tpoly(n, k) = 2t(kn)O(1), which is at
most 2O(k)nO(1) since t ≤ k. Finally note that it is easy to modify the algorithm
so that, given a connectable network it will return a spectrum assignment that
makes it connected. ⊓⊔
Corollary 3 The Spectrum Connectivity problem with trees as poten-
tial graphs is fixed parameter tractable when parameterized by the number of
channels.
5.2 Bounded Treewidth Graphs
In this part we deal with another class of potential graphs, namely the class
of graphs with bounded treewidth. Our main result is the following theorem,
which generalizes Theorem 9 as a tree has treewidth one.
Theorem 10 There is an algorithm that, given a cognitive radio network
whose potential graph has bounded treewidth, checks whether it is connectable
in 2O(k)nO(1) time.
Proof Suppose we are given a cognitive radio network I with potential graph
G = (V, E), which has treewidth tw = O(1). Let (T = (I, F ), {Xi i ∈ I})
be a nice tree decomposition of G of width tw (see Section 2.2 for the related
notions). Recall that T is a rooted binary tree with O(V ) nodes and can be
found in polynomial time. Let r be the root of T . For every non-leaf node
i of T , let iL and iR be the two children of i. (We can always add dummy
leaf-nodes to make every non-leaf node have exactly two children, which at
most doubles the size of T .)
For each i ∈ I, define
Yi := {v ∈ Xj j = i or j is a descendent of i},
and let Ii be a new instance of the problem that is almost identical to I except
that we replace the potential graph with G[Yi], i.e., the subgraph of G induced
on the vertex set Yi ⊆ V .
14
Hongyu Liang et al.
For each i ∈ I, suppose Xi = {v1, v2, . . . , vt} where t = Xi and vj ∈ V
for all 1 ≤ j ≤ t. For each tuple (S1, S2, . . . , St) such that Sj ⊆ SpecMap(vj )
for all 1 ≤ j ≤ t, we use a Boolean variable Bi(S1, S2, . . . , St) to indicate
whether there exists a spectrum assignment SAi that makes Ii connected
such that SAi(vj) = Sj for all 1 ≤ j ≤ t. Notice that for each i, the number
of such variables is at most (2k)Xi ≤ 2k·tw, and we can list them in 2O(k·tw)
time. Initially all variables are set to FALSE. Assume Xr = {w1, w2, . . . , wXr }
(recall that r is the root of T ). Then, clearly, deciding whether I is connectable
is equivalent to checking whether there exists (S1, S2, . . . , SXr ), where Sj ⊆
SpecMap(wj) for all 1 ≤ j ≤ Xr, such that Br(S1, S2, . . . , SXr) is TRUE.
We will compute the values of all possible Bi(S1, S2, . . . , St) by dynamic
programming. For each leaf node l, we can compute the values of all the
variables related to Il in time 2O(k·tw)nO(1) by the brute-force approach.
Now suppose we want to decide the value of Bi(S1, S2, . . . , SXi) for some
non-leaf node i, provided that the variables related to any children of i have
all been correctly computed. Recall that iL and iR are the two children of i.
We define:
-- N EW = Xi \ (YiL ∪ YiR );
-- OLD = Xi \ N EW = Xi ∩ (YiL ∪ YiR);
-- ZL = YiL \ Xi, and ZR = YiR \ Xi.
It is clear that Yi = N EW ∪ OLD ∪ ZL ∪ ZR. By using the properties of a
tree decomposition, we have the following fact:
Lemma 2 N EW , ZL, and ZR are three pairwise disjoint subsets of V , and
there is no edge of G whose endpoints lie in different subsets.
Proof Since N EW ⊆ Xi and ZL = YiL \ Xi, we have N EW ∩ ZL = ∅, and
similarly N EW ∩ ZR = ∅. Assume that ZL ∩ ZR 6= ∅, and let v ∈ ZL ∩ ZR.
Since ZL ⊆ YiL and ZR ⊆ YiR , we have v ∈ YiL ∩ YiR . By the definition
of a tree decomposition, v ∈ Xi, so v ∈ Xi ∩ ZL = Xi ∩ (YiL \ Xi) = ∅, a
contradiction. Therefore ZL ∩ ZR = ∅. This proves the pairwise disjointness of
the three sets.
Now assume that there exists an edge e = (u, v) ∈ E such that u ∈ ZL and
v ∈ ZR. Then, by the definition of a tree decomposition, there exists p ∈ I
such that {u, v} ⊆ Xp. We know that p 6= i. So there are three possibilities:
p lines in the subtree rooted at XiL , or in the subtree rooted at XiR, or it is
not in the subtree rooted at Xi. It is easy to verify that, in each of the three
cases, we can find a path that connects two tree nodes both containing u (or
v) and goes through i, which implies u ∈ Xi or v ∈ Xi by the property of
a tree decomposition. This contradicts our previous result. Thus there is no
edge with one endpoint in ZL and another in ZR. Similarly, we can prove that
there exists no edge with one endpoint in N EW and another in ZL or ZR.
This completes the proof of the lemma. ⊓⊔
We now continue the proof of Theorem 10. Recall that we want to decide
Bi(S1, S2, . . . , SXi), i.e., whether Ii, the network with G[Yi] as the potential
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
15
graph, is connectable under some spectrum assignment SA such that SA(vj) =
Sj for all 1 ≤ j ≤ Xi (we assume that Xi = {v1, v2, . . . , vXi}). Note that
Yi = N EW ∪ OLD ∪ ZL ∪ ZR. Due to Lemma 2, the three subsets N EW ,
ZL and ZR can only be connected through OLD (or, we can think OLD as
an "intermediate" set). Therefore, for any spectrum assignment SA such that
SA(vj) = Sj for all j, Ii is connected under SA if and only if the following
three things simultaneously hold:
-- G[Xi] is connected under SA;
-- BiL (S′
1, . . . , S′
XiL ) is TRUE for some (S′
XiL ) that accords with
(S1, . . . , SXi), i.e., the two vectors coincide on any component correspond-
ing to a vertex in Xi ∩ XiL ;
1, . . . , S′
-- BiR(S′
1, . . . , S′
XiR ) is TRUE for some (S′
XiR ) that accords with
(S1, . . . , SXi), i.e., the two vectors coincide on any component correspond-
ing to a vertex in Xi ∩ XiR.
1, . . . , S′
The first condition above can be checked in polynomial time, and the last
two conditions can be verified in 2O(k·tw)nO(1) time. Thus the time spent
on determining Bi(S1, . . . , SXi) is 2O(k·tw)nO(1). After all such terms have
been computed, we can get the correct answer by checking whether there ex-
ists (S1, . . . , SXr ) such that Br(S1, . . . , SXr) is TRUE, which costs another
2O(k·tw)nO(1) time. Since there are at most O(V ) = O(n) nodes in T , the to-
tal running time of the algorithm is 2O(k·tw)nO(1) = 2O(k)nO(1) as tw = O(1).
The proof is complete. ⊓⊔
Corollary 4 The Spectrum Connectivity problem on bounded treewidth
graphs is fixed parameter tractable when parameterized by the number of chan-
nels.
6 Conclusion and Future Work
In this paper, we initiate a systematic study on the algorithmic complexity
of connectivity problem in cognitive radio networks through spectrum assign-
ment. The hardness of the problem in the general case and several special
cases are addressed, and exact algorithms are also derived to check whether
the network is connectable.
In some applications, when the given cognitive radio network is not con-
nectable, we may want to connect the largest subset of the secondary users.
This optimization problem is NP-hard, since the decision version is already
NP-complete on very restricted instances. Thus it is interesting to design poly-
nomial time approximation algorithms for this optimization problem.
In some other scenarios, we may wish to connect all the secondary users
but keep the antenna budget as low as possible. That is, we want to find
the smallest β such that there exists a spectrum assignment connecting the
graph in which each SU opens at most β channels. It is easy to see that this
problem generalizes the minimum-degree spanning tree problem [8], which
16
Hongyu Liang et al.
asks to find a spanning tree of a given graph in which the maximum vertex
degree is minimized. The latter problem is NP-hard, but there is a polynomial
time algorithm that finds a spanning tree of degree at most one more than
the optimum [6]. It would be interesting to see whether this algorithm can be
generalized to the min-budget version of our connectivity problem, or whether
we can at least obtain constant factor approximations.
Another meaningful extension of this work is to design distributed algo-
rithms to achieve network connectivity. Moreover, due to interference in wire-
less communications, the connected nodes using the same channel may not
be able to communicate simultaneously. Therefore, it is also interesting to in-
vestigate distributed algorithms with channel assignment and link scheduling
jointly considered to achieve some network objective such as connectivity and
capacity maximization, especially under the realistic interference models.
Acknowledgements
The authors would like to give thanks to Dr. Thomas Moscibroda at Microsoft
Research Asia for his introduction of the original problem. This work was
supported in part by the National Basic Research Program of China Grant
2011CBA00300, 2011CBA00302, the National Natural Science Foundation of
China Grant 61073174, 61103186, 61202360, 61033001, and 61061130540.
References
1. P. Bahl, R. Chandra, T. Moscibroda, R. Murty, and M. Welsh. White space networking
with Wi-Fi like connectivity. In ACM SIGCOMM, 2009.
2. H. L. Bodlaender. A linear-time algorithm for finding tree-decompositions of small
treewidth. SIAM J. Comput., 25(6):1305 -- 1317, 1996.
3. A. Cayley. A theorem on trees. Quarterly Journal on Pure and Applied Mathematics,
23:376 -- 378, 1889.
4. S. Dolev, S. Gilbert, R. Guerraoui, and C. C. Newport. Gossiping in a multi-channel
radio network. In DISC, 2007.
5. R. Downey and M. Fellows. Parameterized Complexity. Springer, 1998.
6. M. Furer and B. Raghavachari. Approximating the minimum-degree steiner tree to
within one of optimal. J. Algorithms, 17(3):409 -- 423, 1994.
7. H. Gabow and E. Myers. Finding all spanning trees of directed and undirected graphs.
SIAM J. Comput., 7(3):280 -- 287, 1978.
8. M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman, 1979.
9. J. Hopcroft and R. Karp. An n5/2 algorithm for maximum matchings in bipartite
graphs. SIAM J. Comput., 2(4):225 -- 231, 1973.
10. T. Kloks. Treewidth: Computations and Approximations, volume 842 of LNCS.
Springer, 1994.
11. F. Kuhn, R. Wattenhofer, and A. Zollinger. Ad hoc networks beyond unit disk graphs.
Wireless Netwoks, 14(5):715 -- 729, 2008.
12. X.-Y. Li, P. Yang, Y. Yan, L. You, S. Tang, and Q. Huang. Almost optimal accessing
of nonstochastic channels in cognitive radio networks. In IEEE INFOCOM, 2012.
13. H. Liang, T. Lou, H. Tan, Y. Wang, and D. Yu. Complexity of connectivity in cognitive
radio networks through spectrum assignment. In ALGOSENSORS'12, volume 7718 of
LNCS, pages 108 -- 119, 2013.
On the Complexity of Connectivity in CRNs Through Spectrum Assignment
17
14. D. Lu, X. Huang, P. Li, and J. Fan. Connectivity of large-scale cognitive radio ad hoc
networks. In IEEE INFOCOM, 2012.
15. W. Ren, Q. Zhao, and A. Swami. Connectivity of cognitive radio networks: Proximity
vs. opportunity. In ACM CoRoNet, 2009.
16. W. Ren, Q. Zhao, and A. Swami. Power control in cognitive radio networks: How to
cross a multi-lane highway. IEEE JSAC, 27(7):1283 -- 1296, 2009.
17. N. Robertson and P. D. Seymour. Graph minors. ii. algorithmic aspects of treewidth.
J. Algorithms, 7:309 -- 322, 1986.
18. A. Shukla. A short proof of Cayley's tree formula. arXiv:0908.2324v2.
19. B. Wang and K. J. R. Liu. Advances in cognitive radio networks: A survey. IEEE J.
Selected Topics in Signal Processing, 5(1):5 -- 23, 2011.
20. C. Xu and J. Huang. Spatial spectrum access game: nash equilibria and distributed
learning. In ACM MobiHoc, 2012.
21. J. Yu, H. Roh, W. Lee, S. Pack, and D.-Z. Du. Topology control in cooperative wireless
ad-hoc networks. IEEE JSAC, 30(9):1771 -- 1779, 2012.
22. Y. Yuan, P. Bahl, R. Chandra, T. Moscibroda, and Y. Wu. Allocating dynamic time-
spectrum blocks in cognitive radio networks. In ACM MobiCom, 2007.
23. X. Zhou, S. Gandhi, S. Suri, and H. Zheng. ebay in the sky: Strategyproof wireless
spectrum auctions. In ACM MobiCom, 2008.
|
1612.06016 | 1 | 1612 | 2016-12-19T00:52:17 | A Characterization of Constant-Sample Testable Properties | [
"cs.DS"
] | We characterize the set of properties of Boolean-valued functions on a finite domain $\mathcal{X}$ that are testable with a constant number of samples. Specifically, we show that a property $\mathcal{P}$ is testable with a constant number of samples if and only if it is (essentially) a $k$-part symmetric property for some constant $k$, where a property is {\em $k$-part symmetric} if there is a partition $S_1,\ldots,S_k$ of $\mathcal{X}$ such that whether $f:\mathcal{X} \to \{0,1\}$ satisfies the property is determined solely by the densities of $f$ on $S_1,\ldots,S_k$. We use this characterization to obtain a number of corollaries, namely: (i) A graph property $\mathcal{P}$ is testable with a constant number of samples if and only if whether a graph $G$ satisfies $\mathcal{P}$ is (essentially) determined by the edge density of $G$. (ii) An affine-invariant property $\mathcal{P}$ of functions $f:\mathbb{F}_p^n \to \{0,1\}$ is testable with a constant number of samples if and only if whether $f$ satisfies $\mathcal{P}$ is (essentially) determined by the density of $f$. (iii) For every constant $d \geq 1$, monotonicity of functions $f : [n]^d \to \{0, 1\}$ on the $d$-dimensional hypergrid is testable with a constant number of samples. | cs.DS | cs |
A Characterization of Constant-Sample Testable Properties
Eric Blais∗
Yuichi Yoshida†
David R. Cheriton School of Computer Science
National Institute of Informatics and
University of Waterloo
[email protected]
Preferred Infrastructure, Inc.
[email protected]
April 23, 2018
Abstract
We characterize the set of properties of Boolean-valued functions on a finite domain X that
are testable with a constant number of samples. Specifically, we show that a property P is
testable with a constant number of samples if and only if it is (essentially) a k-part symmetric
property for some constant k, where a property is k-part symmetric if there is a partition
S1, . . . , Sk of X such that whether f : X → {0, 1} satisfies the property is determined solely by
the densities of f on S1, . . . , Sk.
We use this characterization to obtain a number of corollaries, namely:
• A graph property P is testable with a constant number of samples if and only if whether
a graph G satisfies P is (essentially) determined by the edge density of G.
• An affine-invariant property P of functions f : Fn
p → {0, 1} is testable with a constant
number of samples if and only if whether f satisfies P is (essentially) determined by the
density of f .
• For every constant d ≥ 1, monotonicity of functions f : [n]d → {0, 1} on the d-dimensional
hypergrid is testable with a constant number of samples.
1
Introduction
Property testing [22, 17] is concerned with the general question: for which properties of combi-
natorial objects can we efficiently distinguish the objects that have the property from those that
are "far" from having the same property? Consideration of this question has led to surprisingly
powerful results: many natural graph properties [17, 2], algebraic properties of functions on finite
fields [22, 27], and structural properties of Boolean functions [12, 10], for example, can be tested
with a constant number of queries to the object being tested. Nearly all of these results appear
to rely critically on the testing algorithm's ability to query the unknown object at locations of its
choosing. The goal of the present work is to determine to what extent this is actually the case: is
it possible that some (or most!) of these properties can still be tested efficiently even if the tester
has no control over the choice of queries that it makes?
∗Supported by an NSERC Discovery Grant
†Supported by JSPS Grant-in-Aid for Young Scientists (B) (No. 26730009), MEXT Grant-in-Aid for Scientific
Research on Innovative Areas (24106003), and JST, ERATO, Kawarabayashi Large Graph Project.
1
Our main question is made precise in the sample-based property testing model originally intro-
duced by Goldreich, Goldwasser, and Ron [17]. For a finite set X , let {0, 1}X denote the set of
Boolean-valued functions on X endowed with the normalized Hamming distance metric d(f, g) :=
{x ∈ X : f (x) 6= g(x)}/X . A property of functions mapping X to {0, 1} is a subset P ⊆ {0, 1}X .
A function is ǫ-close to P if it is in the set Pǫ := {f ∈ {0, 1}X : ∃g ∈ P s.t. d(f, g) ≤ ǫ}; otherwise,
it is ǫ-far from P. An s-sample ǫ-tester for a property P is a randomized algorithm with bounded
error that observes s pairs (x1, f (x1)), . . . , (xs, f (xs)) ∈ X × {0, 1} with x1, . . . , xs drawn indepen-
dently and uniformly at random from X and then accepts when f ∈ P and rejects when f is ǫ-far
from P. The sample complexity of P for some ǫ > 0 is the minimum value of s such that it has an
s-sample ǫ-tester. When the sample complexity of P is independent of the domain size X of the
functions for every ǫ > 0, we say that P is constant-sample testable.
In the 20 years since the original introduction of the sample-based property testing model,
a number of different properties have been shown to be constant-sample testable, including all
symmetric properties (folklore; see the discussion in [19]), unions of intervals [20], decision trees over
low-dimensional domains [20], and convexity of images [21, 5]. Conversely, many other properties,
such as monotonicity of Boolean functions [16], linearity [3, 18], linear threshold functions [4], and
k-colorability of graphs [18], are known to not be constant-sample testable. Yet, despite a wide-
spread belief that most properties are not testable with a constant number of samples, it remained
open until now to determine whether this is actually the case or not. Our main result settles
this question, and in the process unifies all of the above results and explains what makes various
properties testable with a constant number of samples or not.
1.1 Main result
We show that constant-sample testability is closely tied to a particular notion of symmetry-or
invariance-of properties. Let SX denote the set of permutations on a finite set X , and for any
subset S ⊆ X , let S (S)
denote the set of permutations on X that preserves the elements in S.
X
A permutation π ∈ SX acts on functions f : X → {0, 1} in the obvious way: πf is the function
that satisfies (πf )(x) = f (πx) for every x ∈ X . The property P ⊆ {0, 1}X is invariant under
a permutation π ∈ SX if for every f ∈ P, we also have πf ∈ P. P is (fully) symmetric if it is
invariant under all permutations in SX . The following definition relaxes this condition to obtain a
notion of "partial" symmetry.
Definition 1. For any k ∈ N, the property P ⊆ {0, 1}X is k-part symmetric if there is a partition
of X into k parts X1, . . . , Xk such that P is invariant under all the permutations in SX1 × · · · × SXk ,
where SXi is the set of permutations over X that preserves elements in Xi.
Equivalently, P is k-part symmetric if there exists a partition X1, . . . , Xk of X such that the
event f ∈ P is completely determined by the density f −1(1)∩Xi
of f in each of the sets X1, . . . , Xk.
Our main result shows that O(1)-part symmetry is also essentially equivalent to constant-sample
testability.
Xi
Theorem 1. The property P of a function f : X → {0, 1} is constant-sample testable if and only
if for any ǫ > 0, there exists a constant k = kP (ǫ) that is independent of X and a k-part symmetric
property P ′ such that P ⊆ P ′ ⊆ Pǫ.
In words, Theorem 1 says that constant-sample testable properties are the properties P that
can be covered by some O(1)-part symmetric property P ′ that does not include any function that
2
is ǫ-far from P. Note that this characterization cannot be replaced with the condition that P
itself is k-part symmetric. To see this, consider the function non-identity property NotEq(g) that
includes every function except some non-constant function g : X → {0, 1}. This property is not
k-part symmetric for any k = O(1), but the trivial algorithm that accepts every function is a valid
ǫ-tester for NotEq(g) for any constant ǫ > 0.
Theorem 1 can easily be generalized to apply to properties of functions mapping X to any finite
set Y. We restrict our attention to the range Y = {0, 1} for simplicity and clarity of presentation.
The sample-based property testing model is naturally extended to non-uniform distributions over
the input domain. It appears likely that Theorem 1 can be generalized to this more general setting
as well, though we have not attempted to do so.
The proof of Theorem 1 follows the general outline of previous characterizations of the properties
testable in the query-based model (e.g., [2, 27, 6]). As with those results, the most interesting part
of the proof is the direction showing that constant-sample testability implies coverage by an O(1)-
part symmetric property, and this proof is established with a regularity lemma. Our proof departs
from previous results in both the type of regularity lemma that we use and in how we use it. For
a more detailed discussion of the proof, see Section 2.1.
1.2 Applications
The characterization of constant-sample testability in Theorem 1 can be used to derive a number
of different corollaries. We describe a few of these.
When X is identified with the (cid:0)n
2(cid:1) pairs of vertices in V , the function f : X → {0, 1} represents
a graph G = (V, E) where E = f −1(1). A graph property is a property of these functions that
is invariant under relabelling of the vertices. Observant readers will have noted that no non-
symmetric graph property was present in the list of properties that have been determined to be
constant-sample testable. Using Theorem 1, we can show that this is unavoidable: the only graph
properties that are constant-sample testable are those that are (essentially) fully symmetric.
Corollary 1. For every ǫ > 0, if P is a graph property that is ǫ-testable with a constant number
of samples, then there is a symmetric property P ′ such that P ⊆ P ′ ⊆ Pǫ.
When X is identified with a finite field Fn
invariant under any affine transformation over Fn
properties are essentially the only constant-sample testable affine-invariant properties.
p , a property P ⊆ {0, 1}X is affine-invariant if it is
p . Theorem 1 can be used to show that symmetric
Corollary 2. For every ǫ > 0, if P is an affine-invariant property of functions f : Fn
p → {0, 1}
that is ǫ-testable with a constant number of samples, then there is a symmetric property P ′ such
that P ⊆ P ′ ⊆ Pǫ.
These two corollaries show that Theorem 1 can be used to show that some properties are
In our third application, we show that the other direction of the
not constant-sample testable.
characterization can also be used to show that some properties are constant-sample testable.
Fix a constant d ≥ 1. Two points x, y ∈ [n]d := {1, 2, . . . , n}d satisfy x (cid:22) y when x1 ≤ y1,
. . ., and xd ≤ yd. The function f : [n]d → {0, 1} is monotone if for every x (cid:22) y ∈ [n]d, we have
f (x) ≤ f (y). When d = 1, it is folklore knowledge that monotonicity of Boolean-valued functions
on the line is constant-sample testable. Using Theorem 1, we give an easy proof showing that the
same holds for every other constant dimension d.
3
Corollary 3. For every constant d ≥ 1 and constant ǫ > 0, we can ǫ-test monotonicity of functions
f : [n]d → {0, 1} on the d-dimensional hypergrid with a constant number of samples.
Chen, Servedio, and Tan [8] showed that the number of queries (and thus also of samples)
required to test monotonicity of f : [n]d → {0, 1} must depend on d. (The same result for the case
where n = 2 was first established by Fischer et al. [14].) Combined with the result above, this
shows that monotonicity of Boolean-valued functions on the hypergrid is constant-sample testable
if and only if the number of dimensions of the hypergrid is constant.
1.3 Related work
Sample-based property testing. The first general result regarding constant-sample testable
properties goes back to the original work of Goldreich, Goldwasser, and Ron [17]. They showed that
every property with constant VC dimension (and, more generally, every property that corresponds
to a class of functions that can be properly learned with a constant number of samples) is constant-
sample testable. As they also show, this condition is not necessary for constant-sample testability-
in fact, there are even properties that are testable with a constant number of samples whose
corresponding class require a linear number of samples to learn [17, Prop. 3.3.1].
More general results on sample-based testers were obtained by Balcan et al. [4]. In particular,
they defined a notion of testing dimension of a property P in terms of the total variation distance
between the distributions on the tester's observations when a function is drawn from distributions
πyes and πno essentially supported on P and Pǫ, respectively. They show that this testing dimension
captures the sample complexity of P up to constant factors, and observe that it can be interpreted
as an "average VC dimension"-type of complexity measure. It would be interesting to see whether
the combinatorial characterization in Theorem 1 could be combined with these results to offer new
insights into the connections between invariance and VC dimension-like complexity measures.
Finally, Goldreich and Ron [18] and Fischer et al. [11, 13] established connections between the
query- and sample-based models of property testing giving sufficient conditions for sublinear -sample
testability of properties. The exact bounds between sample complexity and partial symmetry in
the proof of Theorem 1 yield another sufficient condition for sublinear-sample testability: every
property P that can be ǫ-covered by an o(log log X )-part symmetric function P ′ has sublinear
sample complexity o(X ). As far as we can tell, these two characterizations are incomparable.
Symmetry and testability. The present work was heavily influenced by the systematic ex-
ploration of connections between the invariances of properties and their testability initiated by
Kaufman and Sudan [19]. (See also [24].) In that work, the authors showed that such connections
yield new insights into the testability of algebraic properties in the query-based property testing
model, and advocated for further study of the invariance of properties as a means to better un-
derstand their testability. Theorem 1 provides evidence that this approach is a critical tool in the
study of sample-based property testing model as well.
The notion of partial symmetry and its connections to computational efficiency has a long
history-it goes back at least to the pioneering work of Shannon [23]. Partial symmetry also ap-
peared previously in a property testing context in the authors' joint work with Amit Weinstein
on characterizing the set of functions for which isomorphism testing is constant-query testable [7].
However, it should be noted that the notion of partial symmetry considered in [7] does not cor-
respond to the notion of k-part symmetry studied here. In fact, as mentioned in the conference
4
version of that paper, there are 2-part symmetric functions for which isomorphism testing is not
constant-query testable, so the two characterizations inherently require different notions of partial
symmetry.
1.4 Organization
The proof of Theorem 1 is presented in Section 2. The proofs of the application results are in
Section 3. Finally, since the weak regularity lemma that we use in the proof of Theorem 1 is not
completely standard, we include its proof in Section 4 for completeness.
2 Proof of Theorem 1
We prove the two parts (sufficiency and necessity) of Theorem 1 in Sections 2.2 and 2.3, respectively.
But first, we provide a high-level overview of the proof and discuss the connections with regularity
lemmas in Section 2.1.
2.1 Overview of the proof
Symmetry implies testability. The proof of this direction of the theorem is straightforward
and is obtained by generalizing the following folklore proof that symmetric properties can be tested
with a constant number of samples. Let P be any symmetric property. A tester can estimate
the density Ex∈X [f (x)] up to additive accuracy γ for any small γ > 0 with a constant number of
samples. This estimated density can be used to accept or reject the function based on how close it
is to the density of the functions in P. The validity of this tester is established by showing that a
function can be ǫ-far from P only when its density is far from the density of every function in P.
Consider now a property P that is k-part symmetric for some constant k. Let X1, . . . , Xk
be a partition of X associated with P. We show that a tester which estimates the densities
µi(f ) := Ex∈Xi[f (x)] for each i = 1, . . . , k and then uses these densities to accept or reject is a
valid tester for P. We do so by showing that any function that is ǫ-far from P must have a density
vector that is far from those of every function in P.
Testability implies symmetry. To establish the second part of the theorem, we want to show
that the existence of a constant-sample tester T for a property P implies that there is a partition
of X into a constant number of parts for which P is nearly determined by the density of functions
within those parts. We do so by using a variant of the Frieze–Kannan weak regularity lemma [15] for
hypergraphs. An s-uniform weighted hypergraph is a hypergraph G = (V, E) on V vertices where
E : V s → [0, 1] denotes the weight associated with each hyperedge. Given a partition V1, . . . , Vk of
V and a multi-index I = (i1, . . . , is) with i1, . . . , is ∈ [k], the expected weight of hyperedges of G in
VI is wG(VI ) = wG(Vi1, . . . , Vis) = {E(v1, . . . , vs) : v1 ∈ Vi1, . . . , vs ∈ Vis}/Qi∈I Vi. Furthermore,
for any subset S ⊆ V , we define S ∩ VI = (S ∩ Vi1, . . . , S ∩ Vis).
Lemma 1 (Weak regularity lemma). For every ǫ > 0 and every s-uniform weighted hypergraph
G = (V, E), there is a partition V1, . . . , Vk of V with k = 2O(log( 1
ǫ )/ǫ2) parts such that for every
subset S ⊆ V ,
(1)
Qi∈I S ∩ Vi
V s
X
I∈[k]s
(cid:12)(cid:12)wG(S ∩ VI ) − wG(VI )(cid:12)(cid:12) ≤ ǫ.
5
This specific formulation of the weak regularity lemma seems not to have appeared previously
in the literature, but its proof is essentially the same as that of usual formulations of the weak
regularity lemma. For completeness, we provide a proof of Lemma 1 in Section 4.
Lemma 1 is best described informally when we consider the special case where we consider
unweighted graphs. In this setting, the weak regularity lemma says that for every graph G, there
is a partition of the vertices of G into k = O(1) parts V1, . . . , Vk such that for every subset S of
vertices, the density of edges between S ∩ Vi and S ∩ Vj is close to the density between Vi and Vj on
average over the choice of Vi and Vj. Regularity lemmas where this density-closeness condition is
satisfied for almost all pairs of parts Vi and Vj are known as "strong" regularity lemmas. To the best
of our knowledge, all previous characterization results in property testing that relied on regularity
lemmas (e.g., [2, 27, 6]) used strong regularity lemmas. This approach unavoidably introduces
tower-type dependencies between the query complexity and the characterization parameters. By
using a weak regularity lemma instead, we get a much better (though still doubly-exponential)
dependence between the sample complexity and the partial symmetry parameter.
The second point of departure of our proof from previous characterizations is in how we use the
regularity lemma. In the prior work, the regularity lemma was applied to the tested object itself
(e.g., the dense graphs being tested in [2]) and the testability of the property was used to show
that the objects with the given property could be described by some combinatorial characteristics
related to the regular partition whose existence is promised by the regularity lemma. Instead, in
our proof of Theorem 1, we apply Lemma 1 to a hypergraph associated with the tester itself, not
with the tested object.
Specifically, let T be an s-sample ǫ-tester for some property P ⊆ {0, 1}X . We associate T
with an s-uniform weighted hypergraph GT on the set of vertices X × {0, 1}. The weight of each
s-hyperedge of GT is the acceptance probability of T when its s observations correspond to the s
vertices covered by the hyperedge. By associating each function f : X → {0, 1} with the subset
S ⊆ X × {0, 1} that includes all 2X vertices of the form (x, f (x)), we see that the probability
that T accepts f is the expected value of a hyperedge whose s vertices are drawn uniformly and
independently at random from the set S. We can use Lemma 1 to show that there is a partition
of V into a constant number of parts such that for each function f with associated set S, this
probability is well approximated by some function of the density of S in each of the parts. We then
use this promised partition of V to partition the original input domain X into a constant number
of parts where membership in P is essentially determined by the density of a function in each of
these parts, as required.
2.2 Proof that symmetry implies testability
We now begin the proof of Theorem 1 with the easy direction.
Lemma 2. Let P ⊆ {0, 1}X be a property where for every ǫ > 0, there exists a constant k = kP (ǫ)
that is independent of X and a k-part symmetric property P ′ such that P ⊆ P ′ ⊆ Pǫ. Then P is
constant-sample testable.
Proof. Fix ǫ > 0. We show that we can distinguish functions in P from functions that are ǫ-far from
P with a constant number of samples. From the condition in the statement of the lemma, there
exists a k-part symmetric property P ′ with P ⊆ P ′ ⊆ Pǫ/2 for some k = kP (ǫ/2). Let S1, . . . , Sk
be a partition of X such that whether a function f : X → {0, 1} satisfies P ′ is determined by
µS1(f ), . . . , µSk (f ). For a set S ⊆ X , let cS(f ) = µS(f )S be the number of x ∈ S with f (x) = 1.
6
Our algorithm for testing P is as follows. For each i ∈ [k], we draw q := O(k2 log k/ǫ2) samples
f (xj) for each i ∈ [k]. We accept
x1, . . . , xq and computes the estimates ecSi(f ) := X
if there exists g ∈ P ′ such that
q Pj∈[q]:xj∈Si
X
i∈[k]
ecSi(f ) − cSi(g) <
ǫ
4
X ,
and reject otherwise.
Let us now establish the correctness of the algorithm. By Hoeffding's bound, for each i ∈ [k],
3k by choosing the hidden constant
in the definition of q sufficiently large. By union bound, with probability at least 2/3, we have
we have cSi(f ) −ecSi(f ) < ǫ
cSi(f ) −ecSi(f ) < ǫ
If f ∈ P, then the algorithm accepts f because Pi∈[k] ecSi(f ) − cSi(f ) < ǫ
4 X and f ∈ P ⊆ P ′.
If f is ǫ-far from satisfying P, then for any g ∈ P ′, the triangle inequality and the fact that
4k X for every i ∈ [k]. In what follows, we assume this happens.
4k X with probability at least 1 − 1
P ′ ⊆ Pǫ/2 imply that
X
i∈[k]
ecSi(f ) − cSi(g) ≥ X
i∈[k]
(cid:16)cSi(f ) − cSi(g) − ecSi(f ) − cSi(f )(cid:17) >
ǫ
2
X −
ǫ
4
X =
ǫ
4
X
and, therefore, the algorithm rejects f .
2.3 Testability implies symmetry
Suppose that a property P is testable by a tester T with sample complexity s. We want to show
that for any ǫ > 0, there exists a k-part symmetric property P ′ for k = k(ǫ) such that P ⊆ P ′ ⊆ Pǫ
For any x = (x1, . . . , xs) ∈ X s, we define f (x) = (cid:0)f (x1), . . . , f (xs)(cid:1) and we let T (x, f (x)) ∈ [0, 1]
denote the acceptance probability of the tester T of f when the samples drawn are x. The overall
acceptance probability of f by T is
pT (f ) = E
[T (x, f (x))].
x
(2)
We show that there is a family S of a constant number of subsets of X such that, for every
function f : X → {0, 1}, the acceptance probability pT (f ) is almost completely determined by the
density of f on the subsets in S.
Lemma 3. For any ǫ > 0 and any s-sample tester T , there is a family S = {S1, . . . , Sm} of
m ≤ 2O(22s/ǫ2) subsets of X such that for every f : X → {0, 1},
(cid:12)(cid:12)pT (f ) − ϕT (µS1(f ), . . . , µSm(f ))(cid:12)(cid:12) ≤ ǫ
where µS(f ) = Ex∈S[f (x)] is the density of f on the subset S ⊆ X and ϕT : [0, 1]m → [0, 1] is some
fixed function.
Proof. We consider the weighted hypergraph G = (V, E) where V = X ×{0, 1} and E is constructed
by adding a hyperedge ((x1, y1), . . . , (xs, ys)) of weight T (x, y) for each x = (x1, . . . , xs) ∈ X s and
y = (y1, . . . , ys) ∈ {0, 1}s.
7
A function f : X → {0, 1} corresponds to a subset S ⊆ V of size X = V /2, that is,
S := {(x, f (x)) x ∈ X }. The probability that T accepts f is
pT (f ) = E
x∈X s
[T (x, f (x))] = E
v∈V s
[E(v) v ∈ Ss] = X
Qi∈I Vi ∩ S
Ss
wG(S ∩ VI )
= X
I∈[k]s
2sQi∈I Vi ∩ S
V s
wG(S ∩ VI ) = X
I∈[k]s
I∈[k]s
2sQi∈I Vi ∩ S
V s
wG(VI ) ± ǫ
where the last step is by Lemma 1 applied with approximation parameter ǫ/2s. For every part
Vi, let V 1
i = {x ∈ X : (x, 0) ∈ Vi}. Then the value of
k , V 0
k .
w(VI ) is completely determined by the density of f on V 1
i = {x ∈ X : (x, 1) ∈ Vi} and let V 0
2 , . . . , V 1
1 , V 1
1 , V 0
2 , V 0
PI∈[k]s 2s Qi∈I Vi∩S
V s
We are now ready to complete the second part of the proof of Theorem 1.
Lemma 4. Let P ⊆ {0, 1}X be constant-sample testable. Then for every ǫ > 0, there exists
a constant k = kP (ǫ) that is independent of X and a k-part symmetric property P ′ such that
P ⊆ P ′ ⊆ Pǫ.
Proof. Fix any ǫ > 0 and let T be an s-sample ǫ-tester for P. Let γ < 1
less than 1
m = 2O(22s/γ2) sets such that for every f : X → {0, 1},
3 be any constant that is
3 . By Lemma 3 applied with the parameter γ, there is a family S = {S1, . . . , Sm} with
pT (f ) − ϕT (µS1(f ), . . . , µSk (f )) ≤ γ.
(3)
Define
P ′ = {f : X → {0, 1} : ∃g ∈ P s.t. (µS1(f ), . . . , µSk (f )) = (µS1(g), . . . , µSk (g))}.
This construction trivially guarantees that P ′ ⊇ P. Furthermore, (3) guarantees that for every
f ∈ P ′, if we let g ∈ P be one of the elements with the same density profile as f ,
pT (f ) ≥ ϕT (µS1(f ), . . . , µSk (f )) − γ = ϕT (µS1(g), . . . , µSk (g)) − γ ≥ pT (g) − 2γ.
Since γ < 1/3, the fact that T is an ǫ-tester for P and that g ∈ P imply that f ∈ Pǫ.
Let S′
1, . . . , S′
S1, . . . , Sm. Note that S′
, . . . , µS′
by µS′
the partition S′
k be the family of sets obtained by taking intersections and complements of
k forms a partition of X and µS1, . . . , µSm is completely determined
. Furthermore, k = O(2m). Hence, P ′ is a k-part symmetric property induced by
1, . . . , S′
k with P ⊆ P ′ ⊆ Pǫ.
1, . . . , S′
k
1
Theorem 1 follows immediately from Lemmas 2 and 4.
3 Applications
3.1 Affine-invariant properties
For an affine transformation A : Fn
to be the function that satisfies Af (x) = f (Ax) for every x ∈ Fn
p and a function f : Fn
p → Fn
p → {0, 1}, we define Af : Fn
p → {0, 1}
p . A property P of functions
8
p → {0, 1} is affine-invariant if for any affine transformation A : Fn
f : Fn
p and f ∈ P, we
have Af ∈ P. Our characterization shows that the only affine-invariant properties of functions
f : Fn
p → {0, 1} that are testable with a constant number of samples are the (fully symmetric)
properties that are determined by the density of f .
Corollary 2 (Restated). For every ǫ > 0, if P is an affine-invariant property of functions f : Fn
p →
{0, 1} that is ǫ-testable with a constant number of samples, then there is a symmetric property P ′
such that P ⊆ P ′ ⊆ Pǫ.
p → Fn
Proof. By Theorem 1, if P is testable with O(1) samples, then there are subsets S1, . . . , Sk of
p with k = O(1) and a property P ′′ such that P ⊆ P ′′ ⊆ Pǫ and P ′′ is invariant under all
Fn
permutations of S1, . . . , Sk. Let P ′ be the closure of P ′′ under all affine transformations over Fn
p .
(I.e., P ′ = {Af : f ∈ P ′′, A is affine}.) Since P itself is invariant under affine transformations, we
have that P ′ ⊆ Pǫ. We want to show that P ′ is symmetric. To show this, it suffices to show that
p and f ∈ P ′, the function g obtained
P ′ is closed under transpositions. I.e., that for every x, y ∈ Fn
by setting g(x) = f (y), g(y) = f (x), and g(z) = f (z) for every other z /∈ {x, y} is also in P ′. We
write g = (x y)f to denote the action of the transposition (x y) on f .
If x, y ∈ Si for some i ∈ [k], then our conclusion follows immediately from the invariance
of P ′′ over permutations on Si. Otherwise, let w, z ∈ Sj for some j ∈ [k]. Since the set of affine
transformations is a doubly-transitive action on Fn
p , there is a transformation A such that A(x) = w
and A(y) = z. Then (x y)f = A−1(w z)Af ∈ P ′, as we wanted to show.
3.2 Graph properties
Similarly, our characterization shows that the only graph properties that are testable with a constant
number of samples are the (fully symmetric) properties that are determined by the edge density of
the graph.
Corollary 1 (Restated). For every ǫ > 0, if P is a graph property that is ǫ-testable with a constant
number of samples, then there is a symmetric property P ′ such that P ⊆ P ′ ⊆ Pǫ.
Proof. The proof is nearly identical. By Theorem 1, if P is testable with O(1) samples, then there
are subsets S1, . . . , Sk of the edge set with k = O(1) and a property P ′′ such that P ⊆ P ′′ ⊆ Pǫ
and P ′′ is invariant under all permutations of S1, . . . , Sk. Let P ′ be the closure of P ′′ under all
permutations of the vertex set. We again have that P ⊆ P ′ ⊆ Pǫ and we want to show that P ′ is
invariant under every transposition of the edge set.
The one change with the affine-invariant property proof is that the set of permutations of the
vertices of a graph is not a 2-transitive action on the set of edges of this graph. But the same idea
still works because we can always find a vertex permutation to send two edges on disjoint vertices
to a same part Si that also contains a pair of edges on disjoint vertices, and we can find a vertex
permutation to send two edges that share a common vertex to a part Sj that also contains two
edges that share a common vertex. So for every pair of edges e1, e2, we have a vertex permutation
πV and a pair of edges e3, e4 ∈ Sℓ such that if G ∈ P ′, then (e1 e2)G = π −1
V (e3 e4)πV G ∈ P ′.
3.3 Testing monotonicity
With Theorem 1, to show that monotonicity of functions f : [n]d → {0, 1} is constant-sample
testable, it suffices to identify an O(1)-part symmetric function that covers all the monotone func-
tions and does not include any function that is far from monotone. This is what we do below.
9
Corollary 3 (Restated). For every constant d ≥ 1 and constant ǫ > 0, we can ǫ-test monotonicity
of functions f : [n]d → {0, 1} on the d-dimensional hypergrid with a constant number of samples.
Proof. For any ǫ > 0, fix k = ⌈d/ǫ⌉ and let R be a partition of the space [n]d into kd subgrids of
side length (at most) ⌊ǫn⌋ each. We identify the parts in R with the points in [k]d. For an input
x ∈ [n]d, let φR(x) denote the part of R that contains x.
Given some function f : [n]d → {0, 1}, define the R-granular representation of f to be the
function fR : [k]d → {0, 1, ∗} defined by
fR(x) =
0 if ∀y ∈ [n]d with φR(y) = x, f (y) = 0
1 if ∀y ∈ [n]d with φR(y) = x, f (y) = 1
∗ otherwise.
Let P = {f : [n]d → {0, 1} : ∃ monotone g s.t. fR = gR} be the property that includes every
function whose R-granular representation equals that of a monotone function. By construction P
includes all the monotone functions and is invariant under any permutations within the O(1) parts
of R. To complete the proof of the theorem, we want to show that every function in P is ǫ-close to
monotone.
Fix any f ∈ P and let g : [n]d → {0, 1} be a monotone function for which fR = gR. The
distance between f and g is bounded by
d(f, g) ≤
g−1
R (∗)
kd
.
Now consider the poset P on [k]d where x ≺ y iff xi < yi for every i ∈ [d]. We first observe that
the set g−1
R (∗) forms an anti-chain on this poset. Indeed, if there exist x, y ∈ [k]d with x ≺ y and
g(x) = g(y) = ∗, then there exist x′, y′ ∈ [n]d such that φR(x′) = x, φR(y′) = y, g(x′) = 1, and
g(y′) = 0. But this contradicts the monotonicity of g because x′ ≤ y′ holds from x ≺ y.
For x ∈ [k]d, let xmax = maxi∈[d] xi and xmin = mini∈[d] xi. Define 1 = (1, 1, 1, . . . , 1) ∈ [k]d
and S = {x ∈ [k]d : xmin = 1}. We can partition [k]d into S = kd − (k − 1)d ≤ dkd−1 chains
(x, x + 1, x + 2 · 1, . . . , x + (xmax − 1) · 1), one for each x ∈ S. Therefore, by Dilworth's theorem,
every anti-chain on P has size at most dkd−1. In particular, this bound holds for the anti-chain
g−1(∗) so d(f, g) ≤ dkd−1
kd = d
k ≤ ǫ.
4 Proof of the weak regularity lemma
4.1
Information theory
The proof we provide for Lemma 1 is information-theoretic. In this section, we will use bold fonts
to denote random variables. We assume that the reader is familiar with the basic concepts of
entropy and mutual information; a good introduction to these definitions is [9]. The only facts we
use about these concepts is the chain rule for mutual information and the fact that the entropy of
a random variable is at most the logarithm of the number of values it can take.
The one non-basic information-theoretic inequality that we use in the proof is an inequality
established by Tao [25] and later refined by Ahlswede [1].
10
Lemma 5 (Tao [25], Ahlswede [1]). Let y, z, and z′ be discrete random variables where y ∈ [−1, 1]
and z′ = φ(z) for some function φ. Then
Eh(cid:12)(cid:12) E[y z′] − E[y z](cid:12)(cid:12)i ≤ p2 ln 2 · I(y; z z′).
Tao originally used his inequality to offer an information-theoretic proof of Szemer´edi's (strong)
Regularity Lemma. The proof we offer below follows (a simplified version of) the same approach.
The fact that Tao's proof of the strong regularity lemma can also be applied (with simplifications)
to prove the Frieze–Kannan weak regularity lemma was observed previously by Trevisan [26].
4.2 Proof of Lemma 1
For τ > 0, we say that a hypergraph is τ -granular if the weight of each hyperedge is a multiple of
τ . When proving Lemma 1, we can assume that the given graph is τ -granular for τ = Θ(ǫ). To see
this, let G′ = (V, E′) be the hypergraph obtained from G = (V, E) by rounding the weight of each
hyperedge to a multiple of τ . Then, for any set S ⊆ V and a partition V1, . . . , Vk of V , we have
Qi∈I S ∩ Vi
V s
X
I∈[k]s
wG(S ∩ VI ) − wG(VI ) = X
= X
I∈[k]s
I∈[k]s
V s
Qi∈I S ∩ Vi
Qi∈I S ∩ Vi
V s
(wG′(S ∩ VI ) ± τ ) − (wG′(VI ) ± τ )
wG′(S ∩ VI ) − wG′(VI ) ± 2τ.
In order to make the right-hand side less than ǫ, it suffices to show that the claim of Lemma 1
holds for ǫ/3-granular hypergraphs with an error parameter ǫ/3. In what follows, we assume the
input graph G is ǫ/3-granular.
Let V1, . . . , Vℓ be any partition of V . Draw v ∈ V s uniformly at random. Define y = E(v).
Let ψ : V s → [ℓ]s be the function that identifies the parts containing each of s vertices. For any
set S ⊆ V , let 1S : V s → {0, 1}s be the indicator function of S for s-tuples of vertices. Define
zS = (ψ(v), 1S (v)) and z′ = ψ(v). We consider two cases.
First, consider the situation where for every set S ⊆ V ,
I(y; zS z′) ≤
(ǫ/3)2
2 ln 2
.
(4)
Then by Tao's lemma, for every set S we also have
X
I∈[k]s,b∈{0,1}s
Qj∈[s] Sbj ∩ Vij
V s
wG(Sb ∩ VI ) − wG(VI ) = E[E(v) ψ(v)] − E[E(v) ψ(v), 1S (v)]
= E[y z′] − E[y zS] ≤ ǫ/3.
And clearly
Qj∈[s] S ∩ Vij
V s
X
I∈[k]s
wG(S ∩ VI) − wG(VI ) ≤ X
I∈[k]s,b∈{0,1}s
Qj∈[s] Sbj ∩ Vij
V s
wG(Sb ∩ VI ) − wG(VI )
since the expression in the left-hand side is one of the terms in the sum (over b) on the right-hand
side. So in this case the lemma holds.
11
Second, we need to consider the case where there is some set S ⊆ V for which
I(y; zS z′) >
(ǫ/3)2
2 ln 2
.
(5)
Then by the chain rule for mutual information
I(y; zS ) = I(y; z′) + I(y; zS z′) ≥ I(y; z′) +
(ǫ/3)2
2 ln 2
.
Define the information value of a partition with indicator function ψ as I(E(v); ψ(v)). Then the
above observation shows that when (5) holds, we can obtain a refined partition with 2ℓ parts whose
information value increases by at least (ǫ/3)2
2 ln 2 . The information value of any partition is bounded
above by H(y), so after at most 2 ln 2·H(y)
refinements, we must obtain a partition (with at most
(ǫ/3)2
2 ln 2·H(y)
(ǫ/3)2
2
lemma follows.
parts) that satisfies (4). Since G is ǫ/3-granular, we have H(y) = O(log(1/ǫ)), and the
References
[1] Rudolf Ahlswede. The final form of Tao's inequality relating conditional expectation and
conditional mutual information. Advances in Mathematics of Communications, 1(2):239–242,
2007.
[2] Noga Alon, Eldar Fischer, Ilan Newman, and Asaf Shapira. A combinatorial characterization
It's all about regularity. SIAM Journal on Computing,
of the testable graph properties:
39(1):143–167, 2009.
[3] Noga Alon, Rani Hod, and Amit Weinstein. On active and passive testing. Combinatorics,
Probability and Computing, 25(1):1–20, 2016.
[4] Maria-Florina Balcan, Eric Blais, Avrim Blum, and Liu Yang. Active property testing. Pro-
ceedings of the 53rd Annual IEEE Symposium on Foundations of Computer Science (FOCS),
pages 21–30, 2012.
[5] Piotr Berman, Meiram Murzabulatov, and Sofya Raskhodnikova. Tolerant testers of image
properties. In Proceedings of the 43rd International Colloquium on Automata, Languages, and
Programming (ICALP), pages 90:1–90:14, 2016.
[6] Arnab Bhattacharyya and Yuichi Yoshida. An algebraic characterization of testable Boolean
In Proceedings of the 40th International Colloquium Conference on Automata, Lan-
CSPs.
guages, and Programming (ICALP), pages 123–134, 2013.
[7] Eric Blais, Amit Weinstein, and Yuichi Yoshida. Partially symmetric functions are efficiently
isomorphism testable. SIAM Journal on Computing, 44(2):411–432, 2015.
[8] Xi Chen, Rocco A Servedio, and Li-Yang Tan. New algorithms and lower bounds for monotonic-
ity testing. In Proceedings of the 55th Annual IEEE Symposium on Foundations of Computer
Science (FOCS), pages 286–295, 2014.
12
[9] Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley & Sons,
2012.
[10] Ilias Diakonikolas, Homin K. Lee, Kevin Matulef, Krzysztof Onak, Ronitt Rubinfeld, Rocco A.
Servedio, and Andrew Wan. Testing for concise representations. In Proc. 48th Symposium on
Foundations of Computer Science, pages 549–558, 2007.
[11] Eldar Fischer, Yonatan Goldhirsh, and Oded Lachish. Partial tests, universal tests and de-
composability. In Proceedings of the 5th Conference on Innovations in Theoretical Computer
Science (ITCS), pages 483–500, 2014.
[12] Eldar Fischer, Guy Kindler, Dana Ron, Shmuel Safra, and Alex Samorodnitsky. Testing juntas.
J. Comput. Syst. Sci., 68(4):753–787, 2004.
[13] Eldar Fischer, Oded Lachish, and Yadu Vasudev. Trading query complexity for sample-based
testing and multi-testing scalability. In Proceedings of the 56th Annual IEEE Symposium on
Foundations of Computer Science (FOCS), pages 1163–1182, 2015.
[14] Eldar Fischer, Eric Lehman, Ilan Newman, Sofya Raskhodnikova, Ronitt Rubinfeld, and Alex
Samorodnitsky. Monotonicity testing over general poset domains. In Proceedings of the 34th
Annual ACM Symposium on Theory of Computing (STOC), pages 474–483, 2002.
[15] Alan M. Frieze and Ravi Kannan. The regularity lemma and approximation schemes for dense
problems. In Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer
Science (FOCS), pages 12–20, 1996.
[16] Oded Goldreich, Shafi Goldwasser, Eric Lehman, Dana Ron, and Alex Samorodnitsky. Testing
monotonicity. Combinatorica, 20(3):301–337, 2000.
[17] Oded Goldreich, Shafi Goldwasser, and Dana Ron. Property testing and its connection to
learning and approximation. Journal of the ACM, 45(4):653–750, 1998.
[18] Oded Goldreich and Dana Ron. On sample-based testers. In Proceedings of the 6th Conference
on Innovations in Theoretical Computer Science (ITCS), pages 337–345, 2015.
[19] Tali Kaufman and Madhu Sudan. Algebraic property testing: the role of invariance.
In
Proceedings of the 40th Annual ACM Symposium on Theory of Computing (STOC), pages
403–412, 2008.
[20] Michael Kearns and Dana Ron. Testing problems with sublearning sample complexity. Journal
of Computer and System Sciences, 61(3):428–456, 2000.
[21] Sofya Raskhodnikova. Approximate testing of visual properties.
In Proceedings of the 7th
International Workshop on Randomization and Computation (RANDOM), pages 370–381,
2003.
[22] Ronitt Rubinfeld and Madhu Sudan. Robust characterizations of polynomials with applications
to program testing. SIAM Journal on Computing, 25(2):252–271, 1996.
[23] Claude Shannon. The synthesis of two-terminal switching circuits. Bell System Technical
Journal, 28(1):59–98, 1949.
13
[24] Madhu Sudan. Invariance in property testing. In Oded Goldreich, editor, Property Testing:
Current Research and Surveys, pages 211–227. Springer, 2010.
[25] Terence Tao. Szemer´edi's regularity lemma revisited. Contributions to Discrete Mathematics,
1(1), 2006.
[26] Luca
Trevisan.
Entropy
and
the
weak
regularity
lemma.
https://lucatrevisan.wordpress.com/2006/09/10/entropy-and-the-weak-regularity-lemma/,
2006.
[27] Yuichi Yoshida. A characterization of locally testable affine-invariant properties via decompo-
sition theorems. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing
(STOC), pages 154–163, 2014.
14
|
1204.1467 | 1 | 1204 | 2012-04-06T12:42:24 | Learning Fuzzy {\beta}-Certain and {\beta}-Possible rules from incomplete quantitative data by rough sets | [
"cs.DS",
"cs.LG"
] | The rough-set theory proposed by Pawlak, has been widely used in dealing with data classification problems. The original rough-set model is, however, quite sensitive to noisy data. Tzung thus proposed deals with the problem of producing a set of fuzzy certain and fuzzy possible rules from quantitative data with a predefined tolerance degree of uncertainty and misclassification. This model allowed, which combines the variable precision rough-set model and the fuzzy set theory, is thus proposed to solve this problem. This paper thus deals with the problem of producing a set of fuzzy certain and fuzzy possible rules from incomplete quantitative data with a predefined tolerance degree of uncertainty and misclassification. A new method, incomplete quantitative data for rough-set model and the fuzzy set theory, is thus proposed to solve this problem. It first transforms each quantitative value into a fuzzy set of linguistic terms using membership functions and then finding incomplete quantitative data with lower and the fuzzy upper approximations. It second calculates the fuzzy {\beta}-lower and the fuzzy {\beta}-upper approximations. The certain and possible rules are then generated based on these fuzzy approximations. These rules can then be used to classify unknown objects. | cs.DS | cs | Mining Fuzzy β-Certain and β-Possible rules from incomplete quantitative data by rough sets
A.S,Mohammadi
Department of Information
Technology, Payame Noor University,
PO BOX 19395-3697 Tehran, IRAN
Ali_s [email protected]
L. Asadzadeh
Department of Information
Technology, Payame Noor University,
PO BOX 19395-3697 Tehran, IRAN
[email protected]
Abstract—The rough-set theory proposed by
Pawlak, has been widely used in dealing with data
classification problems. The original rough-set
model is, however, quite sensitive to noisy data.
Tzung thus proposed deals with the problem of
producing a set of fuzzy certain and fuzzy possible
rules from quantitative data with a predefined
uncertainty
tolerance
degree
of
and
misclassification. This model allowed
, which
combines the variable precision rough-set model
and the fuzzy set theory, is thus proposed to solve
this problem. This paper thus deals with the
problem of producing a set of fuzzy certain and
fuzzy possible rules from incomplete quantitative
data with a predefined tolerance degree of
uncertainty and misclassification. A new method,
incomplete quantitative data for rough-set model
and the fuzzy set theory, is thus proposed to solve
this problem. It first transforms each quantitative
value into a fuzzy set of linguistic terms using
membership
functions
and
then
finding
incomplete quantitative data with lower and the
fuzzy upper approximations. It second calculates
the
fuzzy β-lower and
the
fuzzy β-upper
approximations. The certain and possible rules
are
then generated based on
these
fuzzy
approximations. These rules can then be used to
classify unknown objects.
Keywords-component: Fuzzy set; Rough set; Data
mining; Certain rule; Possible rule; Quantitative
value; incomplete data
I.
Introduction
Machine learning and data mining techniques have
recently been developed to find implicitly meaningful
the knowledge-acquisition
pat-terns
and
ease
bottleneck. Among
these appr oaches, deriving
inference or association rules from training examples
is the most common [9,12,13]. Given a set of
examples and counterexamples of a concept, the
learning program tries to induce general rules that
describe all or most of the positive training instances
and none or few of the coun terexamples [6,7]. If the
training instances belong to more than two classes,
the learning program tries to induce general rules that
D.D,Rezaee
Department of Information
Technology, Payame Noor University,
PO BOX 19395-3697 Tehran, IRAN
[email protected]
describe each class. Recently, the rough-set theory
has been used
reasoning and knowledge
in
acquisition for expert systems [4,14]. It was proposed
by Pawlak in 1982 [19], with the concept of
equivalence classes as its basic principle. Several
applications and extensions of the rough-set theory
have also been proposed. Examples are [14,18]
reasoning
with
incomplete
information,[2]
knowledge-base reduction, [10] data mining, [16]
rule discovery. Due to the success of the rough-set
theory to knowledge acquisition, many researchers in
database and machine learning fields are interested in
this new research topic because it offers opportunities
to discover useful information in training examples.
[6,17] mentioned that the main issue in the rough-set
approach was the formation of good rules.
II. Review of the variable precision rough-set
model
The rough-set theory, proposed by Pawlak in 1982,
can serve as a new mathematical tool to deal with
data classification problems [19]. It adopt s the
concept of equivalence classes to partition the
training instances according to some criteria. Two
kinds of partitions are formed in the mining process:
lower approximations and upper approximations,
from which certain and possible rules are easily
derived. Let X be an arbitrary subset of the universe
U, and B be an arbitrary subset of the attribute set A.
The lower and the upper approximations for B on X,
denoted B* (X) and B*
(X) respectively, are defined
B*
(X) = {xx ϵ U, B(X)⊆ X} ,
as follows:
(X) = {xx ϵ U and B(X) ∩ X ≠Ø} .
(1)
B
*
(2)
Elements in B* (x) can be classified using attribute
set B as members of the set X with full certainty, and
B * (x) is thus called a lower approximation of X. On
the other hand, elements in B (x) can be classified
using attribute set B as members of the set X only
with partial certainty, and B* (x) is thus called an
uppe r approximation of X. The origina l rough-set
model is quite sensitive to noisy data [7]. When noisy
data exists, the lower and the upper approximations
cannot normally be formed [17]. Let two sets X and
Y be non-empty subsets of the un iversal set (X, Y,
U). The rough inclusion function is then defined as
follows:
(3)
Y)
=
(X,
µ
card(X
Y)
∩
X
card
(
)
If the rough inclusion value equals to 1, then X is
totally included in Y (X,Y). Otherwise, the rough
inclusion value ranges between 0 to 1, and X is said
partially included in Y. Also, the relative degree of
misclassification of the set X with respect to set Y is
defined as :
(4)
c
(X,
(5)
(6)
1 Y)
−=
card(X
Y)
∩
X
card
(
)
Based on the relative degree of misclassification, in
the[17],
generalized
the
lower
and
upper
approximations of the original rough-set model with
a majority inclusion threshold β. The β-lower and the
β-upper approximations are defined as follows:
B*
(X) = {xx ϵ U , c(B(x),X) ≤ β }
(X) = {xx ϵ U , c(B(x),X) <1- β } .
B *β
β
III.
Incomplete data sets
Data sets can be rough ly classified into two
classes: complete and incomplete data sets. All the
objects in a complete data set have known attribute
values. If at least one object in a data set has a
missing value, the data set is incomplete. the symbol
‘*’ denotes an unknown attribute value. Learning
rules from incomplete data sets is usually more
difficult than from complete data sets. In the past,
several methods were proposed
the
to handle
problem of incomplete data sets [21]. For example,
incomplete data sets may be transformed into
complete data sets by similarity measures or by
removing obj ects with unknown values before
learning programs begin. Incomplete data sets may
also be directly processed in a particular way to get
the rules [20,22]. In the [20] proposed a rough-set
approach to directly learn rules from incomplete data
sets without guessing unknown attribute values. In
the [22] modified [20] approach by introducing the
rough entropy to distinguish the power of the
attribute subsets that have the same partition for
similarity relations.
IV. Notation
Notation used in this paper is described as follows :
U universe of all objects
β tolerance degree of noise and misclassification
n total number of training examples (objects) in U
Obj (i)
ith training example (object), 1 ≤i ≤n
A set of all attributes describing U
m total number of attributes in A
B an arbitrary subset of A
A j
j th attribute, 1 ≤ j≤ m
A
j
j number of fuzzy regions for A
jk k th fuzzy region of A j ,1 ≤k ≤Aj
R
(i) quantitative value of A j for Obj
(i)
v
j
(i)
fuzzy set converted from v
f
j
(i)
membership value of v in region R
f
jk
C set of classes to be determined
c total number of classes in C
l th class, 1 ≤ l≤ c
X l
(i)) the fuzzy incomplete equivalence classes in
B(Obj
which Obj(i)
exists
c(Obj (i)) the certain part of the kth fuzzy incomplete
B
k
equivalence class in B(Obj(i)
)
B
(X) the fuzzy incomplete lower approximation for
*
B on X
*
B
(X) the fuzzy incomplete upper approximation for
B on X
When the same linguistic term RRjkR of an attribute ARjR
( i)
(r )
P and ObjP
P with
exists in two fuzzy objects ObjP
( i)
(r)
P larger than zero,
P and fRjk RP
membership values fRjkRP
( i)
(r)
to have a fuzzy
P are said
and ObjP
P
ObjP
indiscernibility
relation
(or
fuzzy equivalence
relation) on attribute ARj R with a membership value
(i)
(r)
P ∩ fRjkRP
P ). Also, if the same
equal to min (fRjkRP
linguistic terms of an attribute subset B exist in both
( i)
(r)
P with membership values larger than
Pand ObjP
ObjP
( i)
(r )
P and ObjP
P are said to have a fuzzy
zero, Obj P
indiscernibility relation (or a fuzzy equivalence
relation) on attribute subset B with a membership
value equal to the minimum of all the membership
values
.These fuzzy equivalence relations
thus
partition the fuzzy object set U into several fuzzy
subsets that may overlap, and the result is denoted by
U/B. The set of fuzzy partitions, based on B and
( i)
( i)
P). Thus,
is denoted B(ObjP
P
,
including ObjP
( i)
(i)
( i)
(r)
P)= {((BR1 R(ObjP
P, µ RB1 R(ObjP
P)) … ((BRrR(ObjP
P, µ
B(ObjP
(r)
RBrR(ObjP
P)) , where r is the number of partitions
(i)
( i)
included in B(ObjP
P) , BRjR(ObjP
P) is the jth partition in
( i)
( r)
P) ,and µ RB jR(ObjP
P) is the membership value of
B(ObjP
the jth partition.
Since an incomplete quantitative data set contains
( i)
P is thus
unknown attribute values, each object ObjP
(i)
P ,symbol), where the
represented as a tuple (ObjP
symbol may be certain (c) or uncertain (u). If an
(i)
P has an uncertain value for attribute ARjR,
object ObjP
( i)
P ,u) is put in each fuzzy equivalence class
then (ObjP
of attribute ARjR. The fuzzy sets formed in this way are
called fuzzy incomplete equi valence classes, which
are not necessarily equi valence classes.The fuzzy
incomplete lower and upper approximations for B on
*
X, denoted BR* R(X) and BP
P(X) respectively, are
P ϵ XRlR ,
defined as follows:
B RkRPc P(ObjP(i)P)⊆ Xl , 1≤k≤B(ObjP(i)P)}. (7)
( i)
( i)
( i)
P)) 1≤i ≤n , objP
P),µRBkR(ObjP
BR* R(X Rl R) = {(BRk R(ObjP
P)) 1≤i ≤n ,
B RkRPc P(ObjP(i)P)∩ Xl ≠ Ø , B RkRPc P(ObjP(i)P)⊄ X Rl R
(i)
*
( i)
P),µRBkR(ObjP
P(X Rl R) = {(BRk R(ObjP
BP
1≤k≤B(ObjP(i)P}. (8)
P)⊄ X to
Here, the definition of the fuzzy incompl ete upp er
c
( i)
P(ObjP
approximation has the constraint BRk RP
exclude the objects in the fuzzy incomplete lower
approximation for avoiding redundant calculation.
The
fu zzy β-lo wer and
the
fu zzy β-upper
appr oximations for B on X, denoted BR*β R(X) and
*
PRβ R(X) respectively, are defined as follows:
BP
BR*β R(X Rl R) = {(BRk R(x),µRBk R(x)) x ϵ U , c(B RkR(x), X)≤ β ,
1≤k≤B(x)} (9)
PRβ R(X Rl R) = {(BRk R(x),µRBk R(x)) x ϵ U , β< c(B RkR(x),
X)≤1- β , 1≤k≤B(x)}.
*
BP
(10)
After the fu zzy β-lower and the fu zzy β-upper
approximations are
found, both β-certain and
β-uncertain rules can thus be derived.
V. The proposed algorithm for incomplete
quantitative data sets
In the section, a learning algorithm based on rough
sets is proposed, which can simultaneously estimate
the missing values and derive fuzzy certain and
possible rules from incomplete quantitative data sets.
The proposed
fuzzy
learning algorithm
first
transforms each quantitative value into a fuzzy set of
linguistic terms using membership functions. The
details of the proposed fuzzy learning algorithm are
described as follows.
The fuzzy mining algorithm for β-certain and β-
possible rules:
Input: A incomplete quantitative data set with n
objects ,each with m attribute values, β is tolerance
degree of noise and misclassification ,and a set of
membership functions.
Output: A set of maximally general fuzzy β-certain
and fuzzy β-possible rules.
Step 1: Partition the object set into disjoint subsets
according to class labels. Denote each set of objects
belonging to the same class CRl R as XRl R.
(i
)
Step 2: Transform the quantitative value vRj RP
P of each
(i)
P ;i =1 to n, for each attribute ARj R; j = 1 to
object ObjP
(i)
P , represented as,
m , into a fuzzy set fRj RP
i
i
i
)(
)(
)(
f
j
1
R
j
1
f
j
2
R
j
f
jl
R
(11)
... ++
+
jl
2
using the given membership functions, where RRjkR is
( i)
( i)
the kth fuzzy region of attribute ARjR ; fRjkRP
P’s
P is vRjRP
fuzzy membership value in region RRjk R, and l (= ARjR)
(i)
is the number of fuzzy regions for ARj R. If ObjP
P has a
missing value for ARjR, keep it with a missing value (*).
Step 3: Find the fuzzy incomplete elementary sets of
(i)
P has a
singleton attributes; That is, if an object ObjP
( i)
P for attribute ARj R,
certain fuzzy membership value fRjkRP
(i)
P ,c) into the fuzzy incomplete equivalence
put (ObjP
( i)
class from ARjR = RRjk R; If ObjP
P has a missing value for
(i)
ARj R, put (ObjP
P
,u) into each fuzzy incomplete
equivalence class from ARjR; The membership value
µRAjk R of a fuzzy incomplete class for ARjR=RRjkR is
calculated as:
f (i)
Min
=µ
jkA
jk
(i)
( i)
P≠ 0.
P is certain and fRjkRP
where ObjP
Step 4: Initialize q = 1, where q is used to count the
number of attributes currently being processed for
fuzzy incomplete lower approximations.
(12)
Step 5: Compute
lower
incomplete
the fuzzy
approximations of each subset B with q attributes for
each class XRl R as:
P ϵ XRlR ,
B RkRPc P(ObjP(i)P)⊆ Xl , 1≤k≤B(ObjP(i)P)}. (13)
( i)
( i)
( i)
P)) 1≤i ≤n , objP
P),µRBkR(ObjP
BR* R(X Rl R) = {(BRk R(ObjP
( i)
P) is the set of fuzzy incomplete
where B(ObjP
( i)
P and derived from
equivalence classes including ObjP
c
( i)
P) is the certain part of the
P(ObjP
attribute subset B, BRk RP
(i)
P) .
kth fuzzy incomplete equivalence class in B(ObjP
r
)(
obj
(14)
r
)(
jk
(i)
in fuzzy
incomplete
P exists
Step 6:
If ObjP
c
(i)
equivalence class BRkRP
the kth region
P) of
P(ObjP
k
P from attribute subset B in a fuzzy
combination RRBRP
incomplete lower approximation, assign the uncertain
( i)
P as:
value of ObjP
∑
r
)(
v
f
.
j
c
i
)(
obj
B
(
)
∈
k
∑
r
)(
f
jk
c
i
)(
obj
B
(
∈
k
(r)
(r)
P is the quantitative value of ObjP
where vRjRP
P for
(r)
(r)
P’s fuzzy membership value
attribute ARjR and fRjk RP
P is vRjRP
k
( i)
in RRBRP
P . Also transform the estimated ObjP
P value into
( i)
P ,u)’s with membership
a fuzzy set, remove (ObjP
values equal to zero from the fuzzy incomplete
(i)
,u)’s with
P
equivalence classes, change
(ObjP
( i)
P ,c)’s
membership values not equal to zero into (ObjP
and re-calculate the membership values of the fuzzy
incomplete equivalence classes including them by the
obj
r
)(
)
P
P
the
to
minimum operation. Besides, backtrack
lower
previously
incomplete
fuzzy
found
approximations for doing the same actions on Obj(i)
.
Step 7: Set q = q+1 and repeat Steps 5–7 unt il q > m.
Step 8: Initialize h = 1, where h is used to count the
number of attributes currently being processed for
fuzzy incomplete upper approximations.
incomplete upper
the fuzzy
Step 9: Compute
approximations of each subset B with q attributes for
each class X l
B RkRPc P(ObjP(i)P)∩ Xl ≠ Ø
, B RkRPc P(ObjP(i)P)⊄ XRlR
(i)
( i)
*
,
P)) 1≤i ≤n
P(X Rl R) = {(BRkR(ObjP
BP
P),µRBkR(ObjP
1≤k≤B(ObjP(i)P}. (15)
( i)
P) is the set of fuzzy incomplete
where B(ObjP
( i)
equivalence classes including ObjP
P and derived from
c
( i)
P) is the certain part of the
P(ObjP
attribute subset B, BRk RP
(i)
P) .
kth fuzzy incomplete equivalence class in B(ObjP
Step 10: Do the following sub steps for each
(i)
P in the fuzzy incomplete
uncertain instance ObjP
upper approximations:
l
(
(
X
X
∈
r
)(
r
)(
r
)(
obj
obj
c
B
∈
k
( i)
P exists in fuzzy incomplete equivalence
a) If ObjP
c
(i)
k
P(ObjP
P) of the kth region combination RRB RP
P
class BRk RP
from attribute subset B in a fuzzy incomplete lower
( i)
P as:
approximation, assign the uncertain value of ObjP
∑
(16)
r
r
)(
)(
v
f
.
j
jk
r
i
)(
)(
obj
obj
&)
∈
∑
r
)(
f
jk
i
)(
obj
obj
&)
c
B
∈
l
k
(r)
(r)
P is the quantitative value of ObjP
P for
where vRjRP
(r)
(r)
attribute ARjR and fRjk RP
P’s fuzzy membership value
P is vRjRP
k
( i)
P . Also transform the estimated ObjP
P value into
in RRBRP
( i)
P ,u)’s with membership
a fuzzy set, remove (ObjP
values equal to zero from the fuzzy incomplete
(i)
P
,u)’s with
equivalence classes, change
(ObjP
( i)
P ,c)’s
membership values not equal to zero into (ObjP
and re-calculate the membership values of the fuzzy
incomplete equivalence classes including them by the
minimum operation. Besides, backtrack
to
the
lower
previously
incomplete
fuzzy
found
(i)
P .
approximations for doing the same actions on ObjP
( i)
P still exists in more than one
b) If an object ObjP
fuzzy incomplete equivalence class in a fuzzy
incomplete upper approximation, use the equivalence
class with the maximum plausibility measure to
( i)
P . The estimation
estimate the uncertain value of ObjP
and processing are the same as those stated in Step
10(a). Calculate the plausibility measures of each
fuzzy incomplete equivalence class in an upper
approximation for each class XRl R as:
(17)
r
)(
µ
jk
(
=
r
)(
r
)(
))
i
)(
&)
obj
obj
c
B
∈
k
X
∈
l
r
)(
µ
jk
c
k obj
BP
(
(
∑
i
)(
obj
∑
c
r
i
)(
)(
obj
obj
B
(
)
⊂
k
( i)
c) If an object ObjP
P exists in more than one fuzzy
incomplete equivalence class in a fuzzy incomplete
upper approximation and them plausibility measure
together are equal, define for determine uncertain
( i)
Pin it equivalence class that number
value of ObjP
certain objects that more of classes other. If number
certain objects classes are equa l, hence define that
class than are include least label.
Step 11: Set h=h+1 and repeat Steps 9–11until h > m.
Step 12 : Initialize l= 1, where l is used to count the
number of classes being processed.
Step 13: Calculate
relative degree of
the
misclassification of each attribute subset BRkR for class
XRl R as:
l
(18)
XxBc
(
(
),
k
∑
y
)(
µ
B
k
X
xBy
)
(
)(
∩
∈−=
l
k
1)
∑
y
)(
µ
B
k
xBy
)(
∈
k
Step 14: Calculate the modified fuzzy β-lower and β-
uppe r approximation of each attribute subset B for
Step 15: Derive the β-certain rules from the fuzzy β-
class XRl R as equation “(9)” and “(10)”.
lower approximation BR*β R(XRl R) of any subset B, set the
membership values in the β-lower approximation as
the effectiveness for future data, and calculate the
plausibility measure of each rule for BRkR(x) as :
(19)
XxBc−
(
1
(
),
)
k
l
Step 16: Remove the β-certain rules with the
condition parts more specific and effectiveness
measure equa l to or smaller than those of some other
β-certain rules.
Step 17: Derive the β-possible rules from the fuzzy β-
*
PRβ R(X ) of any subset B, set the
uppe r approximation BP
membership values in the β-upper approximation as
the effectiveness for future data, and calculate the
plausibility measure of each rule for BRkR(x) as
equa tion “(19)” .
Step 18: Remove the β-possible rules with the
condition parts more
the
specific and both
effectiveness and plausibility equal to or smaller than
those of some other β-certain or β-possible rules.
Step 19: Set l=l+1and Repeat Steps 13 to19 until l≤ c.
Step 20 : Output the β-certain and β-possible fuzzy
rules .
VI. an example
In this section, an example is given to show how
the proposed algorithm can be used to generate
maximally general fuzzy β-certain and fuzzy β-
as:
possible rules from incomplete quantitative data.
Table 1 shows an incomplete quantitative data set.
Assume the membership functions for each attribute
are given by experts as shown in Fig. 1. The proposed
learning
algorithm processes
this
incomplete
quantitative data set as follows.
(1)
ObjP
(2)
ObjP
(3)
ObjP
(4)
ObjP
(5)
ObjP
(6)
ObjP
(7)
ObjP
Table 1 : T he quantita tive data set a s an Example
Systo lic
Diasto lic
Blood
Object
pressure
pressure
pressure
(BP)
(SP)
(DP)
122
80
N
H
90
155
130
92
N
87
68
L
*
93
H
H
100
150
*
95
L
Step 1: Since three classes exist in the data set, three
partitions are thus formed as follows:
)4(
)7(
X
X
X
L
N
H
L
L
=
obj
{
obj
{
=
obj
{
=
)1(
obj
,
obj
,
)2(
obj
,
)3(
},
},
)5(
obj
,
N
)6(
}.
H
140
100
90
Systolic Pressure (SP)
N
160
H
75
70
Diastolic Pressure (DP)
90
95
Fig. 1. The given membe rship functions of each attribute.
Step 2: The quantitative values of each object are
transformed into fuzzy sets. Take the attribute
(2)
P as an example. The
Systolic Pressure (SP) in ObjP
value ‘‘155’’ is converted into a fuzzy set (0.1/N +
0.75/H) using the given membership functions.
Results for all the objects are shown in Table 2.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Obj P
Obj P
Obj P
Obj P
Obj P
Obj P
Obj P
0.9/N
0.1/N+0.75/H
0.85/N
1/L
*
0.3/N+0.5/H
0.5/L+0.1/N
Table 2 : T he fu zzy set transformed from the data in Table 1
Systo lic
Diasto lic
Blood
Object
pressure
pressure (DP)
pressure (SP)
(BP)
N
H
N
L
H
H
L
0.9/N
0.4/N
0.3/N+0.4/H
1/L
0.16/N+0.6/H
1/H
*
Step 3: The elementary fuzzy sets of the singleton
attributes SP and DP are found as follows:
(1)
(2)
(3)
U/{SP}
{{(obj
c)(obj
c)(obj
=
,
,
c),
(5)
(7)
(6)
(obj
c)(obj
c)(obj
,
,
,
u),0.1}
(5)
(6)
(2)
{(obj
c)(obj
c)(obj
u),0.5}
,
,
,
(5)
(7)
(4)
{(obj
c)(obj
c)(obj
,
u),0.5}}
,
,
and
(2)
(1)
{{(obj
c)(obj
U/{DP}
c),
,
=
(5)
(7)
(3)
(obj
c)(obj
u)(obj
c),0.16}
,
,
,
(7)
(6)
(5)
(3)
{(obj
c)(obj
c)(obj
c)(obj
,
,
,
(7)
(4)
{(obj
c)(obj
,
u),0.5}}
,
,
u),0.4}
Step 4: q is initially set at 1, where q is used to count
the number of the attributes currently being processed
for fuzzy incomplete lower approximations.
Step 5: The fuzzy incomplete lower approximation of
(2)
(5)
(6)
one attribute for XRHR={ObjP
P , ObjP
P , ObjP
P } is first
(2)
P,c) of the
calculated. Since only the certain part (ObjP
first incomplete equivalence class for attribute SP is
(5)
included in XRH Rand the uncertain instances ObjP
Pbelong to XRHR, thus:
(XSP
H*
{(obj
= )
(2)
(6)
c)(obj
,
(5)
c)(obj
,
,
u),0.5}
Since the certain part of each fuzzy incomplete
equivalence class for attribute DP is not included in
XRH R, thus:
(XDP
H*
φ = )
Similarly, the fuzzy incomplete lower approximations
of single attributes for XRN R and XRLR are found as
follows:
= )
(XSP
φ
N*
= )
(XDP
φ
N*
(4)
(XSP
{(obj
= )
*
L
(XDP
= )
{(obj
*
L
(7)
c)(obj
c),0.5}
,
,
(4)
(7)
c)(obj
,
,
u),1}
Step 6: Each uncertain object in the above fuzzy
incomplete lower approximations is checked for
change to certain objects. For example in SP * (X H ),
since Obj(5) exist in only one fuzzy incomplete
equivalence class of SP = H, their values can then be
estimated from the certain objects in the same
equivalence class. Since only one certain object Obj(2)
and Obj(6) exists in the fuzzy incomplete equivalence
class of SP = H, the estimated value of Obj(5) is
then (155*0.75+150*0.5)/1.25 (=153), where 153 is
the quantitative value of Obj(2) and Obj(6) for attribute
SP and 0.75 , 0.5 are there fuzzy membership values
for the region of SP = H. The estimated values of
Obj(5)
fuzzy set
the
transformed as
then
is
(0.2/N+0.65/H) . (Obj(5) ,u) is then changed as
(Obj(5)
,c).
The modified SP
* (X H
) and The fuzzy incomplete
elementary set of attribute SP are then:
(2)
(5)
(6)
c)(obj
c)(obj
,
c),0.5}
,
,
{(obj
= )
(XSP
H
*
(3)
(2)
(1)
U/{SP}
c)(obj
c)(obj
{{(obj
,
,
c),
=
(6)
(5)
(7)
(obj
c)(obj
c)(obj
,
c),0.1}
,
,
(2)
(6)
(5)
c)(obj
c)(obj
{(obj
,
,
c),0.5}
,
(7)
c)(obj
{(obj
,
,
c),0.5}}
(4)
Similarly , Since the uncertain object Obj(7) in
DP * (X L ) exist
in only
the
fuzzy
incomplete
equivalence class of DP = L, the estimated value of
Obj(7) is then (68*1)/1 (=68)), where 68 is the
quantitative value of Obj(4) for attribute DP and 1 is
its fuzzy membership value of DP = L. The estimated
value of Obj(7) is then transformed as the fuzzy set
(1/L) for attribute DP. Also, (Obj(7) ,u) is then
changed as (Obj(7) ,c). The modified DP * (X H
) and
The fuzzy incomplete elementary set of attribute DP
are then:
(4)
(7)
c)(obj
c),1}
,
,
{(obj
= )
(XDP
L
*
(1)
(2)
U/{DP}
=
c)(obj
{{(obj
,
c),
(5)
(3)
(obj
c)(obj
,
c),0.16}
,
(6)
(5)
(3)
c)(obj
c)(obj
{(obj
,
,
(7)
c)(obj
{(obj
,
c),0.5}}
,
c),0.4}
(4)
,
Step 7: q = q+1, and Steps 5–7 are repeated. Until the
fuzzy incomplete elementary set of attributes {SP,
DP} is found as follows:
(1)
(2)
c)(obj
c),
U/{SP,
(3)
,
(obj
(3)
{(obj
{(obj
{(obj
{(obj
(5)
(4)
(2)
{{(obj
,
= DP}
(5)
c)(obj
c),0.1}
,
(5)
(6)
c)(obj
c)(obj
,
,
(5)
c)(obj
,
,
c),0.16}
(7)
c)(obj
,
,
c),0.5}{(o
bj
(6)
c)(obj
,
,
c),0.5}}
,
c),0.2}
(7)
,
c),0.1}
Because , It is requirement for next steps .
Step 8: Since all objects in the fuzzy incomplete
lower approximations have become certain, go to the
step12 is executed.
Step 12 : Initialize l=l+1, where l is used to count the
number of classes then beg in find in g β-certain
and β-possible rules processed.
with the relative degree β = 0.2 of misclassification
can be calculated as follows:
for XN
:
=
−
+
=
31.0
SP,
c(
=
28.0
DP c(
N
85.09.0
+
1= )X(x),
SP c(
N
N
85.01.03.02.01.09.0
+
+
+
+
+
1 0-1 )X(x),
SP c(
=
=
N
H
1 0-1 )X(x),
SP c(
=
=
N
L
90.
0.3
+
-1 )X(x),
=
N
0.16
0.9
0.3
0.4
+
+
40.
8.0
-1 )X(x),
DP c(
=
=
N
H
1
0.6
0.4
++
10-1 )X(x),
DP c(
=
=
N
L
90.
0.3
+
-1
)X(x),
DP
=
N
NN
0.16
0.9
0.1
+
+
1 0-1 )X(x),
DP
=
=
N
HN
4.0
-1
)X(x),
=
N
0.2
0.3
+
+
1 0-1 )X(x),
=
=
N
1 0-1 )X(x),
=
=
N
1 0-1 )X(x),
=
=
N
DP
NH
DP
LL
DP
NL
DP
HH
SP,
c(
SP,
c(
SP,
c(
c(
SP,
Step 1 3 : Assume β =0.2. The fu zzy β-lower and β-
upper approximation for class XN=({Obj(1) ,Obj(3) })
is first calculated. Since only {Obj(1) ,Obj(3) } is
, thus:
included in XN
*0.2 (XN ) = {{( Obj(1) ,Obj(2), Obj(3), Obj(5)
SP,DP
),0.1}
SP,
c(
55.0
18.0
0.4
=
+
0.3
β-upper
Similarly,
fuzzy
modified
the
N is
approximations of single attributes for class X
then calculated . Since only {Obj(1) ,Obj(3) } is
, thus:
included in XN
0.2 (X N ) = {( Obj(1) ,Obj(2), Obj(3), Obj(5) ,Obj(6),
*
SP
Obj(7) ),0.1}
),0.16}
0.2 (X N ) = {{( Obj(1) ,Obj(2), Obj(3) ,Obj(5)
DP*
(3) ,Obj(5) ,Obj(6)
),0.4}}
{( Obj
*
0.2 (XN ) = {{( Obj(3) ,Obj(5) ,Obj(6)
SP,DP
Step 1 5 :
in the fu zzy β-lower
Each partition
approximation is used to derived a β-certain rule with
and
plausibility measure
effectiveness
future
measure. From SP * 0.2 (X N
), the following rule is
derived:
),0.2}
1. If Systolic Pressure = Normal and Diastolic Pressure =
Normal Then Blood
Pressure = Normal
, with
plausibility=0.82 and future effectiveness = 0.1 .
Similarly, not exist rules are derived for DP *0.2 (XN )
and SP,DP *0.2 (XN
).
Step 16: Since the condition part of this rule is not
more specific than any other rule in the β-certain rule
sets, it is then kept in the fuzzy β-certain rule set.
Step 1 7 :
in the fu zzy β-upper
Each partition
approximation is used to derived a β-possible rule
with plausibility measure and future effectiveness
measure. From SP*
), the following rule is
0.2 (X N
derived:
2. If Systolic Pressure = Normal Then Blood Pressure =
Normal , with plausibility=0.72 and future effectiveness =
0.1 .
the following rules are derived for
Similarly,
0.2 (X N ) and SP,DP*
DP*
0.2 (XN
) :
3. If Diastolic Pressure = Normal Then Blood Pressure =
Normal, with plausibility =0.6 and future effectiveness =
0.16.
4. If Diastolic Pressure = High Then Blood Pressure
=Normal, with plausibility =0.2 and future effectiveness =
0.4.
5. If Systolic Pressure = Normal and Diastolic Pressure =
High Then Blood Pressure = Normal , with plausibility
=0.45 and future effectiveness = 0.2 .
Step 18: Since the condition part of this rule is not
more specific than any other rule in the β-possible
rule sets, it is then k ept in the fu zzy β-possible rule
set.
Step 19: Steps 13–19 are repeated to find rules for
until l ≤ c .
classes XL and XH
Similarly , perform for X
L and XH
. thus,
6. If Systolic Pressure = Low Then Blood Pressure = Low,
with plausibility =1 and future effectiveness = 0.5.
7. If Diastolic Pressure = Low Then Blood Pressure = Low,
with plausibility = 1 and future effectiveness = 1;
8. If Systolic Pressure =Low and Diastolic Pressure = Low
Then Blood Pressure = Low, with plausibility =1 and future
effectiveness = 0.5.
9. If Systolic Pressure = Normal and Diastolic Pressure =
Low Then Blood Pressure = Low , with plausibility=1 and
future effectiveness = 0.1.
10. If Systolic Pressure = High Then Blood Pressure =
High, with plausibility =1 and future effectiveness = 0.5.
11. If Diastolic Pressure = High Then Blood Pressure =
High, with plausibility = 0.8 and future effectiveness =
0.4;
12. If Systolic Pressure = High and Diastolic Pressure =
Normal Then Blood Pressure = High, with plausibility =1
and future effectiveness = 0.16 .
13. If Systolic Pressure = High and Diastolic Pressure =
High Then Blood Pressure = High , with plausibility=1 and
future effectiveness = 0.5 .
14. If Diastolic Pressure = Normal Then Blood Pressure =
High, with plausibility = 0.32 and future effectiveness =
0.16;
15. If Systolic Pressure = Normal and Diastolic Pressure =
High Then Blood Pressure = High, with plausibility =0.56
and future effectiveness = 0.2 .
16. If Systolic Pressure = Normal Then Blood Pressure =
High , with plausibility=0.25 and future effectiveness =
0.1.
Since the condition part of the eightieth rule is more
specific and both its plausibility (1) and effectiveness
(0.5) is equal to those of the sixth rule, this rule is
thus removed from the β-certain rule set. Similarly,
rules 9,12 and 13 are also removed by fuzzy rules 7
and 10 .
Step 2 0 : all the fu zzy β-certain and β-possible rules
are shown below:
Fuzzy β-certain rules :
1. If Systolic Pressure = Normal and Diastolic Pressure =
Normal Then Blood
, with
Pressure = Normal
plausibility=0.82 and future effectiveness = 0.1 .
2. If Systolic Pressure = Low Then Blood Pressure = Low,
with plausibility =1 and future effectiveness = 0.5.
3. If Diastolic Pressure = Low Then Blood Pressure = Low,
with plausibility = 1 and future effectiveness = 1;
4. If Systolic Pressure = High Then Blood Pressure = High,
with plausibility =1 and future effectiveness = 0.5.
5. If Diastolic Pressure = High Then Blood Pressure =
High, with plausibility = 0.8 and future effectiveness =
0.4.
Fuzzy β-possible rules :
1. If Systolic Pressure = Normal Then Blood Pressure =
Normal , with plausibility=0.72 and future effectiveness =
0.1 .
2. If Diastolic Pressure = Normal Then Blood Pressure =
Normal, with plausibility =0.6 and future effectiveness =
0.16 .
3. If Diastolic Pressure = High Then Blood Pressure =
Normal, with plausibility =0.2 and future effectiveness =
0.4 .
[4]
J. W,Grzymala-Busse,” Knowledge acquisition under
uncertainty: A rough set approach”, Journal of Intelligent Robotic
Systems, 1,pp. 3–16, 1988.
[5] T. P,Hong,C. S,Kuo,, S. C,Chi , ”Mining association rules
from quantitative data”, Intelligent Data Analysis, 3, pp.363–376,
1999.
[6]
T. P,Hong, C.Y,Lee, ”.Induction of
fuzzy rules and
membership functions from training examples”, Fuzzy Sets and
Systems, 84, pp.33–47, 1996.
[7] T. P,Hong, S.S,Tseng, “A generalized version space learning
algorithm for noisy and uncertain data”,IEEE Transactions on
Knowledge and Data Engineering, 9, pp.336–340,1997
[8] T. P,Hong, T.T,Wang, S.L,Wang,” Knowledge acquisition
from quantitative data using the rough-set theory”, Intelligent
Data Analysis, 4, pp.289–304, 2000.
[9] Y,Kodratoff, R.S,Michalski, “Machine learning: An artificial
intelligence artificial intelligence approach”, 3. San Mateo, CA:
Morgan Kaufmann Publishers, 1983 .
[10] P.J,Lingras, Y.Y,Yao,“Data mining using extensions of the
rough set model”, Journal of the American Society for Information
Science, 49(5), pp.415–422, 1998.
[11]
T.P,Hong, L.H ,Tseng, S.L,Wang,”Learning rules from
incomplete training examples by rough sets”, Expert System with
Application, 22,pp. 285–293, 2002.
[12]
R.S,Michalski, J.G, Carbonell, T.M,Mitchell,”Machine
Learning: An Artificial Intelligence Approach 1”, Los Altos, CA:
Morgan Kaufmann Publishers, 1983.
[13]
R.S,Michalski,J.G, Carbonell, T.M,Mitchell,“Machine
learning: An artificial intelligence approach 2”, Los Altos, CA:
Morgan Kaufmann Publishers,1983.
[14] E,Orlowska,“Reasoning with incomplete information: rough
set based information logics”, In V. Alagar, S. Bergler, & F. Q.
Dong (Eds.), Incompleteness and uncertainty in information
systems, pp. 16–33,1993.
[15] Y,Yuan, Y., & M.J,Shaw, “Induction of fuzzy decision trees”,
Fuzzy Sets and Systems, 69, pp.125–139,1995.
[16] N,Zhong, J.Z,Dong, S,Ohsuga, T.Y,Lin,” An incremental,
probabilistic rough set approach to rule discovery”, IEEE
International Conference on Fuzzy Systems, 2,pp. 933–938,1998.
[17] W,Ziarko, “Variable precision rough set model”, Journal of
Computer and System Sciences, 46, pp.39–59,1993.
[18] T.P,Hong, L.H,Tseng, B.C ,Chien,“Mining from incomplete
quantitative data by fuzzy rough sets”, Expert System with
Application, 37, pp.2644–2653,2010.
[19] Z. Pawlak, “Rough set,” International Journal of Computer
and Information Sciences, pp.341-356,1983.
[20] M. Kryszkiewicz, “Rough set approach to incomplete
informationsystems”, Information Science, Vol. 112, pp. 39-
49,1998.
[21] M. R,Chmielewski, J. W,Grzymala-Busse, N. W,Peterson,
S,Than,“The rule induction system LERS – a version for sonal
computers”, Foundations of Computing and Decision Sciences,
Vol .18, pp. 181-212,1993.
[22] J,Liang , Z,Xu ,”Uncertainty measures of roughness of
knowledge and rough sets in incomplete information systems”, The
third world congress on intelligent control and automation ,
pp.2526-2529, 2000 .
4. If Systolic Pressure = Normal and Diastolic Pressure =
High Then Blood Pressure = Normal , with plausibility
=0.45 and future effectiveness = 0.2.
5. If Diastolic Pressure = Normal Then Blood Pressure =
High, with plausibility = 0.32 and future effectiveness =
0.16;
6. If Systolic Pressure = Normal and Diastolic Pressure =
High Then Blood Pressure = High, with plausibility =0.56
and future effectiveness = 0.2 .
7. If Systolic Pressure = Normal Then Blood Pressure =
High , with plausibility=0.25 and future effectiveness =
0.1.
VII. Discussion and conclusion
In this paper, we have proposed a novel data
mining algorithm, which can process incomplete
quantitative data with a predefined tolerance degree
of uncertainty and misclassification. The interaction
between data and approximations helps derive β-
certain and β-possible rules from fuzzy incomplete
data sets and estimate appropriate unknown values.
The fuzzy β-certain rules with misclassification
degrees smaller than β and the fuzzy β-possible rules
with misclassification degrees smaller than 1-β are
derived. Noisy training examples (as outliers) may
then be omitted. The rules thus mined exhibit fuzzy
quantitative regularity in databases and can be used to
provide some sugge stions to appropr iate supervisors.
The proposed algorithm can also solve conventional
incomplete quantitative data problems by using
degraded membership functions. The selection of β
values is remarked here. When β =0, the proposed
approach will be reduced to the one for the original
rough-set model with incomplete quantitative data. A
larger β value represents a larger tolerance of
uncertainty and misclassification, but with a smaller
gap between
lower approximations and upper
approximations. The selection of an appropriate β
value then depends on given training instances.
Acknowledgement
This research was supported by the affiliation must
be Islamic Azad University, Arak Branch of Iran.
References
[1] T.P,Hong, T.T,Wang, S.L,Wang,” Mining fuzzy b-certain and
b-possible rules from quantitative data based on the variable
precision rough-set mode”, Expert System with Application, 32,pp.
223–232, 2007.
[2] L.T,Germano, P,Alexandre ,”Knowledge-base reduction
based on rough set techniques”, Canadian conference on electrical
and computer engineering , pp. 278–281, 1996.
[3]
I.Graham, P.L,Jones
,”Expert systems—knowledge
,uncertainty and decision”, Boston: Chapman and Computing, pp.
117–158 ,1988.
|
1902.04427 | 2 | 1902 | 2019-05-29T09:59:23 | Compressed Range Minimum Queries | [
"cs.DS"
] | Given a string $S$ of $n$ integers in $[0,\sigma)$, a range minimum query RMQ$(i, j)$ asks for the index of the smallest integer in $S[i \dots j]$. It is well known that the problem can be solved with a succinct data structure of size $2n + o(n)$ and constant query-time. In this paper we show how to preprocess $S$ into a compressed representation that allows fast range minimum queries. This allows for sublinear size data structures with logarithmic query time. The most natural approach is to use string compression and construct a data structure for answering range minimum queries directly on the compressed string. We investigate this approach in the context of grammar compression. We then consider an alternative approach. Instead of compressing $S$ using string compression, we compress the Cartesian tree of $S$ using tree compression. We show that this approach can be exponentially better than the former, is never worse by more than an $O(\sigma)$ factor (i.e. for constant alphabets it is never asymptotically worse), and can in fact be worse by an $\Omega(\sigma)$ factor. | cs.DS | cs | Compressed Range Minimum Queries$
Pawe(cid:32)l Gawrychowskia, Seungbum Job, Shay Mozesc, Oren Weimannb
aUniversity of Wroc(cid:32)law
bUniversity of Haifa
cInterdisciplinary Center Herzliya
9
1
0
2
y
a
M
9
2
]
S
D
.
s
c
[
2
v
7
2
4
4
0
.
2
0
9
1
:
v
i
X
r
a
Abstract
Given a string S of n integers in [0, σ), a range minimum query RMQ(i, j) asks for the index of the smallest
integer in S[i . . . j].
It is well known that the problem can be solved with a succinct data structure of
In this paper we show how to preprocess S into a compressed
size 2n + o(n) and constant query-time.
representation that allows fast range minimum queries. This allows for sublinear size data structures with
logarithmic query time. The most natural approach is to use string compression and construct a data
structure for answering range minimum queries directly on the compressed string. We investigate this
approach in the context of grammar compression. We then consider an alternative approach. Instead of
compressing S using string compression, we compress the Cartesian tree of S using tree compression. We
show that this approach can be exponentially better than the former, is never worse by more than an O(σ)
factor (i.e.
for constant alphabets it is never asymptotically worse), and can in fact be worse by an Ω(σ)
factor.
Keywords: RMQ, grammar compression, SLP, tree compression, Cartesian tree.
1. Introduction
Given a string S of n integers in [0, σ), a range minimum query RMQ(i, j) returns the index of the
smallest integer in S[i . . . j]. A range minimum data structure consists of a preprocessing algorithm and a
query algorithm. The preprocessing algorithm takes as input the string S, and constructs the data structure,
whereas the query algorithm takes as input the indices i, j and, by accessing the data structure, returns
RMQ(i, j). The range minimum problem is a fundamental data structure problem that has been extensively
studied, both in theory and in practice (see e.g. [11] and references therein).
Range minimum data structures fall into two categories. Systematic data structures store the input
string S in plain form, whereas non-systematic data structures do not. A significant amount of attention
has been devoted to devising RMQ data structures that answer queries in constant time and require as little
space as possible. There are succinct systematic structures that answer queries in constant time and require
fewer than 2n bits in addition to the n log σ bits required to represent S [11]. Similarly, there are succinct
non-systematic structures that answer queries in constant time, and require 2n + o(n) bits [11, 7].
The Cartesian tree C of S is a rooted ordered binary tree with n nodes. It is defined recursively. The
Cartesian tree of an empty string is an empty tree. Let i be the index of the smallest element of S (if the
smallest element appears multiple times in S, let i be the first such appearance). The Cartesian tree of S is
composed of a root node whose left subtree is the Cartesian tree of S[1, i− 1], and whose right subtree is the
Cartesian tree of S[i + 1, n]. See Figure 1. By definition, the character S[i] corresponds to the i'th node in
an inorder traversal of C (we will refer to this node as node i). Furthermore, for any nodes i and j in C, their
$Preliminary version of this paper appeared in SPIRE 2018. The work is supported in part by Israel Science Foundation
grant 592/17.
Email addresses: [email protected] (Pawe(cid:32)l Gawrychowski), [email protected] (Seungbum Jo),
[email protected] (Shay Mozes), [email protected] (Oren Weimann)
1
lowest common ancestor LCA(i, j) in C corresponds to RMQ(i, j) in S. It follows that the Cartesian tree of S
completely characterizes S in terms of range minimum queries. Indeed, two strings return the same answers
for all possible range minimum queries if and only if their Cartesian trees are identical. This well known
property has been used by many RMQ data structures including the succinct structures mentioned above.
Since there are 22n−O(log n) distinct rooted binary trees with n nodes, there is an information theoretic lower
bound of 2n− O(log n) bits for RMQ data structures. In this sense, the above mentioned 2n + o(n) bits data
structures [11, 7] are nearly optimal.
1.1. Our results and techniques
In this work we present RMQ data structures in the word-RAM model (without using any bit tricks).
The size (in words) of our data structures can be sublinear in the size of the input string and the query
time is O(log n). This is achieved by using compression techniques, and developing data structures that
can answer RMQ/LCA queries directly on the compressed objects. Since we aim for sublinear size data
structures, we focus on non-systematic data structures. We consider two different approaches to achieve this
goal. The first approach is to use string compression to compress S, and devise an RMQ data structure on
the compressed representation. This approach has also been suggested in [1, Section 7.1] in the context of
compressed suffix arrays and the LCP array. See also [7, Theorem 2],
[11, Theorem 4.1], and [3] for steps in
this direction. The other approach is to use tree compression to compress the Cartesian tree C, and devise
an LCA data structure on the compressed representation. To the best of our knowledge, this is the first
time such approach has been suggested. Note that these two approaches are not equivalent. For example,
consider a sorted sequence of an arbitrary subset of n different integers from [1, 2n]. As a string this sorted
sequence is not compressible, but its Cartesian tree is an (unlabeled) path, which is highly compressible. In
a nutshell, we show that the tree compression approach can exponentially outperform the string compression
approach. Furthermore, it is never worse than the string compression approach by more than an O(σ) factor,
and this O(σ) factor is unavoidable. We next elaborate on these two approaches.
Using string compression. In Section 2.1, we show how to answer range minimum queries on a grammar
compression of the input string S. A grammar compression is a context-free grammar that generates only
S. The grammar is represented as a straight line program (SLP) S. I.e., the right-hand side of each rule in
S either consists of the concatenations of two non-terminals or of a single terminal symbol. The size S of
the SLP S is defined as the number of rules in S. Ideally, S (cid:28) S. Computing the smallest possible SLP is
NP-hard [6], but there are many theoretically and practically efficient compression schemes for constructing
S [6, 14, 18] that reasonably approximate the optimal SLP. In particular, Rytter [17] showed an SLP S of
depth log n (the depth of an SLP is the depth of its parse tree) whose size is larger than the optimal SLP by
at most a multiplicative log n factor. Very recently, Ganardi et al. [13] showed that any SLP can be turned
(with no asymptotic overhead) into an equivalent SLP that is of depth log n.
In [1], it was shown how to support range minimum queries on S with a data structure of size O(S) in
time proportional to the depth of the SLP S. Bille et al. [5] designed a data structure of size O(S) that
supports random-access to S (i.e. retrieve the i'th symbol in S) in O(log n) time (i.e. regardless of the depth
of the SLP S). We show how to simply augment their data structure within the same O(S) size bound
to answer range minimum queries in O(log n) time (i.e. how to avoid the logarithmic overhead incurred by
using the solution of [1] on Rytter's SLP).
Theorem 1.1. Given a string S of length n and an SLP-grammar compression S of S, there is a data
structure of size O(S) that answers range minimum queries on S in O(log n) time.
Using tree compression. In Section 2.2, we give a data structure for answering LCA queries on a com-
pressed representation of the Cartesian tree C. By the discussion above, this is equivalent to answering range
minimum queries on S. There are various ways to compress trees. In this work we use DAG compression
of the top-tree of the Cartesian tree C of S. These concepts will be explained in the next paragraph. It is
likely that other tree compression techniques (see, e.g., [15, 16, 12]) can also yield interesting results. We
leave this as future work.
2
A top-tree [2] of a tree T is a hierarchical decomposition of the edges of T into clusters. A cluster is a
connected subgraph of T that has at most two boundary nodes (nodes with neighbors outside the cluster);
the root of the cluster (called the top boundary node), and a leaf of the cluster (called a bottom boundary
node). The intersection of any two clusters is either empty, or consists of exactly one (boundary) node.
Such a decomposition can be described by a rooted ordered binary tree T , called a top-tree, whose leaves
correspond to clusters with individual edges of T , and whose root corresponds to the entire tree T . The
cluster corresponding to a non-leaf node of T is obtained from the clusters of its two children by either
identifying their top boundary nodes (horizontal merge) or by identifying the top boundary node of the left
child with the bottom boundary node of the right child (vertical merge). See Figure 1.
A DAG compression [9] of a tree T is a representation of T by a DAG whose nodes correspond to nodes
of T . All nodes of T with the same subtree are represented by the same node of the DAG. Thus, the DAG
compression of a top-tree has two sinks (due to T being a binary unlabeled tree), corresponding to the two
types of leaf nodes of T (a single edge cluster, either left or right), and a single source, corresponding to the
root of T . If u is the parent of (cid:96) and r in T , then the node in the DAG representing the subtree of T rooted
at u has edges leading to the two nodes of the DAG representing the subtree of T rooted at (cid:96) and the subtree
of T rooted at r. Thus, repeating rooted subtrees in T are represented only once in the DAG. See Figure 1.
A top-tree compression [4] of a tree T is a DAG compression of T 's top-tree T . Bille et al. [4] showed
how to construct a data structure whose size is linear in the size of the DAG of T and supports navigational
queries on T in time linear in the depth of T . In particular, given the preorder numbers of two vertices
u, v in T , their data structure can return the preorder number of LCA(u, v) in T . We show that their data
structure can be easily adjusted to work with inorder numbers instead of preorder, so that, given the inorder
numbers i, j of two vertices in T one can return the inorder number of LCA(i, j) in T . This is precisely
RMQ(i, j) when T is taken to be the Cartesian tree C of S.
Theorem 1.2. Given a string S of length n and a top-tree compression T of the Cartesian tree C, we can
construct a data structure of size O(T ) that answers range minimum queries on S in O(depth(T )) time.
By combining Theorem 1.2 with the greedy construction of T given in [4] (in which depth(T ) = O(log n)),
we can obtain an O(T ) space data structure that answers RMQ in O(log n) time.
We already mentioned that, on some RMQ instances, top-tree compression can be much better than any
string compression technique. As an example, consider the string S = 1, 2, 3,··· , n. Its Cartesian tree is a
single (rightmost, and unlabeled) path, which compresses using top-tree compression into size T = O(log n).
On the other hand, since σ = n, S is uncompressible with an SLP. By Theorem 1.2, this shows that the
tree compression approach to the RMQ problem can be exponentially better than the string compression
approach. In fact, for any string over an alphabet of size σ = Ω(n), any SLP must have S = Ω(n) while for
top-trees T = O(n/ log n) [4]. In Section 3.1 we show that, for small alphabets, T cannot be much larger
nor much deeper than S for any SLP S.
Theorem 1.3. Given a string S of length n over an alphabet of size σ, for any SLP-grammar compression S
of S there is a top-tree compression T of the Cartesian tree C with size O(S·σ) and depth O(depth(S)·log σ).
Observe that in the above theorem, in order to obtain small depth of T , one could use, e.g, the SLP of
Rytter [17] or that of Ganardi et al. [13]. Another way of achieving small depth is to ignore the SLP and
use the top-tree compression of Bille et al. [4] as T . This guarantees T has size O(n/ log n) words and depth
O(log n).
Finally, observe that T can be larger than S by an O(σ) multiplicative factor which can be large for large
alphabets. It is tempting to try and improve this. However, in Section 3.2 we prove a tight lower bound,
showing that this factor is unavoidable.
Theorem 1.4. For every sufficiently large σ and s = Ω(σ2), there exists a string S of integers in [0, σ) that
can be described with an SLP S of size s, such that any top-tree compression T of the Cartesian tree C of S
is of size Ω(s · σ).
3
2. RMQ on Compressed Representations
2.1. Compressing the string
Given an SLP compression S of S, Bille et al. [5] presented a data structure of size O(S) that can report
any S[i] in O(log n) time. We now prove Theorem 1.1 using a rather straightforward extension of this data
structure to support range minimum queries.
The key technique used in [5] is an efficient representation of the heavy path decomposition of the SLP's
parse tree. For each node v in the parse tree, we select the child of v that derives the longer string to be
a heavy node. The other child is light. Ties can be broken arbitrarily. Heavy edges are edges going into a
heavy node and light edges are edges going into a light node. The heavy edges decompose the parse tree
into heavy paths. The number of light edges on any path from a node v to a leaf is O(log v) where v
denotes the length of the string derived from v. A traversal of the parse tree from its root to the i'th leaf
S[i] enters and exists at most log n heavy paths. Bille et al. show how to simulate this traversal in O(log n)
time on a representation of the heavy path decomposition that uses only O(S) space. In other words, their
structure finds the entry and exit vertices of all heavy paths encountered during the root-to-leaf traversal in
total O(log n) time. We elaborate on this now.
Note that we cannot afford to store the entire parse tree as its size is n which can be exponentially
larger than S. Instead, Bille et al. use the following O(S)-space representation H of the heavy paths: H
is a forest of trees whose roots correspond to terminals of S (integers in [0, σ)) and whose non-root nodes
correspond to nonterminals of S. A node u is the parent of v in H iff u is the heavy child of v in the parse
tree (observe that wherever the nonterminal u appears in the parse tree it always has the same heavy child
v). We assign the edge of H from v to its parent u with a left weight (cid:96)(v, u) and right weight r(v, u) defined
as follows. If u is the left child of v in H then the left weight is 0 and the right weight is the subtree size of
the right child of v in the parse tree. Otherwise, the right weight is 0 and the left weight is the subtree size
of the left child of v in the parse tree.
Using H, we can then simulate a root-to-leaf traversal of the parse tree. Suppose we have reached a
vertex u on some heavy path P . Finding out where we need to exit P (and enter another heavy path) easily
translates to a weighted level ancestor query (using the (cid:96)(v, u) or r(v, u) weights) from u on H (see [5] for
details). Given a positive number x, such a query returns the rootmost ancestor of u in H whose distance
from the root is at least x. Bille et al. showed how to answer all such queries (i.e. how to find the entry point
and exit point on all heavy paths visited during a root-to-leaf traversal of the parse tree) in total O(log n)
time.
Extending their structure to support range minimum queries is quite simple. We perform a random access
to the i'th and the j'th leaves in the parse tree. This identifies the entry and exit points of all traversed
heavy paths, and, in particular, the unique heavy path P (cid:48) containing the lowest common ancestor of i and j.
Then, we wish to find the minimum leaf value in all the subtrees hanging to the right (resp. left) of the path
starting from P (cid:48) and going down to i (resp. j). To achieve this for i (the case of j is symmetric), in addition
to the subtree sizes (r(v, u)) we also store the minimum leaf value (r(cid:48)(v, u)) in these subtrees. This way, the
problem now boils down to performing O(log n) bottleneck edge queries on H. Given a forest H with edge
weights (the r(cid:48)(v, u) weights), a bottleneck edge query between two vertices x, y (the entry and exit points)
returns the minimum edge weight on the unique x-to-y path in H. Demaine et al. [8] showed that, after
sorting the edge weights in H, one can construct in O(H) = O(S) time and space a data structure that
answers bottleneck edge queries in constant time. This concludes the proof of Theorem 1.1.
2.2. Compressing the Cartesian tree
We next prove Theorem 1.2, i.e. how to support range minimum queries on S using a compressed
representation of the Cartesian tree [19]. Recall that the Cartesian tree C of S is defined as follows: If the
smallest character in S is S[i] (in case of a tie we choose a leftmost position) then the root of C corresponds
to S[i], its left child is the Cartesian tree of S[1, i − 1] and its right child is the Cartesian tree of S[i + 1, n].
By definition, the i'th character in S corresponds to the node in C with inorder number i (we will refer to
this node as node i). Observe that for any nodes i and j in C, the lowest common ancestor LCA(i, j) of these
4
Figure 1: The string S = "23110122102313" and its corresponding (a) Cartesian tree , (b) top-tree, and (c) DAG representation
of the top-tree.
In (a), each node is labeled by its corresponding character in S (these labels are for illustration only, the
top-tree construction treats the Cartesian tree as an unlabeled tree). In (b) and (c), each node is labeled by el or er (atomic
edge clusters), v (a vertical merge), or h (a horizontal merge). Four clusters are marked with matching colors in (a) and in (b).
nodes in C corresponds to RMQ(i, j) in S. This implies that without storing S explicitly, one can answer
range minimum queries on S by answering LCA queries on C.
In this section, we show how to support
LCA queries on C on a top-tree compression [4] T of C. The query time is O(depth(T )) which can be made
O(log n) using the (greedy) construction of Bille et al. [4] that gives depth(T ) = O(log n). We first briefly
restate the construction of Bille et al., and then extend it to support LCA queries.
The top-tree of a tree T (in our case T will be the Cartesian tree C) is a hierarchical decomposition of
T into clusters. Let v be a node in T with children v1, v2.1 Define T (v) to be the subtree of T rooted at v.
Define F (v) to be the forest T (v) without v. A cluster with top boundary node v can be either (1) T (v), (2)
{v}∪ T (v1), or (3) {v}∪ T (v2). For any node u (cid:54)= v in a cluster with top boundary node v, deleting from the
cluster all descendants of u (not including u itself) results in a cluster with top boundary node v and bottom
boundary node u. The top-tree is a binary tree defined as follows (see Figure 1):
• The root of the top-tree is the cluster T itself.
• The leaves of the top-tree are (atomic) clusters corresponding to the edges of T . An edge (v, parent(v))
of T is a cluster where parent(v) is the top boundary node. If v is a leaf then there is no bottom
boundary node, otherwise v is a bottom boundary node. If v is the right child of parent(v) then we
label the (v, parent(v)) cluster as er and otherwise as e(cid:96).
• Each internal node of the top-tree is a merged cluster of its two children. Two edge disjoint clusters A
and B whose nodes overlap on a single boundary node can be merged if their union A ∪ B is also a
cluster (i.e. contains at most two boundary nodes). If A and B share their top boundary node then
the merge is called horizontal. If the top boundary node of A is the bottom boundary node of B then
the merge is called vertical and in the top-tree A is the left child and B is the right child.
Bille et al. [4] proposed a greedy algorithm for constructing the top-tree: Start with n − 1 clusters, one
for each edge of T , and at each step merge all possible clusters. More precisely, at each step, first do all
possible horizontal merges and then do all possible vertical merges. After constructing the top-tree, the
actual compression T is obtained by representing the top-tree as a directed acyclic graph (DAG) using the
1Bille et al. considered trees with arbitrary degree, but since our tree T is a Cartesian tree we can focus on binary trees.
5
001323211(a) Cartesian tree31221(b) Top-treevervh(c) DAG vvvhvvhhvererererererelelelerelelelvvvvvhherh algorithm of [9]. Namely, all nodes in the top-tree that have a child with subtree X will point to the same
subtree X (see Figure 1). Bille et al. [4] showed that using the above greedy algorithm, one can construct
T of size T that can be as small as log n (when the input tree T is highly repetitive) and in the worst-case
is at most O(n/ log0.19
σ n). Dudek and Gawrychowski [10] have recently improved the worst-case bound to
O(n/ logσ n) by merging in the i'th step only clusters whose size is at most αi for some constant α. Using
either one of these merging algorithms to obtain the top-tree and its DAG representation T , a data structure
of size O(T ) can then be constructed to support various queries on T . In particular, given nodes i and j
in T (specified by their position in a preorder traversal of T ) Bille et al. showed how to find the (preorder
number of) node LCA(i, j) in O(log n) time. Therefore, the only change required in order to adapt their data
structure to our needs is the representation of nodes by their inorder rather than preorder numbers.
The local preorder number uC of a node u in T and a cluster C in T is the preorder number of u in a
preorder traversal of the cluster C. To find the preorder number of LCA(i, j) in O(log n) time, Bille et al.
showed that it suffices if for any node u and any cluster C we can compute uC in constant time from uA or
uB (the local preorder numbers of u in the clusters A and B whose merge is the cluster C) and vice versa.
In Lemma 6 of [4] they show that indeed they can compute this in constant time. The following lemma is a
modification of that lemma to work when uA, uB and uC are local inorder numbers.
Lemma 2.1 (Modified Lemma 6 of [4]). Let C be an internal node in T corresponding to the cluster
obtained by merging clusters A and B. For any node u in C, given uC we can tell in constant time if u is
in A (and obtain uA) in B (and obtain uB) or in both. Similarly, if u is in A or in B we can obtain uC in
constant time from uA or uB.
Proof. We show how to obtain uA or uB when uC is given. Obtaining uC from uA or uB is done similarly.
For each node C, we store a following information:
• (cid:96)(A) (r(A)): the first (last) node visited in an inorder traversal of C that is also a node in A.
• (cid:96)(B) (r(B)): the first (last) node visited in an inorder traversal of C that is also a node in B.
• the number of nodes in A and in B.
• u(cid:48)
C, where u(cid:48) is the common boundary node of A and B.
Consider the case where C is obtained by merging A and B vertically (when the bottom boundary node of A
is the top boundary node of B), and where B includes vertices that are in the left subtree of this boundary
node, the other case is handled similarly:
• if uC < (cid:96)(B) then u is a node in A and uA = uC.
• if (cid:96)(B) ≤ uC ≤ r(B) then u is a node in B and uB = uC − (cid:96)(B) + 1. For the special case when uC = u(cid:48)
C
then u is also the bottom boundary node in A and uA = (cid:96)(B).
• if uc > r(B) then u is a node in A visited after visiting all the nodes in B then uA = uC − B + 1.
When C is obtained by merging A and B horizontally (when A and B share their top boundary node and
A is to the left of B):
• if uC < r(A) then u is a node in A and uA = uC.
• if uC ≥ r(A) then u is a node in B and uB = uC − A + 1. For the special case when uC = u(cid:48)
is also the top boundary node in A and uA = A.
C then u
2
To complete the proof of Theorem 1.2, we now explain how to use Lemma 2.1, given the inorder numbers
of nodes x and y in T , to compute in O(depth(T )) time the inorder number of LCA(x, y) in T . This is
identical to the procedure of Bille et al. (except for replacing preorder with inorder) and is given here for
completeness.
6
We begin with a top-down search on T to find the first cluster whose top boundary node is LCA(x, y) (or
alternatively to reach a leaf cluster whose top or bottom boundary node is LCA(x, y)). At each cluster C in
the search we compute the local inorder numbers xC and yC of x and y in C. Initially, for the root cluster T
we set xT = x and yT = y. If we reach a leaf cluster C we stop the search. Otherwise, C is an internal cluster
with children A and B. If xC and yC are in the same child cluster, we continue the search in that cluster
after computing the new local inorder numbers of x and y in the appropriate child cluster (in constant time
using Lemma 2.1). Otherwise, xC and yC are in different child clusters. If C is a horizontal merge then we
stop the search. If C is a vertical merge (where A's bottom boundary node is B's top boundary node) then
we continue the search in A after setting the local inorder number of the node (either x or y) that is in B to
be the bottom boundary node of A.
After finding the cluster C whose top boundary node v is LCA(x, y), we have the inorder number of v in
C and we need to compute the inorder number of v in the entire tree T . This is done by repeatedly applying
Lemma 2.1 on the path in T from C to the root of T .
3. Compressing the String vs. the Cartesian Tree
In this section we compare the sizes of the SLP compression S and the top-tree compression T .
3.1. An upper bound
We now show that given any SLP S of height h, we can construct a top-tree compression T based on S
(i.e. non-greedily) such that T = O(S· σ) and the height of T is O(h log σ). Using T , we can then answer
range minimum queries on S in time O(h log σ) as done in Section 2.2. Furthermore, we can construct T in
O(n log σ + S · σ) time using Rytter's SLP [17] as S. Then, the height of S is h = log n and the size of S
is larger than the optimal SLP by at most a multiplicative log n factor.
Consider a rule C → AB in the SLP. We will construct a top-tree (a hierarchy of clusters) of C (i.e. of
the Cartesian tree CT (C) of the string derived by the SLP variable C) assuming we have the top-trees of
(the Cartesian trees of the strings derived by) A and of B. We show that the top-tree of C contains only
O(σ) new clusters that are not clusters in the top-trees of A and of B, and that the height of the top-tree
is only O(log σ) larger than the height of the top-tree of A or the top-tree of B. To achieve this, for any
variable A of the SLP, we will make sure that certain clusters (associated with its rightmost and leftmost
paths) must be present in its top-tree. See Figure 2.
The structure of CT (C). We first describe how the Cartesian tree CT (C) of the string derived by
variable C can be described in terms of the Cartesian trees CT (A) and CT (B). We label each node in a
Cartesian tree with its corresponding character in the string. These labels are only used for the sake of this
description, the actual Cartesian tree is an unlabeled tree. By definition of the Cartesian tree, the labels are
monotonically non-decreasing as we traverse any root-to-leaf path. Let (cid:96)(A) (respectively r(A)) denote the
path in CT (A) starting from the root and following left (respectively right) edges. Since we break ties by
taking the leftmost occurrence of the same character we have that the path (cid:96)(A) is strictly increasing (the
path r(A) is just non-decreasing).
Let x be the label of the root of CT (B). To simplify the presentation we assume that the label of the
root of CT (A) is smaller or equal to x (the other case is symmetric). Split CT (A) by deleting the edge
connecting the last node on r(A) that is smaller or equal to x with its right child. The resulting two subtrees
are the Cartesian trees CT (Ap) and CT (As) of a prefix Ap and a suffix As of A whose concatenation is A.
The prefix Ap ends at the last character of A that is at most x. Split CT (B) by deleting the edge connecting
the root to its right child. The resulting two subtrees are the Cartesian trees CT (Bp) and CT (Bs) of a prefix
and a suffix of B. The prefix Bp ends with the first occurrence of x in B.
The Cartesian tree CT (C) of the concatenation C = AB can be described as follows: Consider the
Cartesian trees CT (Ap) of Ap, CT (Bs) of Bs, and CT (AsBp) of the concatenation of As and Bp.
It is
easy to verify that the root of CT (AsBp) has no right child. Attach CT (Bs) as the right child of the root
of CT (AsBp). Then attach the resulting tree as the right child of the last node of the rightmost path of
CT (Ap). See Figure 2.
7
Figure 2: The Cartesian tree of SLP variables A, B, C where C → AB. The single additional cluster Cr
3 (in green) is formed
by merging existing clusters from A (in blue) and from B (in red). First, cluster CAB (corresponding to the tree CT (AsBp)) is
formed by alternating subpaths of the leftmost path in CT (B) and the rightmost path in CT (A). Then, CAB is merged with
Br
3. In this example, As = {Ar
3 , v, and Ar
i i > 3} and Bp = {B(cid:96)
i i > 3}.
8
A`1A` 01 7A`70A`2201Ar03337788Ar1Ar3Ar7Ar8 3849B`4B`8B`9B` 0CT(A)CT(B)334Br4Br31 70201Ar0333Ar17788Ar889B`8B`93Ar34B`4Ar7CT(C)Cr3334Br4Br3vCABA`1A` A`7A`2 The above structural description of CT (C) is not enough for our purposes. In particular, a recursive
computation of CT (AsBp) would lead to a linear O(σ) increase in the height of the top-tree of CT (C)
compared to that of CT (A) and CT (B). In order to guarantee a logarithmic O(log σ) increase, we need to
describe the structure in more detail.
For a node v with label i appearing in (cid:96)(A) other than the root of A, we define A(cid:96)
i subtree rooted at the
v's right child, together with v. Next consider the path r(A). For every label i there can be multiple vertices
with label i that are consecutive on r(A). We define Ar
i to be the subtree of A induced by the union of all
vertices of r(A) that have label i together with the subtrees rooted at their left children. Again, we treat the
first node of r(A) (i.e. the root of CT (A)) differently: if the label of the root is i then Ar
i does not include
the root nor its left subtree. See Figure 2 (left).
j for j ≤ x.
i 's. The structure of CT (AsBp)
It is easy to see that CT (Ap) consists of the subtree of A induced by all the A(cid:96)
See Figure 2 (right). It is also easy to see that CT (Bs) consists of all the Br
is a bit more involved. It consists of alternations of Ar
i 's and all the Ar
i 's which we describe next.
i 's and B(cid:96)
We describe the structure of CT (AsBp) constructively from top to bottom. This constructive procedure
is just for the sake of describing the structure of CT (AsBp). We will later describe a different procedure for
constructing the clusters of the corresponding top-trees. The root of CT (AsBp) is the root of B. Initially,
the root is marked L (indicating that the root can only obtain a left child). Throughout the procedure we
will make sure that exactly one node is marked (by either L or R). For increasing values of i, starting with
i = x + 1 and ending when i exceeds σ, we do the following:
1. if Ar
i is defined:
(a) attach Ar
i as the left child of the marked node if the marked node is marked with L, and as the
right child otherwise,
(b) unmark the marked node, and instead mark the last node on the rightmost path of Ar
i with R.
2. if B(cid:96)
i is defined:
(a) attach B(cid:96)
i as the left child of the marked node if the marked node is marked with L, and as the
right child otherwise,
(b) unmark the marked node, and instead mark the root of B(cid:96)
i with L.
Note that CT (AsBp) consists of all subtrees B(cid:96)
i for i > x, and O(σ) additional edges
(these are the edges that were created when attaching subtrees to a marked node during the construction
procedure). Also observe that each subtree Ar
i is incident to at most two new edges, one incident to the
root of Ar
i is incident to at
most two new edges, both incident to the root of B(cid:96)
i into a single
node. Then the result would be a single "zigzag" path of new edges. We will next use these properties when
describing the clusters of CT (C).
i and the other incident to the rightmost node of Ar
i . Imagine contracting each Ar
i . Similarly, each subtree B(cid:96)
i , the subtrees Ar
i and each B(cid:96)
The clusters of CT (C). We next describe how to obtain the clusters for the top-tree of CT (C) from
the the clusters of the top-trees of CT (A) and CT (B). For each variable (say A) of the SLP S of S, we
require that in the top-tree of S there is a cluster for every A(cid:96)
i only have a
top boundary node. Clusters for Ar
i ), and a bottom boundary
node (the last node on the rightmost path of Ar
i clusters
of C by merging clusters of the form A(cid:96)
i while introducing only O(σ) new clusters, and
with O(log σ) increase in height. First observe that, by the structure of CT (Ap), we have that, for every i,
C (cid:96)
i . Recall that x denotes the label of
the root of CT (B). By the structure of CT (Ap), C r
i for every i < x and, by the structure of CT (Bs),
C r
i , so we already have these clusters. Next consider the clusters C r
i for every i > x. Therefore, the only new cluster we need to create is C r
x.
i have a top boundary node (the root of Ar
i ). We will show how to construct all the C (cid:96)
i . Clusters for A(cid:96)
i and every Ar
i , and Br
i and C r
The cluster C r
it contains the cluster Ar
Finally, Br
Ar
x and Br
cluster C r
x corresponds to a subtree of CT (C) that is composed of the following components: First,
x. Then, CT (AsBp) is connected as the right child of the rightmost node of Ar
x.
x is connected as the right child of the root of CT (AsBp). Since we already have the clusters for
x, we only need to describe how to construct a cluster CAB corresponding to CT (AsBp). The
x will be obtained by merging these three clusters together.
i , Ar
i , B(cid:96)
i = Br
i = Ar
i = A(cid:96)
9
Recall from the structural discussion of CT (AsBp) that CT (AsBp) consists of all subtrees B(cid:96)
i and the
subtrees Ar
i for i > x. We already have the clusters for these O(σ) subtrees. These subtrees are connected
together to form CT (AsBp) by O(σ) new edges that are only incident to the boundary nodes of the clusters.
We will merge the existing clusters to form the cluster CAB by creating O(σ) new clusters, but only increasing
the height of the top-tree by O(log σ). Performing the merges linearly would increase the height of the top-
tree by O(σ). Instead, the merging process consists of O(log σ) phases. In each phase we choose a maximal
set of disjoint pairs of clusters that need to be merged and perform these merges. Since CT (AsBp) is a
binary tree, we can perform half the remaining merges in each phase. This process can be described by a
binary tree of height O(log σ) whose leaves are the O(σ) clusters B(cid:96)
To conclude, given the clusters A(cid:96)
i . Once
we have all clusters of the SLP's start variable, we merge them into a single cluster (i.e. obtain the top-tree of
the entire Cartesian tree of S) by merging all its O(σ) clusters (introducing O(σ) new clusters and increasing
the height by O(log σ)) similarly to the description above. This concludes the proof of Theorem 1.3.
i , Br
i , C r
i , Ar
i , B(cid:96)
i that we started with.
i , we have shown how to compute all clusters C (cid:96)
i and Ar
3.2. A lower bound
We now prove Theorem 1.4. That is, for every sufficiently large σ and s = Ω(σ2) we will construct a
string S of integers in [0, σ) that can be described with an SLP S of size s, such that any top-tree compression
T of the Cartesian tree C of S is of size Ω(s · σ).
Let us first describe the high-level intuition. The shuffle of two strings x[1..(cid:96)] and y[1..(cid:96)] is defined as
x[1]y[1]x[2]y[2] . . . x[(cid:96)]y[(cid:96)]. It is not very difficult to construct a small SLP describing a collection of many
strings Ai and Bj of length (cid:96), and choose s pairs (ik, jk) such that every SLP describing all shuffles of Aik
and Bjk contains Ω(s · (cid:96)) nonterminals. However, our goal is to show a lower bound on the size of a top-tree
compression of the Cartesian tree, not on the size of an SLP. This requires designing the strings Ai and Bj
so that a top-tree compression of the Cartesian tree of AiBj roughly corresponds to an SLP describing the
shuffle of Ai and Bj.
Let σ(cid:48) = (cid:98)(σ − 4)/2(cid:99) and (cid:96) be a parameter such that 2(cid:96) − 1 ≥ σ(cid:48). We start with constructing 2(cid:96) − 1
distinct auxiliary strings Xi over {σ − 2, σ − 1}, each of the same length (cid:96). We construct every such string
except for (σ − 1)(cid:96), so that Cartesian trees corresponding to Xis are all distinct. The total number of Xis
is 2(cid:96) − 1 and there is an SLP X of size O(2(cid:96)) that contains a nonterminal deriving every Xi. Next, let Xi,j
denote the string Xi(σ − 3)Xj. By construction, Cartesian trees corresponding to Xi,js are all distinct, and
all Xi,js are of the same length.
The second and the third step are symmetric. We construct strings Ai of the form:
X1,i2X2,i4··· (2σ
(cid:48) − 4)Xσ(cid:48)−1,i(2σ
(cid:48) − 2)Xσ(cid:48),i(2σ
(cid:48)
)
for every i = 1, 2, . . . , 2(cid:96) − 1. There are 2(cid:96) − 1 such strings Ai, and there is an SLP A of size O((2(cid:96) − 1) · σ(cid:48))
that contains a nonterminal deriving every Ai.
Similarly, we construct strings Bj of the form:
(cid:48) − 1)Xσ(cid:48),j(2σ
(cid:48) − 3) . . . 3X2,j1X1,j.
(2σ
Finally, we obtain S by concatenating the strings AiBj0 for all i, j = 1, 2, . . . , 2(cid:96) − 1. The total size of
an SLP that generates S is O(2(cid:96) · σ(cid:48) + 4(cid:96)). It remains to analyze the size of a top-tree compression T of the
Cartesian tree C of S.
We first need to understand the structure of C. Because all strings AiBj are separated by 0s, the Cartesian
tree of S consists of a right path of length (2(cid:96) − 1)2 and the Cartesian tree of AiBj attached as the left
subtree of the ((i − 1)(2(cid:96) − 1) + j)-th node of the path. The Cartesian tree of a string AiBj consists of a
path of length 2σ(cid:48) starting at the root and consisting of nodes u1 − v1 − u2 − v2 − . . . − uσ(cid:48) − vσ(cid:48) such that
vi is the left child of ui and ui+1 is the right child of vi. For every k ∈ {1, . . . , σ(cid:48)}, the right subtree of uk is
the Cartesian tree of Xk,i and the left subtree of vk is the Cartesian tree of Xk,j. See Figure 3.
We define a zigzag to be an edge u − v such that v is the left child of u. Furthermore, for some
k ∈ {1, . . . , σ(cid:48)} and i, j ∈ {1, . . . , 2(cid:96) − 1}, the right subtree of u should be the Cartesian tree of Xk,j, while
the left subtree of v should be the Cartesian tree of Xk,i.
10
Figure 3: Structure of the Cartesian tree of AiBj for σ(cid:48) = 4.
Proposition 3.1. The Cartesian tree of AiBj contains σ(cid:48) distinct zigzags. Furthermore, a zigzag contained
in the Cartesian tree of AiBj is not contained in the Cartesian tree of Ai(cid:48)Bj(cid:48) for any i(cid:48) (cid:54)= i or j(cid:48) (cid:54)= j.
Lemma 3.2. If T is a top-tree compression of a tree T with x distinct zigzags then T = Ω(x).
Proof. We associate each distinct zigzag with a smallest cluster of the top-tree of T that contains it. We
claim that each cluster obtained by merging clusters A and B is associated with O(1) zigzags. Since the
size of T equals the number of distinct clusters in the top-tree of T , the lemma follows. Consider a zigzag
z = u − v associated with A ∪ B. Hence, z is not contained in A nor in B. We consider two cases.
1. A and B are merged horizontally. Then A and B share the top boundary node b, and in fact u = b. It
follows that z is the only zigzag in A ∪ B that is not in A nor in B.
2. A and B are merged vertically. Then the top boundary node of A is the bottom boundary node b of
B. Then either b = u, b = v, or b is a node of the Cartesian tree of some Xk,x attached as the right
subtree of u or the left subtree of v. Each of the first two possibilities gives us one zigzag associated
with A∪ B that is not in A nor in B. In the remaining two possibilities (i.e. when the Cartesian tree of
Xk,x is attached as the right subtree of u or as the left subtree of v), because the size of the Cartesian
tree of every Xk,x is the same, we can determine u or v, respectively, by navigating up from b as long
2
as the size of the current subtree is too small, and proceed as in the previous two cases.
Combining Proposition 3.1 and Lemma 3.2 we conclude that T = Ω(4(cid:96) · σ(cid:48)). Recall that the size of
an SLP that generates S is O(2(cid:96) · σ(cid:48) + 4(cid:96)), where σ(cid:48) = (cid:98)(σ − 4)/2(cid:99) = Θ(σ) and (cid:96) is parameter such that
2(cid:96) − 1 ≥ σ(cid:48). Given a sufficiently large σ and s = Ω(σ2), we first choose (cid:96) = (cid:100)1/2 log s(cid:101). Observe that then
2(cid:96) − 1 ≥ σ(cid:48) indeed holds because of the assumption s = Ω(σ2). We construct a string S generated by an SLP
of size O(2(cid:96) · σ(cid:48) + 4(cid:96)) = O(s), and any top-tree compression T of the Cartesian tree of S has size Ω(s · σ(cid:48)).
This concludes the proof of Theorem 1.4.
11
u1u2u3v1v2v3v4u4X1,iX2,iX3,iX4,iX4,jX3,jX2,jX1,j References
[1] A. Abeliuk, R. C´anovas, and G. Navarro. Practical compressed suffix trees. Algorithms, 6(2):319 -- 351,
2013.
[2] S. Alstrup, J. Holm, K. de Lichtenberg, and M. Thorup. Maintaining information in fully-dynamic trees
with top trees. ACM Transactions on Algorithms, 1:243 -- 264, 2003.
[3] J. Barbay, J. Fischer, and G. Navarro. LRM-trees: Compressed indices, adaptive sorting, and com-
pressed permutations. Theor. Comput. Sci., 459:26 -- 41, 2012.
[4] P. Bille, I. L. Gørtz, G. M. Landau, and O. Weimann. Tree compression with top trees. Inf. Comput.,
243:166 -- 177, 2015.
[5] P. Bille, G. M. Landau, R. Raman, K. Sadakane, S. R. Satti, and O. Weimann. Random access to
grammar-compressed strings and trees. SIAM J. Comput., 44(3):513 -- 539, 2015.
[6] M. Charikar, E. Lehman, D. Liu, R. Panigrahy, M. Prabhakaran, A. Sahai, and A. Shelat. The smallest
grammar problem. IEEE Trans. Information Theory, 51(7):2554 -- 2576, 2005.
[7] P. Davoodi, R. Raman, and S. R. Satti. On succinct representations of binary trees. Mathematics in
Computer Science, 11(2):177 -- 189, 2017.
[8] E. D. Demaine, G. M. Landau, and O. Weimann. On cartesian trees and range minimum queries.
Algorithmica, 68(3):610 -- 625, 2014.
[9] P. J. Downey, R. Sethi, and R. E. Tarjan. Variations on the common subexpression problem. J. ACM,
27(4):758 -- 771, 1980.
[10] B. Dudek and P. Gawrychowski. Slowing down top trees for better worst-case compression. In 29th
Annual Symposium on Combinatorial Pattern Matching (CPM), pages 16:1 -- 16:8, 2018.
[11] J. Fischer and V. Heun. Space-efficient preprocessing schemes for range minimum queries on static
arrays. SIAM Journal on Computing, 40(2):465 -- 492, 2011.
[12] M. Ganardi, D. Hucke, M. Lohrey, and E. Noeth. Tree compression using string grammars. Al-
gorithmica, 80(3):885 -- 917, 2018. doi: 10.1007/s00453-017-0279-3. URL https://doi.org/10.1007/
s00453-017-0279-3.
[13] M. Ganardi, A. Jez, and M. Lohrey. Balancing straight-line programs. CoRR, abs/1902.03568, 2019.
URL http://arxiv.org/abs/1902.03568.
[14] K. Goto, H. Bannai, S. Inenaga, and M. Takeda. Fast q-gram mining on SLP compressed strings. J.
Discrete Algorithms, 18:89 -- 99, 2013.
[15] A. Jez and M. Lohrey. Approximation of smallest linear tree grammar. Inf. Comput., 251:215 -- 251,
2016.
[16] M. Lohrey. Grammar-based tree compression. In Developments in Language Theory - 19th International
Conference, DLT 2015, pages 46 -- 57, 2015.
[17] W. Rytter. Application of Lempel-Ziv factorization to the approximation of grammar-based compres-
sion. Theor. Comput. Sci., 302(1-3):211 -- 222, 2003.
[18] Y. Takabatake, T. I, and H. Sakamoto. A space-optimal grammar compression. In 25th Annual European
Symposium on Algorithms (ESA), pages 67:1 -- 67:15, 2017.
[19] J. Vuillemin. A unifying look at data structures. Commun. ACM, 23(4):229 -- 239, 1980.
12
|
1606.07425 | 2 | 1606 | 2016-06-27T19:48:50 | Generalized Preconditioning and Network Flow Problems | [
"cs.DS"
] | We consider approximation algorithms for the problem of finding $x$ of minimal norm $\|x\|$ satisfying a linear system $\mathbf{A} x = \mathbf{b}$, where the norm $\|\cdot \|$ is arbitrary and generally non-Euclidean. We show a simple general technique for composing solvers, converting iterative solvers with residual error $\|\mathbf{A} x - \mathbf{b}\| \leq t^{-\Omega(1)}$ into solvers with residual error $\exp(-\Omega(t))$, at the cost of an increase in $\|x\|$, by recursively invoking the solver on the residual problem $\tilde{\mathbf{b}} = \mathbf{b} - \mathbf{A} x$. Convergence of the composed solvers depends strongly on a generalization of the classical condition number to general norms, reducing the task of designing algorithms for many such problems to that of designing a \emph{generalized preconditioner} for $\mathbf{A}$. The new ideas significantly generalize those introduced by the author's earlier work on maximum flow, making them more widely applicable.
As an application of the new technique, we present a nearly-linear time approximation algorithm for uncapacitated minimum-cost flow on undirected graphs. Given an undirected graph with $m$ edges labelled with costs, and $n$ vertices labelled with demands, the algorithm takes $\epsilon^{-2}m^{1+o(1)}$-time and outputs a flow routing the demands with total cost at most $(1+\epsilon)$ times larger than minimal, along with a dual solution proving near-optimality. The generalized preconditioner is obtained by embedding the cost metric into $\ell_1$, and then considering a simple hierarchical routing scheme in $\ell_1$ where demands initially supported on a dense lattice are pulled from a sparser lattice by randomly rounding unaligned coordinates to their aligned neighbors. Analysis of the generalized condition number for the preconditioner follows that of the classical multigrid algorithm for lattice Laplacian systems. | cs.DS | cs |
Generalized Preconditioning and Network Flow Problems
Jonah Sherman∗
University of California, Berkeley
October 13, 2018(preliminary draft)
Abstract
We consider approximation algorithms for the problem of finding x of minimal norm (cid:107)x(cid:107) satisfying a
linear system Ax = b, where the norm (cid:107) · (cid:107) is arbitrary and generally non-Euclidean. We show a simple
general technique for composing solvers, converting iterative solvers with residual error (cid:107)Ax−b(cid:107) ≤ t−Ω(1)
into solvers with residual error exp(−Ω(t)), at the cost of an increase in (cid:107)x(cid:107), by recursively invoking the
solver on the residual problem b = b − Ax. Convergence of the composed solvers depends strongly
on a generalization of the classical condition number to general norms, reducing the task of designing
algorithms for many such problems to that of designing a generalized preconditioner for A. The new
ideas significantly generalize those introduced by the author's earlier work on maximum flow, making
them more widely applicable.
As an application of the new technique, we present a nearly-linear time approximation algorithm for
uncapacitated minimum-cost flow on undirected graphs. Given an undirected graph with m edges labelled
with costs, and n vertices labelled with demands, the algorithm takes −2m1+o(1)-time and outputs a flow
routing the demands with total cost at most (1 + ) times larger than minimal, along with a dual solution
proving near-optimality. The generalized preconditioner is obtained by embedding the cost metric into
(cid:96)1, and then considering a simple hierarchical routing scheme in (cid:96)1 where demands initially supported
on a dense lattice are pulled from a sparser lattice by randomly rounding unaligned coordinates to their
aligned neighbors. Analysis of the generalized condition number for the corresponding preconditioner
follows that of the classical multigrid algorithm for lattice Laplacian systems.
1 Introduction
A fundamental problem in optimization theory is that of finding solutions x of minimum-norm (cid:107)x(cid:107) to
rectangular linear systems Ax = b. Such solvers have extensive applications, due to the fact that many
practical optimization problems reduce to minimum norm problems. When the norm is Euclidean, classical
iterative solvers such as steepest descent and conjugate gradient methods produce approximately optimal
solutions with residual error (cid:107)Ax − b(cid:107) exponentially small in iteration count, The rate of exponential
convergence depends strongly on the condition number of the square matrix AA∗.
Therefore, the algorithm design process for such problems typically consists entirely of efficiently con-
structing preconditioners: easily computable left-cancellable operators P, for which the transformed prob-
lem PA = Pb is well-conditioned so iterative methods converge rapidly. In a seminal work, Spielman and
Teng[17] present a nearly-linear-time algorithm to construct preconditioners with condition number poly-
logarithmic in problem dimension for a large class of operators A, including Laplacian systems and others
arising from discretization of elliptic PDEs.
There are many fundamental optimization problems that can be expressed as minimum-norm problems
with respect to more general, non-Euclidean norms, including some statistical inference problems and network
flow problems on graphs. In particular, the fundamental maximum-flow and uncapacitated minimum-cost
flow problems reduce respectively to (cid:96)∞ and (cid:96)1 minimum-norm problems in the undirected case.
For such problems, there is presently no known black-box iterative solver analagous to those for (cid:96)2 that
converge exponentially with rate independent of problem dimension. Therefore, one is forced to choose
between interior point methods to obtain zero or exponentially small error[4], at the cost of iteration count
depending strongly on problem dimension, or alternative methods that have error only polynomially small
with respect to iterations[15, 9, 14].
∗Research supported by NSF Grant CCF-1410022
An important idea implicitly used by classical iterative (cid:96)2 solvers is residual recursion: reducing the error
of an approximate solution x by recursively applying the solver to b = b − Ax. Such recursion is implicit
in those solvers due to the fact that in (cid:96)2, the map b (cid:55)→ xopt is linear, with xopt = (AA∗)+b. Therefore,
the classical methods are effectively recursing on every iteration, with no distinction between iterations and
recursive solves. In general norms, an xopt may not be unique, and there is generally no linear operator
mapping b to some xopt.
In this paper, we make two primary contibutions. Our first contribution is to extend analysis of residual
recursion to arbitrary norms. We present a general and modular framework for efficiently solving minimum-
norm problems by composing simple well-known base solvers. The main tool is the composition lemma, de-
scribing the approximation parameters of a solver composed via residual recursion of two black-box solvers.
The resulting parameters depend on a generalization of the condition number to arbitrary norms; therefore,
much like (cid:96)2, the algorithm design process is reduced to constructing good generalized preconditioners. Our
second contribution is to present, as a practical application of those tools, a nearly-linear time approximation
algorithm for the uncapacitated minimum-cost flow problem on undirected graphs. Having established the
former framework, the latter algorithm is entirely specified by the construction of a generalized precondi-
tioner, and analyzed by bounding its generalized condition number.
Our framework builds upon earlier work[16], where we leveraged ideas introduced by Spielman and Teng
to obtain nearly-linear time approximation algorithms for undirected maximum flow. We have since realized
that some of the techniques applied to extend (cid:96)2-flows to (cid:96)∞-flows are in no way specific to flows, and indeed
the composition framework presented here is a generalization of those ideas to non-Euclidean minimum-norm
problems. Moreover, a significantly simpler max-flow algorithm may be recovered using black-box solvers
and general composition as presented here, requiring only the truly flow-specific part of the earlier work: the
congestion approximator, now understood to be a specific instance of a more widely-applicable generalized
preconditioner.
We proceed to discuss those contributions separately in more detail, and outline the corresponding
sections of the paper. We also discuss some related work.
1.1 Minimum Norm Problems
In section 2, we define the minimum norm problem more precisely, and introduce a useful bicriteria notion of
approximation we call (α, β)-solutions. In that notion, α quantifies how much (cid:107)x(cid:107) relatively exceeds (cid:107)xopt(cid:107),
and β quantifies the residual error (cid:107)Ax − b(cid:107), with an appropriate relative scale factor. We may succinctly
describe some frequently used algorithms as (α, β)-solvers; we discuss some examples in section 4, including
steepest descent and conjugate gradient for (cid:96)2, and multiplicative weights for (cid:96)1 and (cid:96)∞.
In section 3, we present the composition lemma, which shows that by composing a (α1, β1)-solver with
a (α2, β2)-solver in a black-box manner, we obtain a (α3, β3)-solver with different parameter tradeoffs. The
composed algorithm's parameters (α3, β3) depends crucially on a generalization of the classical condition
number to rectangular matrices in general norms; Demko[7] provides essentially the exact such generalization
needed, so we briefly recall that definition before stating and proving the composition lemma. The proof of
the composition lemma is not difficult; it follows quite easily once the right definitions of (α, β)-solutions and a
generalized condition number have been established. Nevertheless, the modularity of the composition lemma
proves to quite useful. By recursively composing a solver with itself t times, we may decrease β exponentially
with t, at the cost of a single fixed increase in α. By composing solvers with different parameters, better
parameters may be obtained than any single solver alone. A particularly useful example uses a (M, 0)-solver,
where M is uselessly-large on its own, to terminate a chain of t solvers and obtain zero error. As the final
step in the chain, we get all of the benefits (zero error), and almost none of the costs, as it contributes M 2−t
instead of M to the final solver.
By composing the existing solvers discussed in section 4, we obtain solvers for (cid:96)1 and (cid:96)∞ problems with
residual error exponentially small in t. We state the final composed algorithms parameters. As a result, the
algorithm design problem for such problems is reduced to that of designing a generalized preconditioner for
A. We briefly discuss such preconditioning in section 5.
1.2 Minimum Cost Flow
The second part of this paper applies generalized preconditioning to solve a fundamental problem in net-
work optimization. In the uncapacitated minimum-cost flow problem on undirected graphs, we are given a
connected graph G with m edges, each annotated with a cost, and a specified demand at each vertex. The
2
problem is to find an edge flow with vertex divergences equal to the specified demands, of minimal total cost.
Our main result is a randomized algorithm that, given such a problem, takes −2m1+o(1) time and outputs
a flow meeting the demands with cost at most (1 + )-times that of optimal. We define various flow-related
terms and give a more precise statement of the problem and our result in section 6.
We describe and analyze the algorithm in section 7. Having built up our general tools in earlier sections,
our task is reduced to the design and analysis of a generalized preconditioner for min-cost flow. We begin
by interpreting the edge costs as lengths inducing a metric on the graph, and then apply Bourgain's small-
distortion embedding[3] into (cid:96)1. That is, by paying a small distortion factor in our final condition number,
we may focus entirely on designing a preconditioner for min-cost flow in (cid:96)1 space. The construction of our
precondtitioner is based on an extremely simple hierarchical routing scheme, inspired by a combination of
the multigrid algorithm for grid Laplacians, and the Barnes-Hut algorithm for n-body simulation[2]. The
routing scheme proceeds on a sequence of increasingly dense lattices V0, V1, . . . in (cid:96)1, where Vt are the lattice
points with spacing 2−t, and the input demands are specified on the densest level VT . Starting at level t = T ,
the routing scheme sequentially eliminates the demand supported on Vt \ Vt−1 by pulling the demand from
level t − 1 using a random-rounding based routing: each vertex x ∈ Vt picks a random nearest neighbor in
Vt−1, corresponding to randomly rounding its unaligned coordinates, and then pulls a fractional part of its
demand from that neighbor via any shortest path. Afterwards, all demands on Vt \ Vt−1 are met, leaving
a reduced problem with demands supported on Vt−1. The scheme then recurses on the reduced problem.
At level 0, the demands are supported on the corners of a hypercube, for which the simple routing scheme
of pulling all demand from a uniformly random corner performs well-enough. Having described the routing
scheme, we need not actually carry it out. Instead, our preconditioner crudely estimates the cost of that
routing. Furthermore, while the routing has been described on an infinite lattice, we'll observe that if the
demands are only initially supported on a small set of n vertices, the support remains small throughout.
1.3 Related Work
Spielman and Teng present a nearly-linear time algorithm for preconditioning and solving symmetric diag-
onally dominant matrices, including those of graph Laplacians[17]. The ideas introduced in that work have
led to breakthroughs in various network cut and flow algorithms for the case of undirected graphs. Chris-
tiano et. al. apply the solver as a black-box to approximately solve the maximum flow problem in Θ(m4/3)
time. Madry[13] opens the box and uses the ideas directly to give a family of algorithms for approximating
cut-problems, including a mo(1)-approximation in m1+o(1) time. Both the present author[16] and Kelner et.
al.[11] combine madry's ideas with (cid:96)∞ optimization methods to obtain (1 − )-approxmations to maximum
flow in m1+o(1)−O(1) time. While those two algorithms use similar ideas, the actual path followed to achieve
the result is rather different. Kelner et. al. construct an oblivious routing scheme that is no(1)-competetive,
and then show how to use it to obtain a flow. The algorithm maintains a demand-respecting flow at all times,
and minimizes a potential function measuring edge congestion. In contrast, our earlier algorithm short-cuts
the need to explicitly construct the routing scheme by using Madry's construction directly, maintaining a
flow that is neither demand nor capacity respecting, aiming to minimize a certain potential function that
measures both the congestion and the demand error.
For min-cost flow, earlier work considers the more general directed, capacitated case, with integer ca-
pacities in {1, . . . , U}. The Θ(nm log log U )-time double-scaling algorithm of Ahuja, Goldberg, Orlin, and
Tarjan[1] remained the fastest algorithm for solving min-cost flow for 25 years between its publication and
the Laplacian breakthrough. Shortly after that breakthrough, Daitch and Spielman[5] showed how to use the
Laplacian solver with interior-point methods to obtain an Θ(m3/2 log2 U )-time algorithm. Lee and Sidford
O(1) U ) by a general improvement in interior point methods[12]. We are
further reduce this to Θ(m
not aware of prior work
n log
√
2 Approximate Solutions
Let X ,Y be finite dimensional vector spaces, where X is also a Banach space, and let A ∈ Lin(X ,Y) be
fixed throughout this section. We consider the problem of finding a minimal norm pre-image of a specified
b in the image of A; that is, finding x ∈ X with Ax = b and (cid:107)x(cid:107)X minimal.
Let xopt be an exact solution, with Axopt = b, and (cid:107)xopt(cid:107)X minimal. There are multiple notions of
3
approximate solutions for this problem. The most immediate is x ∈ X with Ax = b and
(cid:107)x(cid:107)
(cid:107)xopt(cid:107) ≤ α
(1)
. We call such x an (α, 0)-solution; we shall be interested in finding (1 + , 0)-solutions for small . A weaker
notion of approximation is obtained by further relaxing Ax = b to Ax (cid:117) b. Quantifying that requires
more structure on Y, so let us further assume Y to also be a Banach space. In that case, we say x is an
(α, β)-solution if equation 1 holds and
(cid:107)Ax − b(cid:107)
(cid:107)A(cid:107)(cid:107)xopt(cid:107) ≤ β
(2)
The practical utility of such solutions depends on the application. However, the weaker notion has the
distinct advantage of being approachable by a larger family of algorithms, such as penalty and dual methods,
by avoiding equality constraints. We discuss such existing algorithms in section 4, but mention that they
typically yield (1, )-solutions after some number of iterations. For (cid:96)2, the iteration dependency on is
O(log(1/)), while for some more general norms it is −O(1).
3 Recursive Composition and Generalized Condition Numbers
We now consider how to trade an increase in α for a decrease in β. Residual recursion suggests a natural
strategy: after finding an (α, β)-solution, recurse on b = b − Ax. More precisely, we define the composition
of two algorithms as follows.
Definition 3.1. Let Fi be an (αi, βi)-algorithm for A, for i ∈ {1, 2}. The composition F2 ◦ F1 takes input
b, and first runs F1 on b to obtain x. Next, setting b = b− Ax, it runs F2 on input b to obtain x. Finally,
it outputs x + x.
Success of composition depends on (cid:107)b(cid:107)Y being small implying b has a small-norm pre-image. The extent
to which that is true is quantified by the condition number of A. The condition number in (cid:96)2 may be defined
several ways, resulting in the same quantity. When generalized to arbitrary norms, those definitions differ.
We recall two natural definitions, following Demko[7].
Definition 3.2. The non-linear condition number of A : X → Y is
(cid:26)(cid:107)A(cid:107)X→Y(cid:107)x(cid:107)X
(cid:107)Ax(cid:107)Y
(cid:27)
: Ax (cid:54)= 0
κX→Y (A) = min
The linear condition number is defined by
κX→Y (A) = min{(cid:107)A(cid:107)X→Y(cid:107)G(cid:107)Y→X : G ∈ Lin(Y,X ) : AGA = A}
Of course, κ ≤ κ. Having defined the condition number, we may now state how composition affects the
approximation parameters.
Theorem 3.3 (Composition). Let Fi be an (αi, βi/κ)-algorithm for A : X → Y, where A has non-linear
condition number κ Then, the composition F2 ◦ F1 is an (α1 + α2β1, β1β2/κ-algorithm for the same problem.
Before proving lemma 3.3, we state two useful corollaries. The first concerns the result of recursively
composing a (α, β/κ)-algorithm with itself.
Corollary 3.4. Let F be a (α, β/κ)-algorithm for β < 1. Let F t be the sequence formed by iterated compo-
sition, with F 1 = F , F t+1 = F t ◦ F . Then, F t is a (α/(1 − β), βt/κ)-algorithm.
Proof. By induction on t. To start, an (α, β/κ)-algorithm is trivially a (α/(1−β), β/κ)-algorithm. Assuming
the claim holds for F t, lemma 3.3 implies F t+1 is a (α + αβ/(1 − β), βt+1/κ)-algorithm.
We observe that to obtain (1 + O(), δ)-solution, only the first solver in the chain need be very accurate.
Corollary 3.5. Let F be a (1+, /2κ-solver and G be a (2, 1/2κ)-solver. Then, Gt◦F is a (1+5, 2−t−1/κ-
solver.
For some problems, there is a very simple (M, 0)-solver known. A final composition with that solver
serves to elimate the error.
Corollary 3.6. Let F be a (1 + , δ/κ)-solver and G be a (M, 0)-solver. Then G◦ F is a (1 + (1 + δM ), 0)-
solver.
4
3.1 Proof of lemma 3.3
Since F1 is an (α1, β1/κ)-algorithm, we have
(cid:107)x(cid:107) ≤ α1(cid:107)xopt(cid:107)
(cid:107)b(cid:107)
(cid:107)xopt(cid:107)
(cid:107)A(cid:107) ≤ β1
κ
(cid:107)xopt(cid:107) ≤ κ
(cid:107)b(cid:107)
(cid:107)A(cid:107) ≤ β1(cid:107)xopt(cid:107)
By the definition of κ,
Since F2 is an (α2, β2/κ)-algorithm,
(cid:107)x(cid:107) ≤ α2(cid:107)xopt(cid:107) ≤ α2β1(cid:107)x(cid:107)opt
≤ β2
κ
(cid:107)xopt(cid:107) ≤ β1β2
κ
(cid:107)Ax − b(cid:107)
(cid:107)A(cid:107)
The conclusion follows from (cid:107)x + x(cid:107) ≤ (cid:107)x(cid:107) + (cid:107)x(cid:107).
4 Solvers for (cid:96)p
The recursive composition technique takes existing (α, β)-approximation algorithms and yields new algo-
In this section, we discuss the parameters of existing
rithms with different approximation parameters.
well-known base solvers.
Let A : Rm → Rn be linear and fixed throughout this section. To concisely describe results and ease
comparison, for a norm (cid:107) · (cid:107)X on Rm and a norm (cid:107) · (cid:107)Y in Rn, we write (α, β)X→Y to denote an (α, β)
solution with respect to A : X → Y.
The classical (cid:96)2 solvers provide a useful goalpost for comparison.
Theorem 4.1 ([4]).
The steepest-descent algorithm produces a (1, δ)2→2-solution after O(κ2→2(A)2 log(1/δ)) simple iterations
The conjugate gradient algorithm produces a (1, δ)2→2-solution after O(κ2→2(A) log(1/δ)) simple iterations.
There are many existing algorithms for p-norm minimization, and composing them yields algorithms with
exponentially small error. For this initial manuscript, we simply state the (cid:96)1 version required for min-cost
flow. The general cases consist of describing the many existing algorithms[14] in terms of (α, β)-solvers.
Theorem 4.2. There is a (1 + , δ)1→1-solver with simple iterations totalling
O(cid:0)κ1→1(A)2 log(m)(cid:0)−2) + log(δ−1)(cid:1)(cid:1)
O(cid:0)κ∞→∞(A)2 log(n)(cid:0)−2 + log(δ−1)(cid:1)(cid:1)
Theorem 4.3. There is a (1 + , δ)∞→∞-solver with simple iterations totalling
The preceding theorems follow by combining the multiplicative weights algorithm[15, 9] with the com-
position lemma. The former algorithm applies in a more general online setting; for our applications, the
algorithm yields the "weak" approximation algorithms that shall be composed.
A support oracle for a compact-convex set C takes input y and returns some x ∈ C maximizing x · C.
Let (cid:52)n be the unit simplex in Rn (i.e., the convex hull of the standard basis).
The multiplicative weights method finds approximate solutions to the saddle point problem,
In particular, it obtains an additive -approximation in O(ρ2−2 log n) iterations with each iteration taking
O(n) time plus a call to a support oracle for C.
max
w∈(cid:52)n
min
z∈C
w · z
5
Let Bp be the unit ball in (cid:96)p. A point w ∈ B1 is represented as w = w+ − w− for (w+, w−) ∈ (cid:52)2n. The
base solvers follow by considering the saddle-point problems,
max
y∗∈b1
max
x∈b1
min
x∈b∞
min
y∗∈b∞
y∗ · (Ax − µb)
y∗ · (Ax − µb)
where µ is scaled via binary-search as needed. The search adds a factor O(log κ) to the complexity. This
factor may be avoided by using (cid:96)p-norm for p = log n and p = (log m)/ regularization, respectively, instead.
5 Generalized Preconditioning
Let A : X → Y be linear. The minimum-norm problem for A does not intrinsically require Y to be normed;
nor does the definition of a (α, 0)-solution. Thus, an algorithm designer seeking to solve a class of minimum
norm problems may freely choose how to norm Y. The best choice requires balancing two factors. First, the
norm should be sufficiently "simple" that (1, ) solutions for polynomially-small are easy to find. Second,
the norm should be chosen so that κ(A)X→Y is small. If only the latter constraint existed, the ideal choice
would be the optimal A pre-image norm,1
(cid:107)b(cid:107)opt(A) = min{(cid:107)x(cid:107)X : Ax = b}
By construction, κ(A)X→opt(A) = 1. Of course, simply evaluating that norm is equivalent to the original
problem we aim to solve. It follows that Y should be equipped with the relatively closest norm to opt(A)
that can be efficiently minimized.
Following the common approach in (cid:96)2, when X = (cid:96)p, we propose to choose a generalized pre-conditioner
P : Y → (cid:96)p injective and consider the norm (cid:107)Pb(cid:107)p. Again, P should be chosen such that κ(PA)p→p is
small, and P, P∗ are easy to compute.
6 Graph Problems
In this section we discuss some applications of generalized preconditioning to network flow problems. Let
G = (V, E) be an undirected graph with n vertices and m edges. While undirected, we assume the edges
are oriented arbitrarily. We denote by V∗,E∗ the spaces of real-valued functions f : V → R and f : E → R
on vertices and edges, respectively. We denote by V,E the corresponding dual spaces of demands and flows.
For S ⊆ V , we write 1S ∈ V∗ for the indicator function on S.
The discrete derivative operator D∗ : V∗ → E∗ is defined by (D∗f )(xy) = f (y) − f (x) for each oriented
edge xy. The constant function has zero derivative, D∗1V = 0. The (negative) adjoint −D : E → V is
the discrete divergence operator ; for a flow ∈ E and vertex x ∈ V , −(D)(x) is the net quantity being
transported away from vertex x by the flow.
A single-commodity flow problem is specified by a demand vector b ∈ V, and requires finding a flow ∈ E
satisfying D = b, minimizing some cost function. When the cost-function is a norm on E, the problem is a
minimum-norm pre-image problem. Assuming G is connected, b has a pre-image iff the total demand 1V · b
is zero. We hereafter assume G to be connected and total demands equal to zero.
Having established the generalized preconditioning framework, we may now design fast algorithms for
fundamental network flow problems by designing generalized preconditioners for the corresponding minimum-
norm problems. As a result, we obtain nearly-linear time algorithms for max-flow and uncapacitated min-cost
flow in undirected graphs. The max-flow algorithm was presented previously[16], and involved a combination
of problem-specific definitions (e.g. congestion approximators) and subroutines (e.g. gradient-based (cid:96)∞
minimization). The brief summary in subsection 6.1 shows how the same result may be immediately obtained
by combining theorem ?? with only small part of the earlier flow-specific work[16]: the actual construction
of the preconditioner.
In section 7, we present a nearly-linear time algorithm for uncapacitated min-cost flows, another fun-
damental network optimization problem. Once again, all that is needed is description and analysis of the
preconditioner.
1We remark this is only defined on the image space of A, rather than the entire codomain. However, we only consider b in
that image space
6
6.1 Max-Flow
norm on E∗ by (cid:107)f(cid:107)c∗ =(cid:80)
e c(e)f (e); the dual congestion norm on E is defined by (cid:107)(cid:107)c = maxe
(e)
c(e) .
A capacitated graph associates with each edge e, a capacity c(e) > 0. Given capacities, we define the capacity
The boundary capacity of a set S ⊆ V is (cid:107)D∗1S(cid:107)c, which we abbreviate as c(∂S).
The single-commodity flow problem with capacity norm is undirected maximum flow problem, and is a
fundamental problem in network design and optimization. Following the work of Spielman and Teng for
(cid:96)2-flows[17], and Madry[13] for cut problems, the author has obtained a nearly-linear time algorithm for
maximum flow[16]. We have since realized that only certain parts of the ideas introduced in that paper
are actually specific to maximum-flow, and the present work has been obtained through generalizing the
other parts. The (previously presented as) flow-specific composition and optimization parts of that paper
are entirely subsumed by the earlier sections of this paper. The truly flow-specific part of that work required
to complete the algorithm is the generalized preconditioner construction, which we have previously called a
congestion-approximator. We briefly describe the main ideas of the construction, and some simple illustrative
examples. For the full details, we refer the reader to the third section of [16].
Let F ⊆ 2V be a family of subsets. Such a family, together with a capacity norm on E, induces a cut-
1S·b
congestion semi-norm on V via (cid:107)b(cid:107)c(F) = maxS∈F
c(∂S) . That is, for each set in the family, consider the
ratio of the total aggregate demand in that set to the total capacity of all edges entering that set. If the
indicator functions of sets in F span V, then it is a norm. Note also that we may write (cid:107)b(cid:107)c(F) = (cid:107)Pb(cid:107)∞,
where P : Rn → RF is defined by (Pb)S = 1S·b
c(∂S) .
If F is taken to be the full power set F = 2V , the max-flow, min-cut theorem is equivalent to the statement
that the non-linear condition number κc→c(F)(D∗) is exactly one. If F is taken to be the family of singleton
sets, the non-linear condition number is the inverse of the combinatorial conductance of G. That is, for high-
conductance graphs, composition yields a fast algorithm requiring only a simple diagonal preconditioner.
This is closely analagous to the situation for (cid:96)2, where diagonal preconditioning yields fast algorithms for
graphs of large algebraic conductance.
A particularly illustrative example consists of a unit-capacity 2t × 2t 2-D grid. With low (Θ(2−t) =
√
n) conductance, the singleton family F performs poorly. A simple but dramatically better family
Θ(1/
consists of taking F to be all power-of-two size, power-of-two aligned subgrids. The latter family yields linear
condition number O(t) = O(log n). Furthermore, computing the aggregate demand in all such sets requires
only O(n) time, with each sub-grid aggregate equal to the sum of its four child aggregates. This algorithm
is analagous to an (cid:96)∞ version of multi-grid [?]. Note that the indicator "step" functions 1S are different from
the bilinear "pyramid" functions used by multigrid.
Theorem 6.1 ([16]). There is an algorithm that given a capacitated graph (G, c) with n vertices and m edges,
takes m1+o(1) time and outputs a data structure of size n1+o(1) that efficiently represents a preconditioner P
with κ(PD)c→∞ ≤ no(1). Given the data structure, P and P∗ can be applied in n1+o(1) time.
Let C : Rm → E be the diagonal capacity matrix; (Cx)(e) = c(e)x(e). The nearly-linear time algorithm
follows by using the preconditioner P in the preceding theorem, applying theorem ?? to the problem of
minimizing (cid:107)x(cid:107)∞ subject to PDCx = b. A final minor flow-specific part is needed to achieve zero error:
the simple (m, 0)-approximation algorithm that consists of routing all flow through a maximum-capacity
spanning tree.
7 Uncapacitated Min-Cost Flow
the space of flows E is defined by (cid:107)(cid:107)l = (cid:80)
In this section, we present a nearly-linear time algorithm for the uncapacitated minimum-cost flow problem
on undirected graphs. A cost graph associates a cost l(e) > 0 with each edge e in G. The cost norm on
e (cid:107)(e)(cid:107)l(e). The undirected uncapacitated minimum-cost flow
problem is specified by the graph G, lengths (cid:96), and demands b ∈ V, and consists of finding ∈ E with
D = b minimizing (cid:107)(cid:107)l. Note that, unlike max-flow, the single-source uncapacitated min-cost flow problem
is significantly easier than the distributed source case: the optimal solution is to route along a single-source
shortest path away from the source. The problem only becomes interesting with distributed sources and
sinks.
As a new application of generalized preconditioning, we obtain a nearly-linear time algorithm for approx-
imately solving this problem.
7
Theorem 7.1. There is a randomized algorithm that, given a length-graph G = (V, E, l), takes m1+o(1)
and outputs a (1 + , 0)-solution to the undirected uncapacitated min-cost flow problem.
2
time
As expected, the bulk of the algorithm lies in constructing a good preconditioner. After introducing
some useful definitions and lemmas, we spend the majority of this section constructing and analyzing a
preconditioner for this problem.
Theorem 7.2. There is a randomized algorithm that, given a length-graph G, takes O(m log2 n + n1+o(1))
time and outputs a n1+o(1) × n matrix P with κlto∞(PD) ≤ no(1). Every column of P has no(1) non-zero
entries.
Dual to cost is the stretch norm on E∗ is (cid:107)f(cid:107)l∗ = maxe f (e)/l(e). We say a function φ ∈ V is L-Lipschitz
iff (cid:107)D∗φ(cid:107)l∗ ≤ L; that is, φ(x) − φ(y) ≤ Ll(xy) for all edges xy. The dual problem is to maximize φ · b over
1-Lipschitz φ ∈ V.
A length function l induces an intrinsic metric d : V × V → R, with distances determined by shortest
paths. The preceding definition of Lipschitz is equivalent to φ(x) − φ(y) ≤ Ld(x, y) for all x, y ∈ V . That
is, the Lipschitz constant of φ depends only on the metric, and is otherwise independent of the particular
graph or edge lengths inducing that metric. As the dual problem depends only on the induced metric, so
must the value of the min-cost flow. Therefore, we may unambiguously write
(cid:107)b(cid:107)opt(d) = max{φ · b : φ is 1-Lipschitz w.r.t. d}
The monotonicity of cost with distance follows immediately.
Lemma 7.3. If d(x, y) ≤ d(x, y) for all x, y ∈ V, then for all b ∈ V⊥1,
(cid:107)b(cid:107)opt(d) ≤ (cid:107)b(cid:107)opt(d)
We recall that in some cases, the identity operator is a good preconditioner; for max-flow, such is the case
in constant-degree expanders. For min-cost flow, the analagous case is a metric where distances between any
pair of points differ relatively by small factors.
Definition 7.4. Let U ⊆ V . We say U is r-separated if d(x, y) ≥ r for all x, y ∈ U with x (cid:54)= y. We say U
is R-bounded if d(x, y) ≤ R for all x, y ∈ U .
Lemma 7.5. Let U ⊂ V be R-bounded and r-separated with respect to d. Let b ∈ V⊥1 be demands supported
on U . Then,
r
2
(cid:107)b(cid:107)1 ≤ (cid:107)b(cid:107)opt(d) ≤ R
2
(cid:107)b(cid:107)1
Proof. Note r
the star graph with edge lengths r/2, with a leaf for each x ∈ U ). The inequalities follow by monotonicity.
2(cid:107)b(cid:107)1 is exactly the min-cost of routing b in the metric where all distances are exactly r (consider
To complete the proof of theorem 7.1, assuming theorem 7.2, we also need a simple approximation
algorithm that yields a poor approximation, but with zero error. This is easily achieved by the algorithm
that routes all flow through a minimum cost (i.e., total length) spanning tree T .
Lemma 7.6. Let T be a minimum cost spanning tree with respect to l on G. Consider the algorithm that,
given b ∈ V⊥1, outputs the flow ∈ E routing b using only the edges in T . Then,
(cid:107)(cid:107) ≤ n(cid:107)b(cid:107)opt(d)
Proof. The flow on T is unique, and therefore, optimal for the metric d induced by paths restricted to lie in
T . As T is a minimum cost spanning tree, d(x, y) ≤ nd(x, y).
We now prove theorem 7.1. Let L : Rm → E be the diagonal length matrix (Lx)(e) = (cid:96)(e)x(e), and let P
be the preconditioner from theorem 7.2. Computing P takes O(m log2 n) + n1+o(1) time. The minimum-cost
flow problem is equivalent to the problem of minimizing (cid:107)x(cid:107)1 subject to PDL−1x = Pb. As P has n1+o(1)
non-zero entries and κ1→1(PDL−1) ≤ no(1), the total running time is m1+o(1)
.
2
8
8 Preconditioning Min-Cost Flow
A basic concept in metric spaces is that of an embedding: ψ : V → V , with respective metrics d, d. An
embedding has distortion L if for some µ > 0,
1 ≤ µ
d(ψ(x), ψ(y))
d(x, y)
≤ L
Metric embeddings are quite useful for preconditioning; if d embeds into d with distortion L, then the
monotonicity lemma implies the relative costs of flow problems differ by a factor L. It follows that if we can
construct a preconditioner P for the min-cost flow problem on d, the same P may be used to precondition
min-cost flow on d, with condition number at most L-factor worse.
Our preconditioner uses embeddings in two ways. The preconditioner itself is based on a certain hi-
erarchical routing scheme. Hierarchical routing schemes based on embeddings into tree-metrics have been
studied extensively(see e.g. [8]), and immediately yield preconditioners with polylogarithmic condition num-
ber. However, we are unable to efficiently implement those schemes in nearly-linear time. Instead, we apply
Bourgain's O(log n)-distortion embedding[3] into (cid:96)1. That is, by paying an O(log n) factor in condition
number, we may completely restrict our attention to approximating min-cost flow in (cid:96)1 space.
Theorem 8.1 (Bourgain's Embedding[3]). Any n-point metric space embeds into (cid:96)p space with distortion
O(log n). If d is the intrinsic metric of a length-graph G = (V, E) with m edges, there is a randomized
algorithm that takes O(m log2 n) time and w.h.p. outputs such an embedding, with dimension O(log2 n).
√
log n), incuring another distortion factor of exp(O(
The complexity of constructing and evaluating our preconditioner is exponential in the embedding's di-
mension. Therefore, we reduce the dimension to Θ(
√
log n)).
Theorem 8.2 (Johnson-Lindenstrauss Lemma[10, 6]). Any n points in Euclidean space embed into k-
dimensional Euclidean space with distortion nO(1/k), for k ≤ Ω(log n). In particular, k = O(log n) yields
√
O(1) distortion, and k = O(
an embedding in O(nd) time.
If the original points are in d > k dimensions, there is a randomized algorithm that w.h.p outputs such
√
log n) yields exp(O(
log n)) distortion.
√
k = O(
of exp(O(
log n)).
Our reduction consists of first embedding the vertices into (cid:96)2 via theorem 8.1, reducing the dimension to
√
log n) using lemma 8.2, and then using the identity embedding of (cid:96)2 into (cid:96)1, for a total distortion
To estimate routing costs in (cid:96)1, we propose an extremely simple hierarchical routing algorithm in (cid:96)1, and
show it achieves a certain competetive ratio. The routing scheme applies to the exponentially large graph
corresponding to a lattice discretization of the k-dimensional cube in (cid:96)1; after describing such a scheme, we
show that if the demands in that graph are supported on n vertices, the distributed routing algorithm finds
no demand to route in most of the graph, and the same routing algorithm may be carried out in a k − d-tree
instead of the entire lattice.
On the lattice, the routing algorithm consists of sequentially reducing the support set of the demands
from a lattice with spacing 2−t−1 to a lattice with spacing 2−t, until all demand is supported on the corners
of a cube containing the original demand points. On each level, the demand on a vertex not appearing in
the coarser level is eliminated by pushing its demand to a distribution over the aligned points nearby.
8.1 Lattice Algorithm
For t ∈ Z, let Vt = (2−tZ)k be the lattice with spacing 2−t. We design a routing scheme and preconditioner for
min-cost flow in (cid:96)1, with initial input demands bT supported on a bounded subset of VT . For t = T, T −1, . . .,
the routing scheme recursively reduces a min-cost flow problem with demands supported on Vt to a min-cost
flow problem with demands supported on the sparser lattice Vt−1. Given demands bt supported on Vt, the
scheme consists of each vertex x ∈ Vt pulling its demand bt(x) uniformly from the closest points to x in Vt−1
along any shortest path, the lengths of which are at most k2−t. For example, a point x = (x1, . . . , xk) ∈ V1
with j ≤ k non-integer coordinates and demand b1(x) pulls b1(x)2−j units of flow from each of the 2j points
in V0 corresponding to rounding those j coordinates in any way. By construction, all points in Vt \ Vt−1
have their demands met exactly, and we are left with a reduced problem residual demands bt−1 supported on
Vt−1. As we show, eventually the demands are supported entirely on the 2k corners of a hypercube forming a
9
bounding box of the original support set, at which point the the simple scheme of distributing the remaining
demand uniformly to all corners is a k-factor approximation.
We do not actually carry out the routing; rather, we state a crude upper-bound on its cost, and then
show that bound itself exceeds the true min-cost by a factor of some κ. For the preconditioner, we shall only
require the sequence of reduced demands bt, bt−1, . . ., and not the flow actually routing them.
Theorem 8.3.
(cid:107)bt(cid:107)opt ≤ t(cid:88)
s=0
k2−s(cid:107)bs(cid:107)1 ≤ 2k(t + 1)(cid:107)bt(cid:107)opt
Assuming theorem 8.3 holds, we take our preconditioner to be a matrix P such that,
T(cid:88)
(cid:107)PbT(cid:107)1 =
k2−t(cid:107)bt(cid:107)1
t=0
The proof of theorem 8.3 uses two essential properties of the routing scheme described. The first is that
the cost incurred in the reduction from bt to bt−1 is not much larger than the min-cost of routing bt. The
second is the reduction does not increase costs. That is, the true min-cost for routing the reduced problem
bt−1 is at most that of routing the original demands bt.
Lemma 8.4. Let bt−1 be the reduced demands produced when the routing scheme is given bt. Then,
1. The cost incurred by that level of the scheme is at most k2−t(cid:107)bt(cid:107)1 ≤ 2k(cid:107)bt(cid:107)opt
2. The reduction is non-increasing with respect to min-cost: (cid:107)bt−1(cid:107)opt ≤ (cid:107)bt(cid:107)opt
Assuming lemma 8.4, we now prove 8.3. We argue the sum is an upper bound on the total cost incurred
by the routing scheme, when applied to bt. The left inequality follows immediately, as the true min-cost is
no larger.
The total cost of the scheme is at most the cost incurred on each level, plus the cost of the final routing
of b0. Lemma 8.4(a) accounts for all terms above s = 0. For the final routing, the demands b0 are supported
in {0, 1}k, so the cost of routing all demand to a random corner is at most 1
2 k(cid:107)b0(cid:107)1.
For the right inequality, we observe
t(cid:88)
k2−t(cid:107)bt(cid:107)1 ≤ t(cid:88)
2k(cid:107)bs(cid:107)opt
The inequality follows termwise, using lemma 8.4(a) for s ≥ 1; the s = 0 case follows from lemma 7.5 because
V0 is 1-separated.
s=0
s=0
8.2 Proof of Lemma 8.4
For the first part, each vertex x ∈ Vt routes its demand bt(x) to the closest points in Vt−1 any shortest path.
As such paths have length at most k2−t, each point x contributes at most k2−tbt(x).
For the second part, by scale-invariance it suffices to consider t = 1. Moreover, it suffices to consider
the case where b1 consists of a unit demand from between two points x, y ∈ V1 with (cid:107)x − y(cid:107)1 = 1
2 . To see
this, we observe any flow routing arbitrary demands b1 consists of a sum of such single-edge flows. As the
reduction is linear, if the cost of each single-edge demand-pair is not increased, then the cost their sum is
not increased. By symmetry, it suffices to consider x, y with x = (0, z) and y = ( 1
2 , z). Then, the reduction
distributes x's demand to a distribution over (0, z(cid:48)) where z(cid:48) ∼ Z(cid:48); y's demand is split over (0, z(cid:48)) and (1, z(cid:48))
where z(cid:48) ∼ Z(cid:48). Therefore, half of the demand at (0, z(cid:48)) is cancelled, leaving the residual problem of routing
1/2 unit from (0, z(cid:48)) to (1, z(cid:48)). That problem has cost 1/2, by routing each fraction directly from (0, z(cid:48)) to
(1, z(cid:48)).
References
[1] Ravindra K Ahuja, Andrew V Goldberg, James B Orlin, and Robert E Tarjan. Finding minimum-cost
flows by double scaling. Mathematical programming, 53(1-3):243 -- 266, 1992.
10
[2] Josh Barnes and Piet Hut. A hierarchical o (n log n) force-calculation algorithm. nature, 324(6096):446 --
449, 1986.
[3] J. Bourgain. On lipschitz embedding of finite metric spaces in hilbert space. Israel Journal of Mathe-
matics, 52(1):46 -- 52, 1985.
[4] Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, New York,
NY, USA, 2004.
[5] Samuel I Daitch and Daniel A Spielman. Faster approximate lossy generalized flow via interior point
In Proceedings of the fortieth annual ACM symposium on Theory of computing, pages
algorithms.
451 -- 460. ACM, 2008.
[6] Sanjoy Dasgupta and Anupam Gupta. An elementary proof of a theorem of johnson and lindenstrauss.
Random Structures & Algorithms, 22(1):60 -- 65, 2003.
[7] Stephen Demko. Condition numbers of rectangular systems and bounds for generalized inverses. Linear
Algebra and its Applications, 78:199 -- 206, 1986.
[8] Jittat Fakcharoenphol, Satish Rao, and Kunal Talwar. A tight bound on approximating arbitrary metrics
by tree metrics. In Proceedings of the thirty-fifth annual ACM symposium on Theory of computing, pages
448 -- 455. ACM, 2003.
[9] Yoav Freund and Robert E Schapire. Adaptive game playing using multiplicative weights. Games and
Economic Behavior, 29(1):79 -- 103, 1999.
[10] William B Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space.
Contemporary mathematics, 26(189-206):1, 1984.
[11] Jonathan A Kelner, Yin Tat Lee, Lorenzo Orecchia, and Aaron Sidford. An almost-linear-time algorithm
for approximate max flow in undirected graphs, and its multicommodity generalizations. In Proceedings
of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 217 -- 226. Society for
Industrial and Applied Mathematics, 2014.
[12] Yin Tat Lee and Aaron Sidford. Path finding methods for linear programming: Solving linear programs
in o (vrank) iterations and faster algorithms for maximum flow. In Foundations of Computer Science
(FOCS), 2014 IEEE 55th Annual Symposium on, pages 424 -- 433. IEEE, 2014.
[13] Aleksander Madry. Fast approximation algorithms for cut-based problems in undirected graphs.
In
Foundations of Computer Science (FOCS), 2010 51st Annual IEEE Symposium on, pages 245 -- 254.
IEEE, 2010.
[14] Yu Nesterov. Smooth minimization of non-smooth functions. Math. Program., 103(1):127 -- 152, May
2005.
[15] Serge A. Plotkin, David B. Shmoys, and Eva Tardos. Fast approximation algorithms for fractional
packing and covering problems. Mathematics of Operations Research, 20:257 -- 301, 1995.
[16] Jonah Sherman. Nearly maximum flows in nearly linear time. In Foundations of Computer Science
(FOCS), 2013 IEEE 54th Annual Symposium on, pages 263 -- 269. IEEE, 2013.
[17] Daniel A. Spielman and Shang-Hua Teng. Nearly-linear time algorithms for preconditioning and solving
symmetric, diagonally dominant linear systems. CoRR, abs/cs/0607105, 2006.
11
|
1904.08255 | 1 | 1904 | 2019-04-17T13:10:38 | Online Matching with General Arrivals | [
"cs.DS"
] | The online matching problem was introduced by Karp, Vazirani and Vazirani nearly three decades ago. In that seminal work, they studied this problem in bipartite graphs with vertices arriving only on one side, and presented optimal deterministic and randomized algorithms for this setting. In comparison, more general arrival models, such as edge arrivals and general vertex arrivals, have proven more challenging and positive results are known only for various relaxations of the problem. In particular, even the basic question of whether randomization allows one to beat the trivially-optimal deterministic competitive ratio of $\frac{1}{2}$ for either of these models was open. In this paper, we resolve this question for both these natural arrival models, and show the following.
1. For edge arrivals, randomization does not help --- no randomized algorithm is better than $\frac{1}{2}$ competitive.
2. For general vertex arrivals, randomization helps --- there exists a randomized $(\frac{1}{2}+\Omega(1))$-competitive online matching algorithm. | cs.DS | cs | Online Matching with General Arrivals
Buddhima Gamlath
EPFL
Michael Kapralov∗
EPFL
[email protected]
[email protected]
Andreas Maggiori
EPFL
Ola Svensson
EPFL
David Wajc†
CMU
[email protected]
[email protected]
[email protected]
Abstract
The online matching problem was introduced by Karp, Vazirani and Vazirani nearly three
decades ago. In that seminal work, they studied this problem in bipartite graphs with vertices
arriving only on one side, and presented optimal deterministic and randomized algorithms for
this setting.
In comparison, more general arrival models, such as edge arrivals and general
vertex arrivals, have proven more challenging and positive results are known only for various
relaxations of the problem.
In particular, even the basic question of whether randomization
allows one to beat the trivially-optimal deterministic competitive ratio of 1/2 for either of these
models was open. In this paper, we resolve this question for both these natural arrival models,
and show the following.
1. For edge arrivals, randomization does not help -- no randomized algorithm is better than
1/2 competitive.
2. For general vertex arrivals, randomization helps -- there exists a randomized (1/2 + Ω(1))-
competitive online matching algorithm.
9
1
0
2
r
p
A
7
1
]
S
D
.
s
c
[
1
v
5
5
2
8
0
.
4
0
9
1
:
v
i
X
r
a
∗Supported in part by ERC Starting Grant 759471.
†Work done in part while the author was visiting EPFL. Supported in part by NSF grants CCF-1618280, CCF-
1814603, CCF-1527110, NSF CAREER award CCF-1750808, and a Sloan Research Fellowship.
1
Introduction
Matching theory has played a prominent role in the area of combinatorial optimization, with many
applications [21, 23]. Moreover, many fundamental techniques and concepts in combinatorial op-
timization can trace their origins to its study, including the primal-dual framework [19], proofs
of polytopes' integrality beyond total unimodularity [8], and even the equation of efficiency with
polytime computability [9].
Given the prominence of matching theory in combinatorial optimization, it comes as little sur-
prise that the maximum matching problem was one of the first problems studied from the point
of view of online algorithms and competitive analysis. In 1990, Karp, Vazirani, and Vazirani [18]
introduced the online matching problem, and studied it under one-sided bipartite arrivals. For such
arrivals, Karp et al. noted that the trivial 1/2-competitive greedy algorithm (which matches any
arriving vertex to an arbitrary unmatched neighbor, if one exists) is optimal among deterministic
algorithms for this problem. More interestingly, they provided an elegant randomized online al-
gorithm for this problem, called ranking, which achieves an optimal (1 − 1/e) competitive ratio.
(This bound has been re-proven many times over the years [2, 6, 7, 11, 13].) Online matching and
many extensions of this problem under one-sided bipartite vertex arrivals were widely studied over
the years, both under adversarial and stochastic arrival models. See recent work [5, 15, 16, 17] and
the excellent survey of Mehta [22] for further references on this rich literature.
Despite our increasingly better understanding of one-sided online bipartite matching and its
extensions, the problem of online matching under more general arrival models, including edge
arrivals and general vertex arrivals, has remained staunchly defiant, resisting attacks. In particular,
the basic questions of whether the trivial 1/2 competitive ratio is optimal for the adversarial edge-
arrival and general vertex-arrival models have remained tantalizing open questions in the online
algorithms literature. In this paper, we answer both of these questions.
1.1 Prior Work and Our Results
Here we outline the most relevant prior work, as well as our contributions. Throughout, we say an
algorithm (either randomized or fractional) has competitive ratio α, or equivalently is α-competitive,
if the ratio of the algorithm's value (e.g., expected matching size, or overall value,Pe xe) to OPT
is at least α 6 1 for all inputs and arrival orders. As is standard in the online algorithms literature
on maximization problems, we use upper bounds (on α) to refer to hardness results, and lower
bounds to positive results.
Edge Arrivals. Arguably the most natural, and the least restricted, arrival model for online
matching is the edge arrival model. In this model, edges are revealed one by one, and an online
matching algorithm must decide immediately and irrevocably whether to match the edge on arrival,
or whether to leave both endpoints free to be possibly matched later.
On the hardness front, the problem is known to be strictly harder than the one-sided vertex
arrival model of Karp et al. [18], which admits a competitive ratio of 1− 1/e ≈ 0.632. In particular,
1+ln 2 ≈ 0.591 for this problem, recently improved by
Epstein et al. [10] gave an upper bound of
Huang et al. [16] to 2−√2 ≈ 0.585. (Both bounds apply even to online algorithms with preemption;
i.e., allowing edges to be removed from the matching in favor of a newly-arrived edge.) On the
positive side, as pointed out by Buchbinder et al. [3], the edge arrival model has proven challenging,
and results beating the 1/2 competitive ratio were only achieved under various relaxations, including:
random order edge arrival [14], bounded number of arrival batches [20], on trees, either with or
without preemption [3, 24], and for bounded-degree graphs [3]. The above papers all asked whether
1
1
there exists a randomized (1/2 + Ω(1))-competitive algorithm for adversarial edge arrivals (see also
Open Question 17 in Mehta's survey [22]).
In this work, we answer this open question, providing it with a strong negative answer.
In
particular, we show that no online algorithm for fractional matching (i.e., an algorithm which
immediately and irrevocably assigns values xe to edge e upon arrival such that ~x is in the fractional
matching polytope P = {~x > ~0 Pe∋v xe 6 1 ∀v ∈ V }) is better than 1/2 competitive. As any
randomized algorithm induces a fractional algorithm with the same competitive ratio, this rules
out any randomized online matching algorithm which is better than deterministic algorithms.
Theorem 1.1. No fractional online algorithm is 1/2 + Ω(1) competitive for online matching under
adversarial edge arrivals, even in bipartite graphs.
This result shows that the study of relaxed variants of online matching under edge arrivals is
not only justified by the difficulty of beating the trivial bound for this problem, but rather by its
impossibility.
In the online matching problem under vertex arrivals, vertices are
General Vertex Arrivals.
revealed one at a time, together with their edges to their previously-revealed neighbors. An online
matching algorithm must decide immediately and irrevocably upon arrival of a vertex whether
to match it (or keep it free for later), and if so, who to match it to. The one-sided bipartite
problem studied by Karp et al. [18] is precisely this problem when all vertices of one side of a
bipartite graph arrive first. As discussed above, for this one-sided arrival model, the problem is
thoroughly understood (even down to lower-order error terms [11]). Wang and Wong [25] proved
that general vertex arrivals are strictly harder than one-sided bipartite arrivals, providing an upper
bound of 0.625 < 1 − 1/e for the more general problem, later improved by Buchbinder et al. [3] to
3+φ2 ≈ 0.593. Clearly, the general vertex arrival model is no harder than the online edge arrival
model but is it easier ? The answer is "yes" for fractional algorithms, as shown by combining our
Theorem 1.1 with the 0.526-competitive fractional online matching algorithm under general vertex
arrivals of Wang and Wong [25]. For integral online matching, however, the problem has proven
challenging, and the only positive results for this problem, too, are for various relaxations, such as
restriction to trees, either with or without preemption [3, 4, 24], for bounded-degree graphs [3], or
(recently) allowing vertices to be matched during some known time interval [15, 16].
2
We elaborate on the last relaxation above.
In the model recently studied by Huang et al.
[15, 16] vertices have both arrival and departure times, and edges can be matched whenever both
their endpoints are present. (One-sided vertex arrivals is a special case of this model with all online
vertices departing immediately after arrival and offline vertices departing at ∞.) We note that any
α-competitive online matching under general vertex arrivals is α-competitive in the less restrictive
model of Huang et al. As observed by Huang et al., for their model an optimal approach might
as well be greedy; i.e., an unmatched vertex v should always be matched at its departure time if
possible. In particular, Huang et al. [15, 16], showed that the ranking algorithm of Karp et al. is
optimal in this model, giving a competitive ratio of ≈ 0.567. For general vertex arrivals, however,
ranking (and indeed any maximal matching algorithm) is no better than 1/2 competitive, as is
readily shown by a path on three edges with the internal vertices arriving first. Consequently, new
ideas and algorithms are needed.
The natural open question for general vertex arrivals is whether a competitive ratio of (1/2+Ω(1))
is achievable by an integral randomized algorithm, without any assumptions (see e.g., [25]). In this
work, we answer this question in the affirmative:
Theorem 1.2. There exists a (1/2 + Ω(1))-competitive randomized online matching algorithm for
general adversarial vertex arrivals.
2
1.2 Our Techniques
Edge Arrivals. All prior upper bounds in the online literature [3, 10, 11, 16, 18] can be rephrased
as upper bounds for fractional algorithms;
i.e., algorithms which immediately and irrevocably
assign each edge e a value xe on arrival, so that ~x is contained in the fractional matching polytope,
P = {~x > ~0 Pe∋v xe 6 1 ∀v ∈ V }. With the exception of [3], the core difficulty of these hard
instances is uncertainty about "identity" of vertices (in particular, which vertices will neighbor
which vertices in the following arrivals). Our hardness instances rely on uncertainty about the
"time horizon". In particular, the underlying graph, vertex identifiers, and even arrival order are
known to the algorithm, but the number of edges of the graph to be revealed (to arrive) is uncertain.
Consequently, an α-competitive algorithm must accrue high enough value up to each arrival time
to guarantee a high competitive ratio at all points in time. As we shall show, for competitive ratio
1/2 + Ω(1), this goal is at odds with the fractional matching constraints, and so such a competitive
ratio is impossible. In particular, we provide a family of hard instances and formulate their prefix-
competitiveness and matching constraints as linear constraints to obtain a linear program whose
objective value bounds the optimal competitive ratio. Solving the obtained LP's dual, we obtain
by weak duality the claimed upper bound on the optimal competitive ratio.
General Vertex Arrivals. Our high-level approach here will be to round online a fractional
online matching algorithm's output, specifically that of Wang and Wong [25]. While this approach
sounds simple, there are several obstacles to overcome. First, the fractional matching polytope is
not integral in general graphs, where a fractional matching may have value,Pe xe, some 3/2 times
larger than the optimal matching size. (For example, in a triangle graph with value xe = 1/2 for
each edge e.) Therefore, any general rounding scheme must lose a factor of 3/2 on the competitive
ratio compared to the fractional algorithm's value, and so to beat a competitive ratio of 1/2 would
require an online fractional matching with competitive ratio > 3/4 > 1− 1/e, which is impossible. To
make matters worse, even in bipartite graphs, for which the fractional matching polytope is integral
and offline lossless rounding is possible [1, 12], online lossless rounding of fractional matchings is
impossible, even under one-sided vertex arrivals [5].
Despite these challenges, we show that a slightly better than 1/2-competitive fractional matching
computed by the algorithm of [25] can be rounded online without incurring too high a loss, yielding
(1/2 + Ω(1))-competitive randomized algorithm for online matching under general vertex arrivals.
To outline our approach, we first consider a simple method to round matchings online. When
vertex v arrives, we pick an edge {u, v} with probability zu = xuv/ Pr[u free when v arrives], and
add it to our matching if u is free.
However, we observe that, for the correct set of parameters, the fractional matching algorithm
be in our matching with the right marginal probability, xe, resulting in a lossless rounding. Unfor-
tunately, we know of no better-than-1/2-competitive fractional algorithm for which this rounding
IfPu zu 6 1, this allows us to pick at most one edge per vertex and have each edge e = {u, v}
guaranteesPu zu 6 1.
of Wang and Wong [25] makesPu zu close to one, while still ensuring a better-than-1/2-competitive
algorithm so thatPu zu 6 1+O(ε), while retaining a competitive ratio of 1/2+O(ε). Now consider
u = zu/ max{1,Pu zu} and match if u is free. As the sum of zu's is slightly
the same rounding algorithm with normalized probabilities: I.e., on v's arrival, sample a neighbor
u with probability z′
above one in the worst case, this approach does not drastically reduce the competitive ratio. But
the normalization factor is still too significant compared to the competitive ratio of the fractional
solution, driving the competitive ratio of the rounding algorithm slightly below 1/2.
fractional solution. Namely, as we elaborate later in Section 3.3, we set the parameters of their
3
To account for this minor yet significant loss, we therefore augment the simple algorithm by
allowing it, with small probability (e.g., say √ε), to sample a second neighbor u2 for each arriving
vertex v, again with probabilities proportional to z′
u2: If the first sampled choice, u1, is free, we
match v to u1. Otherwise, if the second choice, u2, is free, we match v to u2. What is the marginal
probability that such an approach matches an incoming vertex v to a given neighbor u? Letting
Fu denote the event that u is free when v arrives, this probability is precisely
Pr[Fu] · z′
u + z′
u · √ε ·Xw
w · (1 − Pr[Fw Fu])! .
z′
(1)
Here the first term in the parentheses corresponds to the probability that v matches to u via the
first choice, and the second term corresponds to the same happening via the second choice (which
is only taken when the first choice fails).
Ideally, we would like (1) to be at least xuv for all edges, which would imply a lossless rounding.
However, as mentioned earlier, this is difficult and in general impossible to do, even in much more
restricted settings including one-sided bipartite vertex arrivals. We therefore settle for showing
that (1) is at least xuv = Pr[Fu] · zu for most edges (weighted by xuv). Even this goal, however,
is challenging and requires a nontrivial understanding of the correlation structure of the random
events Fu. To see this, note that for example if the Fw events are perfectly positively correlated,
i.e., Pr[Fw Fu] = 1, then the possibility of picking e as a second edge does not increase this edge's
probability of being matched at all compared to if we only picked a single edge per vertex. This
results in e being matched with probability Pr[Fu] · z′
does not lead to any gain over the 1/2 competitive ratio of greedy. Such problems are easily shown
not to arise if all Fu variables are independent or negatively correlated. Unfortunately, positive
correlation does arise from this process, and so we the need to control these positive correlations.
The core of our analysis is therefore dedicated to showing that even though positive correlations
do arise, they are by and large rather weak. Our main technical contribution consists of developing
techniques for bounding such positive correlations. The idea behind the analysis is to consider the
primary choices and secondary choices of vertices as defining a graph, and showing that after a
natural pruning operation that reflects the structure of dependencies, most vertices are most often
part of a very small connected component in the graph. The fact that connected components are
typically very small is exactly what makes positive correlations weak and results in the required
lower bound on (1) for most edges (in terms of x-value), which in turn yields our 1/2 + Ω(1)
competitive ratio.
u = Pr[Fu] · zu/Pw zw = xuv/Pw zw, which
2 Edge Arrivals
In this section we prove the asymptotic optimality of the greedy algorithm for online matching
under adversarial edge arrivals. As discussed briefly in Section 1, our main idea will be to provide
a "prefix hardness" instance, where an underlying input and the arrival order is known to the
online matching algorithm, but the prefix of the input to arrive (or "termination time") is not.
Consequently, the algorithm must accrue high enough value up to each arrival time, to guarantee
a high competitive ratio at all points in time. As we show, the fractional matching constraints rule
out a competitive ratio of 1/2 + Ω(1) even in this model where the underlying graph is known.
Theorem 2.1. There exists an infinite family of bipartite graphs with maximum degree n and edge
arrival order for which any online matching algorithm is at best(cid:16) 1
2 + 1
2n+2(cid:17)-competitive.
4
Proof. We will provide a family of graphs for which no fractional online matching algorithm has
better competitive ratio. Since any randomized algorithm induces a fractional matching algorithm,
this immediately implies our claim. The nth graph of the family, Gn = (U, V, E), consists of a
bipartite graph with U = V = n vertices on either side. We denote by ui ∈ U and vi ∈ V the
ith node on the left and right side of Gn, respectively. Edges are revealed in n discrete rounds.
In round i = 1, 2, . . . , n, the edges of a perfect matching between the first i left and right vertices
arrive in some order. I.e., a matching of u1, u2, . . . , ui and v1, v2, . . . , vi is revealed. Specifically,
edges (uj, vi−j+1) for all i > j arrive. (See Figure 1 for example.) Intuitively, the difficulty for
an algorithm attempting to assign much value to edges of OP T is that the (unique) maximum
matching OP T changes every round, and no edge ever re-enters OP T .
U
u1
u2
u3
u4
u5
V
v1
v2
v3
v4
v5
U
u1
u2
u3
u4
u5
V
v1
v2
v3
v4
v5
U
u1
u2
u3
u4
u5
V
v1
v2
v3
v4
v5
U
u1
u2
u3
u4
u5
V
v1
v2
v3
v4
v5
U
u1
u2
u3
u4
u5
V
v1
v2
v3
v4
v5
(a) round 1
(b) round 2
(c) round 3
(d) round 4
(e) round 5
Figure 1: G5, together with arrival order. Edges of current (prior) round are solid (dashed).
Consider some α-competitive fractional algorithm A. We call the edge of a vertex w in the
(unique) maximum matching of the subgraph of Gn following round i the ith edge of w. For i > j,
denote by xi,j the value A assigns to the ith edge of vertex uj (and of vi−j+1); i.e., to (uj, vi−j+1).
By feasibility of the fractional matching output by A, we immediately have that xi,j > 0 for all i, j,
as well as the following matching constraints for uj and vj. (For the latter, note that the ith edge
of vi−j+1 is assigned value xi,j = xi,i−(i−j+1)+1 and so the ith edge of vj is assigned value xi,i−j+1).
xi,j 6 1.
(uj matching constraint)
nXi=j
xi,i−j+1 6 1.
(vj matching constraint)
nXi=j
(2)
(3)
On the other hand, as A is α-competitive, we have that after some kth round -- when the
maximum matching has cardinality k -- algorithm A's fractional matching must have value at
least α · k. (Else an adversary can stop the input after this round, leaving A with a worse than
α-competitive matching.) Consequently, we have the following competitiveness constraints.
kXi=1
iXj=1
xi,j > α · k
∀k ∈ [n].
(4)
Combining constraints (2), (3) and (4) together with the non-negativity of the xi,k yields the
following linear program, LP(n), whose optimal value upper bounds any fractional online matching
algorithm's competitiveness on Gn, by the above.
5
maximize α
subject to: Pn
Pn
Pk
i=1Pi
xi,j > 0
i=j xi,j 6 1
i=j xi,i−j+1 6 1
j=1 xi,j > α · k
∀j ∈ [n]
∀j ∈ [n]
∀k ∈ [n]
∀i, j ∈ [n].
To bound the optimal value of LP(n), we provide a feasible solution its LP dual, which we denote
by Dual(n). By weak duality, any dual feasible solution's value upper bounds the optimal value of
LP(n), which in turn upper bounds the optimal competitive ratio. Using the dual variables ℓj, rj
for the degree constraints of the jth left and right vertices respectively (uj and vj) and dual variable
ck for the competitiveness constraint of the kth round, we get the following dual linear program.
Recall here again that xi,i−j+1 appears in the matching constraint of vj, with dual variable rj, and
so xi,j = xi,i−(i−j+1)+1 appears in the same constraint for vi−j+1.)
minimize Pn
subject to: Pn
j=1 (ℓj + rj)
k=1 k · ck > 1
ℓj + ri−j+1 −Pn
ℓj, rj, ck > 0
k=i ck > 0 ∀i ∈ [n], j ∈ [i]
∀j, k ∈ [n].
We provide the following dual solution.
2
ck =
n(n + 1)
ℓj = rj =( n−2(j−1)
n(n+1)
0
∀k ∈ [n]
if j 6 n/2 + 1
if n/2 + 1 < j 6 n.
We start by proving feasibility of this solution. The first constraint is satisfied with equality.
2(n−i+1)
For the second constraint, as Pn
n(n+1)
for all i ∈ [n], j ∈ [i]. Note that if j > n/2 + 1, then ℓj = rj = 0 > n−2(j−1)
n(n+1) . So, for all j we
n(n+1) + n−2(i−j+1−1)
for all
have ℓj = rj >
i ∈ [n], j ∈ [i]. Non-negativity of the ℓj, rj, ck variables is trivial, and so we conclude that the above
is a feasible dual solution.
n−2(j−1)
n(n+1) . Consequently, ℓj + ri−j+1 >
it suffices to show that ℓj + ri−j+1 >
k=i ck = 2(n−i+1)
= 2(n−i+1)
n(n+1)
n−2(j−1)
n(n+1)
n(n+1)
It remains to calculate this dual feasible solution's value. We do so for n even,1 for which
nXj=1
(ℓj + rj) = 2 ·
nXj=1
ℓj = 2 ·
n/2+1Xj=1
n − 2(j − 1)
n(n + 1)
=
1
2
+
1
2n + 2
,
completing the proof.
Remark 1. Recall that Buchbinder et al. [3] and Lee and Singla [20] presented better-than-
1/2-competitive algorithms for bounded-degree graphs and bounded number of arrival batches. Our
upper bound above shows that a deterioration of the competitive guarantees as the maximum
degree and number of arrival batches increase (as in the algorithms of [3, 20]) is inevitable.
Remark 2. Recall that the asymptotic competitive ratio of an algorithm is the maximum c
such that the algorithm always guarantees value at least ALG > c · OP T − b for some fixed b > 0.
Our proof extends to this weaker notion of competitiveness easily, by revealing multiple copies of
the hard family of Theorem 2.1 and letting xik denote the average of its counterparts over all copies.
1The case of n odd is similar. As it is unnecessary to establish the result of this theorem, we omit it.
6
3 General Vertex Arrivals
In this section we present a (1/2 + Ω(1))-competitive randomized algorithm for online matching
under general arrivals. As discussed in the introduction, our approach will be to round (online) a
fractional online matching algorithm's output. Specifically, this will be an algorithm from the family
of fractional algorithms introduced in [25]. In Section 3.1 we describe this family of algorithms. To
motivate our rounding approach, in Section 3.2 we first present a simple lossless rounding method for
a 1/2-competitive algorithm in this family. In Section 3.3 we then describe our rounding algorithm
for a better-than-1/2-competitive algorithm in this family. Finally, in Section 3.4 we analyze this
rounding scheme, and show that it yields a (1/2 + Ω(1))-competitive algorithm.
3.1 Finding a fractional solution
In this section we revisit the algorithm of Wang and Wong [25], which beats the 1/2 competitiveness
barrier for online fractional matching under general vertex arrivals. Their algorithm (technically,
family of algorithms) applies the primal-dual method to compute both a fractional matching and
a fractional vertex cover -- the dual of the fractional matching relaxation. The LPs defining these
dual problems are as follows.
Primal-Matching
maximize Pe∈E xe
subject to: Pu∈N (v) xuv 6 1 ∀u ∈ V
∀e ∈ E
xe > 0
Dual-Vertex Cover
minimize Pu∈V yu
subject to:
yu + yv > 1 ∀e = {u, v} ∈ E
yu > 0
∀u ∈ V
Before introducing the algorithm of [25], we begin by defining the fractional online vertex cover
problem for vertex arrivals. When a vertex v arrives, if Nv(v) denotes the previously-arrived
neighbors of v, then for each u ∈ Nv(v), a new constraint yu + yv > 1 is revealed, which an
online algorithm should satisfy by possibly increasing yu or yv. Suppose v has its dual value set
to yv = 1 − θ. Then all of its neighbors should have their dual increased to at least θ. Indeed, an
algorithm may as well increase yu to max{yu, θ}. The choice of θ therefore determines an online
fractional vertex cover algorithm. The increase of potential due to the newly-arrived vertex v is
thus 1 − θ +Pu∈Nv(v)(θ − yu)+.2 In [25] θ is chosen to upper bound this term by 1 − θ + f (θ)
for some function f (·). The primal solution (fractional matching) assigns values xuv so as to
guarantee feasibility of ~x and a ratio of β between the primal and dual values of ~x and ~y, implying
1
β -competitiveness of this online fractional matching algorithm, by feasibility of ~y and weak duality.
The algorithm, parameterized by a function f (·) and parameter β to be discussed below, is given
formally in Algorithm 1. In the subsequent discussion, Nv(u) denotes the set of neighbors of u that
arrive before v.
Algorithm 1 is parameterized by a function f and a constant β. The family of functions
considered by [25] are as follows.
As we will see, choices of β guaranteeing feasibility of ~x are related to the following quantity.
Definition 3.1. Let fκ(θ) :=(cid:0) 1+κ
Definition 3.2. For a given f : [0, 1] −→ R+ let β∗(f ) := maxθ∈[0,1] 1 + f (1 − θ) +R 1
2κ (cid:0)θ + κ−1
2 − θ(cid:1) 1+κ
2Here and throughout the paper, we let x+ := max{0, x} for all x ∈ R.
2κ . We define W := {fκ κ > 1}.
2 (cid:1) κ−1
1−t
f (θ) dθ.
θ
7
Input : A stream of vertices v1, v2, . . . vn. At step i, vertex vi and Nvi(vi) are revealed.
Output: A fractional vertex cover solution ~y and a fractional matching ~x.
1 Let yu ← 0 for all u, let xuv ← 0 for all u, v.
2 foreach v in the stream do
3
4
5
6
7
maximize θ
subject to:
θ 6 1
Pu∈Nv(v) (θ − yu)+
(cid:16)1 + 1−θ
f (θ)(cid:17) .
foreach u ∈ Nv(v) do
xuv ←− (θ−yu)+
yu ←− max{yu, θ}.
β
6 f (θ)
yv ←− 1 − θ.
Algorithm 1: Online general vertex arrival fractional matching and vertex cover
For functions f ∈ W this definition of β∗(f ) can be simplified to β∗(f ) = 1 + f (0), due to the
observation (see [25, Lemmas 4,5]) that all functions f ∈ W satisfy
β∗(f ) = 1 + f (1 − θ) +Z 1
θ
1 − θ
f (θ)
dθ
∀θ ∈ [0, 1].
(5)
As mentioned above, the competitiveness of Algorithm 1 for appropriate choices of f and β is
later), one can even bound individual vertices' contributions to these sums. In particular, for any
obtained by relating the overall primal and dual values,Pe xe andPv yv. As we show (and rely on
vertex v's arrival time, each vertex u's contribution to Pe xe, which we refer to as its fractional
degree, xu :=Pw∈Nv(u) xuw, can be bounded in terms of its dual value by this point, yu, as follows.
fractional degree just before v arrives, xu :=Pw∈Nv(u) xuw, is bounded as follows:
Lemma 3.3. For any vertex u, v ∈ V , let yu be the potential of u prior to arrival of v. Then the
yu
β
6 xu 6
yu + f (1 − yu)
.
β
Broadly, the lower bound on xu is obtained by lower bounding the increase xu by the increase
to yu/β after each vertex arrival, while the upper bound follows from a simplification of a bound
given in [25, Invariant 1] (implying feasibility of the primal solution), which we simplify using (5).
See Appendix B for a full proof.
Another observation we will need regarding the functions f ∈ W is that they are decreasing.
Observation 3.4. Every function f ∈ W is non-increasing in its argument in the range [0, 1].
Proof. As observed in [25], differentiating (5) with respect to z yields −f ′(1 − z) − 1−z
f (z) = 0, from
which we obtain f (z) · f ′(1 − z) = z − 1. Replacing z by 1 − z, we get f (1 − z) · f ′(z) = −z, or
f ′(z) = − z
f (1−z) . As f (z) is positive for all z ∈ [0, 1], we have that f ′(z) < 0 for all z ∈ [0, 1].
The next lemma of [25] characterizes the achievable competitiveness of Algorithm 1.
Lemma 3.5 ([25]). Algorithm 1 with function f ∈ W and β > β∗(f ) = 1 + f (0) is 1
β competitive.
Wang and Wong [25] showed that taking κ ≈ 1.1997 and β = β∗(fκ), Algorithm 1 is ≈ 0.526
In later sections we show how to round the output of Algorithm 1 with fκ with
competitive.
κ = 1 + 2ε for some small constant ε and β = 2 − ε to obtain a (1/2 + Ω(1))-competitive algorithm.
But first, as a warm up, we show how to round this algorithm with κ = 1 and β = β∗(f1) = 2.
8
3.2 Warmup: a 1/2-competitive randomized algorithm
In this section we will round the 1/2-competitive fractional algorithm obtained by running Algo-
rithm 1 with function f (θ) = f1(θ) = 1 − θ and β = β∗(f ) = 2. We will devise a lossless rounding
of this fractional matching algorithm, by including each edge e in the final matching with a proba-
bility equal to the fractional value xe assigned to it by Algorithm 1. Note that if v arrives after u,
then if Fu denotes the event that u is free when v arrives, then edge {u, v} is matched by an online
algorithm with probability Pr[{u, v} ∈ M ] = Pr[{u, v} ∈ M Fu]· Pr[Fu]. Therefore, to match each
edge {u, v} with probability xuv, we need Pr[{u, v} ∈ M Fu] = xuv/ Pr[Fu]. That is, we must
match {u, v} with probability zu = xuv/ Pr[Fu] conditioned on u being free. The simplest way of
doing so (if possible) is to pick an edge {u, v} with the above probability zu always, and to match
it only if u is free. Algorithm 2 below does just this, achieving a lossless rounding of this fractional
algorithm. As before, Nv(u) denotes the set of neighbors of u that arrive before v.
Input : A stream of vertices v1, v2, . . . , vn. At step i, vertex vi and Nvi(vi) are revealed.
Output: A matching M .
1 Let yu ← 0 for all u, let xuv ← 0 for all u, v.
2 Let M ← ∅.
3 foreach v in the stream do
4
Update yu's and xuv's using Algorithm 1 with β = 2 and f = f1.
foreach u ∈ Nv(v) do
5
6
7
8
9
zu ←
Pr[u is free when v arrives] .
xuv
// zu is xuv/(1 − yu) as shown later
Sample (at most) one neighbor u ∈ Nv(v) according to zu.
if a free neighbor u is sampled then
Add {u, v} to M .
Algorithm 2: Online vertex arrival warmup randomized fractional matching
Algorithm 2 is well defined if for each vertex v's arrival, z is a probability distribution; i.e.,
Pu∈Nv(v) zu 6 1. The following lemma asserts precisely that. Moreover, it asserts that Algorithm 2
matches each edge with the desired probability.
Lemma 3.6. Algorithm 2 is well defined, since for every vertex v on arrival, z is a valid probability
distribution. Moreover, for each v and u ∈ Nv(v), it matches edge {u, v} with probability xe.
Proof. We prove both claims in tandem for each v, by induction on the number of arrivals. For the
base case (v is the first arrival), the set Nv(v) is empty and thus both claims are trivial. Consider
the arrival of a later vertex v. By the inductive hypothesis we have that each vertex u ∈ Nv(v) is
β = 2, if w arrives after u, then yu and θ at arrival of w satisfy xuw = (θ−yu)+
That is, xuw is precisely the increase in yu following arrival of w. On the other hand, when u arrived
previously matched with probabilityPw∈Nv(u) xwu. But by our choice of f (θ) = f1(θ) = 1 − θ and
f (θ)(cid:17) = (θ−yu)+.
we have that its dual value yu increased by 1 − θ =Pv′∈Nu(u)(θ − yv′)+ =Pv′∈Nu(u) xuv′. To see
this last step, we recall first that by definition of Algorithm 1 and our choice of f (θ) = 1 − θ, the
value θ on arrival of v is chosen to be the largest θ 6 1 satisfying
·(cid:16)1 + 1−θ
β
X∀u∈Nv(v)
(θ − yu)+ 6 1 − θ.
(6)
But the inequality (6) is an equality whether or not θ = 1 (if θ = 1, both sides are zero). We
conclude that yu =Pv′∈Nv(u) xuv′ just prior to arrival of v. But then, by the inductive hypothesis,
9
this implies that Pr[u free when v arrives] = 1 − yu (yielding an easily-computable formula for zu).
Consequently, by (6) we have that when v arrives z is a probability distribution, as
(θ − yu)+
6 1.
Xu∈Nv(v)
zu = Xu∈Nv(v)
(θ − yu)+
1 − yu
6 Xu∈Nv(v): yu6θ
1 − θ
(θ − yu)+
1 − θ
= Xu∈Nv(v)
Finally, for u to be matched to a latter-arriving neighbor v, it must be picked and free when v
arrives, and so {u, v} is indeed matched with probability
Pr[{u, v} ∈ M ] =
Pr[u is free when v arrives] · Pr[u is free when v arrives] = xuv.
xuv
In the next section we present an algorithm which allows to round better-than-1/2-competitive
algorithms derived from Algorithm 1.
3.3 An improved algorithm
In this section, we build on Algorithm 2 and show how to improve it to get a (1/2+Ω(1)) competitive
ratio.
There are two concerns when modifying Algorithm 2 to work for a general function from the
family W . The first is how to compute the probability that a vertex u is free when vertex v arrives,
in Line 6. In the simpler version, we inductively showed that this probability is simply 1−yu, where
yu is the dual value of u as of v's arrival (see the proof of Lemma 3.6). With a general function f ,
this probability is no longer given by a simple formula. Nevertheless, it is easily fixable: We can
either use Monte Carlo sampling to estimate the probability of u being free at v's arrival to a given
inverse polynomial accuracy, or we can in fact exactly compute these probabilities by maintaining
their marginal values as the algorithm progresses. In what follows, we therefore assume that our
algorithm can compute these probabilities exactly.
The second and more important issue is with the sampling step in Line 7. In the simpler algo-
rithm, this step is well-defined as the sampling probabilities indeed form a valid distribution: I.e.,
Pu∈Nv(v) zu 6 1 for all vertices v. However, with a general function f , this sum can exceed one,
rendering the sampling step in Line 7 impossible. Intuitively, we can normalize the probabilities to
make it a proper distribution, but by doing so, we end up losing some amount from the approxi-
mation guarantee. We hope to recover this loss using a second sampling step, as we mentioned in
Section 1.2 and elaborate below.
significant to ignore.
Suppose that, instead of β = 2 and f = f1 (i.e., the function f (θ) = 1 − θ), we use f = f1+2ε
and β = 2 − ε to define xuv and yu values. As we show later in this section, for an ε sufficiently
1 + O(ε). However, since the approximation factor of the fractional solution is only 1/2 + O(ε)
Now suppose that we allow arriving vertices to sample a second edge with a small (i.e., √ε)
probability and match that second edge if the endpoint of the first sampled edge is already matched.
u denote the normalized
zu values. Further let Fw denote the event that vertex w is free (i.e, unmatched) at the arrival of v.
Then the probability that v matches u for some u ∈ Nv(v) using either of the two sampled edges is
small, we then have Pu∈Nv(v) zu 6 1 + O(ε), implying that the normalization factor is at most
for such a solution, (i.e., P{u,v}∈E xuv > (1/β) ·Pu∈V yu), the loss due to normalization is too
Consider the arrival of a fixed vertex v such thatPu∈Nv(v) zu > 1, and let z′
w · (1 − Pr[Fw Fu]) ,
Pr[Fu] ·z′
u√ε · Xw∈Nv(v)
u + z′
(7)
z′
10
which is the same expression from (1) from Section 1.2, restated here for quick reference. Recall
that the first term inside the parentheses accounts for the probability that v matches u via the first
sampled edges, and the second term accounts for the probability that the same happens via the
second sampled edge. Note that the second sampled edge is used only when the first one is incident
to an already matched vertex and the other endpoint of the second edge is free. Hence we have
the summation of conditional probabilities in the second term, where the events are conditioned
on the other endpoint, u, being free. If the probability given in (7) is xuv for all {u, v} ∈ E, we
would have the same guarantee as the fractional solution xuv, and the rounding would be lossless.
This seems unlikely, yet we can show that the quantity in (7) is at least (1 − ε2)· xuv for most (not
by number, but by the total fractional value of xuv's) of the edges in the graph, showing that our
rounding is almost lossless. We postpone further discussion of the analysis to Section 3.4 where we
highlight the main ideas before proceeding with the formal proof.
Input : A stream of vertices v1, v2, . . . , vn. At step i, vertex vi and Nvi(v) are revealed.
Output: A matching M .
1 Let yu ← 0 for all u, let xuv ← 0 for all u, v.
2 Let M ← ∅.
3 foreach v in the stream do
4
*/
*/
5
6
7
8
9
10
11
12
13
14
15
16
Update yu's and xuv's using Algorithm 1 with β = 2 − ε and f = f1+2ε.
foreach u ∈ Nv(v) do
/* Compute Pr[u is free when v arrives] as explained in Section 3.3
zu ←
Pr[u is free when v arrives] .
xuv
foreach u ∈ Nv(v) do
z′
u ← zu/ maxn1,Pu∈Nv(v) zuo.
Pick (at most) one u1 ∈ Nv(v) with probability z′
u1.
if Pu∈Nv(v) zu > 1 then
With probability √ε, pick (at most) one u2 ∈ Nv(v) with probability z′
/* Probability of dropping edge {u, v} can be computed using (7).
Drop u2 with minimal probability ensuring {u2, v} is matched with probability at
most xu2v.
u2.
if a free neighbor u1 is sampled then
Add {u1, v} to M .
else if a free neighbor u2 is sampled then
Add {u2, v} to M .
Algorithm 3: A randomized online matching algorithm under general vertex arrivals.
Our improved algorithm is outlined in Algorithm 3. Up until Line 6, it is similar to Algorithm 2
except that it uses β = 2 − ε and f = f1+2ε where we choose ε > 0 to be any constant small
enough such that the results in the analysis hold. In Line 8, if the sum of zu's exceeds one we
normalize the zu to obtain a valid probability distribution z′
u. In Line 9, we sample the first edge
incident to an arriving vertex v. In Line 11, we sample a second edge incident to the same vertex
with probability √ε if we had to scale down zu's in Line 8. Then in Line 12, we drop the sampled
second edge with the minimal probability to ensure that no edge {u, v} is matched with probability
more than xuv. Since (7) gives the exact probability of {u, v} being matched, this probability of
dropping an edge {u, v} can be computed by the algorithm. However, to compute this, we need the
11
conditional probabilities Pr[Fw Fu], which again can be estimated using Monte Carlo sampling3.
In the subsequent lines, we match v to a chosen free neighbor (if any) among its chosen neighbors,
prioritizing its first choice.
For the purpose of analysis we view Algorithm 3 as constructing a greedy matching on a directed
acyclic graph (DAG) Hτ defined in the following two definitions.
Definition 3.7 (Non-adaptive selection graph Gτ ). Let τ denote the random choices made by the
vertices of G. Let Gτ be the DAG defined by all the arcs (v, u1), (v, u2) for all vertices v ∈ V . We
call the arcs (v, u1) primary arcs, and the arcs (v, u2) the secondary arcs.
Definition 3.8 (Pruned selection graph Hτ ). Now construct Hτ from Gτ by removing all arcs
(v, u) (primary or secondary) such that there exists a primary arc (v′, u) with v′ arriving before v.
We further remove a secondary arc (v, u) if there is a primary arc (v, u); i.e., if a vertex u has at
least one incoming primary arc, remove all incoming primary arcs that came after the first primary
arc and all secondary arcs that came after or from the same vertex as the first primary arc.
It is easy to see that the matching constructed by Algorithm 3 is a greedy matching constructed
on Hτ based on order of arrival and prioritizing primary arcs. The following lemma shows that the
set of matched vertices obtained by this greedy matching does not change much for any change in
the random choices of a single vertex v, which will prove useful later on. It can be proven rather
directly by an inductive argument showing the size of the symmetric difference in matched vertices
in Gτ and Gτ ′ does not increase after each arrival besides the arrival of v, whose arrival clearly
increases this symmetric difference by at most two. See Appendix A for details.
Lemma 3.9. Let Gτ and Gτ ′ be two realizations of the random digraph where all the vertices in the
two graphs make the same choices except for one vertex v. Then the number of vertices that have
different matched status (free/matched) in the matchings computed in Hτ and Hτ ′ at any point of
time is at most two.
3.4 Analysis
In this section, we analyze the competitive ratio of Algorithm 3. We start with an outline of the
analysis where we highlight the main ideas.
3.4.1 High-Level Description of Analysis
As described in Section 3.3, the main difference compared to the simpler 1/2-competitive algorithm
is the change of the construction of the fractional solution, which in turn makes the rounding
more complex.
The majority of the analysis is therefore devoted to such "problematic" vertices since otherwise,
In particular, we may have at the arrival of a vertex v that Pu∈Nv(v) zu > 1.
if Pu∈Nv(v) zu 6 1, the rounding is lossless due to the same reasons as described in the simpler
setting of Section 3.2. We now outline the main ideas in analyzing a vertex v withPu∈Nv(v) zu > 1.
Let Fw be the event that vertex w is free (i.e., unmatched) at the arrival of v. Then, as described
in Section 3.3, the probability that we select edge {u, v} in our matching is the minimum of xuv
(because of the pruning in Line 12), and
Pr[Fu] ·z′
z′
w · (1 − Pr[Fw Fu]) .
3It is also possible to compute them exactly if we allow the algorithm to take exponential time.
u + z′
u√ε · Xw∈Nv(v)
12
By definition, Pr[Fu] · zu = xuv, and the expression inside the parentheses is at least zu (implying
Pr[{u, v} ∈ M ] = xuv) if
1 + √ε · Xw∈Nv(v)
z′
w · (1 − Pr[Fw Fu]) >
zu
z′
u
.
(8)
To analyze this inequality, we first use the structure of the selected function f = f1+2ε and the
(see Lemma 3.10 and Corollary 3.11 in Section 3.4.2). In particular, there are absolute constants
0 < c < 1 and C > 1 (both independent of ε) such that
selection of β = 2 − ε to show that if Pu∈Nv(v) zu > 1 then several structural properties hold
1. Pu∈Nv(v) zu 6 1 + Cε;
2. zu 6 C√ε for every u ∈ Nv(v); and
3. c 6 Pr[Fw] 6 1 − c for every w ∈ Nv(v).
The first property implies that the right-hand-side of (8) is at most 1+ Cε; and the second property
implies that v has at least Ω(1/√ε) neighbors and that each neighbor u satisfies z′
u 6 zu 6 C√ε.
For simplicity of notation, we assume further in the high-level overview that v has exactly 1/√ε
u = √ε. Inequality (8) would then be implied by
neighbors and each u ∈ Nv(v) satisfies z′
Xw∈Nv(v)
(1 − Pr[Fw Fu]) > C .
(9)
To get an intuition why we would expect the above inequality to hold, it is instructive to consider
the unconditional version:
Xw∈Nv(v)
(1 − Pr[Fw]) > cNv(v) = c/√ε ≫ C ,
where the first inequality is from the fact that Pr[Fw] 6 1 − c for any neighbor w ∈ Nv(v). The
large slack in the last inequality, obtained by selecting ε > 0 to be a sufficiently small constant,
is used to bound the impact of conditioning on the event Fu. Indeed, due to the large slack, we
have that (9) is satisfied if the quantity Pw∈Nv(v) Pr[FwFu] is not too far away from the same
summation with unconditional probabilities, i.e., Pw∈Nv(v) Pr[Fw]. Specifically, it is sufficient to
show
Xw∈Nv(v)
(Pr[FwFu] − Pr[Fw]) 6 c/√ε − C .
(10)
We do so by bounding the correlation between the events Fu and Fw in a highly non-trivial manner,
which constitutes the heart of our analysis. The main challenges are that events Fu and Fw can
be positively correlated and that, by conditioning on Fu, the primary and secondary choices of
different vertices are no longer independent.
We overcome the last difficulty by replacing the conditioning on Fu by a conditioning on the
component in Hτ (at the time of v's arrival) that includes u. As explained in Section 3.3, the
matching output by our algorithm is equivalent to the greedy matching constructed in Hτ and so
the component containing u (at the time of v's arrival) determines Fu. But how can this component
look like, assuming the event Fu? First, u cannot have any incoming primary arc since then u would
be matched (and so the event Fu would be false). However, u could have incoming secondary arcs,
13
u
u
Figure 2: Two examples of the component of Hτ containing u. Vertices are depicted from right to
left in the arrival order. Primary and secondary arcs are solid and dashed, respectively. The edges
that take part in the matching are thick.
assuming that the tails of those arcs are matched using their primary arcs. Furthermore, u can have
an outgoing primary and possibly a secondary arc if the selected neighbors are already matched.
These neighbors can in turn have incoming secondary arcs, at most one incoming primary arc
(due to the pruning in the definition of Hτ ), and outgoing primary and secondary arcs; and so
on. In Figure 2, we give two examples of the possible structure, when conditioning on Fu, of u's
component in Hτ (at the time of v's arrival). The left example contains secondary arcs, whereas
the component on the right is arguably simpler and only contains primary arcs.
An important step in our proof is to prove that, for most vertices u, the component is of the
simple form depicted to the right with probability almost one. That is, it is a path P consisting of
primary arcs, referred to as a primary path (see Definition 3.13) that further satisfies:
(i) it has length O(ln(1/ε)); and
(ii) the total z-value of the arcs in the blocking set of P is O(ln(1/ε)). The blocking set is defined
in Definition 3.14. Informally, it contains those arcs that if appearing as primary arcs in Gτ
would cause arcs of P to be pruned (or blocked) from Hτ .
Let P be the primary paths of above type that appear with positive probability as u's component in
Hτ . Further let EQP be the event that u's component equals P . Then we show (for most vertices)
thatPP ∈P Pr[EQP Fu] is almost one. For simplicity, let us assume here that the sum is equal to
one. Then by the law of total probability and sincePP ∈P Pr[EQP Fu] = 1,
(Pr[Fw Fu, EQP ] − Pr[Fw])
Xw∈Nv(v)
(Pr[Fw EQP ] − Pr[Fw]) ,
(Pr[Fw Fu] − Pr[Fw]) = XP ∈P
= XP ∈P
Pr[EQP Fu] Xw∈Nv(v)
Pr[EQP Fu] Xw∈Nv(v)
where the last equality is because the component P determines Fu. The proof is then completed
by analyzing the term inside the parentheses for each primary path P ∈ P separately. As we prove
in Lemma 3.15, the independence of primary and secondary arc choices of vertices is maintained
after conditioning on EQP .4 Furthermore, we show that there is a bijection between the outcomes
of the unconditional and the conditional distributions, so that the expected number of vertices
4To be precise, conditioning on a primary path P with a so-called termination certificate T , see Definition 3.13.
In the overview, we omit this detail and consider the event EQP,T (instead of EQP ) in the formal proof.
14
that make different choices under this pairing can be upper bounded by roughly the length of the
path plus the z-value of the edges in the blocking set. So, for a path P as above, we have that
the expected number of vertices that make different choices in the paired outcomes is O(ln(1/ε))
which, by Lemma 3.9, implies that the expected number of vertices that change matched status is
also upper bounded by O(ln(1/ε)). In other words, we have for every P ∈ P that
(Pr[FwEQP ] − Pr[Fw]) = O(ln(1/ε)),
Xw∈Nv(v)
(Pr[FwEQP ] − Pr[Fw]) 6 Xw∈V
which implies (10) for a small enough choice of ε. This completes the overview of the main steps
in the analysis. The main difference in the formal proof is that not all vertices satisfy that their
component is a short primary path with probability close to 1. To that end, we define the notion
of good vertices in Section 3.4.4, which are the vertices that are very unlikely to have long directed
paths of primary arcs rooted at them. These are exactly the vertices v for which we can perform the
above analysis for most neighbors u (in the proof of the "key lemma") implying that the rounding
is almost lossless for v. Then, in Section 3.4.6, we show using a rather simple charging scheme that
most of the vertices in the graph are good. Finally, in Section 3.4.7, we put everything together
and prove Theorem 1.2.
3.4.2 Useful Properties of W Functions and Algorithm 3
For the choice of f = f1+2ε as we choose, we have f (θ) = (1 + ε − θ)·(cid:16) θ+ε
1+2ε . In Appendix C
we give a more manageable upper bound for f (θ) which holds for sufficiently small ε. Based on
this simple upper bound on f and some basic calculus, we obtain the following useful structural
properties for the conditional probabilities, zu, of Algorithm 3. See Appendix C.
Lemma 3.10. (Basic bounds on conditional probabilities zu) There exist absolute constants c ∈
(0, 1) and C > 1/c > 1 and ε0 ∈ (0, 1) such that for every ε ∈ (0, ε0) the following holds: for every
vertex v ∈ V , if yu is the dual variable of a neighbor u ∈ Nv(v) before v's arrival and θ is the value
chosen by Algorithm 1 on v's arrival, then for zu as defined in Algorithm 3, we have:
1+ε−θ(cid:17) ε
(1) If θ 6∈ (c, 1 − c), thenPu∈Nv(v) zu 6 1,
(2) If θ ∈ [0, 1], thenPu∈Nv(v) zu 6 1 + Cε,
(3) IfPu∈Nv(v) zu > 1, then zu 6 C√ε for every u ∈ Nv(v),
(4) IfPu∈Nv(v) zu > 1, then for every u ∈ Nv(v) such that zu > 0, one has yu ∈ [c/2, 1− c/2], and
(5) For all u ∈ Nv(v), one has zu 6 1/2 + O(√ε).
The following corollary will be critical to our analysis:
Corollary 3.11. There exist absolute constants c > 0 and ε0 > 0 such that for all ε ∈ (0, ε), on
u ∈ Nv(v) we have
arrival of any vertex v ∈ V , if z as defined in Algorithm 3 satisfiesPu∈Nv(v) zu > 1, then for every
Proof. By Lemma 3.10, (1) and (4) we have that ifPu∈Nv(v) zu > 1, then θ ∈ (c, 1 − c) (c is the
constant from Lemma 3.10), and for every u ∈ Nv(v) one has
c 6 Pr[u is free when v arrives] 6 1 − c.
yu ∈ [c/2, 1 − c/2].
15
(11)
On the other hand, by Lemma 3.3 one has
yu
β
6 xu 6
yu + f (1 − yu)
β
,
(12)
where xu is the fractional degree of u when v arrives.
We now note that by Lemma 3.10, (2), we have that Algorithm 3 matches every vertex u with
probability at least xu/(1 + Cε) (due to choices of primary arcs), and thus
Pr[u is free when v arrives] 6 1 −
6 1 −
xu
1 + Cε
yu
β(1 + Cε)
(by (12))
c/2
2(1 + Cε)
6 1 −
6 1 − c/5,
(by (11) and the setting β = 2 − ε 6 2)
as long as ε is sufficiently small.
For the other bound we will use two facts. The first is that the since f (y) is monotone decreasing
by Observation 3.4 and since we picked β > β∗(f ) = 1 + f (0), we have that for any y 6 1− c/2 6 1,
(13)
y + f (1 − y) 6 1 − c/2 + f (0) < β − c/2.
Then, using the fact that by Line 12, Algorithm 3 matches every vertex u with probability at most
xu, we obtain the second bound, as follows.
Pr[u is free when v arrives] > 1 − xu
yu + f (1 − yu)
> 1 −
> 1 −
> c/5.
β
β − c/2
β
(by (12))
(by (11) and (13))
(β = 2 − ε < 2.5)
Choosing c/5 as the constant in the statement of the lemma, we obtain the result.
Finally, for our analysis we will rely on the competitive ratio of the fractional solution maintained
in Line 4 being 1/β. This follows by Lemma 3.5 and the fact that for our choices of β = 2 − ε and
f = f1+2ε we have that β > β∗(f ). See Appendix C for a proof of this fact.
Fact 3.12. For all sufficiently small ε > 0, we have that 2 − ε > β∗(f1+2ε).
3.4.3 Structural Properties of Gτ and Hτ
In our analysis later, we focus on maximal primary paths (directed paths made of primary arcs) in
Hτ , in the sense that the last vertex along the primary path has no outgoing primary arc in Hτ .
The following definition captures termination certificates of such primary paths.
Definition 3.13 (Certified Primary Path). A tuple (P, T ) is a certified primary path in Hτ if P
is a directed path of primary arcs in Hτ and either
(a) the last vertex of P does not have an outgoing primary arc in Gτ and T = ∅, or
16
(b) the last vertex u of P has an outgoing primary arc (u, w) in Gτ and T = (u′, w) is a primary
arc in Hτ such that u′ precedes u in the arrival order.
To elaborate, a certified primary path (P, T ) is made of a (directed) path P of primary arcs
in Hτ and T is a certificate of P 's termination in Hτ that ensures the last vertex u in P has no
outgoing primary arc in Hτ , either due to u not picking a primary arc with T = ∅, or due to the
picked primary arc (u, w) being blocked by another primary arc T = (u′, w) which appears in Hτ .
As described, Gτ and Hτ differ in arcs (u, w) that are blocked by previous primary arcs to their
target vertex w. We generally define sets of arcs which can block an edge, or a path, or a certified
path from appearing in Hτ as in the following definition:
Definition 3.14 (Blocking sets). For an arc (u, w), define its blocking set
B(u, w) := {(u′, w) {u′, w} is an edge and u′ arrived before u}
to be those arcs, the appearance of any of which as primary arc in Gτ blocks (u, v) from being in
Hτ . In other words, an arc (u, v) is in Hτ as primary or secondary arc if and only if (u, v) is in
Gτ and none of the arcs in its blocking set B(u, v) is in Gτ as a primary arc.
The blocking set of a path P is simply the union of its arcs' blocking sets,
B(P ) := [(u,v)∈P
B(u, v) .
The blocking set of a certified primary path (P, T ) is the union of blocking sets of P and T ,
B(P, T ) := B(P ∪ T ).
The probability of an edge, or path, or certified primary path appearing in Hτ is governed
in part by the probability of arcs in their blocking sets appearing as primary arcs in Gτ . As an
arc (v, u) is picked as primary arc by when v arrives with probability roughly zu (more precisely,
u ∈ [zvu/(1 + Cε), zvu], by Lemma 3.10), it will be convenient to denote by z(v, u) and z′(v, u) the
z′
u when v arrives, and by z(S) =Ps∈S z(s) and z′(S) =Ps∈S z(s) the sum of z-
values zu and z′
and z′-values of arcs in a set of arcs S.
Product distributions. Note that by definition the distribution over primary and secondary
arc choices of vertices are product distributions (they are independent). As such, their joint dis-
tribution is defined by their marginals. Let pw and sw denote the distribution on primary and
secondary arc choices of w, respectively. That is, for every u ∈ Nw(w), pw(u) is the marginal prob-
ability that w selects (w, u) as its primary arc, and sw(u) is the marginal probability that w selects
(w, u) as its secondary arc. Given our target bound (8), it would be useful to show that condition-
ing on Fu preserves the independence of these arc choices. Unfortunately, conditioning on Fu does
not preserve this independence. We will therefore refine our conditioning later on the existence of
primary paths in Hτ , which as we show below maintains independence of the arc choices.
Lemma 3.15. For a certified primary path (P, T ) let EQ(P,T ) be the event that the path P equals
a maximal connected component in Hτ and the termination of P is certified by T . Then the
conditional distributions of primary and secondary choices conditioned on EQ(P,T ) are product
distributions; i.e., these conditional choices are independent. Moreover, if we let pw and sw denote
the conditional distribution on primary and secondary choices of w, respectively, then
TV(pw, pw) 6 z(R(w))
and
TV(sw, sw) 6 z(R(w)),
17
where R(w) ⊆ {w} × Nw(w) is the set of arcs leaving w whose existence as primary arcs in Gτ is
ruled out by conditioning on EQ(P,T ), and the union of these R(w), denoted by R(P, T ), satisfies
R(P, T ) :=[w
R(w) ⊆ B(P, T ) ∪ {(w, r) r is root of P} ∪
[
w∈P ∪{w: T =(w,w′)}
{w} × Nw(w).
(14)
Proof. We first bound the total variation distance between the conditional and unconditional dis-
tributions. For primary choices, conditioning on EQ(P,T ) rules out the following sets of primary
arc choices. For vertex w /∈ P arriving before the root r of P this conditioning rules out w picking
any edge in B(P, T ) as primary arc. For vertices w /∈ P with w arriving after the root r of P
this conditioning rules out picking arcs (w, r). Finally, this conditioning rules out some subset of
arcs leaving vertices in P ∪ {w : T = (w, w′)}. Taking the union over these supersets of R(w), we
obtain (14). Now, the probability of each ruled out primary choice (w, u) ∈ R(w) is zero under pw
and z′(w, u) under pw, and all other primary choices have their probability increase, with a total
increase ofP(w,u)∈R(w) z′(w, u), from which we conclude that
TV(pw, pw) =
pw(u) − pw(u) = z′(R(w)) 6 z(R(w)).
1
2 Xu∈Nw(w)
The proof for secondary arcs is nearly identical, the only differences being that the sets of ruled
out secondary arcs can be smaller (specifically, secondary arcs to w′ such that T = (u, w′) are not
ruled out by this conditioning), and the probability of any arc (w, u) being picked as secondary arc
of w is at most √ε · z′(w, u) 6 z(w, u).
Finally, we note that primary and secondary choices for different vertices are independent.
Therefore, conditioning on each vertex w not picking a primary arc in its ruled out set R(w) still
yields a product distribution, and similarly for the distributions over secondary choices.
It is easy to show that a particular certified primary path (P, T ) with high value of z(B(P, T ))
is unlikely to appear in Hτ , due to the high likelihood of arcs in its breaking set being picked as
primary arcs. The following lemma asserts that the probability of a vertex u being the root of any
primary certified path (P, T ) with high z(B(P, T )) value is low.
Lemma 3.16. For any k > 0 and any vertex u, we have the following
Pr[Hτ contains any certified primary path (P, T ) with P rooted at u and z(B(P, T )) > k] 6 e−k/2,
Pr[Hτ contains any primary path P rooted at u with z(B(P )) > k] 6 e−k/2.
Proof. We first prove the bound for certified primary paths. For a certified primary path (P, T )
where the last vertex of P is w, define P ∗ as follows:
P ∗ =(P
P ∪ {(w, w′′)}
if T = ∅
if T = (w′, w′′).
Observe that z(B(P ∗)) > k whenever z(B(P, T )) > k. This is trivial when T = ∅. To see this
for the case T = (w′, w′′), let w be the last vertex of P , and note that B(w′, w′′) ⊆ B(w, w′′), as w
arrives after w′. Also note that for (P, T ) to be in Hτ , we have that P ∗ must be in Gτ .
We say a directed primary path P ′ = u → u1 → ··· → uℓ−1 → uℓ is k-minimal if z(B(P ′)) >
k and z(B(P ′ \ {(uℓ−1, uℓ)})) < k. For such a path P ′, define B∗(P ′) as follows: Initially set
18
B∗(P ′) = B(P \ {(uℓ−1, uℓ)}). Then from B(uℓ−1, uℓ), the breaking set of the last arc of P ′, add
arcs to B∗(P ′) in reverse order of their sources' arrival until z(B∗(P ′)) > k.
Consider a certified primary path (P, T ) with P rooted at u. If a k-minimal path rooted at u
which is not a prefix of P ∗ is contained in Gτ , then (P, T ) does not appear in Gτ , and therefore
it does not appear in Hτ . On the other hand, if z(B(P, T )) > k then for (P, T ) to appear in
Hτ , we must have that the (unique) k-minimal prefix P ′ of P ∗ must appear in Gτ , and that none
of the edges of B∗(P ′) appear in Gτ . Moreover, for any certified primary path with z(B(P, T )),
conditioning on the existence of P ′ in Gτ does not affect random choices of vertices with outgoing
arcs in B∗(P ′), as these vertices are not in P ′. Since by Lemma 3.10 each arc (w, w′) appears in
Gτ with probability z′(v, u) > z(v, u)/(1 + Cε) > z(v, u)/2, we conclude that for any k-minimal
primary path P ′ rooted at u, we have
(1 − Pr[Some primary edge in B∗(P ′) ∩ ({w} × Nw(w)) is in Gτ ])
Pr[Hτ contains any certified primary path (P, T ) with z(B(P, T )) > k P ′ is in Gτ ]
6 Pr[No edge in B∗(P ′) is in Gτ P ′ is in Gτ ]
= Yw /∈P ′
6 Yw /∈P ′
6 exp(−z(B∗(P ′))/2) 6 e−k/2.
exp(cid:16)−P(w,w′)∈B(P,T )×Nw(w) z(w, w′)/2(cid:17)
Taking total probability Pu, the set of all k-minimal primary paths P ′ rooted at u, we get that
indeed, since u is the root of at most one k-minimal primary path in any realization of Gτ ,
Pr[Hτ contains a certified primary path (P, T ) rooted at u with z(B(P, T )) > k]
6 XP ′∈Pu
Pr[Hτ contains a (P, T ) with z(B(P, T )) > k P ′ is in Gτ ]
}
{z
The proof for primary path is essentially the same as the above, taking P ∗ = P .
6 e−k/2
· Pr[P ′ is in Gτ ] 6 e−k/2.
3.4.4 Analyzing Good Vertices
Consider the set of vertices that are unlikely to be roots of long directed paths of primary arcs in
Hτ . In this section, we show that Algorithm 3 achieves almost lossless rounding for such vertices,
and hence we call them good vertices. We start with a formal definition:
Definition 3.17 (Good vertices). We say that a vertex v is good if
Pr
τ
[Hτ has a primary path rooted at v of length at least 2000 · ln(1/ε)] 6 ε6.
Otherwise, we say v is bad.
As the main result of this section, for good vertices, we prove the following:
Theorem 3.18. Let v be a good vertex. Then
Pr[v is matched on arrival] > (1 − ε2) · Xu∈Nv(v)
xuv.
19
Notational conventions. Throughout this section, we fix v and let z, z′ be as in Algorithm 3.
Moreover, for simplicity of notation, we suppose that the stream of vertices ends just before v's
arrival and so quantities, such as Gτ and Hτ , refer to their values when v arrives. For a vertex u,
we let Fu denote the event that u is free (i.e., unmatched) when v arrives. In other words, Fu is
the event that u is free in the stream that ends just before v's arrival.
To prove the theorem, first note that it is immediate ifPu∈Nv(v) zu 6 1: in that case, we have
z′ = z and so the probability to match v by a primary edge, by definition of zu, is simply
Xu∈Nv(v)
zu · Pr[Fu] = Xu∈Nv(v)
xuv.
u = 1,
From now on we therefore assumePu∈Nv(v) zu > 1, which implies
(I) Pu∈Nv(v) z′
and moreover, by Lemma 3.10 and Corollary 3.11, for every u ∈ Nv(v):
(II) zu 6 C√ε,
(III) zu 6 (1 + Cε) · z′
(IV) c 6 Pr[Fu] 6 1 − c ,
where c is the constant of Corollary 3.11 and C is the constant of Lemma 3.10.
u, and
We now state the key technical lemma in the proof of Theorem 3.18:
Lemma 3.19. Consider a neighbor u ∈ Nv(v) such that
Pr
τ
[Hτ has a primary path rooted at u of length at least 2000 · ln(1/ε) Fu] 6 ε2 .
(15)
Then,
Xw∈Nv(v)
z′
w · Pr[Fw Fu] − Xw∈Nv(v)
z′
w · Pr[Fw] 6 ε1/3 .
(16)
Note that the above lemma bounds the quantityPw∈Nv(v) z′
w · Pr[Fw Fu], which will allow us
to show that (8) holds and thus the edge {u, v} is picked in the matching with probability very
close to xuv. Before giving the proof of the lemma, we give the formal argument why the lemma
implies the theorem.
Proof of Theorem 3.18. Define S to be the neighbors u in Nv(v) satisfying
Pr
τ
[Hτ has a primary path rooted at u of length at least 2000 · ln(1/ε) Fu] > ε2 .
In other words, S is the set of neighbors of v that violate (15). As v is good, we have
ε6 > Pr
τ
[Hτ has a primary path rooted at v of length at least 2000 · ln(1/ε)]
z′
u · Pr[Fu] · Pr
τ
[Hτ has a primary path rooted at u of length at least 2000 · ln(1/ε) − 1 Fu]
z′
u · Pr[Fu] · Pr
τ
[Hτ has a primary path rooted at u of length at least 2000 · ln(1/ε) Fu]
> Xu∈Nv(v)
> Xu∈Nv(v)
>Xu∈S
z′
u · Pr[Fu] · ε2.
20
The second inequality holds because v selects the primary arc (u, v) with probability z′
u and,
conditioned on Fu, u cannot already have an incoming primary arc, which implies that (u, v) is
present in Hτ . The last inequality follows from the choice of S.
By Property (III), zu 6 (1 + Cε) · z′
u and so by rewriting we get
Xu∈S
xuv =Xu∈S
zu · Pr[Fu] 6 (1 + Cε) ·Xu∈S
z′
u · Pr[Fu] 6 (1 + Cε) · ε4 6 ε3.
In other words, the contribution of the neighbors of v in S to Pu∈Nv(v) xuv is insignificant
compared to the contribution of all neighbors,
Xu∈Nv(v)
xuv = Xu∈Nv(v)
zu · Pr[Fu] > c,
(17)
where the inequality follows by the assumptionPu∈Nv(v) zu > 1 and Pr[Fu] > c by Property (IV).
We proceed to analyze a neighbor u ∈ Nv(v) \ S. Recall that it is enough to verify (8) to
conclude that edge {u, v} is picked in the matching with probability xuv. We have that
z′
w · (1 − Pr[Fw Fu])
w · (1 − Pr[Fw]) − √ε · ε1/3
w · c − √ε · ε1/3
z′
z′
1 + √ε Xw∈Nv(v)
> 1 + √ε Xw∈Nv(v)
> 1 + √ε Xw∈Nv(v)
= 1 + √εc − √ε · ε1/3
> 1 + Cε
> zu/z′
u.
(by Lemma 3.19)
(Pr[Fw] 6 1 − c by (IV))
Xw∈Nv(v)
z′
w = 1 by (I)
(for ε small enough)
(by (III))
Therefore, by definition of S and Lemma 3.19, we thus have that for every u ∈ Nv(v) \ S, the edge
{u, v} is taken in the matching with probability xuv. Thus, the probability that v is matched on
arrival is, as claimed, at least
Xu∈Nv(v)\S
xuv = Xu∈Nv(v)
xuv −Xu∈S
xuv > Xu∈Nv(v)
xuv − ε3 > (1 − ε2) Xu∈Nv (v)
xuv ,
where the last inequality holds because we havePu∈Nv(v) xuv > c, as calculated in (17).
3.4.5 Proof of the Key Lemma
It remains to prove the key lemma, Lemma 3.19, which we do here.
Proof of Lemma 3.19. For a certified primary path (P, T ) let EQ(P,T ) be the event as defined
in Lemma 3.15, and let IN(P,T ) be the event that P is a maximal primary path in Hτ and the
termination of P is certified by T . Further, let
C = {(P, T ) : (P, T ) is a certified primary path rooted at u with Pr[IN(P,T )] > 0}
21
be the set of certified primary paths rooted at u that have a nonzero probability of being maximal in
on Fu implies in particular that u has no incoming primary arc), we can rewrite the expression to
Hτ . Then, by the law of total probability and sinceP(P,T )∈C Pr[IN(P,T ) Fu] = 1 (since conditioning
bound,Pw∈Nv(v) z′
X(P,T )∈C
w · Pr[Fw Fu] −Pw∈Nv(v) z′
Pr[IN(P,T ) Fu] Xw∈Nv(v)
z′
w · Pr[Fw Fu, IN(P,T )] − Xw∈Nv(v)
z′
w · Pr[Fw] .
w · Pr[Fw], as
We analyze this expression in two steps. First, in the next claim, we show that we can focus on the
case when the certified path (P, T ) is very structured and equals the component of u in Hτ . We
then analyze the sum in that structured case.
Claim 3.20. Let P ⊆ C contain those certified primary paths (P, T ) of C that satisfy: P has length
less than 2000 · ln(1/ε) and z(B(P, T )) 6 2 ln(1/ε). Then, we have
(18)
(18) 6 X(P,T )∈P
Pr[EQ(P,T ) Fu] Xw∈Nv(v)
z′
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
z′
w · Pr[Fw] + ε1/3/2.
Proof. Define the following subsets of certified primary paths rooted at u:
C1 = {(P, T ) ∈ C P is of length at least 2000 · ln(1/ε)}
C2 = {(P, T ) ∈ C \ C1 z(B(P, T )) > 2 ln(1/ε)}
Note that P = C \ (C1 ∪ C2). Since u satisfies (15), we have that
X(P,T )∈C1
Pr[IN(P,T ) Fu] 6 ε2 6 ε1/3/6.
On the other hand, by Lemma 3.16 and Pr[Fu] > c (by Property (IV)), we have that
X(P,T )∈C2
Pr[IN(P,T ) Fu] 6 c−1 · X(P,T )∈C2
Pr[IN(P,T )] 6 c−1 · ε 6 ε1/3/6.
In other words, almost all probability mass lies in those outcomes where one of the certified paths
(P, T ) ∈ P is in Hτ .
It remains to prove that, in those cases, we almost always have that the
component of u in Hτ equals the path P (whose termination is certified by T ). Specifically, let
EQ(P,T ) denote the complement of EQ(P,T ). We show
PrhEQ(P,T ) IN(P,T )i 6 ε1/3/7 .
(19)
To see this, note that by the definition of the event IN(P,T ), if we restrict ourselves to primary
edges then the component of u in Hτ equals P . We thus have that for the event EQ(P,T ) to be true
at least one of the vertices in P must have an incoming or outgoing secondary edge. Hence the
expression PrhEQ(P,T ) IN(P,T )i can be upper bounded by
Pr[a vertex in P has an incoming or outgoing secondary arc in Gτ IN(P,T )]
(20)
22
Note that event IN(P,T ) is determined solely by choices of primary arcs. By independence of these
choices and choices of secondary arcs, conditioning on IN(P,T ) does not affect the distribution of
secondary arcs. So the probability that any of the nodes in P selects a secondary edge is at most
√ε. Thus, by union bound, the probability that any of the P 6 2000 · ln(1/ε) vertices in P pick
a secondary arc is at most √ε · 2000 · ln(1/ε). We now turn our attention to incoming secondary
arcs. First, considering the secondary arcs that go into u, we have
c 6 Pr[Fu] 6 Y(w,u)∈B(v,u)
(1 − z(w, u)/2) 6 exp(−z(B(v, u))/2),
because any arc (w, u) ∈ B(v, u) appears as a primary arc in Gτ independently with probability
at least z(w, u)/2 and the appearance of such an arc implies that u has an incoming primary arc
in Hτ and is therefore matched; i.e., the event Fu is false in this case. We thus have z(B(v, u)) 6
2 ln(1/c). Further, since (P, T ) 6∈ C2, we have z(B(P )) 6 z(B(P, T )) 6 2 ln(1/ε). Again using that
the conditioning on IN(P,T ) does not affect the distribution of secondary edges, we have that the
probability of an incoming secondary arc to any vertex in P is at most √ε · (2 ln(1/c) + 2 ln(1/ε)) .
Thus, by union bound, the probability that any vertex in P has an incoming or outgoing secondary
arc conditioned on IN(P,T ) is at most
√ε · 2000 · ln(1/ε) + √ε · (2 ln(1/c) + 2 ln(1/ε)) 6 ε1/3/7,
for sufficiently small ε, which implies (19) via (20).
We now show how the above concludes the proof of the claim. We have shown that each one of
w = 1). Hence,
the two sets C1,C2 contributes at most ε1/3/6 to (18) (where we use thatPw∈Nv(v) z′
(18) 6 X(P,T )∈P
w · Pr[Fw IN(P,T ), Fu] − Xw∈Nv(v)
Pr[IN(P,T ) Fu] Xw∈Nv(v)
w · Pr[Fw] + 2ε1/3/6.
This intuitively concludes the proof of the claim as (19) says that Pr[EQ(P,T )IN(P,T )] is almost 1.
The formal calculations are as follows. Since the event EQ(P,T ) implies the event IN(P,T ), we have
that
z′
z′
Pr[EQ(P,T )] = Pr[EQ(P,T ) ∧ IN(P,T )] = Pr[IN(P,T )] − Pr[EQ(P,T ) ∧ IN(P,T )],
which by (19) implies
Pr[EQ(P,T )] = Pr[IN(P,T )](cid:16)1 − PrhEQ(P,T ) IN(P,T )i(cid:17) > Pr[IN(P,T )](cid:16)1 − ε1/3/7(cid:17) .
w · Pr[Fw IN(P,T ), Fu] −Pw∈Nv(v) z′
Specifically, by law of total probability, it can be rewritten as the sum of the expressions (22)
and (23) below:
(21)
We use this to rewrite Pr[IN(P,T ) Fu](cid:16)Pw∈Nv(v) z′
Pr[EQ(P,T ) ∧ IN(P,T ) Fu] Xw∈Nv(v)
= Pr[EQ(P,T ) Fu] Xw∈Nv(v)
z′
z′
w · Pr[Fw](cid:17).
w · Pr[Fw]
(22)
w · Pr[Fw EQ(P,T ), IN(P,T ), Fu] − Xw∈Nv(v)
w · Pr[Fw]
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
z′
z′
23
and
Pr[EQ(P,T ) ∧ IN(P,T ) Fu] Xw∈Nv(v)
where (23) can be upper bounded as follows:
z′
w · Pr[Fw EQ(P,T ), IN(P,T ), Fu] − Xw∈Nv(v)
z′
w · Pr[Fw] , (23)
(23) 6 Pr[EQ(P,T ) ∧ IN(P,T ) Fu]
6 c−1 · Pr[EQ(P,T ) ∧ IN(P,T )]
= c−1 · Pr[IN(P,T )] · PrhEQ(P,T ) IN(P,T )i
6 c−1 ·
6 Pr[EQ(P,T )] · ε1/3/6.
1 − ε1/3/7 · (ε1/3/7)
Pr[EQ(P,T )]
(by Xw∈Nv(v)
z′
w 6 1)
(by c 6 Pr[Fu])
(by (19) and (21))
(for ε small enough)
As at most one of the events {EQ(P,T )}(P,T )∈P is true in any realization of Gτ , we have that
P(P,T )∈P Pr[EQ(P,T )∧ IN(P,T ) Fu] 6P(P,T )∈P(cid:16)Pr[EQ(P,T )] · ε1/3/6(cid:17) 6 ε1/3/6. Thus, again using
thatPw∈Nv(v) zw 6 1, we have that
w · Pr[Fw] + 2ε1/3/6
Pr[IN(P,T ) Fu] Xw∈Nv(v)
(18) 6 X(P,T )∈P
w · Pr[Fw] + 3ε1/3/6,
Pr[EQ(P,T ) Fu] Xw∈Nv(v)
6 X(P,T )∈P
w · Pr[Fw IN(P,T ), Fu] − Xw∈Nv(v)
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
z′
z′
z′
z′
as claimed.
The previous claim bounded the contribution of certified primary paths in C \ P to (18). The
following claim bounds the contribution of paths in P.
Claim 3.21. Let P ⊆ C contain those certified primary paths (P, T ) of C that satisfy: P has length
less than 2000 · ln(1/ε) and z(B(P, T )) 6 2 ln(1/ε). Then, we have
X(P,T )∈P
Pr[EQ(P,T )] Xw∈Nv(v)
z′
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
z′
w · Pr[Fw] 6 ε1/3/2.
Proof. We prove the claim in two steps: first we construct a chain of distributions that interpolates
between the unconditional distribution of Hτ and its conditional distribution, and then bound the
expected number of vertices that change their matched status along that chain. For the remainder
of the proof we fix the certified primary path (P, T ).
24
τ
Constructing a chain of distributions. Let H (0)
τ denote the unconditional distribution of Hτ
when v arrives, and let H (n)
denote the distribution of Hτ conditioned on EQ(P,T ) when v arrives.
Here n = V is the number of vertices in the input graph. For every w ∈ V let F (0)
w denote
the indicator of w being free when v arrives (unconditionally) and let F (n)
denote the indicator
w
variables of w being free when v arrives conditioned on EQ(P,T ). Note that F (0) is determined by
H (0)
that interpolate
between H (0)
and F (n) is determined by H (n)
as follows.
. For t = 0, . . . , n, we define distributions H (t)
τ
and H (n+1)
τ
τ
τ
τ
As in Lemma 3.15, for every w ∈ V we denote the unconditional distribution of its primary
choice by pw, and the unconditional distribution of its secondary choice by sw. Similarly, we
t+1, . . . , n are sampled independently from the unconditional distribution pwt. Similarly, secondary
denote the conditional distribution given EQ(P,T ) of the primary choice by epw and the conditional
distribution of the secondary choice by esw. For every t = 0, . . . , n the primary choice of vertices
wj, j = 1, . . . , t are sampled independently from epwj , and the primary choices of vertices wj, j =
choices of vertices wj, j = 1, . . . , t are sampled independently from eswj and secondary choices of
vertices wj, j = t + 1, . . . , n are sampled independently from swj . Note that H (0)
from the unconditional distribution of Hτ , and H (n)
is sampled from the conditional distribution
(conditioned on EQ(P,T )), as required, due to the independence of the conditional probabilities pwj
and swj , by Lemma 3.15. For t = 0, . . . , n let Mt denote the matching constructed by our algorithm
on H (t)
w be the indicator variable for w being free when v arrives in the DAG sampled
from H (t)
τ .
τ , and let F (t)
is sampled
τ
τ
Coupling the distributions of H (t)
Specifically, we will show that for every such t the following holds.
τ . We now exhibit a coupling between the H (t)
τ , t = 0, . . . , n.
(24)
where R(wt+1) is as defined in Lemma 3.15 with regard to the certified primary path R(P, T ).
Recall that z(R(wt+1)) is the total probability assigned to arcs leaving wt+1 which are ruled out
from being primary arcs in Gτ by conditioning on EQ(P,T ).
We construct the coupling by induction. The base case corresponds to t = 0 and is trivial. We
now give the inductive step (t → t + 1). We write w := wt+1 to simplify notation. Let Z p ∈ Nw(w)
denote the primary choice of w in H (t)
τ , and let Z s ∈ Nw(w) denote the secondary choice of w in
Nw(w) (they are sampled according to the unconditional distributions pw and sw respectively). Let
eZ p ∈ Nw(w) and eZ s ∈ Nw(w) be sampled from the conditional distributionsepw andesw respectively,
such that that the joint distributions (Z p,eZ p) and (Z s,eZ s) satisfy
First, we note that if Z p = eZ p and Z s = eZ s, then w = wt+1 is matched to the same neighbor
Pr[Z p 6= eZ p] = TV(pw,epw) and Pr[Z s 6= eZ s] = TV(sw,esw).
under H (t)
, and so Mt = Mt+1, due to the greedy nature of the matching constructed.
Otherwise, by Lemma 3.9, at most two vertices have different matched status in Mt and Mt+1 in
the latter case (in the former case every vertex has the same matched status). To summarize, we
τ and H (t+1)
(25)
τ
25
EXq∈V
F (t+1)
q
− F (t)
q
6 4z(R(wt+1)),
have, for R(w) determined by (P, T ) as in Lemma 3.15, that
EXq∈V
F (t+1)
q
− F (t)
q
This concludes the proof of the inductive step, and establishes (24). In particular, we get
(by (25) and union bound)
(by Lemma 3.15)
6 4z(R(w)).
6 2 · Pr[Z p 6= eZ p or Z s 6= eZ s]
6 2(TV(pw,epw) + TV(sw,esw))
EXq∈V
6
F (n)
q − F (0)
q
EXq∈V
n−1Xt=0
n−1Xt=0
F (t+1)
q
− F (t)
q
6
4z(R(wt+1))
(by (26))
= 4z(R(P, T )),
(26)
(27)
by the definition of R(P, T ) =Sw R(w) in Lemma 3.15.
ln(1/ε) and z(B(P, T )) 6 2 ln(1/ε) one has Pw z(R(w)) = z(R(P, T )) = O(ln(1/ε)). Indeed, by
We now finish the claim. First note that for any (P, T ) such that P has length at most 2000 ·
Lemma 3.15 and linearity of z, recalling that u is the root of P and that no vertex appears after v
(and thus B(v, u) = {(w, u) w arrives between u and v}), we have
z(R(P, T )) 6 z(B(P, T )) + z(B(v, u)) +
Xw∈P ∪{w: T =(w,w′)}
z ({w} × Nw(w)) .
(28)
We now bound the contribution to the above upper bound on Pw z(R(w)) = z(R(P, T )) in (28).
First, we have that z(B(P, T )) 6 2 ln(1/ε) by assumption of the lemma. To bound the contribution
of z(B(v, u)), we note that by Property IV, we have
c 6 Pr[Fu] = Ye∈B(v,u)
(1 − ze) 6 exp− X(w,u)∈B(v,u)
z(w, u)/2 6 exp(−z(B(v, u))/2),
because any arc e = (w, u) appears as a primary arc in Gτ with probability z′(w, u) > z(w, u)/2,
independently of other such arcs, and the appearance of any such an edge implies that u has an
incoming primary edge in Hτ when v arrives and is therefore matched; i.e., the event Fu is false in
this case. We thus have z(B(v, u)) 6 2 ln(1/c). Finally, it remains to note that for every one of the
at most 2000 · ln(1/ε) + 1 vertices w ∈ P ∪ {w : T = (w, w′)} the contribution of z({w} × Nw(w))
to the right hand side of (28) is at most 1 + Cε 6 2, by Lemma 3.10, (2). Putting these bounds
together, we get that for sufficiently small ε,
z(R(P, T )) 6 2 ln(1/ε) + 2 ln(1/c) + 2 · 2000 · ln(1/ε) + 2 = O(ln(1/ε)).
(29)
26
The term we wish to upper bound is at most
z′
z′
w∈Nv(v)
z′
w · Pr[Fw]
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
w(cid:19) · Xw∈Nv(v)(cid:12)(cid:12)(cid:12)Pr[Fw EQ(P,T )] − Pr[Fw](cid:12)(cid:12)(cid:12)
Xw∈Nv(v)
6(cid:18) max
6C√ε · Xw∈Nv(v)(cid:12)(cid:12)(cid:12)Pr[Fw EQ(P,T )] − Pr[Fw](cid:12)(cid:12)(cid:12)
=C√ε · E Xw∈Nv(v)
F (n)
w − F (0)
w
w #
6C√ε · E"Xw∈V
w − F (0)
F (n)
6C√ε · z(R(P, T ))
=O(√ε · log(1/ε))
6ε1/3/2,
then, using (27) and (29), we find that the term we wish to upper bound is at most
(by Lemma 3.10, (3))
(by definition of F (0) and F (n))
(by (27))
(by (29))
completing the proof.
Finally, we obtain Lemma 3.19 by combining Claim 3.20 and Claim 3.21, to find that, as claimed
(18) 6 X(P,T )∈P
Pr[EQ(P,T ) Fu] Xw∈Nv(v)
z′
w · Pr[Fw EQ(P,T )] − Xw∈Nv(v)
z′
w · Pr[Fw] + ε1/3/2
6 ε1/3/2 + ε1/3/2 = ε1/3.
3.4.6 Bounding the Impact of Bad Vertices
In this section, we show that we can completely ignore the bad vertices without losing too much.
From the definition of good vertices, for a bad vertex v, we have that
Pr
τ
[Hτ has a primary path rooted at v of length at least 2000 · ln(1/ε)] > ε6.
As the main result of this section, we prove the following theorem:
To prove this, we first describe a charging mechanism in which, for each bad vertex, a charge of
one is distributed among a subset of other vertices. Then, using the following supplementary lemma,
Theorem 3.22. The number of bad vertices is at most ε3 ·Pe∈E xe.
we show that the total distributed charge over all vertices in the graph is at most ε3 ·P(u,v)∈E xuv.
Lemma 3.23. We call a primary path P a primary predecessor path of v if it ends at v. That is,
P = vℓ → vℓ−1 → ··· → v1 = v. We have
Pr
τ
[v has any primary predecessor path P with z(B(P )) 6 20 · ln(1/ε) and P > 1000 · ln(1/ε)] 6 ε10.
27
Proof. We use the principle of deferred decisions and traverse the path backwards. Let b be the
current vertex, which is initially set to v. Consider all incoming arcs to b, say (a1, b), . . . , (ak, b)
where we index a's by time of arrival; i.e., ai arrives before aj if i < j (and b arrived before any ai).
First consider the random choice of a1 and see if it selected the arc (a1, b).
• If it does, then the path including b in Hτ will use the arc (a1, b).
• Otherwise, if a1 does not select the arc (a1, b), then go on to consider a2 and so on.
If no a1, . . . , ak selects b, then the process stops; i.e., the primary path starts at this vertex since
b has no incoming primary arc. Otherwise let i be the first index so that (ai, b) was selected.
Then (ai, b) is in the primary path ending at v in Hτ . Now, observe that no a1, . . . , ai−1 may be
in the path in this case, because these vertices arrived before ai and after b. Moreover, we have
not revealed any randomness regarding ai+1, . . . , ak that may appear later in the path. We can
therefore repeat the above process with b now set to ai and "fresh" randomness for all vertices we
consider, as the random choices of arcs of all vertices are independent. We now show that this
process, with good probability, does not result in a long predecessor path P of low z(B(P )) value.
Recall from Lemma 3.10, (5), that z(u, v) 6 3/5 for all (u, v) ∈ V × V . Suppose that
i=1 z(ai, b) 6
i=1(1−z(ai, b)) >
i=1 z(ai, b) > 1/5. Consequently, with probability at least 1/5, vertex b either has no prede-
Pk
i=1 z(ai, b) > 4/5. Let j be the first index such that Pj
4/5, and hence the probability that none of the first j vertices select b is at leastQj
1 −Pj
In the other case, we havePk
i=1(1 − z(ai, b)) > 1 −Pk
isQk
Therefore, at any step in the above random process, with probability at least 1/5, we either
stop or increase z(B(P )) by 1/5. Let Zi be an indicator variable for the random process either
stopping or increasing z(B(P )) by at least 1/5 at step i, and notice that according to the above
random process, each Zi is lower bounded by an independent Bernoulli variable with probability
i=1 z(ai, b) > 1/5. ThusPj
i=1 z(ai, b) 6 4/5. Then the probability that b has no predecessor
1/5. Thus if we define Z =Pi∈[1000·ln(1/ε)] Zi, we have E[Z] > 200 · ln(1/ε), and thus by standard
coupling arguments and Chernoff bounds, we have that
cessor or the increase to z(B(P )) is at least 1/5.
i=1 z(ai, b) > 1/5.
Pr[Z 6 100 · ln(1/ε)] 6 Pr [Z 6 (1 − 1/2) · E[Z]] 6 e−(1/2)·(1/2)2 ·200·ln(1/ε) 6 ε10.
But if the path does not terminate within 1000·ln(1/ε) steps and Z > 100·ln(1/ε), then z(B(P )) >
20 · ln(1/ε).
We now prove Theorem 3.22.
Proof of Theorem 3.22. By Lemma 3.16, the probability that Hτ has a primary path P with
z(B(P )) > 20 · ln(1/ε) starting at v is at most ε10. Thus, for a bad vertex u, the probability
that Hτ has some primary path P rooted at u with P > 2000 · ln(1/ε) and z(B(P )) 6 20· ln(1/ε)
is at least ε6 − ε10 > ε6/2.
Let k = 20 · ln(1/ε) and ℓ = 2000 · ln(1/ε). Let Pu be the set of all primary paths P rooted at
u such that z(B(P )) 6 k and P = ℓ starting at u. Since all such primary paths with length more
than ℓ are extensions of those with length exactly ℓ, we have PP ∈Pu Pr[P is in Hτ ] > ε6/2. For
each such path P ∈ Pu, consider the two vertices wP
ℓ−1 at distances ℓ and ℓ− 1 respectively
from u. For each such vertex wP
. Then the
sum of these charges is
j (j ∈ {ℓ − 1, ℓ}), charge (2/ε6) · Pr[P is in Hτ ] · ywP
ℓ and wP
j
XP ∈Pu
(2/ε6) · Pr[P is in Hτ ] · (ywP
ℓ
>1
{z
+ ywP
ℓ−1
28
Pr[P is in Hτ ] > 1.
> (2/ε6) · XP ∈Pu
)
}
Notice that the fact (ywP
vertex cover problem).
ℓ
+ ywP
ℓ−1
) > 1 follows because yw's form a feasible dual solution (to the
As Q = ℓ − 1 > 1000 · ln(1/ε) for all Q ∈ Qw, by Lemma 3.23,PQ∈Qw Pr[Q is in Hτ ] 6 ε10 . For
On the other hand, consider how many times each vertex is charged. For this, for every vertex
w, let Qw be the set of primary predecessor paths Q of u such that Q = ℓ − 1 and z(B(P )) 6 k.
a primary predecessor path Q ∈ Qw (or one of its extensions), the vertex w can be charged at most
twice according to the above charging mechanism. Since any predecessor path of w with length
more than ℓ − 1 must be an extension of one with length exactly ℓ − 1, we have that the amount
w is charged is at most
XQ∈Qw
2 · 2 · Pr[Q is in Hτ ] · yw/ε6 6 4 · (ε10/ε6) · yw 6 4 · ε4 · yw.
Summing over all w ∈ V and using Lemma 3.3, the total charge is at most
Xw∈v
4 · ε4 · yw 6 4 · ε4 · β ·Xe∈E
xe 6 ε3Xe∈E
xe.
3.4.7 Calculating the Competitive Ratio of Algorithm 3
We now show that the competitive ratio of Algorithm 3 is indeed (1/2 + α) competitive for some
sufficiently small absolute constant α > 0, thus proving Theorem 1.2. This essentially combines
the facts that for good vertices, the matching probability is very close to the fractional values of
incident edges, and that the number of bad vertices is very small compared to the total value of
the fractional algorithm (over the entire graph).
Proof of Theorem 1.2. Let OPT denote the size of the maximum cardinality matching in the input
graph G. Then, by Lemma 3.5 and our choice of f = f1+2ε and β = 2 − ε > β∗(f1+2ε), we have
thatPe xe > (1/β)· OPT > (1/2 + ε/4)· OPT, where the xe's are the fractional values we compute
in Algorithm 3.
Now let M be the matching output by Algorithm 3. We have
xuv
Pr[e is matched]
(1 − ε2) · Xu∈Nv(v)
E[M] =Xe∈E
> Xgood v∈V
> (1 − ε2) ·Xe∈E
> (1 − ε2) · Xe∈E
> (1 − ε2) · Xe∈E
> (1 − 2ε2) ·Xe∈E
> (1 − 2ε2) · (1/2 + ε/4) · OPT
> (1/2 + ε/5) · OPT,
xe − Xbad v∈V Xu∈Nv(v)
1!
xe − Xbad v∈V
xe!
xe − ε3Xe∈E
xe
xuv
(By Theorem 3.18)
( Xu∈Nv(v)
xuv 6 1)
(By Theorem 3.22)
where the last line holds for a sufficiently small constant ε > 0.
29
Appendix
A Deferred Proofs of Section 3.3
Here we prove that a change of the realized arc choices of any vertex does not change the matched
status of more than two vertices (at any point in time). This is Lemma 3.9, restated below.
Lemma 3.9. Let Gτ and Gτ ′ be two realizations of the random digraph where all the vertices in the
two graphs make the same choices except for one vertex v. Then the number of vertices that have
different matched status (free/matched) in the matchings computed in Hτ and Hτ ′ at any point of
time is at most two.
Proof. We consider the evolution, following each vertex arrival, of the matchings Mτ and Mτ ′
computed in Hτ and Hτ ′, respectively, as well as the set of vertices with different matched status
in these matchings, denoted by D := (Mτ \ Mτ ′) ∪ (Mτ ′ \ Mτ ). The set D is empty before the first
arrival and remains empty until the arrival of v, as all earlier vertices than v have the same primary
and secondary arcs and have the same set of free neighbors in Hτ and Hτ ′ (as D = ∅, by induction).
Now, if immediately after v arrives it remains free in both Mτ and Mτ ′, or it is matched to the
same neighbor in both matchings, then clearly D remains empty. Otherwise, either v is matched to
different neighbors in Mτ and Mτ ′, or v is matched in one of these matchings but not in the other.
Both these cases result in D = 2. We now show by induction that the cardinality of D does not
increase following subsequent arrivals, implying the lemma.
Let u be some vertex which arrives after v. If when u arrives u is matched to the same neighbor
w in Mτ and Mτ ′ or if u remains free in both matchings, then D is unchanged. If u is matched
to some w on arrival in Mτ , but not in Mτ ′, then since the arcs of u are the same in Gτ and Gτ ′,
this implies that w must have been free in Mτ but not in Mτ ′, and so D ∋ w. Therefore, after u
arrives, we have D ← (D\{w})∪{u}, and so D's cardinality is unchanged. Finally, if u is matched
to two distinct neighbors, denoted by w and w′, respectively, then one of (u, w) and (u, w′) must
be the primary arc of u in both Gτ and Gτ ′. Without loss of generality, say (u, w) is this primary
arc. Since u is matched to w in Mτ but not in Mτ ′, then w must be free in Mτ when u arrives, but
not in Mτ ′, and so D ∋ w. Consequently, we have that after u arrives we have D ← S for some set
S ⊆ (D \ {w}) ∪ {w′}, and so D's cardinality does not increase.
B Deferred Proofs of Section 3.1
Here we prove the bound on the fractional degree xu in terms of its dual value, restated below.
Lemma 3.3. For any vertex u, v ∈ V , let yu be the potential of u prior to arrival of v. Then the
fractional degree just before v arrives, xu :=Pw∈Nv(u) xuw, is bounded as follows:
yu
β
6 xu 6
yu + f (1 − yu)
.
β
Proof. Let y0 be u's potential after u's arrival. For the lower bound, note that it suffices to prove
that every increase in the fractional degree is bounded below by the increase in the potential divided
by β. When vertex u first arrived, we consider two cases.
1. y0 > 0 (thus y0 = 1 − θ > 0, and so θ < 1), then the increase in u's fractional degree was:
Xv∈Nu(u)
(θ − yv)+
β
(cid:18)1 +
1 − θ
f (θ)(cid:19) =
30
f (θ) + 1 − θ
β
=
f (1 − y0) + y0
β
>
y0
β
.
2. y0 = 0 (thus θ = 1), then the increase in u's fractional degree was:
(θ − yv)+
(θ − yv)+
> 0 =
y0
β
.
β
Xv∈Nu(u)
β
(cid:18)1 +
1 − θ
f (θ)(cid:19) = Xv∈Nu(u)
f (θ)(cid:19) >
(cid:18)1 +
1 − θ
(θ − yold
u )+
β
,
(θ − yold
u )+
β
For every subsequent increase of the fractional degree due to a newly-arrived vertex we have that:
Which concludes the proof for the lower bound.
For the upper bound, by [25, Invariant 1], we have that
β · xu 6 yc + f (1 − y0) +Z yc
subtracting 1 + f (1 − yu) and writing the integral R yu
f (x) dx -R 1
R 1
1−x
f (x) dx, and relying on Equation (5), we find that
1−x
yu
y0
y0
This upper bound can be simplified by using Equation (5), as follows. Taking (30), adding and
1−x
f (x) dx as the difference of two integrals
y0
1 − x
f (x)
dx.
(30)
y0
dx
1 − x
f (x)
1 − x
f (x)
β · xu 6 yc + f (1 − y0) +Z yc
=(cid:18)1 + f (1 − y0) +Z 1
= β∗(f ) + yc −(cid:18)1 + f (1 − yc) +Z 1
= β∗(f ) + yc − β∗(f ) + f (1 − yc)
= yc + f (1 − yc),
yc
y0
dx(cid:19) − 1 + yc +Z yc
1
1 − x
f (x)
dx
1 − x
f (x)
dx(cid:19) + f (1 − yc)
from which the lemma follows.
C Deferred Proofs of Section 3.4.2
In this section we present the proofs deferred from Section 3.4.2. We start by presenting a more
manageable form for the function f = f1+2ε which we use.
A function in the WW family is determined by a parameter k > 1 and takes the following form
fκ(θ) =(cid:18) 1 + κ
2 − θ(cid:19) 1+κ
2κ (cid:18)θ +
Letting κ = 1 + 2ε, we get that f := fκ is of the form
2κ
κ − 1
2 (cid:19) κ−1
.
1+ε
ε
1+2ε
f (θ) = (1 + ε − θ)
1+2ε · (θ + ε)
= (1 + ε − θ) ·(cid:18) θ + ε
1 + ε − θ(cid:19) ε
1+2ε
.
Clearly this is water filling when ε = 0 and otherwise we have that the first term is like water filling
and then the second term is less than 1 for z 6 1/2 and greater than 1 if z > 1/2.
By Taylor expansion, we obtain the following more manageable form for f .
31
1+2ε
i!
= (1 + ε − θ) ·
Lemma C.1. There exists ε0 ∈ (0, 1) such that for every ε ∈ (0, ε0) and every θ ∈ [0, 1], we have
Proof. Taking the Taylor expansion of ex, we find that
1 + ε − θ(cid:19)(cid:19) + 1.01ε.
f (θ) = (1 + ε − θ) ·(cid:18) θ + ε
f (θ) 6 (1 − θ)(cid:18)1 + ε ln(cid:18) θ + ε
1 + ε − θ(cid:19) ε
1+2ε(cid:17)i
1+ε−θ(cid:17) ·
(cid:16)ln(cid:16) θ+ε
∞Xi=0
= (1 + ε − θ)(cid:18)1 + ln(cid:18) θ + ε
1 + 2ε(cid:19) + o(ε)
1 + ε − θ(cid:19) ·
1 + ε − θ(cid:19) ·
= (1 + ε − θ) + (1 − θ) ln(cid:18) θ + ε
= (1 + ε − θ) + (1 − θ)ε ln(cid:18) θ + ε
1 + ε − θ(cid:19) + o(ε)
= (1 − θ)(cid:18)1 + ε ln(cid:18) θ + ε
1 + ε − θ(cid:19)(cid:19) + ε + o(ε).
6 2
To be precise, for θ ∈ [0, 1] and 0 < ε 6 ε0 6 1 (implying for example θ+ε
ε ), we will show that
terms dropped in the third, fourth and fifth lines are all at most some O((ln( 1
ε ) · ε)2) = o(ε), from
which the lemma follows as the sum of these terms is at most 0.01ε for ε 6 ε0 and ε0 sufficiently
small.
1+ε−θ
ε
1 + 2ε
+ o(ε)
ε
ε
Indeed, in the third line, we dropped
(1 + ε − θ) ·
∞Xi=2
1+ε−θ(cid:17) ·
(cid:16)ln(cid:16) θ+ε
i!
ε
1+2ε(cid:17)i
6 2 ·
∞Xi=2
(ln( 2
ε ) · ε)i
i!
6 ·
∞Xi=2 (cid:0)ln(cid:0) 2
ε(cid:1) · ε(cid:1)i
i2
= O((ln(1/ε) · ε)2),
where the last step used that ln(1/ε) · ε 6 1 holds for all ε > 0. In the fourth line, we dropped
ε
1 + 2ε
6 ε2 · ln (2/ε) = O((ln(1/ε) · ε)2).
Finally, in the fifth line, we dropped
ε · ln(cid:18) θ + ε
1 + ε − θ(cid:19) ·
1 + 2ε(cid:19) · ln(cid:18) θ + ε
ε
(1 − z) ·(cid:18)ε −
1 + ε − θ(cid:19) 6 1 · (ε2/(1 + 2ε)) · ln (2/ε) = O((ln(1/ε) · ε)2).
Given this more manageable form for f , we can now turn to prove Lemma 3.10, restated below.
Lemma 3.10. (Basic bounds on conditional probabilities zu) There exist absolute constants c ∈
(0, 1) and C > 1/c > 1 and ε0 ∈ (0, 1) such that for every ε ∈ (0, ε0) the following holds: for every
vertex v ∈ V , if yu is the dual variable of a neighbor u ∈ Nv(v) before v's arrival and θ is the value
chosen by Algorithm 1 on v's arrival, then for zu as defined in Algorithm 3, we have:
(1) If θ 6∈ (c, 1 − c), thenPu∈Nv(v) zu 6 1,
(2) If θ ∈ [0, 1], thenPu∈Nv(v) zu 6 1 + Cε,
32
(3) IfPu∈Nv(v) zu > 1, then zu 6 C√ε for every u ∈ Nv(v),
(4) IfPu∈Nv(v) zu > 1, then for every u ∈ Nv(v) such that zu > 0, one has yu ∈ [c/2, 1− c/2], and
(5) For all u ∈ Nv(v), one has zu 6 1/2 + O(√ε).
Proof. We begin by getting a generic upper bound for zu. We note that each edge e is matched by
Algorithm 3 with probability at most xe by Line 12. Therefore, u is matched before v arrives with
probability at most xu :=Pw∈Nv(u)\{v} xwu, the fractional degree of u before v arrives. Therefore,
by Lemma 3.3, the probability that u is free is at least
Pr[u free when v arrives] > 1 − xu > 1 −
yu + f (1 − yu)
β
,
(31)
from which, together with the definition of xuv = 1
upper bound on zu:
β (θ − yu)+(cid:16)1 + 1−θ
zu =
xuv
Pr[u is free when v arrive]
6
1
β (θ − yu)+(cid:16)1 + 1−θ
f (θ)(cid:17)
1 − yu+f (1−yu)
β
=
f (θ)(cid:17), we obtain the following
(θ − yu)(cid:16)1 + 1−θ
f (θ)(cid:17)
(32)
β − (yu + f (1 − yu))
.
We start by upper boundingPu∈Nv(v) zu, giving a bound which will prove useful in the proofs
of both (1) and (2). Recall that θ is defined as the largest θ 6 1 such that
Xu∈Nv(v)
(θ − yu)+ 6 f (θ).
(33)
Summing (32) over all u ∈ Nv(v), we find that
Xu∈Nv(v)
zu 6 Xu∈Nv(v)
(θ − yu)+ · (1 + 1−θ
f (θ) )
β − (θ + f (1 − θ))
(f (·) is non-increasing, by Observation 3.4)
f (θ) + 1 − θ
6
β − (θ + f (1 − θ))
We therefore wish to upper bound f (θ)+1−θ
proceeding to the proof, it would be useful to summarize some properties of the function γ(θ, ε).
(by (33) and β > β∗(f ) = 1 + f (0) > θ + f (1 − θ))
1+ε−θ(cid:17). Before
β−θ−f (1−θ) . To this end let γ(θ, ε) := ε ln(cid:16) θ+ε
1. γ(θ, ε) = −γ(1 − θ, ε) for all θ ∈ [0, 1] .
2. For c, ε0 sufficiently small we have for all θ ∈ [0, c) that γ(θ, ε) 6 ε ln(cid:16) c+ε
for all θ ∈ (1 − c, 1] that γ(θ, ε) > ε ln(cid:16) 1−c+ε
1+ε−(1−c)(cid:17) > 20 · ε.
1+ε−c(cid:17) 6 −20 · ε, and
3. γ(θ, ε) · (1 − 2θ) 6 0 for θ ∈ [0, 1], since γ(θ, ε) 6 0 for θ 6 1/2 and γ(θ, ε) > 0 for θ > 1/2.
4. θ · γ(θ, ε) > −ε for all θ ∈ [0, 1].
The last property follows from ln(cid:16) 1+ε−θ
θ+ε (cid:17) 6 ln(cid:16) 1+ε+θ
implies in particular that θ · γ(θ, ε) = θ · ε ·(cid:16)− ln(cid:16) 1+ε−θ
θ+ε (cid:17) 6 ln(cid:16)1 + 1
θ+ε (cid:17)(cid:17) > −ε.
θ+ε(cid:17) 6 1
θ+ε
6 1
θ , which
33
We will use γ as shorthand for γ(θ, ε). Recalling that β = 2− ε and using Lemma C.1, we have:
f (θ) + 1 − θ
β − (θ + f (1 − θ))
6
6
(1 − θ)(cid:16)1 + ε ln(cid:16) θ+ε
2 − ε − θ − θ(cid:16)1 + ε ln(cid:16) 1−θ+ε
1+ε−θ(cid:17)(cid:17) − θ + 1 + 1.01ε
θ+ε (cid:17)(cid:17) − 1.01ε
(34)
(1 − θ)(2 + γ) + 2ε
2 − 2θ + θγ − 3ε
γ(1 − 2θ) + 5ε
2 − 2θ + θγ − 3ε
.
= 1 +
20 ) = γ
4 + 5 · (− γ
We will continue by proving that the second term is negative. First we prove that the de-
nominator is positive. To this end, first consider the case when θ ∈ [0, c). In this case for ε0, c
sufficiently small one has that: 2 − 2θ + θγ − 2ε > 2 − 2θ − ε − 2ε > 0 from Item 4. More-
over, when θ ∈ (1 − c, 1] one has that θ > 1
2 (since c is small) and γ > 20ε from Item 2. Thus
2 − 2θ + θγ − 2ε > θγ − 2ε > 1
2 · 20ε − 3ε = 7ε > 0. Now, it remains to prove that the numerator
is always negative. When θ ∈ [0, c) we have that 1 − 2θ > 3/4(since c is small) and γ 6 −20ε from
Item 2, therefore γ(1 − 2θ) + 5ε 6 γ · 3
2 < 0. In the case where θ ∈ (1 − c, 1],
we have that 1 − 2θ < −3/4, and θ > 1/2 (since c is small), and γ > 20ε from Item 2, thus
γ(1 − 2θ) + 5ε 6 − 3
We now turn to (2). We assume that θ ∈ (c, 1 − c), since otherwise the claim is trivial, by (1).
2−2θ+θγ−3ε . We have that γ(1− 2θ) + 5ε 6 5ε from Item 3.
We have by (34) that
Furthermore, using Item 4 we have that 2 − 2θ + θγ − 3ε > 2c + −4ε > c for a sufficiently small ε0.
Overall, the second term is bounded above by 5
We now prove (3). Note that by (1), Pu∈Nv(v) zu > 1 implies that θ ∈ (c, 1 − c). Now, for
every u ∈ Nv(v), let αu := (θ−yu)+
definition of αu and our choice of θ, we have Pu∈Nv(v) αu =Pu∈Nv(v)
, so that yu = θ − f (θ) · αu if yu 6 θ. We also note that by
6 1. In the proof
of (3) and (4) we will assume for notational simplicity that all u ∈ Nv(v) have yu 6 θ, implying
zu > 0. Summing up (32) over all u ∈ Nv(v) and substituting in αu, we thus find that
c · ε < C · ε, for C > 5
4 · 20ε + 5ε = −10ε < 0.
6 1 + γ(1−2θ)+5ε
c > 1
c as required.
f (θ)+1−θ
β−(θ−f (1−θ))
(θ−yu)+
f (θ)
f (θ)
Xu∈Nv(v)
zu 6 Xu∈Nv(v)
= Xu∈Nv(v)
6 Xu∈Nv(v)
6 Xu∈Nv(v)
αu ·
αu ·
αu ·
(θ − yu)+(1 + 1−θ
f (θ) )
β − (yu + f (1 − yu))
f (θ) + 1 − θ
β − (yu + f (1 − yu))
f (θ) + 1 − θ
2 − yu − f (1 − yu) − 2.01ε
f (θ) + 1 − θ
2 − 4ε − 2yu
,
(by Lemma C.1 and β = 2 − ε)
In the last transition we used again (as in Item 4) that yu · ε ln(cid:16) 1−yu+ε
f (θ) 6 1 − θ + ε for all θ ∈ [0, 1]. Substituting yu = θ − f (θ) · αu into the above upper bound on
yu+ε (cid:17) 6 ε, which implies
34
Pu∈Nv(v) zu, we get
Xu∈Nv(v)
zu 6 Xu∈Nv(v)
= Xu∈Nv(v)
αu ·
αu ·
f (θ) + 1 − θ
2 − 4ε − 2θ + 2f (θ) · αu
f (θ) + 1 − θ
2 − 4ε − 2θ − Xu∈Nv(v)
a − b
(f (θ) + 1 − θ) · 2f (θ) · α2
u
(2 − 4ε − 2θ) · (2 − 4ε − 2θ + 2f (θ) · αu)
,
(35)
(36)
(37)
using the elementary identity 1
a(a+b) for appropriate a and b. Now, both terms in the
last line of (35) can be significantly simplified, as follows. For the former term, again using that
a+b = 1
f (θ) 6 1 − θ + ε, together withPu∈Nv(v) αu 6 1 noted above, we find that
αu ·(cid:18)1 +
Xu∈Nv(v)
f (θ) + 1 − θ
2 − 4ε − 2θ
2 + ε − 2θ
2 − 4ε − 2θ
= Xu∈Nv(v)
6 Xu∈Nv(v)
αu ·
αu ·
5ε
2 − 4ε − 2θ(cid:19) 6 1 + O(ε),
where in the last step we used that θ 6 1 − c and c is some fixed constant. For the second term in
the last line of (35), we note that
Xu∈Nv(v)
(f (θ) + 1 − θ) · 2f (θ) · α2
u
(2 − 4ε − 2θ) · (2 − 4ε − 2θ + 2f (θ) · αu)
= Ω(1) · Xu∈Nv(v)
α2
u .
u > Ω(α2
1+ε−c(cid:17) ε
u), since f is decreasing by Observation 3.4 and f (c) > 1
To see this, first note that for θ ∈ (c, 1 − c), the numerator of each summand of the LHS is at least
2f (c)2 · α2
2 · (1 + ε − c) > Ω(1)
for c and ε sufficiently small. To verify the first inequality of this lower bound for f (c), recall that
f (c) = (1 + ε − c) ·(cid:16) c+ε
1+2ε
1+2ε . Now, for ε tending to zero and c < 1/2, the term (cid:16) θ+ε
tends to one as ε tends to zero. Therefore for ε sufficiently small we have f (c) > 1
2 · (1 + ε − c)
for all c < 1/2. We now turn to upper bounding the denominator of each summand in the LHS of
Equation (37). Indeed, substituting yu = θ − f (θ) · αu, we find that each such denominator is at
most (2− 4ε− 2θ)· (2− 4ε− 2θ + 2f (θ)· αu) 6 (1/2)· (2− 4ε− 2yu) 6 (1/2)· (2− 4ε− 2c) 6 O(1) for
c and ε sufficiently small. Note that both numerator and denominator are positive for sufficiently
small c and ε0. Substituting the bounds of (36) and (37) into (35), we obtain
1+ε−θ(cid:17) ε
u .
From Eq. (38) andPu∈Nv(v) zu > 1 by assumption of (3), we get that
zu 6 1 + O(ε) − Ω(1) · Xu∈Nv(v)
Xu∈Nv(v)
α2
′
Xu∈Nv(v)
α2
u 6 C
ε
(38)
(39)
35
for an absolute constant C ′ > 1, since otherwisePu∈Nv(v) zu 6 1. Finally, it remains to note that
Xu∈Nv(v)
z2
α2
β − (yu + f (1 − yu))(cid:19)2
u = Xu∈Nv(v)(cid:18) αu · (f (θ) + 1 − θ)
u ·(cid:18) f (θ) + 1 − θ
6 Xu∈Nv(v)
β − (θ + f (1 − θ))(cid:19)2
6 Xu∈Nv(v)
u ·(cid:18) f (θ) + 1 − θ
β − (1 − c + f (c))(cid:19)2
u ·(cid:18) 1 − θ + ε + 1 − θ
6 Xu∈Nv(v)
β − (1 − c + 1 − c + ε)(cid:19)2
6 Xu∈Nv(v)
u ·
2c − 2ε
α2
α2
α2
2
6 Cε,
(by Observation 3.4 and yu 6 θ)
(by Observation 3.4 and θ 6 1 − c)
(f (c) 6 1 − c + ε)
2c−2ε . Thus z2
for some constant C > 2
u 6Pu∈Nv(v) zu 6 Cε and so zu 6 √C · ε 6 C√ε, as claimed.
We now prove (4). SincePu∈Nv(v) zu > 1 implies θ ∈ (c, 1 − c) by (1), using the definition of
αu's from the proof of (3) together with the fact that αu 6 C ′√ε for every u ∈ Nv(v) by (39) and
the fact that f (θ) 6 2 for all θ ∈ [0, 1] (by Lemma C.1), we get that
yu = θ − f (θ) · αu ∈ [c − O(√ε), 1 − c] ⊆ [c/2, 1 − c/2],
for sufficiently small ε0 > 0, as required.
As for (5), simplifying (32) and using the fact that θ − yu 6 f (θ), we get
zu 6
θ − yu + 1 − θ
β − yu − f (1 − yu)
=
1 − yu
.
β − yu − f (1 − yu)
which implies the following:
Recall from Lemma C.1 that for all θ ∈ [0, 1], we have f (θ) 6 (1− θ)(cid:16)1 + ε ln(cid:16) θ+ε
1. For all θ ∈ [0, 1], we have f (θ) 6 1 − θ + √ε, and
2. For θ < e−10, we have f (θ) 6 (1 − θ)(1 + ε(ln((e−10 + ε)/(1 − e−10 + ε)) + 1.01ε 6 1 − θ − 2ε.
1+ε−θ(cid:17)(cid:17)+ 1.01ε,
Suppose that yu 6 1 − e−10. Then using Item 1, we have
zu 6
6
1 − yu
β − yu − f (1 − yu)
2(1 − yu) − 2√ε
1 − yu
6
2 − ε − yu − yu − √ε
1 − yu
2√ε
6 1/2 +
2e−10 − 2√ε
6 1/2 + O(√ε).
Now suppose that yu > 1 − e−10. Then 1 − yu < e−10, and so by Item 2, f (1 − yu) 6 1 − yu − 2ε.
Thus we have
zu 6
1 − y
β − yu − f (1 − yu)
completing the proof.
6
1 − yu
2 − ε − yu − (yu − 2ε)
=
1 − yu
2(1 − yu) + ε
6 1/2,
36
Finally, we rely on Lemma C.1 to prove that the fractional solution maintained by Line 4 is
1/β competitive, as implied by Lemma 3.5 and the following restated fact.
Fact 3.12. For all sufficiently small ε > 0, we have that 2 − ε > β∗(f1+2ε).
Proof. Let us denote as before f = f1+2ε. Recall that β∗(f ) = 1 + f (0). By Lemma C.1, this is at
most 1+f (0) 6 1+(cid:16)1 + ε ln(cid:16) ε
1+ε(cid:17)(cid:17)+1.01ε. But for small enough ε, we have that ln(cid:16) ε
1+ε(cid:17) 6 −2.01,
implying that 1 + f (0) 6 2 − ε, as claimed.
References
[1] Ageev, A. A. and Sviridenko, M. I. 2004. Pipage rounding: A new method of constructing
algorithms with proven performance guarantee. Journal of Combinatorial Optimization 8, 3,
307 -- 328.
[2] Birnbaum, B. and Mathieu, C. 2008. On-line bipartite matching made simple. ACM
SIGACT News 39, 1, 80 -- 87.
[3] Buchbinder, N., Segev, D., and Tkach, Y. 2018. Online algorithms for maximum cardi-
nality matching with edge arrivals. Algorithmica, 1 -- 19.
[4] Chiplunkar, A., Tirodkar, S., and Vishwanathan, S. 2015. On randomized algorithms
In Proceedings of the 23rd Annual European
for matching in the online preemptive model.
Symposium on Algorithms (ESA). 325 -- 336.
[5] Cohen, I. R. and Wajc, D. 2018. Randomized online matching in regular graphs. In Pro-
ceedings of the 29th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). 960 -- 979.
[6] Devanur, N. R., Jain, K., and Kleinberg, R. D. 2013. Randomized primal-dual analysis of
ranking for online bipartite matching. In Proceedings of the 24th Annual ACM-SIAM Symposium
on Discrete Algorithms (SODA). 101 -- 107.
[7] Eden, A., Feldman, M., Fiat, A., and Segal, K. 2018. An economic-based analysis of
ranking for online bipartite matching. arXiv preprint arXiv:1804.06637 .
[8] Edmonds, J. 1965a. Maximum matching and a polyhedron with 0, 1-vertices. Journal of
research of the National Bureau of Standards B 69, 125-130, 55 -- 56.
[9] Edmonds, J. 1965b. Paths, trees, and flowers. Canadian Journal of mathematics 17, 3, 449 --
467.
[10] Epstein, L., Levin, A., Segev, D., and Weimann, O. 2013. Improved bounds for online
preemptive matching. In Proceedings of the 30th International Symposium on Theoretical Aspects
of Computer Science (STACS). 389.
[11] Feige, U. 2018.
Tighter bounds for online bipartite matching.
arXiv preprint
arXiv:1812.11774 .
[12] Gandhi, R., Khuller, S., Parthasarathy, S., and Srinivasan, A. 2006. Dependent
rounding and its applications to approximation algorithms. Journal of the ACM (JACM) 53, 3,
324 -- 360.
37
[13] Goel, G. and Mehta, A. 2008. Online budgeted matching in random input models with
applications to adwords. In Proceedings of the 19th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA). 982 -- 991.
[14] Guruganesh, G. P. and Singla, S. 2017. Online matroid intersection: Beating half for ran-
dom arrival. In Proceedings of the 19th Conference on Integer Programming and Combinatorial
Optimization (IPCO). 241 -- 253.
[15] Huang, Z., Kang, N., Tang, Z. G., Wu, X., Zhang, Y., and Zhu, X. 2018a. How to
match when all vertices arrive online. In Proceedings of the 50th Annual ACM Symposium on
Theory of Computing (STOC). 17 -- 29.
[16] Huang, Z., Peng, B., Tang, Z. G., Tao, R., Wu, X., and Zhang, Y. 2019. Tight
competitive ratios of classic matching algorithms in the fully online model. In Proceedings of the
30th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). 2875 -- 2886.
[17] Huang, Z., Tang, Z. G., Wu, X., and Zhang, Y. 2018b. Online vertex-weighted bipar-
In Proceedings of the 45th International
tite matching: Beating 1-1/e with random arrivals.
Colloquium on Automata, Languages and Programming (ICALP). 1070 -- 1081.
[18] Karp, R. M., Vazirani, U. V., and Vazirani, V. V. 1990. An optimal algorithm for
on-line bipartite matching. In Proceedings of the 22nd Annual ACM Symposium on Theory of
Computing (STOC). 352 -- 358.
[19] Kuhn, H. W. 1955. The hungarian method for the assignment problem. Naval research
logistics quarterly 2, 1-2, 83 -- 97.
[20] Lee, E. and Singla, S. 2017. Maximum matching in the online batch-arrival model.
In
Proceedings of the 19th Conference on Integer Programming and Combinatorial Optimization
(IPCO). 355 -- 367.
[21] Lov´asz, L. and Plummer, M. D. 2009. Matching theory. Vol. 367. American Mathematical
Society.
[22] Mehta, A. 2013. Online matching and ad allocation. Foundations and Trends R(cid:13) in Theoretical
Computer Science 8, 4, 265 -- 368.
[23] Schrijver, A. 2003. Combinatorial optimization: polyhedra and efficiency. Vol. 24. Springer
Science & Business Media.
[24] Tirodkar, S. and Vishwanathan, S. 2017. Maximum matching on trees in the online
preemptive and the incremental dynamic graph models. In Proceedings of the 23rd International
Computing and Combinatorics Conference (COCOON). 504 -- 515.
[25] Wang, Y. and Wong, S. C.-w. 2015. Two-sided online bipartite matching and vertex cover:
Beating the greedy algorithm. In Proceedings of the 42nd International Colloquium on Automata,
Languages and Programming (ICALP). 1070 -- 1081.
38
|
1911.10262 | 1 | 1911 | 2019-11-21T18:18:43 | An Algorithm for Strong Stability in the Student-Project Allocation Problem with Ties | [
"cs.DS"
] | We study a variant of the Student-Project Allocation problem with lecturer preferences over Students where ties are allowed in the preference lists of students and lecturers (SPA-ST). We investigate the concept of strong stability in this context. Informally, a matching is strongly stable if there is no student and lecturer $l$ such that if they decide to form a private arrangement outside of the matching via one of $l$'s proposed projects, then neither party would be worse off and at least one of them would strictly improve. We describe the first polynomial-time algorithm to find a strongly stable matching or to report that no such matching exists, given an instance of SPA-ST. Our algorithm runs in $O(m^2)$ time, where $m$ is the total length of the students' preference lists. | cs.DS | cs |
An Algorithm for Strong Stability in the Student-Project
Allocation Problem with Ties
Sofiat Olaosebikan⋆ and David Manlove⋆⋆
School of Computing Science, University of Glasgow,
e-mail: [email protected], [email protected]
Abstract. We study a variant of the Student-Project Allocation problem with lecturer
preferences over Students where ties are allowed in the preference lists of students and
lecturers (spa-st). We investigate the concept of strong stability in this context. Informally,
a matching is strongly stable if there is no student and lecturer l such that if they decide
to form a private arrangement outside of the matching via one of l's proposed projects,
then neither party would be worse off and at least one of them would strictly improve.
We describe the first polynomial-time algorithm to find a strongly stable matching or to
report that no such matching exists, given an instance of spa-st. Our algorithm runs in
O(m2) time, where m is the total length of the students' preference lists.
Keywords: student-project allocation problem, indifference, strongly stable matching,
polynomial-time algorithm
1 Introduction
Matching problems, which generally involve the assignment of a set of agents to another set
of agents based on preferences, have wide applications in many real-world settings, including,
for example, allocating junior doctors to hospitals [25] and assigning students to projects [16].
In the context of assigning students to projects, each project is proposed by one lecturer and
each student is required to provide a preference list over the available projects that she finds
acceptable. Also, lecturers may provide a preference list over the students that find their projects
acceptable, and/or over the projects that they propose. Typically, each project and lecturer have
a specific capacity denoting the maximum number of students that they can accommodate. The
goal is to find a matching, i.e., an assignment of students to projects that respects the stated
preferences, such that each student is assigned at most one project, and the capacity constraints
on projects and lecturers are not violated -- the so-called Student-Project Allocation problem
(spa) [1,6,20].
Two major models of spa exist in the literature: one permits preferences only from the stu-
dents [16], while the other permits preferences from the students and lecturers [1,15]. In the latter
case, three different variants have been studied based on the nature of the lecturers' preference
lists. These include SPA with lecturer preferences over (i) students [15], (ii) projects [13,22,23],
and (iii) (student, project) pairs [2]. Outwith assigning students to projects, applications of each
of these three variants can be seen in multi-cell networks where the goal is to find a stable
assignment of users to channels at base-stations [3,4,5].
⋆ Supported by a College of Science and Engineering Scholarship, University of Glasgow. Orcid ID:
0000-0002-8003-7887.
⋆⋆ Supported by grant EP/P028306/1 from the Engineering and Physical Sciences Research Council.
Orcid ID: 0000-0001-6754-7308.
2
In this work, we will concern ourselves with variant (i), i.e., the Student-Project Allocation
problem with lecturer preferences over Students (spa-s). In this context, it has been argued in
[25] that a natural property for a matching to satisfy is that of stability. Informally, a stable
matching ensures that no student and lecturer would have an incentive to deviate from their
current assignment. Abraham et al. [1] described two linear-time algorithms to find a stable
matching in an instance of spa-s where the preference lists are strictly ordered. In their paper,
they also proposed an extension of spa-s where the preference lists may include ties, which we
refer to as the Student-Project Allocation problem with lecturer preferences over Students with
Ties (spa-st).
If we allow ties in the preference lists of students and lecturers, three different stability
definitions are possible [9,10,11]. We give an informal definition in what follows. Suppose that
M is a matching in an instance of spa-st. Then M is (i) weakly stable, (ii) strongly stable, or
(iii) super-stable, if there is no student and lecturer l such that if they decide to become assigned
outside of M via one of l's proposed projects, respectively,
(i) both of them would strictly improve,
(ii) one of them would be better off, and the other would not be worse off
(iii) neither of them would be worse off.
These concepts were first defined and studied by Irving [9] in the context of the Stable Marriage
problem with Ties (smt) (the restriction of spa-st in which each lecturer offers only one project,
the capacity of each project and lecturer is 1, the numbers of students and lecturers are equal,
each student finds all projects acceptable, and each lecturer finds all students acceptable). It
was subsequently extended to the Hospitals/Residents problem with Ties (HRT) [10,11] (where
HRT is the special case of spa-st in which each lecturer offers only one project, and the capacity
of each project is the same as the capacity of the lecturer offering the project).
Existing results in spa-st. Every instance of spa-st admits a weakly stable matching, which
could be of different sizes [21]. Moreover, the problem of finding a maximum size weakly stable
matching (MAX-SPA-ST) is NP-hard [12,21], even for the so-called Stable Marriage problem with
Ties and Incomplete lists (smti). Cooper and Manlove [7] described a 3
2 -approximation algorithm
for MAX-SPA-ST. On the other hand, Irving et al. argued in [10] that super-stability is a natural
and most robust solution concept to seek in cases where agents have incomplete information.
Recently, Olaosebikan and Manlove [24] showed that if an instance of spa-st admits a super-
stable matching M , then all weakly stable matchings in the instance are of the same size (equal
to the size of M ), and match exactly the same set of students. The main result of their paper was
a polynomial-time algorithm to find a super-stable matching or report that no such matching
exists, given an instance of spa-st. Their algorithm runs in O(L) time, where L is the total
length of all the preference lists.
Motivation. It was motivated in [11] that weakly stable matching may be undermined by
bribery or persuasion, in practical applications of hrt. In what follows, we give a corresponding
argument for an instance I of spa-st. Suppose that M is a weakly stable matching in I, and
suppose that a student si prefers a project pj (where pj is offered by lecturer lk) to her assigned
is offered by a lecturer different from lk). Suppose further
project in M , say pj ′ (where pj ′
that pj is full and lk is indifferent between si and one of the worst student/s assigned to pj in
M , say si′ . Clearly, the pair (si, pj) does not constitute a blocking pair for the weakly stable
matching M , as lk would not improve by taking on si in the place of si′ . However, si might be
overly invested in pj that she is even ready to persuade or even bribe lk to reject si′ and accept
her instead; lk being indifferent between si and si′ may decide to accept si's proposal. We can
reach a similar argument if the roles are reversed. However, if M is strongly stable, it cannot be
potentially undermined by this type of (student, project) pair.
3
Henceforth, if a spa-st instance admits a strongly stable matching, we say that such an
instance is solvable. Unfortunately not every instance of spa-st is solvable. To see this, consider
the case where there are two students, two projects and two lecturers, the capacity of each
project and lecturer is 1, the students have exactly the same strictly ordered preference list of
length 2, and each of the lecturers preference list is a single tie of length 2 (any matching will
be undermined by a student and lecturer that are not matched together). However, it should be
clear from the discussions above that in cases where a strongly stable matching exists, it should
be preferred over a matching that is merely weakly stable.
Related work. The following are previous results for strong stability in the literature. Irving [9]
gave an O(n4) algorithm for computing strongly stable matchings in an instance of smt, where
n is the number of men (equal to the number of women). This algorithm was subsequently
extended by Manlove [19] to instances of smti, which is a generalisation of smt for which the
preference lists need not be complete. The extended algorithm also has running time O(n4).
Irving et al. [11] described an algorithm to find a strongly stable matching or report that no
such matching exists, given an instance of HRT. The algorithm has running time O(m2), where
m is the total number of acceptable (resident, hospital) pairs. Subsequently, Kavitha et al. [14]
presented two strong stability algorithms with improved running time; one for smti with running
time O(nm), where n is the number of vertices and m is the number of edges, and the other for
hrt with running time O(mPh∈H ph), where H is the set of all hospitals and ph is the capacity
of a hospital h. These two algorithms build on the ones described in [9,11,19]. A recent result
in strong stability is the work of Kunysz [17], where he described an O(nm log(W n)) algorithm
for computing a maximum weight strongly stable matching given an instance of smti, where W
is the maximum weight of an edge.
Our contribution. We present the first polynomial-time algorithm to find a strongly stable
matching or report that no such matching exists, given an instance of spa-st -- thus solving
an open problem given in [1,24]. Our algorithm is student-oriented, which implies that if the
given instance is solvable then our algorithm will output a solution in which each student has
at least as good a project as she could obtain in any strongly stable matching. We note that our
algorithm is a non-trivial extension of the strong stability algorithms for smt [9], smti [19], and
hrt [11] (we discuss this further in Sect. 4.3).
The remainder of this paper is structured as follows. We give a formal definition of the spa-s
problem, the spa-st variant, and the three stability concepts in Sect. 2. We give some intuition
for the strong stability definition in Sect. 3. We describe our algorithm for spa-st under strong
stability in Sect. 4. Further, in Sect. 4, we also illustrate an execution of our algorithm with
respect to an instance of spa-st before moving on to present the algorithm's correctness and
complexity results, along with proof of correctness. Finally, we present some potential direction
for future work in Sect. 5.
2 Preliminary definitions
In this section, we give a formal definition of spa-s as described in the literature [1]. We also give
a formal definition of spa-st -- a generalisation of spa-s in which preference lists can include
ties.
2.1 Formal definition of spa-s
An instance I of spa-s involves a set S = {s1, s2, . . . , sn1 } of students, a set P = {p1, p2, . . . , pn2 }
of projects and a set L = {l1, l2, . . . , ln3} of lecturers. Each student si ranks a subset of P in
4
strict order, which forms her preference list. We say that si finds pj acceptable if pj appears on
si's preference list. We denote by Ai the set of projects that si finds acceptable.
Each lecturer lk ∈ L offers a non-empty set of projects Pk, where P1, P2, . . . , Pn3 partitions
P, and lk provides a preference list, denoted by Lk, ranking in strict order of preference those
students who find at least one project in Pk acceptable. Also lk has a capacity dk ∈ Z+, indicating
the maximum number of students she is willing to supervise. Similarly each project pj ∈ P has
a capacity cj ∈ Z+ indicating the maximum number of students that it can accommodate.
We assume that for any lecturer lk, max{cj : pj ∈ Pk} ≤ dk ≤ P{cj : pj ∈ Pk} (i.e., the
capacity of lk is (i) at least the highest capacity of the projects offered by lk, and (ii) at most the
sum of the capacities of all the projects lk is offering). We denote by Lj
k, the projected preference
list of lecturer lk for pj, which can be obtained from Lk by removing those students that do not
find pj acceptable (thereby retaining the order of the remaining students from Lk).
An assignment M is a subset of S × P such that (si, pj) ∈ M implies that si finds pj
acceptable. If (si, pj) ∈ M , we say that si is assigned to pj, and pj is assigned si. For convenience,
if si is assigned in M to pj, where pj is offered by lk, we may also say that si is assigned to lk,
and lk is assigned si. For any project pj ∈ P, we denote by M (pj) the set of students assigned to
pj in M . Project pj is undersubscribed, full or oversubscribed according as M (pj) is less than,
equal to, or greater than cj, respectively. Similarly, for any lecturer lk ∈ L, we denote by M (lk)
the set of students assigned to lk in M . Lecturer lk is undersubscribed, full or oversubscribed
according as M (lk) is less than, equal to, or greater than dk, respectively. A matching M is
an assignment such that M (si) ≤ 1, M (pj) ≤ cj and M (lk) ≤ dk. If si is assigned to some
project in M , for convenience we let M (si) denote that project.
2.2 Ties in the preference lists
We now give a formal definition, similar to the one given in [24], for the generalisation of spa-s
in which the preference lists can include ties. In the preference list of lecturer lk ∈ L, a set T
of r students forms a tie of length r if lk does not prefer si to si′ for any si, si′ ∈ T (i.e., lk
is indifferent between si and si′ ). A tie in a student's preference list is defined similarly. For
convenience, in what follows we consider a non-tied entry in a preference list as a tie of length
one. We denote by spa-st the generalisation of spa-s in which the preference list of each student
(respectively lecturer) comprises a strict ranking of ties, each comprising one or more projects
(respectively students). An example spa-st instance I1 is given in Fig. 1, which involves the
set of students S = {s1, s2, s3}, the set of projects P = {p1, p2, p3} and the set of lecturers
L = {l1, l2}. Ties in the preference lists are indicated by round brackets.
Student preferences
s1: (p1 p2)
s2: p2 p3
s3: p3 p1
Lecturer preferences
l1:
l2:
s3 (s1
(s3
s2)
s2)
l1 offers p1, p2
l2 offers p3
Project capacities: c1 = c2 = c3 = 1
Lecturer capacities: d1 = 2, d2 = 1
Fig. 1: An example spa-st instance I1.
In the context of spa-st, we assume that all notation and terminology carries over from spa-
s with the exception of stability, which we now define. When ties appear in the preference lists,
three types of stability arise, namely weak stability, strong stability and super-stability [10,11].
5
In what follows, we give a formal definition of these three stability concepts in the context of
spa-st. Henceforth, I is an instance of spa-st, (si, pj) is an acceptable pair in I and lk is the
lecturer who offers pj.
Definition 1 (weak stability [7]). Let M be a matching in I. We say that M is weakly stable if
it admits no blocking pair, where a blocking pair of M is an acceptable pair (si, pj) ∈ (S ×P)\M
such that (a) and (b) holds as follows:
(a) either si is unassigned in M or si prefers pj to M (si);
(b) either (i), (ii), or (iii) holds as follows:
(i) pj is undersubscribed and lk is undersubscribed;
(ii) pj is undersubscribed, lk is full and either si ∈ M (lk), or lk prefers si to the worst
student/s in M (lk);
(iii) pj is full and lk prefers si to the worst student/s in M (pj).
Definition 2 (super-stability [24]). We say that M is super-stable if it admits no blocking
pair, where a blocking pair of M is an acceptable pair (si, pj) ∈ (S × P) \ M such that (a) and
(b) holds as follows:
(a) either si is unassigned in M , or si prefers pj to M (si) or is indifferent between them;
(b) either (i), (ii), or (iii) holds as follows:
(i) pj is undersubscribed and lk is undersubscribed;
(ii) pj is undersubscribed, lk is full, and either si ∈ M (lk) or lk prefers si to the worst
student/s in M (lk) or is indifferent between them;
(iii) pj is full and lk prefers si to the worst student/s in M (pj) or is indifferent between
them.
Definition 3 (strong stability). We say that M is strongly stable in I if it admits no blocking
pair, where a blocking pair of M is an acceptable pair (si, pj) ∈ (S × P) \ M such that either
(1a and 1b) or (2a and 2b) holds as follows:
(1a) either si is unassigned in M , or si prefers pj to M (si);
(1b) either (i), (ii), or (iii) holds as follows:
(i) pj is undersubscribed and lk is undersubscribed;
(ii) pj is undersubscribed, lk is full, and either si ∈ M (lk) or lk prefers si to the worst
student/s in M (lk) or is indifferent between them;
(iii) pj is full and lk prefers si to the worst student/s in M (pj) or is indifferent between
them.
(2a) si is indifferent between pj and M (si);
(2b) either (i), (ii), or (iii) holds as follows:
(i) pj is undersubscribed, lk is undersubscribed and si /∈ M (lk);
(ii) pj is undersubscribed, lk is full, si /∈ M (lk), and lk prefers si to the worst student/s in
M (lk);
(iii) pj is full and lk prefers si to the worst student/s in M (pj).
In the remainder of this paper, any usage of the term blocking pair refers to the version of
this term for strong stability as defined in Definition 3. We give an intuition behind the strong
stability definition is what follows.
6
3 Justification of the strong stability definition
It should be clear from our definition of a blocking pair (si, pj) that if si seeks to become assigned
to pj outside of M , then at most one of si and lk can be indifferent to the switch, whilst at least
one of them must strictly improve. In what follows, we justify our definition in more detail (we
remark that some of the argument is similar to that given for the blocking pair definition in the
spa-s case [1, Sect 2.2]).
We consider the first part of the definition. In Definition 3(1a); clearly if the assignment
between si and pj is permitted, si will improve relative to M . Now, let us consider lk's perspec-
tive. In Definition 3(1b)(i), lk will be willing to take on si for pj, since there is a free space. In
Definition 3(1b)(ii), if si was already assigned in M to a project offered by lk then lk will agree
to the switch, since the total number of students assigned to lk remains the same. However,
if si was not already assigned in M to a project offered by lk, since lk is full, lk will need to
reject some student assigned to her in order to take on si. Obviously, lk will not reject a student
that she prefers to si; thus lk will either improve or be no worse off after the switch. Finally,
in Definition 3(1b)(iii), since pj is full, lk will need to reject some student assigned to pj in
order to take on si. Again, lk will either improve or be no worse off after the switch. Under this
definition, as observed in [1, Sect 2.2], if si was already assigned in M to a project offered by lk,
then the number of students assigned to lk will decrease by 1 (the reason for this was further
justified in [1, Sect 6.1]).
Next, we consider the second part of the definition. In Definition 3(2a), if the assignment
between si and pj is permitted, clearly si will be no worse off after the switch. Again, we consider
lk's perspective (we note that in this case, lk must improve after the switch.). In Definition
3(2b)(i), if pj and lk is undersubscribed, then the only way that lk would improve is if si is not
already assigned in M to a project offered by lk. If this is the case, then lk will agree to the
switch since there is a free space and she will get one more student to supervise, namely si. In
Definition 3(2b)(ii), if pj is undersubscribed and lk is full, the only way lk could improve is first
for si to not be assigned in M to a project offered by lk. If this is the case then lk will need to
reject some student assigned to her in M in order to take on si. Obviously, lk will be willing to
reject a student that is worse than si on her list. Similarly, in Definition 3(2b)(iii), if pj is full
in M , lk will need to reject some student assigned to pj in M in order to take on si. Clearly, lk
must prefer si to such student so that lk can have a better set of students assigned to pj after
the switch. We remark that if si is already assigned in M to a project offered by lk, the number
of students assigned to lk in M will decrease by 1 after the switch.
We illustrate this using the example spa-st instance I2 in Fig. 2. Clearly, M1 = {(s1, p2),
(s2, p1)} is a matching in I2. However, by satisfying the blocking pair (s1, p1), we obtain the
matching M2 = {(s1, p1)}, which is strongly stable. In going from M1 to M2, lecturer l1 got a
better student to take on p1, with l1 having to lose one student, namely s2. Following a similar
argument as in [1, Sect 6.1] for the spa-s case, in practice, l1 might not agree to take on s1 for
p1, since in doing so l1 loses a student. So, one could relax Definition 3(2b)(iii) to prevent such
change from happening. Further, we argue that this could lead to two new problems.
1. We introduce an element of strategy into the problem by allowing a matching such as M1
to be strongly stable. That is, a student could provide a shorter preference list in order to
obtain a more suitable project. To illustrate this, suppose that s1 has only listed p1 in the
example instance I2. Irrespective of how we relax Definition 3(2b)(iii), si would be assigned
to p1. However, this strategy might not be beneficial for a student in all cases, since by not
listing all of her acceptable projects, a student is at risk of not being assigned in the final
strongly stable matching.
Student preferences
s1: (p1 p2)
s2: p1
Lecturer preferences
l1:
s1
s2
7
l1 offers p1, p2
Project capacities: c1 = c2 = 1
Lecturer capacities: d1 = 2
Fig. 2: An example spa-st instance I2.
2. If we allow both M1 and M2 to be strongly stable, this would imply that the instance
admits strongly stable matchings of different sizes. Hence, we would seek a maximum size
strongly stable matching in order to match as many students to projects as possible. However,
we conjecture that this problem is NP-hard, following from related problems of finding
maximum size stable matching in the literature.
4 An algorithm for spa-st under strong stability
In this section we present our algorithm for spa-st under strong stability, which we will refer to
as Algorithm SPA-ST-strong. In Sect. 4.1, we give some definitions relating to the algorithm.
In Sect. 4.2, we give a description of our algorithm and present it in pseudocode form. In
Sect. 4.3, we briefly describe the non-trivial modifications that are involved in extending the
existing strong stability algorithms for smt [9], smti [19] and hrt [11] to our algorithm for the
spa-st case. We illustrate an execution of our algorithm with respect to a spa-st instance in
Sect. 4.4. Finally, in Sect. 4.5, we present our algorithm's correctness results along with proof
of correctness.
4.1 Definitions relating to the algorithm
Given a pair (si, pj) ∈ M , for some strongly stable matching M in I, we call (si, pj) a strongly
stable pair. During the execution of the algorithm, students become provisionally assigned to
projects (and implicitly to lecturers), and it is possible for a project (and lecturer) to be provi-
sionally assigned a number of students that exceeds its capacity. We describe a project (respec-
tively lecturer) as replete if at any time during the execution of the algorithm it has been full or
oversubscribed. We say that a project (respectively lecturer) is non-replete if it is not replete.
As stated earlier, for a project pj, it is possible that dG(pj) > cj at some point during
the algorithm's execution. Thus, we denote by qj = min{cj, dG(pj)} the quota of pj in G,
which is the minimum between pj's capacity and the number of students provisionally assigned
to pj in G. Similarly, for a lecturer lk, it is possible that dG(lk) > dk at some point during
the algorithm's execution. At this point, we denote by αk = P{qj : pj ∈ Pk ∩ P } the total
quota of projects offered by lk that is provisionally assigned to students in G and we denote by
qk = min{dk, dG(lk), αk} the quota of lk in G.
The algorithm proceeds by deleting from the preference lists certain (si, pj) pairs that are
not strongly stable. By the term delete (si, pj), we mean the removal of pj from si's preference
list and the removal of si from Lj
k (the projected preference list of lecturer lk for pj); in addition,
if (si, pj) ∈ E we delete the edge from G. By the head and tail of a preference list at a given
point we mean the first and last tie respectively on that list after any deletions might have
occurred (recalling that a tie can be of length 1). Given a project pj, we say that a student
si is dominated in Lj
k if si is worse than at least cj students who are provisionally assigned to
pj. The concept of a student becoming dominated in a lecturer's preference list is defined in a
slightly different manner.
8
Definition 4 (Dominated in Lk). At a given point during the algorithm's execution, let αk
and dG(lk) be as defined above. We say that a student si is dominated in Lk if min{dG(lk), αk} ≥
1, and si is worse than at least dk students who are provisionally assigned in G to a project
dk
offered by lk.
Definition 5 (Lower rank edge). We define an edge (si, pj) ∈ E as a lower rank edge if si is
in the tail of Lk and min{dG(lk), αk} > dk.
Definition 6 (Bound). Given an edge (si, pj) ∈ E, we say that si is bound to pj if (i) and (ii)
holds as follows:
(i) pj is not oversubscribed or si is not in the tail of Lj
(ii) (si, pj) is not a lower rank edge or si is not in the tail of Lk (or both).
k (or both);
If si is bound to pj, we may also say that (si, pj) is a bound edge. Otherwise, we refer to it as
an unbound edge.2
3 and we reduce the quota of pj in Gr (and intuitively lk
We form a reduced assignment graph Gr = (Sr, Pr, Er) from a provisional assignment graph G
as follows. For each edge (si, pj) ∈ E such that si is bound to pj, we remove the edge (si, pj)
4) by one. Further, we remove
from Gr
all other unbound edges incident to si in Gr. Each isolated student vertex is then removed from
Gr. Finally, if the quota of any project is reduced to 0, or pj becomes an isolated vertex, then pj
is removed from Gr. For each surviving pj in Gr, we denote by q∗
j the revised quota of pj, where
q∗
j is the difference between pj's quota in G (i.e., qj ) and the number of students that are bound
to pj. Similarly, we denote by q∗
k is the difference between
lk's quota in G (i.e., qk) and the number of students that are bound to a project offered by lk.
j : pj ∈ Pk ∩ Pr} − q∗
Further, for each lk who offers at least one project in Gr, we let n = P{q∗
k,
where n is the difference between the total revised quota of projects in Gr that are offered by
lk and the revised quota of lk in Gr. Now, if n ≤ 0, we do nothing; otherwise, we extend Gr as
follows. We add n dummy student vertices to Sr. For each of these dummy vertex, say sdi, and
for each project pj ∈ Pk ∩ Pr that is adjacent to a student vertex in Sr via a lower rank edge,
we add the edge (sdi, pj) to Er.5
k the revised quota of lk in Gr, where q∗
Given a set X ⊆ Sr of students, define N (X), the neighbourhood of X, to be the set of
project vertices adjacent in Gr to a student in X. If for all subsets X of Sr, each student in X
can be assigned to one project in N (X), without exceeding the revised quota of each project
in N (X) (i.e., X ≤ P{q∗
j : pj ∈ N (X)} for all X ⊆ Sr); then we say Gr admits a perfect
matching that saturates Sr.
Definition 7 (Critical set). It is well known in the literature [18] that if Gr does not admit
a perfect matching that saturates Sr, then there must exist a deficient subset Z ⊆ Sr such that
Z > P{q∗
j : pj ∈ N (Z)}. To be precise, the deficiency of Z is defined by δ(Z) = Z − P{q∗
j :
1 If a student si is provisionally assigned in G to two different projects offered by lk then, potentially,
dG(lk) < αk. Thus, it is important that we take the minimum of these two parameters, to avoid
deleting strongly stable pairs.
2 An edge (si, pj) ∈ E can change state from bound to unbound, but not vice versa.
3 We note that we only remove this edge to form Gr, we do not delete the edge from G.
4 If si is bound to more than one projects offered by lk, for all the bound edges involving si and these
projects that we remove from Gr, we only reduce lk's quota in Gr by one.
5 An intuition as to why we add dummy students to Gr is as follows. Given a lecturer lk whose project
j : pj ∈ Pk ∩ Pr}, then we need n dummy
k, so that we don't oversubscribe
is provisionally assigned to a student in Gr. If q∗
j : pj ∈ Pk ∩ Pr} and q∗
students to offset the difference between P{q∗
lk in any maximum matching obtained from Gr.
k < P{q∗
9
pj ∈ N (Z)}. The deficiency of Gr, denoted δ(Gr), is the maximum deficiency taken over all
subsets of Sr. Thus, if δ(Z) = δ(Gr), we say that Z is a maximally deficient subset of Sr, and
we refer to Z as a critical set.
We denote by PR the set of replete projects in G and we denote by P ∗
R a subset of projects in
PR which is obtained as follows. For each project pj ∈ PR, let lk be the lecturer who offers pj.
For each student si such that (si, pj) has been deleted, we add pj to P ∗
R if (i) and (ii) holds as
follows:
(i) either si is unassigned in G, or (si, pj ′ ) ∈ G where si prefers pj to pj ′ , or (si, pj ′ ) ∈ G and
si is indifferent between pj and pj ′ where pj ′ /∈ Pk;
(ii) either lk is undersubscribed, or lk is full and either si ∈ G(lk) or lk prefers si to some
student assigned in G(lk).
Definition 8 (Feasible matching). A feasible matching in the final provisional assignment
graph G is a matching M which is obtained as follows:
1. Let G∗ be the subgraph of G induced by the students who are adjacent to a project in P ∗
R.
First, find a maximum matching M ∗ in G∗;6
2. Using M ∗ as an initial solution, find a maximum matching M in G.
4.2 Description of the algorithm
Algorithm SPA-ST-strong, described in Algorithm 1, begins by initialising an empty bipartite
graph G which will contain the provisional assignments of students to projects (and intuitively
to lecturers). We remark that such assignments (i.e., edges in G) can subsequently be broken
during the algorithm's execution.
k, such that st is dominated in Lj
The while loop of the algorithm involves each student si who is not adjacent to any project
in G and who has a non-empty list applying in turn to each project pj at the head of her
list. Immediately, si becomes provisionally assigned to pj in G (and to lk). If, by gaining a new
provisional assignee, project pj becomes full or oversubscribed then we set pj as replete. Further,
for each student st in Lj
k, we delete the pair (st, pj). As we will
prove later, such pairs cannot belong to any strongly stable matching. Similarly, if by gaining a
new provisional assignee, lk becomes full or oversubscribed then we set lk as replete. For each
student st in Lk, such that st is dominated in Lk and for each project pu ∈ Pk that st finds
acceptable, we delete the pair (st, pu). This continues until every student is provisionally assigned
to one or more projects or has an empty list. At the point where the while loop terminates, we
form the reduced assignment graph Gr and we find the critical set Z of students in Gr (Lemma
3 describes how to find Z). As we will see later, no project pj ∈ N (Z) can be assigned to any
student in the tail of Lj
k in any strongly stable matching, so all such pairs are deleted.
At the termination of the inner repeat-until loop in line 21, i.e., when Z is empty, if some
project pj that is replete ends up undersubscribed, we let sr be any one of the most preferred
students (according to Lj
k) who was provisionally assigned to pj during some iteration of the
algorithm but is not assigned to pj at this point (for convenience, we henceforth refer to such
sr as the most preferred student rejected from pj according to Lj
k). If the students at the tail of
Lk (recalling that the tail of Lk is the least-preferred tie in Lk after any deletions might have
occurred) are no better than sr, it turns out that none of these students st can be assigned to any
6 At the point in the algorithm when we need to construct the feasible matching M from G, if P ∗
R is
R to avoid a potential blocking
non-empty, this phase ensures that we fill up all of the projects in P ∗
pair involving some student that has been rejected by some project in P ∗
R.
10
project offered by lk in any strongly stable matching -- such pairs (st, pu), for each project pu ∈ Pk
that st finds acceptable, are deleted. The repeat-until loop is then potentially reactivated,
and the entire process continues until every student is provisionally assigned to a project or has
an empty list.
At the termination of the outer repeat-until loop in line 30, if a student is adjacent in G to
a project pj via a bound edge, then we may potentially carry out extra deletions as follows. First,
we let lk be the lecturer that offers pj and we let U be the set of projects that are adjacent to si
in G via an unbound edge. For each project pu ∈ U \ Pk, it turns out that the pair (si, pu) cannot
belong to any strongly stable matching, thus we delete all such pairs. Finally, we let M be any
feasible matching in the provisional assignment graph G. If M is strongly stable relative to the
given instance I then M is output as a strongly stable matching in I. Otherwise, the algorithm
reports that no strongly stable matching exists in I. We present Algorithm SPA-ST-strong in
pseudocode form in Algorithm 1.
Finding the critical set. Consider the reduced assignment graph Gr = (Sr, Pr, Er) formed
from G at a given point during the algorithm's execution (at line 15). To find the critical set of
students in Gr, first we need to construct a maximum matching Mr in Gr, with respect to the
revised quota q∗
j , for each pj ∈ Pr. In this context, a matching Mr ⊆ Er is such that Mr(si) ≤ 1
for all si ∈ Sr, and Mr(pj) ≤ q∗
j for all pj ∈ Pr. We describe how to construct Mr as follows:
1. Let G′
r be the subgraph of Gr induced by the dummy students adjacent to a project in Gr.
First, find a maximum matching M ′
r in G′
r.
2. Using M ′
r as an initial solution, find a maximum matching Mr in Gr.7
If a student si is not assigned to any project in Mr, we say that vertex si is free in Gr. Similarly,
if a project pj is such that pj has fewer than q∗
j assignees in Mr, we say that vertex pj is free in
Gr. An alternating path in Gr relative to Mr is any simple path in which edges are alternately
in, and not in, Mr. An augmenting path in Gr is an alternating path from a free student to a
free project. The following lemmas are classical results with respect to matchings in bipartite
graphs.
Lemma 1. A matching Mr in a reduced assignment graph Gr has maximum cardinality if and
only if there is no augmenting path relative to Mr in Gr.
Lemma 2. Let Mr be a maximum matching in the reduced assignment graph Gr. Then Mr =
Sr − δ(Gr).
The classical augmenting path algorithm can be used to obtain Mr, and as explained in [11],
this can be implemented to run in O(min{n,P cj}m) time, where n is the number of students
and m is the total length of the students' preference lists. Now that we have described how to
construct a maximum matching in the reduced assignment graph, the following lemma tells us
how to find the critical set of students.
Lemma 3. Given a maximum matching Mr in the reduced assignment graph Gr, the critical
set Z consists of the set U of unassigned students together with the set U ′ of students reachable
from a student in U via an alternating path.
7 By making sure that all the dummy students are matched in step 1, we are guaranteed that no
lecturer is oversubscribed with non-dummy students in Gr.
Algorithm 1 Algorithm SPA-ST-strong
11
delete (st, pu)
k do
repeat
for each project pj at the head of si's list do
for each student st dominated in Lj
k do
delete (st, pj)
if lk is full or oversubscribed then
for each student st dominated in Lk do
for each project pu ∈ Pk ∩ At do
while some student si is unassigned and has a non-empty list do
lk ← lecturer who offers pj
add the edge (si, pj) to G
if pj is full or oversubscribed then
form the reduced assignment graph Gr
find the critical set Z of students
for each project pu ∈ N (Z) do
lk ← lecturer who offers pu
for each student st at the tail of Lu
Input: spa-st instance I
Output: a strongly stable matching in I or "no strongly stable matching exists in I"
1: G ← ∅
2: repeat
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30: until every unassigned student has an empty list
31: for each student si in G do
32:
33:
34:
35:
36:
37: M ← a feasible matching in G
38: if M is a strongly stable matching in I then
39:
40: else
41:
lk ← lecturer who offers pj
U ← unbound projects adjacent to si in G
for each pu ∈ U \ Pk do
lk ← lecturer who offers pj
sr ← most preferred student rejected from pj in Lj
if the students at the tail of Lk are no better than sr then
k {any if > 1}
for each student st at the tail of Lk do
for each project pu ∈ Pk ∩ At do
delete (st, pu)
until Z is empty
for each pj ∈ P do
delete (si, pu)
return M
delete (st, pu)
if si is adjacent in G to a project pj via a bound edge then
if pj is replete and pj is undersubscribed then
return "no strongly stable matching exists in I"
Proof. First, we note that δ(Gr) = U . Let C = U ∪ U ′, we claim that δ(C) = U . By the
definition of δ(Gr), clearly δ(C) ≤ δ(Gr) = U . Now suppose that δ(C) < U , then
U > δ(C)
= U ∪ U ′ − X
pj ∈N (C)
q∗
j
= U + U ′ − X
pj ∈N (C)
q∗
j
(since U ∩ U ′ = ∅)
U ′ < X
pj ∈N (C)
q∗
j
.
(1)
12
Inequality 1 implies that there is a project pj ∈ N (C) such that pj has fewer than q∗
j assignees
in Mr; thus pj is free in Mr. We claim that every student that is assigned in Gr to a project
pj ′ ∈ N (C) must be in C. For suppose there is a student si /∈ C such that si is assigned to pj ′
in Mr. Since pj ′ ∈ N (C), then pj ′ must be adjacent to some student si′ ∈ C. Now, if si′ ∈ U ,
then there is an alternating path from si′ to si via pj ′ , a contradiction. Otherwise, if si′ ∈ U ′,
since si′ is reachable from a student in U via an alternating path, si is also reachable from the
same student in U via an alternating path. Hence our claim is established. Now, given that pj
has fewer than q∗
j assignees in Mr, since each student in U ′ is reachable from a student in U via
an alternating path, and the students in U ∪ U ′ are collectively adjacent to projects in N (C), we
can find an alternating path from a student in U to pj. Thus Mr admits an augmenting path,
contradicting the maximality of Mr.
Further, the critical set Z must contain every student who is unassigned in some maximum
matching in Gr. For, suppose not. Let M ∗
r =
Sr − δ(Gr)), and suppose there is some student si ∈ Sr \ Z such that si is unassigned in M ∗
r .
There must be δ(Gr) unassigned students, with at most δ(Gr) − 1 of these students contained
in Z (since si /∈ Z). Hence Z contains at least Z − δ(Gr) + 1 assigned students. It follows that
r be an arbitrary such matching (where M ∗
or
X
pj ∈N (Z)
q∗
j ≥ Z − δ(Gr) + 1
Z − X
pj ∈N (Z)
q∗
j ≤ δ(Gr) − 1
contradicting the required deficiency of Z.
But, for every si ∈ U ′, there is a maximum matching in which si is unassigned, obtainable
from Mr via an alternating path from a student in U to si. Hence C ⊆ Z; and since δ(C) = δ(Z),
⊓⊔
this completes the proof.
4.3 The non-triviality of extending Algorithm HRT-strong to spa-st
Algorithm SPA-ST-strong is a non-trivial extension of Algorithm HRT-strong for hrt [11].
Here we outline the major distinctions between our algorithm and Algorithm HRT-strong,
which indicate the challenges involved in extending the earlier approach to the spa-st setting.
1. Given a lecturer lk, it is possible that during some iteration of our algorithm, some pj ∈ Pk
is oversubscribed which causes lk to become full or oversubscribed (see Fig. 4(a) in Sect. 4.4,
at the point where s3 applies to p1). Finding the dominated students in Lk becomes more
complex in spa-st -- to achieve this we introduced the notion of quota (i.e., qk) for lk.
2. To form Gr in the spa-st case, we extended the approach described in the hrt case [11]
by introducing the concept of lower rank edges for each lecturer who offers a project in Gr,
and we also introduced dummy students.
3. Lines 22 - 29 of Algorithm SPA-ST-strong refer to additional deletions that must be car-
ried out in a certain situation; this type of deletion was also carried out in Algorithm
SPA-ST-super for super-stability [24] (see the description corresponding to Fig. 5 in Sect. 4.4
for an example showing why we may need to carry out this type of deletion in the strong
stability context).
4. Constructing a feasible matching M in G in the spa-st setting is much more challenging: we
first identify some replete projects that must be full in M , denoted by P ∗
R (see the description
corresponding to Fig. 6 in Sect. 4.4). Also, in the hrt case, when constructing M from G,
preference is given to a bound edge over an unbound edge; in general, this is not always true
in the spa-st case.
13
4.4 Example algorithm execution
In this section, we illustrate an execution of Algorithm SPA-ST-strong with respect to the
spa-st instance I3 shown in Fig. 3, which involves the set of students S = {si : 1 ≤ i ≤ 8},
the set of projects P = {pj : 1 ≤ j ≤ 6} and the set of lecturers L = {lk : 1 ≤ k ≤ 3}. The
algorithm starts by initialising the bipartite graph G = {}, which will contain the provisional
assignment of students to projects. We assume that the students become provisionally assigned
to each project at the head of their list in subscript order. Figs. 4, 5 and 6 illustrate how this
execution of Algorithm SPA-ST-strong proceeds with respect to I3.
Student preferences
s1: p1 p6
s2: p1 p2
s3: (p1 p4)
s4: p2 (p5 p6)
s5: (p2 p3)
s6: (p2 p4)
s7: p3 p1
s8: p5 p1
s1
s2
s3
s4
s5
s6
s7
s8
p1 : 2
p2 : 2
p3 : 1
p4 : 1
p5 : 1
(a) The provisional as-
signment graph G(1) at
the end of the while
loop, with the quota of
each project labelled be-
side it.
Lecturer preferences
{3} l1:
s8
{2} l2:
s6
{3} l3: (s1
s7 (s1
s5 (s7
s4)
s8
s2
s3)
s3) (s4
s5)
s6
offers
p1, p2
p3, p4
p5, p6
Project capacities: c1 = c2 = c6 = 2, c3 = c4 = c5 = 1
Lecturer capacities: d1 = d3 = 3, d2 = 2
Fig. 3: An instance I3 of spa-st.
s1
s2
s3
s4
sd1
p1 : 1
p2 : 2
s4
s5
s6
s7
s8
p1 : 1
p2 : 2
p3 : 1
p4 : 1
p5 : 1
(b) The reduced assign-
ment graph G(1)
r , with
the revised quota of each
project
labelled beside
it. The collection of the
dashed edges is the maxi-
mum matching M (1)
.
r
(c) The provisional as-
signment graph G(1) at
the termination of itera-
tion (1).
Fig. 4: Iteration (1).
Iteration 1: At the termination of the while loop during the first iteration of the inner repeat-until
loop, every student, except s3, s6 and s7, is provisionally assigned to every project in the first
tie on their preference list. Edge (s3, p4) /∈ G(1) because (s3, p4) was deleted as a result of s6 be-
coming provisionally assigned to p4, causing s3 to be dominated in L4
2. Also, edge (s6, p2) /∈ G(1)
because (s6, p2) was deleted as a result of s4 becoming provisionally assigned to p2, causing s6
to be dominated in L1 (at that point in the algorithm, min{dG(l1), α1} = min{4, 3} = 3 = d1
and s6 is worse than at least d1 students who are provisionally assigned to l1). Finally, edge
(s7, p3) /∈ G(1) because (s7, p3) was deleted as a result of s5 becoming provisionally assigned to
p5, causing s7 to be dominated in L3
2.
14
To form G(1)
r , the bound edges (s5, p3), (s6, p4), (s7, p1) and (s8, p5) are removed from the
graph. We can verify that edges (s4, p2) and (s5, p2) are unbound, since they are lower rank
edges for l1. Also, since p1 is oversubscribed, and each of s1, s2 and s3 is at the tail of L1
1, edges
(s1, p1), (s2, p1) and (s3, p1) are unbound. Further, the revised quota of l1 in G(1)
is 2, and the
total revised quota of projects offered by l1 (i.e., p1 and p2) is 3. Thus, we add one dummy
student vertex sd1 to G1
r, and we add an edge between sd1 and p2 (since p2 is the only project
in G(1)
adjacent to a student in the tail of L1 via a lower rank edge). With respect to the
maximum matching M (1)
, it is clear that the critical set Z (1) = {s1, s2, s3}, thus we delete the
edges (s1, p1), (s2, p1) and (s3, p1) from G(1); and the inner repeat-until loop is reactivated.
r
r
r
s1
s2
s4
s5
s6
s7
s8
p1 : 1
p2 : 2
p3 : 1
p4 : 1
p5 : 1
p6 : 1
(a) The provisional as-
signment graph G(2) at
the end of the while
loop.
s4
p2 : 1
(b) The reduced assign-
ment graph G(2)
r .
Fig. 5: Iteration (2).
s1
s2
s5
s6
s7
s8
p1 : 1
p2 : 1
p3 : 1
p4 : 1
p5 : 1
p6 : 1
(c) The provisional as-
signment graph G(2) at
the termination of itera-
tion (2).
Iteartion 2: At the beginning of this iteration, each of s1 and s2 is unassigned and has a non-
empty list; thus we add edges (s1, p6) and (s2, p2) to the provisional assignment graph obtained
r . It can be verified that every edge in G(2)
at the termination of iteration (1) to form G(2)
r ,
except (s4, p2) and (s5, p2), is a bound edge. Clearly, the critical set Z (2) = ∅, thus the inner
repeat-until loop terminates. At this point, project p1, which was replete during iteration (1),
is undersubscribed in iteration (2). Moreover, the students at the tail of L1 (i.e., s4 and s5) are
no better than s3, where s3 is one of the most preferred students rejected from p1 according to
L1
1; thus we delete edges (s4, p2) and (s5, p2). The outer repeat-until loop is then reactivated
(since s4 is unassigned and has a non-empty list).
15
s1
s2
s4
s5
s6
s7
s8
p1 : 2
p2 : 1
p3 : 1
p4 : 1
p5 : 1
p6 : 2
(a) The provisional as-
signment graph G(3) at
the end of the while
loop.
Fig. 6: Iteration (3).
Iteration 3: At the beginning of this iteration, the only student that is unassigned and has a
non-empty list is s4; thus we add edges (s4, p5) and (s4, p6) to the provisional assignment graph
obtained at the termination of iteration (2) to form G(3)
r . The provisional assignment of s4 to
p5 led to p5 becoming oversubscribed; thus (s8, p5) is deleted (since s8 is dominated on L5
3).
Further, s8 becomes provisionally assigned to p1. It can be verified that all the edges in G(3)
are
bound edges. Moreover, the reduced assignment graph G(3)
r = ∅.
r
Again, every unassigned students has an empty list. We also have that a project p2, which
was replete in iteration (2), is undersubscribed in iteration (3). However, no further deletion is
carried out in line 29 of the algorithm, since the student at the tail of L1 (i.e., s2) is better
than s4 and s5, where s4 and s5 are the most preferred students rejected from p2 according
to L2
R = {p5}, since (s8, p5)
has been deleted, s8 prefers p5 to her provisional assignment in G and l3 is undersubscribed.
Thus we need to ensure p5 fills up in the feasible matching M constructed from G, so as to
avoid (s8, p5) from blocking M . Finally, the algorithm outputs the feasible matching M =
{(s1, p6), (s2, p2), (s4, p5), (s5, p3), (s6, p4), (s7, p1), (s8, p1)} as a strongly stable matching.
1. Hence, the repeat-until loop terminates. We observe that P ∗
4.5 Correctness of the algorithm
We now present the following results regarding the correctness of Algorithm SPA-ST-strong.
The first of these results deals with the fact that no strongly stable pair is ever deleted during
the execution of the algorithm.
Lemma 4. If a pair (si, pj) is deleted during the execution of Algorithm SPA-ST-strong, then
(si, pj) does not belong to any strongly stable matching in I.
In order to prove Lemma 4, we present Lemmas 5, 6 and 7.
Lemma 5. If a pair (si, pj) is deleted within the inner repeat-until loop during the execution
of Algorithm SPA-ST-strong, then (si, pj) does not belong to any strongly stable matching in
I.
Proof. Suppose that (si, pj) is the first strongly stable pair to be deleted within the inner
repeat-until loop during an arbitrary execution E of Algorithm SPA-ST-strong. Let M ∗ be
some strongly stable matching in which si is assigned to pj. Let lk be the lecturer who offers pj.
Suppose that G is the provisional assignment graph immediately after the deletion of (si, pj).
There are three cases to consider.
16
1. Suppose that (si, pj) is deleted (in line 10) because some other student became provisionally
assigned to pj during E, causing pj to become full or oversubscribed, so that si is dominated
in Lj
k. Since (si, pj) ∈ M ∗ \ G, there is some student, say sr, such that lk prefers sr to si
and (sr, pj) ∈ G \ M ∗, for otherwise pj would be oversubscribed in M ∗. We note that sr
cannot be assigned to a project that she prefers to pj in any strongly stable matching, for
otherwise some strongly stable pair must have been deleted before (si, pj), as pj must be in
the head of sr's list when she applied. So sr is either unassigned in M ∗, or sr prefers pj to
M ∗(si) or is indifferent between them. Clearly for any combination of lk and pj being full or
undersubscribed in M ∗, it follows that (sr, pj) blocks M ∗, a contradiction.
2. Suppose that (si, pj) is deleted (in line 14) because some other student became provisionally
assigned to a project offered by lk during E, causing lk to become full or oversubscribed, so
that si is dominated in Lk. We denote by Ck the set of projects that are full or oversubscribed
in G, which are offered by lk. We denote by Dk the set of projects that are undersubscribed
in G, which are offered by lk. Clearly the projects offered by lk that are provisionally assigned
to a student in G at this point can be partitioned into Ck and Dk. We consider two subcases.
(i) Each student who is provisionally assigned in G to a project in Dk (if any) is also
assigned to that same project in M ∗. However, after the deletion of (si, pj), we know
that
X
pt∈Ck∪Dk
qt = X
pt∈Ck
ct + X
pt∈Dk
dG(pt) ≥ dk;
(2)
i.e., the total quota of projects in Ck ∪ Dk is at least the capacity of lk. Now, since pj
has one more assignee in M ∗ than it has provisional assignees in G, namely si, then
some other project pj ′ ∈ Ck must have fewer than cj ′ assignees in M ∗, for otherwise lk
would be oversubscribed in M ∗. This implies that there is some student, say sr, such
that lk prefers sr to si and (sr, pj ′ ) ∈ G \ M ∗. Moreover, sr cannot be assigned to a
project that she prefers to pj ′ in M ∗, as explained in (1) above. Hence, (sr, pj ′ ) blocks
M ∗, a contradiction.
(ii) Each project in Ck at this point ends up full in M ∗. This implies that there is some
project pj ′ ∈ Dk with fewer assignees in M ∗ than provisional assignees in G, for other-
wise lk would be oversubscribed in M ∗. Thus pj ′ is undersubscribed in M ∗ (since Dk is
the set of undersusbcribed projects offered by lk). Moreover, there is some student, say
sr, such that lk prefers sr to si and (sr, pj ′ ) ∈ G \ M ∗. Following a similar argument in
(i) above, (sr, pj ′ ) blocks M ∗, a contradiction.
3. Suppose that (si, pj) is deleted (in line 20) because pj is provisionally assigned to a student
in the critical set Z at some point, and at that point si is in the tail of Lj
k. We refer to
the set of preference lists at that point as the current lists. Let Z ′ be the set of students
in Z who are assigned in M ∗ to a project from the head of their current lists, and let P ′
be the set of projects in N (Z) assigned in M ∗ to at least one student from the tail of
its current list. We have that pj ∈ P ′, so P ′
6= ∅. Consider si′ ∈ Z. Now si′ cannot be
assigned in M ∗ to a project that she prefers to any project in the head of her current list,
for otherwise some strongly stable pair must have been deleted before (si, pj). Hence, any
student si′ in Z who is provisionally assigned to pj must be in Z ′, otherwise (si′ , pj) would
6= ∅, because Z − P{q∗
block M ∗. Thus Z ′
j : pj ∈ N (Z)} > 0 and
Z ′ −P{q∗
j : pj ∈ P ′} ≤ 0, since every student in Z ′ is assigned
in M ∗ to a project in P ′.
6= ∅. Also Z \ Z ′
j : pj ∈ N (Z)} ≤ Z ′ −P{q∗
We now claim that there must be an edge (sr, pj ′ ) in Gr such that sr ∈ Z \ Z ′ and pj ′ ∈ P ′.
For otherwise N (Z \ Z ′) ⊆ N (Z) \ P ′, and
17
Z \ Z ′ − X{q∗
j : pj ∈ N (Z \ Z ′)} ≥ Z \ Z ′ − X{q∗
j : pj ∈ N (Z) \ P ′}
= Z − X{q∗
≥ Z − X{q∗
j : pj ∈ N (Z)} − (cid:16)Z ′ − X{q∗
j : pj ∈ N (Z)},
j : pj ∈ P ′}(cid:17)
j : pj ∈ P ′} ≤ 0. Hence Z \ Z ′ has deficiency at least that of Z, contradicting
since Z ′ −P{q∗
the fact that Z is the critical set. Thus our claim is established, i.e., there is some student
sr ∈ Z \ Z ′ and some project pj ′ ∈ P ′ such that sr is adjacent to pj ′ in Gr. Since sr is
either unassigned in M ∗ or prefers pj ′ to M ∗(sr), and since the lecturer who offers pj ′ is
indifferent between sr and at least one student in M ∗(pj ′ ), we have that (sr, pj ′ ) blocks M ∗,
a contradiction.
⊓⊔
Lemma 6. If a pair (si, pj) is deleted in line 29 during the execution of Algorithm SPA-ST-
strong, then (si, pj) does not belong to any strongly stable matching in I.
Proof. Suppose that (si, pj) was deleted in line 29 of the algorithm. Let pj ′ be some other
project offered by lk which was replete during an iteration of the inner repeat-until loop and
subsequently ends up undersubscribed at the end of the loop, i.e., pj ′ plays the role of pj in
line 23. Suppose that si′ plays the role of sr in line 25, i.e., si′ was the most preferred student
rejected from pj ′ according to Lj ′
k (possibly si′ = si). Then lk prefers si′ to si or is indifferent
between them, since si plays the role of st at some for loop iteration in line 27. Moreover si′ was
provisionally assigned to pj ′ during an iteration of the inner repeat-until loop but (si′ , pj ′ ) /∈ G
at the end of the loop. Thus (si′ , pj ′ ) /∈ M ∗, since no strongly stable pair is deleted within the
inner repeat-until loop, as proved in Lemma 5.
Let lz0 = lk, pt0 = pj ′ and sq0 = si′ . Again, none of the students who are provisionally
assigned to some project in G can be assigned to any project better than their current assignment
in any strongly stable matching as this would mean a strongly stable pair must have been
deleted before (si, pj), as each student apply to projects in the head of her list. So, either (a)
sq0 is unassigned in M ∗ or sq0 prefers pt0 to M ∗(sq0 ), or (b) sq0 is indifferent between pt0 and
M ∗(sq0 ). If (a) holds, then by the strong stability of M ∗, pt0 is full in M ∗ and lz0 prefers the
worst student/s in M ∗(pt0 ) to sq0 ; for if pt0 is undersubscribed in M ∗ then (sq0 , pt0) blocks M ∗,
since si ∈ M ∗(lz0) and lz0 prefers sq0 to si or is indifferent between them, a contradiction. If (b)
holds we claim that lz0 prefers sq0 to si. Thus, irrespective of whether M ∗(sq0 ) is offered by lz0
or not, the argument follows from (a), i.e., pt0 is full in M ∗ and lz0 prefers the worst student/s
in M ∗(pt0 ) to sq0.
Just before the deletion of (si, pj) occurred, pt0 is undersubscribed in G. Since pt0 is full in
M ∗, it follows that there exists some student, say sq1 , such that (sq1 , pt0) ∈ M ∗ \ G. We note
that lz0 prefers sq1 to sq0. Let pt1 = pt0. Since (si, pj) is the first strongly stable pair to be
deleted, sq1 is provisionally assigned in G to a project pt2 such that sq1 prefers pt2 to pt1. For
otherwise, as students apply to projects in the head of their list, that would mean (sq1 , pt1) must
have been deleted during an iteration of the inner repeat-until loop, a contradiction. We note
that pt2 6= pt1, since (sq1 , pt2 ) ∈ G and (sq1 , pt1) /∈ G. Let lz1 be the lecturer who offers pt2. By
the strong stability of M ∗, it follows that either
(i) pt2 is full in M ∗ and lz1 prefers the worst student/s in M ∗(pt2 ) to sq1 , or
18
(ii) pt2 is undersubscribed in M ∗, lz1 is full in M ∗, sq1 /∈ M ∗(lz1 ) and lz1 prefer the worst
student/s in M ∗(lz1) to sq1.
Otherwise (sq1 , pt2 ) blocks M ∗. In case (i), there exists some student sq2 ∈ M ∗(pt2 ) \ G(pt2).
Let pt3 = pt2 . In case (ii), there exists some student sq2 ∈ M ∗(lz1) \ G(lz1 ). We note that lz1
prefers sq2 to sq1 . Now, suppose M ∗(sq2 ) = pt3 (possibly pt3 = pt2). It is clear that sq2 6= sq1 .
Applying similar reasoning as for sq1, sq2 is assigned in G to a project pt4 such that sq2 prefers pt4
to pt3. Let lz2 be the lecturer who offers pt4 . We are identifying a sequence hsqi ii≥0 of students,
a sequence hpti ii≥0 of projects, and a sequence hlziii≥0 of lecturers, such that, for each i ≥ 1
1. sqi prefers pt2i to pt2i−1 ,
2. (sqi , pt2i) ∈ G and (sqi , pt2i−1) ∈ M ∗,
3. lzi prefers sqi+1 to sqi ; also, lzi offers both pt2i and pt2i+1 (possibly pt2i = pt2i+1 ).
First we claim that for each new project that we identify, pt2i 6= pt2i−1 for i ≥ 1. Suppose
pt2i = pt2i−1 for some i ≥ 1. From above sqi was identified by lzi−1 such that (sqi , pt2i−1 ) ∈ M ∗\G.
Moreover (sqi , pt2i) ∈ G. Hence we reach a contradiction. Clearly, for each student sqi for i ≥ 1
we identify, sqi must be assigned to distinct projects in G and in M ∗.
Next we claim that for each new student sqi that we identify, sqi 6= sqt for 1 ≤ t < i. We
prove this by induction on i. For the base case, clearly sq2 6= sq1 . We assume that the claim holds
for some i ≥ 1, i.e., the sequence sq1 , sq2 , . . . , sqi consists of distinct students. We show that the
claim holds for i + 1, i.e, the sequence sq1 , sq2 , . . . , sqi , sqi+1 also consists of distinct students.
Clearly sqi+1 6= sqi since lzi prefers sqi+1 to sqi . Thus, it suffices to show that sqi+1 6= sqj for
1 ≤ j ≤ i − 1. Now, suppose sqi+1 = sqj for 1 ≤ j ≤ i − 1. This implies that sqj was identified
by lzi and clearly lzi prefers sqj to sqj−1 . Now since sqi+1 was also identified by lzi to avoid
) in M ∗, it follows that either (i) pt2i is full in M ∗, or (ii) pt2i is
the blocking pair (sqi , pt2i
undersubscribed in M ∗ and lzi is full in M ∗. We consider each cases further as follows.
(i) If pt2i is full in M ∗, we know that (sqi , pt2i) ∈ G \ M ∗. Moreover sqj was identified by lzi+1
because of case (i). Furthermore (sqj−1 , pt2i ) ∈ G \ M ∗. In this case, pt2i+1 = pt2i and we
have that
(sqi , pt2i+1 ) ∈ G \ M ∗ and (sqi+1 , pt2i+1 ) ∈ M ∗ \ G,
(sqj−1 , pt2i+1) ∈ G \ M ∗ and (sqj , pt2i+1) ∈ M ∗ \ G.
By the inductive hypothesis, the sequence sq1 , sq2 , . . . , sqj−1 , sqj , . . . , sqi consists of distinct
students. This implies that sqi 6= sqj−1 . Thus since pt2i+1 is full in M ∗, lzi should have been
able to identify distinct students sqj and sqi+1 to avoid the blocking pairs (sqj−1 , pt2i+1 ) and
(sqi , pt2i+1) respectively in M ∗, a contradiction.
(ii) pt2i is undersubscribed in M ∗ and lzi is full in M ∗. Similarly as in case (i) above, we have
that
sqi ∈ G(lzi) \ M ∗(lzi) and sqi+1 ∈ M ∗(lzi) \ G(lzi ),
sqj−1 ∈ G(lzi) \ M ∗(lzi) and sqj ∈ M ∗(lzi) \ G(lzi).
Since sqi 6= sqj−1 and lzi is full in M ∗, lzi should have been able to identify distinct students
sqj and sqi+1 corresponding to students sqj−1 and sqi respectively, a contradiction.
This completes the induction step. As the sequence of distinct students and projects is infinite,
⊓⊔
we reach an immediate contradiction.
19
Lemma 7. If a pair (si, pj) is deleted in line 36 during the execution of Algorithm SPA-ST-
strong, then (si, pj) does not belong to any strongly stable matching in I.
Proof. Suppose that (si, pj) is the first strongly stable pair to be deleted during an arbitrary
execution E of Algorithm SPA-ST-strong. Then by Lemmas 5 and 6, (si, pj) was deleted in
line 36. Let lk be the lecturer who offers pj and let M ∗ be some strongly stable matching in
which si is assigned to pj. Let G be the provisional assignment graph. At this point in the
algorithm where the deletion of (si, pj) occured, si is adjacent to some other project pj ′ ∈ G
via a bound edge, where pj ′ is offered by lk′ (note that lk 6= lk′ ). By the definition of a bound
edge, it follows that either pj ′ is not oversubscribed in G or lk′ prefers si to some student in
G(pj ′ ) (or both). Also, either (si, pj ′ ) is not a lower rank edge or lk′ prefers si to some student
in G(lk′ ) (or both).
Let lz0 = lk′ , pt0 = pj ′ and sq0 = si. We note that sq0 is indifferent between pj and pt0. Now,
since (sq0 , pt0 ) ∈ G \ M ∗, by the strong stabiity of M ∗, either (i) or (ii) holds as follows:
(i) pt0 is full in M ∗ and lz0 prefers the worst student/s in M ∗(pt0 ) to sq0 or is indifferent
between them;
(ii) pt0 is undersubscribed in M ∗, lz0 is full in M ∗ and lz0 prefers the worst students in M ∗(lz0)
to sq0 or is indifferent between them.
Otherwise (sq0 , pt0 ) blocks M ∗. In case (i), since sq0 is bound to pt0 in G and since (sq0 , pt0) ∈
G \ M ∗, there exists some student sq1 ∈ M ∗(pt0) \ G(pt0 ). Let pt1 = pt0. In case (ii), since sq0 is
assigned in G to a project offered by lz0 (i.e., pt0 ) via a bound edge, and since sq0 is not assigned
to lz0 in M ∗, there exists some student sq1 ∈ M ∗(lz0 ) \ G(lz0 ). We note that lz0 either prefers
sq1 to sq0 or is indifferent between them; clearly sq1 6= sq0 .
Now, suppose M ∗(sq1 ) = pt1 (possibly pt1 = pt0). Since (sq0 , pj) is the first strongly stable
pair to be deleted, sq1 is provisionally assigned in G to a project pt2 such that sq1 prefers pt2 to
pt1. For otherwise, as students apply to projects in the head of their list and since (sq1 , pt1) /∈ G,
that would mean (sq1 , pt1) must have been deleted during an iteration of the repeat-until
loop, a contradiction. We note that pt2 6= pt1, since (sq1 , pt2) ∈ G and (sq1 , pt1 ) /∈ G. Let lz1 be
the lecturer who offers pt2. Again by the strong stability of M ∗, either (i) or (ii) holds as follows:
(i) pt2 is full in M ∗ and lz1 prefers the worst student/s in M ∗(pt2 ) to sq1 ;
(ii) pt2 is undersubscribed in M ∗, lz1 is full in M ∗ and lz1 prefers the worst students in M ∗(lz1)
to sq1 .
Otherwise (sq1 , pt2 ) blocks M ∗. In case (i), there exists some student sq2 ∈ M ∗(pt2 ) \ G(pt2).
Let pt3 = pt2. In case (ii), there exists some student sq2 ∈ M ∗(lz1) \ G(lz1 ). We note that
lz1 prefers sq2 to sq1; again it is clear that sq2 6= sq1 . Now, suppose M ∗(sq2 ) = pt3 (possibly
pt3 = pt2). Applying similar reasoning as for sq1 , sq2 is provisionally assigned in G to a project
pt4 such that sq2 prefers pt4 to pt3 . Let lz2 be the lecturer who offers pt4 . We are identifying a
sequence hsqi ii≥0 of students, a sequence hptiii≥0 of projects, and a sequence hlziii≥0 of lecturers,
such that, for each i ≥ 1
1. sqi prefers pt2i to pt2i−1 ,
2. (sqi , pt2i) ∈ G and (sqi , pt2i−1) ∈ M ∗,
3. lzi prefers sqi+1 to sqi ; also, lzi offers both pt2i and pt2i+1 (possibly pt2i = pt2i+1 ).
Following a similar argument as in the proof of Lemma 6, we can identify an infinite sequence
⊓⊔
of distinct students and projects, a contradiction.
Lemmas 5, 6 and 7 immediately give rise to Lemma 4. The next two lemmas will be used as
a tool in the proof of the remaining lemmas.
20
Lemma 8. Every student who is assigned to a project (intuitively a lecturer) in the final provi-
sional assignment graph G must be assigned in any feasible matching M .
Proof. Suppose si is a student who is provisionally assigned to some project in G, but si is not
assigned in a feasible matching M , then si must be in the critical set Z, and hence Z 6= ∅, a
⊓⊔
contradiction.
Lemma 9. Let M be a feasible matching in the final provisional assignment graph G and let
M ∗ be any strongly stable matching. Let lk be an arbitrary lecturer; (i) if lk is undersubscribed
in M ∗ then every student who is assigned to lk in M is also assigned to lk in M ∗, and (ii) if lk
is undersubscribed in M then lk has the same number of assignees in M ∗ as in M .
Proof. Let lk be an arbitrary lecturer. First, we show that (i) holds. Suppose otherwise, then
there exists a student, say si, such that si ∈ M (lk) \ M ∗(lk). Moreover, there exists some project
pj ∈ Pk such that si ∈ M (pj) \ M ∗(pj). By Lemma 4, si cannot be assigned to a project that
she prefers to pj in M ∗. Also, by the strong stability of M ∗, pj is full in M ∗ and lk prefers the
worst student/s in M ∗(pj) to si.
Let lz0 = lk, pt0 = pj, and sq0 = si. As pt0 is full in M ∗ and no project is oversubscribed in M ,
there exists some student sq1 ∈ M ∗(pt0) \ M (pt0 ) such that lz0 prefers sq1 to sq0 . Let pt1 = pt0.
By Lemma 4, sq1 is assigned in M to a project pt2 such that sq1 prefers pt2 to pt1. We note that
sq1 cannot be indifferent between pt2 and pt1; for otherwise, as students apply to projects in the
head of their list, since (sq1 , pt1) /∈ M , that would mean (sq1 , pt1) must have been deleted during
the algorithm's execution, contradicting Lemma 4. It follows that sq1 ∈ M (pt2 ) \ M ∗(pt2). Let
lz1 be the lecturer who offers pt2 . By the strong stability of M ∗, either
(i) pt2 is full in M ∗ and lz1 prefers the worst student/s in M ∗(pt2 ) to sq1 , or
(ii) pt2 is undersubscribed in M ∗, lz1 is full in M ∗, sq1 /∈ M ∗(lz1) and lz1 prefers the worst
student/s in M ∗(lz1) to sq1.
Otherwise (sq1 , pt2) blocks M ∗. In case (i), there exists some student sq2 ∈ M ∗(pt2 ) \ M (pt2 ).
Let pt3 = pt2 . In case (ii), there exists some student sq2 ∈ M ∗(lz1) \ M (lz1). We note that lz1
prefers sq2 to sq1 . Now, suppose M ∗(sq2 ) = pt3 (possibly pt3 = pt2). It is clear that sq2 6= sq1 .
Applying similar reasoning as for sq1 , student sq2 is assigned in M to a project pt4 such that sq2
prefers pt4 to pt3 . Let lz2 be the lecturer who offers pt4. We are identifying a sequence hsqi ii≥0
of students, a sequence hptiii≥0 of projects, and a sequence hlzi ii≥0 of lecturers, such that, for
each i ≥ 1
1. sqi prefers pt2i to pt2i−1 ,
2. (sqi , pt2i) ∈ G and (sqi , pt2i−1) ∈ M ∗,
3. lzi prefers sqi+1 to sqi ; also, lzi offers both pt2i and pt2i+1 (possibly pt2i = pt2i+1 ).
Following a similar argument as in the proof of Lemma 6, we can identify an infinite sequence
of distinct students and projects, a contradiction. Hence, if lk is undersubscribed in M ∗ then
every student who is assigned to lk in M is also assigned to lk in M ∗.
Next, we show that (ii) holds. By the first claim, any lecturer who is full in M is also full in
M ∗, and any lecturer who is undersubscribed in M has as many assignees in M ∗ as she has in
M . Hence
X
lk∈L
M (lk) ≤ X
lk∈L
M ∗(lk) .
(3)
21
Let S1 and S2 denote the set of students who are assigned to a project in M and M ∗ respectively.
By Lemma 8, each student who is provisionally assigned to a project in the final provisional
assignment graph G must be assigned to a project in M . Moreover, any student who is not
provisionally assigned to a project in G must have an empty list. It follows that these students
are unassigned in M ∗, since Lemma 4 guarantees that no strongly stable pairs are deleted. Thus
S2 ≤ S1. Further, we have that
X
lk∈L
M ∗(lk) = S2 ≤ S1 = X
lk∈L
M (lk),
From Inequality 3 and 4, it follows that M (lk) = M ∗(lk) for each lk ∈ L.
(4)
⊓⊔
The next three lemmas deal with the case that Algorithm SPA-ST-strong reports the non-
existence of a strongly stable matching in I.
Lemma 10. Let M be a feasible matching in the final provisional assignment graph G. Suppose
that (a) some non-replete lecturer lk has fewer assignees in M than provisional assignees in G,
or (b) some replete lecturer is not full in M . Then I admits no strongly stable matching.
Proof. Suppose that M ∗ is a strongly stable matching for the instance. By Lemma 8, each
student who is provisionally assigned to a project in G must be assigned to a project in M .
Moreover, any student who is not provisionally assigned to a project in G must have an empty
list. It follows that these students are unassigned in M ∗, since Lemma 4 guarantees that no
strongly stable pairs are deleted. Thus M ∗ ≤ M .
Suppose that condition (a) is satisfied. Then some non-replete lecturer lk′ satisfies M (lk′ ) <
dG(lk′ ), where dG(lk′ ) is the number of students provisionally assigned in G to a project offered
by lk′ . As lk′ is non-replete, it follows that dG(lk′ ) < dk′ . Now M (lk) ≤ min{dk, dG(lk)} for all
lk ∈ L. Hence
M = X
lk∈L
M (lk) < X
lk∈L
min{dk, dG(lk)} .
(5)
Now, suppose that M ∗(lk) ≥ min{dk, dG(lk)} for all lk ∈ L. Then M ∗ > M by 5,
a contradiction. Hence M ∗(lk) < min{dk, dG(lk)} for some lk ∈ L. This implies that lk is
undersubcribed in M ∗. Moreover, lk has fewer assignees in M ∗ than provisional assignees in G.
Thus there exists some student si who is provisionally assigned to lk in G but not in M ∗. It
follows that there exists some project pj ∈ Pk such that (si, pj) ∈ G \ M ∗. By Lemma 4, si is
not assigned to a project that she prefers to pj in M ∗. Also, by the strong stability of M ∗, pj
is full in M ∗ and lk prefers the worst student/s in M ∗(pj) to si; for if pj is undersubscribed in
M ∗ then (si, pj) blocks M ∗, a contradiction. Since pj is full in M ∗ with students that are better
than si in Lj
k and (si, pj) ∈ G \ M ∗, then there is some student, say si′ , such that lk prefers si′
to si and (si′ , pj) ∈ M ∗ \ G. For if all the students assigned to pj in M ∗ are also assigned to pj
in G, then si would be dominated in Lj
k and thus (si, pj) would have been deleted in G.
Let lz0 = lk, pt0 = pt1 = pj, sq0 = si, sq1 = si′ . Again, by Lemma 4, sq1 is provisionally
assigned in G to a project pt2 such that sq1 prefers pt2 to pt1. For otherwise, as students apply
to projects in the head of their list, since (sq1 , pt1) /∈ G, that would mean (sq1 , pt1) must have
been deleted during the algorithm's execution, a contradiction. We note that pt2 6= pt1, since
(sq1 , pt2) ∈ G and (sq1 , pt1) /∈ G. Let lz1 be the lecturer who offers pt2 . By the strong stability
of M ∗, either
(i) pt2 is full in M ∗ and lz1 prefers the worst student/s in M ∗(pt2 ) to sq1 , or
22
(ii) pt2 is undersubscribed in M ∗, lz1 is full in M ∗, sq1 /∈ M ∗(lz1) and lz1 prefers the worst
student/s in M ∗(lz1) to sq1.
Otherwise (sq1 , pt2 ) blocks M ∗. In case (i), there exists some student sq2 ∈ M ∗(pt2 ) \ G(pt2).
Let pt3 = pt2 . In case (ii), there exists some student sq2 ∈ M ∗(lz1) \ G(lz1 ). We note that lz1
prefers sq2 to sq1 . Now, suppose M ∗(sq2 ) = pt3 (possibly pt3 = pt2). It is clear that sq2 6= sq1 .
Applying similar reasoning as for sq1 , student sq2 is assigned in G to a project pt4 such that sq2
prefers pt4 to pt3 . Let lz2 be the lecturer who offers pt4. We are identifying a sequence hsqi ii≥0
of students, a sequence hptiii≥0 of projects, and a sequence hlzi ii≥0 of lecturers, such that, for
each i ≥ 1
1. sqi prefers pt2i to pt2i−1 ,
2. (sqi , pt2i) ∈ G and (sqi , pt2i−1) ∈ M ∗,
3. lzi prefers sqi+1 to sqi ; also, lzi offers both pt2i and pt2i+1 (possibly pt2i = pt2i+1 ).
Following a similar argument as in the proof of Lemma 6, we can identify an infinite sequence
of distinct students and projects, a contradiction.
Now suppose condition (b) is satisfied, i.e., some replete lecturer is not full in M . Let L1
and L2 be the set of replete and non-replete lecturers respectively. Then some lk′′ ∈ L1 satisfies
M (lk′′ ) < dk′′ . Condition (a) cannot be satisfied, for otherwise the first part of the proof shows
that M ∗ does not exist. Hence M (lk) = dG(lk) < dk for all lk ∈ L2. Now M (lk) ≤ dk for all
lk ∈ L1. Hence
M = X
lk∈L1
M (lk) + X
lk∈L2
M (lk) < X
lk∈L1
dk + X
lk∈L2
dG(lk) .
(6)
Now suppose that M ∗(lk) = dk for all lk ∈ L1, and M ∗(lk) ≥ dG(lk) for all lk ∈ L2. Then
M ∗ > M by 6, a contradiction. Hence either (i) M ∗(lk′ ) < dk′ for some lk′ ∈ L1, or (ii)
M ∗(lk′ ) < dG(lk′ ) for some lk′ ∈ L2. In case (ii), we reach a similar contradiction to that
arrived at for condition (a). In case (i), we have that lk′ is undersubscribed in M ∗. As lk′ is
replete, there exists some student si who was provisionally assigned to lk′ during the algorithm's
execution, but si is not assigned to lk′ in M ∗. Moreover, there exists some project pj ∈ Pk′ such
that (si, pj) /∈ M ∗ By Lemma 4, si is not assigned to a project in M ∗ that she prefers to pj.
Following a similar argument as above, we can identify a sequence of distinct students and
projects, and as this sequence is infinite, we reach a contradiction. Thus I admits no strongly
⊓⊔
stable matching.
Lemma 11. Suppose that in the final provisional assignment graph G, a student is bound to
two or more projects that are offered by different lecturers. Then I admits no strongly stable
matching.
Proof. Suppose that a strongly stable matching exists for the instance. Let M be a feasible
matching in the final provisional assignment graph G. Denote by L1 and L2 the set of replete
and non-replete lecturers respectively.
Denote by S1 the set of students who are bound to one or more projects in G, and by S2 the
other students who are provisionally assigned to one or more projects in G. By Lemma 8,
Also,
M = S1 + S2 .
M = X
lk∈L1
M (lk) + X
lk∈L2
M (lk) .
(7)
(8)
Moreover, we have that
X
lk∈L1
M (lk) + X
lk∈L2
M (lk) = X
lk∈L1
dk + X
lk∈L2
dG(lk);
23
(9)
for otherwise, no strongly stable matching exists by Lemma 10. Now, if some student is bound
to more than one project offered by different lecturers, by considering how the lecturers' quotas
are reduced when the students in S1 are removed in forming Gr from G, it follows that
X
lk∈L1
(dk − q∗
k) + X
lk∈L2
(dG(lk) − q∗
k) ≥ X
lk∈L1
(dk − q∗
k) + X
lk∈L2
(min{dG(lk), αk} − q∗
k)
= X
lk∈L1
(qk − q∗
k) + X
lk∈L2
(qk − q∗
k)
(qk − q∗
k)
= X
lk∈L1∪L2
> S1,
(10)
(dk − q∗
where Plk∈L1
k) is the number of students that are
bound to at least one project offered by a lecturer in L1 and L2 respectively. By substituting
Equality 9 into Inequality 10, we obtain the following
k) and Plk∈L2
(min{dG(lk), αk} − q∗
X
lk∈L1
M (lk) − X
lk∈L1
k + X
q∗
lk∈L2
M (lk) − X
lk∈L2
q∗
k > S1 .
Combining 7 and 8 into 11, we have
S1 + S2 − X
lk∈L1
k − X
q∗
lk∈L2
X
lk∈L1∪L2
q∗
k > S1,
q∗
k < S2 .
(11)
(12)
Since the total revised quota of lecturers in L1 ∪ L2 whose projects are provisionally assigned
to a student in Gr is strictly less than the number of students in S2, the preceding inequality
⊓⊔
suffices to establish that the critical set is non-empty, a contradiction.
Lemma 12. Let M be a feasible matching in the final provisional assignment graph G. Suppose
that the pair (si, pj), where pj is offered by lk, was deleted during the algorithm's execution.
Suppose further that for any pj ′ ∈ Pk such that si is indifferent between pj and pj ′ , (si, pj ′ ) /∈ M .
Finally, suppose each of pj and lk is undersubscribed in M . Then I admits no strongly stable
matching.
Proof. Suppose for a contradiction that there exists a strongly stable matching M ∗ in I. Let
(si, pj) be a pair that was deleted during an arbitrary execution E of the algorithm. This implies
that (si, pj) /∈ M ∗ by Lemma 4. Let M be the feasible matching at the termination of E. By the
hypothesis of the lemma, lk is undersubscribed in M . This implies that lk is undersubscribed in
M ∗, as proved in Lemma 9. Since pj is offered by lk, and pj is undersubscribed in M , it follows
from the proof of [24, Lemma 5] that pj is undersubscribed in M ∗. Further, by the hypothesis
of the lemma, either si is unassigned in M , or si prefers pj to M (si) or is indifferent between
them, where M (si) is not offered by lk. By Lemma 4, this is true for si in M ∗. Hence (si, pj)
blocks M ∗, a contradiction.
⊓⊔
24
The next lemma shows that the feasible matching M may be used to determine the existence,
or otherwise, of a strongly stable matching in I.
Lemma 13. Let M be a feasible matching in the final provisional assignment graph G. If M is
not strongly stable then there is no strongly stable matching for the instance.
Proof. Suppose that M is not strongly stable. Let (si, pj) be a blocking pair for M , and let lk
be the lecturer who offers pj. We consider two cases.
1. Suppose si is unassigned in M or prefers pj to M (si). Then (si, pj) has been deleted.
Suppose (si, pj) was deleted as a result of pj being full or oversubscribed in line 8. Then pj is
replete. Suppose firstly that pj ends up full in M , then (si, pj) cannot block M irrespective
of whether lk is undersubscribed or full in M , since lk prefers the worst assigned student/s
in M (pj) to si. Thus pj is not full in M . As pj was previously full, each pair (st, pu), for
each st that is no better than si at the tail of Lk and each pu ∈ Pk ∩ At, would have been
deleted in line 29. Thus if lk is full in M , then (si, pj) does not block M . Now suppose lk is
undersubscribed in M . If lk was full at some point during the algorithm's execution, then no
strongly stable matching exists, by Lemma 10. Thus lk was never full during the algorithm's
execution. This implies that lk is non-replete. To recap, we have that pj is a replete project
that is not full in M , which is offered by a non-replete lecturer. Moreover, since (si, pj) has
been deleted, no strongly stable matching exists, by Lemma 12. Thus (si, pj) was not deleted
because pj became replete in line 8.
Now, suppose (si, pj) was deleted as a result of lk being full or oversubscribed in line 11.
Then (si, pj) could only block M if lk ends up undersubscribed in M . If this is the case,
then no strongly stable matching exists for the instance, by Lemma 10.
Next, suppose (si, pj) was deleted because pj was a neighbour of some student si′ ∈ Z
at a point when si was in the tail of Lj
k. Denote by Z ′ the set of students in Z who are
provisionally assigned to pj in Gr. Suppose pj is non-replete, it follows that 0 < Z ′ ≤ q∗
j .
Let Z ′′ = Z \ Z ′. Then N (Z ′′) ⊆ N (Z) \ {pj}, so that
X
pj′ ∈N (Z ′′)
Hence
j ′ ≤
q∗
X
pj′ ∈N (Z)
q∗
j ′
− q∗
j
.
δ(Z ′′) = Z ′′ − X
pj′ ∈N (Z ′′)
q∗
j ′
= Z − Z ′ − X
pj′ ∈N (Z ′′)
q∗
j ′
≥ Z − Z ′ −
X
pj′ ∈N (Z)
q∗
j ′
+ q∗
j
= Z + q∗
j − Z ′ − X
pj′ ∈N (Z)
q∗
j ′
≥ Z − X
pj′ ∈N (Z)
q∗
j ′
= δ(Z)
25
If δ(Z ′′) > δ(Z), then we reach a contradiction to the fact Z is maximally deficient; hence
δ(Z ′′) = δ(Z). Moreover, Z ′′ ⊂ Z, contradicting the minimality of Z. Now, suppose pj
is replete. The argument that no strongly stable matchings exists follows from the first
paragraph.
Next, suppose (si, pj) was deleted (at line 29) because some other project pj ′ offered by lk is
replete and subsequently becomes undersubscribed at line 23. Then lk must have identified
her most preferred student, say sr, rejected from pj ′ such that si is at the tail of Lk and
si is no better than sr in Lk. Moreover, every project offered by lk that si finds acceptable
would have been deleted from si's preference list at the for loop iteration in line 29. Thus if
pj ends up full in M , then (si, pj) does not block M . Suppose pj ends up undersubscribed
in M . If lk is full in M , then (si, pj) does not block M , since si /∈ M (lk) and lk prefers the
worst student/s in M (lk) to si. Suppose lk is undersubscribed in M . Again, if lk is replete,
then no strongly stable matching exists, by Lemma 10. Thus lk is non-replete. Since (si, pj)
has been deleted, no strongly stable matching exists, by Lemma 12.
2. Suppose si is indifferent between pj and M (si). First, suppose pj is undersubscribed in M .
By the strong stability definition, M (si) is not offered by lk. Suppose lk is undersubscribed
in M . If lk is replete, then by Lemma 10, no strongly stable matching exists. Hence lk is
non-replete. First suppose (si, pj) ∈ G, since si /∈ M (lk), this implies that lk has fewer
assignees in M than provisional assignees in G. Thus no strongly stable matching exists by
Lemma 10. Now suppose (si, pj) /∈ G. Then (si, pj) has been deleted. If pj is replete, then
no strongly stable matching exists by Lemma 12. Hence pj is non-replete. To recap, (si, pj)
has been deleted, and each of pj and lk is non-replete. There are two points in this algorithm
where this deletion could have happened; (i) if pj was a neighbour of some student si′ ∈ Z
at a point when si was in the tail of Lj
k, or (ii) at line 29 -- because some other project pj ′
offered by lk is replete and subsequently becomes undersubscribed at line 23. In any of the
cases, the argument that no strongly stable matching exists follows from (1) above. Finally,
suppose pj is full in M . Then (si, pj) cannot block M irrespective of whether lk is full or
undersubscribed in M , since the worst student/s in M (pj) are either better or no worse than
si, according to Lj
k.
Since (si, pj) is an arbitrary pair, this implies that M admits no blocking pair. Hence I
⊓⊔
admits no strongly stable matching.
Lemma 14. Algorithm SPA-ST-strong may be implemented to run in O(m2) time, where m
is the total length of the students' preference lists.
Proof. It is clear that the work done in the inner repeat-until loop other than in finding the
maximum cardinality matchings and critical sets is bounded by a constant times the number
of deleted pairs, and so is O(m) (where m is the total length of the students' preference lists).
We remark that the total amount of work done outside the inner repeat-until loop (i.e., in
deleting pairs when a replete project ends up undersubscribed in line 23) is bounded by the
total length of the students' preference lists, and so is O(m). During each iteration of the inner
repeat-until loop of Algorithm SPA-ST-strong, we need to form the reduced assignment
graph Gr, which takes O(m) time. Further, we need to search for a maximum matching in Gr,
which allows us to use Lemma 3 to find the critical set. Next we show how we can bound the
total amount of work done in finding the maximum matchings.
Suppose that Algorithm SPA-ST-strong finds a maximum matching M (i)
in the reduced
assignment graph G(i)
r at the ith iteration of the inner repeat-until loop. Suppose also that,
during the ith iteration in this loop, xi pairs are deleted because they involve students in the
r
26
r
r
r
critical set Z (i), or students tied with them in the tail of a project in N (Z (i)). Suppose further
that in the (i + 1)th iteration, yi pairs are deleted before the reduced assignment graph G(i+1)
is formed. Note that any edge in G(i)
r which is not one of these xi + yi deleted pairs must be in
G(i+1)
, since an edge in the provisional assignment graph cannot change state from unbound to
bound. In particular, at least M (i)
at the (i + 1)th
iteration. Hence we can start from these edges in that iteration and find a maximum matching
M (i+1)
which
were not in G(i)
r .
in O(min{nm, (xi + yi + zi)m}) time, where zi is the number of edges in G(i+1)
remain in G(i+1)
r − xi − yi edges of M (i)
r
r
r
Suppose that a total of s iterations of the inner repeat-until loop are carried out, and
that in t of these, min{nm, (xi + yi + zi)m}) = nm. Then the time complexity of maximum
matching operations, taken over the entire execution of the algorithm, is O(min{n,P cj}m +
tnm + mP(xi + yi + zi)) where the first term is for the maximum matching computation in
the first iteration, and the sum in the third term is over the appropriate s − t − 1 values of
i. We note that Ps
i=1(xi + yi + zi) ≤ 2m (since each of the total number of deletions and
provisional assignments is bounded by the total length of the students' preference lists), and
since xi + yi + zi ≥ n for the appropriate t values of i, it follows that
tnm + mX(xi + yi + zi) ≤ m
s
X
i=1
(xi + yi + zi) ≤ 2m2.
Thus
mX(xi + yi + zi) ≤ 2m2 − tnm.
At the termination of the outer repeat-until loop, a feasible matching is constructed by
taking the final maximum matching and combining it with the bound (student, project) pairs
in the final provisional assignment graph. This operation is clearly bounded by the number of
bound pairs, hence is O(m). It follows that the overall complexity of Algorithm SPA-ST-strong
is O(m + min{n,P cj}m +2m2) = O(m2).
We remark that the problem of finding a maximum matching in Gr is also called the Upper
Degree Constrained Subgraph problem (UDCS) [8]. An instance of Gr can be viewed as an instance
of UDCS, except that students have no explicit preferences in the UDCS case. Using Gabow's UDCS
j ). However, it is
not clear how this algorithm can be used to improve the overall running time of Algorithm
⊓⊔
SPA-ST-strong.
algorithm [8], we can find a maximum matching in Gr in time O(ErqP q∗
The following theorem collects together Lemmas 4-14 and establishes the correctness of Algorithm
SPA-ST-strong.
Theorem 1. For a given instance I of spa-st, Algorithm SPA-ST-strong determines in O(m2)
time whether or not a strongly stable matching exists in I. If such a matching does exist, all pos-
sible executions of the algorithm find one in which each assigned student is assigned at least as
good a project as she could obtain in any strongly stable matching, and each unassigned student
is unassigned in every strongly stable matchings.
Given the optimality property established by Theorem 1, we define the strongly stable
matching found by Algorithm SPA-ST-strong to be student-optimal. For example, in the spa-
st instance illustrated in Fig. 3 (Page 13), the student-optimal strongly stable matching is
M = {(s1, p6), (s2, p2), (s4, p5), (s5, p3), (s6, p4), (s7, p1), (s8, p1)}.
27
5 Conclusion
We leave open the formulation of a lecturer-oriented counterpart to Algorithm SPA-ST-strong.
From an experimental perspective, an interesting direction would be to carry out an empirical
analysis of Algorithm SPA-ST-strong, to investigate how various parameters (e.g., the density
and position of ties in the preference lists, the length of the preference lists, or the popularity of
some projects) affect the existence of a strongly stable matching, based on randomly generated
and/or real instances of spa-st.
Acknowledgement
The authors would like to convey their sincere gratitude to Adam Kunysz for valuable discussions
and helpful suggestions concerning Algorithm SPA-ST-strong. They would also like to thank
anonymous reviewers for their helpful suggestions.
References
1. D.J. Abraham, R.W. Irving, and D.F. Manlove. Two algorithms for the Student-Project allocation
problem. Journal of Discrete Algorithms, 5(1):79 -- 91, 2007.
2. A.H. Abu El-Atta and M.I. Moussa. Student project allocation with preference lists over (stu-
dent,project) pairs. In Proceedings of ICCEE 09, pp. 375 -- 379, 2009.
3. M. Baidas, Z. Bahbahani, E. Alsusa. User association and channel assignment in downlink multi-cell
NOMA networks: A matching-theoretic approach. EURASIP Journal on Wireless Communications
and Networking, 2019:220, 2019.
4. M. Baidas, M. Bahbahani, E. Alsusa, K. Hamdi and Z. Ding. D2D Group Association and Channel
Assignment in Uplink Multi-Cell NOMA Networks: A Matching Theoretic Approach. To appear in
IEEE Transactions on Communications, 2019.
5. M. Baidas, Z. Bahbahani, N. El-Sharkawi, H. Shehada and E. Alsusa. Joint relay selection and max-
min energy-efficient power allocation in downlink multicell NOMA networks: A matching-theoretic
approach. Transactions on Emerging Telecommunications Technologies, 30:5, 2019.
6. M. Chiarandini, R. Fagerberg, and S. Gualandi. Handling preferences in student-project allocation.
Annals of Operations Research, 275(1):39 -- 78, 2019.
7. Frances Cooper and David Manlove. A 3/2-Approximation Algorithm for the Student-Project
Allocation Problem. In Proceedings of SEA '18, volume 103 of Leibniz International Proceedings in
Informatics (LIPIcs), pages 8:1 -- 8:13, 2018.
8. H.N. Gabow. An efficient reduction technique for degree-constrained subgraph and bidirected
network flow problems. In Proceedings of STOC '83, pages 448 -- 456. ACM, 1983.
9. R.W. Irving. Stable marriage and indifference. Discrete Applied Mathematics, 48:261 -- 272, 1994.
10. R.W. Irving, D.F. Manlove, and S. Scott. The Hospitals/Residents problem with Ties. In Proceedings
of SWAT '00, vol. 1851 of LNCS, pp. 259 -- 271, 2000.
11. R. Irving, D. Manlove, and S. Scott. Strong stability in the Hospitals/Residents problem.
In
Proceedings of STACS '03, vol. 2607 of LNCS, pp. 439 -- 450, 2003.
12. K. Iwama, D. Manlove, S. Miyazaki, and Y. Morita. Stable marriage with incomplete lists and ties.
In Proceedings of ICALP '99, vol. 1644 of LNCS, pp. 443 -- 452, 1999.
13. K. Iwama, S. Miyazaki, and H. Yanagisawa. Improved approximation bounds for the student-project
allocation problem with preferences over projects. Journal of Discrete Algorithms, 13:59 -- 66, 2012.
14. T. Kavitha, K. Mehlhorn, D. Michail, and K. Paluch. Strongly stable matchings in time O(nm)
and extension to the Hospitals-Residents problem. In Proceedings of STACS '04, volume 2996 of
LNCS, pages 222 -- 233. Springer, 2004.
Co-ordination
Manuscript,
University
from
http://www-users.cs.york.ac.uk/kazakov/papers/proj.pdf (last accessed 26 June 2019),
2001.
student-project
of
Computer
allocation.
Science.
15. D.
Kazakov.
Available
of
York,
Department
of
28
16. A. Kwanashie, R.W. Irving, D.F. Manlove, and C.T.S. Sng. Profile-based optimal matchings in the
Student -- Project Allocation problem. In Proceedings of IWOCA '14, volume 8986 of LNCS, pages
213 -- 225. Springer, 2015.
17. A. Kunysz. An Algorithm for the Maximum Weight Strongly Stable Matching Problem. In Pro-
ceedings of ISAAC '18, vol. 123 of LIPIcs, pages 42:1 -- 42:13, 2018.
18. C.L. Liu. Introduction to Combinatorial Mathematics. McGraw-Hill, NY, 1968.
19. D. Manlove. Stable marriage with ties and unacceptable partners. Technical Report TR-1999-29,
University of Glasgow, Department of Computing Science, Jan 1999.
20. D. Manlove. Algorithmics of Matching Under Preferences. World Scientific, 2013.
21. D. Manlove, R.W. Irving, K. Iwama, S. Miyazaki, and Y. Morita. Hard variants of stable marriage.
Theoretical Computer Science, 276(1-2):261 -- 279, 2002.
22. D. Manlove, D. Milne, and S. Olaosebikan. An Integer Programming Approach to the Student-
Project Allocation Problem with Preferences over Projects. In Proceedings of ISCO '18, volume
10856 of LNCS, pp. 313 -- 325. Springer, 2018.
23. D. Manlove and G. O'Malley. Student project allocation with preferences over projects. Journal of
Discrete Algorithms, 6:553 -- 560, 2008.
24. S. Olaosebikan and D. Manlove. Super-Stability in the Student-Project Allocation Problem with
Ties. In Proceedings of COCOA '18, volume 11346 of Lecture Notes in Computer Science, pages
357 -- 371. Springer, 2018.
25. A.E. Roth. The evolution of the labor market for medical interns and residents: a case study in
game theory. Journal of Political Economy, 92(6):991 -- 1016, 1984.
|
1608.07022 | 1 | 1608 | 2016-08-25T06:16:32 | Kernelization and Parameterized Algorithms for 3-Path Vertex Cover | [
"cs.DS"
] | A 3-path vertex cover in a graph is a vertex subset $C$ such that every path of three vertices contains at least one vertex from $C$. The parameterized 3-path vertex cover problem asks whether a graph has a 3-path vertex cover of size at most $k$. In this paper, we give a kernel of $5k$ vertices and an $O^*(1.7485^k)$-time and polynomial-space algorithm for this problem, both new results improve previous known bounds. | cs.DS | cs |
Kernelization and Parameterized Algorithms for
3-Path Vertex Cover ⋆
Mingyu Xiao and Shaowei Kou
University of Electronic Science and Technology of China, Chengdu 611731, China
School of Computer Science and Engineering,
[email protected], kou [email protected]
Abstract. A 3-path vertex cover in a graph is a vertex subset C such
that every path of three vertices contains at least one vertex from C. The
parameterized 3-path vertex cover problem asks whether a graph has a
3-path vertex cover of size at most k. In this paper, we give a kernel of
5k vertices and an O∗(1.7485k)-time polynomial-space algorithm for this
problem, both new results improve previous known bounds.
1
Introduction
A vertex subset C in a graph is called an ℓ-path vertex cover if every path of
ℓ vertices in the graph contains at least one vertex from C. The ℓ-path vertex
cover problem, to find an ℓ-path vertex cover of minimum size, has been studied
in the literature [5,6]. When ℓ = 2, this problem becomes the famous vertex
cover problem and it has been well studied. In this paper we study the 3-path
vertex cover problem. A 3-path vertex cover is also known as a 1-degree-bounded
deletion set. The d-degree-bounded deletion problem [11,25,26] is to delete a
minimum number of vertices from a graph such that the remaining graph has
degree at most d. The 3-path vertex cover problem is exactly the 1-degree-
bounded deletion problem. Several applications of 3-path vertex covers have
been proposed in [6,16,27].
It is not hard to establish the NP-hardness of the 3-path vertex cover problem
by reduction from the vertex cover problem. In fact, it remains NP-hard even
in planar graphs [28] and in C4-free bipartite graphs with vertex degree at most
3 [4]. There are several graph classes, in which the problem can be solved in
polynomial time [2,3,4,6,7,14,15,18,19,20].
The 3-path vertex cover problem has been studied from approximation algo-
rithms, exact algorithms and parameterized algorithms. There is a randomized
approximation algorithm with an expected approximation ratio of 23
11 [16]. In
terms of exact algorithms, Kardos et al. [16] gave an O∗(1.5171n)-time algo-
rithm to compute a maximum dissociation set in an n-vertex graph. Chang et
al. [8] gave an O∗(1.4658n)-time algorithm and the result was further improved
to O∗(1.3659n) later [27].
⋆ This is the version accepted by TAMC 2016. To appear in: TAMC 2016, LNCS 9796,
pp. 1 -- 15, 2016.
2
In parameterized complexity, this problem is fixed-parameter tractable by
taking the size k of the 3-path vertex cover as the parameter. The running
time bound of parameterized algorithm for this problem has been improved at
least three times during the last one year. Tu [22] showed that the problem can
be solved in O∗(2k) time. Wu [24] improved the result to O∗(1.882k) by using
the measure-and-conquer method. The current best result is O∗(1.8172k) by
Katrenic [17]. In this paper we will further improve the bound to O∗(1.7485k).
Another important issue in parameterized complexity is kernelization. A ker-
nelization algorithm is a polynomial-time algorithm which, for an input graph
with a parameter (G, k) either concludes that G has no 3-path vertex cover of
size k or returns an equivalent instance (G′, k′), called a kernel, such that k′ ≤ k
and the size of G′ is bounded by a function of k. Kernelization for the d-degree-
bounded deletion problem has been studied in the literature [11,25]. For d = 1,
Fellows et al.'s algorithm [11] implies a kernel of 15k vertices for the 3-path vertex
cover problem, and Xiao's algorithm [25] implies a kernel of 13k vertices. There
is another closed related problem, called the 3-path packing problem. In this
problem, we are going to check if a graph has a set of at least k vertex-disjoint
3-paths. When we discuss kernelization algorithms, most structural properties
of the 3-path vertex cover problem and the 3-path packing problem are similar.
Several previous kernelization algorithms for the 3-path packing problem are
possible to be modified for the 3-path vertex cover problem. The bound of the
kernel size of the 3-path packing problem has been improved for several times
from the first bound of 15k [21] to 7k [23] and then to 6k [10]. Recently, there is a
paper claiming a bound of 5k vertices for the 3-path packing problem in net-free
graphs [9]. Although the paper [9] provides some useful ideas, the proof in it is
incomplete and the algorithm may not stop. Several techniques for the 3-path
packing problem in [23] and [9] will be used in our kernelization algorithm. We
will give a kernel of 5k vertices for the 3-path vertex cover problem.
Omitted proofs in this extended abstract can be found in the full version of
this paper.
2 Preliminaries
We let G = (V, E) denote a simple and undirected graph with n = V vertices
and m = E edges. A singleton {v} may be simply denoted by v. The vertex set
and edge set of a graph G′ are denoted by V (G′) and E(G′), respectively. For a
subgraph (resp., a vertex subset) X, the subgraph induced by V (X) (resp., X)
is simply denoted by G[X], and G[V \ V (X)] (resp., G[V \ X]) is also written
as G \ X. A vertex in a subgraph or a vertex subset X is also called a X-vertex.
For a vertex subset X, let N (X) denote the set of open neighbors of X, i.e.,
the vertices in V \ X adjacent to some vertex in X, and N [X] denote the set of
closed neighbors of X, ie., N (X) ∪ X. The degree of a vertex v in a graph G,
denoted by d(v), is defined to be the number of vertices adjacent to v in G. Two
vertex-disjoint subgraphs X1 and X2 are adjacent if there is an edge with one
endpoint in X1 and the other in X2. The number of connected components in
3
a graph G is denoted by Comp(G) and the number of connected components of
size i in a graph G is denoted by Compi(G). thus, Comp(G) = Pi Compi(G).
A 3-path, denoted by P3, is a simple path with three vertices and two edges.
A vertex subset C is called a 3-path vertex cover or a P3V C-set if there is no
3-path in G \ C. Given a graph G = (V, E), a P3-packing P = {L1, L2, ..., Lt}
of size t is a collection of vertex-disjoint P3 in G, i.e., each element Li ∈ P is a
3-path in G and V (Li1 ) ∩ V (Li2 ) = ∅ for any two different 3-paths Lii, Li2 ∈ P.
A P3-packing is maximal if it is not properly contained in any strictly larger
P3-packing in G. The set of vertices in 3-paths in P is denoted by V (P).
Let P be a P3-packing and A be a vertex set such that A ∩ V (P) = ∅ and A
induces a graph of maximum degree 1. We use Ai to denote the set of degree-i
vertices in the induced graph G[A] for i = 0, 1. A component of two vertices
in G[A] is called an A1-edge. For each Li ∈ P, we use A(Li) to denote the set
of A-vertices that are in the components of G[A] adjacent to Li. For a 3-path
Li ∈ P, the degree-2 vertex in it is called the middle vertex of it and the two
degree-1 vertices in it are call the ending vertices of it.
3 A Parameterized Algorithm
In this section we will design a parameterized algorithm for the 3-path ver-
tex cover problem. Our algorithm is a branch-and-reduce algorithm that runs
in O∗(1.7485k) time and polynomial space, improving all previous results. In
branch-and-reduce algorithms, the exponential part of the running time is de-
termined by the branching operations in the algorithm. In a branching operation,
the algorithm solves the current instance I by solving several smaller instances.
We will use the parameter k as the measure of the instance and use T (k) to
denote the maximum size of the search tree generated by the algorithm running
on any instance with parameter at most k. A branching operation, which gener-
ates l small branches with measure decrease in the i-th branch being at least ci,
creates a recurrence relation T (k) ≤ T (k−c1)+T (k−c2)+···+T (k−cl)+1. The
largest root of the function f (x) = 1 − Pl
i=1 x−ci is called the branching factor
of the recurrence. Let γ be the maximum branching factor among all branch-
ing factors in the algorithm. The running time of the algorithm is bounded by
O∗(γk). More details about the analysis and how to solve recurrences can be
found in the monograph [13]. Next, we first introduce our branching rules and
then present our algorithm.
3.1 Branching Rules
We have four branching rules. The first branching rule is simple and easy to
observe.
Branching rule (B1): Branch on a vertex v to generate N [v] + 1 branches by
either
(i) deleting v from the graph, including it to the solution set, and decreasing k
by 1, or
4
(ii) deleting N [v] from the graph, including N (v) to the solution set, and de-
creasing k by N (v), or
(iii) for each neighbor u of v, deleting N [{u, v}] from the graph, including N ({u, v})
to the solution set, and decreasing k by N ({u, v}).
A vertex v is dominated by a neighbor u of it if v is adjacent to all neighbors
of u. The following property of dominated vertices has been proved and used
in [27].
Lemma 1. Let v be a vertex dominated by u. If there is a minimum 3-path
vertex cover C not containing v, then there is a minimum 3-path vertex cover
C′ of G such that v, u /∈ C′ and N ({u, v}) ⊆ C′.
Based on this lemma, we design the following branching rule.
Branching rule (B2): Branch on a vertex v dominated by another vertex u to
generate two instances by either
(i) deleting v from the graph, including it to the solution set, and decreasing k
by 1, or
(ii) deleting N [{u, v}] from the graph, including N ({u, v}) to the solution set,
and decreasing k by N ({u, v}) = N (v) − 1.
For a vertex v, a vertex s ∈ N2(v) is called a satellite of v if there is a
neighbor p of v such that N [p] − N [v] = {s}. The vertex p is also called the
parent of the satellite s at v.
Lemma 2. Let v be a vertex that is not dominated by any other vertex. If v
has a satellite, then there is a minimum 3-path vertex cover C such that either
v ∈ C or v, u 6∈ C for a neighbor u of v.
Branching rule (B3): Let v be a vertex that has a satellite but is not dominated
by any other vertex. Branch on v to generate N [v] instances by either
(i) deleting v from the graph, including it to the solution set, and decreasing k
by 1, or
(ii) for each neighbor u of v, deleting N [{u, v}] from the graph, including N ({u, v})
to the solution set, and decreasing k by N ({u, v}).
Lemma 3. Let v be a degree-3 vertex with a degree-1 neighbor u1 and two adja-
cent neighbors u2 and u3. There is a minimum 3-path vertex cover C such that
either C ∪ {u1, v} = ∅ or C ∪ {u1, u2, u3} = ∅.
Branching rule (B4): Let v be a degree-3 vertex with a degree-1 neighbor u1
and two adjacent neighbors u2 and u3. Branch on v to generate two instances
by either
(i) deleting N [{u1, v}] from the graph, including {u2, u3} to the solution set, and
decreasing k by 2, or
(ii) deleting N [{u2, u3}] ∪ {u1} from the graph, including N ({u2, u3}) to the
solution set, and decreasing k by N ({u2, u3}).
5
3.2 The Algorithm
We will use P3VC(G, k) to denote our parameterized algorithm. The algorithm
contains 7 steps. When we execute one step, we assume that all previous steps
are not applicable anymore on the current graph. We will analyze each step after
describing it.
Step 1 (Trivial cases) If k ≤ 0 or the graph is an empty graph, then return
the result directly. If the graph has a component of maximum degree 2, find a
minimum 3-path vertex cover S of it directly, delete this component from the
graph, and decrease k by the size of S.
After Step 1, each component of the graph contains at least four vertices. A
degree-1 vertex v is called a tail if its neighbor u is a degree-2 vertex. Let v be a
tail, u be the degree-2 neighbor of v, and w be the other neighbor of u. We show
that there is a minimum 3-path vertex cover containing w but not containing
any of u and v. At most one of u and v is contained in any minimum 3-path
vertex cover C, otherwise C ∪ {w} \ {u, v} would be a smaller 3-path vertex
cover. If none of u and v is in a minimum 3-path vertex cover C, then w must be
in C to cover the 3-path uvw and then C is a claimed minimum 3-path vertex
cover. If exactly one of u and v is contained in a minimum 3-path vertex cover
C, then C′ = C ∪ {w} \ {u, v} is a claimed minimum 3-path vertex cover.
Step 2 (Tails) If there is a degree-1 vertex v with a degree-2 neighbor u,
then return p3vc(G \ N [{v, u}], k − 1).
Step 3 (Dominated vertices of degree ≥ 3) If there is a vertex v of de-
gree ≥ 3 dominated by u, then branch on v with Rule (B2) to generate two
branches
p3vc(G \ {v}, k − 1) and p3vc(G \ N [{v, u}], k − N ({v, u})).
Lemma 1 guarantees the correctness of this step. Note that N ({v, u}) = d(v)−1.
This step gives a recurrence
T (k) ≤ T (k − 1) + T (k − (d(v) − 1)) + 1,
(1)
where d(v) ≥ 3. For the worst case that d(v) = 3, the branching factor of it is
1.6181.
A degree-1 vertex with a degree-1 neighbor will be handled in Step 1, a degree-
1 vertex with a degree-2 neighbor will be handled in Step 2, and a degree-1 vertex
with a neighbor of degree ≥ 3 will be handled in Step 3. So after Step 3, the
graph has no vertex of degree ≤ 1. Next we consider degree≥ 4 vertices.
Step 4 (Vertices of degree ≥ 4 with satellites) If there is a vertex v of
d(v) ≥ 4 having a satellite, then branch on v with Rule (B3) to generate d(v) + 1
branches
p3vc(G\{v}, k−1) and p3vc(G\N [{v, u}], k−N ({v, u})) for each u ∈ N (v).
6
The correctness of this step is guaranteed by lemma 2. Note that there is no
dominated vertex after Step 3. Each neighbor u of v is adjacent to at least one
vertex in N2(v) and then N ({v, u}) ≥ d(v).
This step gives a recurrence
T (k) ≤ T (k − 1) + d(v) · T (k − d(v)) + 1,
(2)
where d(v) ≥ 4. For the worst case that d(v) = 4, the branching factor of it is
1.7485.
After Step 4, if there is still a vertex of degree ≥ 4, we use the following
branching rule. Note that now each neighbor u of v is adjacent to at least two
vertices in N2(v) and then N ({v, u}) ≥ d(v) + 1.
Step 5 (Normal vertices of degree ≥ 4 ) If there is a vertex v of d(v) ≥
4, then branch on v with Rule (B1) to generate d(v) + 2 branches
p3vc(G \ {v}, k − 1), p3vc(G \ N [v], k − N (v))
and p3vc(G \ N [{v, u}], k − N ({v, u})) for each u ∈ N (v).
Since N ({v, u}) ≥ d(v) + 1, this step gives a recurrence
T (k) ≤ T (k − 1) + T (k − d(v)) + d(v) · T (k − (d(v) + 1)) + 1,
(3)
which d(v) ≥ 4. For the worst case that d(v) = 4, the branching factor of it is
1.6930.
After Step 5, the graph has only degree-2 and degree-3 vertices. We first
consider degree-2 vertices.
A path u0u1u2u3 of four vertices is called a chain if the first vertex u0 is of
degree ≥ 3 and the two middle vertices are of degree 2. Note that there is no
chain with u0 = u3 after Step 3. So when we discuss a chain we always assume
that u0 6= u3. A chain can be found in linear time if it exists. In a chain u0u1u2u3,
u2 is a satellite of u0 with a parent u1.
Step 6 (Chains) If there is a chain u0u1u2u3, then branch on u0 with
Rule (B3). In the branch where u0 is deleted and included to the solution set, u1
becomes a tail and we further handle the tail as we do in Step 2.
We get the following branches
p3vc(G \ N [{u1, u2}], k − 2)
and p3vc(G \ N [{u0, u}], k − N ({u0, u})) for each u ∈ N (u0).
Note that N ({u0, u}) ≥ d(u0) since there is no dominated vertex. We get a
recurrence
T (k) ≤ T (k − 2) + d(u0) · T (k − d(u0)) + 1,
where d(u0) ≥ 3. For the worst case that d(u0) = 3, the branching factor of it is
1.6717.
After Step 6, each degree-2 vertex must have two nonadjacent degree-3 ver-
tices. Note that no degree-2 is in a triangle if there is no dominated vertex.
Step 7 (Degree-2 vertices with a neighbor in a triangle) If there is a
degree-2 vertex v with N (v) = {u, w} such that a neighbor u of it is in a triangle
uu1u2, then branch on w with Rule (B1) and then in the branch w is deleted
and included in the solution set further branch on u with Rule (B4). We get the
following branches
7
p3vc(G \ N [{u, v}], k − N ({u, v})),
p3vc(G \ N [{u1, u2}] ∪ {u, w}, k − N ({u1, u2}) ∪ {w}), and
p3vc(G \ N [{w, u′}], k − N ({w, u′})) for each u′ ∈ N (w).
There two neighbors u and w of v are degree-3 vertices. Since there is no
dominated vertex, for any edge v1v2 it holds N ({v1, v2}) ≥ min{d(v1), d(v2)}.
We know that N ({u, v}) ≥ d(u) = 3, N ({u1, u2}) ∪ {w} ≥ N ({u1, u2}) ≥ 3
(since no degree-2 vertex is in a triangle) and N ({w, u′}) ≥ d(w) for each
u′ ∈ N (w). We get the following recurrence
T (k) ≤ T (k − 3) + T (k − 3) + 3 · T (k − 3) + 1.
The branching factor of it is 1.7100.
After Step 7, no degree-3 vertex in a triangle is adjacent to a degree-2 vertex.
Step 8 (Degree-2 vertices v with a degree-3 vertex in N2(v)) If there
is a degree-2 vertex v such that at least one of its neighbors u and w, say u, has
a degree-3 neighbor u1, then branch on u with Rule (B1) and in the branch where
u is deleted and included to the solution set, branch on w with Rule (B2). We
get the branches
p3vc(G \ {u, v, w}, k − 2), p3vc(G \ N [{w, v}], k − N ({w, v})),
and p3vc(G \ N [{u, u′}], k − N ({u, u′})) for each u′ ∈ N (u).
Note that d(u) = d(w) = 3. It holds N ({w, v}) ≥ d(w) = 3 and N ({u, u′}) ≥
d(u) = 3 for u′ ∈ N (u). Furthermore, we have that N ({u, u1}) ≥ 4 because u
and u1 are degree-3 vertices not in any triangle. We get the following recurrence
T (k) ≤ T (k − 2) + T (k − 3) + 2 · T (k − 3) + T (k − 4).
The branching factor of it is 1.7456.
Lemma 4. After Step 8, if the graph is not an empty graph, then each compo-
nent of the graph is either a 3-regular graph or a bipartite graph with one side
of degree-2 vertices and one side of degree-3 vertices.
Lemma 5. Let G = (V1 ∪ V2, E) be a bipartite graph such that all vertices in V1
are of degree 2 and all vertices in V2 are of degree 3. The set V1 is a minimum
3-path vertex cover of G.
Step 9 (Bipartite graphs) If the graph has a component H being a bi-
partite graph with one side V1 of degree-2 vertices and one side V2 of degree-3
vertices, then return p3vc(G \ H, k − V1).
8
Step 10 (3-regular graphs) If the graph is a 3-regular graph, pick up an
arbitrary vertex v and branch on it with Rule (B1).
Lemma 4 shows that the above steps cover all the cases, which implies the cor-
rectness of the algorithm. Note that all the branching operations except Step 10
in the algorithm have a branching factor at most 1.7485. We do not analyze
the branching factor for Step 10, because this step will not exponentially in-
crease the running time bound of our algorithm. Any proper subgraph of a
connected 3-regular graph is not a 3-regular graph. For each connected compo-
nent of a 3-regular graph, Step 10 can be applied for at most one time and all
other branching operations have a branching factor at most 1.7485. Thus each
connected component of a 3-regular graph can be solved in O∗(1.7485k) time.
Before getting a connected component of a 3-regular graph, the algorithm always
branches with branching factors of at most 1.7485. Therefore,
Theorem 1. The 3-path vertex cover problem can be solved in O∗(1.7485k) time
and polynomial space.
4 Kernelization
In this section, we show that the parameterized 3-path vertex cover problem
allows a kernel with at most 5k vertices.
4.1 Graph decompositions
The kernelization algorithm is based on a vertex decomposition of the graph,
called good decomposition, which can be regarded as an extension of the crown
decomposition [1]. Based on a good decomposition we show that an optimal
solution to a special local part of the graph is contained in an optimal solution
to the whole graph. Thus, once we find a good decomposition, we may be able
to reduce the graph by adding some vertices to the solution set directly. We
only need to find good decompositions in polynomial time in graphs with a large
size to get problem kernels. Some previous rules to kernels for the parameterized
3-path packing problem [10,12,23] are adopted here to find good decompositions
in an effective way.
Definition 1. A good decomposition of a graph G = (V, E) is a decomposition
(I, C, R) of the vertex set V such that
1. the induced subgraph G[I] has maximum degree at most 1;
2. the induced subgraph G[I ∪ C] has a P3-packing of size C;
3. no vertex in I is adjacent to a vertex in R.
Lemma 6. A graph G that admits a good decomposition (I, C, R) has a P3-
vertex cover (resp., P3-packing) of size k if and only if G[R] has a P3-vertex
cover (resp., P3-packing) of size k − C.
9
Lemma 6 provides a way to reduce instances of the parameterized 3-path
vertex cover problem based on a good decomposition (I, C, R) of the graph:
deleting I ∪ C from the graph and adding C to the solution set. Here arise a
question: how to effectively find good decompositions? It is strongly related to
the quality of our kernelization algorithm. The kernel size will be smaller if we
can polynomially compute a good decomposition in a smaller graph. Recall that
we use Comp(G′) and Compi(G′) to denote the number of components and
number of components with i vertices in a graph G′, respectively. For a vertex
subset A that induces a graph of maximum degree at most 1 and j = {1, 2}, we
use Nj(A) ⊆ N (A) to denote the set of vertices in N (A) adjacent to at least one
component of size j in G[A], and N ′
2(A) ⊆ N2(A) be the set of vertices in N (A)
adjacent to at least one component of size 2 but no component of size 1 in G[A].
We will use the following lemma to find good decompositions, which was also
used in [9] to design kernel algorithms for the 3-path packing problem.
Lemma 7. Let A be a vertex subset of a graph G such that each connected
component of the induced graph G[A] has at most 2 vertices. If
Comp(G[A]) > 2N (A) − N ′
2(A),
(4)
then there is a good decomposition (I, C, R) of G such that ∅ 6= I ⊆ A and
C ⊆ N (A). Furthermore, the good decomposition (I, C, R) together with a P3-
packing of size C in G[I ∪ C] can be computed in O(√nm) time.
By using Lemma 7, we can get a linear kernel for the parameterized 3-path
vertex cover problem quickly. We find an arbitrary maximal P3-packing S and let
A = V \ V (S). We assume that S contains less than k 3-paths and then V (S) <
3k, otherwise the problem is solved directly. Note that N (A) ⊆ V (S). If
A > 12k, then Comp(G[A]) ≥ A
2 > 6k > 2V (S) ≥ 2N (A) and we reduce
the instance by Lemma 7. So we can get a kernel of 15k vertices. This bound
can be improved by using a special case of Lemma 7.
For a vertex subset A such that G[A] has maximum degree at most 1. Let A0
be the set of degree-1 vertices in G[A]. Note that Comp(G[A0]) = Comp2(G[A])
and N (A0) = N2(A0) = N2(A). By applying Lemma 7 on A0, we can get
Corollary 1 Let A be a vertex subset of a graph G such that each connected
component of the induced graph G[A] has at most 2 vertices. Let N2(A) ⊆ N (A)
be the set of vertices in N (A) adjacent to at least one vertex in a component of
size 2 in G[A]. If
Comp2(G[A]) > N2(A),
(5)
then there is a good decomposition (I, C, R) of G such that ∅ 6= I ⊆ A and
C ⊆ N (A). Furthermore, the good decomposition (I, C, R) together with a P3-
packing of size C in G[I ∪ C] can be computed in O(√nm) time.
Note that A = Comp1(G[A])+2·Comp2(G[A]). If A > 9k, then Comp1(G[A])+
2 · Comp2(G[A]) = A > 9k > 3V (S) ≥ 3N (A) ≥ (2N (A) − N ′
2(A)) +
10
N2(A) and at least one of (4) and (5) holds. Then by using Lemma 7 and
Corollary 1, we can get a kernel of size 9k + 3k = 12k. It is possible to bound
N (A) by k and then to get a kernel of size 3k + 3k = 6k. To further improve
the kernel size to 5k, we need some sophisticated techniques and deep analyses
on the graph structure.
4.2 A 5k kernel
In this section, we use "crucial partitions" to find good partitions. A vertex
partition (A, B, Z) of a graph is called a crucial partition if it satisfies Basic
Conditions and Extended Conditions. Basic Conditions include the following
four items:
(B1) A induces a graph of degree at most 1;
(B2) B is the vertex set of a P3-packing P;
(B3) No vertex in A is adjacent to a vertex in Z;
(B4) Z ≤ 5 · γ(G[Z]), where γ(G[Z]) is the size of a minimum P3V C-set in the
induced subgraph G[Z].
Before presenting the definition of Extended Conditions, we give some used
definitions. We use Pj to denote the collection of 3-paths in P having j vertices
adjacent to A-vertices (j = 0, 1, 2, 3). Then P = P0 ∪P1 ∪P2 ∪P3. We use P 1 to
denote the collection of 3-paths L ∈ P such that A(L) = 1. We also partition
P1 \ P 1 into two parts:
let PM ⊆ P1 \ P 1 be the collection of 3-paths with the middle vertex adjacent
to some A-vertices;
let PL ⊆ P1 \ P 1 be the collection of 3-paths Li such that A(Li) ≥ 2 and one
ending vertex of Li is adjacent to some A-vertices.
A vertex in a 3-path in P is free if it is not adjacent to any A-vertex. A
3-path in P0 is bad if it has at least two vertices adjacent to some free-vertex in
a 3-path in PL and good otherwise. A 3-path in PL is bad if it is adjacent to a
bad 3-path in P0 and good otherwise.
Extended Conditions include the following seven items:
vertex in A, i.e., P \ P 1 = P0 ∪ P1;
(E1) For each 3-path Li ∈ P \ P 1, at most one vertex in Li is adjacent to some
(E2) No 3-path in PM is adjacent to both of A0-vertices and A1-vertices;
(E3) No free-vertex in a 3-path in PL is adjacent to a free-vertex in another 3-path
in PL;
(E4) No free-vertex in a 3-path in PL is adjacent to a free-vertex in a 3-path in
PM ;
(E5) Each 3-path in P 1 has at most one vertex adjacent to a free-vertex in a
3-path in PL;
(E6) If a 3-path in P 0 has at least two vertices adjacent to some free-vertex in
a 3-path in PL, then all those free-vertices are from one 3-path in PL, i.e.,
each bad 3-path in P 0 is adjacent to free-vertices in only one bad 3-path in
PL;
11
(E7) No free-vertex in a 3-path in PL is adjacent to a vertex in Z.
Lemma 8. A crucial partition of the vertex set of any given graph can be found
in polynomial time.
After obtaining a crucial partition (A, B, Z), we use the following three re-
duction rules to reduce the graph. In fact, Extended Conditions are mainly used
for the third reduction rule and the analysis of the kernel size.
Reduction Rule 1 If the number of 3-paths in P is greater than k −Z/5,
halt and report it as a no-instance.
Note that each P3V C-set of the graph G must contain at least Z/5 vertices
in Z by Basic Condition (B4) and each P3V C-set must contain one vertex from
each 3-path in P. If the number of 3-paths in P is greater than k − Z/5, then
any P3V C-set of the graph has a size greater than k.
Reduction Rule 2 If Comp2(G[A]) > N2(A) (the condition in Corol-
lary 1) holds, then find a good decomposition by Corollary 1 and reduce the
instance based on the good decomposition.
Reduction Rule 2 is easy to observe. Next, we consider the last reduction
rule. Let B∗ be the set of free-vertices in good 3-paths in PL and let A∗ be
the set of A0-vertices adjacent to 3-paths in P 1. Let A′ = A ∪ B∗ \ A∗. By the
definition of crucial decompositions, we can get that
Lemma 9. The set A′ still induces of a graph of maximum degree 1.
Proof. Vertices in B∗ are free-vertices and then any vertex in B∗ is not adjacent
to a vertex in A. Furthermore, no two free-vertices in B∗ from two different 3-
paths in PL are adjacent by Extended Condition (E3). Since A induces a graph
of maximum degree 1, we know that A∪ B∗ induces a graph of maximum degree
1. The set A′ = A∪ B∗ \ A∗ is a subset of A∪ B∗ and then A′ induces of a graph
⊓⊔
of maximum degree 1.
Based on Lemma 9, we can apply the following reduction rule.
Reduction Rule 3 If Comp(G[A′]) > 2N (A′) − N ′
2(A′) (the condition
in Lemma 6 on set A′) holds, then find a good decomposition by Lemma 6 and
reduce the instance based on the good decomposition.
Next, we assume that none of the three reduction rules can be applied and
prove that the graph has at most 5k vertices.
We consider a crucial partition (A, B, Z) of the graph. Let k1 be the number
of 3-paths in P. Since Reduction Rule 1 cannot be applied, we know that
k1 ≤ k − Z/5.
(6)
12
Since Reduction Rule 2 and Reduction Rule 3 cannot be applied, we also have
the following two relations
and
Comp2(G[A]) ≤ N2(A),
Comp(G[A′]) ≤ 2N (A′) − N ′
2(A′).
(7)
(8)
By Extended Condition (E1), we know that P = P0 ∪ P1 ∪ P 1 = P0 ∪
PL ∪ PM ∪ P 1. Let x1 and x2 be the numbers of good and bad 3-paths in PL,
respectively. Let yi (i = 0, 1) be the number of 3-paths in P0 with i vertices
adjacent to some free-vertex in a 3-path in PL, and y2 be the number of 3-paths
in P0 with at least two vertices adjacent to some free-vertex in a 3-path in PL,
i.e., the number of bad 3-paths in P0. Let z1 and z2 be the numbers of 3-paths
in PM adjacent to only A0-vertices and only A1-vertices, respectively. Let w1 be
the number of 3-paths in P 1 adjacent to some free-vertex in a 3-path in PL and
w2 be the number of 3-paths in P 1 not adjacent to any free-vertex in a 3-path
in PL. We get that
k1 = x1 + x2 + y0 + y1 + y2 + z1 + z2 + w1 + w2.
(9)
By Extended Conditions (E1) and (E2), we know that
N (A)2 ≤ x1 + x2 + z2.
(10)
Extended Condition (E6) implies the number of bad 3-paths in PL is at most
the number of bad 3-paths in P0, i.e.,
x2 ≤ y2.
(11)
Each 3-path in P 1 is adjacent to only one A0-vertex. Since A∗ is the set of
A0-vertices adjacent to 3-paths in P 1, we know that A∗ is not greater than
w1 + w2, i.e., the number of 3-paths in P 1. By the definition of A′, we know that
(12)
Comp(G[A′]) ≥ Comp(G[A]) + x1 − (w1 + w2).
Next, we consider N (A′) and N ′
2(A′). Note that each 3-path has at most
one vertex adjacent to vertices in A \ A∗ by Extended Condition (E1). This
property will also hold for the vertex set A′ = (A \ A∗) ∪ B∗. We prove the
following two relations
N (A′) ≤ x1 + x2 + y0 + y1 + z1 + z2 + w1,
and
N ′
2(A′) ≥ y1 + z2 + w1.
(13)
(14)
By Extended Conditions (E1) and (E3), we know that each 3-path in PL has at
most one vertex in N ((A\ A∗)∪ B∗) = N (A′). By the definition of good 3-paths
13
in P0, we know that each good 3-path in P0 has no vertex adjacent to vertices
in A and has at most one vertex adjacent to vertices in B∗ (which will be in a
component of size 2 in G[A′]). There are exactly y1 vertices in good 3-paths in
P0 adjacent to vertices in B∗. No vertex in a bad 3-path in P0 is adjacent to a
vertex in A ∪ B∗ by the definitions of bad 3-paths and B∗. Each 3-path in PM
has at most one vertex adjacent to A′ by Extended Conditions (E1) and (E4).
Only z2 vertices in 3-paths in PM are adjacent to vertices in A′, all of which are
vertices of degree-1 in G[A′]. No vertex in a 3-path in P 1 is adjacent to a vertex
in A \ A∗ ⊇ A′ by the definition of A∗. Furthermore, each 3-path in P 1 has at
most one vertex adjacent to vertices in B∗ (which will be in a component of size
2 in G[A′]) by Extended Condition (E5) and there are exactly w1 vertices in
3-paths in P 1 adjacent to vertices in B∗. No vertex in Z is adjacent to a vertex
in A ∪ B∗ by Basic Condition (B3) and Extended Condition (E7). Summing all
above up, we can get (13) and (14).
Relations (8), (12), (13) and (14) imply
Comp(G[A]) ≤ 2(x2 + y0 + z1 + w1) + x1 + y1 + z2 + w2.
According to (7) and (10), we know that
(15)
(16)
Comp2(G[A]) ≤ x1 + x2 + z2.
Note that A = Comp(G[A]) + Comp2(G[A]), we get
A = Comp(G[A]) + Comp2(G[A])
≤ 2(x1 + x2 + y0 + z1 + z2 + w1) + x2 + y1 + w2
≤ 2(x1 + x2 + y0 + z1 + z2 + w1) + y2 + y1 + w2
≤ 2k1
by (15) and (16)
by (11)
by (9).
Note that B = 3k1 and k1 ≤ k − Z/5 by (6). We get that
V = A + B + Z
≤ 5k1 + Z ≤ 5k.
Theorem 2. The parameterized 3-path vertex cover problem allows a kernel of
at most 5k vertices.
References
1. FN.Abu-Khzam, RL.Collins, MR.Fellows, MA.Langston: Kernelization Algorithms
for the Vertex Cover Problem: Theory and Experiments. ALENEX/ANALC. 62 -- 69
(2004)
2. V.E.Alekseev, R.Boliac, D.V.Korobitsyn, V.V.Lozin: NP-hard graph problems and
boundary classes of graphs. Theoretical Computer Science. 389(1 -- 2), 219 -- 236 (2007)
3. K.Asdre, S.D.Nikolopoulos, C.Papadopoulos: An optimal parallel solution for the
path cover problem on P4-sparse graphs. Journal of Parallel and Distributed Com-
puting. 67(1), 63 -- 76 (2007)
14
4. R.Boliac, K.Cameron, V.V.Lozin: On computing the dissociation number and the
induced matching number of bipartite graphs. Ars Combinatoria. 72, 241 -- 253 (2004)
5. B.Bresar, M.Jakovac, J.Katrenic, G.Semanisin, A.Taranenko: On the vertex k-path
cover. Discrete Applied Mathematics. 161(13 -- 14), 1943 -- 1949 (2013)
6. B.Bresar, F.Kardos, J.Katrenic, G.Semanisin: Minimum k-path vertex cover. Dis-
crete Applied Mathematics. 159(12), 1189 -- 1195 (2011)
7. K.Cameron, P.Hell: Independent packings in structured graphs. Mathematical Pro-
gramming. 105(2 -- 3) , 201 -- 213 (2006)
8. M-S.Chang, L-H.Chen, L-J.Hung, Y-Z.Liu, P.Rossmanith,
S.Sikdar: An
O∗(1.4658n)-time exact algorithm for
the maximum bounded-degree-1 set
problem. In: The 31st Workshop on Combinatorial Mathematics and Computation
Theory. pp, 9 -- 18 (2014)
9. M-S.Chang, L-H.Chen, L-J.Huang: A 5k kernel for P2-packing in net-free graphs.
International Computer Science and Engineering Conference, 12 -- 17, IEEE (2014)
10. J.Chen, H.Fernau, P.Shaw, J.Wang, Z.Yang: Kernels for Packing and Covering
Problems. In FAW-AAIM 2012, LNCS 7285, 199 -- 211. Springer, Heidelberg (2012)
11. MR.Fellows, J.Guo, H.Moser, R.Niedermeier: A generalization of Nemhauser and
Trotter's local optimization theorem. JCSS 77(6), 1141 -- 1158 (2011)
12. H.Fermau, D.Raible: A parameterized perspective on packing paths of length two.
Journal of Combinatorial Optimization. 18(4), 319 -- 341 (2009)
13. F.V.Fomin, D.Kratsch: Exact exponential algorithms. Berlin:Springer (2010)
14. F.Goring, J.Harant, D.Rautenbach, I.Schiermeyer: On F-independence in graphs.
Discussiones Mathematica Graph Theory. 29(2), 377 -- 383 (2009)
15. R-W.Hung, M-S.Chang: Finding a minimum path cover of a distance-hereditary
graph in polynomial time. Discrete Applied Mathematics. 155(17), 2242 -- 2256 (2007)
16. F.Kardos, J.Katrenic: On computing the minimum 3-path vertex cover and disso-
ciation number of graphs. Theoretical Computer Science. 412(50), 7009 -- 7017 (2011)
17. J.Katrenic: A faster FPT algorithm for 3-path vertex cover. Information Processing
Letters. 116(4): 273 -- 278(2016)
18. V.V.Lozin, D.Rautenbach: Some results on graphs without long induced paths.
Information Processing Letters. 88(4), 167 -- 171 (2003)
19. Y.Orlovich, A.Dolgui, G.Finke, V.Gordon, F.Werner: The complexity of dissocia-
tion set problems in graphs. Disc. Appl. Math. 159(13), 1352 -- 1366 (2011)
20. C.H.Papadimitriou, M.Yannakakis: The complexity of restricted spanning tree
problems. Journal of ACM. 29(2), 285 -- 309 (1982)
21. E.Prieto, C.Sloper: Looking at the stars. Theor. Comp. Sci. 351(3), 437 -- 445 (2006)
22. J.Tu: A fixed-parameter algorithm for the vertex cover P3 problem. Information
Processing Letters. 115(2), 96 -- 99 (2015)
23. J.Wang, D.Ning, Q.Feng, J.Chen: An improved kernelization for P2-packing. In-
formation Processing Letters. 110(5), 188 -- 192 (2010)
24. B.Y.Wu: A Measure and Conquer Approach for the Parameterized Bounded
Degree-One Vertex Deletion. In: COCOON 2015. LNCS 9198, 469 -- 480 (2015)
25. M.Xiao: On a Generalization of Nemhauser and Trotter's Local Optimization The-
orem. In: ISAAC 2015. LNCS 9472, pp. 442 -- 452. Springer, Heidelberg (2015)
26. M. Xiao: A Parameterized Algorithm for Bounded-Degree Vertex Deletion. In:
T.N. Dinh and M.T. Thai (Eds.): COCOON 2016, Springer, Heidelberg (2016)
27. M.Xiao, S.Kou: Exact algorithms for the maximum dissociation set and min-
(2016)
imum 3-path vertex cover problems. Theoretical Computer Science.
doi:10.1016/j.tcs.2016.04.043
28. M.Yannakakis: Node-deletion problems on bipartite graphs. SIAM Journal on
Computing. 10(2), 310 -- 327 (1981)
|
1511.00310 | 2 | 1511 | 2017-04-10T11:39:57 | Parameterized Integer Quadratic Programming: Variables and Coefficients | [
"cs.DS"
] | In the Integer Quadratic Programming problem input is an n*n integer matrix Q, an m*n integer matrix A and an m-dimensional integer vector b. The task is to find a vector x in Z^n, minimizing x^TQx, subject to Ax <= b. We give a fixed parameter tractable algorithm for Integer Quadratic Programming parameterized by n+a. Here a is the largest absolute value of an entry of Q and A. As an application of our main result we show that Optimal Linear Arrangement is fixed parameter tractable parameterized by the size of the smallest vertex cover of the input graph. This resolves an open problem from the recent monograph by Downey and Fellows. | cs.DS | cs |
Parameterized Integer Quadratic Programming: Variables and
Coefficients
Daniel Lokshtanov∗
April 11, 2017
Abstract
In the Integer Quadratic Programming problem input is an n × n integer matrix
Q, an m × n integer matrix A and an m-dimensional integer vector b. The task is to find
a vector x ∈ Zn minimizing xT Qx, subject to Ax ≤ b. We give a fixed parameter tractable
algorithm for Integer Quadratic Programming parameterized by n + α. Here α is the
largest absolute value of an entry of Q and A. As an application of our main result we show
that Optimal Linear Arrangement is fixed parameter tractable parameterized by the
size of the smallest vertex cover of the input graph. This resolves an open problem from the
recent monograph by Downey and Fellows.
1
Introduction
While Linear Programming is famously polynomial time solvable [16], most generalizations
are not.
In particular, requiring the variables to take integer values gives us the Integer
Linear Programming problem, which is easily seen to be NP-hard. On the other hand, integer
linear programs (ILPs) with few variables can be solved efficiently. The celebrated algorithm of
Lenstra [19] solves ILPs with n variables in time f (n)LO(1) where f is a (doubly exponential)
function depending only on n and L is the total number of bits required to encode the input
integer linear program. In terms of parameterized complexity this means that Integer Linear
Programming is fixed parameter tractable (FPT) when parameterized by the number n of
variables to the input ILP. In parameterized complexity input instances come with a parameter
k, and an algorithm is called fixed parameter tractable if it solves instances of size L with
parameter k in time f (k)LO(1) for some function f depending only on k. For an introduction
to parameterized complexity we refer to the recent monograph of Downey and Fellows [6], as
well as the textbook by Cygan et al. [5].
Following the algorithm of Lenstra [19] there has been a significant amount of research into
parameterized algorithms for Integer Linear Programming, as well as generalizations of the
problem to (quasi) convex optimization. Highlights include the algorithms for Integer Linear
Programming with improved dependence on n by Kannan [14], Clarkson [4] and Frank and
Tardos [10] and generalizations to N -fold integer programming due to Hemmecke et al. [13], see
also the book by Onn [20]. Heinz [12] generalized the FPT algorithm of Lenstra to quasi-convex
polynomial optimization. More concretely, the algorithm of Heinz finds an integer assignment
to variables x1, . . . , xn minimizing f (x1, . . . xn) subject to the constraints gi(x1, . . . , xn) ≤ 0 for
1 ≤ i ≤ m, where f and g1, . . . , gm are quasi-convex polynomials of degree d ≥ 2. Here a
function f : Rn → R is quasi-convex if for every real λ the set {x ∈ Rn : f (x) ≤ λ} is convex.
The algorithm has running time LO(1)nO(dn)2O(n3), that is, it is fixed parameter tractable in the
dimension n and the degree d of the input polynomials. Khachiyan and Porkolab [15] gave an
∗Department of Informatics, University of Bergen, Norway, e-mail: [email protected]
1
even more general algorithm that covers the case of minimization of convex polynomials over
the integer points in convex semialgebraic sets given by arbitrary (not necessarily quasi-convex)
polynomials, see [17] for more details.
On the other hand, generalizations of Integer Linear Programming to optimization of
possibly non-convex functions over possibly non-convex domains quickly become computation-
ally intractable in the strongest possible sense. In fact, solving a system of quadratic equations
over 232 integer valued variables is undecidable [17], and the same holds for finding an integer
root of a single multi-variate polynomial of degree 4 [17]. Nevertheless, there are interesting spe-
cial cases of non-convex (integer) mathematical programming for which algorithms are known
to exist, and it is an under-explored research direction to investigate the parameterized com-
plexity of these problems. Perhaps the simplest such generalization is the Integer Quadratic
Programming problem. Here the input is a n× n integer matrix Q, an m× n integer matrix A
and an m-dimensional integer vector b. The task is to find a vector x ∈ Zn minimizing xT Qx,
subject to Ax ≤ b. Thus, in this problem, the domain is convex, but the objective function
might not be. It is a major open problem whether there exists a polynomial time algorithm
for Integer Quadratic Programming with a constant number of variables. Indeed, until
quite recently the problem was not even known to be in NP [21], and the first polynomial time
algorithm for Integer Quadratic Programming in two variables was given by Del Pia and
Weismantel [22] in 2014.
In this paper we take a more modest approach to Integer Quadratic Programming,
and consider the problem when parameterized by the number n of variables and the largest
absolute value α of the entries in the matrices Q and A. Our main result is an algorithm for
Integer Quadratic Programming with running time f (n, α)LO(1), demonstrating that the
problem is fixed parameter tractable when parameterized by the number of variables and the
largest coefficient appearing in the objective function and in the constraints.
On one hand Integer Quadratic Programming is a more general problem than In-
teger Linear Programming. On the other hand the parameterization by variables and
coefficients is a much stronger parameterization than parameterizing just by the number n of
variables. By making the largest entry α of Q and A a parameter we allow the running time
of our algorithms to depend in arbitrary ways on essentially all of the input. The only reason
that designing an FPT algorithm for this parameterization is non-trivial is that the entries in
the vector b may be arbitrarily large compared to the parameters n and α. This makes the
number of possible assignments to the variables much too large to enumerate all assignments
by brute force. Indeed, despite being quite restricted our algorithm for Integer Quadratic
Programming allows us to show fixed parameter tractability of a problem whose parameter-
ized complexity was unknown prior to this work. More concretely we use the new algorithm
for Integer Quadratic Programming to prove that Optimal Linear Arrangement
parameterized by the size of the smallest vertex cover of the input graph is fixed parameter
tractable.
In the Optimal Linear Arrangement problem we are given as input an undirected graph
G on n vertices. The task is to find a permutation σ : V (G) → {1, . . . , n} minimizing the cost
of σ. Here the cost of a permutation σ is val(σ, G) = Puv∈E(G) σ(u)− σ(v). The problem was
shown to be NP-complete already in the 70's [11], admits a factor O(√log n log log n) approx-
imation algorithm [2, 7], but no admits no polynomial time approximation scheme, assuming
plausible complexity-theoretic assumptions [1].
We consider Optimal Linear Arrangement parameterized by the size of the smallest
vertex cover of the input graph G. A vertex cover of a graph G is a vertex set C such that
every edge in G has at least one endpoint in C. When Optimal Linear Arrangement
is parameterized by the vertex cover number of the input graph, an integer parameter k is
also given as input together with G and n. An FPT algorithm is allowed to run in time
f (k)nO(1) and only has to provide an optimal layout σ of G if there exists a vertex cover in
2
G of size at most k. We remark that one can compute a vertex cover of size k, if it exists,
in time O(1.2748k + nO(1)) [3]. Hence, when designing an algorithm for Optimal Linear
Arrangement parameterized by vertex cover we may just as well assume that a vertex cover
C of G of size at most k is given as input.
The parameterized complexity of Optimal Linear Arrangement parameterized by ver-
tex cover was first posed as an open problem by Fellows et al. [9]. Fellows et al. [9] showed that
a number of well-studied graph layout problems, such as Bandwidth and Cutwidth can be
shown to be FPT when parameterized by vertex cover, by reducing the problems to Integer
Linear Programming parameterized by the number of variables. For the most natural for-
malization of Optimal Linear Arrangement as an integer program the objective function
is quadratic (and not necessarily convex), and therefore the above approach fails.
Motivated by the lack of progress on this problem, Fellows et al. [8] recently showed an
FPT approximation scheme for Optimal Linear Arrangement parameterized by vertex
cover. In partiular they gave an algorithm that given as input a graph G with a vertex cover
of size at most k and a rational ǫ > 0, produces in time f (k, ǫ)nO(1) a layout σ with cost at
most a factor (1 + ǫ) larger than the optimum. Fellows et al. [8] re-state the parameterized
complexity of Optimal Linear Arrangement parameterized by vertex cover as an open
problem. Finally, the problem was re-stated as an open problem in the recent monograph of
Downey and Fellows [6]. Interestingly, Downey and Fellows motivate the study of this problem
as follows.
"Our enthusiasm for this concrete problem is based on its connection to Integer Lin-
ear Programming. The problem above is easily reducible to a restricted form of Integer
Quadratic Programming which may well be FPT".
We give an FPT algorithm for Optimal Linear Arrangement parameterized by vertex
cover, resolving the open problem of [6, 8, 9]. Our algorithm for Optimal Linear Arrange-
ment works by directly applying the new algorithm for Integer Quadratic Programming
to the most natural formulation of Optimal Linear Arrangement on graphs with a small
vertex cover as an integer quadratic program, confirming the intuition of Downey and Fellows [6].
Preliminaries and notation. In Section 2 lower case letters denote vectors and scalars, while
upper case letters denote matrices. All vectors are column vectors. For an integer p ≥ 2, the
ℓp norm of an n-dimensional vector v = [v1, v2, . . . , vn] is denoted by vp and is defined to be
vp = (vp
n)1/p. The ℓ1 norm of v is v1 = v1 + v2 + . . . + vn, while the ℓ∞
norm of v is v∞ = max(v1,v2, . . . ,vn).
2 + . . . + vp
1 + vp
2 Algorithm for Integer Quadratic Programming
We consider the following problem, called Integer Quadratic Programming. Input consists
of an n × n integer symmetric matrix Q, an m × n integer matrix A and m-dimensional integer
vector b. The task is to find an optimal solution x⋆ to the following optimization problem.
Minimize xT Qx
subject to: Ax ≤ b
x ∈ Zn.
(1)
A vector x ∈ Zn that satisfies the constraints Ax ≤ b is called a feasible solution to the IQP (1).
Given an input on the form (1) there are three possible scenarios. A possible scenario is that
there are no feasible solutions, in which case this is what an algorithm for Integer Quadratic
Programming should report. Another possibility is that for every integer β there exists some
feasible solution x such that xT Qx ≤ β. In that case the algorithm should report that the IQP
is unbounded. Finally, it could be that there exist feasible solutions, and that the minimum
3
value of xT Qx over the set of all feasible x is well defined. This is the most interesting case,
and in this case the algorithm should output a feasible x such that xT Qx is minimized.
Note that the requirement that Q is symmetric can easily be avoided by replacing Q by
Q + QT . This operation does not change the set of optimal solutions, since it multiplies the
objective function value of every solution by 2. We will denote by aT
i the i'th row of the matrix
A, and by bi the i'th entry of the vector b. Thus, Ax ≤ b means that aT
i x ≤ bi for all i. The
maximum absolute value of an entry of A and Q is denoted by α. Using a pair of inequalities
one can encode equality constraints.
It is useful to rewrite the IQP (1) to separate out the
equality constraints explicitely, obtaining the following equivalent form.
Minimize xT Qx
subject to: Ax ≤ b
Cx = d
x ∈ Zn.
(2)
If input is given on the form (2),
Here C is an integer matrix and d is an integer vector.
then we still use α to denote the maximum value of an entry of A and Q. The IQP (2)
could be generalized by changing the objective function from xT Qx to xT Qx + qT x for some n-
dimensional vector q also given as input. This generalization can be incorporated in the original
formulation (2) at the cost of introducing a new variable x, adding the constraint x = 1 to the
system Cx = d and adding [0, q] as the row corresponding to the new variable x in Q.
We will denote by ∆ the maximum absolute value of the determinant of a square submatrix
of C. We may assume without loss of generality that the rows of C are linearly independent;
otherwise we may in polynomial time either conclude that the IQP has no feasible solutions,
or remove one of the equality constraints in the system Cx = d without changing the set of
feasible solutions. Thus C has at most n rows. If the IQP (2) is obtained from (1) by replacing
constraints aT
i x = bi, the maximum entry of C is also upper
bounded by α and then we have ∆ ≤ n! · αn. The next simple observation shows that we can
in polynomial time reduce the number of constraints to a function of n and α.
i x ≤ −bi with aT
i x ≤ bi, −aT
Lemma 1. There is a polynomial time algorithm that given as input the matrix A and vector
b outputs an m′ × n submatrix A′ of A and vector b′ such that m′ ≤ (2α + 1)n and for every
x ∈ Zn, Ax ≤ b if and only if A′x ≤ b′.
Proof. Suppose A has more than (2α + 1)n rows. Then, by the pigeon hole principle the system
Ax ≤ b has two rows aT
j x ≤ bj where i 6= j but ai = aj. Without loss of generality
bi ≤ bj, and then any x ∈ Zn such that aT
j x ≤ bj. Thus we can safely remove
the inequality aT
i x ≤ bi satisfies aT
i x ≤ bi and aT
j x ≤ bj from the system, and the lemma follows.
In the following we will assume that the input is on the form (2). We will give an algorithm
that runs in time f (n, m, α, ∆) · LO(1), where L is the length of the bit-representation of the
input instance. Since we can reduce the input using Lemma 1 first and ∆ is upper bounded in
terms of n and α this will yield an FPT algorithm for Integer Quadratic Programming
parameterized by n and α.
Let r be the dimension of the nullspace of C. Using Cramer's rule (see [18]) we can in
polynomial time compute a basis y1, . . . yr for the nullspace of C, such that each yi is an integer
vector and yi∞ ≤ ∆2. We let Y be the n × r matrix whose columns are the vectors y1, . . . yr.
We will abuse notation and write yi ∈ Y to denote that we chose the i'th column yi of Y . We
will say that a feasible solution x is deep if x + yi and x− yi are feasible solutions for all yi ∈ Y .
A feasible solution that is not deep is called shallow.
Lemma 2. Let x be a shallow feasible solution to (2). Then there exists a row aT
integer b′
from the rows of C.
j ∈ {bj−α·n·∆2, . . . , bj}. Further, aT
j of A and
j is linearly independent
j such that aT
j and b′
j x = b′
4
Proof. We prove the statement for yi ∈ Y such that x + yi is not a feasible solution to (2). Then
there exists a row aT
j of A such that aT
j (x + yi) > bj, and thus
bj − aT
j x ≤ bj.
j yi < aT
Thus aT
j x = b′
j for b′
We now show that aT
j ∈ {bj−α·n·yi∞, bj}. Since yi∞ ≤ ∆2 we have that b′
j ∈ {bj−α·n·∆2, bj}.
j is linearly independent from the rows of C. Suppose not, then there
exists a coefficient vector λ such that λT C = aT
j . But then
aT
j (x + yi) = λT C(x + yi) = λT Cx + λT Cyi = λT Cx = aT
j x ≤ bj,
which contradicts that aT
rows of C. The proof for the case when x − yi is not a feasible solution to (2) is symmetric.
j (x + yi) > bj. We conclude that aT
j
is linearly independent from the
Lemma 2 suggests the following branching strategy: either all optimal solutions are deep
or Lemma 2 applies to some shallow optimal solution x⋆. In the latter case the algorithm can
branch on the choice of row aT
j to the set of constraints.
This decreases the dimension of the nullspace of C by 1. We are left with handling the case
when all optimal solutions are deep.
Lemma 3. For any pair of vectors x, y ∈ Rn and symmetric matrix Q ∈ Rn×n, the following
are equivalent.
j and add the equation aT
j and b′
j x = b′
1. (x + y)T Q(x + y) ≥ xT Qx and (x − y)T Q(x − y) ≥ xT Qx,
2. −yT Qy ≤ 2xT Qy ≤ yT Qy.
Proof. Expanding the inequalities of (1) yields
xT Qx + 2xT Qy + yT Qy ≥ xT Qx,
xT Qx − 2xT Qy + yT Qy ≥ xT Qx.
Cancelling the xT Qx terms and re-organizing yields 2xT Qy ≥ −yT Qy and 2xT Qy ≤ yT Qy.
Since the left hand side of the inequalities is the same we can combine the two inequalities in a
single inequality,
−yT Qy ≤ 2xT Qy ≤ yT Qy,
completing the proof. Note that all the manipulations we did on the inequalities are reversible,
thus the above argument does indeed prove equivalence and not only the forward direction
(1) → (2).
Lemma 3 suggests a branching strategy to find a deep solution x⋆: pick a vector yi in Y
such that yT
i Q is linearly independent of the rows of C, guess the value z of 2(x⋆)T Qyi and add
the linear equation 2(x⋆)T Qyi = z to the set of constraints. In each branch the dimension of
the nullspace of C decreases by 1. Thus we are left with the case that all solutions are deep
and there is no yi in Y such that yT
i Q is linearly independent of the rows of C. We now handle
this case.
Lemma 4. For any deep optimal solution x⋆ of the IQP (2) and yi ∈ Y such that yT
i Q is
linearly dependent of the rows of C, the vectors x⋆ + yi and x⋆ − yi are also optimal solutions.
Proof. We prove the statement for x⋆ + yi. Since x⋆ is deep it follows that x⋆ + yi is feasible,
and it remains to lower bound the objective function value of x⋆ + yi. Since yT
i Q is linearly
i Q = λT C. Therefore,
dependent of the rows of C there exists a coefficient vector λT such that yT
2(x⋆ + yi)T Qyi = 2yT
i Q(x⋆ + yi) = 2λT C(x⋆ + yi)
= 2λT Cx⋆ + 2λT Cyi = 2λT Cx⋆ = 2yT
i Qx⋆ = 2(x⋆)T Qyi
5
Since x⋆ is deep, it follows that both x⋆ + yi and x⋆ − yi are feasible and therefore cannot
have a higher value of the objective function than x⋆. Hence, by Lemma 3 we have that
−yT
i Qyi. Since 2(x⋆ + yi)T Qyi = 2(x⋆)T Qyi, we have that
i Qyi ≤ 2(x⋆)T Qyi ≤ yT
−yT
i Qyi ≤ 2(x⋆ + yi)T Qyi ≤ yT
i Qyi.
Hence, Lemma 3 applied to (x⋆ + yi) implies that
(x⋆)T Qx⋆ = (x⋆ + yi − yi)T Q(x⋆ + yi − yi) ≥ (x⋆ + yi)T Q(x⋆ + yi).
This means that the objective function value of x⋆ + yi is at most that of x⋆, hence x⋆ + yi is
optimal. The proof for x⋆ − yi is symmetric.
Lemma 4 immediately implies the following corollary.
Corollary 1. Suppose the IQP (2) has an optimal solution, all optimal solutions to (2) are
deep, and for every yi ∈ Y , yT
i Q is linearly dependent of the rows of C. Then, for every optimal
solution x⋆ and integer vector λ ∈ Zr, x⋆ + Y λ is also an optimal solution of (2).
Proof. Since x⋆ is optimal and deep, and for every yi ∈ Y , yT
i Q is linearly dependent of the
rows of C, it follows from Lemma 4 that for every yi ∈ Y , x⋆ + yi and x⋆ − yi are also optimal
solutions of (2). Since all optimal solutions are deep, x⋆ +yi and x⋆−yi are deep. The statement
of the corollary now follows by induction on λ1.
We are now ready to state the main structural lemma underlying the algorithm for Integer
Quadratic Programming.
Lemma 5. For any Integer Quadratic Program of the form (2) that has an optimal solution
and any x0 such that Cx0 = d, there exists an optimal solution x⋆ such that at least one of the
following three cases holds.
1. There exists a row aT
j of A and integer b′
j ∈ {bj − α · n · ∆2, . . . , bj} such that aT
j x⋆ = b′
j,
and aT
j
is linearly independent from the rows of C.
2. There exists a yi ∈ Y such that yT
z for z ∈ {−n2∆4α, . . . , n2∆4α}.
3. x⋆ − x01 ≤ ∆2 · n.
i Q is linearly independent of the rows of C and 2yT
i Qx⋆ =
i Qyi, . . . , yT
i Qx⋆ = z for z ∈ {−yT
i Qyi}. Furthermore, yT
Proof. Suppose the integer quadratic program (2) has a shallow optimal solution x⋆. Then, by
Lemma 2 case 1 applies. In the remainder of the proof we assume that all optimal solutions are
deep. Suppose now that there is a yi ∈ Y such that yT
i Q is linearly independent of the rows of
C. Then, since x⋆ is a deep optimal solution, both x⋆ + yi and x⋆ − yi are feasible solutions,
so (x⋆ + yi)T Q(x⋆ + yi) ≥ (x⋆)T Qx⋆ and (x⋆ − yi)T Q(x⋆ − yi) ≥ (x⋆)T Qx⋆. Thus, Lemma 3
implies that 2yT
i Qyi is the sum of n2
terms where each term a product of an element of yi (and thus at most ∆2), another element
of yi, and an element of Q. Thus z ∈ {−n2∆4α, . . . , n2∆4α} and therefore case 2 applies.
Finally, suppose that all yi ∈ Y are linearly dependent of the rows of C. Let x be an
arbitrarily chosen optimal solution to (2). Since C(x0 − x) = 0 and Y forms a basis for the
nullspace of C there is a coefficient vector λ ∈ Rr such that x0 = x + Y λ. Define λ from
In other words, for every i we set
λ by rounding each entry down to the nearest integer.
λi = ⌊λi⌋. Set x⋆ = x + Y λ. By Corollary 1 we have that x⋆ is an optimal solution to (2). But
x⋆ − x01 = Y (λ − λ)1 ≤ (maxi yi1) · n ≤ ∆(C)2 · n, concluding the proof.
Theorem 1. There exists an algorithm that given an instance of Integer Quadratic Pro-
gramming, runs in time f (n, α)LO(1), and outputs a vector x ∈ Zn. If the input IQP has a
feasible solution then x is feasible, and if the input IQP is not unbounded, then x is an optimal
solution.
6
Proof. We assume that input is given on the form (2). The algorithm starts by reducing the
input system according to Lemma 1. After this preliminary step the number of constraints m
in the IQP is upper bounded by (2α + 1)n. We give a recursive algorithm, based on Lemma 5.
The algorithm begins by computing in polynomial time a basis Y = y1, . . . , yr for the nullspace
of C, as described in the beginning of Section 2. In particular Y is a matrix of integers, and for
every i, yi∞ ≤ ∆2.
If the dimension of the nullspace of C is 0 the algorithm solves the system Cx = d of linear
equations in polynomial time. Let x⋆ be the (unique) solution to this linear system. If x⋆ is not
an integral vector, or Ax⋆ ≤ b does not hold the algorithm reports that the input IQP has no
feasible solution. Otherwise it returns x⋆ as the optimum.
If C is not full-dimensional, i.e the dimension of the nullspace of C is at least 1, the algorithm
j ∈ {bj − α · n · ∆2, bj} such that
proceeds as follows. For each row aT
aT
is linearly independent from the rows of C, the algorithm calls itself recursively on the
j
same instance, but with the equation aT
j added to the system Cx = d. Furthermore,
for each yi ∈ Y such that yT
i Q is linearly independent of the rows of C and every integer
z ∈ {−n2∆4α, . . . , n2∆4α} the algorithm calls itself recursively on the same instance, but with
the equation 2yT
i Qx = z added to the system Cx = d. Finally the algorithm computes an
arbitrary (not necessarily integral) solution x0 of the system Cx = d, and checks all (integral)
vectors within ℓ1 distance at most ∆2 · n from x0. The algorithm returns the feasible solution
with the smallest objective function value among the ones found in any of the recursive calls,
and the search around x0.
j of A and integer b′
j x = b′
In the recursive calls, when we add a linear equation to the system Cx = d we extend
the matrix C and vector d to incorporate this equation. The algorithm terminates, as in each
recursive call the dimension of the nullspace of C is decreased by 1. Further, any feasible
solution found in any of the recursive calls is feasible for the original system. Thus, if the
algorithm reports a solution then it is feasible. To see that the algorithm reports an optimal
solution, consider an optimal solution x⋆ satisfying the conditions of Lemma 5 applied to the
quadratic integer program (2) and vector x0. Either x⋆ will be found in the search around x0,
or x⋆ satisfies the linear constraint added in at least one of the recursive calls. In the latter case
x⋆ is an optimal solution to the integer quadratic program of the recursive call, and in this call
the algorithm will find a solution with the same objective function value. This concludes the
proof of correctness.
We now analyze the running time of the algorithm. First, consider the time it takes to
search all integral vectors within ℓ1 distance at most ∆2 · n from x0. It is easy to see that there
are at most 3∆2·n+n such vectors. For the running time analysis only we will treat this search
as at most 3∆2·n+n recursive calls to instances where the dimension of the nullspace of C is 0.
Then the running time in each recursive call is polynomial, and it is sufficient to upper bound
the number of leaves in the recursion tree of the algorithm.
We bound the number of leaves of the recursion tree as a function of n – the number of
variables, m – the number of rows in A, α – the maximum value of an entry in A or Q, ∆ – the
maximum absolute value of the determinant of a square submatrix of C, and r – the dimension
of the nullspace of C. Notice that the algorithm never changes Q or A, and that the number
of variables remains the same throughout the execution of the algorithm. Thus n, m and α do
not change throughout the execution. For a fixed value of n, m and α, we let T (r, ∆) be the
maximum number of leaves in the recursion tree of the algorithm when called on an instance
with the given value of n, m, α, r and ∆.
In each recursive call the algorithm adds a new row to the matrix C. Let C ′ be the new
matrix after the addition of this row, r′ be the dimension of the nullspace of C and ∆′ be the
maximum value of a determinant of a square submatrix of C ′. Since the new added row is
linearly independent of the rows of C it follows that r′ = r − 1 in each of the recursive calls
arising from case 1 and case 2 of Lemma 5. The remaining recursive calls are to leaves of the
7
recursion tree.
When the algorithm explores case 1, it guesses a row aj, for which there are m possibilities,
j, for which there are α· n·∆2 possibilities. This generates m· α· n·∆2 recursive
and a value for b′
calls. In each of these recursive calls aj is the new row of C ′, and so, by the cofactor expansion
of the determinant [18], ∆′ ≤ nα∆.
When the algorithm explores case 2, it guesses a vector yi ∈ Y , and there are at most
n possibilities for yi. For each of these possibilities the algorithm guesses a value for z, for
which there are 2n2∆4α possible choices. Thus this generates 2n3∆4α recursive calls. In each
of the recursive calls the algorithm makes a new matrix C ′ from C by adding the new row
2yT
i Q∞ ≤ n · ∆2 · α, and the cofactor expansion of
the determinant [18] applied to the new row yields ∆′ ≤ n2∆3 · α, where ∆′ is the maximum
value of a determinant of a square submatrix of C ′. It follows that the number of leaves of the
recursion tree is gouverned by the following recurrence.
i Q. We have that yi∞ ≤ ∆2. Thus, 2yT
T (r, ∆) ≤ m · α · n · ∆2 · T (r − 1, nα∆) + 2n3 · ∆4 · α · T (r − 1, n2∆3α) + 3(∆2+1)·n
≤ α · m · n3 · ∆4 · T (r − 1, n2∆3α) + 3(∆2+1)·n
The above recurrence is clearly upper bounded by a function of n, m, ∆ and α. Since m is
upper bounded by (2α + 1)n from Lemma 1, the theorem follows.
2.1 Detecting Unbounded IQPs
Theorem 1 allows us to solve bounded IQPs, and is sufficient for the application to Optimal
Linear Arrangement. However, it is somewhat unsatisfactory that the algorithm of Theo-
rem 1 is unable to detect whether the input IQP is bounded or not. Next we resolve this issue.
Towards this, we inspect the algorithm of Theorem 1. For purely notational reasons we will
consider the algorithm of Theorem 1 when run on an instance on the form (1). The first step
of the algorithm is to put the the IQP on the form (2), and then proceed as described in the
proof of Theorem 1.
The algorithm is recursive, and the only variables that change from one recursive call to the
next are the matrix C and the vector d. Furthermore, when making a recursive call, the new
matrix C ′ is computed from C by either adding the row aT
i Q to C.
The vector yT
is a vector from the basis Y for the nullspace of C. In other words C ′ depends
i
only on Q, A, C and i, and is independent of b and d. Furthermore, the recursion stops when
C has full column rank. Thus, the family C of matrices C that the algorithm of Theorem 1 ever
generates depends only on the input matrices Q and A (and not on the input vector b). At this
point we remark that the only reason we assumed input was on the form (1) rather than (2)
was to avoid the confusing sentence "Thus, the family C of matrices C that the algorithm of
Theorem 1 ever generates depends only on the input matrices Q and A, and C," where the
meaning of the matrix C is overloaded.
j to C or adding the row 2yT
Let ∆ be the maximum absolute value of the determinant of a square submatrix of a matrix
C ∈ C ever generated by the algorithm. In other words, ∆ is the maximum value of the variable
∆ throughout the execution of the algorithm. Because ∆ only depends on C, it follows that ∆
only depends on C, and therefore ∆ is a function of the input matrices A and Q.
We now discuss all the different vectors d ever generated by the algorithm. In each recursive
call, the algorithm adds a new entry to the vector d, this entry is either from the set {bj − α ·
n · ∆2, bj} or from the set {−n2 ∆4α, . . . , n2 ∆4α}. Thus, any vector d ever generated by the
algorithm is at ℓ∞ distance at most n2 ∆4α from some vector whose entries are either 0 or equal
to bj for some j ≤ m. Given the vector b and the integer n, we define the vector set D(b, n)
be the set of all integer vectors in at most n dimensions with entries either 0 or equal to bj for
some j ≤ m. Observe that D(b, n) ≤ (m + 1)n. We have that every vector d ever generated
by the algorithm is at ℓ∞ distance at most n2 ∆4α from some vector in D(b, n).
8
The algorithm of Theorem 1 generates potential solutions x to the input IQP by finding a
(not necessarily integral) solution x0 to the linear system Cx = d, and then lists integral vectors
within ℓ1-distance at most ∆2 · n from x0. From Theorem 1 it follows that if the input IQP
is feasible and bounded, then one of the listed vectors x is in fact an optimum solution to the
IQP. The above discussion proves the following lemma.
Lemma 6. Given an n × n integer matrix Q, and an m × n integer matrix A, let C be the set
of matrices C generated by the algorithm of Theorem 1 when run on the IQP
Minimize xT Qx
subject to: Ax ≤ 0
x ∈ Zn,
and let ∆ be the maximum absolute value of the determinant of a square submatrix of a matrix
C ∈ C. Then, for any m-dimensional integer vector b such that the IQP (1) is feasible and
bounded, there exists a C ∈ C, a vector d0 ∈ D(b, n), and an integer vector d at ℓ∞ distance at
most n2 ∆4α from d0 such that the following is satisfied. For any x0 such that Cx0 = d, there
exists an integer vector x∗ at ℓ1 distance at most ∆2 · n from x0 such that x∗ is an optimal
solution to the IQP (1).
The algorithm in Theorem 1 only adds a row to the matrix C if this row is linearly inde-
pendent from the rows of C. Hence all matrices in C have full row rank, and therefore they
have right inverses. Specifically, for each C ∈ C, we define C −1
right = C T (CC T )−1. It follows that
CC −1
rightd is a solution to the system Cx = d for any vector
d. Note that C −1
right is not necessarly an integer matrix, however Cramer's rule [18] shows that
C −1
right is a matrix with rational entries with common denominator det(CC T ). This leads to the
following lemma.
right = I and that therefore, x0 = C −1
rightd0 + v is an optimal solution to the IQP (1).
Lemma 7. There exists an algorithm that given an n × n integer matrix Q, and an m × n
integer matrix A outputs a set C−1 of rational matrices, and a set V of rational vectors with the
following property. For any m-dimensional integer vector b such that the IQP (1) defined by Q,
A and b is feasible and bounded, there exists a matrix C −1
right ∈ C a vector v ∈ V and a vector
d0 ∈ D(b, n), such that x∗ = C −1
Proof. The algorithm starts by applying the algorithm of Lemma 6 to obtain a set C of matrices
and the integer ∆. The algorithm then computes the set C−1 of matrices, defined as C−1 =
{C −1
: C ∈ C}. Next the algorithm constructs the set V of vectors as follows. For every
right ∈ C−1, we have that C −1
matrix C −1
right = C T (CC T )−1 for some C ∈ C. For every integer
vector v0 with v0∞ ≤ n2 ∆4α, and every rational vector v1 with v11 ≤ ∆2· n and denominator
det(CC T ) in every entry, the algorithm adds C −1
rightv0 + v1 to V. It remains to prove that C−1
and V satisfy the statement of the lemma.
Let b be an m-dimensional integer vector such that the IQP (1) defined by Q, A and b is
feasible and bounded. By Lemma 6 we have that there exists C ∈ C, a vector d0 ∈ D(b, n), and
a vector d at ℓ∞ distance at most n2 ∆4α from d0 such that the following is satisfied. For any
x0 such that Cx0 = d, there exists an integer vector x∗ at ℓ1 distance at most ∆2 · n from x0
such that x∗ is an optimal solution to the IQP (1).
Let C ∈ C, d0 ∈ D(b, n), and d be as guaranteed by Lemma 6, and set v0 = d − d0. We
have that v0∞ ≤ n2 ∆4α. Let C −1
rightd =
rightd0 + C −1
C −1
Thus, there exists an integer vector x∗ at ℓ1 distance at most ∆2 · n from x0 such that
x∗ is an optimal solution to the IQP (1). Let v1 = x∗ − x0, we have that v11 ≤ ∆2 · n.
Further, x∗ is an integer vector, while x0 = C −1
rightd is a rational vector whose entries all have
right be the right inverse of C in C−1, and let x0 = C −1
right
rightv0. We have that Cx0 = d.
9
denominator det(CC T ).
v = C −1
proof.
rightv0 + v1 ∈ V and C −1
It follows that all entries of v1 have denominator det(CC T ). Thus
rightd0 + v is an optimal solution to the IQP (1), completing the
Armed with Lemma 7 we are ready to prove the main result of this section.
Theorem 2. There exists an algorithm that given an instance of Integer Quadratic Pro-
gramming, runs in time f (n, α)LO(1), and determines whether the instance is infeasible, feasi-
ble and unbounded, or feasible and bounded. If the instance is feasible and bounded the algorithm
outputs an optimal solution.
Proof. The algorithm of Theorem 1 is sufficient to determine whether the input instance is
feasible, and to find an optimal solution if the instance is feasible and bounded. Thus, to
complete the proof it is sufficient to give an algorithm that determines whether the input IQP
is unbounded. We will assume that the input IQP is given on the form (1).
Suppose now that the input IQP is unbounded. For a positive integer λ, consider adding
the linear constraints Ix ≤ λ1, and −Ix ≤ λ1 to the IQP. Here 1 is the n-dimensional all-ones
vector. In other words, we consider the IQP where the goal is to minimize xT Qx, subject to
A′x ≤ b′ where A′ is obtained from A by adding 2n new rows containing I and −I, and b′ is
obtained from b by adding 2n new entries with value λ. There exists a λ0 such that for every
λ ≥ λ0 this IQP is feasible. Further, for every λ the resulting IQP is bounded. Note that the
matrix A′ does not depend on λ.
We now apply Lemma 7 on Q and A′, and obtain a set C−1 of rational matrices, and a set V
of rational vectors. We have that for every λ ≥ λ0, there exists a matrix C −1
right ∈ C a v ∈ V and
a d0 ∈ D(b′, n), such that x∗ = C −1
rightd0 + v is an optimal solution to the IQP defined by Q, A′
and b′. Furthermore, the entries of b′ are either entries of b, 0 or equal to λ. Hence d0 = db + λd1
for db ∈ D(b, n) and d1 ∈ D(1, n). We can conclude that there exists a matrix C −1
right ∈ C a
vector v ∈ V, a vector db ∈ D(b, n) and a vector d1 ∈ D(1, n), such that
rightdb + v)
rightd1) + (C −1
C −1
right(db + λd1) + v = λ · (C −1
is an optimal solution to the IQP defined by Q, A′ and b′. Thus, the input IQP is unbounded if
and only if there exists a choice for C −1
right ∈ C, v ∈ V, db ∈ D(b, n) and d1 ∈ D(1, n), such that
following univariate quadratic program with integer variable λ is unbounded.
x = λ · (C −1
rightd1) + (C −1
Minimize xT Qx
subject to: Ax ≤ b
rightdb + v)
λ ∈ Z.
Hence, to determine whether the input IQP is unbounded, it is sufficient to iterate over all
choices of C −1
right ∈ C, v ∈ V, db ∈ D(b, n) and d1 ∈ D(1, n), and determine whether the resulting
univariate quadratic program is unbounded. Since the number of such choices is upper bounded
by a function of Q and A, and univariate (both integer and rational) quadratic programming
is trivially decidable, the theorem follows.
3 Optimal Linear Arrangement Parameterized by Vertex Cover
We assume that a vertex cover C of G of size at most k is given as input. The remaining set
of vertices I = V (G) − C forms an independent set. Furthermore, I can be partitioned into at
most 2k sets as follows: for each subset S of C we define IS = {v ∈ I : N (v) = S}. For every
vertex v ∈ I we will refer to N (v) as the type of v, clearly there are at most 2k different types.
10
Let C = {c1, c2, . . . , ck}. By trying all k! permutations of C we may assume that the optimal
solution σ satisfies σ(ci) < σ(ci+1) for every 1 ≤ i ≤ k − 1. For every i between 1 and k − 1
we define the i'th gap of σ to be the set Bi of vertices appearing between ci and ci+1 according
to σ. The 0'th gap B0 is the set of all vertices appearing before c1, and the k'th gap Bk is the
set of vertices appearing after ck. We will also refer to Bi as "gap i" or "gap number i". For
every gap Bi and type S ⊆ C of vertices we denote by I i
S the set Bi ∩ IS of vertices of type S
appearing in gap i.
We say that an ordering σ is homogenous if, for every gap Bi and every type S ⊆ C the
vertices of I i
S appear consecutively in σ. Informally this means that inside the same gap the
vertices from different sets IS and IS ′ "don't mix". Fellows et al. [8] show that there always
exists an optimal solution that is homegenous.
Lemma 8. [8] There exists a homogenous optimal linear arrangement of G.
For every vertex v we define the force of v with respect to σ to be
δ(v) = {u ∈ N (v) : σ(u) > σ(v)} − {u ∈ N (v) : σ(u) < σ(v)}.
Notice that two vertices of the same type in the same gap have the same force. Fellows et
al. [8] in the proof of Lemma 8 show that there exists an optimal solution that is homogenous,
and where inside every gap, the vertices are ordered from left to right in non-decreasing order
by their force. We will call such an ordering solution super-homogenous. As already noted,
the existence of a super-homogenous optimal linear arrangement σ follows from the proof of
Lemma 8 by Fellows et al. [8].
Lemma 9. [8] There exists a super-homogenous optimal linear arrangement of G.
Notice that a super-homogenous linear arrangement σ is completely defined (up to swapping
positions of vertices of the same type) by specifying for each i and each type S the size I i
S.
For each gap i and each type S we introduce a variable xi
S. Clearly the
variables xi
S ∈ Z representing I i
S need to satisfy
and
∀i ≤ k,∀S ⊆ C xi
S ≥ 0
∀S ⊆ C
k
Xi=0
xi
S = IS.
(3)
(4)
On the other hand, every assignment to the variables satisfying these (linear) constraints corre-
sponds to a super-homogenous linear arrangement σ of G with I i
S for every type S and
gap i.
S = xi
We now analyze the cost of σ as a function of the variables. The goal is to show that
val(σ, G) is a quadratic function of the variables with coefficients that are bounded from above
by a function of k. The coefficients of this quadratic function are not integral, but half-integral,
namely integer multiples of 1
2 . The analysis below is somewhat tedious, but quite straightfor-
ward. For the analysis it is helpful to re-write val(σ, G). For a fixed ordering σ of the vertices
we say that an edge uv flies over the vertex w if
min(σ(u), σ(v)) < σ(w) < max(σ(u), σ(v)).
We define the "fly over" relation ∼ for edges and vertices, i.e uv ∼ w means that uv flies over
w. Since an edge uv with σ(u) < σ(v) flies over the σ(v) − σ(u) − 1 vertices appearing between
σ(u) and σ(v) it follows that
val(σ, G) = E(G) + Xuv∈E(G) Xw∈V (G)
uv∼w
1.
11
We partition the set of edges of G into several subsets as follows. The set EC is the set of
all edges with both endpoints in C. For every gap i with i ∈ {0, . . . , k}, every j ∈ {1, . . . , k}
and every S ⊆ C we denote by ES
S and the other
is cj. Notice that ES
S or to 0 depending on whether vertices of type S
are adjacent to cj or not. We have that
i,j the set of edges whose one endpoint is in I i
i,j is either equal to xi
val(σ, G) = E(G) + Xcicj ∈EC Xw∈V (G)
cicj ∼w
1 + Xi,j,S Xucj ∈ES
i,j
1.
Xw∈V (G)
ucj ∼w
(5)
Further, for each edge cicj ∈ EC (with i < j) we have that
Xp=i XS⊆C
1 = j − i − 1 +
Xw∈V (G)
j−1
cicj ∼w
xp
S.
In other words, the first double sum of Equation 5 is a linear function of the variables. Since
EC ≤ (cid:0)k
2(cid:1) the coefficients of this linear function are integers upper bounded by (cid:0)k
2(cid:1).
We now turn to analyzing the second part of Equation 5. We split the triple sum in three
parts as follows.
Xi,j,S Xucj ∈ES
i,j
1
ucj ∼w
Xw∈V (G)
1
Xw∈I−I i
S
ucj ∼w
= Xi,j,S
Xucj ∈ES
i,j
Xw∈C
ucj ∼w
1 + Xucj ∈ES
i,j
Xw∈I i
S
ucj ∼w
1 + Xucj ∈ES
i,j
(6)
For any fixed i, j and S, any edge ucj ∈ ES
depends solely on i and j. It follows that
i,j the number of vertices w ∈ C such that ucj ∼ w
Xucj ∈ES
i,j
Xw∈C
ucj ∼w
1 = f (i, j) · xi
S
Consider a pair of vertices u, w in I i
for some function f , which is upper bounded by k (since C = k).
S and a vertex cj ∈ C such that vertices of u's and w's
type are adjacent to cj. Either the edge ucj flies over w or the edge wcj flies over u, but both
of these events never happen simulataneously. Therefore,
Xucj ∈ES
i,j
Xw∈I i
S
ucj ∼w
1 = (cid:18)xi
2 (cid:19) =
S
(xi
S)2
2 −
xi
S
2
For the last double sum in Equation 6 consider an edge ucj ∈ ES
In other words, this sum is a quadratic function of the variables with coefficients 1
Further, if vertices in IS are not adjacent to cj this sum is 0.
2 and − 1
2 .
S ′ such
i,j fly over all the vertices in I i′
that S′ 6= S or i′ 6= i. If ucj flies over v then all the edges in ES
S ′.
Let g(i, j, S, i′, S′) be a function that returns 1 if vertices in IS are adjacent to cj and all the
edges in ES
S ′. Otherwise g(i, j, S, i′, S′) returns 0. It follows that
S · X(i′,S ′)6=(i,S)
i,j fly over all the vertices in I i′
i,j and vertex v ∈ I i′
g(i, j, S, i′, S′)xi′
Xw∈I−I i
Xucj ∈ES
1 = xi
S ′ .
i,j
S
ucj ∼w
12
In other words, this sum is a quadratic function of the variables with 0 and 1 as coefficients.
The outer sum of Equation 6 goes over all 2k choices for S, k + 1 choices for i and k choices
for j. Since the sum of quadratic functions is a quadratic function, this concludes the analysis
and proves the following lemma.
Lemma 10. val(σ, G) is a quadratic function of the variables {xi
S} with half-integral coefficients
between −2kk2 and 2kk2. Furthermore, there is a a polynomial time algorithm that given G
computes the coefficients.
For each permutation c1, . . . , ck of C we can make an integer quadratic program for finding
the best super-homegenous solution to Optimal Linear Arrangement which places the
vertices of C in the order c1, . . . , ck from left to right. The quadratic program has variable
set {xi
S} and constraints as in Equations 3 and 4. The objective function is the one given
by Lemma 10, but with every coefficient multiplied by 2. This does not change the set of
optimal solutions and makes all the coefficients integral. This quadratic program has at most
2k·(k +1) variables, 2k·(k +2) constraints, and all coefficients are between −2k+1k2 and 2k+2k2.
Furthermore, since the domain of all variables is bounded the IQP is bounded as well. Thus we
can apply Theorem 1 to solve each such IQP in time f (k) · n. This proves the main result of
this section.
Theorem 3. Optimal Linear Arrangement parameterized by vertex cover is fixed param-
eter tractable.
4 Conclusions
We have shown that Integer Quadratic Programming is fixed parameter tractable when
parameterized by the number n of variables in the IQP and the maximum absolute value α
of the coefficients of the objective function and the constraints. We used the algorithm for
Integer Quadratic Programming to give the first FPT algorithm for Optimal Linear
Arrangement parameterized by the size of the smallest vertex cover of the input graph.
We hope that this work opens the gates for further research on the parameterized com-
plexity of non-linear and non-convex optimization problems. There are open problems abound.
For example, is Integer Quadratic Programming fixed parameter tractable when param-
eterized just by the number of variables? What about the parameterization by n + m, the
number of variables plus the number of constraints? It is also interesting to investigate the
parameterized complexity of Quadratic Programming, i.e. with real-valued variables rather
than integer variables. Finally, there is no reason to stop at quadratic functions. In particular,
investigating the parameterized complexity of special cases of (integer) mathematical program-
ming with degree-bounded polynomials in the objective function and constraints looks like an
exciting research direction. Of course, many of these problems are undecidable [17], but for the
questions that are decidable, parameterized complexity might well be the right framework to
study efficient algorithms.
References
[1] C. Ambuhl, M. Mastrolilli, and O. Svensson, Inapproximability results for maxi-
mum edge biclique, minimum linear arrangement, and sparsest cut, SIAM J. Comput., 40
(2011), pp. 567–596.
[2] M. Charikar, M. T. Hajiaghayi, H. J. Karloff, and S. Rao, l2
for vertex ordering problems, Algorithmica, 56 (2010), pp. 577–604.
2 spreading metrics
13
[3] J. Chen, I. A. Kanj, and G. Xia, Improved upper bounds for vertex cover, Theor.
Comput. Sci., 411 (2010), pp. 3736–3756.
[4] K. L. Clarkson, Las vegas algorithms for linear and integer programming when the di-
mension is small, J. ACM, 42 (1995), pp. 488–499.
[5] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh, Parameterized Algorithms, Springer, 2015.
[6] R. G. Downey and M. R. Fellows, Fundamentals of Parameterized Complexity, Texts
in Computer Science, Springer, 2013.
[7] U. Feige and J. R. Lee, An improved approximation ratio for the minimum linear
arrangement problem, Inf. Process. Lett., 101 (2007), pp. 26–29.
[8] M. R. Fellows, D. Hermelin, F. A. Rosamond, and H. Shachnai, Tractable param-
eterizations for the minimum linear arrangement problem, in Algorithms - ESA 2013 - 21st
Annual European Symposium, Sophia Antipolis, France, September 2-4, 2013. Proceedings,
2013, pp. 457–468.
[9] M. R. Fellows, D. Lokshtanov, N. Misra, F. A. Rosamond, and S. Saurabh,
Graph layout problems parameterized by vertex cover, in Algorithms and Computation,
19th International Symposium, ISAAC 2008, Gold Coast, Australia, December 15-17, 2008.
Proceedings, 2008, pp. 294–305.
[10] A. Frank and ´E. Tardos, An application of simultaneous diophantine approximation in
combinatorial optimization, Combinatorica, 7 (1987), pp. 49–65.
[11] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory
of NP-Completeness, Series of Books in the Mathematical Sciences, W. H. Freeman and
Co., 1979.
[12] S. Heinz, Complexity of integer quasiconvex polynomial optimization, J. Complexity, 21
(2005), pp. 543–556.
[13] R. Hemmecke, S. Onn, and L. Romanchuk, n-fold integer programming in cubic time,
Math. Program., 137 (2013), pp. 325–341.
[14] R. Kannan, Minkowski's convex body theorem and integer programming, Mathematics of
Operations Research, 12 (1987), pp. 415–440.
[15] L. Khachiyan and L. Porkolab, Integer optimization on convex semialgebraic sets,
Discrete & Computational Geometry, 23 (2000), pp. 207–224.
[16] L. G. Khachiyan, Polynomial algorithms in linear programming, USSR Comput. Math.
and Math. Phys., 20 (1980), pp. 53–72.
[17] M. Koppe, On the complexity of nonlinear mixed-integer optimization, in Mixed Integer
Nonlinear Programming, Springer, 2012, pp. 533–557.
[18] D. C. Lay, Linear Algebra and its Applications., Addison-Wesley, 2000.
[19] H. W. Lenstra Jr., Integer programming with a fixed number of variables, Math. of
Operations Research, 8 (1983), pp. 538–548.
[20] S. Onn, Nonlinear discrete optimization, Zurich Lectures in Advanced Mathematics, Eu-
ropean Mathematical Society, (2010).
14
[21] A. D. Pia, S. S. Dey, and M. Molinaro, Mixed-integer quadratic programming is in
NP, Math. Programming Series A, to appear. (2016).
[22] A. D. Pia and R. Weismantel, Integer quadratic programming in the plane, in Proceed-
ings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2014, Portland, Oregon, USA, January 5-7, 2014, 2014, pp. 840–846.
15
|
1707.09545 | 1 | 1707 | 2017-07-29T17:58:35 | Balanced Stable Marriage: How Close is Close Enough? | [
"cs.DS"
] | The Balanced Stable Marriage problem is a central optimization version of the classic Stable Marriage problem. Here, the output cannot be an arbitrary stable matching, but one that balances between the dissatisfaction of the two parties, men and women. We study Balanced Stable Marriage from the viewpoint of Parameterized Complexity. Our "above guarantee parameterizations" are arguably the most natural parameterizations of the problem at hand. Indeed, our parameterizations precisely fit the scenario where there exists a stable marriage that both parties would accept, that is, where the satisfaction of each party is "close" to the best it can hope for. Furthermore, our parameterizations accurately draw the line between tractability and intractability with respect to the target value. | cs.DS | cs |
Balanced Stable Marriage: How Close is Close Enough?∗
Sushmita Gupta†
Sanjukta Roy‡
Saket Saurabh† ‡
Meirav Zehavi†
Abstract
The Balanced Stable Marriage problem is a central optimization version of the
classic Stable Marriage problem. Here, the output cannot be an arbitrary stable match-
ing, but one that balances between the dissatisfaction of the two parties, men and women.
We study Balanced Stable Marriage from the viewpoint of Parameterized Complexity.
Our "above guarantee parameterizations" are arguably the most natural parameterizations
of the problem at hand. Indeed, our parameterizations precisely fit the scenario where there
exists a stable marriage that both parties would accept, that is, where the satisfaction of each
party is close to the best it can hope for. Furthermore, our parameterizations accurately
draw the line between tractability and intractability with respect to the target value.
1
Introduction
Matching under preferences is a rich topic central to both economics and computer science,
which has been consistently and intensively studied for over several decades. One of the main
reasons for interest in this topic stems from the observation that it is extremely relevant to
a wide variety of practical applications modeling situations where the objective is to match
agents to other agents (or to resources).
In the most general setting, a matching is defined
as an allocation (or assignment) of agents to resources that satisfies some predefined criterion
of compatibility/acceptability. Here, the (arguably) best known model is the two-sided model,
where the agents on one side are referred to as men, and the agents on the other side are referred
to as women. A few illustrative examples of real life situations where this model is employed
in practice include matching hospitals to residents, students to colleges, kidney patients to
donors and users to servers in a distributed Internet service. At the heart of all of these
applications lies the fundamental Stable Marriage problem. In particular, the Nobel Prize
in Economics was awarded to Shapley and Roth in 2012 "for the theory of stable allocations
and the practice of market design." Moreover, several books have been dedicated to the study
of Stable Marriage as well as optimization variants of this classical problem such as the
Egalitarian Stable Marriage, Sex-Equal Stable Marriage and Balanced Stable
Marriage problems [11, 14, 16].
The input of the Stable Marriage problem consists of a set of men, M , and a set of
women, W , each person ranking a subset of people of the opposite gender. That is, each person
a has a set of acceptable partners, A(a), whom this person subjectively ranks. Consequently,
each person a has a so-called preference list, where pa(b) denotes the position of b ∈ A(a) in a's
preference list. Without loss of generality, it is assumed that if a person a ranks a person b, then
the person b ranks the person a as well. The sets of preference lists of the men and the women
are denoted by LM and LW , respectively. In this context, we say that a pair of a man and a
woman, (m, a), is an acceptable pair if both m ∈ A(w) and w ∈ A(m). Accordingly, the notion
of a matching refers to a matching between men and women, where two people that are matched
∗The research leading to these results received funding from the European Research Council under the Euro-
pean Unions Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement no. 306992.
†University of Bergen, Norway; Emails: {Sushmita.Gupta, Meirav.Zehavi}@uib.no
‡Institute of Mathematical Sciences, India; Emails: {sanjukta, saket}@imsc.res.in
to one another form an acceptable pair. Roughly speaking, the goal of the Stable Marriage
problem is to find a matching that is stable in the following sense: there should not exist two
people who prefer being matched to one another over their current "status". More precisely, a
matching µ is said to be stable if it does not have a blocking pair, which is an acceptable pair
(m, w) such that (i) either m is unmatched by µ or pm(w) < pm(µ(m)), and (ii) either w is
unmatched by µ or pw(m) < pw(µ(w)). Here, the notation µ(a) indicates the person to whom
µ matches the person a. Note that a person always prefers being matched to an acceptable
partner over being unmatched.
The seminal paper [7] by Gale and Shapely on stable matchings shows that given an instance
of Stable Marriage, a stable matching necessarily exists, but it is not necessarily unique. In
fact, for a given instance of Stable Marriage, there can be an exponential number of stable
matchings, and they should be viewed as a spectrum where the two extremes are known as
the man-optimal stable matching and the woman-optimal stable matching. Formally, the man-
optimal stable matching, denoted by µM , is a stable matching such that every stable matching
µ satisfies the following condition: every man m is either unmatched by both µM and µ or
pm(µM (m)) ≤ pm(µ(m)). The woman-optimal stable matching, denoted by µW , is defined
analogously. These two extremes, which give the best possible solution for one party at the
expense of the other party, always exist and can be computed in polynomial time [7]. Naturally,
it is desirable to analyze matchings that lie somewhere in the middle, being globally desirable,
fair towards both sides or desirable by both sides.
Each notion above of what constitutes a desirable stable matching leads to a natural, dif-
ferent optimization problem. The determination of which notion best describes an appropriate
outcome depends on the specific situation at hand. Here, the quantity pa(µ(a)) is viewed as the
"satisfaction" of a in a matching µ, where a smaller value signifies a greater amount of satisfac-
tion. Under this interpretation, the egalitarian stable matching attempts to be globally desirable
by minimizing e(µ) = P(m,w)∈µ(pm(µ(m)) + pw(µ(w))) over the set of all stable matchings,
which we denote by SM. The problem of finding an egalitarian stable matching, called Egali-
tarian Stable Marriage, is known be solvable in polynomial time due to Irving et al. [12].
Roughly speaking, this problem does not distinguish between men and women, and therefore it
does not fit scenarios where it is necessary to differentiate between the individual satisfaction
of each party. In such scenarios, the Sex-Equal Stable Marriage and Balanced Stable
Marriage problems come into play. Before we define each of these two problems, we would
like to remark that a survey of results related to Egalitarian Stable Marriage and Sex-
Equal Stable Marriage is outside the scope of this paper, and we refer interested readers
to the books [11, 14, 16]. Here, we consider these two problems only to understand the context
of Balanced Stable Marriage.
In the Sex-Equal Stable Marriage problem, the objective is to find a stable match-
ing that minimizes the absolute value of δ(µ) over SM, where δ(µ) = P(m,w)∈µ pm(µ(m))
−P(m,w)∈µ pw(µ(w)).
It is thus clear that Sex-Equal Stable Marriage seeks a stable
matching that is fair towards both sides by minimizing the difference between their individ-
ual amounts of satisfaction. Unlike the Egalitarian Stable Marriage, the Sex-Equal
Stable Marriage problem is known to be NP-hard [13]. On the other hand, in Balanced
Stable Marriage, the objective is to find a stable matching that minimizes balance(µ) =
max{P(m,w)∈µ pm(w),P(m,w)∈µ pw(m)} over SM. At first sight, this measure might seem con-
ceptually similar to the previous one, but in fact, the two measures are quite different. Indeed,
Balanced Stable Marriage does not attempt to find a stable matching that is fair, but one
that is desirable by both sides. In other words, Balanced Stable Marriage examines the
amount of dissatisfaction of each party individually, and attempts to minimize the worse one
among the two. This problem fits the common scenario in economics where each party is selfish
in the sense that it desires a matching where its own dissatisfaction is minimized, irrespective
of the dissatisfaction of the other party, and our goal is to find a matching desirable by both
2
parties by ensuring that each individual amount of dissatisfaction does not exceed some thresh-
old. In some situations, the minimization of balance(µ) may indirectly also minimize δ(µ), but
in other situations, this may not be the case. Indeed, McDermid [17] constructed a family of
instances where there does not exist any matching that is both a sex-equal stable matching and
a balanced stable matching (the construction is also available in the book [16]).
The Balanced Stable Marriage problem was introduced in the influential work of
Feder [6] on stable matchings. Feder [6] proved that this problem is NP-hard and that it
admits a 2-approximation algorithm. Later, it was shown that this problem also admits a
(2 − 1/ℓ)-approximation algorithm where ℓ is the maximum size of a set of acceptable partners
[16]. O'Malley [19] phrased the Balanced Stable Marriage problem in terms of constraint
programming. Recently, McDermid and Irving [18] expressed interest in the design of fast
exact exponential-time algorithms for Balanced Stable Marriage. In this paper, we study
this problem in the realm of fast exact exponential-time algorithms as defined by the field of
Parameterized Complexity (see Section 2). Recall that SM is the set of all stable matchings.
In this context, we would like to remark that Egalitarian Stable Roommates problem
is NP-complete [6]. Recently, Chen et al. [1] showed that it is FPT parameterized by the
egalitarian cost. McDermid and Irving [18] showed that Sex-Equal Stable Marriage is
NP-hard even if it is only necessary to decide whether the target ∆ = minµ∈SM δ(µ) is 0 or
not [18]. In particular, this means that Sex-Equal Stable Marriage is not only W[1]-hard
with respect to ∆, but it is even paraNP-hard with respect to this parameter.1 In the case
of Balanced Stable Marriage, however, fixed-parameter tractability with respect to the
target Bal = minµ∈SM balance(µ) trivially follows from the fact that this value is lower bounded
by max{M , W }.2
Our Contribution. We introduce two "above-guarantee parameterizations" of Balanced
Stable Marriage. To this end, we consider the minimum value OM of the total dissat-
isfaction of men that can be realized by a stable matching, and the minimum value OW
of the total dissatisfaction of women that can be realized by a stable matching. Formally,
OM = P(m,w)∈µM
pw(m), where µM and µW are the man-
optimal and woman-optimal stable matchings, respectively. An input integer k would indicate
that our objective is to decide whether Bal ≤ k. In our first parameterization, the parameter is
k − min{OM , OW }, and in the second one, it is k − max{OM , OW }. In other words, we would
like to answer the following questions (recall that Bal = minµ∈SM balance(µ)).
pm(w), and OW = P(m,w)∈µW
Above-Min Balanced Stable Marriage (Above-Min BSM)
Input: An instance (M, W, LM , LW ) of Balanced Stable Marriage, and a non-negative
integer k.
Question: Is Bal ≤ k?
Parameter: t = k − min{OM , OW }
Above-Max Balanced Stable Marriage (Above-Max BSM)
Input: An instance (M, W, LM , LW ) of Balanced Stable Marriage, and a non-negative
integer k.
Question: Is Bal ≤ k?
Parameter: t = k − max{OM , OW }
Before stating our results, let us explain the choice of these parameterizations. Here, note
that the best satisfaction the party of men can hope for is OM , and the best satisfaction the
party of women can hope for is OW . First, consider the parameter t = k − min{OM , OW }.
1If a parameterized problem cannot be solved in polynomial time even when the value of the parameter is a
fixed constant (that is, independent of the input), then the problem is said to be paraNP-hard.
2We assume that any stable matching µ is perfect. We justify this assumption later.
3
Whenever we have a solution such that the amounts of satisfaction of both parties are close
enough to the best they can hope for, this parameter is small. Indeed, the closer the satisfaction
of both parties to the best they can hope for (which is exactly the case where both parties would
find the solution desirable), the smaller the parameter is, and the smaller the parameter is, the
faster a parameterized algorithm is. In other words, if there exists a solution that is desirable
by both parties, our parameter is small.
In the parameterization above, as we take the min of {OM , OW }, we need the satisfaction of
both parties to be close to optimal in order to have a small parameter. As we are able to show that
Balanced Stable Marriage is FPT with respect to this parameter, it is very natural to next
examine the case where we take the max of {OM , OW }. In this case, the closer the satisfaction of
at least one party to the best it can hope for, the smaller the parameter is. In other words, now
the demand from a solution-while not changing the definition of a solution in any way, that is,
a solution is still a stable matching µ minimizing balance(µ)-in order to have a small parameter
is weaker. In the jargon of Parameterized Complexity, it is said that the parameterization by
t = k − max{OM , OW } is "above a higher guarantee" than the parameterization by t = k −
min{OM , OW }, since it is always the case that max{OM , OW } ≥ min{OM , OW }. Unfortunately,
as we show in this paper, the parameterization by k − max{OM , OW } results in a problem that
is W[1]-hard. Hence, the complexities of the two parameterizations behave very differently. We
remark that in Parameterized Complexity, it is not at all the rule that when one takes an "above
a higher guarantee" parameterization, the problem would suddenly become W[1]-hard, as can
be evidenced by the most classical above guarantee parameterizations in this field, which are
of the Vertex Cover problem. For that problem, three above guarantee parameterizations
were considered in [3, 10, 15, 20], each above a higher guarantee than the previous one that was
studied, and each led to a problem that is FPT. In that context, unlike our case, it is still
not clear whether the bar can be raised higher. Overall, our results accurately draw the line
between tractability and intractability with respect to the target value in the context of two
very natural, useful parameterizations.
Finally, to be more precise, we note that our work proves three main theorems:
• First, we prove (in Section 3) that Above-Min BSM admits a kernel where the number of
people is linear in t. For this purpose, we introduce notions that might be of independent
interest in the context of a "functional" variant of Above-Min BSM. Our kernelization
algorithm consists of several phases, each simplifying a different aspect of Above-Min
BSM, and shedding light on structural properties of the Yes-instances of this problem.
Note that this result already implies that Above-Min BSM is FPT.
• Second, we prove (in Section 4) that Above-Min BSM admits a parameterized algorithm
whose running time is single exponential in the parameter t. This algorithm first builds
upon our kernel, and then incorporates the method of bounded search trees.
• Third, we prove (in Section 5) that Above-Max BSM is W[1]-hard. This reduction
is quite technical, and its importance lies in the fact that it rules out (under plausible
complexity-theoretic assumptions) the existence of a parameterized algorithm for Above-
Max BSM. Thus, we show that although Above-Max BSM seems quite similar to
Above-Min BSM, in the realm of Parameterized Complexity, these two problems are
completely different.
2 Preliminaries
Let f be a function f : A → B. For a subset A′ ⊆ A, we let f A′ denote the restriction of f to
A′. That is, f A′ : A′ → B, and f A′(a) = f (a) for all a ∈ A′.
Throughout the paper, whenever the instance I of Balanced Stable Marriage under
discussion is not clear from context or we would like to put emphasis on I, we add "(I)" to the
4
appropriate notation. For example, we use the notation t(I) rather than t. When we would
like to refer to the balance of a stable matching µ in a specific instance I, we would use the
notation balanceI(µ). A matching is called a perfect matching if it matches every person (to
some other person).
While designing our kernelization algorithm (see "Parameterized Complexity"), we might
be able to determine whether the input instance is a Yes-instance or a No-instance. For the
sake of clarity, in the first case, we simply return Yes, and in second case, we simply return
No. To properly comply with the definition of a kernel, the return of Yes and No should be
interpreted as the return of a trivial Yes-instance and a trivial No-instance, respectively. Here,
a trivial Yes-instance can be the one in which M = W = ∅ and k = 0, where the only stable
matching is the one that is empty and whose balance is 0, and a trivial No-instance can be the
one where M = {m}, W = {w}, A(m) = {w}, A(w) = {m} and k = 0.
Parameterized Complexity. A parameterization of a problem is the association of an integer
k with each input instance, which results in a parameterized problem. For our purposes, we need
to recall three central notions that define the parameterized complexity of a parameterized
problem. The first one is the notion of a kernel. Here, a parameterized problem is said to admit
a kernel of size f (k) for some function f that depends only on k if there exists a polynomial-
time algorithm, called a kernelization algorithm, that translates any input instance into an
equivalent instance of the same problem whose size is bounded by f (k) and such that the value
of the parameter does not increase. In case the function f is polynomial in k, the problem is
said to admit a polynomial kernel. Hence, kernelization is a mathematical concept that aims to
analyze the power of preprocessing procedures in a formal, rigorous manner. The second notion
that we use is the one of fixed-parameter tractability ( FPT). Here, a parameterized problem Π
is said to be FPT if there is an algorithm that solves it in time f (k) · IO(1), where I is the
size of the input and f is a function that depends only on k. Such an algorithm is called a
parameterized algorithm. In other words, the notion of FPT signifies that it is not necessary
for the combinatorial explosion in the running time of an algorithm for Π to depend on the
input size, but it can be confined to the parameter k. Finally, we recall that Parameterized
Complexity also provides tools to refute the existence of polynomial kernels and parameterized
algorithms for certain problems (under plausible complexity-theoretic assumptions), in which
context the notion of W[1]-hard is a central one. It is widely believed that a problem that is
W[1]-hard is unlikely to be FPT, and we refer the reader to the books [2,5] for more information
on this notion in particular, and on Parameterized Complexity in general. The notation O∗ is
used to hide factors polynomial in the input size.
Reduction Rule. To design our kernelization algorithm, we rely on the notion of a reduction
rule. A reduction rule is a polynomial-time procedure that replaces an instance (I, k) of a
parameterized problem Π by a new instance (I ′, k′) of Π. Roughly speaking, a reduction rule
is useful when the instance I ′ is in some sense "simpler" than the instance I. In particular, it
is desirable to ensure that k′ ≤ k. The rule is said to be safe if (I, k) is a Yes-instance if and
only if (I ′, k′) is a Yes-instance.
A Functional Variant of Stable Marriage. To obtain our kernelization algorithm for
Above-Min BSM, it will be convenient to work with a "functional" definition of preferences,
resulting in a "functional" variant of this problem which we call Above-Min FBSM. Here,
instead of the sets of preferences lists LM and LW , the input consists of sets of preference
functions FM and FW , where FM replaces LM and FW replaces LW . Specifically, every person
a ∈ M ∪ W has an injective (one-to-one) function fa : A(a) → N, called a preference function.
Intuitively, a lower function value corresponds to a higher preference. Since every preference
function is injective, it defines a total order over a set of acceptable partners. Note that all of
the definitions presented in the introduction extend to our functional variant in the natural way.
5
For the sake of formality, we specify the required adaptations below.
The input of the Functional Stable Marriage problem consists of a set of men, M ,
and a set of women, W . Each person a has a set of acceptable partners, denoted by A(a), and
an injective function fa : A(a) → N called a preference function. Without loss of generality,
it is assumed that if a person a belongs to the set of acceptable partners of a person b, then
the person b belongs to the set of acceptable partners of the person a. The set of preference
functions of the men and the women are denoted by FM and FW , respectively. A pair of a
man and a woman, (m, a), is an acceptable pair if both m ∈ A(w) and w ∈ A(m). Accordingly,
the notion of a matching refers to a matching between men and women, where two people that
are matched to one another form an acceptable pair. A matching µ stable if it does not have a
blocking pair, which is an acceptable pair (m, w) such that (i) either m is unmatched by µ or
fm(w) < fm(µ(m)), and (ii) either w is unmatched by µ or fw(m) < fw(µ(w)). The goal of
the Functional Stable Marriage problem is to find a stable matching.
The man-optimal stable matching, denoted by µM , is a stable matching such that every
stable matching µ satisfies the following condition: every man m is either unmatched by both
µM and µ or fm(µM (m)) ≤ fm(µ(m)). The woman-optimal stable matching, denoted by µW ,
is defined analogously. Given a stable matching µ, define
balance(µ) = max{ X
fm(w), X
fw(m)}.
(m,w)∈µ
(m,w)∈µ
Moreover, Bal = minµ∈SM balance(µ), where SM is the set of all stable matchings, OM =
P(m,w)∈µM
fw(m). Finally, Above-Min FBSM is defined as
fm(w), and OW = P(m,w)∈µW
follows.
Above-Min Functional Balanced Stable Marriage (Above-Min FBSM)
Input: An instance (M, W, FM , FW ) of Functional Balanced Stable Marriage, and
a non-negative integer k.
Question: Is Bal ≤ k?
Parameter: t = k − min{OM , OW }
From the above discussions it is straightforward to turn an instance of Above-Min BSM
into an equivalent instance of Above-Min FBSM as stated in the following observation.
Observation 1. Let I = (M, W, LM , LW , k) be an instance of Above-Min BSM. For each
a ∈ M ∪ W , define fa : A(a) → N by setting fa(b) = pa(b) for all b ∈ A(a). Then, I is a
Yes-instance of Above-Min BSM if and only if (M, W, FM = {fm}m∈M , FW = {fw}w∈W , k)
is a Yes-instance of Above-Min FBSM.
Known Results. Finally, we state several classical results, which were originally presented in
the context of Stable Marriage. By their original proofs, these results also hold in the context
of Functional Stable Marriage. To be more precise, given an instance of Functional
Stable Marriage, we can construct an equivalent instance of Stable Marriage, by ranking
the acceptable partners in the order of their function values, where a smaller value implies a
higher preference. The instances are equivalent in the sense that they give rise to the exact
same set of stable matchings. Hence, all the structural results about stable matchings in the
usual setting (modeled by strict preference lists) apply to the generalized setting, modeled by
injective functions.
Proposition 1 ( [8]). For any instance of Stable Marriage (or Functional Stable Mar-
riage), there exist a man-optimal stable matching, µM , and a woman-optimal stable matching,
µW , and both µM and µW can be computed in polynomial time.
The following powerful proposition is known as the Rural-Hospital Theorem.
6
Proposition 2 ( [9]). Given an instance of Stable Marriage (or Above-Min FBSM), the
set of men and women that are matched is the same for all stable matchings.
We further need a proposition regarding the man-optimal and woman-optimal stable match-
ings that implies Proposition 2 [9].
Proposition 3 ( [11]). For any instance of Stable Marriage (or Functional Stable
Marriage), every stable matching µ satisfies the following conditions: every woman w is
either unmatched by both µM and µ or pw(µM (w)) ≥ pw(µ(w)), and every man m is either
unmatched by both µW and µ or pm(µW (m)) ≥ pm(µ(m)).
3 Kernel
In this section, we design a kernelization algorithm for Above-Min BSM. More precisely, we
prove the following theorem.
Theorem 3.1. Above-Min BSM admits a kernel that has at most 3t men, at most 3t women,
and such that each person has at most 2t + 1 acceptable partners.
3.1 Functional Balanced Stable Marriage
To prove Theorem 3.1, we first prove the following result for the Above-Min FBSM problem.
Lemma 3.1. Above-Min FBSM admits a kernel with at most 2t men, at most 2t women,
and such that the image of the preference function of each person is a subset of {1, 2, . . . , t + 1}.
To obtain the desired kernelization algorithm, we execute the following plan.
1. Cleaning Prefixes and Suffixes. Simplify the preference functions by "cleaning" suf-
fixes and thereby also "cleaning" prefixes.
2. Perfect Matching. Zoom into the set of people matched by every stable matching.
3. Overcoming Sadness. Bound the number of "sad" people. Roughly speaking, a "sad"
person a is one whose best attainable partner, b, does not reciprocate by considering a as
the best attainable partner.
4. Marrying Happy People. Remove "happy" people from the instance.
5. Truncating High-Values. Obtain "compact" preference functions by truncating "high-
values".
6. Shrinking Gaps. Shrink some of the gaps created by previous steps.
Each of the following subsections captures one of the steps above. In what follows, we let I
denote our current instance of Above-Min FBSM. Initially, this instance is the input instance,
but as the execution of our algorithm progresses, the instance is modified. The reduction rules
that we present are applied exhaustively in the order of their presentation.
In other words,
at each point of time, the first rule whose condition is true is the one that we apply next.
In particular, the execution terminates once the value of t drops below 0, as implied by the
following rule.
Reduction Rule 1. If k < max{OM , OW }, then return No.
Lemma 3.3. Reduction Rule 1 is safe.
Proof. For every µ ∈ SM, it holds that balance(µ) ≥ max{OM , OW }. Thus, if k < max{OM , OW },
then every µ ∈ SM satisfies balance(µ) > k. In this case, we conclude that Bal > k, and therefore
I is a No-instance.
Note that if k < 0, then it also holds that t < 0, and that if t < 0, then k < min{OM , OW }.
We remark that by Proposition 1, it would be clear that each of our reduction rules can indeed
be implemented in polynomial time.
7
Cleaning Prefixes and Suffixes. We begin by modifying the images of the preference func-
tions. We remark that it is necessary to perform this step first as otherwise the following steps
would not be correct. To clean prefixes while ensuring both safeness and that the parameter t
does not increase, we would actually need to clean suffixes first. Formally, we define suffixes as
follows.
Definition 3.1. Let (m, w) denote an acceptable pair. If m is matched by µW and fm(w) >
fm(µW (m)), then we say that w belongs to the suffix of m. Similarly, if w is matched by µM
and fw(m) > fw(µM (w)), then we say that m belongs to the suffix of w.
By Proposition 3, we have the following observation.
Observation 2. Let (m, w) denote an acceptable pair such that one of its members belongs to
the suffix of the other member. Then, there is no µ ∈ SM(I) that matches m with w.
For every person a, let worst(a) be the person in A(a) to whom fa assigns its worst preference
value. More precisely, worst(a) = argmaxb∈A(a)fa(b). We will now clean suffixes.
Reduction Rule 2. If there exists a person a such that worst(a) belongs to the suffix of a, then
define the preference functions as follows.
• f ′
a = faA(a)\{worst(a)}.
• f ′
worst(a) = fworst(a)A(worst(a))\{a}.
• For all b ∈ M ∪ W \ {a, worst(a)}: f ′
The new instance is J = (M, W, {f ′
b = fb.
m′ }m′∈M , {f ′
w′}w′∈W , k).
Lemma 3.4. Reduction Rule 2 is safe, and t(I) = t(J ).
Proof. By the definition of the new preference functions, we have that for every µ ∈ SM(I) ∩
SM(J ), it holds that P(m,w)∈µ fm(w) = P(m,w)∈µ f ′
In particular, this means that to conclude that Bal(I) = Bal(J ) (which implies safeness) as well
as that OM (I) = OM (J ) and OW (I) = OW (J ) (which implies that t(I) = t(J )), it is suf-
ficient to show that SM(I) = SM(J ). For this purpose, first consider some µ ∈ SM(I). By
Observation 2, it holds that (a, worst(a)) /∈ µ. Hence, µ is a matching in J . Moreover, if µ has
a blocking pair in J , then by the definition of the new preference functions, it is also a blocking
pair in I. Since µ is stable in I, we have that µ ∈ SM(J ).
m(w) and P(m,w)∈µ fw(m) = P(m,w)∈µ f ′
w(m).
In the second direction, consider some µ ∈ SM(J ). Then, it is clear that µ is a matching
in I. Moreover, if µ has a blocking pair (m, w) in I that is not (a, worst(a)), then (m, w) is
an acceptable pair in J , and therefore by the definition of the new preference functions, we
have that (m, w) is also a blocking pair in J . Hence, since µ is stable in J , the only pair
that can block µ in I is (a, worst(a)). Thus, to show that µ ∈ SM(I), it remains to prove that
(a, worst(a)) cannot block µ in I. Suppose, by way of contradiction, that (a, worst(a)) blocks
µ in I. In particular, this means that fa(worst(a)) < fa(µ(a)). However, this contradicts the
definition of worst(a).
By cleaning suffixes, we actually also accomplish the objective of cleaning prefixes, which
are defined as follows.
Definition 3.2. Let (m, w) denote an acceptable pair. If m is matched by µM and fm(w) <
fm(µM (m)), then we say that w belongs to the prefix of m. Similarly, if w is matched by µW
and fw(m) < fw(µW (w)), then we say that m belongs to the prefix of w.
Let us now claim that we have indeed succeeded in cleaning prefixes.
8
Lemma 3.5. Let I be an instance of Above-Min FBSM on which Reduction Rules 1 to 2
have been exhaustively applied. Then, there does not exist an acceptable pair (m, w) such that
one of it members belongs to the prefix of the other one.
Proof. Suppose, by way of contradiction, that there exists an acceptable pair (m, w) such that
one of its members belongs to the prefix of the other one. Without loss of generality, assume
that w belongs to the prefix of m. Then, fm(w) < fm(µM (m)). Since µM is a stable matching,
it cannot be blocked by (m, w), which means that w is matched by µM and fw(µM (w)) <
fw(m). Thus, we have that m belongs to the suffix of w, which contradicts the assumption that
Reduction Rule 2 was applied exhaustively.
As a direct result of this lemma, we have the following corollary.
Corollary 3.1. Let I be an instance of Above-Min FBSM on which Reduction Rules 1 to
2 have been exhaustively applied. Then, for every acceptable pair (m, w) in I where m and w
are matched (not necessarily to each other) by both µM and µW , it holds that fm(µM (m)) ≤
fm(w) ≤ fm(µW (m)) and fw(µW (w)) ≤ fw(m) ≤ fw(µM (w)).
Perfect Matching. Having Corollary 3.1 at hand, we are able to provide a simple rule that
allows us to assume that every solution matches all people.
Reduction Rule 3. If there exists a person unmatched by µM , then let M ′ and W ′ denote
the subsets of men and women, respectively, who are matched by µM . For each a ∈ M ′ ∪
W ′, denote A′(a) = A(a) ∩ (M ′ ∪ W ′), and define f ′
v = fvA′(v). The new instance is J =
(M ′, W ′, {f ′
m}m∈M ′ , {f ′
w}w∈W ′, k).
To prove the safeness of this rule, we first prove the following lemma.
Lemma 3.6. Let I be an instance of Above-Min FBSM on which Reduction Rules 1 to 2 have
been exhaustively applied. Then, for every person a not matched by µM , it holds that A(a) = ∅.
Proof. Let a be a person not matched by µM . Then, by Proposition 2, it holds that a is not
matched by any stable matching. Hence, we can assume w.l.o.g. that a is man m. First, note
that A(m) cannot contain any woman w that is not matched by some stable matching, else
(m, w) would have formed a blocking pair for that stable matching. Second, we claim that
A(m) cannot contain a woman w that is matched by some stable matching. Suppose, by way
of contradiction, that this claim is false. Then, by Proposition 2, it holds that A(m) contains
a woman w that is matched by µM . We have that w prefers µM (w) over m, else (m, w) would
have formed a blocking pair for µM , which is impossible as µM is a stable matching. However,
this implies that m belongs to the suffix of w, which contradicts the supposition that Reduction
Rule 2 has been exhaustively applied. We thus conclude that A(a) = ∅.
Lemma 3.7. Reduction Rule 3 is safe, and t(I) = t(J ).
Proof. By the definition of the new preference functions, we have that for every µ ∈ SM(I) ∩
SM(J ), it holds that P(m,w)∈µ fm(w) = P(m,w)∈µ f ′
m(w) and P(m,w)∈µ fw(m) = P(m,w)∈µ f ′
w(m).
To conclude that the lemma is correct, it is thus sufficient to argue that SM(I) = SM(J ). For
this purpose, first consider some µ ∈ SM(I). By Proposition 2, we have that µ is also a match-
ing in J . Moreover, if µ has a blocking pair in J , then by the definition of the new preference
functions, it is also a blocking pair in I. Since µ is stable in I, we have that µ ∈ SM(J ).
In the second direction, consider some µ ∈ SM(J ). Then, it is clear that µ is a matching in
I. By Lemma 3.6, if µ has a blocking pair (m, w) in I, then both m ∈ M ′ and w ∈ W ′. However,
for such a blocking pair (m, w), we have that (m, w) is an acceptable pair in J , and therefore
by the definition of the new preference functions, we have that (m, w) is also a blocking pair in
J . Hence, since µ is stable in J , we conclude that µ is also stable in I.
9
By Proposition 2, from now onwards, we have that for the given instance, any stable match-
ing is a perfect matching. Due to this observation, we can denote n = M = W , and for any
stable matching µ, we have the following equalities.
X
fm(w) = X
(m,w)∈µ
m∈M
fm(µ(m));
X
fw(m) = X
(m,w)∈µ
w∈W
fw(µ(w)).
(I)
Overcoming Sadness. As every stable matching is a perfect matching, every person is
matched by every stable matching, including the man-optimal and woman-optimal stable match-
ings. Thus, it is well defined to classify the people who do not have the same partner in the
man-optimal and woman-optimal stable matchings as "sad". That is,
Definition 3.3. A person a ∈ M ∪ W is sad if µM (a) 6= µW (a).
We let MS and WS denote the sets of sad men and sad women, respectively. People who
are not sad are termed happy. Accordingly, we let MH and WH denote the sets of happy men
and happy women, respectively. Note that MS = ∅ if and only if WS = ∅. Moreover, note that
by the definition of µM and µW , for a happy person a it holds that a and µM (a) = µW (a) are
matched to one another by every stable matching. Let us now bound the number of sad people
in a Yes-instance.
Reduction Rule 4. If MS > 2t or WS > 2t, then return No.
Lemma 3.8. Reduction Rule 4 is safe.
Proof. We only prove that if MS > 2t, then I is a No-instance, as the proof of the other case
is symmetric to this one. Let us assume that MS > 2t. Suppose, by way of contradiction, that
I is a Yes-instance. Then, there exists a stable matching µ such that balance(µ) ≤ k. Partition
S ⊎ M ′′
MS = M ′
S to be the set of all m in MS such that m is not the partner
of µ(m) in µW .
S as follows. Set M ′
Accordingly, set M ′′
larger than t. Let us first handle the case where M ′
S = MS \M ′
M ′
S = {m ∈ MS fµ(m)(m) > fµ(m)(µW (µ(m)))}.
S. Since MS > 2t, at least one among M ′
S > t. Then,
S and M ′′
S is (strictly)
X
w∈W
{w µ(w)∈M ′
S }
fw(µ(w)) = X
≥ X
= X
fw(µ(w)) + X
[fw(µW (w)) + 1] + X
{w µ(w) /∈M ′
S}
fw(µ(w))
fw(µW (w))
{w µ(w)∈M ′
S }
fw(µW (w)) + X
S }
{w µ(w) /∈M ′
1 = OW + M ′
S
w∈W
{w µ(w)∈M ′
S }
> OW + t ≥ k.
Here, the first inequality followed directly from the definition of M ′
S > t. However, we now have that
S. As we have reached a
X
contradiction, it must hold that M ′′
fm(µ(m)) = X
≥ X
= X
m∈M
m∈M ′′
S
m∈M ′′
S
fm(µ(m))
m /∈M ′′
S
fm(µ(m)) + X
[fm(µM (m)) + 1] + X
fm(µM (m)) + X
m /∈M ′′
S
fm(µM (m))
1 = OM + M ′′
S
m∈M
{m µ(m)∈M ′′
S }
10
> OM + t ≥ k.
Here, the first inequality followed from the definition of M ′′
S , we have
that fµ(m)(m) ≤ fµ(m)(µW (µ(m))), else m would have belonged to M ′
S. However, in this case
we deduce that m = µW (µ(m)), and since m ∈ MS, we have that µ(m) 6= µM (m), which implies
that fm(µ(m)) ≥ fm(µM (m)) + 1. As we have again reached a contradiction, we conclude the
proof.
S . Indeed, for all m ∈ M ′′
Marrying Happy People. Towards the removal of happy people, we first need to handle the
special case where there are no sad people. In this case, there is exactly one stable matching,
which is the man-optimal stable matching (that is equal, in this case, to the woman-optimal
stable matching). This immediately implies the safeness of the following rule.
Reduction Rule 5. If MS = WS = ∅, then return Yes if balance(µM ) ≤ k and No otherwise.
Observation 3. Reduction Rule 5 is safe.
We now turn to discard happy people. When we perform this operation, we need to ensure
that the balance of the instance is preserved. More precisely, we need to ensure that Bal(I) =
Bal(J ), where J denotes the new instance resulting from the removal of some happy people.
Towards this, we let (mh, wh) denote a happy pair, which is simply a pair of a happy man and
happy woman who are matched to each other in every stable matching. Then, we redefine the
preference functions in a manner that allows us to transfer the "contributions" of mh and wh
from Bal(I) to Bal(J ) via some sad man and woman. We remark that these sad people exist
because Reduction Rule 5 does not apply. The details are as follows.
Reduction Rule 6. If there exists a happy pair (mh, wh), then proceed as follows. Select an
arbitrary sad man ms and an arbitrary sad woman ws. Denote M ′ = M \ {mh} and W ′ =
W \ {wh}. For each person a ∈ M ′ ∪ W ′, the new preference function f ′
a : A(a) \ {mh, wh} → N
is defined as follows.
• The preference function of ms. For each w ∈ A(ms) \ {wh}: f ′
ms(w) = fms(w) +
fmh(wh).
• The preference function of ws. For each m ∈ A(ws) \ {mh}: f ′
ws(m) = fws(m) +
fwh(mh).
• For each w ∈ W ′ \ {ws}: f ′
w = fwM ′.
• For each m ∈ M ′ \ {ms}: f ′
m = fmW ′.
The new instance is J = (M ′, W ′, {f ′
m}m∈M ′, {f ′
w}w∈W ′, k).
Let us first state a lemma concerning the proof of the forward direction of the safeness of
Reduction Rule 6.
Lemma 3.9. Let µ ∈ SM(I). Then, µ′ = µ \ {(mh, wh)} is a stable matching in J such that
balanceJ (µ′) = balanceI (µ).
Proof. We first show that µ′ ∈ SM(J ). By Reduction Rule 3, it holds that µ is a perfect
matching in I. Since (mh, wh) is a happy pair, it is clear that (mh, wh) ∈ µ, and therefore µ′
is a perfect matching in J . Let (m, w) /∈ µ′ be some acceptable pair in J . Since µ ∈ SM(I)
and it is a perfect matching, it holds that fm(w) > fm(µ(m)) or fw(m) > fw(µ(w)). Let us
consider these two possibilities separately.
• Suppose that fm(w) > fm(µ(m)). If m 6= ms, then f ′
m(µ′(m)). Else, f ′
fm(µ(m)), and therefore f ′
m(µ′(m)) = fm(µ(m)) + fmh(wh), and therefore again f ′
f ′
m(w) > f ′
m(w) = fm(w) and f ′
m(µ′(m)) =
m(w) = fm(w) + fmh(wh) and
m(w) > f ′
m(µ′(m)).
11
• Suppose that fw(m) > fw(µ(w)). Analogously to the previous case, we get that f ′
w(m) >
w(µ′(w)).
f ′
Since the choice of (m, w) was arbitrary, we conclude that µ′ does not have a blocking pair
in J , and therefore µ′ ∈ SM(J ). To show that balanceJ (µ′) = balanceI (µ), note that
balanceJ (µ′) = max{ X
m(µ′(m)), X
f ′
f ′
w(µ′(w))}
= max{f ′
m∈M \{mh}
ms(µ′(ms)) + X
w∈W \{wh}
f ′
m(µ′(m)),
m∈M \{mh,ms}
ws(µ′(ws)) + X
f ′
= max{fms(µ(ms)) + fmh(wh) + X
w∈W \{wh,ws}
f ′
w(µ′(w))}
fm(µ(m)),
fws(µ(ws)) + fwh(mh) + X
m∈M \{mh,ms}
fw(µ(w))}
= max{ X
fm(µ(m)), X
m∈M
w∈W
w∈W \{wh,ws}
fw(µ(w))} = balanceI (µ).
This concludes the proof.
To prove a lemma addressing the reverse direction, we first state the following observation,
whose correctness follows directly from Corollary 3.1.
Observation 4. Let I be an instance of Above-Min FBSM on which Reduction Rules 1 to 2
have been exhaustively applied. Then, for every happy pair (mh, wh), it holds that A(mh) = {wh}
and A(wh) = {mh}.
Lemma 3.10. Let µ′ ∈ SM(J ). Then, µ = µ′ ∪ {(mh, wh)} is a stable matching in I such that
balanceI(µ) = balanceJ (µ′).
Proof. We first show that µ ∈ SM(I). By Reduction Rule 3, it holds that µ′ is a perfect
matching in J . Since (mh, wh) is a happy pair and by Observation 4, we have that µ is a
perfect matching in I such that neither mh nor wh participate in any pair that blocks µ (if
such a pair exists). Let (m, w) /∈ µ′ be some acceptable pair in I such that m 6= mh and
w 6= wh. Since µ′ ∈ SM(J ) and it is a perfect matching, it holds that f ′
m(µ′(m)) or
f ′
w(m) > f ′
w(µ′(w)). Let us consider these two possibilities separately.
m(w) > f ′
• Suppose that f ′
m(w) > f ′
m(µ(m)).
If m 6= ms, then fm(w) = f ′
m(µ′(m)), and therefore fm(w) > fm(µ(m)). Else, fm(w) = f ′
f ′
fm(µ(m)) = f ′
m(µ′(m)) − fmh(wh), and therefore again fm(w) > fm(µ(m)).
m(w) and fm(µ(m)) =
m(w) − fmh(wh) and
• Suppose that f ′
w(m) > f ′
w(µ′(w)). Analogously to the previous case, we get that fw(m) >
fw(µ(w)).
Since the choice of (m, w) was arbitrary, we conclude that µ does not have a blocking pair
in I, and therefore µ ∈ SM(I). To show that balanceI(µ) = balanceJ (µ′), we follow the exact
same argument as the one present in the proof of Lemma 3.9.
We now turn to justify the use of Reduction Rule 6.
Lemma 3.11. Reduction Rule 6 is safe, and t(I) = t(J ).
12
Proof. By Lemmata 3.9 and 3.10, it holds that Bal(I) = Bal(J ), and that both balanceI(µM (I)) =
balanceJ (µM (J )) and balanceI (µW (I)) = balanceJ (µW (J )). As the argument k remained un-
touched, we have that Reduction Rule 6 is safe as well as that t(I) = t(J ).
Before we proceed to examine preference functions more closely, let us take a step back and
prove the following result.
Lemma 3.12. Given an instance I of Above-Min FBSM, one can exhaustively apply Reduc-
tion Rules 1 to 6 in polynomial time to obtain an instance J such that t(J ) ≤ t(I). All people
in J are sad and matched by every stable matching, and there exist at most 2t men and at most
2t women.
Proof. Notice that each rule among Reduction Rules 1 to 6 can be applied in polynomial time,
and it either terminates the execution of the algorithm or shrinks the size of the instance. Hence,
it is clear that the instance J is obtained in polynomial time, and the claim that t(J ) ≤ t(I)
follows from Lemmata 3.7 and 3.11. Due to Reduction Rules 3 and 6, we have that all people
in J are sad and matched by every stable matching. Thus, due to Reduction Rule 4, we also
have that there exist at most 2t men and at most 2t women.
Truncating High-Values. Up until now, we have bounded the number of people. However,
the images of the preference functions can contain integers that are not bounded by a function
polynomial in the parameter. Thus, even though the number of people is upper bounded by 4t,
the total size of the instance can be huge. Accordingly, in what follows, we need to process the
images of the preference functions. Recall that we have already modified preference functions
so that they would not contain irrelevant information in their prefixes and suffixes. Our current
goal is to truncate "high-values"' of preference functions. To understand the intuition behind
the rule we present next, suppose that there exists a stable matching µ and a man m such
that fm(µ(m)) > t + fm(µM (m)). That is, µ matches m to a woman whose position is larger
than the position of the woman with whom m is matched by µM by at least t units. Then,
balance(µ) ≥ Pm′∈M fm′(µ(m′)) > OM + t ≥ k. Hence, irrespective of whether or not the
current instance is a Yes-instance, we know that µ is not a yes-certificate. We thus observe
that we should delete all those acceptable pairs whose presence in any stable matching prevents
its balance from being upper bounded by k. Formally,
Reduction Rule 7. If there exists an acceptable pair (m, w) such that fm(w) > (k − OM ) +
fm(µM (m)) or fw(m) > (k − OW ) + fw(µW (w)), then define the preference functions as follows:
• f ′
m = fmA(m)\{w}.
• f ′
w = fwA(w)\{m}.
• For all a ∈ M ∪ W \ {m, w}: f ′
a = fa.
The new instance is J = (M, W, {f ′
m′ }m′∈M , {f ′
w′}w′∈W , k).
Lemma 3.13. Reduction Rule 7 is safe, and t(I) ≥ t(J ).
Proof. Without loss of generality, suppose that fm(w) > (k − OM (I)) + fm(µM (m)). Due to
Reduction Rule 1, we have that k − OM (I) ≥ 0, and therefore fm(w) > fm(µM (m)), which
implies that w 6= µM (m). We thus have that µM (I) is also a matching in J , and due to
Corollary 3.1, we deduce that µM (I) = µM (J ). First, we would like to show that t(I) ≥ t(J ).
For this purpose, it is sufficient to show that OM (I) ≤ OM (J ) and OW (I) ≤ OW (J ). Since
µM (I) = µM (J ), it is clear that OM (I) = OM (J ). By Reduction Rule 3, µM (I) matches all
people in I. Thus, by Proposition 2 and since µM (I) = µM (J ), we have that µW (J ) matches
all people in J , and hence all people in I. By Corollary 3.1, for any woman w and the man m
13
most preferred by w in J , it holds that w does not prefer m over µW (w) in I. Thus, by the
definition of the preference functions, we have that OW (I) ≤ OW (J ).
To show that the rule is safe, we need to show that Bal(I) ≤ k if and only if Bal(J ) ≤ k.
For this purpose, let us first suppose that Bal(I) ≤ k. Then, there exists µ ∈ SM(I) such that
balanceI(µ) ≤ k. Notice that if µ(m) = w, then since fm(w) > (k − OM ) + fm(µM (m)) and by
the equations in "Perfect Matching", we have that
k ≥ balanceI(µ)
fm′(µ(m′))
≥ X
≥ (k − OM (I)) + X
m′∈M
fm′(µM (m′)) > (k − OM (I)) + OM (I),
m′∈M
which is a contradiction. Hence, (m, w) /∈ µ. Therefore, µ is a matching in J . By the definition
of the new preference functions, if µ has a blocking pair in J , then this pair also blocks µ in
I. Since µ is stable in I, we have that µ is also stable in J . Now, by our definition of the new
preference functions, we have that balanceI (µ) = balanceJ (µ). Since balanceI (µ) ≤ k, we thus
conclude that Bal(J ) ≤ k.
In the second direction, suppose that Bal(J ) ≤ k. Then, there exists µ ∈ SM(J ) such
that balanceJ (µ) ≤ k. Clearly, µ is also a matching in I. Moreover, by the definition of the
preference functions, every acceptable pair in I that is also present in J cannot block µ in I,
else it would have also blocked µ in J . Thus, if µ has a blocking pair in I, then this pair must be
(m, w). We claim that (m, w) cannot block µ in I, which would imply that µ ∈ SM(I). Suppose,
by way of contradiction, that this claim is not true. Recall that we have already proved that
µM (I) = µM (J ). Let us denote µM = µM (I). We have that µ matches m, which implies that
fm(µ(m)) > fm(w). Since fm(w) > (k−OM (I))+fm(µM (m)), we deduce that fm(µ(m)) > (k−
OM (I)) + fm(µM (m)). Furthermore, since f ′
m(µM (m)) = fm(µM (m))
and OM (I) = OM (J ), we get that f ′
m(µM (m)). However, we then
have that
m(µ(m)) > (k − OM (J )) + f ′
m(µ(m)) = fm(µ(m)), f ′
k ≥ balanceJ (µ)
f ′
m′(µ(m′))
≥ X
≥ (k − OM (J )) + X
m′∈M
m′(µM (m′)) > (k − OM (J )) + OM (J ),
f ′
m′∈M
which is a contradiction. Therefore, µ ∈ SM(I). The definition of the new preference functions
imply that balanceI (µ) = balanceJ (µ). Since balanceJ (µ) ≤ k, we conclude that Bal(I) ≤ k.
Shrinking Gaps. Currently, there might still exist a man m or a woman w such that
fm(µM (m)) > 1 or fw(µW (w)) > 1, respectively.
In the following rule, we would like to
decrease some values assigned by the preference functions of such men and women in a manner
that preserves equivalence.
Reduction Rule 8. If there exist m ∈ M and w ∈ W such that fm(µM (m)) > 1 and
fw(µW (w)) > 1, then define the preference functions as follows.
• The preference function of m. For all w′ ∈ A(m): f ′
m(w′) = fm(w′) − 1.
• The preference function of w. For all m′ ∈ A(w): f ′
w(m′) = fw(m′) − 1.
• For all a ∈ M ∪ W \ {m, w}: f ′
a = fa.
The new instance is J = (M, W, {f ′
m′ }m′∈M , {f ′
w′}w′∈W , k − 1).
14
Lemma 3.14. Reduction Rule 8 is safe, and t(I) = t(J ).
Proof. Let us first observe that the set of acceptable pairs in I is equal to the set of acceptable
pairs of J . Furthermore, for every person a, including the cases where this person is either m or
w, any two acceptable partners b and b′ of a that satisfy fa(b) < fa(b′) also satisfy f ′
a(b′),
and vice versa. Indeed, this observation follows directly from our definition of the new preference
functions. In other words, if a person prefers some person over another in I, then this person
also has the same preference order in J , and vice versa. We thus deduce that SM(I) = SM(J ).
We proceed by claiming that for all µ ∈ SM(I), we have that balanceI(µ) = balanceJ (µ) +
Indeed, by the definition of the new preference functions and the equations in "Perfect
a(b) < f ′
1.
Matching", we have that
balanceI (µ) = max{ X
fm′(µ(m′)), X
m′∈M
= max{fm(µ(m)) + X
w′∈W
fw′(µ(w′))}
fm′(µ(m′)), fw(µ(w)) + X
fw′(µ(w′))}
m′∈M \{m}
= max{f ′
= max{ X
m(µ(m)) + 1 + X
m′(µ(m′)), X
f ′
m′(µ(m′)), f ′
f ′
w′∈W \{w}
w(µ(w)) + 1 + X
w′(µ(w′))}
f ′
m′∈M \{m}
w′(µ(w′))} + 1 = balanceJ (µ) + 1.
f ′
w′∈W \{w}
m′∈M
w′∈W
Furthermore, the arguments above also show that OM (I) = OM (J ) + 1 and OW (I) =
OW (J ) + 1. Hence, we have that Bal(I) = Bal(J ) + 1. Since k was decreased by 1, we conclude
that the rule is safe and that t(I) = t(J ).
The usefulness of Reduction Rule 8 lies in the observation that after the exhaustive appli-
cation of this rule, at least one of the two parties does not have any member without a person
assigned 1 by his/her preference function. More precisely,
Observation 5. Let I be an instance of Above-Min FBSM that is reduced with respect to
Reduction Rules 1 to 8. Then, either (i) for every m ∈ M , we have that fm(µM (m)) = 1, or
(ii) for every w ∈ W , we have that fw(µW (w)) = 1. In particular, either (i) OM = M or (ii)
OW = W .
This concludes the description of our reduction rules. We are now ready to prove Lemma 3.1.
Proof of Lemma 3.1. Given an instance I of Above-Min FBSM, our kernelization algorithm
exhaustively applies Reduction Rules 1 to 8, after which it outputs the resulting instance, J , as
the kernel. Notice that each rule among Reduction Rules 1 to 8 can be applied in polynomial
time, and it either terminates the execution of the algorithm or shrinks the size of the instance.
Hence, it is clear that the instance J is obtained in polynomial time. The claims that J and
I are equivalent and that t(J ) ≤ t(I) follow directly from the lemmata that prove the safeness
of each rule as well as argue with respect to the parameter. By Lemma 3.12, we also have that
the instance contains at most 2t (sad) men and at most 2t (sad) women.
It remains to show that the image of the preference function of each person is a subset
of {1, 2, . . . , t + 1}. By Reduction Rule 7, every acceptable pair (m, w) satisfies fm(w) ≤
(k − OM ) + fm(µM (m)) and fw(m) ≤ (k − OW ) + fw(µW (w)). Moreover, for every acceptable
pair (m, w), it holds that fm(µM (m)) ≤ OM −(M −1) and fw(µW (w)) ≤ OW −(W −1). Thus,
every acceptable pair (m, w) satisfies fm(w) ≤ k − (M − 1) and fw(m) ≤ k − (W − 1). By
Reduction Rule 3 and Observation 5, we have that t = k − min{OM , OW } = k − M = k − W .
Hence, we further conclude that every acceptable pair (m, w) satisfies fm(w) ≤ t + 1 and
fw(m) ≤ t + 1.
15
3.2 Balanced Stable Marriage
Having proved Lemma 3.1, we have a kernel for Above-Min FBSM. We would like to employ
this kernelization algorithm to design one for Above-Min BSM. For this purpose, we need to
remove gaps from preference functions. Once we do this, we can view preference functions as
preference lists and obtain the desired kernel. In what follows, we describe our kernelization
algorithm for Above-Min BSM.
Let K = (M ′, W ′, LM ′, LW ′, k′) be the input instance, which is an instance of Above-Min
BSM. Our algorithm begins by applying the reduction given by Observation 1 to translate
K into an instance I ′ = (M ′, W ′, FM ′, FW ′, k′) of Above-Min FBSM. Then, our algorithm
applies the kernelization algorithm given by Lemma 3.1 to I ′, obtaining a reduced instance
I = (M, W, FM , FW , k) of Above-Min FBSM. By Lemma 3.1, this instance has at most 2t
men, at most 2t women, and the image of the preference function of each person is a subset
of {1, 2, . . . , t + 1}. To eliminate "gaps" in the preference functions, the algorithm proceeds as
described below. Note that we no longer apply any reduction rule from Section 3.1 (even if
its condition is satisfied), as we currently give a new kernelization procedure rather than an
extension of the previous one. Let us first formally define the notion of a gap.
Definition 3.4. Let a ∈ M ∪ W , and i be positive integer outside the image of f . If there exists
an integer j > i that belongs to the image of f , then fa is said to have a gap at i.
Inserting Dummies. We have ensured that the largest number in the image of any preference
function is at most t + 1. As every person is sad, it has at least two acceptable partners, and
hence it has at most t − 1 ≤ t gaps. To handle the gaps of all people, we create a set of t dummy
men and t dummy women. Our objective is to introduce these dummy people as acceptable
partners for people who have gaps in their preference functions, such that the function values of
the dummy people would fill the gaps. In the context of the following rule, note that currently
there are no happy people, and hence this rule would be applied (only once).
Reduction Rule 9. If there do not exist happy people, then let X = {x1, x2, . . . , xt} denote a
set of new (dummy) men, and Y = {y1, y2, . . . , yt} denote a set of new (dummy) women. For
each i ∈ {1, 2, . . . , t}, initialize A(xi) = {yi}, A(yi) = {xi} and fxi(yi) = fyi(xi) = 1. The new
instance is J = (M ∪ X, W ∪ Y, {fm}m∈M ∪X , {fw}w∈W ∪Y , k + t).
We note that for all i ∈ {1, 2, . . . , t}, it holds that (xi, yi) is a happy pair.
Lemma 3.15. Reduction Rule 9 is safe, and t(I) = t(J ).
Proof. For all i ∈ {1, 2, . . . , t}, it holds that (xi, yi) is a happy pair, and therefore it is present
in every stable matching in J . By our definition of the new preference functions, it is clear that
if µ is a stable matching in I, then µ′ = µ ∪ {(x1, y1), . . . , (xt, yt)} is a stable matching in J .
Moreover, if µ′ is a stable matching in J , then µ = µ′\{(x1, y1), . . . , (xt, yt)} is a stable matching
in I. Hence, since for all i ∈ {1, 2, . . . , t}, it holds that fxi(yi) = fyi(xi) = 1, our definition of
the new preference functions directly implies that Bal(I) + t = Bal(J ), OM (I) + t = OM (J )
and OW (I) + t = OW (J ). Hence, t(I) = t(J ), which concludes the proof.
Reduction Rule 10. [Male version] If there exists m ∈ M such that fm has a gap at some j,
then select some yi ∈ Y \ A(m), and set A′(m) = A(m) ∪ {yi} and A′(yi) = A(yi) ∪ {m}. The
preference functions are defined as follows.
• The preference function of m. f ′
m(yi) = j, and for all a ∈ A(m), f ′
m(a) = fm(a).
• The preference function of yi. f ′
yi(m) = max
m′∈A(yi)
(fyi(m′) + 1), and for all a ∈ A(yi),
f ′
yi(a) = fyi(a).
16
• For all a ∈ (M ∪ W ) \ {m, yi}: f ′
a = fa.
The new instance is J = (M, W, {f ′
w′}w′∈W , k).
Lemma 3.16. Reduction Rule 10 is safe, and t(I) = t(J ).
m′ }m′∈M , {f ′
Proof. The only modifications that are performed are the insertion of m intro the set of ac-
ceptable partners of yi as the least preferred person, and the insertion of yi into the set of
acceptable partners of m in a location that previosuly contained a gap. Let us first observe that
since fxi(yi) = fyi(xi) = 1 and f ′
yi(xi) = 1, it holds that (xi, yi) is a happy pair in
both I and J . Hence, it is clear that SM(I) = SM(J ), OM (I) = OM (J ), OW (I) = OW (J ),
and that the balance of any stable matching in I is equal to its balance in J . We thus conclude
that the rule is safe and that t(I) = t(J ).
xi(yi) = f ′
Analogously, we have a female version of Reduction Rule 10 where we fill a gap in the
preference function of some woman w ∈ W . We do not repeat our arguments again, and
straightaway state the following result, which follows directly from the safeness of Reduction
Rule 9 and the male and female versions of Reduction Rule 10.
Lemma 3.17. Above-Min FBSM admits a kernel that has at most 3t men among whom at
most 2t are sad, at most 3t women among whom at most 2t are sad, and such that each happy
person has at most 2t + 1 acceptable partners and each sad person has at most t + 1 acceptable
partners.
Finally, we translate the kernel for Above-Min FBSM to an instance of Above-Min
BSM as follows. For all a ∈ M ∪ W and b ∈ A(a), we set pa(b) = fa(b). The new instance
is J = (M, W, {pm}m∈M , {fw}w∈W , k). Clearly, we thus obtain an equivalent instance, which
leads us to the following somewhat stronger version of Theorem 3.1, relevant to Appendix 4.
Lemma 3.18. Above-Min BSM admits a kernel that has at most 3t men among whom at 2t
are sad, at most 3t women among whom at most 2t are sad, and such that each happy person
has at most 2t + 1 acceptable partners and each sad person has at most t + 1 acceptable partners.
Moreover, every stable matching in the kernel is a perfect matching.
This concludes the proof of Theorem 3.1.
4 Parameterized Algorithm
In this section, we design a parameterized algorithm for Above-Min BSM. More precisely, we
prove the following theorem.
Theorem 4.1. Above-Min BSM can be solved in time O∗(8t).
As our algorithm is based on the method of bounded search trees, we first give a brief
description of this technique.
4.1 Bounded Search Tree
The running time of an algorithm that uses bounded search trees can be analyzed as follows (see,
e.g., [2]). Suppose that the algorithm executes a branching rule which has ℓ branching options
(each leading to a recursive call with the corresponding parameter value), such that, in the ith
branch option, the current value of the parameter decreases by bi. Then, (b1, b2, . . . , bℓ) is called
the branching vector of this rule. We say that α is the root of (b1, b2, . . . , bℓ) if it is the (unique)
= xb∗−b1 + xb∗−b2 + · · · + xb∗−bℓ, where b∗ = max{b1, b2, . . . , bℓ}. If r > 0
positive real root of xb∗
is the initial value of the parameter, and the algorithm (a) returns a result when (or before)
the parameter is negative, and (b) only executes branching rules whose roots are bounded by a
constant c > 0, then its running time is bounded by O∗(cr).
17
4.2 Description of the Algorithm
Given an instance bI = (cM ,cW , bLM , bLW ,bk) of Above-Min BSM, we begin by using the pro-
cedure given by Lemma 3.18 to obtain (in polynomial time) a kernel I = (M, W, LM , LW , k)
of Above-Min BSM such that I has at most 3t men among whom at most 2t are sad, at
most 3t women among whom at most 2t are sad. Let us denote the happy pairs in I by
(x1, y1), (x2, y2), . . . , (xh, yh) for the appropriate h ≤ t, and the set of sad men by MS. Note
that MS ≤ 2t.
We proceed by executing a loop where each iteration corresponds to a different subset
M ′ ⊆ MS. For a specific iteration, our goal is to determine whether there exists a stable
matching µ such that the following conditions are satisfied:
• balance(µ) ≤ k.
• For all m ∈ M ′: µ(m) 6= µM (m).
• For all m ∈ MS \ M ′: µ(m) = µM (m).
A stable matching satisfying the conditions above (in the context of the current iteration)
is said to be valid. We denote r = k − OM , and observe that r ≤ t.
Let us now consider some specific iteration. To determine whether there exists a valid stable
matching, our plan is to execute a branching procedure, called Branch, which outputs every set
S of pairs of a man in M ′ and a woman, such that the following conditions are satisfied.
1. Every man m in M ′ participates in exactly one pair (m, w) of S, and for that unique pair,
it holds that w ∈ A(m) and pm(w) > pm(µM (m)).
2. X
m∈M ′
(pm(µ(m)) − pm(µM (m))) ≤ r.
The description of this procedure is given in the following subsection. Here, let us argue
that by having this procedure, we can conclude the proof of the correctness of the algorithm.
In the current iteration, we examine each set S in the outputted family of sets. Then, we
check whether the pairs in S, together with (x1, y1), (x2, y2), . . . , (xh, yh) and every pair in
{(m, µM (m)) : m ∈ M ′} form a stable matching whose balance is at most k. If the answer is
positive, then we terminate the execution and accept, which is clearly correct.
At the end, if we did not accept in any iteration, we reject. To see why this decision is
correct, suppose that there exists a stable matching µ whose balance is at most k.
In this
case, due to our exhaustive search, there exists some iteration in which µ is also valid. In that
iteration, associated with some M ′ ⊆ M −S, observe that the set of pairs {(m, µ(m)) : m ∈ M ′}
is one of the outputted sets of pairs. Indeed, the satisfaction of Condition 1 follows from the
fact that µ is a stable matching satisfying the last two conditions of validity, and Condition 2
follows from the fact that µ satisfies the first condition of validity.
Let us denote by T the running of the procedure Branch. Then, the total running time of
our algorithm is bounded by O∗(2MS · T ) = O∗(4t · T ), and therefore to derive the running
time in Theorem 4.1, we will need to ensure that T = O∗(2t). For this purpose, it is sufficient
to ensure that T = O∗(2r).
4.3 The Branching Procedure
We now present the description of the procedure Branch in the context of some set M ′ ⊆ MS.
Let us denote M ′ = {m1, m2, . . . , mp} for the appropriate p.
Each call to our procedure is of the form Branch(i, S) where i ∈ {0, 1, . . . , p + 1} and S is a
set of pairs of a man in {m1, m2, . . . , mi} and a woman, such that the following conditions are
satisfied.
18
1. Every man m in {m1, m2, . . . , mi} participates in exactly one pair (m, w) of S, and for
that unique pair, it holds that w ∈ A(m) and pm(w) > pm(µM (m)). We define µS as
the function whose domain is {m1, m2, . . . , mi} and which assigns to each man m in its
domain the unique woman w such that (m, w) ∈ S.
2. Define balance(S) = X
(pm(µS(m)) − pm(µM (m))). Then, balance(S) ≤ r.
m∈{m1,m2,...,mi}
Note that in case i = 0, we have that S = ∅. The call to Branch that is performed by
the algorithm given in Section 4.2 is precisely with these arguments, that is, Branch(0, ∅). The
objective of a call Branch(i, S) is to return a family F of sets, where each set F ∈ F is a set
of pairs of a man in {mi+1, mi+2, . . . , mp} and a woman, such that the following conditions are
satisfied.
1. Every man m in {mi+1, mi+2, . . . , mp} participates in exactly one pair (m, w) of F , and
for that unique pair, it holds that w ∈ A(m) and pm(w) > pm(µM (m)). We define µF as
the function whose domain is {mi+1, mi+2, . . . , mp} and which assigns to each man m in
its domain the unique woman w such that (m, w) ∈ F .
2. balance(S) +
X
m∈{mi+1,mi+2,...,mp}
(pm(µF (m)) − pm(µM (m))) ≤ r.
Clearly, by accomplishing this goal, we also achieve the objective posed in Section 4.2. The
measure that we employ to analyze our procedure is (r − balance(S)) (recall that balance(S)
is defined in Condition 2 of the specification of a call), which is initially equal to r. Hence,
according to the method of bounded search trees (see Section 4.1), to derive the running time
O∗(2r), it is sufficient to ensure that Branch (a) returns a result when (or before) the measure
r is negative, and (b) only executes branching rules whose roots are bounded by 2. When the
measure r is negative, we should simply return F = ∅, as there does not exist a set F satisfying
the conditions above. Otherwise, when i = p + 1, we can simply return F = {∅}.
Let us now consider a call Branch(i, S) where r ≥ 0 and i ≤ p. Denote fW = {w ∈ A(mi+1) :
pmi+1(w) > pmi+1(µM (mi+1))}. We further refine fW be letting W ∗ denote the set of those r
women in fW who are the most preferred by mi+1. In case there are no such r women since
fW < r, we simply denote W ∗ = fW . Let us also denote W ∗ = {w1, w2, . . . , wq} for the
appropriate q ≤ r. Then, our procedure executes r branches. At the jth branch, Branch calls
itself recursively with (i + 1, S ∪ {(mi+1, wj)}). Eventually, Branch returns Sq
j=1{{(mi+1, wj)} ∪
Fj : F ∈ Fj} where for all j ∈ {1, 2 . . . , q}, we set Fj to be the family of sets of pairs that was
returned by the recursive call of the jth branch. The correctness follows from the observation
that the process that we execute is an exhaustive search. More precisely, if there exists a set
F satisfying Conditions 1 and 2, then it must include exactly one of the pairs in {(mi+1, wj) :
j ∈ {1, 2, . . . , q}}. Now, let us observe that at the jth branch, the measure changes from
r−balance(S) to r−balance(S ∪{(mi+1, wj)}). By our definition of wj, we have that pmi+1(wj)−
pmi+1(µM (mi+1))) = j. Hence, at the worst case, the branching vector is (1, 2, . . . , r). Since the
root of such a branching vector upper bounded by 2, our proof is complete.
5 Hardness
In this section, we prove the following theorem.
Theorem 5.1. Above-Max BSM is W[1]-hard.
For this purpose, we consider the Clique problem, which is defined as follows.
19
Clique
Input: A graph G = (V, E), and a positive integer k.
Question: Does G contain a clique on k vertices?
Parameter: k.
The Clique problem is known to be W[1]-hard [4]. Thus, to prove Theorem 5.1, it is
sufficient to prove the following result.
Lemma 5.1. Given an instance I = (G = (V, E), k) of Clique, an equivalent instance bI =
(M, W, LM , LW ,bk) of Above-Max BSM such that t = 6(k +
time O(f (k) · IO(1)) for some function f .
) can be constructed in
k(k − 1)
2
The rest of this section focuses on the proof of Lemma 5.1. To this end, we let I =
(G = (V, E), k) be some instance of Clique. In Section 5.1, we construct (in "FPT time") an
explanation of the intuition underlying the construction. Then, in Section 5.2, we verify that
instance bI = (M, W, LM , LW ,bk) of Above-Max BSM. This section also contains an informal
the parameter t associated with bI is equal to 6(k +
prove that the input instance I of Clique and our instance bI of Above-Max BSM are
equivalent.
denote V = {v1, v2, . . . , vV } and E = {e1, e2, . . . , eE}.
In what follows, we select arbitrary orders on V and E, according to which we
). Finally, in Section 5.3, we
k(k − 1)
2
5.1 Reduction
First, to construct the sets M and W , we define three pairwise-disjoint subsets of M , called
MV , ME and fM , and three pairwise-disjoint subsets of W , called WV , WE and fW . Then, we
set M = MV ∪ ME ∪ fM ∪ {m∗} and W = WV ∪ WE ∪fW ∪ {w∗}, where m∗ and w∗ denote a new
man and a new woman (which do not belong to the six sets defined previously), respectively.
• MV = {mi
v : v ∈ V, i ∈ {1, 2}}; WV = {wi
v : v ∈ V, i ∈ {1, 2}}.
• ME = {mi
e : e ∈ E, i ∈ {1, 2}}; WE = {wi
e : e ∈ E, i ∈ {1, 2}}.
• Let δ = 2(V + E + V E + V E2) − k(4 + 4k + 2E + (k − 1)V E).
Then, fM = {emi : i ∈ {1, 2, . . . , δ}} and fW = {ewi : i ∈ {1, 2, . . . , δ}}.
Note that M = W . We remark that in what follows, we assume w.l.o.g. that δ ≥ 0 and
k(k − 1)
V > k +
k and can therefore, by using brute-force, be solved in FPT time.
, else the size of the input instance I of Clique is bounded by a function of
2
v and m2
Before we proceed, let us discuss the intuition behind the definition of the subsets of men
and women above, and the definitions of the preference lists that will follow. Roughly speaking,
each pair of men, m1
v, represents a vertex, and we aim to ensure that either both men
will be matched to their best partners (in the man-optimal stable matching) or both men will
be matched to other partners (where there would be only one choice for these other partners
that preserves stability). Accordingly, we will guarantee that the choice of matching these two
men to their best partners translates to not choosing the vertex they represent into the clique,
and the other choice translates to choosing this vertex into the clique.
Now, having just the set MV , we can encode selection of vertices into the clique, but we
cannot ensure that the vertices we select indeed form a clique. For this purpose, we also have
the set ME which, in a manner similar to MV , encodes selection of edges into the clique. By
designing the instance in a way that the situation of the men in the man-optimal stable matching
is significantly worse than that of the women in the women-optimal stable matching, we are able
20
k(k − 1)
2
to ensure that at most 2(k +
) men in MV ∪ ME will not be assigned their best partners
(here, we exploit the condition that balance(µ) ≤ bk for a solution µ). We remark that here the
man m∗ plays a crucial role-by using dummy men and women (in the sets fM and fW ) that
prefer each other over all other people, we ensure that the situation of m∗ is always "extremely
bad" (from his viewpoint), while the situation of his partner, w∗, is always "excellent" (from
her viewpoint).
At this point, we first need to ensure that the edges that we select indeed connect the
vertices that we select. For this purpose, we carefully design our reduction so that when a pair
of men representing some edge e obtain partners worse than those they have in the man-optimal
stable matching, it must be that the men representing the endpoints of e have also obtained
partners worse than those they have in the man-optimal stable matching, else stability will not
be preserved-the partners of the men represting the endpoints of e will form blocking pairs
together with the men representing e.
k(k − 1)
2
) distinguished
Finally we observe that we still need to ensure that among our 2(k +
men in MV ∪ME, which are associated with k +
selected elements (vertices and edges),
there will be exactly 2k distinguished men from MV and exactly k(k − 1) distinguished men
2
k(k − 1)
k(k − 1)
edges. For this purpose,
from ME, which would mean we have chosen k vertices and
we construct an instance where for the women, it is only somewhat "beneficial" that the men
in MV will not be matched to their best partners, but it is extremely beneficial that the men in
ME will not be matched to their best partners. This objective is achieved by carefully placing
dummy men (from cM ) in the preference lists of women in WE. By again exploiting the condition
that balance(µ) ≤ bk for a solution µ, we are able to ensure that there would be at least k(k − 1)
distinguished men from ME.
Next, we proceed with the formal presentation of our reduction by defining pm for every
2
man m ∈ M , and thus constructing LM .
• For all m1
v ∈ MV : A(m1
– pm1
v
(w1
v) = 1; pm1
v
v, ew1, ew2, w2
v}.
(ew2) = 3; pm1
v
(w2
v) = 4.
• For all m2
v ∈ MV : A(m2
v, w1, w2, w1
v}.
v
v) = {w1
(ew1) = 2; pm1
v) = {w2
(ew1) = 2; pm2
v
– pm2
v
(w2
v) = 1; pm2
v
(w1
v) = 4.
v
(ew2) = 3; pm2
{u,v}) = {w1
• For all m1
{u,v} ∈ ME where u < v:3 A(m1
{u,v}, w1
u, w1
v, w2
{u,v}}.
– pm1
{u,v}
(w1
{u,v}) = 1; pm1
{u,v}
(w1
u) = 2; pm1
{u,v}
(w1
v) = 3; pm1
{u,v}
(w2
{u,v}) = 4.
• For all m2
{u,v} ∈ ME where u < v: A(m2
{u,v}) = {w2
{u,v}, w2
u, w2
v, w1
{u,v}}.
– pm2
{u,v}
(w2
{u,v}) = 1; pm2
{u,v}
(w2
u) = 2; pm2
{u,v}
(w2
v) = 3; pm2
{u,v}
(w1
{u,v}) = 4.
• For all emi ∈ fM such that i ≤ V E: A(emi) = {ewi} ∪ WE ∪ {wi
v ∈ WV , the set A(wi
v) is determined later.
v ∈ MV : emi ∈ A(wi
v)},
where for all wi
– p emi(ewi) = 1.
3Recall that we have defined an order on V .
21
– For all j ∈ {1, 2, . . . , E}: p emi(w1
– Denote X = {wi
bijection. Then, for all wi
v ∈ MV : emi ∈ A(wi
v ∈ X: p emi(wi
ej ) = j + 1; p emi(w2
ej ) = E + j + 1.
v)}. Let f : X → X be an arbitrarily chosen
v) = 2E + f (wi
v) + 1.
• For all emi ∈ fM such that i > V E: A(emi) = {ewi}.
– p emi(ewi) = 1.
• For m∗: A(m∗) = fW ∪ {w∗}.
– For all i ∈ {1, 2, . . . , δ}: pm∗(ewi) = i.
– pm∗ (w∗) = δ + 1.
Accordingly, we define pw for every woman w ∈ W , and thus construct LW .
• For all w1
v ∈ WV : A(w1
v) = {m1
e ∈ ME : v ∈ e} ∪ {m2
v, m1
v} ∪ {emi ∈ fM : i ≤ E −
degreeG(v)}.
(m2
v
v) = 1
– pw1
– For all i ∈ {1, 2, . . . , E}: pw1
– pw1
v) = E + 2.
(m1
v
v
(m1
ei) = i + 1.
• For all w2
v ∈ WV : A(w2
v) = {m2
e ∈ ME : v ∈ e} ∪ {m1
v, m2
v} ∪ {emi ∈ fM : i ≤ E −
degreeG(v)}.
(m1
v
v) = 1.
– pw2
– For all i ∈ {1, 2, . . . , E}: pw2
– pw2
v) = E + 2.
(m2
v
v
(m2
ei) = i + 1.
• For all w1
e ∈ WE: A(w1
e ) = {emi ∈ fM : i ∈ {1, 2, . . . , V E}} ∪ {m2
e, m1
e}.
(m2
e
e) = 1.
– pw1
– For all i ∈ {1, 2, . . . , V E}: pw1
– pw1
e) = V E + 2.
(m1
e
e
(emi) = i + 1.
• For all w2
e ∈ WE: A(w2
e ) = {emi ∈ fM : i ∈ {1, 2, . . . , V E}} ∪ {m1
e, m2
e}.
(m1
e
e) = 1.
– pw2
– For all i ∈ {1, 2, . . . , V E}: pw2
– pw2
e) = V E + 2.
(m2
e
e
(emi) = i + 1.
• For all ewi ∈ fW such that i ∈ {1, 2}: A(ewi) = {emi} ∪ {m∗} ∪ MV .
– p ewi(emi) = 1; p ewi(m∗) = 2.
– For all j ∈ {1, 2, . . . , V }: p ewi(m1
vj ) = j + 2; p ewi(m2
• For all ewi ∈ fW such that i > 2: A(ewi) = {emi} ∪ {m∗}.
vj ) = V + j + 2.
– p ewi(emi) = 1; p ewi(m∗) = 2.
22
• For w∗: A(w∗) = {m∗}.
– pw∗(m∗) = 1.
Finally, we define bk = M + δ + 6(k +
(under the assumptions that δ ≥ 0 and V > k +
time.
5.2 The Parameter
2
k(k − 1)
2
). It is clear that the entire construction
k(k − 1)
) can be performed in polynomial
Our current objective is to verify that t is indeed bounded by a function of k. For this purpose,
we first observe that for all i ∈ {1, 2, . . . , δ}, it holds that p emi(ewi) = p ewi(emi) = 1. Therefore, for
all µ ∈ SM(bI) and i ∈ {1, 2, . . . , δ}, we have that µ(emi) = ewi, else (emi, ewi) would have formed
a blocking pair for µ in bI.
Observation 6. For all µ ∈ SM (bI) and i ∈ {1, 2, . . . , δ}, it holds that µ(emi) = ewi.
Now, note that A(m∗) = fW ∪{w∗}. Thus, by Observation 6, we have that for all µ ∈ SM (bI),
either m∗ is unmatched or µ(m∗) = w∗. However, A(w∗) = {m∗}, which implies that in the
former case, (m∗, w∗) forms a blocking pair. Thus, we also have the following observation.
Observation 7. For all µ ∈ SM (bI), it holds that µ(m∗) = w∗.
Let us proceed by identifying the man-optimal µM and the woman-optimal µW stable match-
ings. For this purpose, we first define a matching µ′
M as follows.
• For all mi
• For all mi
v ∈ MV : µ′
e ∈ ME: µ′
• For all emi ∈ fM : µ′
M (mi
v) = wi
v.
e) = wi
M (mi
e.
M (emi) = ewi.
• µ′
M (m∗) = w∗.
Lemma 5.2. µM = µ′
M .
M (emi) = ewi and p emi(ewi) = p ewi(emi) = 1, we
Proof. Since for all i ∈ {1, 2, . . . , δ}, it holds that µ′
have that there cannot exist a blocking pair with at least one person from fM ∪fW . Now, notice
that for every m ∈ M , including m∗, the woman most preferred by m who is outside fW is also
the one with whom it is matched. Therefore, there cannot exist any blocking pair for µ′
M , and
by Observation 6, we further conclude that indeed µM = µ′
M .
Now, we define a matching µ′
W as follows.
• For all wi
• For all wi
v ∈ WV : µ′
e ∈ WE: µ′
• For all ewi ∈ fW : µ′
W (wi
W (wi
W (ewi) = emi.
v) = m3−i
e) = m3−i
.
v
e
• µ′
W (w∗) = m∗.
.
Lemma 5.3. µW = µ′
W .
Proof. In the matching µ′
it is immediate that µW = µ′
W .
W , every woman is matched with the man she prefers the most. Thus,
As a corollary to Lemmata 5.2 and 5.3, we obtain the following result.
Corollary 5.1. OM = M + δ and OW = W .
23
Proof. First, note that
OM = X
pm(w)
(m,w)∈µM
= pm∗ (µM (m∗)) + X
m∈M \{m∗}
Second, note that
pm(µM (m)) = (δ + 1) + (M − 1) = M + δ.
OW = X
pw(m) = X
(m,w)∈µW
w∈W
pw(µW (w)) = W .
We are now ready to bound t.
Lemma 5.4. The parameter t associated with bI is equal to 6(k +
k(k − 1)
2
).
Proof. By the definition of t, we have that
t = bk − max{OM , OW }
= M + δ + 6(k +
k(k − 1)
2
) − max{M + δ, W } = 6(k +
k(k − 1)
2
).
5.3 Correctness
First, from Lemma 5.2 and Proposition 2 we derive the following useful observation.
Observation 8. Every µ ∈ SM (bI) matches all people in M ∪ W .
Next, we proceed to state the first direction necessary to conclude that the input instance
I of Clique and our instance bI of Above-Max BSM are equivalent.
Lemma 5.5. If I is a Yes-instance, then bI is a Yes-instance.
Proof. Suppose that I is a Yes-instance, and let U be the vertex-set of a clique on k vertices
in G. We denote M U
{u,v} ∈ ME : u, v ∈ U }. Then, we
define a matching µ as follows.
v ∈ MV : v ∈ U } and M U
V = {mi
E = {mi
• For all mi
v ∈ MV :
– If mi
v ∈ M U
– Else: µ(mi
V : µ(mi
v) = w3−i
v) = wi
v.
v
.
• For all mi
e ∈ ME:
e ∈ M U
– If mi
– Else: µ(mi
e) = w3−i
E : µ(mi
e) = wi
e.
e
• For all emi ∈ fM : µ(emi) = ewi.
• µ(m∗) = w∗.
.
24
On the one hand, notice that for every m ∈ (MV \ M U
We claim that µ ∈ SM (bI) and balance(µ) ≤ bk, which would imply that bI is a Yes-instance.
To this end, we first show that µ ∈ SM (bI). Since for all i ∈ {1, 2, . . . , δ}, it holds that
µ(emi) = ewi and p emi(ewi) = p ewi(emi) = 1, we have that there cannot exist a blocking pair with
at least one person from fM ∪ fW . Thus, there can also not be a blocking pair with any person
from {m∗, w∗}.
E ) ∪ {m∗}, the woman
most preferred by m who is outside fW is also the one with whom it is matched. Thus, no man
in (MV \M U
E )∪{m∗} can belong to a blocking pair. Moreover, the set of acceptable
E is a subset of fM ∪ (ME \ M U
partners of any woman in WE matched to a man in ME \ M U
E ),
and therefore such a woman cannot belong to a blocking pair. On the other hand, let W ′ denote
the set of every woman that is matched to a man m ∈ M U
E . Then, for every w ∈ W ′, the
man most preferred by w is also the one with whom she is matched. Therefore, no woman in
W ′ can belong to a blocking pair. Hence, we also conclude that no woman in WE can belong
to a blocking pair.
V ) ∪ (ME \ M U
V )∪(ME \M U
V ∪ M U
v ∈ M U
V ∪ M U
Thus, if there exists a blocking pair, it must consist of a man m ∈ M U
E and a woman
w ∈ WV \ W ′. Suppose, by way of contradiction, that there exists such a blocking pair (m, w).
v, all women in A(mi
First, let us assume that m = mi
v)
belong to fW ∪ {µ(mi
v, and
thus we reach a contradiction. Next, we assume that m = mi
E . In this case, it must
hold that w is either wi
v. However,
since mi
over
mi
u. Without loss of generality, we assume that w = wi
E , we have that v ∈ U . Therefore, µ(wi
V . In this case, since apart from wi
v)}, we deduce that w = wi
{u,v}, we reach a contradiction.
It remains to prove that balance(µ) ≤ bk. To this end, we need to show that
v prefers m3−i
{u,v} ∈ M U
v. However, wi
v prefers µ(wi
v) over mi
{u,v} ∈ M U
v) = m3−i
v
. Since wi
v or wi
v
max{ X
pm(w), X
pw(m)} ≤ M + δ + 6(k +
k(k − 1)
).
2
(m,w)∈µ
(m,w)∈µ
First, note that
X
pm(w) = X
pm(µ(m)) + X
pm(µ(m)) + X
pm(µ(m))
(m,w)∈µ
m∈M U
V
+ X
m∈MV \M U
V
pm(µ(m)) + X
m∈M U
E
pm(µ(m)) + pm∗ (µ(m∗))
m∈ME \M U
E
= 4M U
V + MV \ M U
E + ME \ M U
m∈ fM
V + 4M U
= M + δ + 3(M U
V + M U
E ) = M + δ + 6(k +
E + fM + δ + 1
k(k − 1)
).
2
Second, note that
X
pw(m) = X
pµ(m)(m) + X
pµ(m)(m) + X
pµ(m)(m)
(m,w)∈µ
m∈M U
V
+ X
m∈MV \M U
V
pµ(m)(m) + X
m∈M U
E
pµ(m)(m) + pµ(m∗)(m∗)
m∈ME\M U
E
V + MV \ M U
= M U
V (E + 2) + M U
= M + 2(V − k)(E + 1) + 2(E −
m∈ fM
E + ME \ M U
k(k − 1)
E (V E + 2) + fM + 1
)(V E + 1)
2
= M + δ + 6(k +
k(k − 1)
2
).
This concludes the proof of the lemma.
We now turn to prove the second direction.
25
Lemma 5.6. If bI is a Yes-instance, then I is a Yes-instance.
Proof. Suppose that bI is a Yes-instance, and let µ be a stable matching such that balance(µ) ≤
bk. By Observations 6 and 7, it holds that
• For all i ∈ {1, 2, . . . , δ}: µ(emi) = ewi.
• µ(m∗) = w∗.
Thus, since Observation 8 implies that all vertices in MV should be matched by µ, we deduce
that
• For all v ∈ V : Either both µ(m1
v) = w1
v and µ(m2
v) = w2
v or both µ(m2
v) = w1
v and
µ(m1
v) = w2
v.
Let U denote the set of every v ∈ V such that µ(m2
v. Moreover,
v ∈ MV : v ∈ U }. By the item above, and since all vertices in ME should also
v and µ(m1
v) = w1
v) = w2
denote M U
be matched by µ, we further deduce that
V = {mi
• For all e ∈ E: Either both µ(m1
e) = w1
e and µ(m2
e) = w2
e or both µ(m2
e) = w1
e and
µ(m1
e) = w2
e .
Let S denote the set of every e ∈ E such that µ(m2
e . Moreover,
denote M S
{u,v}, w1
u)
would have formed a blocking pair, which contradicts the fact that µ is a stable matching. Thus,
we have that the set of endpoints of the edges in S is a subset of U .
e ∈ ME : e ∈ S}. If there existed {u, v} ∈ S such that u /∈ U , then (m1
e and µ(m1
E = {mi
e) = w2
e) = w1
We claim that U = k and that U is the vertex-set of a clique in G, which would imply that
I is a Yes-instance. Since we have argued that the set of endpoints of the edges in S is a subset
of U , it is sufficient to show that U ≤ k and S ≥
implies
that U ≥ k), as this would imply that U is indeed the vertex-set of a clique on k vertices in
(note that S ≥
2
2
k(k − 1)
k(k − 1)
G. First, since balance(µ) ≤ bk, we have that X
(m,w)∈µ
pm(w) ≤ M + δ + 6(k +
k(k − 1)
2
). Now,
note that
X
pm(w) = X
pm(µ(m)) + X
pm(µ(m)) + X
pm(µ(m))
(m,w)∈µ
m∈M U
V
+ X
m∈MV \M U
V
pm(µ(m)) + X
m∈M S
E
pm(µ(m)) + pm∗(µ(m∗))
m∈ME \M S
E
= 4M U
= M + δ + 6(U + S).
V + MV \ M U
m∈ fM
V + 4M S
E + ME \ M S
E + fM + δ + 1
Thus, we deduce that U + S ≤ k +
k(k − 1)
. Now, observe that since balance(µ) ≤ bk, we
2
also have that X
(m,w)∈µ
pw(m) ≤ M + δ + 6(k +
). Here, on the one hand we note that
k(k − 1)
2
X
pw(m) = X
pµ(m)(m) + X
pµ(m)(m) + X
pµ(m)(m)
(m,w)∈µ
m∈M U
V
+ X
m∈MV \M U
V
pµ(m)(m) + X
m∈M S
E
pµ(m)(m) + pµ(m∗)(m∗)
m∈ME\M S
E
V + MV \ M U
m∈ fM
E(V E + 2) + fM + 1
= M U
= M + 2(V − U )(E + 1) + 2(E − S)(V E + 1)
= M + 2(V + E + V E + V E2) − 2U (E + 1) − 2S(V E + 1).
V (E + 2) + M S
E + ME \ M S
26
On the other hand, we note that
bk = M + δ + 6(k +
k(k − 1)
2
)
= M + 2(V + E + V E + V E2) − k(4 + 4k + 2E + (k − 1)V E) + 6(k +
= M + 2(V + E + V E + V E2) − 2k(E + 1) − k(k − 1)(V E + 1).
Thus, we have that
U (E + 1) + S(V E + 1) ≥ k(E + 1) +
k(k − 1)
2
(V E + 1)
k(k − 1)
2
)
Recall that we have also shown that U +S ≤ k +
k(k − 1)
2
. Thus, since U ≤ k +
V , to satisfy the above equation it must hold that S ≥
. Since U + S ≤
k(k − 1)
2
<
k(k − 1)
2
k +
k(k − 1)
2
, we deduce that U ≤ k. This, as we have argued earlier, finished the proof.
This concludes the proof of Theorem 5.1.
References
[1] J. Chen, D. Hermelin, M. Sorge, and H. Yedidsion, How hard is it to satisfy (almost)
all roommates?, arXiv preprint arXiv:1707.04316, (2017).
[2] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh, Parameterized Algorithms, Springer, 2015.
[3] M. Cygan, M. Pilipczuk, M. Pilipczuk, and J. O. Wojtaszczyk, On multiway cut
parameterized above lower bounds, TOCT, 5 (2013), pp. 3:1–3:11.
[4] R. G. Downey and M. R. Fellows, Fixed-parameter tractability and completeness II:
On completeness for W[1], Theoretical Computer Science, 141 (1995), pp. 109–131.
[5]
, Fundamentals of parameterized complexity, Springer, 2013.
[6] T. Feder, Stable networks and product graphs, PhD thesis, Stanford University, 1990.
[7] D. Gale and L. S. Shapley, College admissions and the stability of marriage, The
American Mathematical Monthly, 69 (1962), pp. 9–15.
[8]
, College admissions and the stability of marriage, American Mathematical Monthly,
69 (1962), pp. 9–15.
[9] D. Gale and M. Sotomayor, Some remarks on the stable matching problem, Discrete
Applied Mathematics, 11 (1985), pp. 223–232.
[10] S. Garg and G. Philip, Raising the bar for vertex cover: Fixed-parameter tractability
above A higher guarantee, in Proceedings of the Twenty-Seventh Annual ACM-SIAM Sym-
posium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016,
2016, pp. 1152–1166.
[11] D. Gusfield and R. W. Irving, The Stable marriage problem - structure and algorithms,
Foundations of computing series, MIT Press, 1989.
[12] R. W. Irving, P. Leather, and D. Gusfield, An efficient algorithm for the "optimal"
stable marriage, Journal of ACM, 34 (1987), pp. 532–543.
27
[13] A. Kato, Complexity of the sex-equal stable marriage problem, Japan Journal of Industrial
and Applied Mathematics, 10 (1993), p. 1.
[14] D. E. Knuth, Stable marriage and its relation to other combinatorial problems : an
introduction to the mathematical analysis of algorithms, CRM proceedings & lecture notes,
Providence, R.I. American Mathematical Society, 1997.
[15] D. Lokshtanov, N. S. Narayanaswamy, V. Raman, M. S. Ramanujan, and
S. Saurabh, Faster parameterized algorithms using linear programming, ACM Trans. Al-
gorithms, 11 (2014), pp. 15:1–15:31.
[16] D. F. Manlove, Algorithmics of Matching Under Preferences, vol. 2 of Series on Theo-
retical Computer Science, WorldScientific, 2013.
[17] E. McDermid, in Personal communications between McDermid and Manlove, 2010.
[18] E. McDermid and R. W. Irving, Sex-equal stable matchings: Complexity and exact
algorithms, Algorithmica, 68 (2014), pp. 545–570.
[19] G. O'Malley, Algorithmic aspects of stable matching problems, PhD thesis, University of
Glasgow, 2007.
[20] V. Raman, M. S. Ramanujan, and S. Saurabh, Paths, flowers and vertex cover, in Al-
gorithms - ESA 2011 - 19th Annual European Symposium, Saarbrucken, Germany, Septem-
ber 5-9, 2011. Proceedings, 2011, pp. 382–393.
28
|
1608.06325 | 1 | 1608 | 2016-08-22T21:50:07 | A PTAS for the Steiner Forest Problem in Doubling Metrics | [
"cs.DS"
] | We achieve a (randomized) polynomial-time approximation scheme (PTAS) for the Steiner Forest Problem in doubling metrics. Before our work, a PTAS is given only for the Euclidean plane in [FOCS 2008: Borradaile, Klein and Mathieu]. Our PTAS also shares similarities with the dynamic programming for sparse instances used in [STOC 2012: Bartal, Gottlieb and Krauthgamer] and [SODA 2016: Chan and Jiang]. However, extending previous approaches requires overcoming several non-trivial hurdles, and we make the following technical contributions.
(1) We prove a technical lemma showing that Steiner points have to be "near" the terminals in an optimal Steiner tree. This enables us to define a heuristic to estimate the local behavior of the optimal solution, even though the Steiner points are unknown in advance. This lemma also generalizes previous results in the Euclidean plane, and may be of independent interest for related problems involving Steiner points.
(2) We develop a novel algorithmic technique known as "adaptive cells" to overcome the difficulty of keeping track of multiple components in a solution. Our idea is based on but significantly different from the previously proposed "uniform cells" in the FOCS 2008 paper, whose techniques cannot be readily applied to doubling metrics. | cs.DS | cs | A PTAS for the Steiner Forest Problem in Doubling Metrics
T-H. Hubert Chan∗
Shuguang Hu∗
Shaofeng H.-C. Jiang∗
Abstract
We achieve a (randomized) polynomial-time approximation scheme (PTAS) for the Steiner
Forest Problem in doubling metrics. Before our work, a PTAS is given only for the Euclidean
plane in [FOCS 2008: Borradaile, Klein and Mathieu]. Our PTAS also shares similarities
with the dynamic programming for sparse instances used in [STOC 2012: Bartal, Gottlieb and
Krauthgamer] and [SODA 2016: Chan and Jiang]. However, extending previous approaches
requires overcoming several non-trivial hurdles, and we make the following technical contribu-
tions.
(1) We prove a technical lemma showing that Steiner points have to be "near" the terminals
in an optimal Steiner tree. This enables us to define a heuristic to estimate the local behavior
of the optimal solution, even though the Steiner points are unknown in advance. This lemma
also generalizes previous results in the Euclidean plane, and may be of independent interest for
related problems involving Steiner points.
(2) We develop a novel algorithmic technique known as "adaptive cells" to overcome the
difficulty of keeping track of multiple components in a solution. Our idea is based on but
significantly different from the previously proposed "uniform cells" in the FOCS 2008 paper,
whose techniques cannot be readily applied to doubling metrics.
6
1
0
2
g
u
A
2
2
]
S
D
.
s
c
[
1
v
5
2
3
6
0
.
8
0
6
1
:
v
i
X
r
a
∗Department of Computer Science, the University of Hong Kong. {hubert,sghu,sfjiang}@cs.hku.hk
1
Introduction
We consider the Steiner Forest Problem (SFP) in a metric space (X, d). An instance of the problem
is given by a collection W of n terminal pairs {(ai, bi) : i ∈ [n]} in X, and the objective is to find
a minimum weight graph F = (V, E) (where V is a subset of X and the edge weights are induced
by the metric space) such that every pair in W is connected in F .
1.1 Problem Background
The problem is well-known in the computer science community. In general metrics, Chleb´ık and
Chleb´ıkov´a [CC08] showed that SFP is NP-hard to approximate with ratio better than 96
95 . The
best known approximation ratio achievable in polynomial time is 2 [GW95, AKR95]. Recently,
Gupta and Kumar [GK15] gave a purely combinatorial greedy-based algorithm that also achieves
constant ratio. However, it is still an open problem to break the 2-approximation barrier in general
metrics for SFP.
In light of the aforementioned hardness re-
SFP in Euclidean Plane and Planar Graphs.
sult [CC08], restrictions are placed on the metric space to achieve (1 + ǫ) approximation in polyno-
mial time. In the Euclidean plane, a randomized polynomial-time approximation scheme (PTAS)
was obtained in [BKM08], using the dynamic programming framework proposed by Arora [Aro98].
Later on, a simpler analysis was presented in [BH12], in which a new structural property is proved
and additional information is incorporated in the dynamic programming algorithm. It was only
suggested that similar techniques might be applicable to higher-dimensional Euclidean space.
Going beyond the Euclidean plane, a PTAS for planar graphs is obtained in [BHM11] and more
generally, on bounded genus graphs. As a building block, they also obtained a PTAS for graphs
with bounded treewidth.
Steiner Tree Problems. A notable special case of SFP is the Steiner Tree Problem (STP), in
which all terminals are required to be connected. In general metrics, the MST on the terminal points
simply gives a 2-approximation. There is a long line of research to improve the 2-approximation,
and the state-of-the-art approximation ratio 1.39 was presented in [BGRS10] via an LP rounding
approach. On the other hand, it is NP-hard to approximate STP better than the ratio 96
95 [CC08].
For the group STP in general metrics, it is NP-hard to approximate within log2−ǫ n [HK03]
unless NP ⊆ ZTIME(npolylog(n)). On the other hand, it is possible to approximate within O(log3 n)
as shown in [GKR00]. Restricting to planar graphs, the group STP can be approximated within
O(log n poly log log n) [DHK14], and very recently, this result is improved to a PTAS [BDHM16].
For more related works, we refer the reader to a survey by Hauptmann and Karpi´nski [HK13],
who gave a comprehensive literature review of STP and its variations.
PTAS's for Other Problems in Doubling Metrics. Doubling dimension captures the local
growth rate of a metric space. A k-dimensional Euclidean dimension has doubling dimension O(k).
A challenge in extending algorithms for low-dimensional Euclidean space to doubling metrics is the
lack of geometric properties in doubling metrics. Although QPTAS's for various approximation
problems in doubling metrics, such as the Traveling Salesman Problem (TSP) and STP, were
presented in [Tal04], a PTAS was only recently achieved for TSP [BGK12]. Subsequently, a PTAS
is also achieved for group TSP in doubling metrics [CJ16]. Before this work, the existence of a
PTAS for SFP (or even the special case STP) in doubling metrics remains an open problem.
1.2 Our Contribution and Techniques
Although PTAS's for TSP (and its group variant) are known, as we shall explain later, the nature
of SFP and TSP-related problems are quite different. Hence, it is interesting to investigate what
1
new techniques are required for SFP. Fundamentally, it is an important question that whether the
notion of doubling dimension captures sufficient properties of a metric space to design a PTAS
for SFP, even without the geometric properties that are crucially used in obtaining approximation
schemes for SFP in the Euclidean plane [BKM08].
In this paper, we settle this open problem by giving a (randomized) PTAS for SFP in doubling
metrics. We remark that previously even a PTAS for SFP in higher-dimensional Euclidean space
is not totally certain.
Theorem 1.1 (PTAS for SFP in Doubling Metrics). For any 0 < ǫ < 1, there is a (randomized)
algorithm that takes an instance of SFP with n terminal pairs in a metric space with doubling
dimension at most k, and returns a (1 + ǫ)-approximate solution with constant probability, running
in time O(nO(1)k
) · exp(√log n · O( k
ǫ )O(k)).
We next give an overview of our techniques. On a high level, we use the divide and conquer
framework that was originally used by Arora [Aro98] to achieve a PTAS for TSP in Euclidean
space, and was extended recently to doubling metrics [BGK12].
However, we shall explain that it is non-trivial to adapt this framework to SFP, and how we
overcome the difficulties encountered. Moreover, we shall provide some insights regarding the
relationship between Euclidean and doubling metrics, and discuss the implications of our technical
lemmas.
Summary of Framework. As in [BGK12], a PTAS is designed for a class of special instances
known as sparse instances. Then, it can be shown that the general instances can be decomposed
into sparse instances. Roughly speaking, an instance is sparse, if there is an optimal solution such
that for any ball B with radius r, the portion of the solution in B has weight that is small with
respect to r.
The PTAS for the sparse instances is usually based on a dynamic program, which is based on
a randomized hierarchical decomposition as in [Tal04, BGK12]. This framework has also been suc-
cessfully applied to achieve a PTAS for group TSP in doubling metrics [CJ16]. Intuitively, sparsity
is used to establish the property that with high enough probability, a cluster in the randomized
decomposition cuts a (near) optimal tour only a small number of times [BGK12, Lemma 3.1]. How-
ever, SFP brings new significant challenges when such a framework is applied. We next describe
the difficulties and give an overview of our technical contributions.
Challenge 1: It is difficult to detect a sparse instance because which Steiner points are
used by the optimal solution are unknown. Let us first consider STP, which is a special case
of SFP in which all (pairs of) terminals are required to be connected. In other words, the optimal
Steiner tree is the minimum weight graph that connects all terminals. Unlike TSP in which the
points visited by a tour are clearly known in advance, it is not known which points will be included
in the optimal Steiner tree.
In [BGK12], a crucial step is to estimate the sparsity of a ball B, which measures the weight
of the portion of the optimal solution restricted to B. For TSP tour, this can be estimated from
the points inside B that have to be visited. However, for solution involving Steiner points, it is
difficult to analyze the solution inside some ball B, because it is possible that there are few (or
even no) terminals inside B, but the optimal solution could potentially have lots of Steiner points
and a large weight inside B.
Our Solution: Analyzing the Distribution of Steiner Points in an Optimal Steiner Tree
in Doubling Metrics. We resolve this issue by showing a technical characterization of Steiner
points in an optimal Steiner tree for doubling metrics. This technical lemma is used crucially in
2
our proofs, and we remark that it could be of interest for other problems involving Steiner points
in doubling metrics.
Lemma 1.1 (Formal version in Lemma 3.1). For a terminal set S with diameter D, if an optimal
Steiner tree spanning S has no edge longer than γD, then every Steiner point in the solution is
within O(√γ) · D distance to some terminal in S, where the big O hides the dependence on the
doubling dimension.
We observe that variants of Lemma 1.1 have been considered on the Euclidean plane.
In
[DHC85, DHW87], it is shown that if the terminal set consists of n evenly distributed points on a
unit circle, then for large enough n, there is no Steiner points in an optimal Steiner tree. To see
how this relates to our lemma, when n is sufficiently large, it follows that adjacent points in the
circle are very close to each other. Hence, any long edge in a Steiner tree could be replaced by
some short edge between adjacent terminals in the circle. Our lemma then implies that all Steiner
points must be near the terminals, which is a weaker statement than the conclusion in [DHC85],
but is enough for our purposes. We emphasize that the results in [DHC85, DHW87] rely on the
geometric properties of the Euclidean plane. However, in our lemma, we only use that the doubling
dimension is bounded.
Implication of Lemma 1.1 on Sparsity Heuristic. We next demonstrate an example of how
we use this technical lemma. In Lemma 3.3, we argue that our sparsity heuristic provides an upper
bound on the weight of the portion of an optimal solution F within some ball B.
The idea is that we remove the edges in F within B and add back some edges of small total
weight to maintain connectivity. We first add a minimum spanning tree H on some net-points N
within B of an appropriate scale γ · D. Using the property of doubling dimension, we argue that
the number of points in H is bounded and so is its weight. In one of our case analysis, there are
two sets S and T of terminals that are far apart d(S, T ) ≥ D, and we wish to argue that in the
optimal Steiner tree F connecting S and T , there is an edge {u, v} of length at least Ω(γ) · D. If
this is the case, we could remove this edge and connect u and v to their corresponding net-points
directly. For contradiction's sake, we assume there is no such edge, but Lemma 1.1 implies that
every Steiner point must be close to either S and T . Since S and T are far apart, this means that
there is a long edge after all.
Conversely, in Lemma 3.4, we also use this technical lemma to show that if the sparsity heuristic
for some ball B is large, then the portion of the optimal solution F inside B is also large.
Challenge 2: In doubling metrics, the number of cells for keeping track of connectivity
in each cluster could be too large. Unlike the case for TSP variants [BGK12, CJ16], the
solution for SFP need not be connected. Hence, in the dynamic programming algorithm for SFP, in
addition to keeping track of what portals are used to connect a cluster to points outside, we need
to keep information on which portals the terminals inside a cluster are connected to. In previous
works [BKM08], the notion of cells is used for this purpose.
Previous Technique: Cell Property. The idea of cell property was first introduced in [BKM08],
which gave a PTAS for SFP in the Euclidean plane using dynamic programming. Since there would
have been an exponential number of dynamic program entries if we keep information on which
portal is used by every terminal to connect to its partner outside the cluster, the high level idea
is to partition a cluster into smaller clusters (already provided by the hierarchical decomposition)
known as cells. Loosely speaking, the cell property ensures that every terminal inside the same
cell must be connected to points outside the cluster in the same way. More precisely, a solution F
satisfies the cell property if for every cluster C and every cell e inside C, there is only one component
in the portion of F restricted to C that connects e to points outside C.
3
A great amount of work was actually needed in [BKM08] and subsequent work [BH12] to show
that it is enough to consider cells whose diameters are constant times smaller than that of its
cluster. This allows the number of dynamic program entries to be bounded, which is necessary for
a PTAS.
Difficulty Encountered for Doubling Metrics. When the notion of cell is applied to the
dynamic program for SFP in doubling metrics, an important issue is that the diameters of cells
need to be about Θ(log n) times smaller than that of its cluster, because there are around Θ(log n)
levels in the hierarchical decomposition. Hence, the number of cells in a cluster is Ω(poly log n),
which would eventually lead to a QPTAS only. A similar situation is observed when dynamic
programming was first used for TSP on doubling metrics [Tal04]. However, the idea of using
sparsity as in [BGK12] does not seem to immediately provide a solution.
Our Solution: Adaptive Cells. Since there are around Θ(log n) levels in the hierarchical decom-
position, it seems very difficult to increase the diameter of cells in a cluster. Our key observation
is that the cells are needed only for covering the portion of a solution inside a cluster that touches
the cluster boundary. Hence, we use the idea of adaptive cells. Specifically, for each connected
component A in the solution crossing a cluster C, we define the corresponding basic cells such that
if the component A has larger weight, then its corresponding basic cells (with respect to cluster C)
will have larger diameters. Combining with the notion of sparsity and bounded doubling dimension,
we can show that we only need to pay attention to a small number of cells.
Further Cells for Refinement. Since the dynamic program entries are defined in terms of the
hierarchical decomposition and the entries for a cluster are filled recursively with respect to those of
its child clusters, we would like the cells to have a refinement property, i.e., if a cluster C has some
cell e (which itself is some descendant cluster of C), then the child C ′ containing e has either e or
all children of e as its cells.
At first glance, a quick fix may be to push down each basic cell in C to its child clusters.
Although we could still bound the number of relevant cells, it would be difficult to bound the cost
to achieve the cell property. The reason is that the basic cells from higher levels are too large for
the descendant clusters. When more than one relevant component intersects such a large cell, we
need to add edges to connect the components. However, if the diameter of the cell is too large
compared to the cluster, these extra edges would be too costly.
We resolve this issue by introducing non-basic cells for a cluster: promoted cells and virtual
cells. These cells are introduced to ensure that every sibling of a basic cell is present. Moreover,
only non-basic cells of a cluster will be passed to its children. We show in Lemma 5.5 that the
total number of effective cells for a cluster is not too large. Moreover, Lemma 5.3 shows that the
refinement property still holds even if we only pass the non-basic cells down to the child clusters.
More importantly, we show that as long as we enforce the cell property for the basic cells, the cell
property for all cells are automatically ensured. This means that it is sufficient to bound the cost
to achieve the cell property with respect to only the basic cells.
Further Techniques: Global Cell Property. We note that the cell property in [BKM08] is
localized. In particular, for each cluster C, we restrict the solution inside C, which could have com-
ponents disconnected within C but are actually connected globally. In order to enforce the localized
cell property as in [BKM08], extra edges would need to be added for these locally disconnected
components. Instead, we enforce a global cell property, in which for every cell e in a cluster C,
there is only one (global) connected component in the solution that intersects e and crosses the
boundary of cluster C. A consequence of this is that if there are m components in the solution,
then at most m − 1 extra edges are needed to maintain the global cell property. This implication
is crucially used in our charging argument to bound the cost for enforcing the cell property for
the basic cells. However, this would imply that in the dynamic program entries, we need to keep
4
additional information on how the portals of a cluster are connected outside the cluster.
Combining the Ideas: A More Sophisticated Dynamic Program. Even though our ap-
proaches to tackle the encountered issues are intuitive, it is a non-trivial task to balance between
different tradeoffs and keep just enough information in the dynamic program entries, but still ensure
that the entries can be filled in polynomial time.
2 Preliminaries
We consider a metric space M = (X, d) (see [DL97, Mat02] for more details on metric spaces). For
x ∈ X and ρ ≥ 0, a ball B(x, ρ) is the set {y ∈ X d(x, y) ≤ ρ}. The diameter Diam(Z) of a
set Z ⊂ X is the maximum distance between points in Z. For S, T ⊂ X, we denote d(S, T ) :=
min{d(x, y) : x ∈ S, y ∈ T}, and for u ∈ X, d(u, T ) := d({u}, T ). Given a positive integer m, we
denote [m] := {1, 2, . . . , m}.
A set S ⊂ X is a ρ-packing, if any two distinct points in S are at a distance more than ρ away
from each other. A set S is a ρ-cover for Z ⊆ V , if for any z ∈ Z, there exists x ∈ S such that
d(x, z) ≤ ρ. A set S is a ρ-net for Z, if S is a ρ-packing and a ρ-cover for Z. We assume that a
ρ-net for any ball in X can be constructed efficiently.
We consider metric spaces with doubling dimension [Ass83, GKL03] at most k; this means that
for all x ∈ X, for all ρ > 0, every ball B(x, 2ρ) can be covered by the union of at most 2k balls
of the form B(z, ρ), where z ∈ X. The following fact captures a standard property of doubling
metrics.
Fact 2.1 (Packing in Doubling Metrics [GKL03]). Suppose in a metric space with doubling dimen-
sion at most k, a ρ-packing S has diameter at most R. Then, S ≤ ( 2R
ρ )k.
Given an undirected graph G = (V, E), where V V ⊂ X, E ⊆ (cid:0)V
2(cid:1), and an edge e = {x, y} ∈ E
receives weight d(x, y) from the metric space M . The weight w(G) or cost of a graph is the sum of
its edge weights. Let V (G) denote the vertex set of a graph G.
We consider the Steiner Forest Problem (SFP). Given a collection W = {(ai, bi) i ∈ [n]}
of terminal pairs in X, the goal is to find an undirected graph F (having vertex set in X) with
minimum cost such that each pair of terminals are connected in F . The non-terminal vertices in
V (F ) are called Steiner points.
Rescaling Instance. Fix constant ǫ > 0. Since we consider asymptotic running time to obtain
(1+ǫ)-approximation, we consider sufficiently large n > 1
ǫ . Suppose R > 0 is the maximum distance
between a pair of terminals. Then R is a lower bound on the cost of an optimal solution. Moreover,
the optimal solution F has cost at most nR, and hence, we do not need to consider distances larger
than nR. Since F contains at most 4n vertices, if we consider an ǫR
32n2 -net S for X and replace
every point in F with its closest net-point in S, the cost increases by at most ǫ · OPT. Hence,
after rescaling, we can assume that inter-point distance is at least 1 and we consider distances
up to O( n3
ǫ ) = poly(n). By the property of doubling dimension (Fact 2.1), we can hence assume
X ≤ O( n
ǫ )O(k) ≤ O(n)O(k).
k ≥ 4, where 0 < c < 1
Hierarchical Nets. As in [BGK12], we consider some parameter s = (log n)
is a universal constant that is sufficiently small (as required in Lemma 5.11). Set L := O(logs n) =
O( k log n
log log n ). A greedy algorithm can construct NL ⊆ NL−1 ⊆ ··· ⊆ N1 ⊆ N0 = N−1 = ··· = X
such that for each i, Ni is an si-net for X, where we say distance scale si is of height i.
Net-Respecting Solution. As defined in [BGK12], a graph F is net-respecting with respect to
{Ni}i∈[L] and ǫ > 0 if for every edge {x, y} in F , both x and y belong to Ni, where si ≤ ǫ· d(x, y) <
si+1.
c
5
Given an instance W of a problem, let OPT(W ) be an optimal solution; when the context is
clear, we also use OPT(W ) to denote the cost w(OPT(W )) as well; similarly, OPTnr(W ) refers to
an optimal net-respecting solution.
2.1 Overview
As in [BGK12, CJ16], we achieve a PTAS for SFP by the framework of sparse instance decompo-
sition.
Sparse Solution and Dynamic Program. Given a graph F and a subset S ⊆ X, FX is the
subgraph induced by the vertices in V (F ) ∩ X. A graph F is called q-sparse, if for all i ∈ [L] and
all u ∈ Ni, w(FB(u,3si)) ≤ q · si.
We show that for SFP (in Section 5) there is a dynamic program DP that runs in polynomial
time such that if an instance W has an optimal net-respecting solution that is q-sparse for some
small enough q, DP(W ) returns a (1 + ǫ)-approximation with high probability (at least 1− 1
poly(n) ).
Sparsity Heuristic. Since one does not know the optimal solution in advance, we estimate
the local sparsity with a heuristic. For i ∈ [L] and u ∈ Ni, given an instance W , the heuristic
H(i)
u (W ) is supposed to estimate the sparsity of an optimal net-respecting solution in the ball B′ :=
B(u, O(si)). We shall see in Section 3 that the heuristic actually gives a constant approximation
to some appropriately defined sub-instance W ′ in the ball B′.
Generic Algorithm. We describe a generic framework that applies to SFP. Similar framework is
also used in [CJ16, BGK12] to obtain PTAS's for TSP related problems. Given an instance W , we
describe the recursive algorithm ALG(W ) as follows.
force, recalling that X ≤ O( n
ǫ )O(k).
2. Sparse Instance.
1. Base Case. If W = n is smaller than some constant threshold, solve the problem by brute
u (W ) is at most q0 · si, for some
3. Identify Critical Instance. Otherwise, let i be the smallest height such that there exists
u (W ) is
appropriate threshold q0, call the subroutine DP(W ) to return a solution, and terminate.
If for all i ∈ [L], for all u ∈ Ni, H(i)
u ∈ Ni with critical H(i)
u (W ) > q0 · si; in this case, choose u ∈ Ni such that H(i)
maximized.
4. Decomposition into Sparse Instances. Decompose the instance W into appropriate sub-
instances W1 and W2 (possibly using randomness). Loosely speaking, W1 is a sparse enough
sub-instance induced in the region around u at distance scale si, and W2 captures the rest. We
note that H(i)
u (W2) ≤ q0 · si such that the recursion will terminate. The union of the solutions
to the sub-instances will be a solution to W . Moreover, the following property holds.
E[OPT(W1)] ≤
1 − ǫ · (OPTnr(W ) − E[OPTnr(W2)]),
1
where the expectation is over the randomness of the decomposition.
(1)
5. Recursion. Call the subroutine F1 := DP(W1), and solve F2 := ALG(W2) recursively; return
the union F1 ∪ F2.
Analysis of Approximation Ratio. We follow the inductive proof as in [BGK12] to show that
with constant probability (where the randomness comes from DP), ALG(W ) returns a tour with
expected length at most 1+ǫ
1−ǫ·OPTnr(W ), where expectation is over the randomness of decomposition
into sparse instances in Step 4.
As we shall see, in ALG(W ), the subroutine DP is called at most poly(n) times (either explicitly
in the recursion or the heuristic H(i)). Hence, with constant probability, all solutions returned by
all instances of DP have appropriate approximation guarantees.
6
1−ǫ · OPTnr(W2).
In Step 4, equation (1) guarantees that E[OPT(W1)] ≤ 1
Suppose F1 and F2 are solutions returned by DP(W1) and ALG(W2), respectively. Since we
assume that W1 is sparse enough and DP behaves correctly, w(F1) ≤ (1 + ǫ) · OPT(W1). The
induction hypothesis states that E[w(F2)W2] ≤ 1+ǫ
1−ǫ · (OPTnr(W ) − E[OPTnr(W2)]).
Hence, it follows that E[w(F1) + w(F2)] ≤ 1+ǫ
1−ǫ · OPTnr(W ) = (1 + O(ǫ)) · OPT(W ), achieving the
desired ratio.
Analysis of Running Time. As mentioned above, if H(i)
u (W ) is found to be critical, then in the
decomposed sub-instances W1 and W2, H(i)
u (W2) should be small. Hence, it follows that there will
be at most X · L = poly(n) recursive calls to ALG. Therefore, as far as obtaining polynomial
running times, it suffices to analyze the running time of the dynamic program DP. The details are
in Section 5.3.
2.2 Paper Organization
In order to apply the above framework to obtain a PTAS for SFP, we shall describe in details the
following components.
1. (Section 3.) Design a heuristic H such that for each i ∈ [L] and u ∈ Ni, the heuristic H(i)
2. (Section 4.) When a critical H(i)
gives an upper bound for OPTnr(W )B(u,3si).
u (W ) is found, decompose W into instances W1 and W2 such
u (W )
that equation (1) holds.
3. (Section 5.) Design a dynamic program DP that gives (1+ǫ)-approximation to sparse instances
in polynomial time.
3 Sparsity Heuristic for SFP
u
u
Suppose a collection W of terminal pairs is an instance of SFP. For i ∈ [L] and u ∈ Ni, recall that
we wish to estimate OPTnr(W )B(u,3si) with some heuristic H(i)
u (W ). We consider a more general
heuristic T(i,t)
associated with the ball B(u, tsi), for t ≥ 1. The following auxiliary sub-instance
deals with terminal pairs that are separated by the ball.
Auxiliary Sub-Instance. Fix δ := Θ( ǫ
k ), where the constant depends on the proof of Lemma 4.2.
For i ∈ [L], u ∈ Ni and t ≥ 1, the sub-instance W (i,t)
is induced by each pair {a, b} ∈ W as follows.
(a) If both a, b ∈ B(u, tsi), or if exactly one of them is in B(u, tsi) and the other in B(u, (t + δ)si),
(b) Suppose j is the index such that sj < δsi ≤ sj+1. If a ∈ B(u, tsi) and b /∈ B(u, (t + δ)si),
(c) If both a and b are not in B(u, tsi), then the pair is excluded.
then {a, b} is also included in W (i,t)
then {a, a′} is included in W (i,t)
Defining Heuristic. We define H(i)
(W ) in terms of a more general heuristic, where
(W ) is the cost of a constant approximate net-respecting solution of SFP on the instance W (i,t)
T(i,t)
.
For example, we can first apply the primal-dual algorithm in [GW95] that gives a 2-approximation
of SFP, and then make it net-respecting and we have T(i,t)
, where a′ is the nearest point to a in Nj.
u (W ) := T(i,4)
).
u
u
u
.
u
(W ) ≤ 2(1 + Θ(ǫ)) · OPT(W (i,t)
u
One potential issue is that OPTnr(W ) might use Steiner points in B(u, tsi), even if W (i,t)
is
empty. We shall prove a structural property of Steiner tree in Lemma 3.1, and Lemma 3.1 implies
Lemma 3.2 which helps us to resolve this issue. Recall that the Steiner tree problem is a special
case of SFP where the goal is to return a minimum cost tree that connects all terminals.
u
u
u
7
Lemma 3.1 (Distribution of Steiner Points in The Optimal Steiner Tree). Suppose S is a terminal
set with Diam(S) ≤ D, and suppose F is an optimal Steiner tree with terminal set S. If the longest
edge in F has weight at most γD (0 < γ ≤ 1), then for any Steiner point r in F , d(r, S) ≤
4kγ log2
4
γ · D.
Proof. Since F is an optimal solution, all Steiner points in F have degree at least 3. Fix any Steiner
point r in F .
Denote K := ⌈log2(γD)⌉. Suppose we consider r as the root of the tree F . We shall show
that there is a path of small weight from r to some terminal. Without loss of generality, we can
assume that all terminals are leaves, because once we reach a terminal, there is no need to visit its
descendants. For simplicity, we can assume that each internal node (Steiner point) has exactly two
children, because we can ignore extra branches if an internal has more than two children.
u = M (i)
u1 + M (i)
For i ≤ K, let Ei be the set of edges in F that have weights in the range (2i−1, 2i], and we say
that such an edge is of type i. For each node u in F , denote Fu as the subtree rooted at u. Suppose
we consider Fu and remove all edges in ∪j≥iEj from Fu; in the resulting forest, let M (i)
u be the
number of connected components that contain at least one terminal. We shall prove the following
statement by structural induction on the tree bF .
For each node u ∈ F , there exists a leaf x ∈ Fu such that d(x, u) ≤ Pi≤K 2i log2 M (i)
u .
u2 . Then, we can pick x1 to be the desired leaf, because the extra distance
for j 6= i. More precisely,
u1 ≤ Pj≤K 2j log2 M (j)
u ,
Base Case. If u is a leaf, then the statement is true.
Inductive Step. Suppose u has children u1 and u2 such that {u, u1} ∈ Ei and {u, u2} ∈ Ei′,
where i ≥ i′. Suppose x1 and x2 are the leaves in Fu1 and Fu2, respectively, from the induction
hypothesis. Observe that M (i)
(1) Suppose M (i)
d(u1, u) ≤ 2i can be accounted for, as 2M (i)
u1 ≤ M (j)
u1 ≤ M (i)
d(x1, u) ≤ d(x1, u1) + d(u1, u) ≤ 2i · (1 + log2 M (i)
where the second inequality follows from the induction hypothesis for u1.
(2) Suppose M (i)
distance is d(u2, u) ≤ 2i′ ≤ 2i. This completes the inductive step.
Next, it suffices to give an upper bound for each M (i) := M (i)
for root r. Suppose after removing
r
all tree edges in ∪j≥iEj, P and Q are two clusters each containing at least one terminal. Then,
observe that the path in F connecting P and Q must contain an edge e with weight at least 2i−1. It
follows that d(P, Q) ≥ 2i−1; otherwise, we can replace e in F with another edge of length less than
2i−1 to obtain a Steiner tree with strictly less weight. It follows that each cluster has a terminal
representative that form a 2i−1-packing. Hence, we have M (i) ≤ ( 4D
2i )k, by the packing property of
doubling metrics (Fact 2.1).
u , and M (j)
u1 ) +Pj≤K:j6=i 2j log2 M (j)
u1 . Then, similarly we pick x2 to be the desired leaf, because the extra
u2 . We consider two cases.
u1 ≤ M (i)
u2 < M (i)
u
Therefore, every Steiner point r in bF has a terminal within distance Pi≤K k · 2i log2
4D
2i ≤
4kγD log2
4
γ .
Given a graph F , a chain in F is specified by a sequence of points (p1, p2, . . . , pl) such that there
is an edge {pi, pi+1} in F between adjacent points, and the degree of an internal point pi (where
2 ≤ i ≤ l − 1) in F is exactly 2.
Lemma 3.2 (Steiner Tree of Well-Separated Terminals Contains A Long Chain). Suppose S and T
are terminal sets in a metric space with doubling dimension at most k such that Diam(S ∪ T ) ≤ D,
and d(S, T ) ≥ τ D, where 0 < τ < 1. Suppose F is an optimal net-respecting Steiner tree connecting
4096k2 ·D such that any internal
the points in S∪ T . Then, there is a chain in F with weight at least
point in the chain is a Steiner point.
τ 2
8
8
distance at most 4kγ log2
defined to be the weight of the corresponding chain.
use the optimality of the solutions. Specifically, in Lemma 3.1 we use the fact that when an edge
e connects point sets P and Q that both contain at least one terminal (i.e. removing e results in
Proof. Denote γ := τ 2
4096k2 . Suppose for contradiction's sake that all chains in F have weight less
than γD. We consider a minor bF that is obtained from F by merging Steiner points of degree 2
with adjacent points. Hence, the vertex set of bF are the terminals together with Steiner points in
F with degree at least 3. Moreover, an edge in bF corresponds to a chain in F , and its weight is
Then by using the argument in Lemma 3.1, We can prove that every point u in bF is within
γ · D to a terminal. Precisely, we shall replace the F in the argument
of Lemma 3.1 with bF . We observe that the only difference caused by this replacement is when we
the dis-connectivity of P and Q), it has to be d(P, Q) ≥ w(e), while the corresponding fact for bF
is d(P, Q) ≥ w(e)
Obtaining Contradiction. Recall that the terminal sets S and T are well-separated d(S, T ) ≥
τ D. Since all Steiner points in bF are at distance at most 4kγ log2
γ ·D from the terminals, it follows
γ ·D > τ D−32k√γD > γD.
that there must be an edge in bF with length at least τ D−8kγ log2
any i and u ∈ Ni and t ≥ 1, w(FB(u,tsi)) ≤ T(i,t+1)
Lemma 3.3. Suppose F is an optimal net-respecting solution for an SFP instance W . Then, for
2 because of the net-respecting property.
1+Θ(ǫ) ≥ w(e)
(W ) + O( skt
ǫ )O(k)si.
Proof. Given an optimal net-respecting solution F , we shall construct another net-respecting solu-
tion in the following steps.
u
8
8
1. Remove edges in FB(u,tsi).
2. Add edges corresponding to the heuristic T(i,t+1)
3. Add edges in a minimum spanning tree H of Nj ∩ B(u, (t + 2)si), where sj ≤ Θ(
(t+1)k2 )· si <
sj+1, where the constant in Theta depends on Lemma 3.2; convert each added edge into
a net-respecting path if necessary. Observe that the weight of edges added in this step is
O( stk
(W ).
u
ǫ
ǫ )O(k) · si.
4. To ensure feasibility, replace some edges without increasing the weight.
If we can show that the resulting solution is feasible for W , then the optimality of F implies
u
and
Feasibility.
(W ). If a ∈ bB and b /∈ bB, then edges for the heuristic T(i,t+1)
the result. We denote B := B(u, tsi) and bB := B(u, (t + 1)si).
Define bV1 := {x : x ∈ B ∃{x, y} ∈ F s.t. y /∈ B and y is connected in FX\B to some point outside bB},
bV2 := {x : x ∈ bB \ B x is connected in F bB to some point in bV1,∃{x, y} ∈ F s.t. y /∈ bB}. In Step
4, we will ensure that all points in bV1 ∪ bV2 are connected to the MST H.
If a pair {a, b} ∈ W has both terminals in bB, then they will be connected by the edges corre-
sponding to T(i,t+1)
(W ) ensures that
a is connected to H; moreover, in the original tree F , if the path from a to b does not meet any
node in bV2, then this path is preserved, otherwise there is a portion of the path from a point in bV2
to b that is still preserved. If both a and b are outside bB, then they might be connected in F via
points in bV2; however, since all points in bV2 are connected to H, feasibility is ensured.
We next elaborate how Step 4 is performed. Consider a connected component U in FbV1∪( bB\B)
that contains a point in bV1. Let S1 := U ∩ bV1 and S2 := U ∩ bV2. If S2 = ∅, then there is an edge
connecting S1 directly to a point outside bB. This means that both its end-points are in Nj by the
Next, if there is a point z /∈ bB connected directly to some point y ∈ S2 such that d(y, z) ≥ si
2 ,
then by the net-respecting property, y ∈ Nj and so again U is connected to H. Otherwise, we
net-respecting property, and hence S1 is already connected to H.
u
9
have d(S1, S2) ≥ si
S1 ∪ S2. Since U itself is net-respecting, this does not increase the cost.
a chain in bU from some point u to v such that its length is at least Θ(
2 . We next replace U with an optimal net-respecting Steiner tree bU connecting
Observing that Diam(S1 ∪ S2) ≤ 2(t + 1)si, we can use Lemma 3.2 to conclude that there exists
k2(t+1) ) · si. Hence, we
can remove this chain, and use its weight to add a net-respecting path from each of u and v to its
nearest point in Nj. This does not increase the cost, and ensures that both S1 and S2 are connected
to H.
1
Therefore, we have shown that Step 4 ensures that all points in cV1 and cV2 are connected to
It is because of Lemma 3.3 that we choose H(i)
(W ) to be the heuristic.
u (W ) := T(i,4)
u
H.
Corollary 3.1 (Threshold for Critical Instance). Suppose F is an optimal net-respecting solution
for an SFP instance W , and q ≥ Θ( sk
u (W ) ≤ qsi, then F is
2q-sparse.
Lemma 3.4. Suppose W is an SFP instance. Consider i ∈ [L], u ∈ Ni, and t ≥ t′ ≥ 1. Suppose
F is a net-respecting solution for W (i,t)
ǫ )Θ(k). If for all i ∈ [L] and u ∈ Ni, H(i)
. Then, T(i,t′)
(W ) ≤ 4(1 + ǫ) · w(F ) + O( skt′
ǫ )O(k)si.
u
u
u
u
u
u
u
ǫ )O(k) · si.
t′k2 ) · si < sj+1.
ǫ )O(k)si. Then, the heuristic T(i,t′)
ǫ )O(k)si.
We first include F in the solution. It suffices to handle the terminal pairs in W (i,t′)
with weight at most 2 · w(F ) +
(W ) gives the weight of a net-respecting solution with cost
Proof. We first show that there is a feasible solution for W (i,t′)
O( skt′
at most 4(1 + ǫ) · w(F ) + O( skt′
\ W (i,t)
.
Such a pair {a, a′} must be induced from {a, b} ∈ W (i,t)
such that a ∈ B(u, t′si) and b ∈ B(u, (t +
δ)si) \ B(u, (t′ + δ)si). We next add more edges such that a is connected to a′, which lies in Nj,
where sj ≤ Θ( δ2
We add a minimum spanning tree H on the points in Nj ∩ B(u, (t′ + δ)si). This has cost at
most O( skt′
Consider a connected component U of F . Consider the terminal pairs {a, b} ∈ W (i,t)
connected
by U such that a ∈ B(u, t′si) and b ∈ B(u, (t + δ)si); let S1 be those terminals a's, and S2 be those
terminal b's. Suppose bU is an optimal net-respecting Steiner tree connecting S1 ∪ S2. Since U is
also net-respecting, it follows that the weight of bU is at most that of U .
Since d(S1, S2) ≥ δsi and Diam(S1 ∪ S2) ≤ Θ(t′)si, it follows from Lemma 3.2 that there exists
a chain from p to q in bU with weight at least Θ( δ2
t′k2 )· si. Hence, we can remove this chain, and use
this weight to connect p and q to each of their closest points in Nj. This ensures that each point
a ∈ S1 is connected to its closest point in Nj via the minimum spanning tree H.
is at most w(F ). Hence, we have shown that there is a feasible solution to W (i,t′)
2 · w(F ) + O( skt′
If we perform this operation on each connected component U of F , the weight of edges added
with cost at most
ǫ )O(k)si, as required.
u
u
4 Decomposition into Sparse Instances
In Section 3, we define a heuristic H(i)
u (W ) to detect a critical instance around some point u ∈ Ni
at distance scale si. We next describe how the instance W can be decomposed into W1 and W2
such that equation (1) in Section 2.1 is satisfied.
Since the ball centered at u with radius around si could potentially separate terminal pairs in
W , we use the idea in Section 3 for defining the heuristic to decompose the instance.
10
ǫ )Θ(k) according to Corol-
Decomposing a Critical Instance. We define a threshold q0 := Θ( sk
lary 3.1. As stated in Section 2.1, a critical instance is detected by the heuristic when a smallest
i ∈ [L] is found for which there exists some u ∈ Ni such that H(i)
(W ) > q0si. More-
over, in this case, u ∈ Ni is chosen to maximize H(i)
u (W ). To achieve a running time with an
exp(O(1)k log(k)) dependence on the doubling dimension k, we also apply the technique in [CJ16]
to choose the cutting radius carefully.
u (W ) = T(i,4)
u
Claim 4.1 (Choosing Radius of Cutting Ball). Denote T(λ) := T(i,4+2λ)
0 ≤ λ < k such that T(λ + 1) ≤ 30k · T(λ).
Proof. Suppose the contrary is true. Then, it follows that T(k) > (30k)k · T(0). We shall obtain
a contradiction by showing that there is a solution for the instance W (i,4+2k)
corresponding to
T(k) = T(i,4+2k)
(W ). Then, there exists
(W ) with small weight.
u
u
u
Define N ′
i to be the set of points in Ni that cover B(u, (2k + 5)si), and similarly define N ′
j,
u
v over v ∈ N ′
j · 2(2k + 5) · si + N ′
Define edge set F to be the union of a minimum spanning tree on N ′
i. It follows that F is a feasible solution for the instance W (i,4+2k)
j together with the union of
. By
i · T(0) ≤ q0si + (4k + 10)k · T(0) ≤
Hence, we have an upper bound for the heuristic T(k) ≤ 2(1 + Θ(ǫ))· w(F ) ≤ (30)k · T(0), which
where sj ≤ δ · si ≤ sj+1.
the edge sets H(i)
the choice of u and q0, we have w(F ) ≤ N ′
(15k)k · T(0).
gives us the desired contradiction.
Cutting Ball and Sub-Instances. Suppose λ ≥ 0 is picked as in Claim 4.1, and sample h ∈ [0, 1
2 ]
k ). Define B := B(u, (4 + 2λ + h)si) and bB := B(u, (4 +
uniformly at random. Recall that δ := Θ( ǫ
2λ + h + δ)si). The instances W1 and W2 are induced by each pair {a, b} ∈ W as follows.
(a) If a ∈ B and b ∈ bB, then include {a, b} in W1.
(b) If a ∈ B and b /∈ bB, then include {a, a′} in W1 and {a′, b} in W2, where a′ is the closest point
(c) If both a and b are not in B, then include {a, b} in W2.
Lemma 4.1 (Sub-Instances Are Sparse). The sub-instances W1 and W2 satisfy the following.
in Nj to a and sj ≤ δ · si < sj+1.
(i) If F1 is feasible for W1 and F2 is feasible for W2, then the union F1 ∪ F2 is feasible for W .
(ii) The sub-instance W2 does not have a critical instance with height less than i, and H(i)
u (W2) =
(iii) H(i)
0.
u (W1) ≤ O(s)O(k) · q0 · si.
Proof. The first two statements follow immediately from the construction. For the third statement,
we use the fact that there is no critical instance at height i − 1 to show that there is a solution to
W1 with small cost.
Specifically, we consider a minimum spanning tree H on Nj∩B(u, 5si), where sj ≤ δ·si−1 < sj+1.
Moreover, we consider the union of solutions corresponding to H(i−1)
(W ), over v ∈ Ni−1 ∩
v
Then, we have w(H) ≤ q0 · si.
B(u, 5si). The cost is O(s)O(k) · q0 · si.
implies that H(i)
u (W1) ≤ O(s)O(k) · q0 · si.
Hence, the union of H together with the edges for the H(i−1)
v
(W )'s is feasible for W1, and this
Lemma 4.2 (Combining Costs of Sub-Instances). Suppose F is an optimal net-respecting solution
for W . Then, for any realization of the decomposed sub-instances W1 and W2 as described above,
11
there exist net-respecting solutions F1 and F2 for W1 and W2, respectively, such that (1 − ǫ) ·
E[w(F1)] + E[w(F2)] ≤ w(F ), where the expectation is over the randomness to generate W1 and
W2.
Proof. Let B and bB be defined as above, and denote B := B(u, (4+2λ+1)·si). Hence, B ⊂ bB ⊂ B.
We start by including FB in T1, and including the remaining edges in F in F2. We will then
show how to add extra edges with expected weight at most ǫ· E[w(F1)] to make F1 and F2 feasible.
This will imply the lemma.
Define N to be the subset of Nj that cover the points in B, where sj < δsi ≤ sj+1. We include
a copy of a minimum spanning tree H of N in each of F1 and F2, and make it net-respecting. This
costs at most N · O(k) · si ≤ O( ks
cost at most δ · w(FB).
Connecting Crossing Points. To ensure the feasibility of F1, we connect the following sets of
points to N . We denote:
We next include the edges of F in the annulus bB \ B (of width δ) into F1. This has expected
ǫ )O(k) · si.
is incident to some edge in F with weight at least si
V1 := {x ∈ B ∃y ∈ bB \ B,{x, y} ∈ F}, V2 := {y ∈ bB \ B ∃x ∈ B,{x, y} ∈ F}, and
V3 := {x ∈ bB ∃y /∈ bB,{x, y} ∈ F}.
We shall connect each point in V1 ∪ V2 ∪ V3 to its closest point in N . Note that if such a point x
4 , then the net-respecting property of F implies
that x is already in N . Otherwise, this is because some edge {x, y} in F is cut by either B or bB,
which happens with probability at most O( d(x,y)
). Hence, each edge {x, y} ∈ FB has an expected
contribution of δsi · O( d(x,y)
to N . Denote cV1 := {x ∈ B ∃y /∈ B,{x, y} ∈ F}. By the same argument, the expected cost to
connect each point to N is also at most O(δ) · w(FB).
Charging the Extra Costs to F1. Apart from using edges in F , the extra edges come from two
copies of the minimum spanning tree H, and other edges with cost O(δ)· w(FB ). We charge these
extra costs to F1.
, by Lemma 3.4,
Similarly, to ensure the feasibility of F2, we ensure each point in the following set is connected
) = O(δ) · d(x, y).
(W ) > q0 · si and F1 is a net-respecting solution for W (i,4+2λ+h)
8 · si, by choosing large enough q0.
ǫ )O(k) · si) > q0
4(1+ǫ) (T (i)(u, 4) − O( sk
w(F1) ≥ 1
Since T (i,4)
si
si
u
u
ǫ
2 · w(F1).
Therefore, the cost for the two copies of the minimum spanning tree H is at most O( ks
We next give an upper bound on w(FB), which is at most T(i,4+2(λ+1))
ǫ )O(k)·si ≤
ǫ )O(k) · si,
(W ) ≤ 30k · T(i,4+2λ+1)
by Lemma 3.3. By the choice of λ, we have T(i,4+2(λ+1))
(W ). Moreover,
by Lemma 3.4, T(i,4+2λ+1)
ǫ )O(k) · si. Hence, we can conclude that
(W ) ≤ 4(1 + ǫ) · w(F1) + O( sk
w(FB) ≤ O(k) · w(F1).
k ), we can conclude that the extra costs O(δ)·w(FB ) ≤
2 · w(F1).
Therefore, we have shown that E[w(F1)] + E[w(F2)] ≤ w(F ) + ǫ · w(F1), where the right hand
side is a random variable. Taking expectation on both sides and rearranging gives the required
result.
Hence, by choosing small enough δ = Θ( ǫ
(W ) + O( sk
u
u
u
u
ǫ
5 A PTAS for Sparse SFP Instances
Our dynamic program follows the divide and conquer strategy as in previous works on TSP [Aro98,
Tal04, BGK12] that are based on hierarchical decomposition. However, to apply the framework to
12
SFP, we need a version of the cell property that is more sophisticated than previous works [BKM08,
BH12].
We shall first give a review of the hierarchical decomposition techniques in Section 5.1. Then
in Section 5.2, we shall define our cell property precisely, and also prove that there exist good
solutions that satisfy the cell property (in Lemma 5.6). Finally, we shall define DP in Section 5.3,
and conclude a PTAS for sparse SFP instances (in Corollary 5.2).
5.1 Review on Hierarchical Decomposition
Definition 5.1 (Single-Scale Decomposition [ABN06]). At height i, an arbitrary ordering πi is
imposed on the net Ni. Each net-point u ∈ Ni corresponds to a cluster center and samples random
χ−1 · ln χ
hu from a truncated exponential distribution Expi having density function t 7→ χ
for
t ∈ [0, si], where χ = O(1)k. Then, the cluster at u has random radius ru := si + hu.
The clusters induced by Ni and the random radii form a decomposition Πi, where a point p ∈ V
belongs to the cluster with center u ∈ Ni such that u is the first point in πi to satisfy p ∈ B(u, ru).
We say that the partition Πi cuts a set P if P is not totally contained within a single cluster.
The results in [ABN06] imply that the probability that a set P is cut by Πi is at most β·Diam(P )
,
· e− t ln χ
si
si
si
where β = O(k).
Definition 5.2 (Hierarchical Decomposition). Given a configuration of random radii for {Ni}i∈[L],
decompositions {Πi}i∈[L] are induced as in Definition 5.1. At the top height L − 1, the whole space
is partitioned by ΠL−1 to form height-(L − 1) clusters. Inductively, each cluster at height i + 1 is
partitioned by Πi to form height-i clusters, until height 0 is reached. Observe that a cluster has
K := O(s)k child clusters. Hence, a set P is cut at height i iff the set P is cut by some partition
Πj such that j ≥ i; this happens with probability at most Pj≥i
β·Diam(P )
si
= O(k)·Diam(P )
si
.
Portals. As in [Aro02, Tal04, BGK12], each height-i cluster U is equipped with portals such that a
solution F is portal-respecting, if for every edge {x, y} in F between a point x in U and some point
y outside U , at least one of x and y must be a portal of cluster U . As mentioned in [BGK12], the
portals of a cluster need not be points of the cluster itself, but are just used as connection points.
For a height-i cluster C, its portals is the subset of net-points in Ni′ that cover C, where i′ is the
maximum index such that si′ ≤ max{1,
4βL · si}. As noted in [Tal04, BGK12, CJ16], any solution
can be made to be portal-respecting with a multiplicative factor of 1 + O(ǫ) in cost.
ǫ )k
Since a height-i cluster has diameter O(si), by Fact 2.1, the cluster has at most m := O( βLs
portals.
(m, r)-Light Solution. A solution F is called (m, r)-light, if it is portal-respecting for a hierarchical
decomposition in which each cluster has at most m portals, and for each cluster, at most r of its
portals are used in F to connect points in the cluster to the points outside.
ǫ
5.2 Structural Property
In this section, we shall define the cell property (Definition 5.13) with respect to the effective cells
(Definition 5.9), where the effective cells are carefully chosen to implement our adaptive cells idea
which is discussed in Section 1. Specifically, the effective cells are defined by the union of the basic
cells (Definition 5.5) and the non-basic cells (Definition 5.8). Moreover, the virtual cells and the
promoted cells (Definition 5.7) are introduced in order to define the non-basic cells. Finally, we
shall prove the structural property in Lemma 5.6.
Notations and Parameters. Let ht(C) denote the height of a cluster C, des(C) denote the
collection of all descendant clusters of C (including C), and par(C) denote the parent cluster of C.
13
For x ∈ R+, let ⌊x⌋s denote the largest power of s that is at most x, and ⌈x⌉s denote the smallest
s2 ). Define γ0 such that
power of s that is at least x. Define γ0 := Θ(
1
:= ⌊ 1
γ0
γ0⌉s, and define γ1 such that 1
γ1⌋s. We note that γ0 < γ1.
ks2L ), and define γ1 := Θ( ǫ
:= ⌈ 1
γ1
ǫ
Definition 5.3 (Cell). Suppose C is a cluster of height i. A p-cell of C is a height-logs p sub-cluster
of C.
Definition 5.4 (Crossing Component). Suppose C is some cluster, and F is a solution for SFP.
We say that a subset A crosses C, if there exists points x, y ∈ A such that x ∈ C and y /∈ C. A
component A in F is called a crossing component of C if A crosses C.
In the following, we shall introduce the notions of the basic cells, owner of basic cells, promoted
cells, virtual cells, non-basic cells, and effective cells. All of these are defined with respect to some
feasible solution to SFP. We assume there is an underlying feasible solution F when talking about
these definitions.
Adaptive Cells. For each cluster C, we shall define its basic cells whose heights depend on the
weights l of the crossing components of C in the solution F . We consider three cases.
si ≤ ⌊l⌋s < si} and I3(l) := {i i ≤ L,⌊l⌋s < γ0
γ1
si}.
γ1
Define I1(l) := {i ⌊l⌋s ≥ si}, I2(l) := {i γ0
Define a function h : [L] × R+ → R+, such that
γ1si,
γ1⌊l⌋s,
γ0si,
≤ h(i, l) ≤ h(i + 1, l).
Lemma 5.1. h(i+1,l)
h(i, l) =
s
for i ∈ I1(l)
for i ∈ I2(l)
for i ∈ I3(l)
Proof. If both i and i + 1 lie in the same Ij(l) (j ∈ {1, 2, 3}), then it holds immediately.
Otherwise, it is either i ∈ I2(l) but i + 1 ∈ I3(l), or i ∈ I1(l) but i + 1 ∈ I2(l).
• If i ∈ I1(l) and i+1 ∈ I2(l). This implies si = ⌊l⌋s. Hence, h(i, l) = γ1si = γ1⌊l⌋s = h(i+1, l).
• If i ∈ I2(l) and i + 1 ∈ I3(l). This implies si = γ1
γ0⌊l⌋s. Hence, s· h(i, l) = s· γ1⌊l⌋s = γ0si+1 =
h(i + 1, l).
This implies the inequality.
Definition 5.5 (Basic Cell). Suppose C is a cluster of height i, and A is a crossing component
of C. Define l := w(A). Define the basic cells of A in C, BasA(C), to be the collection of the
h(i, l)-cells of C that intersect A. Define the basic cells of C, Bas(C), to be the union of BasA(C)
for all crossing components A of C.
Definition 5.6 (Owner of a Basic Cell). For some cluster C, define the owner of e ∈ Bas(C) to
be the minimum weight crossing component A such that e ∈ BasA(C).
Definition 5.7 (Promoted Cell and Virtual Cell). Suppose C is a cluster of height i. Let S be the
set of sub-clusters of C that is not in Bas(C) but has a sibling in Bas(C).
Consider each e ∈ S.
• If there exists a sub-cluster C ′ of C such that e ∈ Bas(C ′), then define Proe(C) := des(e) ∩
Bas(C ′), and define Vire(C) := ∅, where C ′ ⊂ C is any one that satisfies e ∈ Bas(C ′).
• Otherwise, define Proe(C) := ∅, and define Vire(C) := e.
Finally, Pro(C) := Se∈S Proe(C), and Vir(C) := Se∈S Vire(C), and elements in Pro(C) and
Vir(C) are called promoted cells and virtual cells respectively.
14
Lemma 5.2. For any cluster C, if e ∈ Vir(C), then for any cluster C ′ ⊂ C (C ′ may equal C),
e\{e′ ∈ Bas(C ′) e′ ( e} has no intersection with any crossing component of C ′.
Proof. Suppose not. Then, there exists a cluster C ′ ⊂ C, and a crossing component A of C ′, such
that A intersects u := e\{e′ ∈ Bas(C ′) e′ ( e}. This implies that there exists u′ ∈ BasA(C ′), such
that e ⊂ u′. By Lemma 5.1, and the fact that h(ht(C ′), w(A)) ≥ sht(e) and that h(0, w(A)) < sht(e),
we know that there exists a cluster C ′′ ⊂ C ′ ⊂ C, such that e ∈ BasA(C ′′). This contradicts with
the definition of virtual cells.
Definition 5.8 (Non-basic Cell). We define the non-basic cells NBas(C) for a cluster C. If C
is the root cluster, then NBas(C) = Pro(C) ∪ Vir(C)\Bas(C). For any other cluster C, define
NBas(C) := {e ∩ C e ∈ Pro(C) ∪ Vir(C) ∪ NBas(par(C))\Bas(C)}.
Definition 5.9 (Effective Cell). For a cluster C, define the effective cells of C as Eff(C) :=
Bas(C) ∪ NBas(C).
Definition 5.10 (Refinement). Suppose S1 and S2 are collections of clusters. We say S1 is a
refinement of S2, if for any e ∈ S2, either e ∈ S1, or all child clusters of e are in S1.
Lemma 5.3. Suppose C is a cluster that is not a leaf. Define {Ci}i to be the collection of all the
child clusters of C. Then Si Eff(Ci) is a refinement of Eff(C).
Proof. Define S := Si Eff(Ci). It is sufficient to prove that for any e ∈ Eff(C), either e ∈ S, or all
child clusters of e are in S.
If e ∈ NBas(C) and e 6= C, then e ∈ S follows from Definition 5.8 and Definition 5.9.
If
e ∈ NBas(C) but e = C, then also by Definition 5.8 and Definition 5.9, C ∩ Ci = Ci ⊂ Eff(Ci), and
this implies that all child clusters of e are in S.
Otherwise, e ∈ Bas(C), then by Lemma 5.1, we know that either e ∈ S, or there exists e′ ⊂ e
such that ht(e′) = ht(e)−1 and e′ ∈ S. Then all siblings of e′ are in S, by the definition of promoted
cells and virtual cells. This implies that all child clusters of e are in S.
Definition 5.11 (Candidate Center). Suppose C is a cluster of height i. The set of candidate
0 si Nj that may become a center of C's
centers of C, denoted as Can(C), is the subset of Si
j=logs γ2
child cluster in the hierarchical decomposition.
Lemma 5.4. For any cluster C, the centers of clusters in Eff(C) are chosen from Can(C), and
Can(C) ≤ κ, where κ := O( 1
Proof. We first prove that centers of cluster in Eff(C) are chosen from Can(C).
)O(k).
γ0
• For e ∈ Bas(C), by the definition of the basic cells, we have ht(e) ≥ logs γ0si.
• For e ∈ Pro(C), we have that e is a basic cell of some cluster C ′, and hence ht(e) ≥ logs γ2
• For e ∈ Vir(C), since it is a sibling of a basic cell, so ht(e) ≥ logs γ0si.
• For e ∈ NBas(C), there is a cluster C ′′ such that C ⊂ C ′′ and e ∈ Pro(C ′′) ∪ Vir(C ′′).
Hence ht(e) ≥ logs γ2
We then bound Can(C). Suppose i := ht(C). Observe that a center of height j ≤ i that may
become a center of a child cluster of C are contained in a ball of diameter O(si). Moreover, Nj is
an sj packing. Hence, by packing property, Can(C) ≤ O( 1
Lemma 5.5. Suppose Eff is defined in terms of a solution that is (m, r)-light. Then for each cluster
C, Eff(C) ≤ ρ, where ρ := O(logs
0si. Therefore, centers of clusters in Eff(C) are in Can(C).
) · r2 · O( s
)O(k).
0 si.
)O(k).
γ0
1
γ0
γ1
15
Pro(C),
Proof. Suppose C is of height i. We give upper bounds for Bas(C),
Vir(C) and
NBas(C) respectively.
Bounding Bas(C). Fix a crossing component A of C, and suppose l := w(A). We upper bound
BasA(C).
• If i ∈ I1(l), then BasA(C) is a subset of γ1si-cells of C. By packing property, BasA(C) ≤
O( 1
γ1
• If i ∈ I2(l), then BasA(C) is a subset of γ1⌊l⌋s-cells of C. Since all the γ1⌊l⌋s-cells that
)O(k).
• If i ∈ I3(l), then BasA(C) is a subset of γ0si-cells of C. Since all the γ0si-cells that intersect
intersect A are inside a ball of diameter O(l), by packing property, BasA(C) ≤ O( s
A are inside a ball of diameter O( γ0
γ1
si), by packing property, BasA(C) ≤ O( 1
Since the solution is r-light, there are at most r crossing components. Therefore,
)k.
)k.
γ1
γ1
Bas(C) ≤ r · O(
s
γ1
)O(k).
Bounding Pro(C) and Vir(C). Recall that for e ∈ Bas(C), and for e′ /∈ Bas(C) that is a sibling
of e, we either include e to Vir(C), or include des(e)∩ Bas(C ′) to Pro(C), for some sub-cluster C ′ of
C. In either cases, the number of added elements is at most r · O( s
)O(k), and we charge this to e.
We observe that for each e ∈ Bas(C), it has at most O(s)k siblings, by packing property.
Therefore, each e is charged at most O(s)k times. We conclude that
γ1
Pro(C) ∪ Vir(C) ≤ O(s)k · r2 · O(
s
γ1
)O(k).
We shall first prove that if ht(p) − ht(C) > 2 logs
Bounding NBas(C). Suppose P is the set consisting of C and all its ancestor clusters. By
definition, NBas(C) is a subset of the inside C clusters of Sp∈P (Pro(p) ∪ Vir(p)).
, then there is no element in Pro(p) ∪ Vir(p)
that can appear in NBas(C), for any p ∈ P . Suppose not. Then there exists some p such that
. Let j := ht(p). We observe that all elements in Pro(p)∪ Vir(p) have height
ht(p)− ht(C) > 2 logs
at least logs γ2
, by Definition 5.7 and Definition 5.5. However, if some element in
Pro(p)∪ Vir(p) appears in C, then it has height less than j − 2 logs
.
This is a contradiction. Therefore,
, by ht(C) < ht(p)− 2 logs
0 sj = j − 2 logs
1
γ0
1
γ0
1
γ0
1
γ0
1
γ0
NBas(C) ≤ O(logs
Hence Eff(C) ≤ Bas(C) + NBas(C) ≤ O(logs
Definition 5.12 (Disjointification). For any collection of clusters S, define Dis(S) := {e\Se′∈S:e′(e e′}e∈S.
We say e is induced by u in S, if u ∈ S and e = u\Se′∈S:e′(u e′, and the height of e is defined as
1
s
) · r2 · O(
γ1
γ0
1
) · r2 · O( s
)O(k).
γ0
)O(k).
γ1
the height of u.
Definition 5.13 (Cell Property). Suppose F is an SFP solution, and suppose f maps a cluster
C to a collection of sub-clusters of C. We say that f satisfies the cell property in terms of F if
for all clusters C, for all e ∈ Dis(f (C)), there is at most one crossing component of C in F that
intersects e.
Lemma 5.6 (Structural Property). Suppose an instance has a q-sparse optimal net-respecting
solution F . Moreover, for each i ∈ [L], for each u ∈ Ni, point u samples O(k log n) independent
random radii as in Definition 5.1. Then, with constant probability, there exists a configuration from
the sampled radii that defines a hierarchical decomposition, under which there exists an (m, r)-light
solution F ′ that includes all the points in F , and Eff defined in terms of F ′ satisfies the cell property,
where
16
ǫ )k and r := O(1)k · q logs log n + O( k
• E[w(F ′)] ≤ (1 + O(ǫ)) · w(F ),
• m := O( skL
bF with the desired m and r, and also satisfies Ehw(bF )i ≤ (1 + ǫ) · w(F ).
ǫ )k + O( s
ǫ )k.
Proof. We observe that the argument in [BGK12, Lemma 3.1] readily gives an (m, r)-light solution
We shall first show additional steps with additional cost at most ǫw(F ) in expectation, so that
Bas defined in terms of the resultant solution satisfies the cell property. And then, we shall show
that this implies Eff defined in terms of the resultant solution also satisfies the cell property (hence
no more additional cost caused).
Maintaining Cell Property: Basic Cells. For i := L, L−1, L−2, . . . , 0, for each height-i cluster
C, we examine e ∈ Dis(Bas(C)) in the non-decreasing order of its height. If there are at least two
crossing components that intersect e, we add edges in e to connect all crossing components that
intersect e. We note that each added edge connects two components in F , and edges added are of
length at most Diam(e). At the end of the procedure, we define the solution as F ′. We observe
that Bas defined in terms of F ′ satisfies the cell property.
Recall that each added edge connects two components. We charge the cost of the edge to one of
the components that it connects to. Moreover, after a rearrangement (at the end of the procedure),
we can make sure each edge is charged to one of the components it connects to and each component
is charged at most once.
Bounding The Cost. We shall show that for a fixed component A, the expected cost it takes
charge of is at most ǫ · w(A). Define l := w(A). The expected cost that A takes is at most the
following (up to contant)
LX
i=1
Pr[A takes an edge in a cell of height i] · si+1.
Define pi := Pr[A takes an edge in a cell of height i]. Then,
LX
i=0
pi · si+1 ≤ X
si+1 + X
pisi+1
i:si>2γ1l
pisi+1
i:si≤2γ1l
≤ O(γ1s)l + X
≤ O(ǫ)l + X
i:si>2γ1l
i:si>2γ1l
pisi+1.
Fix an i such that si > 2γ1l, and we shall upper bound pi. Suppose in the event corresponding
to pi, A takes charge of an edge inside a cell e that is a basic cell of some height-h cluster. Note
that h and e are random and recall that the edge is inside a cell of height i. We shall give a lower
bound of h.
Claim 5.1. sh ≥ si
Proof. Define the weight of the owner of e to be l′. We first show that h must be in I3(l′). By the
procedure of maintaining cell property, we know that l′ ≤ l.
If h ∈ I1(l′), then ⌊l′⌋s ≥ sh, and si ≤ 2γ1sh by e is of height i and the choice of radius in the
single-scale decomposition. This implies that si ≤ 2γ1sh ≤ 2γ1l, which cannot happen since we
assume si > 2γ1l.
2γ0
.
17
If h ∈ I2(l′), then si ≤ 2γ1⌊l′⌋s. This implies that si ≤ 2γ1l, which cannot happen as well.
Therefore, h ∈ I3(l′). This implies that 2γ0sh ≥ si.
Since the event that the edge is taken by A automatically implies that A is cut by a height-h
sj for j ∈ [L], we
cluster, and the probability that A is cut at a height-j cluster is at most O(k) ·
conclude that
l
pi ≤ X
j:sj≥ si
2γ0
Pr[A is cut at height j] ≤ O(k) · X
j:sj≥ si
2γ0
l
sj ≤ O(γ0k) ·
l
si .
Hence Pi:si>2γ1l pisi+1 ≤ O(γ0ksL) · l ≤ O(ǫ)l.
Maintaining Cell Property: Effective Cells. Next we show that Bas defined in terms of F ′
satisfies the cell property implies that Eff defined in terms of F ′ also satisfies the cell property.
component of C that intersects e in F ′. Suppose e is induced by u in Eff(C).
Fix a cluster C and fix e ∈ Dis(Eff(C)). We shall prove that there is at most one crossing
Lemma 5.7. If there is no cluster bC such that C ⊂ bC and u ∈ Vir(bC), then there exists cluster C ′
such that u ∈ Bas(C ′), ht(C ′) ≤ ht(C) and Eff(C) is a refinement of des(u) ∩ Bas(C ′).
Proof. If u ∈ Bas(C), then we define C ′ = C, and the Lemma follows.
If u ∈ NBas(C), then there exists C ′′ such that C ⊂ C ′′ and u ∈ Pro(C ′′). This is by the
definition of non-basic cells, and by the assumption that there is not cluster bC such that C ⊂ bC
and u ∈ Vir(bC). Then by the definition of the promoted cells, there exists cluster C ′ such that
u ∈ Bas(C ′), ht(C ′) < ht(C ′′), and des(u) ∩ Bas(C ′) ⊂ Eff(C ′′). Since u ∈ NBas(C) ⊂ Eff(C) and
by Lemma 5.3, we know that Eff(C) is a refinement of des(u)∩ Bas(C ′). Hence, it remains to show
ht(C ′) ≤ ht(C).
Suppose for contradiction that ht(C ′) > ht(C), so ht(C) < ht(C ′) < ht(C ′′). By the definition
of non-basic cells, we have that NBas(C ′) ∩ Bas(C ′) = ∅. Since u ∈ Bas(C ′), we know that
u /∈ NBas(C ′). However, this implies that u /∈ NBas(C), which contradicts with the assumption
that u ∈ NBas(C).
crossing component of C in F ′ that intersects e.
If there exists cluster bC such that u ∈ Vir(bC) and C ⊂ bC, then by Lemma 5.2, there is no
Otherwise, there is no cluster bC such that u ∈ Vir(bC) and C ⊂ bC. By Lemma 5.7, there exists
a cluster C ′ such that u ∈ Bas(C ′), ht(C ′) ≤ ht(C) and Eff(C) is a refinement of des(u) ∩ Bas(C ′).
We pick any one of such C ′. Define e′ ∈ Dis(Bas(C ′)) as the one induced by u in Bas(C ′). Since
Bas defined in terms of F ′ satisfies the cell property, there is at most one crossing component of C ′
that intersects e′.
Lemma 5.8. e ⊂ e′.
Proof. Recall that e ∈ Dis(Eff(C)) is induced by u in Eff(C), and e′ ∈ Dis(Bas(C ′)) is induced by
u in Bas(C ′). Then we can write e = u\P and e′ = u\P ′ such that P ⊂ Eff(C) and P ′ ⊂ Bas(C ′).
Since P ′ = des(u) ∩ Bas(C ′), and Eff(C) is a refinement of des(u) ∩ Bas(C ′), we know that P ′ ⊂ P .
This implies that e ⊂ e′.
Since ht(C) ≥ ht(C ′), any crossing component of C is also a crossing component of C ′. Moreover,
Lemma 5.8 implies that e ⊂ e′. Hence, if there are two crossing components A1, A2 of C that
intersect e, then A1 and A2 are also crossing components of C ′ and both of them intersect e′.
However, this cannot happen since Bas satisfies the cell property, and there is at most one crossing
18
component in C ′ that intersects e′. Therefore, there is at most one crossing component of C that
intersects e.
5.3 Dynamic Program
Recall that the input of DP is an instance that has a q-sparse optimal net-respecting solution,
where q ≤ O(s)O(k) · q0, by Lemma 4.1 and Corollary 3.1. In the DP algorithm, O(k log n) random
radii are independently sampled for each u ∈ Ni, i ∈ [L], and then a dynamic programming based
algorithm is used to find a near optimal SFP solution over all hierarchical decompositions defined
by the radii. In this section, we shall describe in detail the dynamic program and an algorithm that
solves the dynamic program efficiently. For completeness, we shall also analyze the correctness of
the dynamic program.
We first describe the information needed to identify each cluster at each height.
Information to Identify a Cluster. Each cluster is identified by the following information.
O(nk).
1. Height i and cluster center u ∈ Ni. This has L · O(nk) combinations, recalling that Ni ≤
2. For each j ≥ i, and v ∈ Nj such that d(u, v) ≤ O(sj), the random radius chosen by (v, j).
Observe that the space around B(u, O(si)) can be cut by net-points in the same or higher
heights that are nearby with respect to their distance scales. As argued in [BGK12], the
= nO(1)k
number of configurations that are relevant to (u, i) is at most O(k log n)L·O(1)k
,
where L = O(logs n) and s = (log n)Θ( 1
k ).
3. For each j > i, which cluster at height j (specified by the cluster center vj ∈ Nj) contains
the current cluster at height i. This has O(1)kL = nO(
log log n ) combinations.
k2
To define the dynamic program, we start by defining the entries.
Entries of DP. We define entries as (C, (R, Y ), (BAS, NBAS), (g, P )). Define U := Dis(BAS ∪
NBAS). We define the following internal constraints for entries, where the parameters m, r are as
defined in Lemma 5.6, and ρ is as defined in Lemma 5.5.
connected inside C.
• C is a cluster.
• R is a subset of the m pre-defined portals, such that R ≤ r. This intends to denote the
active portals.
• Y ⊂ 2R is a partition of R. We intend to use it to record the subsets of portals that are
• BAS and NBAS are collections of sub-clusters of C such that BAS ∩ NBAS = ∅ and BAS ∪
NBAS ≤ ρ, and the centers of the clusters in BAS∪ NBAS are chosen from Can(C). Moreover,
e ∈ BAS implies that any sibling cluster of e is in BAS ∪ NBAS. We intend to use this to
record the basic cells and non-basic cells.
• g is a mapping from U to 2Y . For some e ∈ U , we intend to use g(e) to denote the portals
• P ⊂ 2Y is a partition of Y , such that ∀e ∈ U , g(e) = Q implies that Q is a subset of a part
We only consider the entries that satisfy the internal constraints. We capture the intended use of
an entry formally as follows.
in P . The intended use of P is to denote the portals that are to be connected outside C.
that e connects to inside C.
Definition 5.14 (Compatibility). Suppose F is a graph on the metric space, and E is an entry.
Let E := (C, (R, Y ), (BAS, NBAS), (g, P )). Define F ′ := FC∪R. We say F is compatible to E, if
F ′ satisfies the following.
1. A part y is in Y , if and only if F ′ connects all the portals in y.
2. BAS covers all components of F ′ that intersect R.
19
3. For e ∈ U , g(e) is exactly the collection of subsets of Y that e is connected to by F ′.
4. Every terminal in C is visited by F ′.
5. Every isolated terminal of C is connected to at least one portal in R by F ′.
6. Every terminal pair that both lie in C is either in the same component of F ′, or they are
connected to y1 and y2 in Y by F ′ and {y1, y2} is a subset of a part in P .
We bound the number of entries in the following lemma.
) · O(κmr)O(k)k·ρr number of en-
Lemma 5.9 (Number of Entries). There are at most O(nO(1)k
tries. Moreover, for any fixed cluster C, the number of entries with C as the cluster is at most
O(κmr)O(k)k·ρr. (κ is defined as in Lemma 5.4.)
Proof. Since R is a set of at most r portals chosen from m pre-defined portals, there are at most
O(mr) possibilities of R. Then after R is fixed, there are O(rr) possibilities of Y , since Y is a
partition of Y and R ≤ r.
To count the number of BAS and NBAS, we count the union S := BAS ∪ NBAS of them, and
then for any fixed S we count the number of ways to assign elements in S to BAS and NBAS. Since
it is required that the centers of clusters in S are chosen from Can(C), to form S, we first choose at
most ρ centers from Can(C). There are at most O(κρ) possibilities for this, by Lemma 5.4. For each
chosen center u that is of height iu, we count the number of configurations of the cluster Cu centered
at u. Since C is already fixed, we only need to consider relevant radii for clusters of height less than
ht(C) and at least iu. Since u ∈ Can(C), and for j ≥ iu there are O(1)k clusters of height-j can
configurations
affect u, we conclude that there are at most O(k log n)
for Cu. Since S ≤ ρ, there are at most O(k log n)O(k)k·ρ configurations for all clusters in S, for
any given the centers. Therefore, there are O(κ)O(k)k·ρ possibilities for S in total. Then we assign
elements in S to one of BAS, NBAS, and there are at most 2S ≤ 2ρ number of them. In conclusion,
the number of possibilities for BAS and NBAS is at most O(κ)O(k)k·ρ.
With S fixed, we count the number of possibilities of g. Since g is a mapping from U to 2Y , the
number of such a mapping is at most O((2Y )U ) ≤ O(2ρ·r). Finally, observe that P is a partition
of Y , and Y ≤ r. This implies that P has at most O(rr) possibilities.
≤ O(k log n)O(k)k
O(1)k·logs ( 1
γ2
0
)
Therefore, after fixing C, there are at most O(κmr)O(k)k·ρr possibilities for ((R, Y ), (BAS, NBAS), (g, P )).
We then count the number of possibilities of C. Observe that there are O(n)O(k) centers for C.
For a fixed center, since the number of configurations is at most nO(1)k , we conclude that there are
at most O(nO(1)k
) · O(κmr)O(k)k·ρr entries in total.
After we define the entries, we shall (recursively) define the value that is associated with each
entry. The intended value of an entry E is the weight of the minimum graph that is recursively
compatible to E (see definition 5.18).
Definition 5.15 (Child Entry Collection). Suppose E := (C, (R, Y ), (BAS, NBAS), (g, P )) is an en-
try. We say a collection of entries {(Ci, (Ri, Yi), (BASi, NBASi), (gi, Pi))}i is a child entry collection
of E, if {Ci}i is a partition of C with ht(Ci) = ht(C) − 1 for all i.
Definition 5.16 (Portal Graph). We say a graph G is a portal graph of a collection of entries
I := {(Ci, (Ri, Yi), (BASi, NBASi), (gi, Pi))}i, if the vertex set of G is Si Ri.
Definition 5.17 (Consistency Checking). Suppose E := (C, (R, Y ), (BAS, NBAS), (g, P )) is an
entry, and I := {(Ci, (Ri, Yi), (BASi, NBASi), (gi, Pi))}i is a child entry collection of E and G is a
portal graph of I. We say G and I are consistent with E, if all checks in the following procedure
are passed.
20
cluster of e.
i to be a mapping from Ui to 2Y , where g′
1. Check if Si (Ci ∪ Ri) = C ∪ R.
2. We shall define Y ′ to be a partition of R′ := Si Ri. Initialize Y ′ := Si Yi, and whenever there
are y1, y2 ∈ Y ′ connected by G or y1 ∩ y2 6= ∅, replace them by the union of them. Check if
Y ′ restricted to R is exactly Y .
3. For each e ∈ BAS, check if there exists i and e′ ∈ BASi, such that e′ = e or e′ is a child
4. For each e ∈ NBAS, check if either there exists i and e′ ∈ BASi ∪ NBASi such that e = e′, or
all child clusters of e are in Si (BASi ∪ NBASi).
i(e) := {y ∩ R y ∈ Y ′ ∧ ∃y′ : (y′ ∈
5. Define g′
gi(e) ∧ y ∩ y′ 6= ∅)}, for e ∈ Ui. Here g′
i(e) intends to mean the parts in Y that e connects
to, which is defined by "extending" gi(e) with respect to G. For each i and u ∈ BASi, if there
i(e) 6= ∅, then check if there exists u′ ∈ BAS such that
exists e ∈ Ui such that e ⊂ u and g′
u = u′ or u is a child cluster of u′.
i(e′), for e ∈ U . Check if g′
is exactly g. We observe that here we consider e′ ⊂ e only, and we shall see later why this is
sufficient.
7. For each i, for each y1, y2 ∈ Yi (y1 6= y2) such that y1, y2 are in the same part of Pi, check if
1 6= y′
either there exists y ∈ Y ′ such that y1 ∪ y2 ⊂ y, or there exists y′
2,
y1 ⊂ y′
2 ∩ R} is a subset of a part in P . This
intends to check if the parts in Pi are connected by G, or the information in Pi's is passed to
P .
6. Define a mapping g′ from U to 2Y , where g′(e) := SiSe′∈Ui:e′⊂e g′
2 ∈ Y ′ such that y′
1, y2 ⊂ y′
2, y′
1 ∩ R 6= ∅, y′
2 ∩ R 6= ∅, and {y′
1 ∩ R, y′
1, y′
8. For each terminal pair (a, b) such that a ∈ Ci and b ∈ Cj for i 6= j, suppose a ∈ ei and
i(ei) 6= ∅,
j(ej) is a subset of a part in P . This intends to check if (a, b) are already
b ∈ ej for ei ∈ Ui and ej ∈ Uj. Check if gi(ei) is connected by G to gj(ej), or if g′
g′
j(ej) 6= ∅, g′
connected by G, or otherwise they will be connected outside C.
i(ei)∪ g′
9. For each isolated terminal a in C, check if there exists i and e ∈ Ui, such that a ∈ e and g′
i(e)
is non-empty.
Definition 5.18 (Recursive Compatibility). Suppose E := (C, (R, Y ), (BAS, NBAS), (g, P )) is an
entry, and F is some graph on the metric space. F is recursively compatible with E, if there exists
a set S of entries with E ∈ S and with a unique entry in S that corresponds to each descendant
cluster of C, such that the following requirements hold.
• For each E′ := (C ′, (R′, Y ′), (BAS′, NBAS′), (g′, P ′)) in S, we require F ′ := FC ′∪R′ be com-
• For each E′ := (C ′, (R′, Y ′), (BAS′, NBAS′), (g′, P ′)) in S, suppose the child entry collection
that consisting of elements in S is I ′, and define I ′ := {(Ct, (Rt, Yt), (BASt, NBASt), (gt, Pt))}t.
Define G′ := FSt Rt. (Note that G′ is a portal graph of I ′.) We require I ′ and G′ be consistent
with E′.
patible to E′.
Value of Entries. For any entry E := (C, (R, Y ), (BAS, NBAS), (g, P )), we shall define its value
val(E). The height-0 clusters are corresponding to the base cases. In particular, for any C := {x}
that is a height-0 cluster, we define entries with such C and with BAS := {C}, NBAS := ∅, R := C,
Y := {R}, g(C) := Y , P := {Y } to be the base entries. All base entries have value 0. All other
(non-base) entries with height-0 clusters have value ∞.
We then define val(E) when ht(C) 6= 0. Define IE to be the set of tuples (I, G), such that I is a
child entry collection of E and G is a portal graph of I, and I, G are consistent. The value of E is
defined as val(E) := min(I,G)∈IE{w(G) + val(I)}, where val(I) = PE ′∈I val(E′). As we shall see in
Lemma 5.12, for any entry E, if val(E) 6= ∞, then there actually exists a graph that is recursively
compatible to E with weight val(E).
21
, where κ is defined as in Lemma 5.4.
Lemma 5.10 (Counting IE). For any entry E, the number of possibilities of IE is at most
O(k log n)O(s)k · O(κmr)O(sk)O(k)·ρr2
Proof. Define E := (C, (R, Y ), (BAS, NBAS), (g, P )). We first bound the number of possibilities of
child entry collections I := {(Ci, (Ri, Yi), (BASi, NBASi), (gi, Pi))}i of C. To define I, we start by
defining {Ci}i. By packing property, there are at most O(s)k centers for the child clusters of C.
For each center u of the child cluster, there are at most O(k log n) possible radii. Hence, there are
at most O(k log n)O(s)k
By Lemma 5.9, there are at most Z possibilities for ((Ri, Yi), (BASi, NBASi), (gi, Pi)) for any
possibil-
fixed Ci, where Z := O(κmr)O(k)k·ρr. Therefore, there are at most O(k log n)O(s)k · Z O(s)k
ities of I.
possibilities for {Ci}i.
For a fixed I, the vertex set of the portal graph G of I is fixed, and there are at most O(s)k · r
vertices in G. Then the number of possibilities of G for a fixed I is at most the number of edge
sets, and it is at most 2O(s)O(k)·r2
since there are at most (O(s)k · r)2 edges.
In conclusion, there are at most O(k log n)O(s)k · Z O(s)k · 2O(s)O(k)·r2
.
at most O(k log n)O(s)k · O(κmr)O(sk)O(k)·ρr2
Final Entry. The final entry is the entry with C being the root cluster, R, BAS, NBAS to be ∅,
and Y, g, P being uniquely defined from R, BAS, NBAS = ∅. We use the value of the final entry as
the output of DP.
Evaluating The Final Entry Although we only care about the value of the final entry, it may
be necessary to evaluate the value of other entries. We shall define a (recursive) algorithm in
Definition 5.19 that takes an entry and returns the value of the input. To get the value of the final
entry which is the output of DP, we invoke the algorithm with the final entry as the input.
possibilities of IE, which is
We note that the counting argument in Lemma 5.9 and Lemma 5.10 can both be naturally
implemented as algorithms, with additional O(nO(k)) factors in the running time compared with
the corresponding counting bounds. We will make use of these implementations as subroutines in
Definition 5.19. Moreover, the natural implementation of the consistency checking procedure in
Definition 5.17 runs in time O(nO(k)).
Definition 5.19 (Algorithm for Evaluating Value of Entries). We define a recursive procedure that
evaluates the value of an input entry E := (C, (R, Y ), (BAS, NBAS), (g, P )).
• If ht(C) = 0, then the value of it is already defined, and we return its value.
• If ht(C) > 0 and val(E) is already calculated, then we return the calculated value.
• Otherwise, ht(C) > 0 and val(E) has not yet calculated. The following procedure is executed.
1. Set the default value for val(E) := ∞.
2. Calculate IE.
3. For each element (I, G) ∈ IE, use the consistency checking procedure defined in Defini-
tion 5.17 to check if I and G are consistent with E. If they are consistent, then recur-
sively use this procedure to calculate val(I) + w(G), and update val(E) if val(E) + w(G)
is smaller than val(E).
4. Finally, return val(E) as the output.
Lemma 5.11 (Running Time). The running time for the algorithm defined in Definition 5.19 is
at most O(nO(1)k
) · exp(√log n · O( k
ǫ )O(k)).
Proof. Suppose the input is E := (C, (R, Y ), (BAS, NBAS), (g, P )). We observe that once the value
for some entry is calculated, it would not be calculated again, and recalling the value takes constant
time. Then we shall bound the time when val(E) is not yet calculated and ht(C) 6= 0.
22
Observe that for any given I with val(E′) for all E′ ∈ I known and a graph G such that
(I, G) ∈ IE, evaluating val(I) + w(G) takes O(n)O(k) time. Therefore, combining with Lemma 5.9
and Lemma 5.10, there are at most O(nO(1)k ) · Z entries, and it takes O(n)O(k) · O(k log n)O(s)k ·
O(κmr)O(sk)O(k)·ρr2
In conclusion, the time for evaluating all the entries is at
most O(nO(1)k
Substituting Parameters. Recall that we consider q ≤ O(s)O(k) · q0. Observe that
γ0⌉s ≤ O( ks3L
⌈ 1
ρ ≤ O( sk
) · O(k log n)O(s)k · O(κmr)O(sk)O(k)·ρr2
), and 1
γ1
ǫ )O(k). Moreover,
ǫ ). Substituting γ0 and γ1, we have κ ≤ O( ksL
:=
ǫ )O(k) and
γ1⌋s ≤ O( s2
to evaluate each.
:= ⌊ 1
1
γ0
.
ǫ
r := O(1)k · q logs log n + O(
k
ǫ
)k + O(
s
ǫ
)k ≤ O(
sk
ǫ
)O(k), m ≤ O(
skL
ǫ
)k.
By definition, s := (log n)
c
k , L := O(logs n) = O( k log n
c log log n ). Therefore, the running time is at most
O(nO(1)k
≤ O(nO(1)k
≤ O(nO(1)k
≤ O(nO(1)k
· O(κmr)O(sk)O(k)·ρr2
) · O(k log n)O(s)k
ǫ )O(k)
) · O(k log n)O(s)k
· O(
k
) · exp(O(s)O(k) · O(
ǫ
) · exp(O(
ǫ
)O(k) · log
)O(k) · O(log n)O(c) · log log n).
)O( sk
k log n
ksL
k
ǫ
ǫ
)
By choosing constant c to be sufficiently small so that O(log n)O(c) · log log n ≤ O(√log n), we
conclude that the running time is at most O(nO(1)k ) · exp(√log n · O( k
ǫ )O(k)).
Lemma 5.12 (Characterizing the Value of Entries). For each entry E := (C, (R, Y ), (BAS, NBAS), (g, P ))
with val(E) 6= ∞, val(E) is the weight of the minimum weight graph that is recursively compatible
to the entry and uses points in C ∪ R only.
Proof. For the clusters of height 0, the Lemma holds trivially.
Assuming the Lemma holds for all entries with the clusters of height i− 1, we prove the Lemma
for an entry E with C of height i centered at u ∈ Ni, where i ≥ 1. We shall first show that val(E)
is the weight of some graph that is recursively compatible to the entry and uses points in C ∪ R
only. Then we shall show that the value is minimum.
Feasibility. Suppose (I, G) := arg min(I ′,G′)∈IE{val(I ′) + w(G′)}. Define I = {Ej}j, where Ej :=
(Cj, (Rj, Yj), (BASj, NBASj), (gj , Pj )). Since val(E) 6= ∞, we have val(I ′) 6= ∞. For Ej ∈ I, by
assumption, there exists a graph that is recursively compatible to Ej and uses points in Cj ∪ Rj
only, and we denote it as Fj. We define a graph F that is the union of Fj for all j, and G. Then
w(F ) = val(I) + w(G).
We shall show that F is recursively compatible to E. Since (I, G) ∈ IE, I and G are consistent
with E. Since Fj is recursively compatible to Ej for all j, it remains to verify F is compatible to
E. When we say "consistency checking procedure", we refer to Definition 5.17.
• F uses points in C ∩ R only. This is by definition.
• A part y is in Y , if and only if F connects all the portals in the part y. This is by the step 2
• BAS covers all components of F that intersect R. This is by step 5 of the consistency checking
of the consistency checking procedure.
procedure.
23
• For e ∈ U , the collection of subsets of Y that e is connected to by F is exactly g(e). We note
that step 3, 4 of the consistency checking procedure, together with the internal constraint
that e′ ∈ BASj implies any sibling cluster of e′ is in BASj ∪ NBASj for all j. This implies
that Sj (BASj ∪ NBASj) is a refinement of BAS ∪ NBAS. Then, for each j, for each e′ ∈ Uj,
and for each e ∈ U , either e′ ⊂ e or e′ ∩ e = ∅. Therefore, step 6 is sufficient to ensure this
item. (If e′ is not a subset of e but e′ ∩ e 6= ∅, then the gj mappings in the sub-entries have
not sufficient information to determine the portals that e′ ∩ e is connected to.)
• Every terminal in C ∪ R is visited by F . This is by the construction of F , and by Fj is
recursively compatible to Ej for all j.
• Every isolated terminal of C is connected to at least one portal in R by F . This is by step 9
• Every terminal pair that both lie in C is either in the same component of F , or they are
connected to y1 and y2 in Y by F and {y1, y2} is a subset of a part in P . This is by step 7, 8
of the consistency checking procedure, and by Fj is recursively compatible to Ej for all j.
of the consistency checking procedure.
This implies that F is recursively compatible to E.
Optimality. Then we shall show that val(E) is minimum. Suppose not. Define l as the weight
of the minimum weight graph that is recursively compatible to E and uses points in C ∪ R only.
Define F ′ to be the corresponding graph recursively compatible to E with weight l. Since F ′ is
recursively compatible to E, there exists I ′ := {Et := (Ct, (Rt, Yt), (BASt, NBASt), (gt, Pt))}t and
a portal graph G′ of I ′ that are consistent with E. Moreover, there exists a graph Ft that is
recursively compatible to Et, for all t.
We note that (I ′, G′) ⊂ IE. Therefore, Pt w(Ft) + w(G′) = l < val(E) ≤ val(I ′) + w(G′).
This implies that Pt w(Ft) < PE ′∈I ′ val(E′), and hence there exists t such that w(Ft) < val(Et).
However, we know that Ft is recursively compatible to Et, and by assumption, val(Et) ≤ w(Ft).
This is a contradiction.
Corollary 5.1. There exists a feasible solution to SFP whose weight is the value of the final entry.
Lemma 5.13 (Good Solution is Recursively Compatible). Suppose for each i ∈ [L] and u ∈ Ni,
O(k log n) radii are fixed. Suppose F is an (m, r)-light solution such that Eff satisfies the cell
property in terms of F under one of the hierarchical decompositions defined by the radii. Then the
value of the final entry is at most w(F ).
Proof. We shall show that F is recursively compatible to the final entry, and then Lemma 5.12
implies that the value of the final entry is at most w(F ).
Suppose we fix a hierarchical decomposition induced from the given radii, such that F is (m, r)-
light and Eff satisfies the cell property in terms of F . For each cluster C in the decomposition,
we define FC := FC∪R, define an entry EC := (C, (RC , YC ), (BASC , NBASC), (gC , PC )) as follows,
where R is the set of active portals for C.
• RC := R.
• YC contains a part y, if and only if portals in y is connected by FC.
• BASC := Bas(C), NBASC := NBas(C).
• For each e ∈ UC, let gC(e) := Q, where Q is the collection of parts in YC that e is connected
• Define PC to be any one that satisfies
1. for each e ∈ UC , gC (e) = Q implies Q is a subset of PC ;
2. for each terminal pair (a, b) that both lie in C, if they are not connected by FC then the
to by FC.
subsets of portals that a and b are connected by FC are in a same part of PC .
24
The internal constraints for an entry is satisfied, from the definition of the cells, the fact that Eff
satisfies the cell property, Lemma 5.4 and Lemma 5.5.
Then we (uniquely) define IC := {(Ci, (Ri, Yi), (BASi, NBASi), (gi, Pi))}i as the child collection
of EC, and define GC := FSi Ri as the portal graph of IC.
We then check that IC and GC are consistent with EC . Step 1, 2, 7, 8, 9 are immediate. Step
3, 4, 5 follow from the definition of the basic cells and non-basic cells. Inside step 6, we observe
that g′(e) is evaluated by looking at e′ ⊂ e only (instead of considering all e′ ∈ Ui), for e′ ∈ Ui for
some i, and e ∈ U . However, this is indeed sufficient, since Lemma 5.3 asserts that for any e ∈ U ,
e′ ∈ Ui for any i, either e′ ⊂ e′ or e ∩ e′ = ∅.
is by definition.
It remains to check the following for EC , for each cluster C.
• A part y ∈ YC, if and only if FC connects all the portals in the part y. This is by definition.
• BAS covers all components of FC that intersect RC . This is by definition.
• For e ∈ UC, the collection of subsets of YC that e is connected to by FC is exactly gC (e). This
• Every terminal in C ∪ RC is visited by FC . This is by the feasibility of F .
• Every isolated terminal of C is connected to at least one portal in RC by FC . This is by the
• Every terminal pair that both lie in C is either in the same component of FC , or they are
connected to y1 and y2 in YC by FC and {y1, y2} is a subset of a part in PC . This is by
definition.
feasibility of F .
This finishes the proof.
Combining Lemma 5.6, Lemma 5.13, Corollary 5.1 and Lemma 5.11, we conclude a PTAS for
sparse SFP instances.
Corollary 5.2 (PTAS for Sparse SFP Instances). For an instance of SFP that has a q-sparse
optimal net-respecting solution, algorithm DP returns a (1 + ǫ) solution with constant probability,
running in time O(nO(1)k ) · exp(√log n · O( k
ǫ )O(k)), for q ≤ O(s)O(k) · q0.
References
[ABN06]
Ittai Abraham, Yair Bartal, and Ofer Neiman. Advances in metric embedding theory.
In STOC, pages 271 -- 286. ACM, 2006.
[AKR95] Ajit Agrawal, Philip N. Klein, and R. Ravi. When trees collide: An approximation al-
gorithm for the generalized steiner problem on networks. SIAM J. Comput., 24(3):440 --
456, 1995.
[Aro98]
[Aro02]
Sanjeev Arora. Polynomial time approximation schemes for euclidean traveling sales-
man and other geometric problems. J. ACM, 45(5):753 -- 782, 1998.
S. Arora. Approximation algorithms for geometric TSP.
In The traveling salesman
problem and its variations, volume 12 of Comb. Optim., pages 207 -- 221. Kluwer Acad.
Publ., Dordrecht, 2002.
[Ass83]
P. Assouad. Plongements lipschitziens dans Rn. Bull. Soc. Math. France, 111(4):429 --
448, 1983.
[BDHM16] MohammadHossein Bateni, Erik D. Demaine, MohammadTaghi Hajiaghayi, and D´aniel
Marx. A PTAS for planar group steiner tree via spanner bootstrapping and prize
collecting. In STOC, pages 570 -- 583. ACM, 2016.
25
[BGK12] Yair Bartal, Lee-Ad Gottlieb, and Robert Krauthgamer. The traveling salesman prob-
lem: low-dimensionality implies a polynomial time approximation scheme. In STOC,
pages 663 -- 672. ACM, 2012.
[BGRS10] Jaroslaw Byrka, Fabrizio Grandoni, Thomas Rothvoss, and Laura Sanit`a. An improved
lp-based approximation for steiner tree. In STOC, pages 583 -- 592. ACM, 2010.
[BH12]
MohammadHossein Bateni and MohammadTaghi Hajiaghayi.
collecting steiner forest. Algorithmica, 62(3-4):906 -- 929, 2012.
Euclidean prize-
[BHM11] MohammadHossein Bateni, Mohammad Taghi Hajiaghayi, and D´aniel Marx. Approx-
imation schemes for steiner forest on planar graphs and graphs of bounded treewidth.
J. ACM, 58(5):21, 2011.
[BKM08] Glencora Borradaile, Philip N. Klein, and Claire Mathieu. A polynomial-time approx-
imation scheme for euclidean steiner forest. In FOCS, pages 115 -- 124. IEEE Computer
Society, 2008.
[CC08]
[CJ16]
Miroslav Chleb´ık and Janka Chleb´ıkov´a. The steiner tree problem on graphs: Inap-
proximability results. Theor. Comput. Sci., 406(3):207 -- 214, 2008.
T.-H. Hubert Chan and Shaofeng H.-C. Jiang. Reducing curse of dimensionality: Im-
proved PTAS for TSP (with neighborhoods) in doubling metrics.
In SODA, pages
754 -- 765. SIAM, 2016.
[DHC85] DZ Du, FK Hwang, and SC Chao. Steiner minimal tree for points on a circle. Proceed-
ings of the American Mathematical Society, 95(4):613 -- 618, 1985.
[DHK14]
Erik D. Demaine, Mohammad Taghi Hajiaghayi, and Philip N. Klein. Node-weighted
steiner tree and group steiner tree in planar graphs. ACM Trans. Algorithms,
10(3):13:1 -- 13:20, 2014.
[DHW87] Ding-Zhu Du, Frank K. Hwang, and JF Weng. Steiner minimal trees for regular poly-
gons. Discrete & Computational Geometry, 2(1):65 -- 84, 1987.
[DL97]
[GK15]
M. M. Deza and M. Laurent. Geometry of cuts and metrics, volume 15 of Algorithms
and Combinatorics. Springer-Verlag, Berlin, 1997.
Anupam Gupta and Amit Kumar. Greedy algorithms for steiner forest.
pages 871 -- 878. ACM, 2015.
In STOC,
[GKL03] Anupam Gupta, Robert Krauthgamer, and James R. Lee. Bounded geometries, fractals,
In FOCS, pages 534 -- 543. IEEE Computer Society,
and low-distortion embeddings.
2003.
[GKR00] Naveen Garg, Goran Konjevod, and R. Ravi. A polylogarithmic approximation algo-
rithm for the group steiner tree problem. J. Algorithms, 37(1):66 -- 84, 2000.
[GW95]
[HK03]
Michel X. Goemans and David P. Williamson. A general approximation technique for
constrained forest problems. SIAM J. Comput., 24(2):296 -- 317, 1995.
Eran Halperin and Robert Krauthgamer. Polylogarithmic inapproximability. In STOC,
pages 585 -- 594. ACM, 2003.
26
[HK13]
[Mat02]
[Tal04]
Mathias Hauptmann and Marek Karpi´nski. A compendium on steiner tree problems.
Inst. fur Informatik, 2013.
J. Matousek. Lectures on discrete geometry, volume 212 of Graduate Texts in Mathe-
matics. Springer-Verlag, New York, 2002.
Kunal Talwar. Bypassing the embedding: algorithms for low dimensional metrics. In
STOC, pages 281 -- 290. ACM, 2004.
27
|
1706.02202 | 1 | 1706 | 2017-06-07T14:14:50 | Isomorphic coupled-task scheduling problem with compatibility constraints on a single processor | [
"cs.DS",
"cs.CC"
] | The problem presented in this paper is a generalization of the usual coupled-tasks scheduling problem in presence of compatibility constraints. The reason behind this study is the data acquisition problem for a submarine torpedo. We investigate a particular configuration for coupled tasks (any task is divided into two sub-tasks separated by an idle time), in which the idle time of a coupled task is equal to the sum of durations of its two sub-tasks. We prove -completeness of the minimization of the schedule length, we show that finding a solution to our problem amounts to solving a graph problem, which in itself is close to the minimum-disjoint-path cover (min-DCP) problem. We design a (3a+2b)/(2a+2b)-approximation, where a and b (the processing time of the two sub-tasks) are two input data such as a>b>0, and that leads to a ratio between 3/2 and 5/4. Using a polynomial-time algorithm developed for some class of graph of min-DCP, we show that the ratio decreases to 1.37 . | cs.DS | cs | Noname manuscript No.
(will be inserted by the editor)
Isomorphic coupled-task scheduling problem with
compatibility constraints on a single processor
G. Simonin · B. Darties · R. Giroudeau · J.-C. König
Received: date / Accepted: date
Abstract The problem presented in this paper is a
generalization of the usual coupled-tasks scheduling pro-
blem in presence of compatibility constraints. The rea-
son behind this study is the data acquisition problem
for a submarine torpedo. We investigate a particular
configuration for coupled-tasks (any task is divided into
two sub-tasks separated by an idle time), in which the
idle time of a coupled-task is equal to the sum of dura-
tions of its two sub-tasks. We prove NP-completeness
of the minimization of the schedule length, we show that
finding a solution to our problem amounts to solving a
graph problem, which in itself is close to the minimum-
disjoint path cover (min-DCP) problem. We design a
2a+2b(cid:17)- approximation, where a and b (the process-
(cid:16) 3a+2b
ing time of the two sub-tasks) are two input data such
as a > b > 0, and that leads to a ratio between 3
2 and 5
4 .
Using a polynomial-time algorithm developed for some
class of graph of min-DCP, we show that the ratio de-
creases to 1+√3
Keywords coupled-tasks · complexity · compatibility
graph · polynomial-time approximation
2 ≈ 1.37.
1 Introduction
In this paper, we present a scheduling problem of coupled-
tasks subject to compatibility constraints, which is a
G. Simonin · R. Giroudeau · J.-C. König
LIRMM UMR 5506, rue Ada,
34392 Montpellier Cedex 5 - France
E-mail: {simonin,rgirou,konig}@lirmm.fr
B. Darties
LE2I UMR 5158
9 Rue Alain Savary
21000 Dijon - France
E-mail: [email protected]
generalization of the scheduling problem of coupled-
tasks first introduced by Shapiro [19]. This problem is
motivated by the problem of data acquisition in a sub-
marine torpedo. The aim amounts to treating various
environmental data coming from sensors located on the
torpedo, that collect information which must be pro-
cessed on a single processor. A single acquisition task
can be described as follows: a sensor of the torpedo
emits a wave at a certain frequency (according to the
data that must be collected) which propagates in the
water and reflects back to the sensor. This acquisition
task is divided into two sub-tasks: the first task con-
sists in sending an ultrasound pulse while the second
receives returning echo. Between them, there is an in-
compressible idle time which represents the spread of
the echo under the water. Thus acquisition tasks may
be assigned to coupled-tasks.
In order to use idle time, other sensors can send
more echoes. However, the proximity of the waves causes
disruptions and interferences. In order to handle infor-
mation error-free, a compatibility graph between acqui-
sition tasks is created. In this graph, which describes
the set of tasks, we have an edge between two compat-
ible tasks. A task is compatible with another if at least
one of its sub-tasks can be executed during the idle
time of another task. Given a set of coupled-tasks and
such a compatibility graph, the aim is to schedule the
coupled-tasks in order to minimize the time required
for the completion of all the tasks.
1.1 Notations
First we present some common notations:
-- Let G be an undirected graph. We note V (G) the
set of its vertices and E(G) the set of its edges;
7
1
0
2
n
u
J
7
]
S
D
.
s
c
[
1
v
2
0
2
2
0
.
6
0
7
1
:
v
i
X
r
a
PSfrag replacements
2
-- we note n (resp. m) the cardinality of set V (G)
(resp. E(G));
-- a path is a non-empty graph C with V (C) = {x0, x1,
. . . , xk} and E(C) ={x0x1, . . . , xk−1, xk}, where all
the xi are distinct;
-- the length of a path is the number of edges that the
path uses.
Then, we introduce the notations relative to coupled-
tasks we will use in the rest of the paper: we note
A = {A1, A2, . . . , An} the set of n coupled-tasks. Us-
ing the notation proposed by Shapiro [19], each task
Ai ∈ A is composed of two sub-tasks ai and bi. For clar-
ity we use the same notations for the processing time of
these tasks: ai and bi have processing time ai ∈ N and
bi ∈ N), and separated by a fixed idle time Li ∈ N (see
Figure 1(a)). For each i the second sub-task bi must
start its execution exactly Li time units after the com-
pletion time of ai.
According to the torpedo problem, a task may be
started during the idle time of a running task if it uses
another frequency, is not dependant on the execution
of the running task (and reciprocally), or does not re-
quire to access the resources used by the running tasks.
Formally, we say that two tasks Ai and Aj are compat-
ible if and only if we can execute at least a sub-task
of Ai during the idle time of Aj (see Figure 1(b)). On
the other side, some tasks cannot be compatible due to
previously cited reasons.
PSfrag replacements
PSfrag replacements
aj
bj
ai
Li
bi
(a)
.
ai
aj
bi
bj
Li
(b)
Fig. 1 A single coupled-task and two compatible coupled-tasks.
1.2 Main problem formulation
We aim at scheduling a set of coupled-tasks with com-
patibility constraints on a monoprocessor. The input
of the general problem is described with the set A =
{A1, A2, . . . , An} of coupled-tasks and a compatibility
graph Gc, with V (Gc) = A and E(Gc) the edges which
represent all pairs of compatible tasks, ie an edge exists
between Ai and Aj if and if only aj can be scheduled
between ai and bi (Fig. 1(b)). Note that compatibility
is symmetrical, thus here ai could be scheduled between
aj and bj.
A2
A1
A3
Compatibility graph
L2
L3
L1
a2
a1
b2
a3
b1
b3
Fig. 2 Link between the compatibility graph and the scheduling
The solution of an instance consists in determin-
ing the starting time of each sub-task ai of each task
Ai ∈ A. The tasks have to be processed on a single
processor while preserving the constraints given by the
compatibility graph (see Figure 2). Formally, we need
to find a valid schedule σ : A → N where the notation
σ(Ai) denotes the starting time of the task Ai. We use
the following abuse of notation: σ(ai) = σ(Ai) (resp.
σ(bi) = σ(Ai) + ai + Li) denotes the starting time of
the first sub-task ai (resp. the second sub-task bi).
Let Cmax = maxAi∈A(σ(Ai) + ai + Li + bi) be the
required time to complete all the tasks. Then the ob-
jective is to find a feasible schedule which minimizes
Cmax. We use the notation scheme αβγ proposed by
Graham and al. [10], where α denotes the environment
processors, β the characteristics of the jobs and γ the
criteria. The main problem denoted as Π will be defined
by:
Π = 1coupled − task, (ai, bi, Li), GcCmax
1.3 Related work
The problem of coupled-tasks has been studied in re-
gard to different conditions on the values of ai, bi, Li
for 1 ≤ i ≤ n, and precedence constraints [4,1,14,17].
Note that, in the previous works, all tasks are compati-
ble by considering a complete graph [4,1,14,17]. More-
over, in presence of any compatibility graph, we find
several complexity results [20,21,22], which are sum-
marized in Table 1. The notation ai = a implies that
for all 1 ≤ i ≤ n, ai is equal to a constant a ∈ N.
This notation can be extended to bi and Li with the
constants b, L and p ∈ N.
Problem
1coupled−task, (ai = bi = Li), GcCmax
Complexity
N P-complete
1coupled−task, (ai = a, bi = b, Li = L), GcCmax N P-complete
N P-complete
1coupled−task, (ai = bi = p, Li = L), GcCmax
O(n2m)
O(n2m)
1coupled−task, (ai = Li = p, bi), GcCmax
1coupled−task, (ai, bi = Li = p), GcCmax
Table 1 Complexity for scheduling problems with coupled-tasks
and compatibility constraints
ref
[20]
[20]
[22]
[20]
[20]
1.4 Contribution and organization of this paper
2 Computational complexity
3
Our work consists in measuring the impact of the com-
patibility graph on the complexity and approximation
of scheduling problems with coupled-tasks on a mono-
processor. In this way, we focus our work on establishing
the limits between polynomiality and NP-completeness
of these problems according to some parameters, when
the compatibility constraints is introduced. In [21,22],
we have studied the impact of the parameter L, and
have shown that the problem 1coupled − task, (ai =
bi = p, Li = L), GcCmax was NP-complete as soon as
L ≥ 2, and polynomial otherwise.
In this work, we complete complexity results with
the study of other special cases according to the value of
ai and bi, and we propose several approximation algo-
rithms for them. We restrict our study to a special case,
by adding new hypotheses to the processing time and
idle time of the tasks. For any task Ai, i ∈ {1, . . . , n},
the processing time ai (resp. bi) of sub-task ai (resp.
bi) is equal to a constant a (resp. b), and the length of
the idle time between ai and bi is L. Considering ho-
mogeneous tasks is a realistic hypothesis according to
the tasks that the torpedo has to execute. Let Π1 be
this new problem. Formally:
Π1 = 1coupled−task, (ai = a, bi = b, Li = L), GcCmax
PSfrag replacements
This paper is organized as follows: in section 2, we
establish the complexity of Π1 according to the val-
ues of a, b and L, and we show that the problem is
polynomial for any L < a + b; then we consider in
the rest of the paper that L = a + b. In that case,
the problem can be considered as a new graph problem
we call Minimum Schedule-Linked Disjoint-Path
Cover (Min-SLDPC). We present the proof of NP-
completeness of Min-SLDPC and we conclude this sec-
tion by the study of a specific sub-case with a = b =
L/2. In Section 3, we show that Min-SLDPC is imme-
diately 2-approximated by a simple approach: we design
a polynomial-time approximation algorithm with per-
formance guarantee lower than 3
2 . In fact, we show that
the approximation ratio obtained by this algorithm is
between 3
4 , according to the values of a and b.
The last section is devoted to the study of Π1 for some
particular topology of the graph Gc. First we present a
well-known graph problem, Minimum Disjoint-Path
Cover, (Min-DPC). Then we show the relation be-
tween Min-DPC and Min-SLDPC and evaluate how
results from the first one can be applied to solve the
second problem on specific topologies. This implies the
reduction of the performance ratio we can obtain on
some restricted instances from 3
2 and 5
2 to ≈ 1.37.
First, we prove that Π1 is polynomial when L < a + b:
it is obvious that a maximum matching in the graph
Gc gives an optimal solution. Indeed, during the idle
time L of a coupled-task Ai, we can process at most
one sub-task aj or bk. Since the idle time L is identical,
so it is obvious that finding an optimal solution consists
in computing a maximum matching. Thus, the problem
1coupled− task, (ai = a, bi = b, Li = L < a+ b), GcCmax
admits a polynomial-time algorithm with complexity
O(mp(n)) where n is the number of tasks and m the
number of edges of Gc (see [18]).
The rest of the paper is devoted to the case L =
a + b. Without loss of generality, we consider the case1
of b < a. The particular case b = a will be discussed in
subsection 2.2.
2.1 From a scheduling problem to a graph problem
Let us consider a valid schedule σ of an instance (A, Gc)
of Π1 with b < a, composed of a set of coupled-tasks
A and a compatibility graph Gc. For a given task Ai,
at most two sub-tasks may be scheduled between the
completion time of ai and the starting time of bi, and
in this case the only available schedule consists in exe-
cuting a sub-task bj and a sub-task ak during the idle
time Li with i 6= j 6= k such that σ(bj ) = σ(ai) + a and
σ(ak) = σ(ai) + a + b. Figure 3 shows a such configura-
tion.
a
aj
a
ai
b
bj
a
ak
b
bi
Li
b
b
bk
a
Fig. 3 At most 2 sub-tasks may be scheduled between ai and bi
We can conclude that any valid schedule σ can be
viewed as a partition {T1, T2, . . . , Tk} of A, such that
for any Ti the subgraph Pi = Gc[Ti] of Gc induced by
vertices Ti is a path (here, isolated vertices are con-
sidered as paths of length 0). Clearly, {P1, P2 . . . Pk} is
a partition of Gc into vertex-disjoint paths. Figure 4
shows an instance of Π1 (Figure 4(a)), a valid schedule
(Figure 4(c)) - not necessarily an optimal one -, and the
corresponding partition of Gc into vertex-disjoint paths
(Figure 4(b)).
For a given feasible schedule σ, let us analyse the
relation between the length of the schedule Cmax and
1 The results we present here can be symmetrically extended
to instances with b > a.
ak
bk
a
a + b
Ak
Li
Ai
Aj
b
a − b
ai
aj
bi
bj
idle time
idle time
(a) Chains of length 1 increase tidle by a
Aj
Ai
Ak
PSfrag replacements
a − b
a + b
b
a
aj
ai
bj
ak
bi
bk
Li
idle time
idle time
(b) Chains of length > 2 increase tidle by a + b
A7
A3
A4
A5
A6
(a) An input graph Gc with 7 tasks
A1
A2
A7
A3
A4
A5
A6
b3
a4
b4
PSfrag replacements
a5
a1
b5
b1
a6
a2
b6
b2
a7
a3
b7
b3
4
PSfrag replacements
A1
a4
b4
a5
b5
A2
a6
b6
a7
b7
A1
A2
A3
A4
A5
A6
A7
(b) A vertex-disjoint partition
Fig. 5 Impact of the length of the paths on the idle time
a1
a2
b1
a3
b2
a4
b3
b4 a5
a6
b5
b6 a7
b7
(c) A schedule on a monoprocessor
Fig. 4 Relation between a schedule and a partition into vertex-
disjoint paths
the corresponding partition {P1, P2, . . . Pk} into vertex-
disjoint paths. Clearly, we have Cmax = tseq + tidle
where tseq = n(a + b) and tidle is the inactivity time
of the processor. Since tseq is fixed for a given instance,
tidle obviously depends on the partition. We propose
the following lemma:
Lemma 1 Let {P1, P2, . . . Pk} be the partition of vertex-
disjoint paths corresponding to a schedule σ.
1. A path of length 0 corresponds to a single task sched-
uled in σ, tidle is incremented by L = a + b;
2. for any path of length 1, tidle is increased by a;
3. for any path of length strictly greater than 1, tidle is
incremented by (a + b).
Proof Point 1 is obvious. Fig. 5 illustrates points 2 and
3: a path of length 1 represents 2 tasks that may be
imbricated as on Figure 5(a). Paths of length strictly
greater than 1 represent more than two tasks. These
tasks can be scheduled in order to get an idle time of
length b at the beginning of the schedule and one of
length a at the end of it (as on Figure 5(b)). The reader
could check there is no other way to imbricate tasks in
order to reduce the idle time for paths of any length.
Thus, there exists a link between finding an op-
timal schedule and a graph problem which is called
Minimum Schedule-Linked Disjoint-Path Cover
(Min-SLDPC) defined in Table 2: Clearly, Min-SLDPC
is equivalent to Π1 with b < a and L = a+b, and can be
viewed as the graph problem formulation of a schedul-
ing problem. In any solution, each path increments the
Instance: a graph G = (V, E) of order n, two natural integers
a and b, b < a.
Result: a partition P of G into vertex-disjoint paths (can be
of length 0)
Objective: Minimize n(a+ b)+Pp∈P w(p) where w : P → N
is a cost function with w(p) = a if and only if E(p) = 1, and
w(p) = a + b otherwise.
Table 2 Minimum Schedule-Linked Disjoint-Path Cover
(Min-SLDPC)
cost of idle time by at least a (when the path has a
length 1), and at most a + b < 2a. So, we can deduce
that an optimal solution to Min-SLDPC consists in
finding a partition P with a particular cardinality k∗,
and a maximal number of paths of length 1 among all
possible k∗-partitions. The following immediate theo-
rem establishes the complexity of Min-SLDPC:
Theorem 1 Min-SLDPC is an NP-hard problem.
Proof We consider the decision problem associated to
Min-SLDPC. We will prove that the the problem of
deciding whether an instance of SLDPC has a schedule
of length at most (n + 1)(a + b) is NP-complete. Our
proof is based on the polynomial-time transformation
Hamiltonian Path ∝ SLDPC. We keep the graph
and the vertices is the task to schedule. Let us consider
a graph G.
This transformation can be clearly computed in poly-
nomial time.
• Assume that the length of the optimal schedule is
C opt
max = (n + 1)(a + b). We will prove that the graph
G possess a Hamiltonian path i.e. k = 1. Recall first
that k is the number of partition and that b < a.
We know, from the previous discussion, that tseq =
n(a + b) and that a chain of length one increase
tidle by a (see illustration given by Figures 5(b) and
5(a)). It is clear that the graph G must be covered
by paths of different lengths.
Suppose that the graph G is covered by k paths
with k1 paths of length one (the set of these paths
is denoted by P1), and k2 paths of lenght greater
than one (resp. by P≥2).
So the length of schedule given by this covering is:
processing times
idle time for P1
idle time for P≥2
C h
max =
n(a + b) +
z } {
a × k1 +
= (n + k2)(a + b) + k1a > C∗max
{
z } {
if (k16= 0 and k2≥ 1) or (k1 > 1 and k2 = 0)
(a + b)k2
}
z
Thus, the only schedule requiring exactly (n+1)(a+
b) units of time implies that the graph possess a
Hamiltonian path i.e. k1 = 0 and k2 = 1.
• Reciprocally, we suppose that the graph G possess
a Hamiltonian path, we will prove the existence of
a schedule of length Cmax = (n + 1)(a + b).
G contains an Hamiltonian path, we can deduce a
schedule with Cmax = (n + 1)(a + b): as tseq =
n(a + b), tidle must be equal to (a + b), which is
possible if and only if the schedule is represented
with only one chain.
2.2 A particular case Π2 : 1coupled − task, (ai = bi = p,
Li = L = 2p), GcCmax
In this subsection only, we suppose that both sub-tasks
are equal to a constant p and that the inactivity time
is equal to a constant L = 2p.
The previous proof cannot be used for this case.
Indeed the structure of these tasks allows to schedule
three compatible tasks together without idle time (see
Figure 6). Another solution consists in covering vertices
of Gc by triangles and paths (length 0 allowed), where
we minimize the number of paths and then maximize
the number of path of length 1.
PSfrag replacements
This problem is a generalization of Triangle Pack-
ing [8] since an optimal solution without idle time con-
sists in partitioning into triangles the vertices of Gc.
This problem is well known to be NP-complete, and
leads to the NP-completeness of problem Π2.
5
a covering of the graph Gc by triangles and paths, which
minimize the idle time. In the following section, we will
develop an efficient polynomial-time approximation al-
gorithm for the general problem Min-SLDPC.
3 Approximation algorithm for Min-SLDPC
Notice that the following algorithm, executing sequen-
tially the tasks, admits a ratio equal to two.
We also develop a polynomial-time 3
2 -approximation
algorithm based on a maximum matching in the graph
Gc. In fact, we show that this algorithm has an approx-
imation ratio of at most 3a+2b
2a+2b , which leads to a ratio
between 3
4 according to the values of a and b
(with b < a). This result, which depends on the values
a and b, will be discussed in Section 4, in order to pro-
pose a better ratio on some class of graphs.
2 and 5
max = tseq +topt
ule has a length C opt
For any instance of Min-SLDPC, an optimal sched-
idle where tseq = n(a+b).
Remark 1 For any solution of length Cmax, we neces-
sarily have2 tidle ≥ (a + b) and also3 tidle ≤ n(a + b).
Then, for any solution h of Min-SLDPC we have a
performance ratio ρ(h) such that:
ρ(h) ≤
C h
max
max ≤
C opt
2n(a + b)
(n + 1)(a + b)
< 2.
(1)
Indeed, we have C opt
max ≥ Tseq = n(a + b) + (a + b)
In the following, we develop a polynomial-time ap-
proximation algorithm based on a maximum matching
in the graph Gc, with performance guaranty in [ 5
4 , 3
2 ]
according to the values of a and b.
Let I be an instance of our problem. An optimal
solution is a disjoint-paths cover. The n vertices are
partitioned in three disjoint sets: n1 uncovered vertices,
n2 vertices covered by α2 = n2
2 paths of length 1, and n3
vertices covered by exactly α3 paths of length strictly
greater than 1 (see illustration Figure 7). The cost of
an optimal solution is equal to the sum of sequential
time and idle time:
processing times
idle time for matched vertices
n(a + b) +
z } {
z
}
{
(a + b)n1
+
2
a
z}{n2
z
+
(a + b)α3
}
{
A1
2p
C opt
max =
A2
A3
a1
a2
a3
b1
b2
b3
Fig. 6 Illustration of a schedule without idle time
A correct approximation algorithm for this problem
is an algorithm close to the general case. Indeed, finding
an optimal solution to this problem amounts to finding
idle time for isolated vertices
idle time for path of length > 1
2 The equality is obtained when the graph Gc possesses an
hamiltonian path, otherwise we need at least two paths to cover
Gc (where Gc is not only an edge), which leads to increase tidle
by at least 2a ≥ a + b units of time.
3 The worst case consists in executing tasks sequentially with-
out scheduling any sub-task aj or bj of task Aj during the idle
time of a task Ai.
6
α3 paths
n2 vertices
n1 vertices
PSfrag replacements
. . .
. . .
. . .
n3
vertices
. . .
Fig. 7 Illustration of the optimal solution on an instance I
Now, we propose a polynomial-time approximation
algorithm with non trivial ratio on an instance I. This
algorithm is based on a maximum matching in Gc in
order to process two coupled-tasks at a time. For two
coupled-tasks Ai and Aj connected by an edge of the
matching, we obtain an idle time of length a (see Figure
5(a)).
Let M∗ be the cardinality of a maximum matching.
In the worst case, the α3 paths are all odd and a match-
ing of the paths leaves α3 isolated vertices. So, we have
by hypothesis:
M∗ ≥
n2
2
+ (
n3
2 − α3) = Γ (worst case)
(2)
Indeed, based on the decomposition given by the
Figure 7, we may deduce a matching within this cardi-
nality: the α2 paths of length one are contained in the
matching; for any path of length greater than one, we
include odd edges to the matching. In the worst case, all
α3 paths have odd length and the number of uncovered
nodes is α3. If the tasks of the matching are processed
first and the isolated vertices in second, the length of
schedule is4:
processing times
idle time 1
idle time 2
C h
max ≤
n(a + b) +
z } {
z } {
a × Γ +
z
(a + b)(α3 + n1)
}
{
Since a optimal length is C opt
max = n(a + b) +n1(a +
2 a+ α3(a+ b), we obtain, with the last equation
max, the following ratio of the polynomial-
b)+ n2
involved C h
time approximation algorithm:
max + (
n3 − α3
2
)a
C h
max ≤ C opt
ρ(h) ≤ 1 +
ρ(h) ≤ 1 +
ρ(h) ≤ 1 +
ρ(h) ≤ 1 +
)a
n(a+b)+n1(a+b)+ n2
2 a +α3(a+b)
( n3−α3
2
( n3−α3
)a
2
(n + n1 + α3)(a+b)+ n2
2 a
n3
2 a
(n + n1)(a+b)+ n2
2 a
n
2 a
n3
2 a
n(a + b) ≤ 1 +
n(a + b)
, max obtained for α3 = 0
, since n3 ≤ n
ρ(h) ≤ 1 +
a
2(a + b)
=
3a + 2b
2a + 2b
4 Instances with particular topologies
We conclude this work by a study of Min-SLDPC
when Gc admits a particular topology. First we present
a related problem: Minimum Disjoint Path Cover
problem (Min-DPC). This problem has some inter-
esting results on restricted topologies. We establish a
link between Min-DPC and Min-SLDPC and we show
that finding a ρdpc-approximation for Min-DPC on
Gc allows to find a strategy with performance ratio
ρsldpc ≤ min{ρdpc × ( a+b
2a+2b}. This leads to pro-
pose, independently from the values a and b, a 1+√3
-
approximation for Min-SLDPC when Min-DPC can
be polynomially solved on Gc.
a ), 3a+2b
2
4.1 A related problem: Min-DPC
The graph problem Min-SLDPC is very close to the
well-known problem Minimum Disjoint Path Cover
(Min-DPC) which consists in covering the vertices of a
graph with a minimum number of vertex-disjoint paths5.
This problem has been studied in depth in several graph
classes: it is known that this problem is polynomial on
cographs [16], blocks graphs and bipartite permutation
graphs [23], distance-hereditary graph [12], and on in-
terval graphs [2]. In [5] and [9], the authors have pro-
posed (independently) a polynomial-time algorithm in
the case where the graph is a tree. Few years later,
in [13] the authors showed that this algorithm can be
implemented in linear time. Among the other results,
there is a polynomial-time algorithm for the cacti [15],
and another for the line graphs of a cactus [7]. In circular-
arc graphs the authors [11] have proposed an approxi-
mation algorithm of complexity O(n), which returns an
4 Idle time 1 (resp. 2) represents the idle time for matched
vertices (resp. isolated vertices).
5 Sometimes referenced as the Path-Partition problem (PP).
and let us evaluate it cost. Since each path of a Min-
SLDPC solution increments the cost of the solution by
at most a + b, then we have:
7
1
a
(5)
OP Tsldpc
according to (3)
b
a
a + b
OP Tsldpc + OP Tsldpc according to (4)
w(p) + n(a + b) ≤ (a + b)P∗1 + n(a + b)
Pp∈P ∗
≤ (a + b)P∗2 + n(a + b)
≤ bP∗2 + OP Tsldpc
≤
≤
The same proof may be applied if there exists a
ρdpc-approximation for Min-DPC, and then there ex-
ists a ρdpc×( a+b
a )-approximation for Min-SLDPC. Let
us suppose that we know a constant ρdpc such that
there exists a ρdpc-approximation for Min-DPC on Gc.
Let S1 be the strategy, which consists in determining a
ρdpc×( a+b
a )-approximation for Min-SLDPC from ρdpc,
and S2 the strategy, which consists in using the algo-
rithm introduced in section 3. Clearly, S1 is particu-
larly relevant when b is very small in comparison with
a. Whereas S2 gives better ratio when b is close to a.
Both strategies are complementary along the value of b,
which varies from 0 to a. Choosing the best result be-
tween the execution of S1 and S2, gives a performance
ratio ρsldpc such that:
a (cid:19) ,
ρsldpc ≤ min(cid:8)ρdpc ×(cid:18) a + b
3a + 2b
2a + 2b(cid:9).
(6)
optimal number of paths to a nearly constant additive
equal to 1.
The problem Min-DPC is directly linked to Hamil-
tonian Completion [8], which consists in finding the
minimum number of edges, noted HC(G), that must be
added to a given graph G, in order to make it hamilto-
nian (to guarantee the existence of a Hamiltonian cy-
cle). It is known that if G is not hamiltonian, then the
cardinality of a minimum disjoint path cover is clearly
equal to HC(G).
The dual of Min-DPC is Maximum Disjoint-Path
Cover [8]. It consists in finding in G a collection of
vertex-disjoint paths of length at least 1, which max-
imises the edges covered in G. This problem is known
to be 7
6 -approximable [3].
4.2 Relation between Min-DPC and Min-SLDPC
From the literature, we know that Min-DPC is polyno-
mial on trees [9,13,5], distance-hereditary graphs [12],
bipartite permutation graphs [23], cactis [15] and many
others classes. There are currently no result about the
complexity of Min-SLDPC on such graphs: since the
values of of a and b have a high impact, techniques
used to prove the polynomiality of Min-DPC cannot
be adapted to prove the polynomiality of Min-SLDPC.
Despite all our effort, the complexity of Min-SLDPC
remains an open problem. However using known results
on Min-DPC, we show how the approximation ratio
can be decreased for these class of graphs. We propose
the following lemma:
Lemma 2 If Min-DPC can be solved polynomial com-
putation time, then there exists a polynomial-time (a+b)
-
approximation for Min-SLDPC.
a
Proof Let I1 = (G) be an instance of Min-DPC, and
I2 = (G, a, b) an instance of Min-SLDPC. Let P∗1 be
an optimal solution of Min-DPC of cost P∗1 , and P∗2
an optimal solution of Min-SLDPC of cost OP Tsldpc.
According to the definition of Min-SLDPC, we have
P∗2 ≥ P∗16. Since each path of a Min-SLDPC solu-
tion increments the cost of the solution by at least a,
then we have:
OP Tsldpc = Xp∈P ∗
2
w(p)+n(a+b)
≥ aP∗2+n(a+b)≥ aP∗2
⇒
≥ P∗2
OP Tsldpc
(4)
Let us consider the partition P∗1 as a solution (not
necessarily optimal) to the instance I2 of Min-SLDPC,
a
Compared to executing S1 only, this new strategy
increases the obtained results if and only if ρdpc is lower
PSfrag replacements
than 3
2 (see Figure 8).
approximation ratio
3
2
bound
5
4
ρdpc
1
0
S1
S2
a
4
a
2
3a
4
a
value of b
Fig. 8 Finding a good ρdcp-approximation helps to increase the
results of Section 3
(3)
We propose the following remark, which is not good
news:
6 The best solution for Min-SLDPC is not necessarily a solu-
tion with a minimum cardinality of k.
Remark 2 There is no ρdpc-approximation for Min-DPC
in general graphs for some ρdpc < 2
8
This result is a consequence of the Impossibility
Theorem [6]. It can be checked by considering an ins-
tance of Min-DPC which has an hamiltonian path:
the optimal solution of Min-DPC has cost 1, thus any
polynomial-time ρdpc - approximation algorithm with
ρdpc < 2 will return a solution of cost 1, which is not
allowed under the assumption that P 6= NP. This re-
sult also implies that constant-factor approximation al-
gorithms for max-DPC do not necessarily give the same
performance guarantees on min-DCP, since the best ap-
proximation ratio for max-DCP is 7
6 , which is lower
than the inapproximability bound for min-DCP.
It is good news that Min-DPC is polynomial for
some compatibility graphs such as trees [9,13,5], distance-
hereditary graphs [12], bipartite permutation graphs
[23], cactis [15] and many others classes; thus ρdpc = 1.
For all these graphs we obtain an approximation ratio
of min{ (a+b)
a =
a
3a+2b
2a+2b , i.e.:
2a+2b} which is maximal when (a+b)
, 3a+2b
(a + b)
=
3a + 2b
2a + 2b ⇔ −a2 + 2ab + 2b2 = 0.
(7)
a
The only solution of this equation with a and b ≥ 0
is a = b(1 + √3). By replacing a by this new value on
2a+2b , we show that in the worst case the
2 ≈
approximation ratio is reduced from 3
1.37.
2 down to 1+√3
or on 3a+2b
(a+b)
a
5 Conclusion
We investigate a particular coupled-tasks scheduling
problem Π1 in presence of a compatibility graph. We
have shown how this scheduling problem can be reduced
to a graph problem. We have proved that adding the
compatibility graph leads to the NP-completeness of
Π1, whereas the problem is obviously polynomial when
there is a complete compatibility graph (each task is
compatible with each other). We have proposed a ρ-
approximation of Π1 where ρ is between 3
4 ac-
cording to value of a and b. We have also decreased the
upper bound of 3
2 down to ≈ 1.37 on instances where
Minimum Disjoint Path Cover can be polynomially
solved on the compatibility graph.
2 and 5
As perspectives of this work, we plan to test the
pertinence of our polynomial-time approximation algo-
rithm through simulations in order to determine the av-
erage gap between the optimal solution and the results
obtained with our strategy on significant instances. We
also aim to classify the complexity of other configura-
tions, especially Π1 : 1coupled − task, (ai = a, bi =
b, Li = L)Cmax with a complete compatibility graph.
6 Acknowledgements
Thanks to our reviewers for the thorough review of this
paper and many helpful comments and suggestions.
References
1. D. Ahr, J. Békési, G. Galambos, M. Oswald, and G. Reinelt.
An exact algorithm for scheduling identical coupled-tasks.
Mathematical Methods of Operations Research, 59:193 --
203(11), June 2004.
2. S. Rao Arikati and C. Pandu Rangan. Linear algorithm for
optimal path cover problem on interval graphs. Information
Processing Letters, 35(3):149 -- 153, 1990.
3. P. Berman and M. Karpinski. 8/7-approximation algorithm
for (1,2)-TSP. In SODA '06: Proceedings of the seventeenth
annual ACM-SIAM symposium on Discrete algorithm, pages
641 -- 648, New York, NY, USA, 2006. ACM.
4. J. Blażewicz, K. Ecker, T. Kis, C.N. Potts, M. Tanas, and
J. Whitehead. Scheduling of coupled tasks with unit process-
ing times. Technical report, Poznan University of Technol-
ogy, 2009.
5. F. T. Boesch S. Chen and B. McHugh. On covering the points
of a graph with point-disjoint paths. Graphs and Combina-
torics, 406:201 -- 212, 1974.
6. Ph. Chrétienne and C. Picouleau. Scheduling with commu-
nication delays: a survey. In Scheduling theory and its appli-
cations, pages 641 -- 648. John Wiley & Sons, 1995.
7. P. Detti and C. Meloni. A linear algorithm for the hamilto-
nian completion number of the line graph of a cactus. Dis-
crete Applied Mathematics, 136(2-3):197 -- 215, 2004.
8. M. R. Garey and D. S. Johnson. Computers and Intractabil-
ity: A guide to the theory of NP-completeness. Freeman,
1979.
9. S. E. Goodman, S. T. Hedetniemi, and P. J. Slater. Ad-
vances on the hamiltonian completion problem. Journal of
the ACM, 22(3):352 -- 360, 1975.
10. R.L. Graham, E.L. Lawler, J.K. Lenstra, and A.H.G. Rin-
nooy Kan. Optimization and approximation in determinis-
tic sequencing and scheduling: a survey. Annals of Discrete
Mathematics, 5:287 -- 326, 1979.
11. R.-W. Hung and M.-S. Chang. Solving the path cover prob-
lem on circular-arc graphs by using an approximation algo-
rithm. Discrete Applied Mathematics, 154(1):76 -- 105, 2006.
12. R.-W. Hung and M.-S. Chang. Finding a minimum path
cover of a distance-hereditary graph in polynomial time. Dis-
crete Applied Mathematics, 155(17):2242 -- 2256, 2007.
13. S. Kundu. A linear algorithm for the hamiltonian completion
number of a tree. Information Processing Letters, 5:55 -- 57,
1976.
14. V. Lehoux-Lebacque, N. Brauner, and G. Finke.
Identical
coupled task scheduling: polynomial complexity of the cyclic
case. Les Cahiers Leibniz, 179, 2009.
15. S. Moran and Y. Wolfstahl. Optimal covering of cacti by
vertex-disjoint paths. Theoretical Computer Science, 84:179 --
197, 1988.
16. K. Nakano, S. Olariu, and A. Y. Zomaya. A time-optimal
solution for the path cover problem on cographs. Theoretical
Computer Science, 290(3):1541 -- 1556, 2003.
17. A.J. Orman and C.N. Potts. On the complexity of coupled-
task scheduling. Discrete Applied Mathematics, 72:141 -- 154,
1997.
18. A. Schrijver. Combinatorial Optimization : Polyhedra and
Efficiency (Algorithms and Combinatorics). Springer, July
2004.
9
19. R.D. Shapiro. Scheduling coupled tasks. Naval Research
Logistics Quarterly, 27:477 -- 481, 1980.
20. G. Simonin. Impact du graphe de compatibilité sur la com-
plexité et l'approximation des problèmes d'ordonnancement
en présence de tâches-couplées. Ph.D thesis, LIRMM, De-
cember 2009.
21. G. Simonin, R.Giroudeau, and J.-C. König. Complexity and
approximation for scheduling problem for a torpedo. CIE'39
: The 39th International Conference on Computers and In-
dustrial Engineering, IEEE, Troyes, France, pages 300 -- 304,
2009.
22. G. Simonin, R.Giroudeau, and J.-C. König.
Extended
matching problem for a coupled-tasks scheduling problem.
TMFCS'09 : International Conference on Theoretical and
Mathematical Foundations of Computer Science, Orlando,
Florida, pages 082 -- 089, 2009.
23. R. Srikant, R. Sundaram, K. Sher Singh, and C. Pandu Ran-
gan. Optimal path cover problem on block graphs and bi-
partite permutation graphs. Theoretical Computer Science,
115(2):351 -- 357, 1993.
|
1211.6195 | 1 | 1211 | 2012-11-27T02:37:21 | Time-Darts: A Data Structure for Verification of Closed Timed Automata | [
"cs.DS",
"cs.LO"
] | Symbolic data structures for model checking timed systems have been subject to a significant research, with Difference Bound Matrices (DBMs) still being the preferred data structure in several mature verification tools. In comparison, discretization offers an easy alternative, with all operations having linear-time complexity in the number of clocks, and yet valid for a large class of closed systems. Unfortunately, fine-grained discretization causes itself a state-space explosion. We introduce a new data structure called time-darts for the symbolic representation of state-spaces of timed automata. Compared with the complete discretization, a single time-dart allows to represent an arbitrary large set of states, yet the time complexity of operations on time-darts remain linear in the number of clocks. We prove the correctness of the suggested reachability algorithm and perform several experiments in order to compare the performance of time-darts and the complete discretization. The main conclusion is that in all our experiments the time-dart method outperforms the complete discretization and it scales significantly better for models with larger constants. | cs.DS | cs | Time-Darts: A Data Structure for Veri fication of
Closed Timed Automata∗
Kenneth Y. Jørgensen
Ji r´ı Srba
Kim G. Larsen
Department of Computer Science, Aalborg University
Selma Lagerl ofs Vej 300, 9220 Aalborg East, Denmark
{kyrke,kgl,srba}@cs.aau.dk
Symbolic data structures for model checking timed systems have been subject to a significant re-
search, with Difference Bound Matrices (DBMs) still being the preferred data structure in several
mature verification tools. In comparison, discretization o ffers an easy alternative, with all opera-
tions having linear-time complexity in the number of clocks, and yet valid for a large class of closed
systems. Unfortunately, fine-grained discretization caus es itself a state-space explosion. We intro-
duce a new data structure called time-darts for the symbolic representation of state-spaces of timed
automata. Compared with the complete discretization, a single time-dart allows to represent an ar-
bitrary large set of states, yet the time complexity of operations on time-darts remain linear in the
number of clocks. We prove the correctness of the suggested reachability algorithm and perform sev-
eral experiments in order to compare the performance of time-darts and the complete discretization.
The main conclusion is that in all our experiments the time-dart method outperforms the complete
discretization and it scales significantly better for model s with larger constants.
1
Introduction
Timed automata [2] are a well studied formalism for modelling and verification of real-time systems.
Over the years extensive research effort has been made towards the design of data structures and algo-
rithms allowing for efficient model checking of this modelin g formalism. These techniques have by now
been implemented in a number of mature tools (e.g. U PPAA L [4], IF [9], Kronos [14], PAT [21], Rabbit
[6], RED [16]), with zone-based analysis [15, 5] still being predominant, stemming from that fact that
Difference Bound Matrices (DBMs) offer a very compact data structure for efficient implementation of
the various operations required for the state-space exploration. Still the DBM data structure suffers from
the fact that all operations have at least quadratic—and the
crucial closure operation even cubic—time
complexity in the number of clocks (though for diagonal-free constraints the operations can be imple-
mented in quadratic time [23]). In contrast, as advocated in [10, 19], the use of discretization offers an
easy alternative, with all operations having linear complexity in the number of clocks, and yet valid for
the large—and in practice often sufficient—class of closed s
ystems that contain only nonstrict guards;
moreover for reachability checking the continuous and discrete semantics coincide on this subclass.
As an example consider the timed automaton shown in Figure 1, containing n clocks and n self-loops
where the i’th loop has the guard xi = i and resets the clock xi . We are interested in whether or not we
can reach the Goal location. For this to happen, all clocks x1 , . . . , xn must simultaneously have the value
zero, corresponding effectively to calculating the least common multiple of the numbers from 1 to n. In
Figure 1 we compare the verification times of the zone-based r eachability performed in U PPAA L with
that of a simple Python based implementation of discrete time reachability checker for timed automata.
∗The paper was supported by VKR Center of Excellence MT-LAB.
F. Cassez, R. Huuck, G. Klein and B. Schlich (eds.):
Systems Software Veri fication Conference 2012 (SSV 2012)
EPTCS 102, 2012, pp. 141–155 , doi:10.4204/EPTCS.102.13
c(cid:13) K.Y. Jørgensen, K.G. Larsen & J. Srba
This work is licensed under the
Creative Commons Attribution License.
142
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
x2 = 2, x2 := 0
· · ·
x1 = 1, x1 := 0
xn = n, xn := 0
x1 = x2 = · · · = xn = 0 ∧ y ≥ 1
Goal
Size
4
5
6
7
8
Discrete Uppaal
<0.1
0.2
1.1
2.2
27.5
5.5
>300
97
out of memory
-
Figure 1: Discrete vs. zone-based reachability algorithm (time in seconds)
Opposite to what one might expect, it turns out that in this case the naive discrete implementation without
any speed optimizations outperforms a state-of-the-art model checking tool.
On the other hand, the disadvantage of discretization is that the number of states to be considered ex-
plodes when the size of the constants appearing in the constraints of the timed automaton are increased.
In fact, the experimental results of Lamport [19] show that the zone-based methods outperform discreter-
ized methods when the maximum constant in the timed automaton exceeds 10. Also in [19] the BDD-
based model checker SMV was applied to symbolically represent the discreterized state-space. This
representation is less sensitive to the maximum constant of the model, yet in experimental results [7, 3]
it appears that the zone-based method is still superior for constants larger than 16.
Inspired by the success of discretization reported in Figure 1, we revisit the problem of finding effi-
cient data structures for the analysis of timed automata. In particular, we introduce a new data structure
called time-darts for the symbolic representation of the state-spaces of timed automata. Compared with
the complete discretization, a single time-dart allows us to represent an arbitrary large set of states, yet
the time complexity of operations remain linear in the number of clocks, providing a potential advantage
compared to DBMs.
We propose a symbolic reachability algorithm based on a forward search. To ensure the termination
of the forward search the so-called extrapolation of time darts with respect to the maximum constant
appearing in the model is required. Given the subtleties of extrapolation,1 we prove the termination
and correctness of the proposed algorithm. We perform several experiments in order to compare the
performance of time-darts versus the complete discretization representation. The main conclusion is
that the time-dart method consistently outperforms the complete discretization and it is particularly well
suited for scaling up the constants used in the model. Given the simplicity of implementing discrete-
time algorithms compared to the DBM-based ones, our method can be in practice well suited for the
verification of closed time systems with moderately large co nstants.
2 Timed Automata
¥
= N ∪ {¥}. The comparison and addition operators
Let N be the set of nonnegative integers and let N
are defined as expected, in particular n < ¥ and n + ¥ = ¥ for n ∈ N.
A Discrete Timed Transition System (DTTS) is a pair T = (S, −→) where S is a set of states, and
−→⊆ S × (N ∪ {t}) × S is a transition relation written s
d
−→ s′ if (s, d , s′ ) ∈−→ where d ∈ N for delay
1Despite several earlier claims, it was not before [8] that a complete—and a quite non-trivial —proof of correctness of
zone-based forward reachability was given.
K.Y. Jørgensen, K.G.Larsen& J.Srba
143
t
−→ s′ if (s, t, s′ ) ∈−→ for switch actions. By −→∗ we denote the reflexive and transitive
actions, and s
t
def
d
closure of the relation −→
−→.
=
−→ ∪ Sd∈N
Let C be a finite set of clocks. A (discrete) clock valuation of clocks from C is a function v : C → N.
The set of all clock valuations is denoted by V . Let v ∈ V . We define the valuation v + d after a delay
def
= v(x) + d for every x ∈ C . For a subset R ⊆ C of clocks we define
of d ∈ N time units by (v + d )(x)
def
the valuation v[R := 0] where all clocks from R are reset to zero by v[R := 0](x)
= v(x) for x ∈ C \ R and
def
v[R := 0](x)
= 0 for x ∈ R.
A nonstrict (or closed) time interval I is of the form [a, b] or [a, ¥) where a, b ∈ N and a ≤ b. The
set of all time intervals is denoted by I . We use the functions ub, lb : I −→ N to return the upper resp.
lower bound of a given interval. A clock guard over the set of clocks C is a function g : C −→ I that
assigns a time interval to each clock. We denote the set of all clock guards over C by G (C). We write
v = g for a valuation v ∈ V and a guard g ∈ G (C) whenever v(x) ∈ g(x) for all x ∈ C .
Timed Automaton A timed automaton (TA) is a tuple A = (L,C, −→, ℓ0 ) where L is a finite set of
g,R
locations, C is a finite set of clocks, −→⊆ L × G (C) × 2C × L is a finite transition relation written ℓ
−→ ℓ′
for (ℓ, g, R, ℓ′ ) ∈−→, and ℓ0 ∈ L is an initial location.
Note that we do not consider clock invariants as they can be substituted by adding corresponding clock
guards to the outgoing transitions while preserving the answers to location-reachability checking.
A configuration of a timed automaton A is a pair (ℓ, v) where ℓ ∈ L and v ∈ V . We denote the set
def
of all configurations of A by Conf (A). The initial configuration of A is (ℓ0 , v0 ) where v0 (x)
= 0 for all
x ∈ C .
(ℓ, v)
(ℓ, v)
def
= (Conf (A), −→DS ) where
Discrete Semantics A TA A = (L,C, −→, ℓ0 ) generates a DTTS TDS (A)
states are configurations of A and the transitions are given by
t
g,R
−→DS (ℓ′ , v[R := 0])
−→ ℓ′ such that v = g
if ℓ
d
if d ∈ N.
−→DS (ℓ, v + d )
The discrete semantics clearly yields an infinite state spac e due to unbounded time delays. We will
now recall that the reachability problem for a TA A can be solved by looking only at a finite prefix of
the state space up to some constant determining the largest possible delay. Let MC be the largest integer
that appears in any guard of A. Two valuations v, v′ ∈ V are equivalent up to the maximal constant MC,
written v ≡MC v′ , if
∀x ∈ C. v(x) = v′ (x) ∨ (v(x) > MC ∧ v′ (x) > MC).
Observe that the equivalence relation ≡MC has only finitely many equivalence classes as there are
finitely many clocks and each of them is bounded by the constan t MC.
Lemma 2.1 Let v, v′ ∈ V s.t. v ≡MC v′ and let g ∈ G (C) be a guard where 0 ≤ lb(g(x)) ≤ MC, and
ub(g(x)) = ¥ or 0 ≤ ub(g(x)) ≤ MC for all x ∈ C . Then v = g iff v′ = g.
Moreover, any two configurations with the same location and e quivalent valuations are timed bisim-
ilar (for the definition of timed bisimilarity see e.g [20]).
Lemma 2.2 The relation B = {((ℓ, v), (ℓ, v′ ) v ≡MC v′ } is a timed bisimulation for any timed automaton
with its maximum constant MC .
Proof Let ((ℓ, v), (ℓ, v′ )) ∈ B. We analyse only the switch and delay actions from (ℓ, v); the situation for
the transitions from (ℓ, v′ ) is symmetric.
144
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
t
g,R
−→DS (ℓ′ , v[R := 0]) via a transition ℓ
−→ ℓ′ . Due to Lemma 2.1 and the fact
• Assume that (ℓ, v)
t
−→DS (ℓ′ , v′ [R := 0]) and it is easy to verify that
that v = g, we get v′ = g. Hence also (ℓ, v′ )
v[R := 0] ≡MC v′ [R := 0].
d
d
−→DS (ℓ, v + d ). We want to argue that also (ℓ, v′ )
−→DS (ℓ, v′ + d ) such that
• Assume that (ℓ, v)
v + d ≡MC v′ + d , however, this is easy to see from that facts that (i) if v(x), v′ (x) > MC then also
(v + d )(x), (v′ + d )(x) > MC and (ii) if v(x) = v′ (x) ≤ MC then (v + d )(x) = (v′ + d )(x).
We now define an alternative discrete semantics of TA with onl y finitely many reachable configura-
tions. First, for the maximum constant MC, we define a bounded addition operator
= (MC + 1 if n + m > MC ,
n ⊕MC m def
otherwise.
n + m
The operation ⊕MC is in a natural way extended to functions and tuples.
Bounded Discrete Semantics A TA A = (L,C, −→, ℓ0 ) with the maximal constant MC generates a
def
= (Conf (A), −→BDS ) where states are configurations of A and the transition relation
DTTS TBDS (A)
−→ is defined by
(ℓ, v)
(ℓ, v)
t
−→BDS (ℓ′ , v[R := 0])
d
−→BDS (ℓ, v ⊕MC d )
g,R
−→ ℓ′ such that v = g
if ℓ
if d ∈ N.
We say that a location ℓg is reachable in TDS (A) resp.
valuation v where −→ is −→DS resp. −→BDS .
We conclude that the bounded semantics preserves reachability of locations, the main problem we
are interested in. This fact follows from Lemma 2.2.
in TBDS (A) if (ℓ0 , v0 ) −→∗ (ℓg , v) for some
Theorem 2.3 A location ℓ is reachable in TDS (A) iff ℓ is reachable in TBDS (A).
3 Naive Reachability Algorithm
We can now describe the naive search algorithm that explores in a standard way, point by point, the
finite state-space of the bounded semantics and provides the
answer to the location reachability prob-
lem. Algorithm 1 searches through all reachable states, starting from the initial location, until a goal
configuration is found (returning true) or all configuration
s are visited (returning false). Notice that the
algorithm is nondeterministic as it is not specified what ele ment should be removed from Waiting at
line 5 (such choice depends on the concrete search strategy like DFS or BFS). The next theorem states
that Algorithm 1 is correct.
Theorem 3.1 Let A be a timed automaton and let ℓg be a location. Algorithm 1 terminates, and it returns
true iff ℓg is reachable in the discrete semantics TDS (A).
Proof First notice that the algorithm terminates because there is only a finite number of configurations
that can be possibly added to Waiting: the number of locations is finite and due to the bounded addit ion
at line 9 the total number of configurations is finite too. When
ever a configuration is removed from the
set Waiting, it is added to Passed (line 6) and can never be inserted into Waiting again due to the test
at line 12. As we remove one element from Waiting each time the body of the while-loop is executed,
K.Y. Jørgensen, K.G.Larsen& J.Srba
145
Algorithm 1: Naive reachability algorithm
Input: A timed automaton A = (L,C, −→, ℓ0 ) and a location ℓg ∈ L
Output: true if ℓg is reachable in TDS (A), false otherwise
1 begin
Passed := /0; Waiting := /0;
2
AddToPW(ℓ0, v0 )
3
while Waiting 6= /0 do
4
remove some (ℓ, v) from Waiting
5
Passed := Passed ∪ {(ℓ, v)}
6
forall the (ℓ′ , v′ ) such that (ℓ, v)
AddToPW(ℓ′, v′ )
AddToPW(ℓ, (v ⊕MC 1))
return false
t
−→BDS (ℓ′ , v′ ) do
7
8
9
10
11 AddToPW(ℓ, v)
12 if (ℓ, v) /∈ Passed ∪ Waiting then
if ℓ = ℓg then
13
return true
14
else
15
Waiting := Waiting ∪ {(ℓ, v)}
16
/* and terminate the whole algorithm */
the algorithm necessarily terminates either at line 10 or even earlier at line 14 if the goal location is
reachable.
Now we prove the correctness part. By Theorem 2.3 we can equivalently argue that the algorithm
returns true iff ℓg is reachable in TBDS (A).
“ ⇒ ”: Assume that Algorithm 1 returns true. We want to show that t he location ℓg is reachable in
TBDS (A). This can be established by the following invariant: any call of AddToPW with the argument
(ℓ, v) implies that the configuration (ℓ, v) is reachable in TBDS (A). For the initialisation at line 3 this
clearly holds. In the while-loop the calls to AddToPW are at lines 8 and 9. At line 8 we know by the
t
−→BDS (ℓ′ , v′ ),
invariant that (ℓ, v) is reachable and we call AddToPW only with (ℓ′ , v′ ) such that (ℓ, v)
so the invariant is preserved. Similarly at line 9 for the argument (ℓ, (v ⊕MC 1)) of AddToPW holds that
1
−→BDS (ℓ, (v ⊕MC 1)), so it is reachable as well.
(ℓ, v)
(ℓ′ , v′ ) is reachable via n transitions in TBDS (A), formally
“ ⇐ ”: Assume that a configuration
(ℓ0 , v0 ) −→n
BDS (ℓ′ , v′ ), where (without loss of generality) all delay transitions in the sequence are of
the form 1
−→BDS , in other words they add exactly one-unit time delay. By induction on n we will estab-
lish that during any execution of the algorithm there is eventually a call of AddToPW with the argument
(ℓ′ , v′ ), unless the algorithm already returned true. If n = 0 then the claim is trivial due to the call at
t
line 3. If n > 0 then either (i) (ℓ0 , v0 ) −→n−1
−→BDS (ℓ′ , v′ ) with the last switch action or (ii)
BDS (ℓ, v)
1
(ℓ0 , v0 ) −→n−1
−→BDS (ℓ′ , v′ ) with the last one-unit delay action. By induction hypothesis, unless
BDS (ℓ, v)
the algorithm already returned true, there will be eventually a call of AddToPW with the argument (ℓ, v)
and this element is added to the set Waiting. Because the algorithm terminates, the element (ℓ, v) will
be eventually removed from Waiting at line 5 and its switch successors and the one-unit delay successor
will become arguments of the call to AddToPW at lines 8 and 9. Hence the induction hypothesis for the
146
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
y
k
c
o
l
c
time-dart ((2, 0), 2, 5)
p
w
g
clock x
Figure 2: A time-dart (g, w, p) where g(x) = 2, g(y) = 0, w = 2 and p = 5
cases (i) and (ii) is established.
4 Time-Dart Data Structure
We shall now present a novel symbolic representation of the discrete state space. The symbolic structure,
we call it a time-dart, allows us to represent a number of concrete configurations i n a more compact way
so that time successors of a configuration are stored without being explicitly enumerated. We start with
the definition of an anchor point, denoting the beginning of a time-dart.
An anchor point over a set of clocks C is a clock valuation g : C −→ N where g(x) = 0 for at least
one x ∈ C . We denote the set of all anchor points over a set of clocks C by Anchors(C). Now we are
ready to define time-darts.
Time-Dart A time-dart over a set of clocks C is a triple (g, w, p) where g ∈ Anchors(C) is an anchor
¥ is a passed distance such that w ≤ p.
point, w ∈ N is a waiting distance, and p ∈ N
The intuition is that a time-dart describes the corresponding passed and waiting sets in a given lo-
cation. Figure 2 shows a dart example with two clocks x and y, anchor point (2, 0), waiting distance
2 and passed distance 5. The empty circles represent the points in the waiting set and the filled circles
represent the points in the passed set, formally defined by: Waiting(g, w, p) = {(g+ d ) w ≤ d < p} and
Passed (g, w, p) = {(g+ d ) d ≥ p}.
The passed-waiting list is represented as a function from locations and anchor points to the corre-
sponding waiting and passed distances (here ⊥ represents the undefined value):
PW : L × Anchors −→ (N × N
¥
) ∪ {⊥} .
Such a structure can be conveniently implemented as a hash map. A given passed-waiting list PW defines
the sets of passed and waiting configurations.
Waiting(PW ) = {(ℓ, v) ∃g.PW (ℓ, g) = (w, p) 6= ⊥ and v ∈ Waiting(g, w, p)}
Passed (PW ) = {(ℓ, v) ∃g.PW (ℓ, g) = (w, p) 6= ⊥ and v ∈ Passed (g, w, p)}
K.Y. Jørgensen, K.G.Larsen& J.Srba
147
Algorithm 2: Time-dart reachability algorithm
Input: A timed automaton A = (L,C, −→, ℓ0 ) and a location ℓg ∈ L
Output: true if ℓg is reachable in TDS (A), false otherwise
1 begin
PW (ℓ, g) := ⊥ for all (ℓ, g) /* default value */
2
AddToPW(ℓ0, g0 , 0, ¥) where g0 (x) := 0 for all x ∈ C
3
while ∃(ℓ, g). PW (ℓ, g) = (w, p) and w < p do
4
PW (ℓ, g) := (w, w)
5
foreach (ℓ, g, R, ℓ′ ) ∈−→ do
6
start := max(w, max({lb(g(x)) − g(x) x ∈ C}))
7
end := min({ub(g(x)) − g(x) x ∈ C})
8
if (start < p ∧ start ≤ end ) then
9
if R = /0 then
10
AddToPW(ℓ′, (g⊕MC start ) − start, start , ¥)
11
else
stop := max{start , MC + 1 − minx∈CrR g(x)}
for n := start to min(end , p − 1, stop) do
AddToPW(ℓ′ , (g⊕MC n)[R := 0], 0, ¥)
12
13
14
15
16
return false
17 AddToPW(ℓ, g, w, p)
18 if ℓ = ℓg then
return true
/* and terminate the whole algorithm */
19
20 if PW (ℓ, g) = ⊥ then
PW (ℓ, g) := (w, p)
21
22 else
(w′ , p′ ) := PW (ℓ, g)
23
PW (ℓ, g) := (min(w, w′ ), min( p, p′ ))
24
5 Reachability Algorithm Based on Time-Darts
We can now present Algorithm 2 showing us how time-darts can be used to compute the set of reachable
states of a timed automaton in a compact and efficient way. The algorithm repeatedly selects from the
waiting list a location ℓ with a time-dart (g, w, p) that still contains some unexplored points (w < p).
g,R
−→ ℓ′ in the timed automaton it computes the start and end delays from the anchor
Then for each edge ℓ
point such that start is the minimum delay where the guard g gets first enabled and end is the maximum
possible delay so that g is still enabled. Depending on the concrete situation it will add a new time-dart
(or a set of darts) with location ℓ′ to the waiting list by calling AddToPW. A switch transition is always
followed by a delay transition that is computed symbolically (including in a single step all possible
delays). There are several cases that determine what kinds of new time-darts are generated. Figure 3
gives a graphical overview of the different situations. In Figure 3a we illustrate the produced time-dart
that serves as the argument for the call to AddToPW at line 11 of the algorithm (no clocks are reset). Here
the anchor point g is not modified because ((g⊕MC start) − start ) = g. Figure 3b shows another example
148
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
y
((2, 0), 2, 5)
MC
((2, 0), 4, ¥)
y
MC
4
ℓ
y
MC
4
3 ≤ x ≤ 7 ∧ y ≥ 4
x
MC
7
g 3
start = 4, end = 5
g⊕MC start = (6, 4)
g⊕MC end = (7, 5)
ℓ′
g
6
MC
x
(a) Unchanged anchor point
((3, 0), 2, 5)
y
((2, 0), 4, ¥)
MC
x ≥ 4 ∧ y ≥ 4
x
ℓ′
gg′
MC
x
g
ℓ
MC = 5
start = 4, end = ¥
g+ start = (7, 4)
g⊕MC start = (6, 4)
(g⊕MC start) − start = (2, 0) = g′
((2, 0), 2, 5)
MC
y
6
(b) Shift of anchor point
((4, 0), 0, ¥)
((5, 0), 0, ¥)
((6, 0), 0, ¥)
y
MC
y ≤ 6, y:=0
ℓ
g 4 5 6
start = 2, end = 6
x
MC
ℓ′
4 5 6
x
MC
(c) Reset of a clock
Figure 3: Successor generation for a selected time-dart
K.Y. Jørgensen, K.G.Larsen& J.Srba
149
x := 0
x ≥ 2, y ≥ 2 x := y := 0
x ≥ 2
ℓ0
ℓ1
x ≥ 1
ℓ2
x ≤ 1, y ≥ 2
ℓ3
Location Anchor
(0, 0)
ℓ0
(0, 0)
ℓ1
(0, 1)
ℓ1
(0, 2)
ℓ1
(0, 3)
ℓ1
(0, 0)
ℓ2
2
1
0
(0, ¥)
(0, 0)
(0, 0)
⊥ (2, ¥)
(2, 2)
⊥
⊥
⊥
⊥ (0, ¥)
⊥
⊥ (0, ¥)
⊥
⊥ (0, ¥)
⊥
3
(0, 0)
(2, 2)
⊥
(0, 0)
(0, ¥)
(0, ¥)
4
(0, 0)
(2, 2)
⊥
(0, 0)
(0, 0)
(0, ¥)
6
5
(0, 0)
(0, 0)
(1, 2)
(1, 1)
⊥ (0, ¥)
(0, 0)
(0, 0)
(0, 0)
(0, 0)
(0, 0)
(0, 0)
7
(0, 0)
(1, 1)
(0, 0)
(0, 0)
(0, 0)
(0, 0)
Figure 4: Example of an execution of the algorithm (columns represent the number of iterations of the
main while-loop; all unlisted pairs of locations and anchor points are constantly having the value ⊥)
of a call at line 11 where the anchor point changes. Finally, Figure 3c explains the case where some
clocks are reset and several new darts are added in the body of the for-loop at line 15 of the algorithm
(the for-loop starts from start and stops as soon as either end , the beginning of the passed list, or the
number stop —used for performance optimization–is reached). We note th
at the figures show the time-
darts that the function AddToPW is called with; inside the function the information already stored in the
passed-waiting list for the concrete anchor point and location is updated so that we take the minimum of
the current and new waiting and passed distances (line 24 of the algorithm).
Let us now demonstrate the execution of Algorithm 2 on the automaton depicted in Figure 4, where
we ask if the goal location ℓ3 is reachable from the initial state (ℓ0 , v0 ) where v0 (x) = v0 (y) = 0. The
values stored in the passed-waiting list after each iteration of the while-loop are shown in the table such
that a column labelled with a number i is the status of the passed-waiting list after the i’th execution of
the body of the while-loop; all values for anchor points not listed in the table are constantly ⊥.
Initially we set PW (ℓ0 , (0, 0)) = (0, ¥), meaning that all points reachable from the initial valuation
after an arbitrary delay action belong to the waiting list and should be explored. As ℓ0 is not the goal
state, the algorithm continues with the execution of the main while-loop. In the first iteration of the
loop we pick the only element in the waiting list so that ℓ = ℓ0 , g = (0, 0), w = 0 and p = ¥. Then we
update PW (ℓ0 , (0, 0)) to (0, 0)2 according to line 5 of the algorithm, meaning that all points on the dart
are now in the passed list. After this we consider the transition from ℓ0 to ℓ1 with the guard x ∈ [2, ¥)
(and the implicit guard y ∈ [0, ¥)) and calculate the values of start (minimum delay from the anchor
point to satisfy the guard and at the same time having at least the delay w where the waiting list starts)
and end (maximum delay from the anchor point so that the guard is still satisfied). In our example we
have start = max(0, (2 − 0), (0 − 0)) = 2 and end = min((¥ − 2), (¥ − 0)) = ¥.
Next we consider the test at line 9 that requires that the minimum delay start to enable all guards is
not in the region of already passed points (start < p) and at the same time that it is below the maximum
delay after which the guard become disabled (start ≤ end).
If this test fails, there is no need to do
2 In each column we mark by bold font the element that is picked in the next iteration of the while-loop.
150
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
anything with the currently picked element from the waiting list. As the values in our example satisfy the
condition at line 9 and no clocks are reset, we update according to line 11 of the algorithm the value of
(ℓ1 , (0, 0)) to (2, ¥). This means that in the future iterations we have to explore in location ℓ1 all points
(2, 2), (3, 3), (4, 4), . . . . Note that the addition and subtraction of start at line 11 had no effect as none of
the clocks after the minimum delay exceeded the maximum constant 2; should this happen the values
exceeding the maximum constant get truncated to MC + 1.
In the second iteration of the while-loop we select the location and anchor point (ℓ1 , (0, 0)), with
w = 2 and p = ¥, set it to (2, 2) in the table and mark it in bold as the selected point in the previous
column. This time we have to explore two edges. First, we select the self-loop that resets the clock x
and we get start = 2 and end = ¥. Now we execute the lines 10 to 15 as the edge contains a reset. The
for-loop will be run for the value of n from 2 to 3. The upper-bound of 3 for the for-loop follows from
the fact that MC = 2 and the maximum value of a clock that is not reset in the anchor point is 0. In
the for-loop we add two successors (line 15) at the location ℓ1 with the anchor points (0, 2) and (0, 3).
Second, if we consider the edge from ℓ1 to ℓ2 we can see that in location ℓ2 the anchor point (0, 0) is set
to (0, ¥).
The remaining values stored in the passed-waiting list are computed in the outlined way. We can
notice that after the 7th iteration of the while-loop the set Waiting(PW ) is empty and the algorithm
terminates. As the location ℓ3 has not been discovered during the search, the algorithm returns false.
The correctness theorem requires a detailed technical treatment and its complete proof is given in the
full version of this paper. Termination follows from the fact that newly added anchor points are computed
as (g⊕MC start) − start or (g⊕MC n)[R := 0] which ensures a finite size of the passed-waiting list and tha t
every time-dart (g, w, p) on the list satisfies 0 ≤ w ≤ MC, w < p, and p ≤ MC or p = ¥. Soundness proof
is by a case analysis establishing a loop-invariant that every call to AddToPW only adds time-darts that
represent reachable configurations in the bounded semantic s. Finally, the completeness proof is done
by induction on the length of the computation leading to a reachable configuration, taking into account
the nondeterministic nature of the algorithm, the fact that ≡MC is a timed bisimulation, and it makes a
full analysis of the different cases for adding new time-darts present in the algorithm for its performance
optimization.
Theorem 5.1 Let A be a timed automaton and let ℓg be a location. Algorithm 2 terminates, and it returns
true iff ℓg is reachable in the discrete semantics TDS (A).
6 Experiments
We have conducted a number of experiments in order to test the performance of the time-dart state-space
representation. The experiments were done within the project opaal [12], a model-checking framework
designed explicitly for fast prototyping and testing of verification algorithms using the programming lan-
guage Python. The tool implements the pseudocode of both the fully discrete (called naive in the tables)
as well as the time-dart reachability algorithms based on passed-waiting list presented in Section 4.
The experiments were conducted on Intel Core 2 Duo [email protected] running Ubuntu linux.
The verification was interrupted after five minutes or when th
e memory limit of 2GB RAM was ex-
ceeded (marked in the tables as OOM). The number of discovered symbolic states corresponds to
the total number of calls to the function AddToPW (including duplicates) and the number of stored
states is the size of the passed-waiting list at the termination of the algorithm. Verification times
(in seconds) are highlighted in the bold font. The examples and tool implementation are available at
http://people.cs.aau.dk/~kyrke/download/timedart/timedart.tar.gz.
K.Y. Jørgensen, K.G.Larsen& J.Srba
151
Model
T55
T55
T55
T55
T55
T55
T55
T125
T125
T125
T125
T125
T125
T125
T125
T155
T155
T155
T155
T155
T155
T155
T155
T155
T155
T155
T155
T155
# Naive Discovered
3.8
81,062
4
13.0
254,969
5
43.8
6
727,712
95.7
1,431,665
7
8 OOM
-
9
-
10
0.3
4
1.6
5
5.5
6
20.3
7
93.4
8
9 OOM
-
10
-
11
0.4
4
1.1
5
19.9
6
28.5
7
32.3
8
34.1
9
60.8
10
11 OOM
-
12
-
13
-
14
-
15
-
16
-
-
2,609
18,394
61,242
205,808
82,4630
-
-
-
5,796
23,454
433,674
577,179
620,138
626,100
1,035,226
-
-
-
-
-
Stored Darts Discovered
1.9
34,012
38,906
5.9
111,543
110,907
17.6
336,527
297,026
32.0
524,270
607,483
91.8
1,740,066
255.7
4,700,607
-
- >300
-
0.2
2,050
198
0.2
713
14,772
0.5
48,600
2,916
1.4
10,337
161,394
4.7
529,032
39,242
13.7
111,438
-
34.5
274,939
-
- OOM
-
0.3
3,048
1,532
0.4
4,572
9,740
3.6
62,771
14,2861
4.5
74,105
18,7857
4.8
203,178
78,093
4.9
79,111
205,646
7.5
241,44
329,193
31.2
514,959
-
37.7
-
608,974
105.7
1,684,525
-
158.3
2,316,474
-
164.5
2,409,417
-
- OOM
-
Stored
3,654
10,739
29,378
51,730
136,639
347,136
-
139
503
1,769
6,102
20,392
56,191
126,895
-
467
1,195
8,859
10,725
11,508
11,753
18,574
67,592
80,634
205,087
284,859
298,288
-
Figure 5: Results for three different TGS scaled by the number of tasks
6.1 Task Graph Scheduling
The task graph scheduling problem (TGS) is the problem of find ing a feasible schedule for a number of
parallel tasks with given precedence constraints and processing times on a fixed number of homogeneous
processors [17]. The chosen task graphs for two processors were taken from the benchmark [22] such that
several scheduling problems with different degree of concurrency are included. The models are scaled
by the number of tasks in the order given by the benchmark and the verification query performed a full
state-space search. The experimental results are displayed in Figure 5. The data confirm that the time-
dart verification technique saves both the number of stored/ discovered states and noticeably improves
the verification speed, in particular in the model T155.
6.2 Bridge Crossing Vikings
The bridge crossing Vikings is a slightly modified version of
the standard planning problem available in
the official distribution of U PPAA L; we only eliminated the used integer variables that are not supported
in our opaal implementation and are simulated by new locations. The query searched the whole state-
152
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
# Naive Discovered
0.2
295
2
0.2
2,614
3
0.8
16,114
4
4.9
5
90,743
33.2
501,958
6
7 OOM
-
-
8
-
Stored Darts Discovered
0.1
87
152
0.2
754
1,263
0.5
4,902
7,588
2.5
42,294
29,144
13.4
165,535
235,635
74.8
900,439
-
- >300
-
Stored
46
336
1,759
8,308
38,367
177,807
-
Figure 6: Results for bridge crossing scaled by the number of Vikings
# Naive Discovered
0.2
1
173
1.5
16684
2
3 OOM
-
-
4
-
Stored Darts Discovered
0.2
139
21
0.3
1140
11042
4.6
40671
-
- OOM
-
Stored
15
647
21721
-
Figure 7: Results for train level crossing scaled by the number of trains
space. Verification results are given in Figure 6. The perfor mance of the time-dart algorithm is again
better than the full discretization, even though in this case the constants in the model are relatively small
(proportional to the number of Vikings), meaning that the potential of time-darts is not fully exploited.
6.3 Train Level Crossing
In train level crossing we consider auto-generated timed automata templates constructed via automatic
translation [11] from timed-arc Petri net model of a train level-crossing example. The auto-generated
timed automata were produced by the tool TAPAAL [13] and have a rather complex structure that human
modelers normally never design and hence we can test the potential of the discrete-time engine also for
the models translated from other time-dependent formalisms. The query we asked searches the whole
state-space. We list the results in Figures 7 and the experiment demonstrates again the advantage of the
time-dart verification method.
6.4 Fischer’s Protocol
The discrete-time techniques are sensitive to the size of the constants present in the model. We have
therefore scaled our next experiment by the size of the maximal constant (MC) that appears in the model
in order to demonstrate the main advantage of the time-dart algorithm. For this we use the well known
Fischer’s protocol for ensuring a mutual exclusion between two or more parallel processes [18]. It is
a standard model for testing the performance of verification tools; we replaced one open interval in
the model with a closed one such that mutual exclusion is still guaranteed. The concrete version of
the protocol we verified was created by a translation from tim ed-arc Petri net model of the protocol [1]
available as a demo example in the tool TAPAAL [13]. We searched the whole state-space and the results
are summarized in Figure 8. It is clear that time-darts are superior w.r.t. the scaling of the constants in
the model, allowing us to verify (within the given limit of 300 seconds) models where the maximum
constant is 66, opposed to only 18 when the full discretization is used.
K.Y. Jørgensen, K.G.Larsen& J.Srba
153
MC Naive Discovered
3.1
36,774
3
4.5
53,655
4
6.2
74,415
5
17.8
9
202,965
64.5
569,280
15
111.8
850,164
18
OOM
-
19
-
25
-
-
-
38
-
-
50
-
66
-
Stored Darts Discovered
1.1
9,464
25,882
1.4
14,341
38,570
1.8
20,226
54,513
3.8
157,555
53,846
10.4
133,796
466,888
14.2
187,595
710,857
15.7
207,544
-
26.1
-
348,406
61.3
778,095
-
114.7
1,325,931
-
217.2
-
2,214,846
Stored
6,238
8,725
11,548
26,200
58,258
78,823
86,350
138,568
293,203
486,343
795,808
Figure 8: Experimental results for Fischer’s mutual exclusion protocol
7 Conclusion
We have introduced a new data structure of time-darts in order to represent the reachable state-space of
closed timed automata models. We showed on a number of experiments that our time-dart reachability
algorithm achieves a consistently better performance than the explicit search algorithm, improving both
the speed and memory requirements. This is obvious in particular on models with larger constants
(as demonstrated in the Fischer’s experiment or T155 task graph) where time-darts provide a compact
representation of the delay successors and considerably improve both time and memory.
The algorithms were implemented in the interpreted language Python without any further optimiza-
tions techniques like partial order and symmetry reductions and advanced extrapolation techniques and
with only one global maximum constant. This does not allow us to compare its performance directly
with the state-of-the-art optimized tools for real-time systems.
An advantage of time-darts and explicit state-space methods in general is that it is relatively easy to
extend them with additional modelling features like clock invariants and diagonal guards. In our future
work we will implement the time-dart algorithm in C++ with additional optimizations (e.g. considering
local constants instead of the global ones) and we shall also consider the verification of liveness proper-
ties. It is clear that for large enough constants the DBM-search engine will always combat the explicit
methods (see [19]); our technique can be so seen as a practical alternative to the DBM-engines on the
subset of models that for example use counting features (like in our introductory example) and where
DBM state-space representation explodes even for models with small constants. Another line of research
will focus on further optimizations of the time-dart technique by considering federations of time-darts so
that the data structure becomes even less sensitive to the scaling of the constants.
Acknowledgements. We would like to thank the anonymous reviewers for their comments.
References
[1] P.A. Abdulla & A. Nyl ´en (2001): Timed Petri Nets and BQOs. In: Proceedings of the 22nd International
ConferenceonApplicationandTheoryofPetriNets(ICATPN’01), LNCS 2075, Springer-Verlag, pp. 53 –70,
doi:10.1007/3-540-45740-2 5.
154
Time-Darts: ADataStructure forVerification ofClosedTime dAutomata
[2] R. Alur & D. Dill (1994): A Theory of Timed Automata. Theoretical Computer Science (TCS) 126(2), pp.
183 –235, doi:10.1016/0304-3975(94)90010-8.
[3] E. Asarin, M. Bozga, A. Kerbrat, O. Maler, A. Pnueli & A. Rasse (1997): Data-Structures for the
In Oded Maler, editor: HART, LNCS 1201, Springer, pp. 346 –360,
Verification of Timed Automata .
doi:10.1007/BFb0014737.
[4] G. Behrmann, A. David & K.G. Larsen (2004): A Tutorial on U P PAAL .
In: Formal Methods for the
Design of Real-Time Systems: 4th International School on Formal Methods for the Design of Com-
puter, Communication, and Software Systems (SFM-RT’04), LNCS 3185, Springer-Verlag, pp. 200 –236,
doi:10.1007/978-3-540-30080-9 7.
[5] B. Berthomieu & M. Menasche (1983): An Enumerative Approach for Analyzing Time Petri Nets. In: IFIP
Congress, pp. 41 –46.
[6] D. Beyer, C. Lewerentz & A. Noack (2003): Rabbit: A Tool for BDD-Based Verification of Real-Time
Systems.
In Warren A. Hunt Jr. & Fabio Somenzi, editors: CAV, LNCS 2725, Springer, pp. 122 –125,
doi:10.1007/978-3-540-45069-6 13.
[7] D. Beyer & A. Noack (2003): Can Decision Diagrams Overcome State Space Explosion in Real-Time Ver-
In Hartmut K onig, Monika Heiner & Adam Wolisz, editors: FORTE, LNCS 2767, Springer, pp.
ification?
193 –208, doi:10.1007/978-3-540-39979-7 13.
[8] P. Bouyer (2003): Untameable Timed Automata!
In Helmut Alt & Michel Habib, editors: STACS, LNCS
2607, Springer, pp. 620 –631, doi:10.1007/3-540-36494-3 54.
[9] M. Bozga, S. Graf & L. Mounier (2002): IF-2.0: A Validation Environment for Component-Based Real-Time
Systems. In Ed Brinksma & Kim Guldstrand Larsen, editors: CAV, LNCS 2404, Springer, pp. 343 –348,
doi:10.1007/3-540-45657-0 26.
[10] M. Bozga, O. Maler & S. Tripakis (1999): Efficient Verification of Timed Automata Using Dense and Disc
rete
Time Semantics. In: Proceedings of Correct Hardware Design and Verification Met hods (CHARME’99),
LNCS 1703, Springer, pp. 125 –141, doi:10.1007/3-540-48153-2 11.
[11] J. Byg, K.Y. Jørgensen & J. Srba (2009): An Efficient Translation of Timed-Arc Petri Nets to Networks of
Timed Automata. In: Proc. of the 11th International Conf. on Formal Engineering Methods (ICFEM’09),
LNCS 5885, Springer-Verlag, pp. 698 –716, doi:10.1007/978-3-6 42-10373-5 36.
[12] A.E. Dalsgaard, R.R. Hansen, K.Y. Jørgensen, K.G. Lars en, M.Chr. Olesen, P. Olsen & J. Srba (2011): opaal:
A Lattice Model Checker. In: Proceedingsof the3rdNASAFormalMethodsSymposium(NFM’11), LNCS
6617, Springer-Verlag, pp. 487 –493, doi:10.1007/978-3-6 42-20398-5 37.
[13] A. David, L. Jacobsen, M. Jacobsen, K.Y. Jørgensen, M.H . Møller & J. Srba (2012): TAPAAL 2.0: Integrated
Development Environment for Timed-Arc Petri Nets. In: Proceedings of the 18th International Conference
onToolsandAlgorithmsfor theConstructionandAnalysisofSystems (TACAS’12), LNCS 7214, Springer-
Verlag, pp. 492 –497, doi:10.1007/978-3-642-28756-5 36.
[14] C. Daws, A. Olivero, S. Tripakis & S. Yovine (1996): The Tool KRONO S. In: Proc. Hybrid Systems III:
VerificationandControl(1995) , LNCS 1066, Springer-Verlag, pp. 208 –219, doi:10.1007/BFb0020 947.
[15] D.L. Dill (1989): Timing Assumptions and Verification of Finite-State Concur rent Systems.
In Joseph
Sifakis, editor: AutomaticVerificationMethodsforFiniteStateSystems , LNCS 407, Springer, pp. 197 –212,
doi:10.1007/3-540-52148-8 17.
[16] P.-A. Hsiung, F. Wang & R.-Ch. Chen (2000): On the verification of Wireless Transaction Protocol using
SGM and RED. In: RTCSA, IEEE Computer Society, pp. 379 –383, doi:10.1109/RTCSA.2 000.896414.
[17] Y.-K. Kwok & I. Ahmad (1999): Benchmarking and Comparison of the Task Graph Scheduling Algorithms.
JournalofParallelandDistributedComputing 59(3), pp. 381 – 422, doi:10.1006/jpdc.1999.1578.
[18] L. Lamport (1987): A Fast Mutual Exclusion Algorithm. ACMTransactionsonComputerSystems 5(1), pp.
1 –11, doi:10.1145/7351.7352.
K.Y. Jørgensen, K.G.Larsen& J.Srba
155
[19] L. Lamport (2005): Real-Time Model Checking Is Really Simple. In: CHARME, LNCS 3725, Springer, pp.
162 –175, doi:10.1007/11560548 14.
[20] K.G. Larsen & Y. Wang (1997): Time-Abstracted Bisimulation: Implicit Specifications an d Decidability.
InformationandComputation 134(2), pp. 75 – 101, doi:10.1006/inco.1997.2623.
[21] Y. Liu, J. Sun & J.S. Dong (2008): An Analyzer for Extended Compositional Process Algebras. In: ICSE
Companion, ACM, pp. 919 –920, doi:10.1145/1370175.1370187.
[22] Kasahara
Laboratory
at
Waseda
University:
Http://www.kasahara.elec.waseda.ac.jp/schedule/.
[23] J. Zhao, X. Li & G. Zheng (2005): A quadratic-time DBM-based successor algorithm for checking timed
automata. InformationProcessingLetters 96(3), pp. 101 –105, doi:10.1016/j.ipl.2005.05.027.
Standard
Set.
Task
Graph
|
1806.02207 | 1 | 1806 | 2018-06-06T14:19:52 | Online Makespan Minimization: The Power of Restart | [
"cs.DS"
] | We consider the online makespan minimization problem on identical machines. Chen and Vestjens (ORL 1997) show that the largest processing time first (LPT) algorithm is 1.5-competitive. For the special case of two machines, Noga and Seiden (TCS 2001) introduce the SLEEPY algorithm that achieves a competitive ratio of $(5 - \sqrt{5})/2 \approx 1.382$, matching the lower bound by Chen and Vestjens (ORL 1997). Furthermore, Noga and Seiden note that in many applications one can kill a job and restart it later, and they leave an open problem whether algorithms with restart can obtain better competitive ratios.
We resolve this long-standing open problem on the positive end. Our algorithm has a natural rule for killing a processing job: a newly-arrived job replaces the smallest processing job if 1) the new job is larger than other pending jobs, 2) the new job is much larger than the processing one, and 3) the processed portion is small relative to the size of the new job. With appropriate choice of parameters, we show that our algorithm improves the 1.5 competitive ratio for the general case, and the 1.382 competitive ratio for the two-machine case. | cs.DS | cs | Online Makespan Minimization: The Power of Restart∗
Zhiyi Huang†
Ning Kang‡
Zhihao Gavin Tang§
Xiaowei Wu¶
Yuhao Zhang(cid:107)
Abstract
We consider the online makespan minimization problem on identical machines. Chen and
gorithm that achieves a competitive ratio of (5 − √
Vestjens (ORL 1997) show that the largest processing time first (LPT) algorithm is 1.5-competitive.
For the special case of two machines, Noga and Seiden (TCS 2001) introduce the SLEEPY al-
5)/2 ≈ 1.382, matching the lower bound by
Chen and Vestjens (ORL 1997). Furthermore, Noga and Seiden note that in many applications
one can kill a job and restart it later, and they leave an open problem whether algorithms with
restart can obtain better competitive ratios.
We resolve this long-standing open problem on the positive end. Our algorithm has a natural
rule for killing a processing job: a newly-arrived job replaces the smallest processing job if 1)
the new job is larger than other pending jobs, 2) the new job is much larger than the processing
one, and 3) the processed portion is small relative to the size of the new job. With appropriate
choice of parameters, we show that our algorithm improves the 1.5 competitive ratio for the
general case, and the 1.382 competitive ratio for the two-machine case.
8
1
0
2
n
u
J
6
]
S
D
.
s
c
[
1
v
7
0
2
2
0
.
6
0
8
1
:
v
i
X
r
a
Hong Kong RGC under the grant HKU17202115E.
∗A preliminary version of the paper to appear in APPROX 2018.
†Department of Computer Science, The University of Hong Kong. [email protected]. Partially supported by the
‡Department of Computer Science, The University of Hong Kong. [email protected].
§Department of Computer Science, The University of Hong Kong. [email protected].
¶Department of Computing, The Hong Kong Polytechnic University. [email protected]. Part of the work was
done when the author was a postdoc at the University of Hong Kong.
(cid:107)Department of Computer Science, The University of Hong Kong. [email protected].
1
1
Introduction
We study in this paper the classic online scheduling problem on identical machines. Let there be m
identical machines, and a set of jobs that arrive over time. For each job j, let rj denote its release
time (arrival time), and pj denote its processing time (size). We assume without loss of generality
that all rj's and pj's are distinct. We seek to schedule each job on one of the m machines such that
the makespan (the completion time of the job that completes last) is minimized.
We adopt the standard assumption that there is a pending pool such that jobs released but not
scheduled are in the pending pool. That is, the algorithm does not need to assign a job to one of
the machines at its arrival; it can decide later when a machine becomes idle. Alternatively, the
immediate-dispatching model has also been considered in some papers (e.g., Avrahami and Azar
[2007]).
We consider the standard competitive analysis of online algorithms. An algorithm is (1 + γ)-
competitive if for any online sequence of jobs, the makespan of the schedule made by the algorithm
is at most (1+γ) times the minimum makespan in hindsight. Without loss of generality, (by scaling
the job sizes) we assume the minimum makespan OPT = 1 (for analysis purpose only).
Chen and Vestjens [1997] consider a greedy algorithm called largest processing time first (LPT):
whenever there is an idle machine, schedule the largest job in the pending pool. They prove that
the LPT algorithm is 1.5-competitive and provide a matching lower bound (consider m jobs of size
0.5 followed by a job of size 1). They also show that no online algorithm can achieve a competitive
ratio better than 1.3473. For the special case when there are only two machines, Noga and Seiden
5)/2 ≈ 1.382 competitive ratio,
[2001] introduce the SLEEPY algorithm that achieves a tight (5−√
due to a previous lower bound given by Chen and Vestjens [1997].
The 1.382 lower bound (for two machines) holds under the assumption that whenever a job
is scheduled, it must be processed all the way until its completion. However, as noted in Noga
and Seiden [2001], many applications allow restart: a job being processed can be killed (put into
pending) and restarted later to make place for a newly-arrived job; a job is considered completed
only if it has been continuously processed on some machine for a period of time that equals to its
size. In other words, whenever a job gets killed, all previous processing of this job is wasted.
Note that the restart setting is different from the preemptive setting, in which the processed
portion is not wasted. Noga and Seiden [2001] leave the following open problem: Is it possible to
beat the 1.382 barrier with restart?
In this paper, we bring an affirmative answer to this long-standing open problem.
We propose a variant of the LPT algorithm (with restart) that improves the 1.5 competitive
ratio for the general case, and the 1.382 competitive ratio for the two-machine case.
Our Replacement Rule. A naıve attempt for the replacement rule would be to replace a job
whenever the newly-arrived job has a larger size. However, it is easy to observe that the naıve
attempt fails even on one machine: the worst case competitive ratio is 2 if we keep replacing
jobs that are almost completed (with jobs of slightly larger size). Hence we should prevent a job
from being replaced if a large portion has been processed. Moreover, we allow a newly-arrived
job to replace a processing job only if it has a much larger size, in order to avoid a long chain of
replacements. As we will show by an example in Section 7, the worst case competitive ratio is 1.5
if a job of size 1 is replaced by a job of size 1 + , which is in turn replaced by a job of size 1 + 2,
etc.
We hence propose the following algorithm that applies the above rules.
1
LPT with Restart. As in the LPT algorithm, our algorithm schedules the largest pending job
whenever there is an idle machine. The main difference is that our algorithm may kill a processing
job to make place for a newly-arrived job according to the following rule. Upon the arrival of a job
j, we kill a processing job k (i.e., put k into pending) and schedule j if:
1. j is the largest pending job and k is the smallest among the m processing jobs;
2. the processed portion of k is less than αpj;
3. the size of j is more than 1 + β times larger than k, i.e., pj > (1 + β)pk,
2 are parameters of the algorithm. We call such an operation a replacement (i.e.
where 0 < α, β < 1
j replaces k).
Intuitively, the parameter α provides a bound on the total amount of wasted processing (in
terms of the total processing time); while the parameter β guarantees an exponential growth in the
processing time of jobs if there is a chain of replacements. With appropriate choice of parameters,
we show the following results.
Theorem 1.1 LPT with Restart, with parameters α = 1
200 and β =
competitive for the Online Makespan Minimization problem with restart.
√
2 − 1, is (1.5 − 1
20000 )-
Theorem 1.2 LPT with Restart, with parameters α = β = 0.2, is 1.38-competitive for the Online
Makespan Minimization problem with restart on two machines.
There are many other natural candidate replacement rules. We list some candidate algorithms
that we have considered and their counter examples in Sec 7.
Our Techniques. The main focus of our paper is the general case, i.e., on m machines. The
analysis for the two-machine case is built on the general case by refining some of the arguments.
We adopt an idea from Chen and Vestjens [1997] to look at the last completed job in our
schedule. Intuitively, only jobs with size comparable to that of the last job matter. We develop
two kinds of arguments, namely the bin-packing argument and the efficiency argument.
2 − 1
Assume for contrary that the algorithm has a makespan strictly larger than 1 + γ, where
γ := 1
20000 , we use the bin-packing argument to give an upper bound on the size of the last
completed job. Assume that the last job is large, we will find a number of large jobs that cannot
be packed into m bins of size 1 (recall that we assume OPT = 1). In other words, to schedule this
set of jobs, one of the m machines must get a total workload strictly greater than 1. For example,
finding 2m + 1 jobs of size strictly greater than 1
3 would suffice. We refer to such a set of large jobs
as an infeasible set of jobs.
We then develop an efficiency argument to handle the case when the last job is of small size. The
central of the argument is a Leftover Lemma that upper bounds the difference of total processing
done by the algorithm and by OPT. As our main technical contribution, the lemma is general
enough to be applied to all schedules.
Fix any schedule (produced by some algorithm ALG) and a time t. Let M denote the set
of machines. For each machine M ∈ M, let W(M, x) ∈ {0, 1} be the indicator function of the
event that "at time x, machine M is not processing while there are pending jobs". Define Wt =
0 W(M, x)dx to be the total waste (of processing power) before time t. We show (in
(cid:80)
Section 3) the following lemma that upper bounds the leftover workload.
(cid:82) t
M∈M
Lemma 1.1 (Leftover Lemma) For all time t, let ∆t be the difference in total processing time
before time t between ALG and OPT. We have ∆t ≤ 1
4 tm + Wt.
2
Observe that the total processing power (of m machines) before time t is tm. The Leftover
Lemma says that compared to any schedule (produced by algorithm ALG), the extra processing
the optimal schedule can finish before time t, is upper bounded by the processing power wasted by
the schedule (e.g., due to replacements), plus a quarter of the total processing power, which comes
from the sub-optimal schedule of jobs.
Consider applying the lemma to the final schedule1 produced by our algorithm. Since our
algorithm schedules a job whenever a machine becomes idle, the waste Wt comes only from the
processing (before time t) of jobs that are replaced. Thus (by our replacement) we can upper bound
Wt by α fraction of the total size of jobs that replace other jobs.
We remark that the above bound on the leftover workload is tight for LPT (for which Wt = 0).
Consider m jobs of size 0.5 arriving at time 0, followed by m/2 jobs of size 1 − arriving at time .
The optimal schedule uses m
2 machines to process
the size 0.5 jobs (two per machine), finishing all jobs at time 1. LPT would schedule all the size 0.5
2 size (1− ) jobs have half of their workload unprocessed at time 1. Therefore,
jobs first; all of the m
the amount of leftover workload at time t = 1 is m
4 .
2 machines to process the size (1 − ) jobs and m
Other Work. The online scheduling model with restart has been investigated in the problem
of scheduling jobs on a single machine to maximize the number of jobs completed before their
deadlines. Hoogeveen et al. [2000] study the general case and propose a 2-competitive algorithm
with restart. Subsequently, Chrobak et al. [2007] consider the special case when jobs have equal
lengths. They propose an improved 3
2 -competitive algorithm with restart for this special case, and
prove that this is optimal for deterministic algorithms. However, the restart rule and its analysis
in our paper do not bear any obvious connections to those in Hoogeveen et al. [2000] and Chrobak
et al. [2007] due to the different objectives.
Other settings of the online makespan minimization problem have been studied in the literature.
A classic setting is when all machines are identical and all jobs have release time 0, but the algorithm
must immediately assign each job to one of the machines at its arrival (immediate dispatching).
This is the same as online load balancing problem. Graham [1969] proves that the natural greedy
algorithm that assigns jobs to the machine with the smallest workload is (2 − 1
m )-competitive in
this setting, which is optimal for m ≤ 3 (due to folklore examples). A series of research efforts
have then been devoted to improving the competitive ratio when m is large (e.g., Albers [1999],
Bartal et al. [1995], Karger et al. [1996]). For m = 4, the best upper bound is 1.7333 Chen et al.
[1994a], while the best lower bound stands at 1.7321 RudinIII and Chandrasekaran [2003]. For m
that tends to infinity, the best upper bound is 1.9201 Fleischer and Wahl [2000], while the best
lower bound is 1.880 RudinIII [2001].
A variant of the above setting is that a buffer is provided for temporarily storing a number of
jobs; when the buffer is full, one of the jobs must be removed from the buffer and allocated to a
machine (e.g., Li et al. [2007], D´osa and Epstein [2010]). Kellerer et al. [1997] and Zhang [1997] use
algorithms with a buffer of size one to achieve an improved 4/3 competitive ratio for two machines.
Englert et al. [2014] characterize the best ratio achievable with a buffer of size Θ(m), where the
ratio is between 4/3 and 1.4659 depending on the number of machines m. When both preemption
and migration are allowed, Chen et al. [1995] give a 1.58-competitive algorithm without buffer,
matching the previous lower bound by Chen et al. [1994b]. D´osa and Epstein [2011] achieve a ratio
of 4/3 with a buffer of size Θ(m).
Finally, if the machines are related instead of identical, the best known algorithm is 4.311-
competitive by Berman et al. [2000], while the best lower bound is 2 by Epstein and Sgall [2000].
1Since a job can be scheduled and replaced multiple times, its start time is finalized only when it is completed.
3
When preemption is allowed, Ebenlendr et al. [2009] show that the upper bound can be improved
to e. For the special case of two related machines, the current best competitive ratio is 1.53 by
Epstein et al. [1999] without preemption, and 4/3 with preemption by Ebenlendr et al. [2009] and
Wen and Du [1998].
Organization. We first provide some necessary definitions in Section 2. Then we prove the most
crucial structural property (Lemma 1.1, the Leftover Lemma) in Section 3, which essentially gives
a lower bound on the efficiency of all schedules. We present the details of the bin-packing argument
and efficiency argument in Section 4, where our main result Theorem 1.1 is proved. The special
case of two machines is considered in Section 6, where Theorem 1.2 is proved. Finally, we prove
in Section 8 that no deterministic algorithm, even with restart, can get a competitive ratio better
than
√
1.5 ≈ 1.225.
2 Preliminaries
Consider the online makespan minimization with m identical machines and jobs arriving over time.
Recall that for each job j, rj denotes its release time and pj denotes its size. Let OPT and ALG be
the makespan of the optimal schedule and our schedule, respectively. Recall that we assume without
loss of generality that OPT = 1 (for analysis purpose only). Hence we have rj +pj ≤ 1 for all jobs j.
Further, let sj and cj := sj + pj denote the start and completion time of job j, respectively, in the
final schedule produced by our online algorithm. Note that a job can be scheduled and replaced
multiple times. We use sj(t) to denote the last start time of j before time t.
We use n to denote the job that completes last, i.e., we have ALG = cn = sn + pn.
We consider the time horizon as continuous, and starts from t = 0. Without loss of generality
(by perturbing the variables slightly), we assume that all ri's, pi's and si's are different.
Definition 2.1 (Processing Jobs) For any t ≤ ALG, we denote by J(t) the set of jobs that are
being processed at time t, including the jobs that are completed or replaced at t but excluding the
jobs that start at t.
Note that J(t) is defined based on the schedule produced by the algorithm at time t.
It is
possible that jobs in J(t) are replaced at or after time t.
Idle and Waste. We say that a machine is idle in time period (a, b), if for all t ∈ (a, b), the
machine is not processing any job according to our algorithm, and there is no pending job. We
call time t idle if there exists at least one idle machine at time t. Whenever a job k is replaced by
a job j (at rj), we say that a waste is created at time rj. The size of the waste is the portion of k
that is (partially) processed before it is replaced. We can also interpret the waste as a time period
on the machine. We say that the waste comes from k, and call j the replacer.
Definition 2.2 (Total Idle and Total Waste) For any t ∈ [0, 1], define It as the total idle time
before time t, i.e., the summation of total idle time before time t on each machine. Similarly, define
Wt as the total waste before time t in the final schedule, i.e., the total size of wastes located before
time t, where if a waste crosses t, then we only count its fractional size in [0, t].
4
3 Bounding Leftover: Idle and Waste
(recall from Section 1) Wt is defined as(cid:80)
In this section, we prove Lemma 1.1, the most crucial structural property. Recall that we define Wt
as the total waste located before time t. For applying the lemma to general scheduling algorithms,
0 W(M, x)dx, the total time during which machines
are not processing while there are pending jobs. It is easy to check that the proofs hold under both
definitions. We first give a formal definition of the leftover ∆t at time t.
M∈M
(cid:82) t
Definition 3.1 (Leftover) Consider the final schedule and a fixed optimal schedule OPT. For
any t ∈ [0, 1], let ∆t be the total processing OPT does before time t, minus the total processing our
algorithm does before time t.
Since the optimal schedule can process a total processing at most m(1− t) after time t, we have
the following useful observation.
Observation 3.1 The total processing our algorithm does after time t is at most m(1 − t) + ∆t.
We call time t a marginal idle time if t is idle and the time immediately after t is not. We first
define At, which is designated to be an upper bound on the total processing that could have been
done before time t, i.e., the leftover workload due to sub-optimal schedule.
Definition 3.2 (At) For all t ∈ [0, 1], if there is no idle time before t, then define At = 0, otherwise
j∈J(t(cid:48)) min{δj, pj}, where δj := Tj = {θ ∈
[rj, t(cid:48)] : job j is pending at time θ} is the total pending time of job j ∈ J(t(cid:48)) before time t(cid:48).
let t(cid:48) ≤ t be the last idle time before t. Define At =(cid:80)
(while dAt
dt = dIt
dt
We show the following claim, which (roughly) says that the extra processing OPT does (com-
pared to ALG) before time t, is not only upper bounded by total idle and waste (It + Wt), but also
by the total size or pending time of jobs currently being processed (At + Wt).
Claim 3.1 We have ∆t ≤ min{At, It} + Wt for all t ∈ [0, 1].
Proof: First observe that we only need to prove the claim for marginal idle times, as we have
d∆t
dt ≤ dWt
which the algorithm is not processing (in the final schedule). Next we show that ∆t ≤ At + Wt.
Let ∆t(t), At(t) and Wt(t) be the corresponding variables when the algorithm is run until time
t. Observe that for a job j ∈ J(t), if it is replaced after time t, then it contributes a waste to Wt
but not to Wt(t). Moreover, it has the same contribution to ∆t − ∆t(t) and to Wt. Thus we have
Wt − Wt(t) = ∆t − ∆t(t). By definition we have At = At(t).
It is easy to see that ∆t is at most It + Wt, the total length of time periods before t during
dt = 0) for non-idle time t. Now suppose t is a marginal idle time.
Hence it suffices to show that ∆t(t) ≤ At + Wt(t).
Since t is idle, there is no pending job at time t. Thus the difference in total processing at time
t, i.e., ∆t, must come from the difference (between ALG and OPT) in processing of jobs in J(t) that
has been completed. For each j ∈ J(t), the extra processing OPT can possibly do on j (compared
to ALG) is at most min{sj(t) − rj, pj}. Hence we have ∆t(t) ≤(cid:80)
Thus at every time t ∈ [rj, sj(t))\Tj, j is being processed (and replaced later). Hence(cid:12)(cid:12)[rj, sj(t))\
(cid:12)(cid:12) is at most the total wastes from j that are created before sj(t) < t, which implies
Recall by Definition 3.2, we have Tj ⊂ [rj, sj(t)) is the periods during which j is pending.
j∈J(t) min{sj(t) − rj, pj}.
Tj
min{δj, pj} + Wt(t) = At + Wt(t),
∆t(t) ≤ (cid:88)
min{sj(t) − rj, pj} ≤ (cid:88)
j∈J(t)
j∈J(t)
5
as desired.
We prove the following technical claim.
Claim 3.2 For any integer k ≥ 1, given any three sequences of positive reals {ai}i∈[k], {bi}i∈[k] and
{hi}i∈[k] satisfying conditions
(1) 0 ≤ h1 ≤ h2 ≤ . . . ≤ hk ≤ 1;
(2) for all j ∈ [k], we have(cid:80)
we have(cid:80)
(cid:80)
i∈[j] aihi ≥ 1
4
i∈[k](ai + bi).
i∈[k] bi(1 − hi) ≤ 1
4
(cid:80)
i∈[j](ai + bi),
Proof: We prove the claim by induction on k. We first show that the claim holds true when
k = 1. Note that we have a1h1 · b1(1 − h1) ≤ ( a1+b1
. Combine with
property (2) we know that b1(1 − h1) ≤ 1
)2 · ( h1+(1−h1)
)2 = (a1+b1)2
16
2
2
4 (a1 + b1).
Figure 1: graph representation of Claim 3.2 for k = 4
aihi.
Now suppose the claim is true for all values smaller than k. Using induction hypothesis on
{ai}i∈[k−1],{bi}i∈[k−1] and {hi}i∈[k−1], we have
bi(1 − hi) ≤ 1
4
(cid:88)
(cid:88)
(ai + bi) ≤ (cid:88)
i∈[k−1]
i∈[k−1]
k−1 = bk−1 + φ.
k = bk − φ and b(cid:48)
i∈[k−1]
i∈[k−1](4ai · hi − ai − bi)}. Let b(cid:48)
Define φ = min{bk,(cid:80)
Note that {ai}i∈[k],{bi}i∈[k−2]∪{b(cid:48)
(cid:88)
k−1 + b(cid:48)
k−1) ≤ 1
4
i∈[k−2]
Applying the induction hypothesis on {ai}i∈[k−1],{bi}i∈[k−2] ∪ {b(cid:48)
k} and {hi}i∈[k] (and their prefixes) satisfy the conditions
k−1, b(cid:48)
k−1 ≥ bk−1 > 0; second, since {ai}i∈[k] and
k > 0 and b(cid:48)
of the claim: first, by definition we have b(cid:48)
k = bk−1 + bk, if suffices to check condition (2) for j = k − 1:
{hi}i∈[k] are not changed, and b(cid:48)
(cid:88)
(cid:88)
(ai · hi).
(ak−1 + b(cid:48)
1
4
(ai + bi) + φ) ≤ (cid:88)
(cid:88)
i∈[k−1]
k−1} and {hi}i∈[k−1],
bi(1 − hi) + φ(1 − hk−1) ≤ 1
4
(ai · hi − ai + bi
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(ai + bi) +
(ai + bi) +
i∈[k−1]
i∈[k−1]
i∈[k−1]
i∈[k−1]
i∈[k−1]
) =
4
aihi.
1
4
(
If φ = bk, then immediately we have
bi(1 − hi) ≤ (cid:88)
(cid:88)
as desired. Otherwise we have φ =(cid:80)
i∈[k−1]
i∈[k]
(ai + bi),
i∈[k]
bi(1 − hi) + bk(1 − hk−1) ≤ 1
4
(
i∈[k−1]
(ai + bi) + bk) <
1
4
i∈[k−1](4ai · hi − ai − bi), and hence we have
(ak + b(cid:48)
(ak + bk + φ) =
ai · hi =
(ai + bi) − (cid:88)
(cid:88)
k),
1
4
1
4
i∈[k−1]
ak · hk ≥ 1
4
i∈[k]
6
h3b4b3b2b1a4a3a2a11h4h2h1 which implies b(cid:48)
(cid:88)
k(1 − hk) ≤ 1
bi(1 − hi) =
(cid:88)
i∈[k]
(ai + bi) +
(cid:88)
4 (ak + b(cid:48)
i∈[k−2]
≤ 1
4
i∈[k−2]
k). Hence we have
bi(1 − hi) + b(cid:48)
k−1(1 − hk−1) + b(cid:48)
k(1 − hk) + φ(hk−1 − hk)
ak−1 + b(cid:48)
k−1
4
ak + b(cid:48)
k
4
=
1
4
+
(ai + bi),
(cid:88)
i∈[k]
which completes the induction.
Given Claim 3.2, we are now ready to proof the Leftover Lemma.
dt
d( 1
(while
Proof of Lemma 1.1: As before, it suffices to prove the lemma for marginal idle times, as we
4 tm)
have d∆t
dt > 0) for non-idle time t. Now suppose t is a marginal idle time. As
before, let ∆t(t) and Wt(t) be the values of variables when the algorithm is run until time t.
dt ≤ dWt
We prove a stronger statement that ∆t(t) ≤ 1
4 tm + Wt(t), by induction on the number k of
marginal idle times at or before time t. Note that the stronger statement implies the lemma, as we
have ∆t − ∆t(t) = Wt − Wt(t).
In the following, we use a weaker version of Claim 3.1: we only need At ≤(cid:80)
processed from time g to t. By definition we have It ≤ (t− g)(m−J). Recall that At ≤(cid:80)
Base Case: k = 1. Since t is the first marginal idle time, let g be the first idle time, we know
that [g, t] is the only idle period. Define J := {j ∈ J(t) : sj(t) ≤ g} to be the set of jobs that are
j∈J(t) δj,
where δj is the total pending time of job j before time t. Hence we have δj ≤ g if j ∈ J, and δj = 0
otherwise. By Claim 3.1 we have
j∈J(t) δj.
∆t(t) ≤ min{At, It} + Wt(t) ≤ min{gJ, (t − g)(m − J)} + Wt(t) ≤ 1
4
Induction. Now suppose the statement holds for all marginal idle times 0 < t1 < t2 < . . . <
tk−1, and consider the next marginal idle time tk. We show that ∆tk (tk) ≤ 1
4 tkm + Wtk (tk). First
of all, observe that the difference in ∆tk (tk) and ∆tj (tj) must come from the idle periods in [tj, tk]
and wastes created in [tj, tk]. Hence for all j < k we have
tm + Wt(t).
∆tk (tk) − ∆tj (tj) ≤ (Itk − Itj ) + Wtk (tk) − Wtj (tj).
Hence, if there exists some j such that Itk − Itj ≤ 1
4 (tk − tj)m, then by induction hypothesis,
∆tk (tk) ≤ ∆tj (tj) +
1
4
(tk − tj)m + Wtk (tk) − Wtj (tj) ≤ 1
4
tkm + Wtk (tk),
Note that for all j ∈ Ji, we have δj ≤(cid:80)k
and we are done. Now suppose otherwise.
For all i ≤ k, let gi ∈ (ti−1, ti) be the first idle time after ti−1 (assume g0 = t0 = 0), i.e.,
[g1, t1], [g2, t2], . . . , [gk, tk] are the disjoint idle periods. Define Ji := {j ∈ J(tk) : rj ∈ [ti−1, gi)}.
x=i(gx − tx−1), as j is not pending during idle periods; for
j ∈ J(rk) \ ∪i≤kJi, we have δi = 0. For all i ∈ [k], define
ai := tk−i+1 − gk−i+1, bi := gk−i+1 − tk−i, hi := 1 − 1
We show that the three sequences of positive reals {ai}i∈[k], {bi}i∈[k],{hi}i∈[k] satisfy the con-
ditions of Claim 3.2, which implies ∆tk (tk) ≤ Atk + Wtk (tk) ≤ 1
(cid:88)
(cid:32) k(cid:88)
(cid:33)
δj ≤ (cid:88)
(gx − tx−1)
x∈[k−i+1] Jx.
m
(cid:80)
(gi − ti−1) ·(cid:88)
4 tkm + Wtk (tk) as
Jx
Ji =
(cid:88)
i∈[k]
i∈[k]
x∈[i]
(ai + bi) =
1
4
tkm.
Atk ≤ (cid:88)
(cid:88)
j∈J(rk)
= m
i∈[k]
i∈[k]
x=i
(bk−i+1 · (1 − hk−i+1)) ≤ 1
4
m
7
observe that Iti − Iti−1 ≤ (ti − gi) · (m −(cid:80)
Finally, we check the conditions of Claim 3.2. Condition (1) trivially holds. For condition (2),
x∈[i] Jx) = ak−i+1 · hk−i+1 · m. Hence we have
(cid:88)
i∈[j]
(cid:88)
i∈[j]
(cid:88)
i∈[j]
(ai · hi) ≥ 1
m
(Itk−i+1 − Ik−i) =
(Itk − Itk−j ) >
1
m
1
4
(tk − tk−j) =
1
4
(ai + bi),
as required.
4 Breaking 1.5 on Identical Machines
In this section, we prove Theorem 1.1. We will prove by contradiction: assume for contrary that
ALG > 1 + γ, we seek to derive a contradiction, e.g., no schedule could complete all jobs before
time 1 (Recall that we assume OPT = 1). To do so, we introduce two types of arguments: we use
a bin-packing argument to show that the last job must be of small size, as otherwise there exists
a set of infeasible large jobs; then we use an efficiency argument (built on the Leftover Lemma) to
show that the total processing (excluding idle and waste periods) our algorithm completes exceed
m, the maximum possible processing OPT does.
For convenience of presentation, in the rest of the paper, we adopt the minimum counter-example
assumption Noga and Seiden [2001], i.e., we consider the instance with the minimum number of
jobs such that ALG > 1 + γ and OPT = 1. As an immediate consequence of the assumption, we
get that no job arrives after sn. This is because such jobs do not affect the start time of n and
therefore could be removed to obtain a smaller counter example.
Recall that in our algorithm, we set α = 1
20000 .
We first provide some additional structural properties of our algorithm, which will be the building
blocks of our later analysis.
200 and β =
2− 1. Define γ := 1
2 − , where = 1
√
4.1 Structural Properties
Observe that if a job is replaced, then it must be the minimum job among the m jobs that are
currently being processed. Hence immediately we have the following lemma, since otherwise we
can find m + 1 jobs (including the replacer) of size larger than 1
2 .
Fact 4.1 (Irreplaceable Jobs) Any job with size at least 1
2 cannot be replaced.
Next, we show that if a job i is pending for a long time, then each of the jobs processed at time
si must be of (relatively) large size.
Lemma 4.1 For any job i, we have pj > min{si − ri, pi} for all j ∈ J(si).
It suffices to consider the non-trivial case when si > ri. Consider any j ∈ J(si). If pj > pi
Proof:
2, then we have pj > min{si − ri, pi}. Otherwise, we consider time sj(si), at which
or sj(si) < ri
job j is scheduled. Since pj < pi and sj(si) > ri (i is already released), we know that pi must be
processed at sj(si). Hence we know that i is replaced during (sj(si), si), which is impossible since
j (which is of smaller size than i) is being processed during this period.
Specifically, since ALG = sn +pn > 1+γ and rn +pn ≤ OPT = 1, we have sn−rn > γ. Applying
Lemma 4.1 to job n gives the following.
2Note that we use sj(si) here instead of sj as j can possibly be replaced after time si.
8
Corollary 4.1 (Jobs Processed at Time sn) We have pj > min{γ, pn} for all j ∈ J(sn).
In the following, we show two lemmas, one showing that if a job released very early is not
scheduled, then all jobs processed at that time are (relatively) large; the other showing that if a
job is replaced, then the next time it is scheduled must be the completion time of a larger job.
1+β for all j ∈ J(rk).
Lemma 4.2 (Irreplaceable Jobs at Arrival) If a job k is not scheduled at rk and rk < αpk,
then pj ≥ pk
Proof: By our replacement rule, k is not scheduled at rk either because k is not the largest
pending job at rk, or k is the largest pending job, but the minimum job in J(rk) is not replaceable.
For the second case, since the minimum job i in J(rk) is processed at most rk < αpk, job i must
violate our third replacement rule, that is pi ≥ pk
1+β . For the first case, let k(cid:48) be the first job of size
at least pk that is not scheduled at its release time. Then we have rk(cid:48) < rk < αpk < αpk(cid:48). Hence
pk
by the above argument every job in J(rk(cid:48)) is of size at least
1+β > rk. Thus every job in J(rk) is
also of size at least pk
1+β .
Lemma 4.3 (Reschedule Rule) Suppose some job k is replaced, then the next time k is scheduled
must be the completion time of a job j such that pk < pj ≤ sk.
Suppose k is replaced at time t and rescheduled at time t(cid:48). Since k can only replace
Proof:
other jobs at rk, the next time k is scheduled must be when some machine becomes idle. And this
happens only if the job j processed before t(cid:48) on this machine is completed. Moreover, since k is
pending from t to t(cid:48), if sj ≥ t, i.e., k is pending when j is scheduled, then (by greedy scheduling
rule) we have pj > pk; otherwise j is being processed at time t, and we also have pj > pk as k is the
smallest job among all jobs in J(t) by the replacement rule. Hence, we have sk ≥ cj ≥ pj > pk.
We present the central lemma for our bin-packing argument as follows. Intuitively, our bin-
packing argument applies if there exists time t1 and t2 that are far apart, and the jobs in J(t1) and
3 ): if J(t1) ∩ J(t2) = ∅, then together with job n, we have found
J(t2) are large (e.g. larger than 1
an infeasible set of 2m + 1 large jobs; otherwise (since t1 and t2 are far apart) we show that the
jobs in J(t1) ∩ J(t2) must be even larger, e.g. larger than 2
3 .
Lemma 4.4 (Bin-Packing Constraints) Given non-idle times t1, t2 such that t1 < t2 and n /∈
J(t1)∪J(t2), let a = minj∈J(t1){pj} and b = minj∈J(t2){pj}, none of the following cases can happen:
(1) a > 1
(2) min{a, b, pn} > 1
2+β , and t2 − t1 > 1 − min{a, b, pn}.
3(1+β) , t2 − t1 > 2
3 and pn > 1
3 ;
3 , b > 2
Proof: We show that if any of the cases happens, then we have the contradiction that OPT > 1.
We first consider case (1). We show that we can associate jobs to machines such that every
machine is associated with either a job of size larger than 2
3 , or two jobs of size larger than 1
3 .
Moreover, we show that every job is associated once, while n is not associated. Note that since
pn > 1
done with this machine; otherwise (when pj ≤ 2
associate another job of size larger than 1
3 , such an association would imply the contradiction that OPT > 1.
First, we associate every j ∈ J(t2) to the machine that it is processed on. If pj > 2
3 then we are
3 ), we have sj(t2) > t1 and we show that we can
• If the job i ∈ J(t1) processed on this machine is not replaced, or pi ≤ 2
3 to this machine.
3(1+β) , then we associate
i with this machine (it is easy to check that i has not been associated before);
9
• otherwise the first job that completes after t1 must be of size larger than (1 + β)pi > 2
3 , which
In both cases we are able to do the association, as claimed.
association argument similar as before: let M be the machine that job i is processed on.
also has not been associated before. Thus we associate it to this machine.
Next we consider case (2). Consider any job i ∈ J(t1), we have pi > 1
2+β . We apply an
• If i is replaced, then we associate the first job that completes on this machine after i is
• otherwise if pi > 1 − min{a, b, pn} then we associate i to M ;
• otherwise we know that i completes before t2, and we can further associate to M the job in
2+β > 1 − min{a, b, pn}, to machine M ;
replaced, which is of size larger than 1+β
J(t2) processed on M .
It is easy to check that every job is associated at most once. Hence every machine is associated
with either a job of size larger than 1 − min{a, b, pn}, or two jobs of size larger than min{a, b, pn},
which (together with pn > 1
2+β ) gives OPT > 1, a contradiction.
4.2 Upper Bounding pn: Bin-Packing Argument
We show in this section how to apply the structural properties from Section 4.1 to provide an
upper bound 1
2+α , then
Lemma 4.4 leads us to a contradiction. We first prove the following lemma (which will be further
used in Section 4.4), under a weaker assumption, i.e., pn > 1
2+α on pn. Recall that we assume ALG > 1 + γ. We show that if pn > 1
2+β .
Lemma 4.5 If pn > 1
2+β , then job n is never replaced.
Proof: Assume the contrary and consider the last time when n is replaced. Note that by Fact 4.1,
we have pn < 1
2 . Suppose n is replaced by job l at time rl. Then by our replacement rule, we have
pl ≥ (1 + β)pn. As rl + pl ≤ 1 and sn + pn > 1 + γ, we have
sn − rl > (1 + γ − pn) − (1 − pl) > γ + βpn > 1 − pn > pn,
2+β and pn > 1
where the second last inequality holds since γ > 1
2+β . Since rn < rl, by Lemma 4.1,
we have minj∈J(sn){pj} > min{sn − rn, pn} = pn. Thus we can apply Lemma 4.4(2), with t1 = rl,
t2 = sn, a = minj∈J(t1){pj} = pn and b = minj∈J(t2){pj} > pn, and derive a contradiction.
Lemma 4.6 (Upper Bound on Last Job) We have pn ≤ 1
Proof: We first show a weaker upper bound: pn ≤ 1+β
2 > 1
2 .
2 > 1 − γ;
As shown in the proof of Lemma 4.1, for all j ∈ J(sn), if sj > rn, we have pj > pn > 1+β
otherwise sj < rn and we have pj > sn − rn ≥ (1 + γ − pn) − (1 − pn) = γ. Among the m + 1 jobs
J(sn) ∪ {n}, there exist two jobs, say k and j, that are scheduled on the same machine in OPT.
Moreover, we have sk, sj < rn, since otherwise one of them is larger than 1 − γ and they cannot be
completed in the same machine within makespan 1. Let k be the one with a smaller release time,
i.e., rk < rj. Then we have rk ≤ 1 − pk − pj < 1 − 2γ < αpk.
2 . Assume the contrary that pn > 1+β
2+α .
Observe that k is never replaced, as otherwise (by Lemma 4.3, the reschedule rule) we have
sk > pk, which implies rn + pn > sk + pn > pk + pn > γ + 1+β
2 > 1, a contradiction.
By Lemma 4.2, we know that sk = rk (k is scheduled at its arrival time rk), as otherwise the
1+β , which is also a contradiction
1+β . Thus we have sk ≥ γ
γ
minimum job in J(rk) is of size at least
as rn + pn > sk + pn > γ
1+β + 1+β
2 > 1 (recall that γ = 1
2 − and β =
√
2 − 1).
10
By rk + pk + pj ≤ 1 and rk + pk + pn > 1 + γ, we have pn > 2γ. Hence rn < 1 − 2γ < αpn. By
2 . Hence we have
Lemma 4.2, we know that the minimum job in J(rn) is of size at least
the contradiction that there are m + 1 jobs, namely J(rn) ∪ {n}, of size larger than 1
2 .
Hence we have that pn ≤ 1+β
2 . Now assume that pn > 1
By Lemma 4.5, we know that n is never replaced. By Corollary 4.1, all jobs in J(sn) ∪ {n} are
of size at least min{γ, pn} ≥ 1
2+α . Let k, j ∈ J(sn)∪{n} be scheduled on the same machine in OPT
such that rk ≤ 1 − pk − pj < 1 − 2
2+α ≤ αpk. Note that different from the previous analysis (when
pn > 1+β
1+β ≥
3 . Then we can apply Lemma 4.4(1) with t1 = rk and t2 = sn to derive a contradiction
By Lemma 4.2, if k is not scheduled at rk, then each job i ∈ J(rk) has size pi ≥ pk
2 ), it is possible that n ∈ {k, j}.
pn
1+β > 1
2+α .
1
(2+α)·(1+β) > 1
(observe that t2 − t1 = sn − rk > (1 + γ − pn) − (1 − pk − pj) > 3
2 ≥ 2
3 ).
Hence we know that k is scheduled at time rk (thus we conclude that k (cid:54)= n).
We show that k must be replaced (say, by job l at time rl) , as otherwise by rk + pk + pn > 1 + γ
2+α , as otherwise rn ≤
By Lemma 4.3, we have sk > pk ≥ 1
2+α > 1
2 ,
and rk + pk + pj ≤ 1, we have pn > γ + pj > γ + 1
1− pn < 1
contradicting Fact 4.1 (jobs larger than 1
2+α < sk (k is scheduled at sk, when n is pending), which implies pk > pn > 1− 1
2 cannot be replaced). Since pl > (1 + β)pk, we have
2+α . We also have that pn ≤ 1 − 1
2 , which is a contradiction.
2+α − 1+β
2+α > 1+β
sn − rl > (1 + γ − pn) − (1 − (1 + β)pk) = (1 + β)pk + γ − pn
>
1 + β
2 + α
+ γ − 1 + α
2 + α
>
1 + α
2 + α
.
Then we can apply Lemma 4.4(2), with t1 = rl, t2 = sn, a = pk, b ≥ 1
contradiction.
2+α > 1
2+β , to derive a
Given the upper bound 1
any job of size larger than 1
2+α on pn, we show the following stronger version of Lemma 4.5 that
2+β cannot be replaced.
Corollary 4.2 (Irreplaceable Threshold) Any job of size larger than 1
2+β cannot be replaced.
Proof: Assume the contrary that some job j of size larger than 1
at rk). Then we have pk > (1 + β)pj and mini∈J(rk){pi} = pj > 1
Corollary 4.1, we have mini∈J(sn){pi} > min{pn, γ} > 1
2+β is replaced (say, by job k
2+β . On the other hand, by
2+α < γ, we have
sn − rk > (1 + γ − pn) − (1 − 1 + β
2 + β
2+β . Since pn < 1
) = γ − pn +
1 + β
2 + β
>
1 + β
2 + β
.
Hence we can apply the bin-packing argument to derive a contradiction: by Lemma 4.4(2), with
t1 = rk, t2 = sn, a = pj and b > 1
2+β , we have a contradiction.
4.3 Lower Bounding pn: Efficiency Argument
Next we establish a lower bound on pn, applying the Leftover Lemma. First, observe that if n is
never replaced, then n is pending from rn to sn, where rn ≤ 1 − pn; if n is replaced, then n is
pending from rl to sn, where rl is the last time n is replaced. Since rl ≤ 1 − (1 + β) · pn < 1 − pn,
in both cases n is pending during time period [1 − pn, sn). Hence we have the following fact.
Fact 4.2 (Non-idle Period) Every t ∈ [1 − pn, sn) is non-idle.
11
As a warm-up, we show the following simple lower bound on pn using the Leftover Lemma.
Lemma 4.7 (Simple Lower Bound) We have pn > 1
Proof: Let t = 1− pn. Since there is no waste after time 1, the total waste located after time t is
W1 − Wt. Since no machine is idle during time [1− pn, sn), the total processing our algorithm does
after time t is at least m(sn− t)− (W1− Wt). On the other hand, the total processing our algorithm
does after time t is upper bounded by m(1 − t) + ∆t (Observation 3.1). Applying Lemma 1.1 (the
Leftover Lemma) on ∆t, we have
3 − 2α.
m(sn − t) < m(1 − t) + (W1 − Wt) + ∆t ≤ m(1 − t) + W1 +
tm ≤ m(1 − t) + αm +
1
4
1
4
tm.
Note that the last inequality follows since (by our replacement rule) each job j can only create
a waste of size at most α · pj, and the total size of jobs is at most m. Thus we have
ALG = sn + pn ≤ 1 + α + 1−pn
which implies that (recall that we assume ALG > 1 + γ = 3
pn ≥ 4
3 · (ALG − 5
4 − α) > 4
3 · ( 3
2 − − 5
where the last inequality holds by our choices of parameter, i.e., = 1
4 + α + 3
4 pn,
4 + pn = 5
2 − )
4 − α) = 1
3 − 4
3 ( + α) > 1
3 − 2α,
20000 and α = 1
200 .
Observe that the Leftover Lemma provides tighter upper bounds for smaller values of t. Thus
in the above proof, if we can find a smaller t such that there is no (or very little) idle time from t
to sn, then we can obtain a stronger lower bound on pn.
Lemma 4.8 (Lower Bound on Last Job) We have pn > 1
2+β . Then by Corollary 4.1, we have pj > min{γ, pn} = pn
Proof: Assume for contrary that pn ≤ 1
for all job j ∈ J(sn). Hence at least two jobs k, j ∈ J(sn)∪{n} are scheduled on the same machine
in OPT, such that rk < 1 − 2pn. We prove the following claim3, which enables us to use a refined
efficiency argument, i.e., apply the Leftover Lemma on a earlier time t = rk. For continuity of
presentation, we defer its proof to the end of this subsection.
Claim 4.1 The total idle time Isn − Irk during time period [rk, sn] is at most 3α · m.
2+β .
By the above claim, the total processing ALG does after time t = rk is at least m(sn − t) −
(W1 − Wt) − 3αm. On the other hand, by Observation 3.1 and Lemma 1.1 we have
4 tm + Wt.
m(sn − t) − (W1 − Wt) − 3αm ≤ m(1 − t) + ∆t ≤ m(1 − t) + 1
Rearranging the inequality, we have (recall that t = rk < 1 − 2pn)
1 + γ < ALG = sn + pn ≤ 1 + 1−2pn
which is a contradiction by our choice of parameters.
4 + 4α + pn ≤ 5
4 + 4α + β
2(2+β) ,
It remains to prove Claim 4.1.
Proof of Claim 4.1: Recall that we assume 1
2+β , and there exist two jobs
k, j ∈ J(sn) ∪ {n} (of size at least pn) scheduled on the same machine in OPT, such that rk < rj.
Observe that sk − rk > (1 + γ − pn − pk)− (1− pk − pj) > γ. We first upper bound the total idle
time in [rk, sk) by 3α · m, and then show that there is no idle time after time sk (until time sn).
3We remark that the proof of Claim 4.1 relies on the fact that pn is not too small. Hence Lemma 4.7 (the warm-up
3 − 2α < pn ≤ 1
lower bound) is necessary for achieving the improved lower bound (Lemma 4.8).
12
Lemma 4.9 The total idle time in [rk, sk] is at most 3α · m.
Proof: As k is released at rk while (eventually) scheduled at sk, we know that if there is an idle
period [a, b] ⊆ [rk, sk), then k must be processed during time period [a, b] and replaced after time
b. Since k cannot be replaced if it is processed α, immediately we have b − a < α.
Now consider any fixed machine M . Suppose a1 ≥ rk is the first time machine M becomes idle.
We know that k is being processed on some other machine at time a1 (as no job is pending), and
replaced at some rl < a1 + α. Then the job processed on M at time rl must be of size larger than
pk, which implies that there is no idle period in [rl, a1 + pk].
By Fact 4.2, there is no idle time after 1 − pn < 2
machine M during time period [rk, sk) is at most (cid:100) 2
time in [rk, sk) (on all m machines) is at most 3α · m.
pk
3 + 2α. Hence the number of idle periods on
(cid:101) ≤ 3, which implies that the total idle
3 +2α
Next we show that every time t ∈ (sk, sn) is non-idle.
Suppose otherwise, let t ∈ (sk, sn) be the last idle time before sn. Since there is no pending job
at time t, we know that n is either being processed at time t, or is not released.
For the first case, since t < sn, we know that n is replaced after time t. Let rl be the last time
n is replaced, we have rl ≤ 1 − pl < 1 − (1 + β)pn. Note that n is pending from rl to sn, which is
of length
sn − rl > (1 + γ − pn) − (1 − (1 + β)pn) > γ + β · pn.
Since sk < t < rl, the total processing our algorithm does after time rk is m(sn − rl) + m(sk −
rk) − 4α · m > m, contradicting OPT = 1.
never replaced, i.e., n is pending from rn to sn, which is of length larger than γ.
Now we consider the second case, i.e., sk < t < rn. By the same reasoning, we know that n is
Recall that we have sk − rk > γ, and since k ∈ J(sn), we have pk > sn − rn > γ.
Claim 4.2 We have rn − sk > 6α + 2 (by the efficiency argument).
Proof: Assume the contrary that sn − rk ≤ 1 + 6α, we know that the total idle time after sk is
at most rn − sk ≤ 1 + 6α − 2γ = 6α + 2. Then we have (by Observation 3.1, Lemma 1.1:
ALG = sn + pn ≤ 1 + α +
1
4
≤ 1 + 10α + 2 +
≤ 1 + (10α +
rk + 3α + 6α + 2 + pn
(1 − pn − γ) + pn
1
8
4(2 + β)
+
1
4
9
4
+
3
) ≤ 1 + γ,
contradicting our assumption that ALG > 1 + γ.
On the other hand, we have the following contradicting claim.
Claim 4.3 We have rn − sk < 6α + 2 (mainly by the bin-packing argument).
Proof: Observe that k is processed from sk to sn. Hence we have pk ≥ sn − sk. If pk < 1
then we are done as rn − sk ≤ pk − (sn − rn) < 1
2 + 6α − γ = 6α + .
argument similar to Lemma 4.4 to show that on every machine M , we can find
Assume for contrary that rn − sk ≥ 6α + 2, we have pk ≥ 1
• either a job of size larger than 2
• a job a /∈ J(sn) of size pa > 1
3 + 2α, or
3 + 4α, and a job b ∈ J(sn) (of size pb > 1
3 − 2α),
2 + 6α. We apply an association
2 + 6α,
13
that are completed on M , and none of them is n.
Let x ∈ J(sk) and y ∈ J(sn) be processed on M . By Lemma 4.1 we have px > γ > 1
3 + 4α and
3 − 2α. If x (cid:54)= y, then we are done with this machine, i.e., a is the first job completed
py > pn > 1
after sk on M and b = y. Note that we have a (cid:54)= b as either a = x or a is a replacer, which cannot
be processed at sn > 1.
If x = y, then we are also done if px > 2
Now suppose px ≤ 2
3 + 2α. Since x is processed at sn > 1, we know that the job z processed
3 + 2α.
2 ) is pending during [rk, sk], we have
before sx must be completed. Moreover, since k (of size > 1
pz ≥ (sn − rk) − px > 2γ + 6α + 2 − 2
3
− 2α =
1
3
+ 4α.
Hence we have found a = z and b = x completed on M and the association is completed.
Thus in OPT (which completes all jobs before time 1), three jobs of size in ( 1
3 , as mini∈J(sn){pi} > pn.
3 + 4α)
from J(sn) are scheduled on the same machine, which means that at least one of them, say x, is
released before 6α < sk. Observe that we also have pn ≤ 1
Recall that t ∈ (sk, rn) is the last idle time before rn. As there is no pending jobs at time t, x
must be processed at time t, but replaced later (as otherwise px > sn − rn > γ). Hence we know
3 − 2α), which implies that total idle time in [rk, sn] can be
that t ≤ 1 − (1 + β)px < 1 − (1 + β)( 1
upper bounded by
3 − 2α, 1
(t − sk)m < 1 − (1 + β)(
− 2α) − γ.
1
3
Hence we have the following contradiction:
rk + (1 − (1 + β)(
+ (1 − (1 + β)(
1
3
where the second inequality holds since rk ≤ 1 − pk − pj ≤ 1
ALG = sn + pn ≤ 1 + α +
3
4
≤ 1 + α +
2 − pn, and pn ≤ 1
3 .
As the two claims are contradicting, there is no idle time during (sk, sn).
− 2α) − γ) + pn
1
3
− 2α) − γ) ≤ 1 + γ,
1
4
· 1
3
+
1
4
· 1
2
4.4 A Hybrid Argument
1
We have shown that assuming ALG > 1 + γ, the size of the last job n can be bounded as
2+β <
pn ≤ 1
2+α . In the remaining part of this section, we use a hybrid argument to show that we can
either use the bin-packing argument to find a set of infeasible large jobs; or derive a contradiction
using the efficiency argument.
1
2+β < pn ≤ 1
2+α < γ, we have sn = ALG − pn > 1. Thus we
General Framework. Given that
have sj (cid:54)= rj for all j ∈ J(sn) (as they are processed at time sn > 1). Moreover, by Corollary 4.1,
2+β . In other words, there exists a set J(sn)∪{n} of large
we have minj∈J(sn){pj} > min{γ, pn} > 1
jobs, each of which is never replaced (by Corollary 4.2), and none of them is scheduled at its release
time. We know that at least two of them, say k and j (assume k is released earlier), are scheduled
on the same machine in OPT. Since sk − rk > (1 + γ − pn − pk) − (1 − pk − pj) ≥ γ, we know that
k is pending from rk to sk, which is a period of length γ. Thus either our algorithm finishes a lot
of processing during this period (then we can use the efficiency argument), or there are many idle
and waste periods during this period (then we can use the bin-packing argument to find another
m large jobs, e.g., larger than 1
3 ) .
We first show that rk and sn cannot be too far apart.
14
Lemma 4.10 We have sn − rk ≤ 1.
Proof: Assume for contrary that sn − rk > 1. We show that can apply the bin-packing argument
to find a set of infeasible large jobs. We apply an association argument as in the proof of Lemma 4.4
to show that every machine is associated with either a job of size larger than 1+β
2+β , or two jobs of
size larger than 1
2+β )
is not associated, which implies a contradiction.
2+β . Moreover, every job is associated at most once, while n (recall that pn > 1
Fix any machine M . Consider i ∈ J(rk) and x ∈ J(sn) processed on M .
• If pi > 1
• if pi ≤ 1
since sn − rk > 1, we have i (cid:54)= x.
2+β , then we associate i and x (both of them cannot be replaced) to M . Observe that
2+β , we consider the job l processed on M after i. Since rk + pi < sk, we know that l
2+β and is completed on M . If l (cid:54)= x then we associate
starts during (rk, sk), hence pl > pk > 1
l and x to M ; otherwise px > (sn − rk) − pi > 1+β
2+β and we can associate x to M .
In both cases we can associate jobs to machines as claimed, which gives a contradiction.
Lemma 4.10 immediately implies that the following stronger lower bound on pn. Note that the
3 (thus more convenient
new lower bound 1
to use the bin-packing argument), while for the previous lower bound we have
2 − 3α is crucial in the sense that we have
2 − 3α) > 1
1+β ( 1
1
1
(2+β)(1+β) < 1
3 .
Corollary 4.3 We have pn > 1
2 − 3α.
Proof:
If sk > rn, then we know that there is no idle time between rk and sn, as during this
period, either k or n is pending. Hence by Observation 3.1, we have m(sn − rk) < m(1 − rk) +
∆rk + (W1 − Wrk ), which implies sn < 1 + 1
4 rk + α by the Leftover Lemma. Therefore, we have
(recall that we have rk < 1 − pk − pn < 1 − 2pn)
1
4
− 3α) = 1 + (
ALG = sn + pn < 1 + α +
(1 − 2pn) + pn
≤ 1 + α +
) ≤ 1 + γ,
+
(
1
4
1
2
1
2
1
2
− α
2
contradicting our assumption that ALG > 1 + γ.
If sk < rn, then we have pk > sn − rn > γ (recall that pk is processed at time sn). By
Lemma 4.10, the total idle time between sk and rn is at most (1 − 2γ)m = 2m. Then by the
Leftover Lemma,
ALG = sn + pn < 1 + α +
1
4
(1 − γ − pn) + pn + 2 ≤ 3
2
− 1
4
(5α − 9) ≤ 1 + γ,
which is also a contradiction.
2 − 3α is still less than 1
1
Unfortunately, 1
2 − 3α < pn ≤ 1
and [rn, sn], both of length at least γ, and contain no idle time.
2+α . Hence it remains to consider the subtle case when
2+α . Note that so far we have proved that there exists two periods, namely [rk, sk]
From the proof of Corollary 4.3, we observe that depending on whether the two intervals overlap,
the analysis can be quite different. Hence we divide the discussion into two parts. As we will show
later, the central of the analysis is to give strong upper bounds on W1.
4.4.1 Overlapping Case: when sk ≥ rn
Note that in this case, from rk to sn, the largest pending job is always at least pn. Hence there
is no idle time in [rk, sn]. Moreover, every job that starts in [rk, sn] must be larger than pn (and
hence cannot be replaced). First we show that rk ≥ αpk.
15
Suppose otherwise, then by Lemma 4.2, we have pi > 1
we can apply Lemma 4.4(1) with t1 = rk, t2 = sn, a > 1
1+β ( 1
2 − 3α) > 1
3 and b > pn > 1
3 for all i ∈ J(rk). Hence
2 − 3α > 2
3(1+β) , for which
t2 − t1 = sn − rk > (1 + γ − pn) − (1 − pk − pj) > γ + pn >
2
3
,
to derive a contradiction.
Lemma 4.11 There exists a time r ≤ rk, at which the minimum job processed has been processed
at least α · pk; moreover, from r to sn, the largest pending job is always at least pn.
If k is the largest pending job at rk, then let x be the job of minimum size processed at
Proof:
time rk, i.e., x = arg minj∈J(rk){pj}. For the same reason as argued above, we have px ≤ 1
1+β .
Since k does not replace x, we conclude that x must be processed at least αpk. Hence the corollary
holds with r = rk.
Otherwise we consider the earliest time r before rk such that from r to rk, the largest pending
job is always at least pk. Note that we must have r = rk(cid:48) for some job k(cid:48) of size pk(cid:48) > pk. Moreover,
k(cid:48) is the largest pending job at rk(cid:48), but not scheduled. Let x = arg minj∈J(rk(cid:48) ){pj}. We have
px ≤ 1
3 > 1− pk − pj is a contradiction. Hence by a similar argument as above,
the corollary holds with r = rk(cid:48).
3 , as otherwise rk > 1
3 < pk
By the above lemma, every job that starts in [r, sn] must be larger than pn > 1
2+β , which cannot
be replaced. Thus, from r to sn, there is no idle time, and if there is any waste, then it must come
from the jobs in J(r).
the following upper bound on W − P.
Let P be the total processing our algorithm does after time sn, and let W = W1. We show that
Lemma 4.12 (Total Waste) We have
W − P ≤ 2α · rm +
1
2 + β
(rm − Ar − Ir).
We first show how to use Lemma 4.12 to prove the desired competitive ratio, and defer the
proof of lemma (which is long and contains many cases) to Section 5.
Corollary 4.4 When sk ≥ rn, we have ALG ≤ 1 + γ.
Proof: First by Observation 3.1, we have (since there is no idle time after r)
m(sn − r) + P − (W − Wr) ≤ m(1 − r) + ∆r,
which (by Lemma 4.12) implies
sn ≤ 1 +
(W − P + ∆r − Wr) ≤ 1 + 2α · r +
1
m
r
2 + β
+
1
m
(∆r − Wr − Ar + Ir
2 + β
).
(Leftover Lemma) and ∆r − Wr ≤ min{Ar, Ir} (Claim 3.1), we have
Since ∆r − Wr ≤ rm
4
sn ≤ 1 + 2αr +
r
2 + β
Given α = 1
200 and β =
2 min{Ar, Ir} − (Ar + Ir)
1
+
(
m
√
2 − 1, it is easy to check that 2α + 4+β
2 + β
2 + β
+
β
) ≤ 1 + (2α +
4 + β
4(2 + β)
)r.
2.14 . Hence we have
rm
4
4(2+β) < 1
0.14
ALG = sn + pn ≤ 1 +
1
+
where in the second inequality we use the fact that r ≤ rk < 1 − 2pn.
(r + 2pn) +
2.14
2.14
pn ≤ 1 +
1
0.14
2.14
≤ 1 + γ,
2.14(2 + α)
16
4.4.2 Disjoint Case: when sk < rn
Note that in this case we have two disjoint time periods, namely [rk, sk] and [rn, sn], both of length
at least γ, and during which there is a pending job of size at least pn. Moreover, by Lemma 4.10,
we have rn − sk ≤ 1 − 2γ = 2.
Lemma 4.13 We have minj∈J(rn){pj} > 1
3 .
Proof: Let x be the minimum job in J(rn). Suppose px ≤ 1
1+β . Then we have sx(rn) > sk,
which means that x is processed at most rn − sk < αpn. Hence we know that n is not the largest
pending job at rn, as otherwise x would be replaced.
3 < pn
Then we consider the first job i of size pi > pn that is released in [sx(rn), rn] and not scheduled.
Since x is the largest pending job at sx(rn), we know that i must be the largest pending job at
ri. We also know that minj∈J(ri){pi} ≤ px < pi
1+β , and the minimum job y has been processed less
than αpi, which is impossible as i should have replaced y.
Let P be the total processing our algorithm does after time rn + γ, and let W = W1. We show
the following lemma, which is an analogy (weaker) version of Lemma 4.12 in previous section.
Lemma 4.14 (Total Waste) We have
W − (Wrn − Wsk ) − P ≤ (
1
2
+ α) · rk · m + 8α2m.
For continuity of presentation, we defer its proof to Section 5.
Lemma 4.15 When sk < rn, we have ALG ≤ 1 + γ.
Proof: Since there is no idle time in [rk, sk] ∪ [rn, sn], by Observation 3.1, we have
m(1 − rk) ≥ 2γm + P − (W − (Wrn − Wsk )) ≥ (1 − 2)m − (
+ α) · rk · m + 8α2m,
1
2
which implies rk ≤ 2+8α2
2−α
1
< αpk, contradicting the assumption rk ≥ αpk.
5 Bounding Total Waste
In this section, we prove Lemma 4.12 and 4.14, which give upper bounds on the total waste W.
Proof of Lemma 4.12: We upper bound the total waste by partitioning it into four parts.
only be scheduled after sn. Hence this part of wastes can be upper bounded by(cid:80)
If any job i ∈ J(r) is replaced, then pi ≤ 1
Part-1: wastes created after r.
2+β and hence can
i∈J(r):si>sn
pi.
Part-2: wastes created by jobs never processed from r to sn. As sn − r > 1− r, the total
processing time of jobs that are never processed in [r, sn] is at most r · m. Hence the total waste
created by these jobs is at most α · r · m.
Note that excluding Part-1 and Part-2, the wastes created by jobs that are ever scheduled
2+β and will not be
from r to sn can only be created by jobs in J(r), as other jobs are larger than 1
replaced (hence cannot be a replacer before r).
17
Part-3: wastes created by i ∈ J(r) at ri < si(r). As i is processed at r, we know that i is
replaced during [ri, si(r)], and rescheduled at si(r). Note that the waste is at most α · pi, since it
is replaced by i. Consider the job completed at si(r), by Lemma 4.3 we know that it is larger than
pi. Thus we can construct a one to one mapping from this kind of wastes to jobs completed before
r. Moreover, each waste is bounded by α fraction of its image. Hence the total waste of this part
is at most α · r · m.
Part-4: wastes created by i ∈ J(r) at ri = si(r). We will show that this part of wastes can
2+β (rm− Ar − Ir). Let t be the last idle time before r. We
be upper bounded by(cid:80)
have Ar = At =(cid:80)
i /∈J(r):si>sn
pi + 1
u∈J(t) min{δu, pu}, where δu is the total pending time of u before time t.
Consider any waste (of size) w created by i ∈ J(r) at time ri = si(r) on machine M .
We interpret the waste as the time period [ri − w, ri] on M . In the following, we charge the
waste to a job of size at least w that is not in J(r) and starts after sn, or to a set of time periods of
total length at least (2 + β)w located before r. We show that (over all wastes) every job and time
period will be charged at most once, and none of them overlaps with the idle periods. Moreover,
we show that on every machine M(cid:48), we can find a non-idle time period before r of length at least
min{δu, pu} that is not charged, where u ∈ J(t) is the job processed on M(cid:48). (It is easy to check
whether the charged time period overlap with the idle periods. However, it is more involved to
check the disjointness with pending periods, which is the main focus of our charging argument.)
Note that such a charging argument gives an upper bound(cid:80)
2+β (rm − Ar − Ir) on
l /∈J(r):sl>sn
pl + 1
the total wasted in this part.
px ≤ 1
argument.
Let x be the minimum job in J(r), and y be the job where w comes from. Recall that we have
3 and x is processed at least α · pk, i.e., rx = sx(r) ≤ r − α · pk. Now we present our charging
1. If r − ri ≥ (1 + β)w, then we can charge w to itself, together with the time period [ri, r] on
machine M . It is easy to see that the charged periods are of total length at least (2 + β)w,
and disjoint from the idle periods. Define u(M ) ∈ J(t) as the job processed on M at the last
marginal idle time before time r. If u = i, then since su(t) = ru, we have δu = 0; otherwise
su(t) ≤ ri − w and δu is at most the total length of non-idle periods before ri − w. In both
cases we find a non-idle time period, namely Tu(M(cid:48)), of length at least δu that is not charged.
2. If py < px (which implies y /∈ J(r)) and sy > sn, then let w(cid:48) be the first waste from y that
has not been charged.
We charge w(cid:48) to py > w(cid:48) (processed after sn), and charge the remaining wastes from y, which
is of total size at most (1 + β)w(cid:48), to themselves, w(cid:48), and the processing (before r) of the
job that creates w(cid:48). Note that the charged periods are of total length at least (3 + 2β)w(cid:48) >
(2 + β)(1 + β)w(cid:48), and disjoint from the idle periods. Since every charged period is either
processing of jobs in J(r), or the wastes they create, applying the same argument as above,
on every machine M(cid:48) that has a charged waste, we can find a non-idle time period of length
at least δu(M(cid:48)) that is not charged.
18
sypysnw'r 3. If py < px and sy < sn, then we know that y is rescheduled (say, on machine M(cid:48)) after ri,
and completed before r. Hence we have r − ri > py > w. Moreover, by Lemma 4.3, the job z
completed at sy on M(cid:48) must be larger than py.
We charge w to w, [ri, r] on machine M , and [sy, cy] on machine M(cid:48). The total length of
charged periods is at least w + 2py > 3w. As before, a non-charged non-idle time period of
length δu(M ) can be found on M . Now consider u(M(cid:48)).
• If pz > pu(M(cid:48)), then [sz, cz] is the desired non-charged non-idle time on M(cid:48);
• otherwise we know that u(M(cid:48)) is not pending during [sy, cy], as y is smaller than u(M(cid:48))
but is scheduled and completed. Hence Tu(M(cid:48)) is the desired period on M(cid:48).
4. If py > px, then we know that ri < sx(r) ≤ r − αpk, as otherwise x would be replaced instead
of y. Moreover, the next time y is scheduled (say, on machine M(cid:48)) must be before sx(r), as x
cannot be scheduled while y is pending, and we have pz > py for the job z completed before
y on M(cid:48).
We charge w to w, [ri, r] on M , and [sz, ri] on M(cid:48). The charged periods are of total length
w + r − sz > w + pz + αpk > (2 + β)w, and are disjoint from idle periods. As before, a size
δu(M ) non-charged non-idle time period can be found on M . Now consider u(M(cid:48)).
• If [sz, ri] ∩ Tu(M(cid:48)) = ∅, then Tu(M(cid:48)) is the desired period;
• otherwise we know that u(M(cid:48)) (cid:54)= z. Moreover, u(M(cid:48)) (cid:54)= y, as y is not pending during
[sz, ri], which means that u(M(cid:48)) is a job processed after y. First observe that y cannot
be replaced after cz, as otherwise the replacer l must be of size larger than (1 + β)py >
(1 + β)w > r − ri, which implies δu(M(cid:48)) = 0: either uM(cid:48) is the replacer, or the replacer
of the replacer, etc. Hence [cz, cy] is the desired non-charged non-idle period on M(cid:48): we
have py > pu(M(cid:48)) (if uM(cid:48) is ever replaced, then we use Lemma 4.3, otherwise we use the
fact that uM(cid:48) is pending at cz).
Combing the four cases above, we have
W ≤ (cid:88)
i:si>sn
pi + 2α · rm +
1
2 + β
(rm − Ar − Ir) < P + 2α · rm +
1
2 + β
(rm − Ar − Ir),
19
M'Mcyszsyripipzpywr≥αpkpxsx(r)M'Mszczripipzpywr j /∈J(rk):sj >sn
Define P(cid:48) =(cid:80)
as claimed.
Proof of Lemma 4.14: Note that W− (Wrn − Wsk ) is the total wastes located at [0, sk]∪ [rn, sn].
pj to be the total processing of jobs not processed at rk that start after
sn. We show that Wrk (rk) − P(cid:48) ≤ ( 1
2 + 2α) · rk · m and (W − Wrn) + (Wsk − Wrk (rk)) ≤ P − P(cid:48),
combing the two upper bounds we have the lemma.
We first show that (W− Wrn) + (Wsk − Wrk (rk)) ≤ P− P(cid:48). Note that by definition (W− Wrn) +
(Wsk − Wrk (rk)) is at most the total size of wastes created in [rk, sk] ∪ [rn, sn]. Moreover, any such
waste w must come from jobs in J(rk)∪ J(rn)\(J(sk)∪ J(sn)), as otherwise the job is irreplaceable.
Consider any i ∈ J(rn)\(J(sk)∪J(sn)) processed on machine M . Let x ∈ J(sn) be processed on
3 . Hence we have cx−(rn +γ) >
3 − − 3α > 2α, i.e., the contribution of i to P − P(cid:48) is larger than to
Moreover, we have J(rn)\(J(sk) ∪ J(sn)) ≤ 4αm, as otherwise the total processing after rk is
3 − − 3α)m− αm > m, contradicting OPT = 1. This fact will be used later.
Now consider any i ∈ J(rk)\(J(rn) ∪ J(sk) ∪ J(sn)). Then we have si > sn. Hence the
2 + α) · rk · m + 9α2m. The strategy is similar to the proof of
Next we show that Wrk − P(cid:48) ≤ ( 1
M , then we know that the job completed at sx must be larger than 1
(sk + 1
(W − Wrn) + (Wsk − Wrk (rk)) (note that i can be replaced at most twice).
larger than 2γm + 4α( 1
contribution of i to P − P(cid:48) is larger than to (W − Wrn) + (Wsk − Wrk (rk)).
3 + px) − (rn + γ) > 1
job at ri. We show that(cid:80)
Lemma 4.12 (but simpler). We partition Wrk into three parts.
Part-1: wastes created by J(rk). Let R be the set of jobs i /∈ J(rk) that replaced some other
i∈R pj ≤ 4αm, which implies that the total wastes created by jobs not
in J(rk) is at most 4α2m. It suffices to show that on every machine M , we can find a set of jobs
completed on M that are of total size at least 1 − 4α and not in R. Fix any machine M and
consider j1 ∈ J(sk) and j2 ∈ J(sn) processed on M . Note that we have pj1 > γ, pj2 > pn, and
both jobs cannot be replaced. If j1 (cid:54)= j2, then we find two jobs completed on M of total size larger
than γ + pn > 1 − 4α that are not in R; otherwise we know that j1 did not replace any job (as it is
processed at sn > 1). Then we know that the total size of j1 and the job completed at sj1 is larger
than (sn − sk) + min{sk − rk, γ} > 1 − 4α, which completes the analysis.
Part-2: wastes created by i ∈ J(rk) at ri < si(rk). Applying a similar argument as in the
proof of Lemma 4.12 (Part-3), the total waste of this part can be upper bounded by α · rk · m.
Part-3: wastes created by i ∈ J(rk) at ri = si(rk). Note that there is at most one waste on
every machine. Consider any waste of size w created at ri on machine M . Suppose w comes from
job x.
2 , then we can charge w to machine M ;
• If w ≤ rk
• if x ∈ J(rk), then we can charge w to M and the machine that processes x at rk;
• if sx > sn, then the contribution of x to Wrk − P(cid:48) is non-positive;
• otherwise we have x ∈ J(rn)\(J(sk) ∪ J(sn)). As J(rn)\(J(sk) ∪ J(sn)) ≤ 4αm, the total
waste of this part is at most 4α2m.
Hence in total we have
as claimed.
Wrk − P(cid:48) ≤ 4α2m + α · rk · m +
· rk · m + 4α2m,
1
2
20
6 Breaking 5−√
2
5
on Two machines with Restart
We prove Theorem 1.2 in this section, that is, we show that running our algorithm with β = α = 0.2
on two machines, i.e., m = 2, achieves a competitive ratio at most 1.38, strictly better than the
5) ≈ 1.382 for the problem without restart on two machines
best possible competitive ratio 1
(for deterministic algorithms) Noga and Seiden [2001].
2 (5−√
As before, we adopt the minimum counter-example assumption Noga and Seiden [2001], i.e.,
we consider the instance with the minimum number of jobs such that ALG > 1.38 and OPT = 1,
and proceed to derive a contradiction.
6.1 Overview of Techniques
Our analysis for the two-machine case follows the same framework as the general case: we use a
bin-packing argument to upper bound the size of last job, and use an efficiency argument to lower
bound the size of last job, and finally use a hybrid argument to handle the boundary case.
In
this section, we will overview two technical ingredients that are specifically developed for the two-
machine case, namely, a structural result that upper bounds the number of times that large jobs
are replaced, and a refined efficiency argument. Similar to the general case, most of the difficulties
arise when the last job has medium size. For concreteness, readers may consider the size of the last
job n as slightly larger than γ = 0.38, say, pn = 0.4.
Recall that for the two-machine case we fix the parameters β = α = 0.2.
6.1.1 Bounding Major Replacements
We are interested in jobs that have size at least pn. We call such jobs major jobs and refer to the
replacements of major jobs as major replacements. In the case that there are only two machines,
the number of major jobs is at most 4 by a simple bin-packing argument. (Recall that we focus on
medium size last job, say pn = 0.4.) Further, only major jobs can replace major jobs. Hence, we
can show sharp bound on the number of major replacements. In particular, we show that when the
last job is of medium size, there is either no major replacement, or at most one major replacement,
depending on the size of the last job. This is formulated as the following lemmas.
Lemma 6.1 If ALG > 1.38 and pn > 1
2+α , then there is no major replacement.
Proof: Recall that for the two machines case, we set β = α. First we show that job n cannot be
replaced given pn > 1
2+α , which
cannot be scheduled on the same machine with n in the optimal schedule. Suppose n is replaced
(which can happen at most once) by l. Then we know that n is pending from rl to sn, where
2+α . Observe that the replacer of n must be of size (1 + α)pn > 1− 1
sn − rl > (1.38 − pn) − (1 − (1 + α)pn) = 0.38 + αpn >
1
2 + α
.
Hence by Lemma 4.1, we have pk, pj > 1
2+α , where {k, j} = J(sn) are the two jobs processed
If l /∈ {k, j}, then none of k, j, n can be scheduled on the same machine with l in
at time sn.
OPT, which yields a contradiction; otherwise suppose l = k. Then we must have J(rl) = {n, j} as
otherwise we also have the same contradiction as just argued. Then n and j must be scheduled on
the same machine in OPT. If rn < rj, then rn < 1 − pn − pj < 1 − 2
2+α < αpn. Hence n must be
scheduled at rn, as otherwise by Lemma 4.2 the two jobs in J(rn) are of size larger than pn
1+α > 1
3 .
Since
rl − rn > sj − rn > (1.38 − pn − pj) − (1 − pn − pj) = 0.38 > α,
21
it is impossible for l to replace n as n has been processed a portion larger than α. If rj < rn, then
for the same reasoning j is scheduled at rj. As sj − rj > (1.38 − pn − pj) − (1 − pn − pj) = 0.38,
we know that j is replaced, which is impossible as by Lemma 4.3, the job completed at sj is a job
of size larger than 1
Hence we know that n is never replaced. Now suppose some other job of size larger than 1
2+α , apart from l, j and n.
2+α is
replaced at time rx. Then we have px > 1− 1
2+α , while the two jobs in J(rx) are of size larger than
1
2+α . Note that since n is never replaced, it is not scheduled before sn. Further by our assumption
that no job arrives after sn, we have rx < sn. Thus n /∈ J(rx), which implies a contradiction.
Lemma 6.2 If ALG > 1.38 and pn ∈ (0.38,
1
2+α ], then there is at most one major replacement.
The proof of the lemma is deferred to Section 6.3. The main idea is that, when there are two
or more major replacements, then we can find at least four major jobs. Hence as long as we can
find a job of size "not too small", then the bin-packing argument yields a contradiction.
Why is bounding the number of major replacements useful? Recall that in the efficiency
argument, we upper bound the total waste by αm = 2α using the fact that each job can only create
waste once at its release time if it replaces some other job, and the amount of waste created is at
most α times the size of the replacer. Suppose we can find one major job that does not replace
any other job at its arrival. Then, the upper bound of the total waste will significantly decrease by
αpn. (Recall that we focus on medium size last job, say pn = 0.4.)
How do we find major jobs that do not replace other jobs at their arrival? It turns out we can
argue that in order to have ALG > 1 + γ, there must exists some major jobs whose final start times
do not equal their release times. For example, the last job n's final start time definitely does not
equal its release time. For each of these jobs, if it replaces some other job at its arrival, it must
be replaced later on in order to get rescheduled at its final start time. Hence, a major replacement
must occur for each of these jobs. By bounding the number of major replacements, we get that
some of these major jobs must not replace other jobs at their arrivals.
6.1.2 Refined Efficiency Argument
The second technical ingredient is a more careful efficiency argument. Let t be the last idle time
before sn. The total amount of work done in the optimal schedule after time t is at most 2(1 − t).
(Recall that OPT = 1 and there are m = 2 machines.)
How much work does the algorithm process after time t? The algorithm is fully occupied
from time t to time sn. Further, let P be the total processing our algorithm does after sn. Then,
the total amount of processing power after time t by the algorithm is 2(sn − t) + P. However, some
of the processing power is wasted (due to replacements) and some of the workload could have been
done before time t in OPT (the leftover). Recall that Wt is the amount of waste before time t.
Hence, the amount of waste located after time t is W1 − Wt. Also recall that the amount of leftover
is denoted as ∆t. So we have: (the first inequality comes from Observation 3.1)
2(1 − t) ≥ 2(sn − t) + P − ∆t − (W1 − Wt) ≥ 2(sn − t) + P − t
2
− W1,
where the second inequality follows by the leftover lemma. Rearranging terms, the above implies:
ALG = sn + pn ≤ 1 + pn − P
2
+
t
4
+
W1
2
.
22
Then, we will bound each of the terms P, t, and W1: P is trivially lower bounded by pn; t is trivially
upper bounded by 1, while a more careful argument shows that t ≤ 1 − pn (Fact 4.2); and W1 is
trivially upper bounded by 2α. If we plug in the trivial bounds, we get:
ALG ≤ 1.25 + α +
pn
2
.
Hence, we recover the efficiency argument similar to what we have used in Section 4.1 (except that
m pn ≤ pn in the general case). However, if we can obtain improved bounds on
we further relax m−1
P, t, or W1, we would have a better efficiency argument for upper bounding ALG. It is usually
impossible to get better bounds for all of these three quantities. Nonetheless, we manage to do so
for at least one of them in all cases.
We have already provided an argument for getting a better upper bound of W1 by bounding the
number of major replacements. Next, we present some intuitions why it is possible to get improved
bounds for P and t.
Since the jobs are scheduled greedily, if a replaced job x is of size smaller than pn, then it often
happens that job x is rescheduled after sn. Hence while we suffer a loss in the total waste W1 due
to the waste that comes from x, we have an extra gain of px in P. In general, we will develop a
unified upper bound on W1 − P (using the same spirit as in the proof of Lemma 4.12 and 4.14),
which measures the net waste due to replacements.
Apart from the trivial upper bound 1 − pn on t (by Fact 4.2), we can often derive better upper
bounds on t if we do not have a good upper bound on W1 − P. In the case when the total net waste
is large, we can usually find many major jobs processed at the end of the schedule. As we have
only m = 2 machines, during idle periods, only one job can be processed while no job is pending.
Suppose we find four major jobs processed at the end of the schedule, then at least three of them
must be released after the last idle time t. Since OPT = 1, we must have t ≤ 1 − 2pn, which gives
a better upper bound on t.
6.2 Some Basic Facts for Scheduling on Two Machines
As before, let n be the job completed last and we assume all release times, processing times, start
times and completion times are distinct. By the minimum counter-example assumption, no job
arrives after time sn. Since ALG = sn + pn > 1.38, we have sn − rn > (1.38− pn)− (1− pn) = 0.38.
Let J(sn) = {k, j} be the two jobs processed before n is scheduled. By Lemma 4.1, we have
pk, pj > min{0.38, pn}.
Let t be the last idle time before sn. By Fact 4.2, we have t ≤ 1 − pn. Let P be the total
processing our algorithm does after sn. By Observation 3.1, we have (note that m = 2)
2(1 − t) ≥ 2(sn − t) + P − ∆t − (W1 − Wt) ≥ 2(sn − t) + P − t
2
− W1,
which implies
ALG = sn + pn ≤ 1 + pn − P
2
+
t
4
+
W1
2
.
(1)
Definition 6.1 (Uncharged Jobs) We call a job j uncharged, if j does not replace any job at
rj, or the job it replaces is rescheduled strictly after sn.
Note that if a job j is uncharged, then its contribution to the RHS of (1) is non-positive. Let
Q be the total size of uncharged jobs, and define d := 1 − pn − t ≥ 0, we have
ALG = sn + pn ≤ 1 +
3pn
4
− (cid:101)P
2
1 − d
4
+
α
2
+
(2 − Q) = 1.45 − (
Q
10
+
23
2(cid:101)P + d − 3pn
4
),
(2)
j:sj∈[sn−pj sn] max{cj − sn, 0} is the total processing our algorithm does after sn,
≥ 0.07,
10 + 2(cid:101)P+d−3pn
excluding the jobs start strictly after sn. Hence as long as we can show that Q
we have ALG ≤ 1.45 − 0.07 = 1.38, contradicting our initial assumption.
4
where (cid:101)P = (cid:80)
Observe that since (cid:101)P ≥ pn, we always have
2(cid:101)P + d − 3pn
Q
10
+
≥ Q
10
+
d − pn
4
.
4
Observation 6.1 (Bin-packing Constraint) It is impossible to find a set of jobs whose sizes
cannot be packed into two bins of size 1. For example, if we can find five jobs or size larger than
1
3 , or three jobs of size larger than 1
2 , then we have the contradiction that OPT > 1. Similarly, it is
impossible to have three jobs of size at least p > 1
3 , and another job of size larger than 1 − p.
2+α in this section. As it will be convenient for future analysis, we first rule
6.3 Upper Bounding Last Job
We show that pn ≤ 1
out the case when pn > 1+α
2 .
Lemma 6.3 We have pn ≤ 1+α
2 .
Proof: Assume the contrary that pn > 1+α
OPT = 1 (Observation 6.1).
2 . We show that min{pk, pj, pn} > 1
Suppose pk = min{pk, pj, pn} ≤ 1
2 , which contradicts
2 , then we have sk ≤ rn. We show that sj > rn, which implies
pj > pn. Suppose otherwise, then k is the minimum job processed at time rn. As pk < pn
1+α and
rn − sk ≤ 0.5 − 0.38 = 0.12 < αpn, k should have been replaced by n (note that n must be the
largest pending job at rn, if sj ≤ rn).
Hence we know that at sk, both n and j are not released, which gives rk ≤ sk < min{rn, rj}.
As k must be scheduled with one of j, n on the same machine in OPT, we have sk − rk > (1.38 −
pn − pk) − (1 − pk − pn) = 0.38. Hence by Lemma 4.1, the two jobs in J(sk) are of size larger than
min{sk − rk, pk} ≥ 0.38, which by Observation 6.1 also contradicts OPT = 1.
Lemma 6.4 We have pn ≤ 1+α
2 .
Proof: Assume for contrary that pn > 1
2+α . Recall from Lemma 6.1 that under this assumption,
any job of size larger than 1
2+α (including n) cannot be replaced. Hence n is pending from rn to
sn, and is uncharged. Recall that we have pk, pj > 0.38 for k, j ∈ J(sn). We first show that if n is
scheduled with one of k, j on the same machine in OPT, then n is not the one released earlier.
Suppose otherwise, then we have sn − rn ≥ (1.38 − pn) − (1 − pn − 0.38) = 0.76, which implies
min{pk, pj} > min{sn − rn, pn} > pn. Hence both k and j cannot be replaced. Moreover, we have
rn < 1 − 2pn, which gives d > pn. Suppose sk > sj, then k is not charged: either sk (cid:54)= rk, or the
4 ≥ 0.2pn ≥ 0.09, which implies
job replaced by k is rescheduled after sn. Then we have Q
ALG ≤ 1.38, a contradiction.
10 + d−pn
We show that min{pk, pj} < 1
2+α . Suppose otherwise, then k, j, n are all of size lager than 1
2+α .
Hence one of them, suppose k, is scheduled before one of j, n on the same machine in OPT, which
2+α ≤ αpk. Since sk (cid:54)= rk, by Lemma 4.2, the two jobs in J(rk) are both of size
implies rk < 1 − 2
larger than 1
3 . Hence there exist five
jobs of size larger than 1
2+α .
Hence we have rj ≤ sj < min{rk, rn}, which implies sj − rj > (1.38− pn− pj)− (1− pj − pn) = 0.38.
Next we show that sk ≤ rn and sj ≤ rn. Suppose sk > rn, then we have pk > pn and pj < 1
3 . Moreover, we have j /∈ J(rk), as otherwise pj > sn − rk > 2
3 , contradicting OPT = 1.
24
Thus the two jobs in J(sj) (note that k, j, n /∈ J(sn)) are both of size larger than 0.38, which
contradicts OPT = 1. Hence we know that both k and j starts before rn. Moreover, we know that
n is the largest pending job at time rn, which implies min{pk, pj} ≥ min{0.38 + αpn, pn
1+α} = pn
1+α .
We proceed to show that k, j are never replaced.
Observe that by Lemma 4.3 at most one of k, j is ever replaced. Suppose k is replaced and x is the
job completed at sk. Then as we already have four jobs (x, k, j, n) of size larger than 0.38, k can only
be replaced by one of x and j, while the other job is being processed while k is replaced. Hence k can
only be replaced by j, as otherwise j is of size pj > sn− rx > (sn− rn) + (cx− rx) > 0.38 + (1 + α)pn,
which cannot be scheduled on the same machine with any of x, k, n in OPT.
However, if j replaces k (at rj = sj), then we have pj > (1 + α)pk > pn. First observe that j
cannot be schedule with n on the same machine in OPT, as sj = rj < rn, and sj + pj + pn > 1.38.
Hence j must be scheduled with one of k and x on the same machine in OPT.
Figure 2: Case when j replaces k, when x is being processed.
Note that as we already have px > pk > 0.38, and pj > pn > 1
2+α , any other job must have size
strictly smaller than 1 − 0.38 − 1
2+α = 0.166.
If j is scheduled with k, then we have
pn
1 + α
sk(rj) − rk > sj − αpj − rk > (1.38 − pn − pj) − αpj − (1 − pk − pj)
≥0.38 − pn − αpj +
≥ 0.38 − 0.2 × (1 − 0.38) − (1 − 0.2
1.2
× 0.6) = 0.156,
which implies that the job completed at sk(rj) is of size at least min{0.156 + αpk, pk
k is the largest pending job at rk, but not scheduled. Then we have a contradiction.
k is released before n), we have rk ≤ 1 − pk − pn.
1+α} > 0.2, as
If j is scheduled with x in OPT, then we have rx ≤ 1 − px − pj. As k is scheduled with n (and
• If sx (cid:54)= rx, then x is uncharged. As sk(rj) − rk > (1.38 − pn − pj) − αpj − (1 − pk − pn) >
0.38 + pk − (1 + α)pn > 0, we know that k is uncharged. Moreover, there is no idle time after
max{rx, rk} < 1 − pn − 0.38, as at most one job is processed at idle time. Hence d > 0.38,
which implies Q
• If sx = rx (refer to Figure 2), then we have pk+pn ≥ (sn−cx)+pn > 1.38−((1−px−pj)+px) =
10 + d−pn
1+α ) − 1
4 ≥ ( 1
4 )pn + 0.38
4 > 0.095.
10 (1 + 2
0.38 + pj. Hence we have
sk(rj) − rk > (1.38 − pj − pn) − αpj − (1 − 0.38 − pj)
≥ 0.76 − pn − αpj ≥ 0.76 − pn − α(1 − pn
1 + α
≥ 0.76 − (1 − α
1 + α
− α > 0.1,
1 + α
2 + α
)
)
where we use pn ≤ 1+α
larger than min{0.1 + αpk, pk
2+α as otherwise pk > 1
2+α . Hence the job completed at sk(rj) is of size
1+α} > 0.176, which is also a contradiction.
25
sj = rjkpxpjpksx = rxsksnpnALG > 1.38 1+α . Hence we have d = 1 − pn − t ≥ 1−α
Hence we conclude that none of k, j is ever replaced. Assume sk < sj.
If j is scheduled on the same machine before one of k, n in OPT, then j is uncharged, and there
is no idle time after rj ≤ 1 − 2pn
1+α pn. We show that k is also
uncharged: if k is charged, for the job x replaced by k must be rescheduled before sj. Let y be the
job completed at sx. Then we have py > px > pj > 0.38, contradicting OPT = 1. Hence we have
10 + d−pn
Q
Otherwise k is scheduled before one of j, n in OPT, and k is pending from rk to sk. Let
J(sk) = {x, y}. Note that k, j, n /∈ J(sk). If k is scheduled on the same machine with n in OPT,
then sk−rk > (1.38−pn−pk)−(1−pk−pn) > 0.38, which (by Lemma 4.1) implies px, py > 0.38 and
contradicts OPT = 1; if k is scheduled with j, then sk−rk > (1.38−pn−pk)−(1−pk− pn
1+α ) > 0.28.
Hence px, py > 0.28 > max{1 − 2 × 0.38, 1
1+α − 1)pn ≥ 0.083, which implies ALG < 1.38.
4 ( 1−α
2 (1 − pn)}, which also contradicts OPT = 1.
4 ≥ 1
10 (1 + 2
1+α )pn + 1
6.4 Medium Last Job
In this section, we show that if pn > 0.38, then jobs of size larger than 0.38 are "almost irreplaceable"
(Lemma 6.5). We prove the following lemma, which is stronger than the one claimed in Section 6.1.
Different from Section 6.1, here we call a replacement major if the job replaced is of size larger
than 0.38.
Lemma 6.5 If pn ∈ (0.38,
Proof: Suppose there are two major replacements.
1
2+α ], then there is at most one major replacement.
Consider the first major replacement. Let z be the replacer, J(rz) = {x, y} and x be the job
replaced. Then we have py > px > 0.38, and pz > (1 + α)0.38 = 0.456. Let l be the replacer in
the second major replacement. Then we know that the job replaced by l must be one of x, y, z.
Observe that n ∈ {x, y}, since we have min{pz, pl} > 0.456 > 1
2+α , and it is impossible to have five
jobs of size larger than 0.38.
We first show that l and z cannot be scheduled on the same machine in OPT.
Suppose otherwise, then we have rz + pz + pl ≤ 1 (as rz < rl). Hence the job replaced by l is not
z, as otherwise pz + pl > ((1 + α)2 + (1 + α))0.38 > 1; the job replaced by l is not x, as otherwise
y is completed at sx(rl) and n = x, which gives rz + pz + pl > rz + pz + px ≥ ALG > 1.38. Thus y
is replaced by l, which gives pl ≥ (1 + α)py. However, since rz + pz + px + py > 1.38, we have
rz + pz + pl > (rz + pz + px + py) + αpy − px > 1.38 + α · 0.38 − 1 − (1 + α) · 0.38
1 + α
> 1,
which is also a contradiction.
and l, any other job must be of size less than 1 − (2 + α) · 0.38 = 0.164 < 0.38
y are released before z and l, we have rx ≤ 1 − px − min{pz, pl} and ry ≤ 1 − py − min{pz, pl}.
As l and z cannot be scheduled on the same machine in OPT, we know that apart from x, y, z
1+α . Moreover, as x and
Depending on which job is replaced by l, we divide our analysis into three cases.
Case-1: x is replaced.
n = x, which is scheduled after z or l. As rz − sn(rz) ≤ αpz, we have
sn(rz) − rn > (1.38 − pn − pz) − αpz − (1 − pn − min{pz, pl})
If x is replaced by l, then we know that y is completed at sx(rl). Hence
= 0.38 + min{pz, pl} − (1 + α)pz > 2.2 × 0.38 − 1.2 × 0.62 = 0.092.
26
Figure 3: Case when n = x is replaced by l.
Hence we know that n is pending from rn to sn(rz), which means that the job completed at
sn(rz) is of size at least min{0.092 + αpn, pn
Case-2: y is replaced. First note that n (cid:54)= y, as otherwise (similar to Case-1) we have
1+α} ≥ 0.168, contradicting OPT = 1.
sn(rl) − rn > 0.38 + min{pz, pl} − (1 + α) max{pz, pl} > 2.2 × 0.38 − 1.2 × 0.62 = 0.092,
which implies a contradiction. Thus we have n = x. Consider which job is completed at sn:
• if it is z, then by Case-1 we have sn(rz) − rn > 0.092;
• if it is l, then sy(rl) − ry > (1.38 − px − pl) − αpl − (1 − py − min{pz, pl}) > 0.092;
• if it is y, then we know that the job processed on the other machine at sn is either z or l. If it is
z, then we have sn(rz) > 1.38−pn−(1+α)pz; otherwise we have sy(rl) > 1.38−pn−(1+α)pl.
Thus we go to one of the above two cases.
Hence in all cases, we can find a job completed at either sn(rz) or sy(rl) that is of size larger than
0.168, which contradicts OPT = 1.
Case-3: z is replaced.
If z is replaced, then we know that y is not replaced (thus n = x).
Moreover, y must be processed at rl, which gives py > pz. Observe that pl ≥ (1 + α)pz > 1 − pz >
1 − py. Hence in OPT, l cannot be scheduled with z or y. Then we have py ≤ 1 − pz < pl, which
implies cy < cl. Hence we know that z is rescheduled at cy. Then again, we have
sn(rz) − rn > (1.38 − pn − pl) − α(pl + pz) − (1 − pn − pl)
= 0.38 − α(pl + pz) ≤ 0.38 − 0.2 × (0.62 + 0.5) = 0.156,
which implies that the job completed at sn(rz) is of size at least 0.232, contradicting OPT = 1.
Now with the help of Lemma 6.5, we show that we can push the upper bound of pn from 1
2+α
2+α and n is never replaced, then ALG ≤ 1.38.
to 0.38. Depending on whether n is ever replaced, we use different proof strategies.
Lemma 6.6 If 0.38 < pn ≤ 1
Proof: Note that in this case n is uncharged. Recall that we have pk > 0.38 and pj > 0.38.
If n is scheduled before one of k, j in OPT, then we have rn ≤ 1 − pn − 0.38, which implies
sn − rn > 0.76. Hence we have pk > pn, pj > pn and last idle time t < rn < 1 − 2pn. Suppose
sk > sj. As it is impossible to have pk > 0.76 and pj > 0.76, we have sk > rn. We show that k
is uncharged: if the job replaced by k (at rk) is of size less than pn, then it cannot be rescheduled
before sn; otherwise it is a major replacement, which (by Lemma 6.5) implies that k is not replaced.
Hence we have rk = sk, and the job replaced by k is rescheduled before sn, which is impossible, as
j is processed from sk to sn.
27
rnsl = rlnpyplnpzsysz = rzsn(rz)snpnALG > 1.38 10 + d−pn
4 > 0.2pn > 0.076, which gives ALG ≤ 1.38.
Then we have Q
Hence two of k, j, n are scheduled together in OPT, while n is not the one released earlier.
As before, we show that none of k, j is ever replaced.
Suppose k is replaced. Let x be the job completed at sk. We have px > pk, and we know that
k must be replaced by one of x and j, while the other job is being processed when k is replaced.
If k is replaced by x, then either k is uncharged (if sk(rx) > sj), or j is uncharged (if sk(rx) < sj),
because the job replaced must be rescheduled after sn. Moreover, we have ck − sn > px + pk − pj >
(2 + α)0.38 − 0.62 > 0.216, which gives
≥ 1
10
(pn + 0.38) +
1
4
(2 × 0.216 − pn) ≥ 0.38
10
+
0.216
2
− 0.15
2.2
> 0.0778.
after max{rk, rx} < 1 − 0.38 − pn, which gives d > 0.38.
If k is replaced by j, then we have pj > (1 + α)pk > 0.456 > pn. Note that there is no idle time
If sk(rj) < sx, then x is uncharged, and ck − sn > px + pk − (1 + α)pj > 0.016, which implies
2(cid:101)P − 3pn
4
Q
10
+
2(cid:101)P + d − 3pn
4
Q
10
+
≥ 1
10
≥ 0.38
10
(pn + 0.38) +
(2 × 0.016 + 0.38 − pn)
1
4
+
0.412
4
− 0.15
2.2
> 0.0728.
Hence we have sk(rj) > sx, which means that k is uncharged. Observe that if x is also uncharged
then we are done as
d − pn
Q
10
+
4
≥ 1
10
(pn + 0.38 + 0.38) +
1
4
(0.38 − pn) ≥ 0.38
5
+
0.38
4
− 0.15
2.2
> 0.1.
Hence we have rx = sx < sk(rj), and the job replaced by x is completed during (sx, sk(rj)).
Observe that we have px > pk > 0.38 and pj > pn > 0.38, hence any other job must be of size less
than 0.24. Hence we have rk > sx, as otherwise the job replaced by x (which is of size less than
0.24) will be rescheduled after sn.
Since rk < 1 − pk − pn, we have sk(rj) − rk > (1.38 − pn − pj)αpj − (1 − pk − pn) = 0.38 + pk −
(1 + α)pj > 0, which means that k is not scheduled at rk. Hence the job y (apart from x) processed
at rk is processed at least αpk (as 0.24 < pk
1+α ). Moreover, we have sy(rk) > sx, as otherwise x is
uncharged. Hence we have t < sy(rk) < rk − αpk < 1 − pn − 0.456, which gives
− 0.15
2.2
(0.456 − pn) ≥ 0.38
10
(pn + 0.38) +
≥ 1
10
d − pn
> 0.083.
0.456
Q
10
+
1
4
+
4
4
Hence we can assume that none of k, j has been replaced. Assume sk < sj (j is uncharged).
If j is scheduled before one of k, n on the same machine in OPT, then there is no idle time after
rj ≤ 1 − 2 × 0.38 = 0.24, and k is also not charged: if k is charged, then for the job x replaced by
k and the job y completed at sx we have py > px > pj > 0.38, contradicting OPT = 1. Hence
Q
10
+
d − pn
4
≥ 1
10
(pn + 0.76) +
1
4
(0.76 − 2pn) ≥ 0.76
10
+
0.76
4
− 0.4
2.2
> 0.084.
Otherwise k is scheduled before one of j, n in OPT. Observe that sk−rk > 0. Let J(sk) = {x, y}.
Note that k, j, n /∈ J(sk). If k is scheduled with n, then sk−rk > (1.38−pn−pk)−(1−pk−pn) > 0.38,
which (by Lemma 4.1) implies min{px, py} > 0.38 and contradicts OPT = 1; if k is scheduled
with j, then sk − rk > (1.38 − pn − pk) − (1 − pk − 0.38) = 0.76 − pn. Hence min{px, py} >
min{1 − 2 × 0.38, 1
2 (1 − pn)}, which also contradicts OPT = 1.
28
2+α and n is ever replaced, then ALG ≤ 1.38.
Lemma 6.7 If 0.38 < pn ≤ 1
Proof: Observe that since pn > 0.38 and n is ever replaced, by Lemma 6.5 we know that n is
replaced exactly once, and any other job of size larger than 0.38 is never replaced.
Suppose n is replaced by l. Let J(rl) = {n, x}. Then we must have either l ∈ J(sn) or x ∈ J(sn),
We first consider the case when J(sn) = {l, x}. Note that l, x are never replaced.
as otherwise we have five jobs of size larger than 0.38.
1. If sx > sn(rl), then we have rx > sn(rl), as x cannot be pending at sn(rl). Consider the
instance with jobs released after sn(rl) removed. Let OPT(cid:48) be the new optimal makespan and
ALG(cid:48) be the makespan of our algorithm on the new instance. We have OPT(cid:48) ≤ 1− min{px, pl}
(as x, l are released after sn(rl) and cannot be scheduled on the same machine in OPT), while
ALG(cid:48) = sn(rl) + pn = ALG − (sn − sn(rl)) ≥ ALG − 1.38 · min{px, pl}, which gives a smaller
counter-example (the inequality holds since sn − rl ≤ min{px, pl} and rl − sn(rl) ≤ αpl ≤
α(px + pn − 0.38) < 0.38px).
2. If rx ≤ sx < sn(rl), then we compare the release times of n and x. Observe that rl >
If rn < rx, then t < rn < 1 − 2pn. We show that x is uncharged: any job
max{rn, rx}.
y replaced by x at rx < rn cannot be scheduled before sn, as there is at most one major
4 > 0.2 × 0.38 =
replacement. As sn(rl) > sx > rn, n is uncharged, which implies Q
0.076. If rx < rn, then n is uncharged, as any job replaced by n can only be scheduled after
one of x, n is completed.
10 + d−pn
• If x is scheduled together with one of n, l in OPT, we have sx − rx > (1.38 − px −
pn) − (1 − px − pn) = 0.38, which means that the two jobs processed at sx (note that
n, x, l /∈ J(sx)) are of size larger than 0.38, contradicting OPT = 1;
• otherwise we have rx < rn < 1− pn − pl, which implies d > pl. If x is uncharged then we
4 > 0.095; otherwise we have px > 1.38 − pn − rx > 0.38 + pl. Moreover,
have Q
we have sn(rl) − rn > (1.38 − pn − pl) − αpl − (1 − pn − pl) > 0.092, which means that
the job completed at sn(rl) is of size larger than 0.168. Then we have a contradiction as
min{0.168 + px, 0.168 + pn + pl} > 0.168 + (2 + α)0.38 > 1.
10 + d−pn
Hence we have J(sn) (cid:54)= {l, x}, which means that at least three of the four jobs in {n, x, l}∪J(sn)
We first consider the case when l /∈ J(sn) and x ∈ J(sn). Suppose x = j, then we have
are released after t (as t < rl), which implies t < 1 − 2pn and d > pn.
ck − sn > pl + pk − px > (2 + α)0.38 − 0.62 > 0.216. Hence we have
2(cid:101)P + d − 3pn
4
Q
10
+
≥ pn
10
+
2 × 0.216
4
> 0.146.
Next we consider the case when l ∈ J(sn) and x /∈ J(sn). Suppose l = k. Observe that at any
time from max{sn(rl), sx} to sn, the minimum job being processed is of size at least pn. Hence any
job replaced at or after max{sn(rl), sx} must be rescheduled after sn. Thus job j and one of n, x
are uncharged. Then we have Q
5 > 0.2 × 0.38 = 0.076.
10 + d−pn
4 ≥ pn
6.5 Refined Efficiency Argument
Observe that the upper bound (2) on ALG is quite loose when pn and d are very small. Actually,
since there are only two machines, if time t is idle, then there is one job being processed (it is
impossible to have two idle machines in the minimum counter-example). Hence we should have a
better upper bound on ∆t, compared to Lemma 1.1 (The Leftover Lemma).
29
Let t be the last idle time before sn, and let J(t) = {i}. Let OPTt be the makespan of the
optimal schedule of the jobs released before t. Then we have ci(t) := si(t) + pi ≤ 1.38 · OPTt, as
otherwise we can remove all jobs released after t, and obtain a smaller counter-example. Define
p := max{OPTt−t, 0}. Then we have p ≤ OPTt ≤ t+p. Note that by optimality of OPTt, the total
j:rj≥t pj. Hence the total size of jobs released
processing OPT does after time t is at least p +(cid:80)
after time t is(cid:80)
j:rj≥t pj ≤ 2(1 − t) − p.
Claim 6.1 There exists a job of size at least pn released after t that does not replace other jobs at
its release time.
Proof:
Suppose otherwise, then every job of size at least pn released after t must be scheduled
immediately (but can possibly be replaced later). If there is any job of size px < pn released after
t, then we argue that the instance with x removed is a smaller counter-example: the schedule
produced by our algorithm on the new instance is identical to the original scheduled (projected on
jobs of size at least pn), as the behavior of every job of size at least pn is unchanged.
Hence in the minimum counter-example, all jobs released after t are of size at least pn. Then
the first job released after t does not replace any job, as there is an idle machine.
By Observation 3.1 we have (recall that P is the total processing our algorithm does after sn)
2(1 − t) ≥ 2(sn − t) + P − ∆t(t) − (W1 − Wt(t)).
(3)
Note that W1 − Wt(t) is the total waste created by jobs released after time t, which is at most
α(2 − 2t − p − pn). Rearranging the inequality and by P ≥ pn, we have
ALG = sn + pn ≤ 1.2 + 0.4pn + 0.5∆t(t) − 0.2t − 0.1p.
Note that we have ∆t(t) ≤ ci(t) − t − p ≤ 0.38 · OPTt ≤ 0.38(t + p). Applying the upper bound
on ∆t(t), we have
ALG = sn + pn ≤ 1.2 + 0.4pn + 0.09(t + p) − 0.1t.
(4)
Observe that we have t + p ≤ 1 and the following lower bound on t (from (1)):
which implies t > 4(0.38 − pn
2 (2 − pn)) = 0.72 − 1.6pn. Hence we have
1.38 < ALG ≤ 1 +
2 − α
pn
2
+
t
4
+
α
2
(2 − pn),
ALG = sn + pn ≤ 1.2 + 0.4pn + 0.09 − 0.1(0.72 − 1.6pn) = 1.218 + 0.56pn.
Thus immediately we can show that pn cannot be too small, as otherwise we have the contra-
diction that ALG ≤ 1.218 + 0.56 × 0.28 = 1.3748.
Lemma 6.8 (Lower Bound on pn) We have pn > 0.28.
We show the following lemma, which will be the main framework towards deriving a contradic-
tion, given that 0.28 < pn ≤ 0.38.
Lemma 6.9 Given that pn ≤ 0.38, if we have W1 − Wt(t) − (P − pn) ≤ α(2 − 2t − p − 2pn), then
we can show that ALG ≤ 1.38.
30
Proof: Applying the upper bounds on W1 − Wt(t) − (P − pn) and ∆t(t) to (3), we obtain the
following stronger version of (4).
ALG = sn + pn ≤ 1
2
(2 − pn + 0.38(t + p) + α(2 − 2t − p − 2pn)) + pn
= 1.2 + 0.3pn + 0.09(t + p) − 0.1t.
Since t + p ≤ 1 (we do not use the upper bound given by Lemma 6.10), we have
ALG = sn + pn ≤ 1.2 + 0.3pn + 0.09 − 0.1(0.72 − 1.2pn)
= 1.218 + 0.42pn ≤ 1.218 + 0.42 × 0.38 ≤ 1.3776,
where in the first inequality we use a stronger lower bound on t: t > 4(0.38 − pn
0.72 − 1.2pn, which holds only when W1 − Wt(t) − (P − pn) ≤ α(2 − 2t − p − 2pn).
2 − α
2 (2 − 2pn)) =
Notice that the "if" condition of Lemma 6.9 holds if there exists two jobs of size at least pn
Next we show an upper bound on OPTt. Note that if p (cid:54)= 0, then we have t + p = OPTt. Hence
released after t that are uncharged.
the upper bound holds for t + p when p (cid:54)= 0.
Lemma 6.10 We have OPTt ≤ 1 − pn.
Suppose OPTt > 1 − pn, then we know that any job of size at least pn released after
Proof:
time t must be scheduled on the same machine in OPT (otherwise OPTt is not optimal). Also by
optimality of OPTt, we have ci(t) ≥ OPTt.
Recall that we have min{pk, pj, pn} ≥ pn. Since only one job is processed at time t, we know
that at least two of k, j, n are released after n. Suppose l (cid:54)= n is the largest job released after time
t, then we have pi > ci(t)− t > (1− pn)− (1− pn− pl) = pl, which means that pi cannot be replaced
after t, and ci(t) = ci. Observe that if there are two jobs of size at least pn released after t that are
uncharged, then by Lemma 6.9 we have ALG ≤ 1.38.
Next we prove the existence of uncharged jobs.
If i ∈ J(sn), suppose i = k, then we have ci = ck > sn > 1. Note that n is uncharged:
any job replaced by n must be rescheduled after sn. We show that j is also uncharged. Suppose
otherwise, then the job replaced by j must be n, and j must also be replaced, as cj > 1. Hence
we have pj ≥ (1 + α)pn and the replacer of j is of size at least (1 + α)2pn, which is impossible, as
pn + (1 + α)pn + (1 + α)2pn > 0.28 × (1 + 1.2 + 1.44) > 1.
If i /∈ J(sn), then k, j, n are all released after time t. Note that it is impossible to have four
jobs of size at least pn released after t, as 4pn > 1. Observe that none of k, j or n can be replaced.
Suppose otherwise, let l be the replacer and M be the machine where the replacement happens.
Then the first job completed on M after rl is a job of size at least (1 + α)pn that is not k, j or n.
Hence all of k, j and n are never replaced, thus uncharged (as sn > 1).
Lemma 6.10 helps us to improve the lower bound on pn.
Corollary 6.1 (Improved Lower Bound) We have pn > 1
3 .
Proof: Assume for contrary that pn ≤ 1
otherwise by Lemma 6.10 we have t + p = OPTt ≤ 1 − pn. Hence we have
3 . If p = 0, then by (4), we have ALG ≤ 1.2 + 0.4pn ≤ 1.34;
ALG = sn + pn ≤ 1.2 + 0.4pn + 0.09(1 − pn) − 0.1(0.72 − 1.6pn)
= 1.218 + 0.47pn ≤ 1.218 + 0.47 × 1
3
≤ 1.3747,
31
where in the first inequality we use t > 0.72 − 1.6pn.
It remains to prove the following lemma. Recall that so far we have shown that pn ∈ ( 1
3 , 0.38].
3 , 0.38].
Lemma 6.11 We have W1 − Wt(t) − (P − pn) ≤ α(2 − 2t − p − 2pn), given that pn ∈ ( 1
Proof:
If n is never replaced, i.e., is pending from rn to sn, then n is uncharged. Note that
n (cid:54)= i. Hence n is released after t. We show that at least one of k, j is uncharged. Note that
it is impossible that both k, j are replaced. Hence one of them, suppose k, is not replaced, thus
uncharged. Moreover, k must be released after t: otherwise k = i, and we have the contradiction
that ci(t) ≥ sn = ALG − pn > 1.38(1 − pn) ≥ 1.38 · OPTt.
Otherwise let rl be the last time n is replaced, and J(rl) = {n, x}. Then we have px > pn.
We show that both k, j are never replaced. Suppose the contrary that k is replaced. Then k
cannot be replaced by n, l or j (otherwise j must also be replaced). As there cannot exist five jobs
of size larger than 1
3 , we must have k = l and x = j, i.e., k replaces n while j is being processed, and
then k is replaced by some job y. If ck − sn ≥ 2αpn, then we already have W1 − Wt(t)− (P− pn) ≤
α(2 − 2t − p − 2pn); otherwise pj > py + pk − (ck − sn) ≥ (2 + α)(1 + α)pn − 2αpn > 2
3 , which
also contradicts OPT = 1. Hence both k, j are not charged. Moreover, for the same reason argued
above, both k, j are released after t.
7 Other Candidate Algorithms
All the candidate algorithms are based on LPT. That is, whenever there is an idle machine, we
always schedule the largest job. The only difference is the replacement rule. We will show that
none of the them can beat the ratio of 1.5.
Candidate Algorithm 1. Fix any constant 0 < ρ < 1. Upon the arrival of a job j, job k can
be replaced by job j if k is the smallest processing job, pk < pj and job k has been processed no
larger than ρ fraction.
Counter example. At t = 0, m identical jobs come with p1 = p2 = ··· = pm = 1. Each of them
is scheduled on a machine. At t = ρ, job (m + 1) comes with pm+1 = 1 + ξ (ξ is an infinitesimal
amount). Then one of jobs 1 to m is replaced by job (m + 1). At t = 2ρ, job (m + 2) comes with
pm+2 = 1 + 2ξ, then job (m + 1) is replaced by job (m + 2). The same thing goes on and on, and
at time t = 1, job (m − 1 + (cid:100) 1
ρ(cid:101)) jobs
ρ(cid:101)) is replaced by job (m + (cid:100) 1
ρ(cid:101)+2 = ··· = p2m = 1. Then there are m pending jobs but only (m− 1)
come with pm+(cid:100) 1
ρ(cid:101)+1 = pm+(cid:100) 1
idle machines, so ALG = 3. In the optimal solution, no replacement happens, and OPT = 2 +(cid:100) 1
ρ(cid:101)ξ.
ρ(cid:101)). After then, at t = 1, (m − (cid:100) 1
Candidate Algorithm 2. Fix 0 < ρ < 1 and 1 < µ < 2. Upon the arrival of a job j, job k can
be replaced by job j if µpk < pj and job k has been processed no larger than ρ fraction.
We show a counter example using µ = 3/2 and ρ = 1/2. This counter example can be generalized
to any µ and ρ such that µ + ρ ≤ 2.
Counter example. At t = 0, m jobs come, with p1 = m + 1, p2 = m + 2,··· , pm = 2m. At
time t = m, job (m + 1) comes with pm+1 = 3m. Then, as job m is the only job that is processed
at most half, job m is replaced by job (m + 1). After the replacement, (m − 1) jobs come with
pm+2 = 2m + 1, pm+3 = 2m + 2,··· , p2m = 3m− 1. Then ALG = 6m, while in the optimal solution,
no replacement happens, thus OPT = 4m + 1.
32
Candidate Algorithm 3. Fix a target performance ratio 1 + γ. When a job j comes, schedule
it virtually to ALG, and calculate the current optimal solution with all the jobs that have been
released. If the ratio can still be bounded in 1 + γ, do not replace any jobs; otherwise choose one
job to replace.
Counter example. At t = 0, m jobs comes first, with p1 = 2m, p2 = 2m + 1,··· , pm = 3m − 1.
After each of them has been scheduled on a machine, at t = 0, another m jobs come, with pm+1 =
3m, pm+2 = 3m + 1,··· , p2m = 4m− 1. At this time, since the local ALG and local OPT are exactly
the same, no replacement happens. Then at t = 3m − 1, another job comes with p2m+1 = 3m. So
for this instance, ALG = 6m− 1; while in the optimal solution, three smallest jobs (jobs 1, 2, 3) are
scheduled on the same machine, while all other jobs are paired up with the smallest with largest,
and OPT = 4m + 3.
For our algorithm LPT with Restart, some may wonder what happens if α or β is not in (0, 1/2).
we know that if α = 0, it is exactly the same with LPT, which cannot beat 1.5; if β = 0, a counter
example can be given that is similar to that of Candidate Algorithm 1. Next we show that when
β ≥ 1/2 or α ≥ 1/2, LPT with Restart could not beat 1.5, no matter what the value of the other
parameter is.
Candidate Algorithm 4 In LPT with Restart, set β ≥ 1/2, and α to be any constant.
Counter example. At t = 0, m(m ≥ 4) jobs come first, with p1 = p2 = ··· = pm = 1. After all
these jobs are scheduled, still at time 0, another m jobs come, with pm+1 = pm+2 = ··· = p2m =
3/2+ξ. Then at t = 1, another job comes with p2m+1 = 2. Since β ≥ 1/2, no replacement happens,
and ALG = 9/2 + ξ, while in optimal schedule, all jobs could be completed at or before time 3 + ξ.
Candidate Algorithm 5 In LPT with Restart, set α ≥ 1/2, and β to be any constant such that
0 < β < 1/2.
Counter example. We consider the special case with only one machine. At t = 0, job 1 comes
with p1 = 1. At t = 1− ξ (again, ξ is an infinitesimal amount), job 2 comes with p2 = 2, then job 1
is replaced by job 2. At t = 3 − 2ξ, job 3 comes with p3 = 4, job 2 is replaced by job 3. The same
thing goes on and on. Each time a new job comes, ALG would replace the previous job, while OPT
would wait for the previous job to end. The final ratio would be arbitrarily close to 1.5.
8 Hardness for deterministic algorithms with restart
1.5 ≈ 1.225 for any deterministic algorithms
In this section, we present a simple lower bound of
with restart. Given any deterministic algorithm, consider the following instance with two machines.
6, another job comes with
At t = 0, two jobs come with size p1 = p2 = 1. Then, at t = 3 − √
√
√
6 − 1.
size p3 =
1. If the algorithm starts processing job 3 after time 1, i.e., after completing the two jobs of size
√
1, no more jobs arrive in the instance. We have OPT = 2 as we could have scheduled job 1 and
job 2 on the same machine and job 3 on the other one. On the other hand, ALG ≥ 1+p3 =
6.
√
2. If the algorithm starts processing job 3 before time 1, e.g., it restarts one of the size-1 jobs,
6
let there be a fourth job that arrives at time 1 with size p4 =
by scheduling job 1 and 3 on one machine, and 2 and 4 on the other. On the other hand,
we have ALG ≥ 3 since at time 1 at least one of jobs 1 and 2 is pending, and job 3 does not
complete until time 2.
6 − 1. We have OPT =
√
33
References
Susanne Albers. Better bounds for online scheduling. SIAM Journal on Computing, 29(2):459–473,
1999.
Nir Avrahami and Yossi Azar. Minimizing total flow time and total completion time with immediate
dispatching. Algorithmica, 47(3):253–268, 2007.
Y. Bartal, A. Fiat, H. Karloff, and R. Vohra. New algorithms for an ancient scheduling problem.
Journal of Computer and System Sciences, 51(3):359 – 366, 1995.
Piotr Berman, Moses Charikar, and Marek Karpinski. On-line load balancing for related machines.
Journal of Algorithms, 35(1):108–121, 2000.
Bo Chen and Arjen P. A. Vestjens. Scheduling on identical machines: How good is LPT in an
on-line setting? Operations Research Letters, 21(4):165–169, 1997.
Bo Chen, Andr van Vliet, and Gerhard J. Woeginger. New lower and upper bounds for on-line
scheduling. Operations Research Letters, 16(4):221 – 230, 1994a.
Bo Chen, Andr´e van Vliet, and Gerhard J. Woeginger. A lower bound for randomized on-line
scheduling algorithms. Information Processing Letters, 51(5):219–222, 1994b.
Bo Chen, Andr´e van Vliet, and Gerhard J. Woeginger. An optimal algorithm for preemptive on-line
scheduling. Operations Research Letters, 18(3):127–131, 1995.
Marek Chrobak, Wojciech Jawor, Jir´ı Sgall, and Tom´as Tich´y. Online scheduling of equal-length
jobs: Randomization and restarts help. SIAM Journal on Computing, 36(6):1709–1728, 2007.
Gyorgy D´osa and Leah Epstein. Online scheduling with a buffer on related machines. Journal of
Combinatorial Optimization, 20(2):161–179, 2010.
Gyorgy D´osa and Leah Epstein. Preemptive online scheduling with reordering. SIAM Journal on
Discrete Mathematics, 25(1):21–49, 2011.
Tom´as Ebenlendr, Wojciech Jawor, and Jir´ı Sgall. Preemptive online scheduling: Optimal algo-
rithms for all speeds. Algorithmica, 53(4):504–522, 2009.
Matthias Englert, Deniz Ozmen, and Matthias Westermann. The power of reordering for online
minimum makespan scheduling. SIAM Journal on Computing, 43(3):1220–1237, 2014.
Leah Epstein and Jir´ı Sgall. A lower bound for on-line scheduling on uniformly related machines.
Operations Research Letters, 26(1):17–22, 2000.
Leah Epstein, John Noga, Steven S. Seiden, Jir´ı Sgall, and Gerhard J. Woeginger. Randomized
online scheduling on two uniform machines. In SODA, pages 317–326. ACM/SIAM, 1999.
Rudolf Fleischer and Michaela Wahl. On-line scheduling revisited. Journal of Scheduling, 3(6):
343–353, 2000.
Ronald L. Graham. Bounds on multiprocessing timing anomalies. SIAM Journal of Applied Math-
ematics, 17(2):416–429, 1969.
34
Han Hoogeveen, Chris N Potts, and Gerhard J Woeginger. On-line scheduling on a single machine:
maximizing the number of early jobs. Operations Research Letters, 27(5):193–197, 2000.
David R. Karger, Steven J. Phillips, and Eric Torng. A better algorithm for an ancient scheduling
problem. Journal of Algorithms, 20(2):400 – 430, 1996.
Hans Kellerer, Vladimir Kotov, Maria Grazia Speranza, and Zsolt Tuza. Semi on-line algorithms
for the partition problem. Operations Research Letters, 21(5):235 – 242, 1997.
Shisheng Li, Yinghua Zhou, Guangzhong Sun, and Guoliang Chen. Study on parallel machine
scheduling problem with buffer. In IMSCCS, pages 278–273. IEEE Computer Society, 2007.
John Noga and Steven S. Seiden. An optimal online algorithm for scheduling two machines with
release times. Theoretical Computer Science, 268(1):133–143, 2001.
J. F. RudinIII. Improved Bound for the Online Scheduling Problem. PhD thesis, University of
Texas at Dallas, 2001.
J. F. RudinIII and R. Chandrasekaran. Improved bounds for the online scheduling problem. SIAM
Journal on Computing, 32(3):717–735, 2003.
Jianjun Wen and Donglei Du. Preemptive on-line scheduling for two uniform processors. Operations
Research Letters, 23(3-5):113–116, 1998.
Guochuan Zhang. A simple semi on-line algorithm for p2//c {max} with a buffer. Information
Processing Letters, 61(3):145–148, 1997.
35
|
1104.2230 | 1 | 1104 | 2011-04-12T14:39:13 | Subexponential Parameterized Algorithm for Minimum Fill-in | [
"cs.DS"
] | The Minimum Fill-in problem is to decide if a graph can be triangulated by adding at most k edges. Kaplan, Shamir, and Tarjan [FOCS 1994] have shown that the problem is solvable in time O(2^(O(k)) + k2 * nm) on graphs with n vertices and m edges and thus is fixed parameter tractable. Here, we give the first subexponential parameterized algorithm solving Minimum Fill-in in time O(2^(O(\sqrt{k} log k)) + k2 * nm). This substantially lower the complexity of the problem. Techniques developed for Minimum Fill-in can be used to obtain subexponential parameterized algorithms for several related problems including Minimum Chain Completion, Chordal Graph Sandwich, and Triangulating Colored Graph. | cs.DS | cs |
Subexponential Parameterized Algorithm for
Minimum Fill-in ∗
Fedor V. Fomin †
Yngve Villanger †
Abstract
The Minimum Fill-in problem is to decide if a graph can be triangu-
lated by adding at most k edges. Kaplan, Shamir, and Tarjan [FOCS 1994]
have shown that the problem is solvable in time O(2O(k)+k2nm) on graphs
with n vertices and m edges and thus is fixed parameter tractable. Here,
we give the first subexponential parameterized algorithm solving Mini-
mum Fill-in in time O(2O(√k log k) + k2nm). This substantially lower the
complexity of the problem. Techniques developed for Minimum Fill-in
can be used to obtain subexponential parameterized algorithms for sev-
eral related problems including Minimum Chain Completion, Chordal
Graph Sandwich, and Triangulating Colored Graph.
1 Introduction
A graph is chordal (or triangulated) if every cycle of length at least four con-
tains a chord, i.e. an edge between nonadjacent vertices of the cycle. The
Minimum Fill-in problem (also known as Minimum Triangulation and
Chordal Graph Completion) is
Minimum Fill-in
Input: A graph G = (V, E) and a non-negative integer k.
Question: Is there F ⊆ [V ]2, F ≤ k, such that graph H = (V, E ∪ F ) is chordal?
The name fill-in is due to the fundamental problem arising in sparse matrix
computations which was studied intensively in the past. During Gaussian elim-
inations of large sparse matrices new non-zero elements called fill can replace
original zeros thus increasing storage requirements and running time needed to
solve the system. The problem of finding the right elimination ordering mini-
mizing the amount of fill elements can be expressed as the Minimum Fill-in
problem on graphs [45, 47]. See also [15, Chapter 7] for a more recent overview
∗This research was partially supported by the Research Council of Norway.
†Department
Informatics,
Bergen,
University
of
of
Bergen,
Norway,
{fedor.fomin,yngve.villanger}@ii.uib.no
1
of related problems and techniques. Besides sparse matrix computations, appli-
cations of Minimum Fill-in can be found in database management [2], artificial
intelligence, and the theory of Bayesian statistics [13, 27, 40, 51]. The survey
of Heggernes [30] gives an overview of techniques and applications of minimum
and minimal triangulations.
Minimum Fill-in (under the name Chordal Graph Completion) was
one of the 12 open problems presented at the end of the first edition of Garey
and Johnson's book [26] and it was proved to be NP-complete by Yannakakis
[52]. Kaplan et al. proved that Minimum Fill-in is fixed parameter tractable
by giving an algorithm of running time O(m16k) in [36] and improved the
running time to O(k616k + k2mn) in [37], where m is the number of edges and
n is the number of vertices of the input graph. There were several algorithmic
improvements resulting in decreasing the constant in the base of the exponents.
In 1996, Cai [11], reduced the running time to O((n + m) 4k
k+1 ). The fastest
parameterized algorithm known prior to our work is the recent algorithm of
Bodlaender et al. with running time O(2.36k + k2mn) [4].
In this paper we give the first subexponential parameterized algorithm for
Minimum Fill-in. The last chapter of Flum and Grohe's book [21, Chap-
ter 16] concerns subexponential fixed parameter tractability, the complexity
class SUBEPT, which, loosely speaking -- we skip here some technical conditions --
is the class of problems solvable in time 2o(k)nO(1), where n is the input length
and k is the parameter. Subexponential fixed parameter tractability is inti-
mately linked with the theory of exact exponential algorithms for hard prob-
lems, which are better than the trivial exhaustive search, though still exponen-
tial [22]. Based on the fundamental results of Impagliazzo et al. [33], Flum and
Grohe established that most of the natural parameterized problems are not in
SUBEPT unless Exponential Time Hypothesis (ETH) fails. Until recently, the
only notable exceptions of problems in SUBEPT were the problems on planar
graphs, and more generally, on graphs excluding some fixed graph as a minor
[16]. In 2009, Alon et al.
[1] used a novel application of color coding to show
that parameterized Feedback Arc Set in Tournaments is in SUPEPT.
Minimum Fill-in is the first problem on general graphs which appeared to be
in SUBEPT.
General overview of our approach. Our main tool in obtaining subexpo-
nential algorithms is the theory of minimal triangulations and potential maximal
cliques of Bouchitt´e and Todinca [8]. This theory was developed in contest of
computing the treewidth of special graph classes and was used later in exact
exponential algorithms [23, 24]. A set of vertices Ω of a graph G is a potential
maximal clique if there is a minimal triangulation such that Ω is a maximal
clique in this triangulation. Let Π be the set of all potential maximal cliques in
graph G. The importance of potential maximal cliques is that if we are given
the set Π, then by using the machinery from [8, 23], it is possible compute an
optimum triangulation in time up to polynomial factor proportional to Π.
If (G, k) is a YES
instance of the Minimum Fill-in problem, then every maximal clique of every
Let G be an n-vertex graph and k be the parameter.
2
optimum triangulation is obtained from some potential maximal clique of G by
adding at most k fill edges. We call such potential maximal clique vital. To
give a general overview of our algorithm, we start with the approach that does
not work directly, and then explain what has to be changed to suceed. The
algorithm consists of three main steps.
Step A. Apply a kernelization algorithm that in time nO(1) reduces the problem
an instance to instance of size polynomial in k;
Step B. Enumerate all vital potential maximal cliques of an n-vertex graph in
time no(k/ log k). By Step A, n = kO(1), and thus the running time of
enumeration algorithm and the number of vital potential maximal cliques
is 2o(k);
Step C. Apply the theory of potential maximal clique to solve the problem in time
proportional to the number of vital potential maximal cliques, which is
2o(k).
Step A, kernelization for Minimum Fill-in, was known prior to our work.
In 1994, Kaplan et al. gave a kernel with O(k5) vertices. Later the kernelization
was improved to O(k3) in[37] and then to 2k2 + 4k in [43]. Step C, with some
modifications, is similar to the algorithm from [8, 23]. This is Step B which does
not work and instead of enumerating vital potential maximal cliques we make
a "detour". We use branching (recursive) algorithm that in subexponential
time outputs subexponential number of graphs avoiding a specific combinato-
rial structure, non-reducible graphs.
In non-reducible graphs we are able to
enumerate vital potential maximal cliques. Thus Step B is replaced with
Step B1. Apply branching algorithm to generate nO(√k) non-reducible instances
such that the original instance is a YES instance if and only if at least one
of the generated non-reducible instances is a YES instance;
Step B2. Show that if G is non-reducible, then all vital potential maximal cliques
of G can be enumerated in time nO(√k).
Putting together Steps A, B1, B2, and C, we obtain the subexponential algo-
rithm.
It follows from our results that several other problems belong to SUBEPT.
We show that within time O(2O(√k log k) +k2nm) it is possible to solve Minimum
Chain Completion, and Triangulating Colored Graph. For Chordal
Graph Sandwich, and we show that deciding if a sandwiched chordal graph
G can be obtained from G1 by adding at most k fill edges, is possible in time
O(2O(√k log k) + k2nm).
A chain graph is a bipartite graph where the sets of neighbors of vertices
form an inclusion chain. In the Minimum Chain Completion problem, we
are asked if a bipartite graph can be turned into a chain graph by adding at
most k edges. The problem was introduced by Golumbic [28] and Yannakakis
3
[52]. The concept of chain graph has surprising applications in ecology [41, 46].
Feder et al. in [20] gave approximation algorithms for this problem.
The Triangulating Colored Graph problem is a generalization of Min-
imum Fill-in. The instance is a graph with some of its vertices colored; the
task is to add at most k fill edges such that the resulting graph is chordal and no
fill edge is monochromatic. We postpone the formal definition of the problem
till Section 7. The problem was studied intensively because of its close relation
to Perfect Phylogeny Problem -- fundamental and long-standing problem
for numerical taxonomists [7, 10, 35]. The Triangulating Colored Graph
problem is N P -complete [6] and W [t]-hard for any t, when parametrized by the
number of colours [5]. However, the problem is fixed parameter tractable when
parameterized by the number of fill edges.
In Chordal Graph Sandwich we are given two graphs G1 and G2 on the
same vertex set, and the question is if there is a chordal graph G which is a
supergraph of G1 and a subgraph of G2. The problem is a generalization of
Triangulating Colored Graph. We refer to the paper of Golumbic et al.
[29] for a general overview of graph sandwich problems.
The remaining part of the paper is organized as follows. Section 2 contains
definitions and preliminary results. In Section 3, we give branching algorithm,
Step B1. Section 4 provides algorithm enumerating vital potential maximal
cliques in non-reducible graphs, Step B2. This is the most important part of
our algorithm. It is based on a new insight into the combinatorial structure of
potential maximal cliques. In Section 5, we show how to adapt the algorithm
from [8, 23] to implement Step C. The main algorithm is given in Section 6.
In Section 7, we show how the ideas used for Minimum Fill-in can be used
to obtain subexponential algorithms for other problems. To implement our
strategy for Chordal Graph Sandwich, we have to provide a polynomial
kernel for this problem. We conclude with open problems in Section 8.
2 Preliminaries
We denote by G = (V, E) a finite, undirected and simple graph with vertex set
V (G) = V and edges set E(G) = E. We also use n to denote the number of
vertices and m the number of edges in G. For a nonempty subset W ⊆ V , the
subgraph of G induced by W is denoted by G[W ]. We say that a vertex set
W ⊆ V is connected if G[W ] is connected. The open neighborhood of a vertex v is
N (v) = {u ∈ V : uv ∈ E} and the closed neighborhood is N [v] = N (v)∪{v}. For
a vertex set W ⊆ V we put N (W ) = Sv∈W N (v) \ W and N [W ] = N (W )∪ W .
Also for W ⊂ V we define fillG(W ), or simple fill(W ), to be the number of
non-edges of W , i.e. the number of pairs u 6= v ∈ W such that uv 6∈ E(G). We
use GW to denote the graph obtained from graph G by completing its vertex
subset W into a clique. We refer to Diestel's book [17] for basic definitions from
Graph Theory.
Chordal graphs and minimal triangulations. Chordal or triangulated
4
graphs is the class of graphs containing no induced cycles of length more than
three. In other words, every cycle of length at least four in a chordal graph con-
tains a chord. Graph H = (V, E ∪ F ) is said to be a triangulation of G = (V, E)
if H is chordal. The triangulation H is called minimal if H′ = (V, E ∪ F ′) is
not chordal for every edge subset F ′ ⊂ F and H is a minimum triangulation
if H′ = (V, E ∪ F ′) is not chordal for every edge set F ′ such that F ′ < F.
The edge set F for the chordal graph H is called the fill of H, and if H is a
minimum triangulation of G, then F is the minimum fill-in for G.
Minimal triangulations can be described in terms of vertex eliminations (also
known as Elimination Game) [45, 25]. Vertex elimination procedure takes as
input a vertex ordering π : {1, 2, . . . , n} → V (G) of graph G and outputs a
chordal graph H = Hn. We put H0 = G and define Hi to be the graph obtained
from Hi−1 by completing all neighbours v of π(i) in Hi−1 with π−1(v) > i into a
clique. An elimination ordering π is called minimal if the corresponding vertex
elimination procedure outputs a minimal triangulation of G.
Proposition 2.1 ([44]). Graph H is a minimal triangulation of G if and only if
there exists a minimal elimination ordering π of G such that the corresponding
procedure outputs H.
We will also need the following description of the fill edges introduced by
vertex eleminations.
Proposition 2.2 ([48]). Let H be the chordal graph produced by vertex elim-
ination of graph G according to ordering π. Then uv 6∈ E(G) is a fill edge of
H if and only if there exists a path P = uw1w2 . . . wℓv such that π−1(wi) <
min(π−1(u), π−1(v)) for each 1 ≤ i ≤ ℓ.
Minimal separators. Let u and v be two non adjacent vertices of a graph G.
A set of vertices S ⊆ V is an u, v-separator if u and v are in different connected
components of the graph G[V \ S]. We say that S is a minimal u, v-separator of
G if no proper subset of S is an u, v-separator and that S is a minimal separator
of G if there are two vertices u and v such that S is a minimal u, v-separator.
Notice that a minimal separator can be contained in another one. If a minimal
separator is a clique, we refer to it as to a clique minimal separator. A connected
component C of G[V \ S] is a full component associated to S if N (C) = S. The
following proposition is an exercise in [28].
Proposition 2.3 (Folklore). A set S of vertices of G is a minimal a, b-separator
if and only if a and b are in different full components associated to S.
In
particular, S is a minimal separator if and only if there are at least two distinct
full components associated to S.
Potential Maximal Cliques are combinatorial objects which properties are
crucial for our algorithm. A vertex set Ω is defined as a potential maximal
clique in graph G if there is some minimal triangulation H of G such that Ω
is a maximal clique of H. Potential maximal cliques were defined by Bouchitt´e
and Todinca in [8, 9].
5
The following proposition was proved by Kloks et al. for minimal separators
[38] and by Bouchitt´e and Todinca for potential maximal cliques [8].
Proposition 2.4 ([8, 38]). Let X be either a potential maximal clique or a
minimal separator of G, and let GX be the graph obtained from G by completing
X into a clique. Let C1, C2, . . . , Cr be the connected components of G \ X.
Then graph H obtained from GX by adding a set of fill edges F is a minimal
triangulation of G if and only if F = Sr
i=1 Fi, where Fi is the set of fill edges
in a minimal triangulation of GX [N [Ci]].
The following result about the structure of potential maximal cliques is due
to Bouchitt´e and Todinca.
Proposition 2.5 ([8]). Let Ω ⊆ V be a set of vertices of the graph G. Let
{C1, C2, . . . , Cp} be the set of the connected components of G\Ω and {S1, S2, . . . , Sp},
where Si = N (Ci) for i ∈ {1, 2, . . . , p}. Then Ω is a potential maximal clique of
G if and only if:
1. G \ Ω has no full component associated to Ω, and
2. the graph on the vertex set Ω obtained from G[Ω] by completing each Si,
i ∈ {1, 2, . . . , p}, into a clique, is a complete graph.
Moreover, if Ω is a potential maximal clique, then {S1, S2, . . . , Sp} is the set of
minimal separators of G contained in Ω.
We will need also the following proposition from [23].
Proposition 2.6 ([23]). Let Ω be a potential maximal clique of G. Then for
every y ∈ Ω, Ω = NG(Y )∪{y}, where Y is the connected component of G\ (Ω\
{y}) containing y.
A naive approach of deciding if a given vertex subset is a potential maximal
clique would be to try all possible minimal triangulations. There is a much
faster approach of recognizing potential maximal cliques due to Bouchitt´e and
Todinca based on Proposition 2.5.
Proposition 2.7 ([8]). There is an algorithm that, given a graph G = (V, E)
and a set of vertices Ω ⊆ V , verifies if Ω is a potential maximal clique of G in
time O(nm).
Parameterized complexity. A parameterized problem Π is a subset of Γ∗× N
for some finite alphabet Γ. An instance of a parameterized problem consists of
(x, k), where k is called the parameter. A central notion in parameterized com-
plexity is fixed parameter tractability (FPT) which means, for a given instance
(x, k), solvability in time f (k) · p(x), where f is an arbitrary function of k and
p is a polynomial in the input size. We refer to the book of Downey and Fellows
[19] for further reading on Parameterized Complexity.
Kernelization. A kernelization algorithm for a parameterized problem Π ⊆
Γ∗ × N is an algorithm that given (x, k) ∈ Γ∗ × N outputs in time polynomial
6
in x + k a pair (x′, k′) ∈ Γ∗ × N, called kernel such that (x, k) ∈ Π if and only
if (x′, k′) ∈ Π, x′ ≤ g(k), and k′ ≤ k, where g is some computable function.
The function g is referred to as the size of the kernel. If g(k) = kO(1) then we
say that Π admits a polynomial kernel.
There are several known polynomial kernels for the Minimum Fill-in prob-
lem [36, 37]. The best known kernelization algorithm is due to Natanzon et al.
[42, 43], which for a given instance (G, k) outputs in time O(k2nm) an instance
(G′, k′) such that k′ ≤ k, V (G′) ≤ 2k2 + 4k, and (G, k) is a YES instance if
and only if (G′, k′) is.
Proposition 2.8 ([42, 43]). Minimum Fill-in has a kernel with vertex set of
size O(k2). The running time of the kernelization algorithm is O(k2nm).
3 Branching
In our algorithm we apply branching procedure (Rule 1) whenever it is possible.
To describe this rule, we need some definitions. Let u, v be two nonadjacent
vertices of G, and let X = N (u) ∩ N (v) be the common neighborhood of u and
v. Let also P = uw1w2 . . . wℓv be a chordless uv-path.
In other words, any
two vertices of P are adjacent if and only if they are consecutive vertices in P .
We say that visibility of X from P is obscured if X \ N (wi) ≥ √k for every
i ∈ {1, . . . , ℓ}. Thus every internal vertex of P is nonadjacent to at least √k
vertices of X.
The idea behind branching is based on the observation that every fill-in of
G with at most k edges should either contain fill edge uv, or should make at
least one internal vertex of the path to be adjacent to all vertices of X.
Rule 1 (Branching Rule). If instance (G = (V, E), k) of Minimum Fill-in
contains a pair of nonadjacent vertices u, v ∈ V and a chordless uv-path P =
uw1w2 . . . wℓv such that visibility of X = N (u) ∩ N (v) from P is obscure, then
branch into ℓ + 1 instances (G0, k0), (G1, k1), . . . , (Gℓ, kℓ). Here
• G0 = (V, E ∪ {uv}), k0 = k − 1;
• For i ∈ {1, . . . , ℓ}, Gi = (V, E ∪ Fi), ki = k − Fi, where Fi = {wixx ∈
X ∧ wix 6∈ E}.
(G, k) is a YES instance if and only if
Lemma 3.1. Rule 1 is sound, i.e.
(Gi, ki) is a YES instance for some i ∈ {0, . . . , ℓ}.
Proof. If for some i ∈ {0, . . . , ℓ}, (Gi, ki) is a YES instance, then G can be
turned into a chordal graph by adding at most ki + Fi = k edges, and thus
(G, k) is a YES instance.
Let (G, k) be a YES instance, and let F ⊆ [V ]2 be such that graph H =
(V, E ∪ F ) is chordal and F ≤ k. By Proposition 2.1, there exists an ordering
π of V , such that the Elimination Game algorithm on G and π outputs H.
Without loss of generality, we can assume that π−1(u) < π−1(v). If for some
7
x ∈ X, π−1(x) < π−1(u), then by Proposition 2.2, we have that uv ∈ F . Also
by Proposition 2.2, if π−1(wi) < π−1(u) for each i ∈ {1, . . . , ℓ}, then again
uv ∈ F . In both cases (G0, k0) is a YES instance.
The only remaining case is when π−1(u) < π−1(x) for all x ∈ X, and there
is at least one vertex of P placed after u in ordering π. Let i ≥ 1 be the
smallest index such that π−1(u) < π−1(wi). Thus for every x ∈ X, in the
path xuw1, w2, . . . , wi all internal vertices are ordered by π before x and wi. By
Proposition 2.2, this imply that wi is adjacent to all vertices of X, and hence
(Gi, ki) is a YES instance.
The following lemma shows that every branching step of Rule 1 can be
performed in polynomial time.
Lemma 3.2. Let (G, k) be an instance of Minimum Fill-in. It can be identified
in time O(n4) if there is a pair u, v ∈ V (G) satisfying conditions of Rule 1.
Moreover, if conditions of Rule 1 hold, then a pair u, v of two nonadjacent
vertices and a chordless uv-path P such that visibility of N (u) ∩ N (v) from P
is obscured, can be found in time O(n4).
Proof. For each pair of nonadjacent vertices u, v, we compute X = N (u)∩N (v).
We compute the set of all vertices W ⊆ V (G) \ {u, v}, such that every vertex
of W is nonadjacent to at least √k vertices of X. Then conditions of Rule 1
do not hold for u and v if in the subgraph Guv induced by W ∪ {u, v}, u and
v are in different connected components. If u and v are in the same connected
component of Guv, then a shortest (in Guv) uv-path P is a chordless path and
the visibility of X from P is obscured. Clearly, all these procedures can be
performed in time O(n4).
We say that instance (G, k) is non-reducible if conditions of Rule 1 do not
hold. Thus for each pair of vertices u, v of non-reducible graph G, there is no
uv-path with obscure visibility of N (u) ∩ N (v).
Lemma 3.3. Let t(n, k) be the maximum number of non-reducible problem in-
stances resulting from recursive application of Rule 1 starting from instance
(G, k) with V (G) = n. Then t(n, k) = nO(√k) and all generated non-reducible
instances can be enumerated within the same time bound.
Proof. Let us assume that we branch on the instances corresponding to a pair
u, v and path P = uw1w2 . . . wℓv such that the visibility of N (u) ∩ N (w) is
obscure from P . Then the value of t(n, k) is at most Pℓ
i=0 t(n, ki). Here k0 = 1
and for all i ≥ 1, ki = k−Fi ≤ k−√k. Since the number of vertices in P does
not exceed n, we have that t(n, k) ≤ t(n, k − 1) + n · t(n, k − √k). By making
use of standard arguments on the amount of leaves in branching trees (see, for
example [34, Theorem 8.1]) it follows that t(n, k) = nO(√k). By Lemma 3.2,
every recursive call of the branching algorithm can be done in time O(n4), and
thus all non-reducible instances are generated in time O(nO(√k) · n4) = nO(√k).
8
4 Listing vital potential maximal cliques
Let (G, k) be a YES instance of Minimum Fill-in. It means that G can be
turned into a chordal graph H by adding at most k edges. Every maximal clique
in H corresponds to a potential maximal clique of G. The observation here is
that if a potential maximal clique Ω needs more than k edges to be added to
become a clique, then no solution H can contain Ω as a maximal clique.
In
Section 5 we prove that the only potential maximal cliques that are essential
for a fill-in with at most k edges are the ones that miss at most k edges from a
clique.
A potential maximal clique Ω is vital if the number of edges in G[Ω] is at
least Ω(Ω− 1)/2− k. In other words, the subgraph induced by vital potential
maximal clique can be turned into a complete graph by adding at most k edges.
In this section we show that all vital potential maximal cliques of an n-vertex
non-reducible graph can be enumerated in time nO(√k).
We will first show how to enumerate potential maximal cliques which are,
in some sense, almost cliques. This enumeration algorithm will be used as a
subroutine to enumerate vital potential maximal cliques. A potential maximal
clique Ω is quasi-clique, if there is a set Z ⊆ Ω of size at most 5√k such that
Ω \ Z is a clique. In particular, if Ω ≤ 5√k, then Ω is also a quasi-clique. The
following lemma gives an algorithm enumerating all quasi-cliqes.
Proof. We will prove that while a quasi-clique can be very large, it can be
Lemma 4.1. Let (G, k) be a problem instance on n vertices. Then all quasi-
cliques in G can be enumerated within time nO(√k).
reconstructed in polynomial time from a small set of O(√k) vertices. Hence
all quasi-cliques can be generated by enumerating vertex subsets of size O(√k).
Because the amount of vertex subsets of size O(√k) is nO(√k), this will prove
Let Ω be a potential maximal clique which is a quasi-clique, and let Z ⊆ Ω
be the set of size at most 5√k such that X = Ω \ Z is a clique. Depending on
the amount of full components associated to X in G\ (Z∪ X), we consider three
cases: There are at least two full components, there is exactly one, and there is
no full component.
the lemma.
Consider first the case when X has at least two full components, say C1
and C2. In this case, by Proposition 2.3, X is a minimal clique separator of
G \ Z. Let H be some minimal triangulation of G \ Z. By Proposition 2.4,
clique minimal separators remain clique minimal separators in every minimal
triangulation. Therefore, X is a minimal separator in H. It is well known that
every chordal graph has at most n − 1 minimal separators and that they can
be enumerated in linear time [12]. To enumerate quasi-cliques we implement
the following algorithm. We construct a minimal triangulation H of G \ Z. A
minimal triangulation can be constructed in time O(nm) or O(nω log n), where
ω < 2.37 is the exponent of matrix multiplication and m is the number of edges
in G [32, 48]. For every minimal separator S of H, where G[S] is a clique, we
9
check if S ∪ Z is a potential maximal clique in G. This can be done in O(km)
time by Proposition 2.7. Therefore, in this case, the time required to enumerate
all quasi-cliques of the form X ∪ Z, up to polynomial multiplicative factor is
proportional to the amount of sets Z of size at most 5√k. The total running
time to enumerate quasi-cliques of this type is nO(√k).
Now we consider the case when no full component in G\ (Z ∪ X) associated
to X. It means that for every connected component C of G \ (Z ∪ X), there is
x ∈ X\N (C). By Proposition 2.5, X is also a potential maximal clique in G\Z.
We construct a minimal triangulation H of G\ Z. By Proposition 2.4, X is also
a potential maximal clique in H. By the classical result of Dirac [18] chordal
graph H contains at most n maximal cliques and all the maximal cliques of H
can be enumerated in linear time [3]. For every maximal clique K of H such
that K is also a clique in G, we check if K ∪ Z is a potential maximal clique
in G, which can be done in O(nm) time by Proposition 2.7. As in the previous
case, the enumeration of all such quasi-cliques boils down to enumerating sets
Z, which takes time nO(√k).
Last case, vertex set X has unique full component Cr in G\(Z∪X) associated
to X. Since Ω = Z ∪ X, we have that each of the connected components
C1, C2, . . . , Cr of G \ (Z ∪ X) is also a connected component of G \ Ω. Then for
every i ∈ {1, . . . , r − 1}, Si = NG\Z(Ci) is a clique minimal separator in G \ Z
because Si ⊆ X is a clique, and Ci together with the component of G \ (Z ∪ Si)
containing X \ Si, are full components associated to Si. Let H be a minimal
triangulation of G \ Z. Vertex set X is a clique in G \ Z and thus is a clique in
H. Let K be a maximal clique of H containing X. By Proposition 2.4, for every
i ∈ {1, . . . , r − 1}, Si is a minimal separator in H. By Proposition 2.4, G \ Z
has no fill edges between vertices separated by Si and thus Ci is a connected
component of H \ K.
Because Ω is a potential maximal clique in G, by Proposition 2.5, there is
y ∈ Ω such that y 6∈ NG(Cr). Since Cr is a full component for X, we have
that y ∈ Z. Moreover, every connected component C 6= Cr of G \ (Z ∪ X)
is also a connected component of H \ K. Thus every connected component of
H \ K containing a neighbor of y in G is also a connected component of G \ Ω
containing a neighbor of y.
Let B1, B2, . . . , Bℓ be the set of connected components in G \ (K ∪ Z) with
y ∈ NG(Bi). We define
Y = [
1≤i≤ℓ
Bi ∪ {y}.
By Proposition 2.6, Ω = NG(Y ) ∪ {y} and in this case potential maximal
clique is characterized by y and Y .
To summarize, to enumerate all quasi-cliques corresponding to the last case,
we do the following. For every set Z of size at most 5√k, we construct a minimal
triangulation H of G \ Z. The amount of maximal cliques in a chordal graph
H is at most n, and for every maximal clique K of H and for every y ∈ Z,
we compute the set Y . We use Propositions 2.7 to check if NG(Y ) ∪ {y} is a
10
u
Y
Nv
Nuv
v
Nu
1:
Figure
N u, N v, Nuv,{u},{v}, and Y .
Partitioning
of potential maximal
clique Ω into
sets
potential maximal clique. The total running time to enumerate quasi-cliques in
this case is bounded, up to polynomial factor, by the amount of subsets of size
O(√k) in G which is nO(√k).
Now we are ready to prove the result about vital potential maximal cliques
in non-reducible graphs.
Lemma 4.2. Let (G, k) be a non-reducible instance of the problem. All vital
potential maximal cliques in G can be enumerated within time nO(√k), where n
is the number of vertices in G.
Proof. We start by enumerating all vertex subsets of G of size at most 5√k + 2
and apply Proposition 2.7 to check if each such set is a vital potential maximal
clique or not.
Let Ω be a vital potential maximal clique with at least 5√k + 3 vertices and
let Y ⊆ Ω be the set of vertices of Ω such that each vertex of Y is adjacent in
G to at most Ω − 1 − √k vertices of Ω. To turn Ω into a complete graph, for
each vertex v ∈ Y , we have to add at least √k fill edges incident to v. Hence
Y ≤ 2√k. If Ω \ Y is a clique, then Ω is a quasi-clique. By Lemma 4.1, all
quasi-cliques can be enumerated in time nO(√k).
If Ω \ Y is not a clique, there is at least one pair of non-adjacent vertices
u, v ∈ Ω \ Y . By Proposition 2.5, there is a connected component C of G \ Ω
such that u, v ∈ N (C).
Claim 1. There is w ∈ C such that Ω \ N (w) ≤ 5√k + 2.
Proof. Targeting towards a contradiction, we assume that the claim does not
hold. We define the following subsets of Ω \ Y .
• N u ⊆ Ω \ Y is the set of vertices which are not adjacent to u,
• N v ⊆ Ω \ Y is the set of vertices which are not adjacent to v, and
• Nuv = Ω \ (Y ∪ N u ∪ N v) is the set of vertices adjacent to u and to v.
11
See Fig. 1 for an illustration. Let us note that
Ω = N u ∪ N v ∪ Nuv ∪ {u} ∪ {v} ∪ Y.
Since u, v 6∈ Y , we have that max{N u,N v} ≤ √k.
We claim that Nuv ≤ √k. Targeting towards a contradiction, let us assume
that Nuv > √k. By our assumption, every vertex w ∈ C is not adjacent to at
least 5√k+2 vertices of Ω. Since Y ∪N u∪N v∪{u}∪{v} ≤ 2√k+√k+√k+2 =
4√k + 2, we have that each vertex of C is nonadjacent to at least √k vertices of
Nuv. We take a shortest uv-path P with all internal vertices in C. Because C
is a connected component and u, v ∈ N (C), such a path exists. Every internal
vertex of P is nonadjacent to at least √k vertices of Nuv ⊆ N (u) ∩ N (v), and
thus the visibility of Nuv from P is obscured. But this is a contradiction to the
assumption that (G, k) is non-reducible. Hence Nuv ≤ √k.
Thus if the claim does not hold, we have that
Ω = N u ∪ N v ∪ Nuv ∪ {u} ∪ {v} ∪ Y ≤ 5√k + 2,
but this contradicts to the assumption that Ω ≥ 5√k + 3. This concludes the
proof of the claim.
We have shown that for every vital potential maximal clique Ω of size at least
5√k + 3, there is a connected component C and w ∈ C such that Ω \ N (w) ≤
5√k + 2. Let H be the graph obtained from G by completing N (w) into a
clique. The graph H[Ω] consist of a clique plus at most 5√k + 2 vertices. We
want to show that Ω is a quasi-clique in H, by arguing that Ω is a potential
maximal clique in H. Vertex set Ω is a potential maximal clique in G, and thus
by Proposition 2.5, there is no full component associated to Ω in G\ Ω. Because
N (w) ∩ Ω ⊆ N (C) ⊂ Ω, there is no full component associated to Ω in H. We
use Proposition 2.5 to show that Ω is a potential maximal clique in H as well.
Hence Ω is a quasi-clique in H.
To conclude, we use the following strategy to enumerate all vital potential
maximal cliques. We enumerate first all quasi-cliques in G in time nO(√k) by
making use of Lemma 4.1, and for each quasi-clique we use Proposition 2.7 to
check if it is a vital potential maximal clique. We also try all vertex subsets of
size at most 5√k + 2 and check if each of such sets is a vital potential maximal
clique. All vital potential maximal cliques which are not enumerated prior to
this moment should satisfy the condition of the claim. As we have shown, each
such vital potential maximal clique is a quasi-clique in the graph H obtained
from G by selecting some vertex w and turning NG(w) into clique. Thus for
every vertex w of G, we construct graph H and then use Lemma 4.1 to enumerate
all quasi-cliques in H. For each quasi-clique of H, we use Proposition 2.7 to
check if it is a vital potential maximal clique in G. The total running time of
this procedure is nO(√k).
12
5 Exploring remaining solution space
For an instance (G, k) of Minimum Fill-in, let Πk be the set of all vital po-
tential maximal cliques. In this section, we give an algorithm of running time
O(nmΠk), where n is the number of vertices and m the number of edges in G.
The algorithm receives (G, k) and Πk as an input, and decides if (G, k) is a YES
instance. The algorithm is a modification of the algorithm from [23]. The only
difference is that the algorithm from [23] computes an optimum triangulation
from the set of all potential maximal cliques while here we have to work only
with vital potential maximal cliques. For reader's convenience we provide the
full proof, but first we need the following lemma.
Lemma 5.1. Let S be a minimal separator in G and let C be a full connected
component of G\ S associated to S. Then every minimal triangulation H of GS
contains a maximal clique K such that S ⊂ K ⊆ S ∪ C.
Proof. By Proposition 2.4, H is a minimal triangulation of GS if and only if
H[S ∪ C] is a minimal triangulation of GS[S ∪ C]. Because S is a clique GS[S]
in, S is a subset of some maximal clique K of H[S ∪ C]. By definition K is a
potential maximal clique in GS[S ∪ C], and by Proposition 2.5 K is a potential
maximal clique in G. Since GS[S ∪ C]\ S has a full component associated to S,
we have that by Proposition 2.5, S is not a potential maximal clique in GS[S∪C]
and thus S ⊂ K.
Lemma 5.2. Given a set of all vital potential maximal cliques Πk of G, it can
be decided in time O(nmΠk) if (G, k) is a YES instance of Minimum Fill-in.
Proof. Let mfi(G) be the minimum number of fill edges needed to triangulate G.
Let us remind that by fillG(Ω) we denote the number of non-edges in G[Ω] and
by GΩ the graph obtained from G by completing Ω into a clique. If mfi(G) ≤ k,
then by Proposition 2.4, we have that
mfi(G) = min
Ω∈Πk
[fillG(Ω) +
X
C is a component of G\Ω
mfi(GΩ[C ∪ NG(C)])].
(1)
Formula (1) can be used to compute mfi(G), however by making use of this
formula we are not able to obtained the claimed running time. To implement
the algorithm in time O(nmΠk), we compute mfi(GΩ[C∪NG(C)])] by dynamic
programming.
By Proposition 2.5, for every connected component C of G\Ω, where Ω ∈ Πk,
S = NG(C) ⊂ Ω is a minimal separator. We define the set ∆k as the set of all
minimal separators S, such that S = N (C) for some connected component C
in G \ Ω for some Ω ∈ Πk. Since for every Ω ∈ Πk the number of components
in G \ Ω is at most n, we have that ∆k ≤ nΠk.
For S ∈ ∆k and a full connected component C of G\ S associated to S. We
define ΠS,C as the set of potential maximal cliques Ω ∈ Πk such that S ⊂ Ω ⊆
S ∪ C. The triple (S, C, Ω) was called a good triple in [23].
13
For every Ω ∈ Πk, connected component C of G \ Ω, and S = N (C), we
compute mfi(F ), where F = GΩ[C ∪ S]. We start dynamic programming by
computing the values for all sets (S, C) such that Ω = C ∪ S is an inclusion-
minimal potential maximal clique. In this case we put mfi(F ) = fill(C ∪ S).
Observe that GS[C ∪ S] = GΩ[C ∪ S]. Hence by Lemma 5.1, for every minimal
triangulation H of GS, there exists a potential maximal clique Ω in G such that
Ω is a maximal clique in H and S ⊂ Ω ⊆ S ∪ C. Thus Ω ∈ ΠS,C. Using this
observation, we write the following formula for dynamic programming.
mfi(F ) = min
Ω′∈ΠS,C
[fillF (Ω′) +
X
C ′ is a component of F\Ω′
mfi(FΩ′ [C′ ∪ N (C′)])]. (2)
The fact S ⊂ Ω′ ensures that the solution in (2) can be reconstructed from
instances with S ∪ C of smaller sizes. By (1) and (2) we can decide if there
exists a triangulation of G using at most k fill edges, and to construct such a
triangulation. It remains to argue for the running time.
Finding connected components in G \ Ω and computing fill(Ω) can easily
been done in O(n + m) time. Furthermore, (1) is applied Πk times in total.
The running time of dynamic programming using (2) is proportional to the
amount of states of dynamic programming, which is
X
S∈∆k
X
C∈G\S
ΠS,C
The graph G \ Ω contains at most n connected components and thus for every
minimal separator, each potential maximal clique is counted at most n times,
and thus the amount of the elements in the sum does not exceed nΠk. The
total running time is O(nmΠk).
6 Putting things together
Now we are in the position to prove the main result of this paper.
Theorem 6.1. The Minimum Fill-in problem is solvable in time O(2O(√k log k)+
k2nm).
Proof. Step A. Given instance (G, k) of the Minimum Fill-in problem, we use
Proposition 2.8 to obtain a kernel (G′, k′) on O(k2) vertices and with k′ ≤ k.
Let us note that (G, k) is a YES instance if and only if (G′, k′) is a YES instance.
This step is performed in time O(k2nm).
Step B1. We use Branching Rule 1 on instance (G′, k′). Since the number of
vertices in G′ is O(k2), we have that by Lemma 3.3, the result of this procedure is
the set of (k2)O(√k) = 2O(√k log k) non-reducible instances (G1, k1), . . . , (Gp, kp).
For each i ∈ {1, 2, . . . , p}, graph Gi has O(k2) vertices and ki ≤ k. Moreover,
by Lemma 3.1, (G′, k′), and thus (G, k), is a YES instance if and only if at least
14
one (Gi, ki) is a YES instance. By Lemma 3.3, the running time of this step is
2O(√k log k).
Step B2. For each i ∈ {1, 2, . . . , p}, we list all vital potential maximal cliques
of graph Gi. By Lemma 4.2, the amount of all vital potential maximal cliques
in non-reducible graph Gi is 2O(√k log k) and they can be listed within the same
running time.
Step C. At this step for each i ∈ {1, 2, . . . , p}, we are given instance (Gi, ki)
together with the set Πki of vital potential maximal cliques of Gi computed in
Step B2. We use Lemma 5.2 to solve the Minimum Fill-in problem for instance
(Gi, ki) in time O(k6Πki) = 2O(√k log k). If at least one of the instances (Gi, ki)
is a YES instance, then by Lemma 3.1, (G, k) is a YES instance. If all instances
(Gi, ki) are NO instances, we conclude that (G, k) is a NO instance. Since p =
2O(√k log k), we have that Step C can be performed in time 2O(√k log k). The total
running time required to perform all steps of the algorithm is O(2O(√k log k) +
k2nm).
Let us remark that our decision algorithm can be easily adapted to output
the optimum fill-in of size at most k.
7 Applications to other problems
The algorithm described in the previous section can be modified to solve several
related problems. Problems considered in this section are Minimum Chain
Completion, Chordal Graph Sandwich, and Triangulating Colored
Graph.
Minimum Chain Completion. Bipartite graph G = (V1, V2, E) is a chain
graph if the neighbourhoods of the nodes in V1 forms a chain, that is there is an
ordering v1, v2, . . . , vV1 of the vertices in V1, such that N (v1) ⊆ N (v2) ⊆ . . . ⊆
N (vV1).
Minimum Chain Completion
Input: A bipartite graph G = (V1, V2, E) and integer k ≥ 0.
Question: Is there F ⊆ V1 × V2, F ≤ k, such that graph H = (V1, V2, E ∪ F ) a chain
graph?
Yannakakis in his famous NP-completeness proof of Minimum Fill-in [52]
used the following observation. Let G be a bipartite graph with bipartitions V1
and V2, and let G′ be cobipartite (the complement of bipartite) graph formed
by turning V1 and V2 into cliques. Then G can be transformed into a chain
graph by adding k edges if and only if G′ can be triangulated with k edges.
By Theorem 6.1, we have that Minimum Chain Completion is solvable in
O(2O(√k log k) + k2nm) time.
15
In the chordal graph sandwich problem we are
Chordal graph sandwich.
given two graphs G1 = (V, E1) and G2 = (V, E2) on the same vertex set V , and
with E1 ⊂ E2. The Chordal Graph Sandwich problem asks if there exists
a chordal graph H = (V, E1 ∪ F ) sandwiched in between G1 and G2, that is
E1 ∪ F ⊆ E2.
Chordal Graph Sandwich
Input: Two graphs G1 = (V, E1) and G2 = (V, E2) such that E1 ⊂ E2, and k =
E2 \ E1.
Question: Is there F ⊆ E2 \ E1, F ≤ k, such that graph H = (V, E1 ∪ F ) is a
triangulation of G1?
Let us remark that the Chordal Graph Sandwich problem is equivalent
to asking if there is a minimal triangulation of G1 sandwiched between G1 and
G2.
To solve Chordal Graph Sandwich we cannot use the algorithm from
Section 6.1 directly. The reason is that we are only allowed to add edges from
E2 as fill edges. We need a kernelization algorithm for this problem as well.
Lemma 7.1. Chordal Graph Sandwich has a kernel with vertex set of size
O(k3). The running time of kernelization algorithm is O(k2nm).
Proof. To simplify notations, we denote an instance of the Chordal Graph
Sandwich problem by (G, E′, k), where G1 = G = (V, E), E′∩E = ∅, k ≤ E′,
and G2 = (V, E ∪ E′).
We first define two reduction rules and prove their correctness.
No-cycle-vertex Rule. If instance (G, E′, k) has u ∈ V such that for each
connected component C of G[V \ N [u]], N (C) is a clique, then replace instance
(G, E′, k) with instance (G \ {u}, E′′, k′), where E′′ ⊆ E′ are the edges not
incident to u and k′ = k − E′ \ E′′.
Claim 2. No-cycle-vertex Rule is sound, i.e. (G \ {u}, E′′, k′) is an YES in-
stance of the chordal sandwich problem if and only if (G, E′, k) is an YES in-
stance.
Proof. Let Gu be the graph G[V \ {u}]. Chordallity is a hereditary property so
if H = (V, E ∪ F ) is a triangulation of G where F ≤ k, then H[V \ {u}] is a
triangulation of Gu where the set of fill edges are the edges of F not incident to
u.
For the opposite direction assume that Hu = (V (Gu), E(Gu) ∪ Fu) is a
minimal triangulation of Gu where Fu ≤ k. Our objective now is to argue that
H = (V, E ∪ Fu) is a triangulation of G. Targeting towards a contradiction, let
us assume that W = uaw1w2 . . . wℓbu is a chordless cycle of length at least four
in H. Then notice that u is a vertex of W as Hu = H[V \{u}] is a chordal graph
by our assumption. Vertices a and b of W are not adjacent by definition, and let
vertices w1w2 . . . wℓ be contained in the connected component C of G[V \ N [u].
16
Nonadjacent vertices a, b are now contained in N (C) which is a contradiction
to the condition of applying No-cycle-vertex Rule.
For each pair of nonadjacent vertices x, y ∈ V , we define Ax,y as the set of
vertices such that w ∈ Ax,y if x, y ∈ NG(w) and vertices x, y are contained in
the same connected component of G[(V \ N [w]) ∪ {x, y}].
Safe-edge Rule.
instance (G, E′, k) then
If 2k < Ax,y for some pair of vertices x, y in a problem
• replace instance (G, E′, k) with a trivial NO instance if xy 6∈ E′, and
• otherwise with instance (G = (V, E ∪ {xy}), E′ \ {xy}, k − 1).
the instance outputted by the rule is
Claim 3. Safe-edge Rule is sound, i.e.
a YES instance of the chordal sandwich problem if and only if (G, E′, k) is an
YES instance.
Proof. By the definition of Ax,y, it is not hard to see that there exists an induced
cycle of length at least four consisting of x, w, y and a shortest induced path from
x to y in G[(V \ N [w])∪{x, y}]. A trivial observation is that every triangulation
of G either has xy as a fill edge, or there exists a fill edge incident to w. Since
2k < Axy, we have that every minimal triangulation not using xy as a fill edge
has at least one fill edge incident to each vertices in Auv, and thus (G, E′, k)
is a NO instance if xy 6∈ E′ and xy ∈ F for every edge set F ⊆ E′ such that
H = (V, E ∪ F ) is chordal.
It is possible to show that exhausting application of both No-cycle-vertex
and Safe-edge rules either results in a polynomial kernel or a trivial recognizable
NO instance. However, the running time of the algorithm would be O(kn2m).
In what follow, we show that with much more careful implementation of rules, it
is possible to obtain a kernel of size O(k3) in time O(k2nm). The algorithm uses
the same approach as the kernel algorithms given in [37] and [43] for Minimum
Fill-in.
Let (G, E′, k) be an instance of the problem. We give an algorithm with
running time O(k2nm) that outputs an instance (G′, E′′, k′) such that V (G′) ≤
32k3 + 4k and (G′, E′′, k′) is a YES instance if and only if (G, E′, k) is a YES
instance. Let us remark that we put no efforts to optimize the size of the kernel.
Let A, B be a partitioning of the vertices of graph G in the given instance
(G, E′, k).
Initially we put A = ∅ and B = V (G). There is a sequence of
procedures. To avoid a confusing nesting of if-then-else statements, we use the
following convention: The first case which applies is used first in the procedure.
Thus, inside a given procedure, the hypotheses of all previous procedures are
assumed false.
P1: If 4k < A then return a trivial NO instance, else if G[B] contains a
chordless cycle W of length at least four, move V (W ) from B to A.
17
P2: If 4k < A then return a trivial NO instance, else if G[B] contains a
chordless path P of at least two vertices which is also an induced subgraph
of a chordless cycle W of G of length at least four, move V (P ) from B to
A.
P3: Compute Ax,y for each pair of vertices in A.
P4: If Ax,y ≤ 2k then move Ax,y from B to A, else if xy 6∈ E′ return a trivial
NO instance, else add edge xy to F .
P5: Delete every vertex of B.
Now we argue on the correctness and running time required to implement the
procedures. P1: Every cordless cycle of length 4 ≤ ℓ requires at least ℓ − 3 fill
edges to be triangulated [37]. Thus, at least one fill edge has to be added for
each 4 vertices moved to A, and we have a NO instance if 4k < A. A chordless
cycle of a non chordal graph can be obtained in O(n + m) time [49]. Total
running time is O(k(n + m)).
P2: Let P be an induced path which is an induced subgraph of a chordless
cycle W in a graph G. Then any triangulation of G will add at least V (P )−1 fill
edges incident to vertices in P [37]. Thus, 2 end points of a fill edge(equivalently
to one fill edge) has to be added for each 4 vertices moved to A, and we have
a NO instance if 4k < A. A path P can be obtained as follows: Let u be an
end vertex of P and let x be the unique neighbour of u in A on the cycle W .
We have a chordless cycle W satisfying the conditions if there is a path from
(N (u)∩ B)\ N (x) to a vertex in N (x)\ N (u) not using vertices of N (x)∩ N (u).
Such a path can trivially be found in O(m) time. Thus the time required to
this step is O(knm).
P3: Let x, y be nonadjacent vertices of A. For each vertex w ∈ N (x)∩N (y)∩
B check if there is a path from x to y avoiding N [w], if so there is a chordless
cycle containing xwy as consecutive vertices. Running time is O(k2nm) because
A is O(k).
P4: If Ax,y > 2k then by the Safe-edge Rule, it is safe to add edge xy if
xy ∈ E′, and to return a trivial NO instance if xy 6∈ E′. Running time for this
step is O(k2n).
P5: For every remaining vertex u ∈ B, No-cycle-vertex Rule can be applied,
and vertex u can safely be deleted from the graph M = (V, E∪F ). Let us on the
contrary assume that this is not the case. Then one can find two nonadjacent
vertices a, b ∈ NM (u) and connected component C of M \ NM [w] such that
a, b ∈ NM (C). Let W be the chordless cycle of length at least four, obtained
by yux and a shortest chordless path from x to y in M [C ∪ {x, y}]. By P1, at
least one vertex of W is contained in A and by P2, no two consecutive vertices
of W are contained in B. By our assumption u ∈ B, so x, y ∈ A by P2. Vertex
u is contained in N (x) ∩ N (y) because during P3 we add edges only between
vertices in A. Furthermore, u ∈ Ax,y because otherwise by [37, Theorem 2.10],
u would already have been moved to A by P2. Since xy 6∈ F and u ∈ Ax,y ∩ B,
we have a contradiction to P4.
18
Let G′ = (A, E(G[A] ∪ F ) and let E′′ be the edges in E′ \ F where both
endpoints are incident to vertices of A. Since (G′, E′′, k − F) is obtained by
applying Safe-edge Rule on each edge in F and No-cycle-vertex Rule on each
vertex in B, we have that instance (G′, E′′, k − F) is a YES instance if and
only if (G, E′, k) is a YES instance, Finally, A ≤ 32k3 + 4k because during P1
and P2 we move in total at most 4k vertices from B to A, and during P3 at
most 2k · (4k)2 vertices to A.
Theorem 7.2. Chordal Sandwich Problem is solvable in time O(2O(√k log k)+
k2nm).
Proof. Let (G1, G2, k) be an instance of the problem. We sketch the proof
by following the steps of the proof of Theorem 6.1 and commenting on the
differences. Step A: We use Lemma 7.1, to obtain in time O(k2nm) kernel
(G′1, G′2, k′) such that V (G′1) = O(k3) and k′ ≤ k. Step B1: On kernel we use
Branching Rule 1 exhaustively, with the adaptation that every instance defined
by fill edge set Fi where Fi 6⊆ E2 is discarded. Thus we obtain 2O(√k log k)
non-reducible instances. Step B2: For each non-reducible instance (Gi
2, ki),
we enumerate vital potential maximal cliques of Gi
1 but discard all potential
maximal cliques that are not cliques in Gi
2. Step C: Solve the remaining problem
in time proportional to the number of vital potential maximal cliques in G′1 that
are also cliques in Gi
2. This step is almost identical to Step C of Theorem 6.1.
1, Gi
Triangulating Colored Graph. In the Triangulating Colored Graph
problem we are given a graph G = (V, E) with a partitioning of V into sets
V1, V2, . . . , Vc, a colouring of the vertices. Let us remark that this coloring is
not necessarily a proper coloring of G. The question is if G can be triangulated
without adding edges between vertices in the same set (colour).
Triangulating Colored Graph
Input: A graph G = (V, E), a partitioning of V into sets V1, V2, . . . , Vc, and an integer
k.
Question: Is there F ⊆ [V ]2, F ≤ k, and such that for each uv ∈ F , u, v 6∈ Vi,
1 ≤ i ≤ c, and graph H = (V, E1 ∪ F ) is a triangulation of G?
Triangulating Colored Graph can be trivially reduced to Chordal
Graph Sandwich by defining G1 = G, and the edge set of graph G2 as the edge
set of G1 plus the set of all vertex pairs of different colors. Thus by Theorem 7.2,
Triangulating Colored Graph is solvable in time O(2O(√k log k) + k2nm).
8 Conclusions and open problems
In this paper we gave the first parameterized subexponential time algorithm
solving Minimum Fill-in in time O(2O(√k log k) + k2nm). It would be interest-
ing to find out how tight is the exponential dependence, up to some complexity
19
assumption, in the running time of our algorithm. We would be surprised to
hear about time 2o(√k)nO(1) algorithm solving Minimum Fill-in. For example,
such an algorithm would be able to solve the problem in time 2o(n). However,
the only results we are aware in this direction is that Minimum Fill-in can-
not be solved in time 2o(k1/6)nO(1) unless the ETH fails [14]. See [33] for more
information on ETH. Similar uncertainty occurs with a number of other graph
problems expressible in terms of vertex orderings. Is it possible to prove that
unless ETH fails, there are no 2o(n) algorithms for Treewidth, Minimum In-
terval Completion, and Optimum Linear Arrangement? Here the big
gap between what we suspect and what we know is frustrating.
On the other hand, for the Triangulated Colored Graph problem,
which we are able to solve in time O(2O(√k log k) + k2nm), Bodlaender et al. [6]
gave a polynomial time reduction that from a 3-SAT formula on p variables and
q clauses constructs an instance of Triangulated Colored Graph. This
instance has 2 + 2p + 6q vertices and a triangulation of the instance respecting
its colouring can be obtained by adding of at most (p + 3q) + (p + 3q)2 + 3pq
edges. Thus up to ETH, Triangulated Colored Graph and Chordal
Graph Sandwich cannot be solved in time 2o(√k)nO(1).
The possibility of improving the nm factor in the running time O(2O(√k log k)+
k2nm) of the algorithm is another interesting open question. The factor nm
appears from the running time required by the kernelization algorithm to iden-
tify simplicial vertices. Identification of simplicial vertices can be done in time
O(min{mn, nω log n}), where ω < 2.37 is the exponent of matrix multiplication
[32, 39]. Is the running time required to obtain a polynomial kernel for Min-
imum Fill-in is at least the time required to identify a simplicial vertex in a
graph and can search of a simplicial vertex be done faster than finding a triangle
in a graph?
Finally, there are various problems in graph algorithms, where the task is
to find a minimum number of edges or vertices to be changed such that the
resulting graph belongs to some graph class. For example, the problems of
completion to interval and proper interval graphs are fixed parameter tractable
[31, 36, 37, 50]. Can these problems be solved by subexponential parameterized
algorithms? Are there any generic arguments explaining why some FPT graph
modification problems can be solved in subexponential time and some can't?
Acknowledgement We are grateful to Marek Cygan and Saket Saurabh for
discussions and useful suggestions.
References
[1] N. Alon, D. Lokshtanov, and S. Saurabh, Fast FAST, in ICALP, vol. 5555
of LNCS, Springer, 2009, pp. 49 -- 58.
[2] C. Beeri, R. Fagin, D. Maier, and M. Yannakakis, On the desirability of
acyclic database schemes, J. ACM, 30 (1983), pp. 479 -- 513.
20
[3] J. R. S. Blair and B. W. Peyton, An introduction to chordal graphs and
clique trees, in Graph Theory and Sparse Matrix Computations, Springer, 1993,
pp. 1 -- 30. IMA Volumes in Mathematics and its Applications, Vol. 56.
[4] H. Bodlaender, P. Heggernes, and Y. Villanger, Faster parameterized
algorithms for minimum fill-in, Algorithmica, (to appear).
[5] H. L. Bodlaender, M. R. Fellows, M. T. Hallett, T. Wareham, and
T. Warnow, The hardness of perfect phylogeny, feasible register assignment and
other problems on thin colored graphs, Theor. Comput. Sci., 244 (2000), pp. 167 --
188.
[6] H. L. Bodlaender, M. R. Fellows, and T. Warnow, Two strikes against
perfect phylogeny, in ICALP, vol. 623 of LNCS, Springer, 1992, pp. 273 -- 283.
[7] M. L. Bonet, C. A. Phillips, T. Warnow, and S. Yooseph, Constructing
evolutionary trees in the presence of polymorphic characters, in STOC, ACM,
1996, pp. 220 -- 229.
[8] V. Bouchitt´e and I. Todinca, Treewidth and minimum fill-in: Grouping the
minimal separators, SIAM J. Comput., 31 (2001), pp. 212 -- 232.
[9]
, Listing all potential maximal cliques of a graph, Theor. Comput. Sci., 276
(2002), pp. 17 -- 32.
[10] P. Buneman, A characterisation of rigid circuit graphs, Discrete Math., 9 (1974),
pp. 205 -- 212.
[11] L. Cai, Fixed-parameter tractability of graph modification problems for hereditary
properties, Inf. Process. Lett., 58 (1996), pp. 171 -- 176.
[12] L. S. Chandran and F. Grandoni, A linear time algorithm to list the minimal
separators of chordal graphs, Discrete Mathematics, 306 (2006), pp. 351 -- 358.
[13] F. R. K. Chung and D. Mumford, Chordal completions of planar graphs, J.
Comb. Theory, Ser. B, 62 (1994), pp. 96 -- 106.
[14] M. Cygan, Private communication. 2011.
[15] T. A. Davis, Direct methods for sparse linear systems, vol. 2 of Fundamentals of
Algorithms, Society for Industrial and Applied Mathematics (SIAM), Philadel-
phia, PA, 2006.
[16] E. D. Demaine, F. V. Fomin, M. Hajiaghayi, and D. M. Thilikos, Subex-
ponential parameterized algorithms on graphs of bounded genus and H-minor-free
graphs, J. Assoc. Comput. Mach., 52 (2005), pp. 866 -- 893.
[17] R. Diestel, Graph theory, vol. 173 of Graduate Texts in Mathematics, Springer-
Verlag, Berlin, third ed., 2005.
[18] G. A. Dirac, On rigid circuit graphs, Abh. Math. Sem. Univ. Hamburg, 25
(1961), pp. 71 -- 76.
[19] R. G. Downey and M. R. Fellows, Parameterized complexity, Springer-Verlag,
New York, 1999.
[20] T. Feder, H. Mannila, and E. Terzi, Approximating the minimum chain
completion problem, Inf. Process. Lett., 109 (2009), pp. 980 -- 985.
[21] J. Flum and M. Grohe, Parameterized Complexity Theory, Texts in Theoretical
Computer Science. An EATCS Series, Springer-Verlag, Berlin, 2006.
21
[22] F. V. Fomin and D. Kratsch, Exact Exponential Algorithms, An EATCS Series:
Texts in Theoretical Computer Science, Springer, 2010.
[23] F. V. Fomin, D. Kratsch, I. Todinca, and Y. Villanger, Exact algorithms
for treewidth and minimum fill-in, SIAM J. Comput., 38 (2008), pp. 1058 -- 1079.
[24] F. V. Fomin and Y. Villanger, Finding induced subgraphs via minimal trian-
gulations, in STACS, vol. 5 of LIPIcs, Schloss Dagstuhl - Leibniz-Zentrum fuer
Informatik, 2010, pp. 383 -- 394.
[25] D. R. Fulkerson and O. A. Gross, Incidence matrices and interval graphs,
Pacific J. Math., 15 (1965), pp. 835 -- 855.
[26] M. R. Garey and D. S. Johnson, Computers and Intractability, A Guide to
the Theory of NP-Completeness, W.H. Freeman and Company, New York, 1979.
[27] D. Geman, Random fields and inverse problems in imaging, in ´Ecole d'´et´e de
Probabilit´es de Saint-Flour XVIII -- 1988, vol. 1427 of Lecture Notes in Math.,
Springer, Berlin, 1990, pp. 113 -- 193.
[28] M. C. Golumbic, Algorithmic Graph Theory and Perfect Graphs, Academic
Press, New York, 1980.
[29] M. C. Golumbic, H. Kaplan, and R. Shamir, Graph sandwich problems, J.
Algorithms, 19 (1995), pp. 449 -- 473.
[30] P. Heggernes, Minimal triangulations of graphs: A survey, Discrete Mathemat-
ics, 306 (2006), pp. 297 -- 317.
[31] P. Heggernes, C. Paul, J. A. Telle, and Y. Villanger, Interval completion
with few edges, in STOC, 2007, pp. 374 -- 381.
[32] P. Heggernes, J. A. Telle, and Y. Villanger, Computing minimal trian-
gulations in time O(nα log n) = o(n2.376), SIAM J. Discrete Math., 19 (2005),
pp. 900 -- 913.
[33] R. Impagliazzo, R. Paturi, and F. Zane, Which problems have strongly expo-
nential complexity?, J. Comput. Syst. Sci., 63 (2001), pp. 512 -- 530.
[34] M. Jurdzinski, M. Paterson, and U. Zwick, A deterministic subexponential
algorithm for solving parity games, SIAM J. Comput., 38 (2008), pp. 1519 -- 1532.
[35] S. Kannan and T. Warnow, Triangulating 3-colored graphs, SIAM J. Discrete
Math., 5 (1992), pp. 249 -- 258.
[36] H. Kaplan, R. Shamir, and R. E. Tarjan, Tractability of parameterized com-
pletion problems on chordal and interval graphs: Minimum fill-in and physical
mapping, in FOCS, IEEE, 1994, pp. 780 -- 791.
[37] H. Kaplan, R. Shamir, and R. E. Tarjan, Tractability of parameterized com-
pletion problems on chordal, strongly chordal, and proper interval graphs, SIAM
J. Comput., 28 (1999), pp. 1906 -- 1922.
[38] T. Kloks, D. Kratsch, and J. Spinrad, On treewidth and minimum fill-in of
asteroidal triple-free graphs, Theor. Comput. Sci., 175 (1997), pp. 309 -- 335.
[39] D. Kratsch and J. Spinrad, Between O(nm) and O(nα), SIAM J. Comput.,
36 (2006), pp. 310 -- 325.
[40] S. L. Lauritzen and D. J. Spiegelhalter, Local computations with proba-
bilities on graphical structures and their application to expert systems, J. Royal
Statist. Soc. B, 50 (1988), pp. 157 -- 224.
22
[41] H. Mannila and E. Terzi, Nestedness and segmented nestedness, in KDD '07,
New York, NY, USA, 2007, ACM, pp. 480 -- 489.
[42] A. Natanzon, R. Shamir, and R. Sharan, A polynomial approximation algo-
rithm for the minimum fill-in problem, in STOC, 1998, pp. 41 -- 47.
[43] A. Natanzon, R. Shamir, and R. Sharan, A polynomial approximation algo-
rithm for the minimum fill-in problem, SIAM J. Comput., 30 (2000), pp. 1067 --
1079.
[44] T. Ohtsuki, L. K. Cheung, and T. Fujisawa, Minimal triangulation of a graph
and optimal pivoting ordering in a sparse matrix, J. Math. Anal. Appl., 54 (1976),
pp. 622 -- 633.
[45] S. Parter, The use of linear graphs in Gauss elimination, SIAM Review, 3
(1961), pp. 119 -- 130.
[46] B. Patterson and W. Atmar, Nested subsets and the structure of insular mam-
malian faunas and archipelagos, Biological Journal of the Linnean Society, 28
(1986).
[47] D. J. Rose, A graph-theoretic study of the numerical solution of sparse positive
definite systems of linear equations, in Graph Theory and Computing, R. C. Read,
ed., Academic Press, New York, 1972, pp. 183 -- 217.
[48] D. J. Rose, R. E. Tarjan, and G. S. Lueker, Algorithmic aspects of vertex
elimination on graphs, SIAM J. Comput., 5 (1976), pp. 266 -- 283.
[49] R. E. Tarjan and M. Yannakakis, Simple linear-time algorithms to test
chordality of graphs, test acyclicity of hypergraphs, and selectively reduce acyclic
hypergraphs, SIAM J. Comput., 13 (1984), pp. 566 -- 579.
[50] Y. Villanger, P. Heggernes, C. Paul, and J. A. Telle, Interval completion
is fixed parameter tractable, SIAM J. Comput., 38 (2009), pp. 2007 -- 2020.
[51] S. Wong, D. Wu, and C. Butz, Triangulation of bayesian networks: A re-
lational database perspective, in Rough Sets and Current Trends in Computing,
vol. 2475 of LNCS, Springer, 2002, pp. 950 -- 950.
[52] M. Yannakakis, Computing the minimum fill-in is NP-complete, SIAM J. Alg.
Disc. Meth., 2 (1981), pp. 77 -- 79.
23
|
1904.12804 | 1 | 1904 | 2019-04-29T16:39:50 | The I/O complexity of hybrid algorithms for square matrix multiplication | [
"cs.DS"
] | Asymptotically tight lower bounds are derived for the I/O complexity of a general class of hybrid algorithms computing the product of $n \times n$ square matrices combining ``\emph{Strassen-like}'' fast matrix multiplication approach with computational complexity $\Theta{n^{\log_2 7}}$, and ``\emph{standard}'' matrix multiplication algorithms with computational complexity $\Omega\left(n^3\right)$. We present a novel and tight $\Omega\left(\left(\frac{n}{\max\{\sqrt{M},n_0\}}\right)^{\log_2 7}\left(\max\{1,\frac{n_0}{M}\}\right)^3M\right)$ lower bound for the I/O complexity a class of ``\emph{uniform, non-stationary}'' hybrid algorithms when executed in a two-level storage hierarchy with $M$ words of fast memory, where $n_0$ denotes the threshold size of sub-problems which are computed using standard algorithms with algebraic complexity $\Omega\left(n^3\right)$.
The lower bound is actually derived for the more general class of ``\emph{non-uniform, non-stationary}'' hybrid algorithms which allow recursive calls to have a different structure, even when they refer to the multiplication of matrices of the same size and in the same recursive level, although the quantitative expressions become more involved. Our results are the first I/O lower bounds for these classes of hybrid algorithms. All presented lower bounds apply even if the recomputation of partial results is allowed and are asymptotically tight.
The proof technique combines the analysis of the Grigoriev's flow of the matrix multiplication function, combinatorial properties of the encoding functions used by fast Strassen-like algorithms, and an application of the Loomis-Whitney geometric theorem for the analysis of standard matrix multiplication algorithms.
Extensions of the lower bounds for a parallel model with $P$ processors are also discussed. | cs.DS | cs |
The I/O complexity of hybrid algorithms
for square matrix multiplication
Lorenzo De Stefani∗
Department of Computer Science, Brown University
April 30, 2019
Abstract
Asymptotically tight lower bounds are derived for the I/O complexity of a general
class of hybrid algorithms computing the product of n × n square matrices combin-
ing "Strassen-like" fast matrix multiplication approach with computational complexity
Θ(cid:0)nlog2 7(cid:1), and "standard " matrix multiplication algorithms with computational com-
M(cid:19)
plexity Ω(cid:0)n3(cid:1). We present a novel and tight Ω(cid:18)(cid:16)
lower bound for the I/O complexity a class of "uniform, non-stationary" hybrid algo-
rithms when executed in a two-level storage hierarchy with M words of fast memory,
where n0 denotes the threshold size of sub-problems which are computed using standard
M }(cid:1)3
(cid:0)max{1, n0
n
max{√M ,n0}(cid:17)log2 7
algorithms with algebraic complexity Ω(cid:0)n3(cid:1).
The lower bound is actually derived for the more general class of "non-uniform, non-
stationary" hybrid algorithms which allow recursive calls to have a different structure,
even when they refer to the multiplication of matrices of the same size and in the same
recursive level, although the quantitative expressions become more involved. Our results
are the first I/O lower bounds for these classes of hybrid algorithms. All presented
lower bounds apply even if the recomputation of partial results is allowed and are
asymptotically tight.
The proof technique combines the analysis of the Grigoriev's flow of the matrix
multiplication function, combinatorial properties of the encoding functions used by fast
Strassen-like algorithms, and an application of the Loomis-Whitney geometric theorem
for the analysis of standard matrix multiplication algorithms. Extensions of the lower
bounds for a parallel model with P processors are also discussed.
∗[email protected]
1
Introduction
Data movement plays a critical role in the performance of computing systems, in terms of
both time and energy. This technological trend [28] appears destined to continue, as physical
limitations on minimum device size and on maximum message speed lead to inherent costs
when moving data, whether across the levels of a hierarchical memory system or between
processing elements of a parallel system [10]. While the communication requirements of
algorithms have been widely investigated in literature, obtaining significant and tight lower
bounds based on such requirements remains an important and challenging task.
In this paper, we focus on the I/O complexity of a general class of hybrid algorithms for
the computing the product of square matrices which combine. Such algorithms combine fast
algorithms with base case 2 × 2 similar to Strassen's matrix multiplication algorithm [35]
with algebraic (or computational) complexity O(cid:0)nlog2 7(cid:1) with standard (or classic) matrix
multiplication algorithms with algebraic complexity Ω(cid:0)n3(cid:1). Further, these algorithms allow
recursive calls to have a different structure, even when they refer to the multiplication of
matrices in the same recursive level and of the same input size. These algorithms are
referred in literature as "non-uniform, non-stationary". This class includes, for example,
algorithms that optimize for input sizes [15, 16, 20]. Matrix multiplication is a pervasive
primitive utilized in many applications.
While of actual practical importance, to the best of our knowledge, no characterization
of the I/O complexity of such algorithms has presented before this work. This is likely due to
the the fact that the irregular nature of hybrid algorithms and, hence, the irregular structure
of the corresponding Computational Directed Acyclic Graphs (CDAGs), complicates the
analysis of the combinatorial properties of the CDAG which is the foundation of many of
I/O lower bound technique presented in literature (e.g., [8, 18, 29]).
The technique used in this work overcomes such challenges and yields asymptotically
tight I/O lower bounds which hold even if recomputation of intermediate values is allowed.
Previous and Related Work: Strassen [35] showed that two n × n matrices can be
multiplied with O(nω) operations, where ω = log2 7 ≈ 2.8074, hence with asymptotically
fewer than the n3 arithmetic operations required by the straightforward implementation of
the definition of matrix multiplication. This result has motivated a number of efforts which
have lead to increasingly faster algorithms, at least asymptotically, with the current record
being at ω < 2.3728639 [24].
I/O complexity has been introduced in the seminal work by Hong and Kung [18]; it is
essentially the number of data transfers between the two levels of a memory hierarchy with a
fast memory of M words and a slow memory with an unbounded number of words. Hong and
Kung presented techniques to develop lower bounds to the I/O complexity of computations
modeled by computational directed acyclic graphs (CDAGs). The resulting lower bounds
apply to all the schedules of the given CDAG, including those with recomputation, that
is, where some vertices of the CDAG are evaluated multiple times. Among other results,
they established a Ω(cid:16)n3/√M(cid:17) lower bound to the I/O complexity of standard, definition-
based matrix multiplication algorithms, which matched a known upper bound [13]. The
techniques of [18] have also been extended to obtain tight communication bounds for the
definition-based matrix multiplication in some parallel settings [4, 21, 33] and for the special
1
case of "sparse matrix multiplication" [27]. Ballard et al. generalized the results on matrix
multiplication of Hong and Kung [18] in [7, 6] by using the approach proposed in [21] based
on the Loomis-Whitney geometric theorem [25, 36].
In an important contribution, Ballard et al. [8], obtained an Ω((n/√M )log2 7M ) I/O
lower bound for Strassen's algorithm, using the "edge expansion approach". The authors
extend their technique to a class of "Strassen-like" fast multiplication algorithms and to
fast recursive multiplication algorithms for rectangular matrices [5]. This result was later
generalized to increasingly broader classes of "Strassen-like" algorithms by Scott et. al [31]
using the "path routing" technique, and De Stefani [14] using a combination the concept
of Grigoriev's flow of a function and the "dichotomy width" technique [9]. While the pre-
viously mentioned results hold only under the restrictive assumption that no intermediate
result may be more than once (i.e., the no-recomputation assumption), in [11] Bilardi and
De Stefani introduced the first asymptotically tight I/O lower bound which holds if re-
computation is allowed. Their technique was later extended to the analysis of Strassen-like
algorithms with base case 2×2 [26], and to the analysis of Toom-Cook integer multiplication
algorithms [12].
A parallel, "communication avoiding" implementation of Strassen's algorithm whose
performance matches the known lower bound [8, 31], was proposed by Ballard et al. [3]. A
communication efficient algorithm for the special case of sparse matrices based on Strassen's
algorithm was presented in [22].
In [32], Scott derived a lower bound for the I/O complexity of a class of uniform, non-
stationary algorithms combining Strassen-like algorithm with recursive standard algorithms.
This result holds only under the restrictive no-recomputation assumption.
To the best of our knowledge, ours is the first work presenting asymptotically tight I/O
lower bounds for non-uniform, non-stationary hybrid algorithms for matrix multiplication
that hold when recomputation is allowed.
Our results: We present the first I/O lower bound for a class H of non-uniform, non-
stationary hybrid matrix multiplication algorithms when executed in a two-level storage
hierarchy with M words of fast memory. Algorithms in H combine fast Strassen-like algo-
rithms with base case 2 × 2 with algebraic complexity Θ(cid:0)nlog2 7(cid:1), and standard algorithms
based on the definition with algebraic complexity Ω(cid:0)n3(cid:1). These algorithms allow recursive
calls to have a different structure, even when they refer to the multiplication of matrices in
the same recursive level and of the same input size. The result in Theorem 7 relates the I/O
complexity of algorithms in H to the number and the input size of an opportunely selected
set of the sub-problems generated by the algorithms themselves.
n
max{√M,n0}(cid:17)log2 7
M }(cid:1)3 M(cid:19) lower
bound for the I/O complexity of algorithms in a subclass UH (n0) of H composed by uniform
non-stationary hybrid algorithms where n0 denotes the threshold size of sub-problems which
We also present, in Theorem 9, a novel Ω(cid:18)(cid:16)
(cid:0)max{1, n0
are computed using standard algorithms with algebraic complexity Ω(cid:0)n3(cid:1).
The previous result by Scott [32] covers only a sub-class of UH (n0) composed by uniform,
non-stationary algorithms combining Strassen-like algorithms with the recursive standard
algorithm, and holds only assuming that no intermediate value is recomputed. Instead, all
our bounds allow for recomputation of intermediate values and are asymptotically tight.
2
As the matching upper bounds do not recompute any intermediate value, we conclude that
using recomputation may reduce the I/O complexity of the considered classes of hybrid
algorithms by at most a constant factor.
Our proof technique is of independent interest since it exploits to a significant extent
the "divide and conquer " nature exhibited by many algorithms. Our approach combines
elements from the "G-flow " I/O lower bound technique originally introduced by Bilardi
and De Stefani, with an application of the Loomis-Whitney geometric theorem, which has
been used by Irony et al. to study the I/O complexity of standard matrix multiplication
algorithms [21], to recover an information which relates to the concept of Minimum set
introduced in Hong and Kung's method. We follow the dominator set approach pioneered
by Hong and Kung in [18]. However, we focus the dominator analysis only on a select set
of target vertices, which, depending on the algorithm structure, correspond either to the
outputs of the sub-CDAGs that correspond to sub-problems of a suitable size (i.e., chosen
as a function of the fast memory capacity M ) computed using a fast Strassen-like algorithm,
or to the the vertices corresponding to the elementary products evaluated by a standard
(definition) matrix multiplication algorithm.
We derive our main results for the hierarchical memory model (or external memory
model). Our results generalize to parallel models with P processors. For these parallel
models, we derive lower bounds for the "bandwith cost", that is for the number of messages
(and, hence, the number of memory) that must be either sent or received by at least one
processor during the CDAG evaluation.
Paper organization:
In Section 2 we outline the notation and the computational models
used in the rest of the presentation. In Section 3 we rigorously define the class of hybrid ma-
trix multiplication algorithms H being considered. In Section 4 we discuss the construction
and important properties of the CDAGs corresponding to algorithms in H. In Section 5 we
introduce the concept of Maximal Sup-Problem (MSP) and describe their properties which
lead to the I/O lower bounds for algorithms in H in Section 6.
2 Preliminaries
We consider algorithms that compute the product of two square matrices A × B = C with
entries from a ring R. We use A to denote the set variables each corresponding to an entry
of matrix A. We refer to the number of entries of a matrix A as its "size" and we denote
it as A. We denote the entry on the i-th row of the j-th column of matrix A as A[i][j].
In this work we focus on algorithms whose execution can be modeled as a computational
directed acyclic graph (CDAG) G = (V, E). Each vertex v ∈ V represents either an input
value or the result of a unit-time operation (i.e., an intermediate result or one of the output
values) which is stored using a single memory word. For example, each of the input (resp.,
output) vertices of G corresponds to one of the 2n2 entries of the factor matrices A and
B (resp., to the n2 entries of the product matrix C). The directed edges in E represent
data dependencies. That is, a pair of vertices u, v ∈ V are connected by an edge (u, v)
directed from u to v if and only if the value corresponding to u is an operand of the unit
time operation which computes the value corresponding to v. A directed path connecting
3
vertices u, v ∈ V is an ordered sequence of vertices starting with u and ending with v, such
that there is an edge in E directed from each vertex in the sequence to its successor.
We say that G′ = (V ′, E′) is a sub-CDAG of G = (V, E) if V ′ ⊆ V and E′ ⊆ E∩(V ′×V ′).
Note that, according to this definition, every CDAG is a sub-CDAG of itself. We say that
two sub-CDAGs G′ = (V ′, E′) and G′′ = (V ′′, E′′) of G are vertex disjoint if V ′ ∩ V ′′ = ∅.
Analogously, two directed paths in G are vertex disjoint if they do not share any vertex.
When analyzing the properties of CDAGs we make use of the concept of dominator set
originally introduced in [18]. We use the following -- slightly different -- definition:
Definition 1 (Dominator set). Given a CDAG G = (V, E), let I ⊂ V denote the set of its
input vertices. A set D ⊆ V is a dominator set for V ′ ⊆ V with respect to I′ ⊆ I if every
path from a vertex in I′ to a vertex in V ′ contains at least a vertex of D. When I′ = I, D
is simply referred as "a dominator set for V ′".
I/O model and machine models: We assume that sequential computations are exe-
cuted on a system with a two-level memory hierarchy, consisting of a fast memory or cache
of size M (measured in memory words) and a slow memory of unlimited size. An operation
can be executed only if all its operands are in cache. We assume that each entry of the input
and intermediate results matrices (including entries of the output matrix) is maintained in
a single memory word (the results trivially generalize if multiple memory words are used).
Data can be moved from the slow memory to the cache by read operations and, in the
other direction, by write operations. These operations are also called I/O operations. We
assume the input data to be stored in slow memory at the beginning of the computation.
The evaluation of a CDAG in this model can be analyzed by means of the "red-blue pebble
game" [18]. The number of I/O operations executed when evaluating a CDAG depends
on the "computational schedule", that is, it depends on the order in which vertices are
evaluated and on which values are kept in/discarded from cache.
The I/O complexity IOG(M ) of a CDAG G is defined as the minimum number of I/O
operations over all possible computational schedules. We further consider a generalization
of this model known as the "External Memory Model" by Aggarwal and Vitter [2], where
B ≥ 1 values can be moved between cache and consecutive slow memory locations with a
single I/O operation. For B = 1, this model clearly reduces to the red-blue pebble game.
Given an algorithm A, we only consider "parsimonious execution schedules", that is
schedules such that: (i) each time an intermediate result (excluding the output entries of
C) is computed, such value is then used to computed to compute at least one of the values
of which it is an operand before being removed from the memory (either the cache or slow
memory); and (ii) any time an intermediate result is read from slow to cache memory,
such value is then used to computed to compute at least one of the values of which it
is an operand before being removed from the memory or moved back to slow memory
using a write I/O operation. Clearly, any non-parsimonious schedule C can be reduced
to a parsimonious schedule C′ by removing all the steps which violate the definition of
parsimonious computation. C′ has therefore less computational and I/O operations than C.
Hence, restricting the analysis to parsimonious computations leads to no loss of generality.
We also consider a parallel model where P processors, each with a local memory of size
2n2/P ≤ M < n2, are connected by a network. We do not, however, make any assumption
4
on the initial distribution of the input data nor regarding the balance of the computational
load among the P processors. Processors can exchange point-to-point messages, with every
message containing up to Bm memory words. For this parallel model, we derive lower
bounds for the number of messages that must be either sent or received by at least one
processor during the CDAG evaluation. The notion of "parsimonious execution schedules"
straightforwardly extends to this parallel model.
3 Hybrid matrix multiplication algorithms
In this work, we consider a family of hybrid matrix multiplication algorithms obtained by
hybridizing the two following classes of algorithms:
Standard matrix multiplication algorithms: This class includes all the square matrix
multiplication algorithms which, given the input factor matrices A, B ∈ Rn×n, satisfy the
following properties:
• The n3 elementary products A[i][j]B[j][i], for i, j = 0, . . . , n−1, are directly computed;
• Each of the C[i][j] is computed by summing the values of the n elementary prod-
ucts A[i][z]B[z][j], for z = 0, . . . , n − 1, through a summation tree by additions and
subtractions only;
• The evaluations of the C[i][j]'s are independent of each other. That is, internal vertex
sets of the summation trees of all the C[i][j]'s are disjoint from each other.
This class, also referred in literature as classic, naive or conventional algorithms, correspond
to that studied for the by Hong and Kung [18] (for the sequential setting) and by Irony et
al. [21] (for the parallel setting). Algorithms in this class have computational complexity
Ω(cid:0)n3(cid:1). This class includes, among others, the sequential iterative definition algorithm, the
sequential recursive divide and conquer algorithm based on block partitioning, and parallel
algorithms such as Cannon's "2D " algorithm [13], the "2.5D " algorithm by Solomonik and
Demmel [34], and "3D " algorithms [1, 23].
Fast Strassen-like matrix multiplication algorithms with base case 2×2: This class
includes algorithms following a structure similar to that of Strassen's [35] (see Appendix B)
and Winograd's variation [37] (which reduces the leading coefficient of the arithmetic com-
plexity reduced from 7 to 6). Algorithms in this class generally follow three steps:
1. Encoding: Generate the inputs, of size n/2 × n/2 of seven sub-problems, as linear
sums of the input matrices;
2. Recursive multiplications: Compute (recursively) the seven generated matrix mul-
tiplication sub-products;
3. Decoding: Computing the entries of the product matrix C via linear combinations
of the output of the seven sub-problems.
5
Algorithms in this class have algebraic complexity O(cid:0)nlog2 7(cid:1), which is optimal for algo-
rithms with base case 2 × 2 [37].
Remarkably, the only properties of relevance for the analysis of the I/O complexity of
algorithms in these classes are those used in the characterization of the classes themselves.
In this work we consider a general class of non-uniform, non-stationary hybrid square
matrix multiplication algorithms, which allow mixing of schemes from the fast Strassen-
like class with algorithms from the standard class. Given an algorithm A let P denote
the problem corresponding to the computation of the product of the input matrices A and
B. Consider an "instruction function" fA(P ), which, given as input P returns either (a)
indication regarding the algorithm from the standard class which is to be used to compute
P , or (b) indication regarding the fast Strassen-like algorithm to be used to recursively
generate seven sub-problems P1, P2, . . . , P7 and the instruction functions fA(Pi) for each of
the seven sub-problems. We refer to the class of non-uniform, non-stationary algorithms
which can be characterized by means of such instruction functions as H. Algorithms in H
allow recursive calls to have a different structure, even when they refer to the multiplication
of matrices in the same recursive level. E.g., some of the sub-problems with the same size
may be computed using algorithms form the standard class while others may be computed
using recursive algorithms from the fast class. This class includes, for example, algorithms
that optimize for input sizes, (for sizes that are not an integer power of a constant integer).
We also consider a sub-class UH (n0) of H constituted by uniform, non-stationary hybrid
algorithms which allow mixing of schemes from the fast Strassen-like class for the initial
ℓ recursion levels, and then cut the recursion off once the size the generated sub-problems
is smaller or equal to a set threshold n0 × n0, and switch to using algorithm form the
standard class. Algorithms in this class are uniform, i.e., sub-problems of the same size
are all either recursively computed using a scheme form the fast class, or are all computed
using algorithms from the standard class.
This corresponds to actual practical scenarios, as the use of Strassen-like algorithms is
mostly beneficial for large input size. As the size of the input of the recursively generated
sub-problems decreases, the asymptotic benefit of fast algorithms is lost due to the increasing
relative impact of the constant multiplicative factor, and algorithms in the standard class
exhibit lower actual algebraic complexity. For a discussion on such hybrid algorithms and
their implementation issues we refer the reader to [16, 20] (sequential model) and [15]
(parallel model).
4 The CDAG of algorithms in H
Let GA = (VA, EA) denote the CDAG that corresponds to an algorithm A ∈ H used to
multiply input matrices A, B ∈ Rn×n. The challenge in the characterization of GA comes
from the fact that rather than considering to a single algorithm, we want to characterize the
CDAG corresponding to the class H. Further, the class H is composed by a rich variety of
vastly different and irregular algorithm. Despite such variety, we show a general template for
the construction of GA and we identify some of it properties which crucially hold regardless
of the implementation details of A and, hence, of GA.
6
C[0][0] C[0][1] C[1][0] C[1][1]
C1,1 C1,2 C2,1 C2,2
Dec
M7 M5 M4 M1 M3 M2 M6
n2 × Dec
GAP7
GAP5
GAP4
GAP1
GAP3
GAP2
GAP6
EncA
EncB
n2 × EncA
n2 × EncB
A[0][0] A[0][1] A[1][0] A[1][1]
B[0][0] B[0][1] B[1][0] B[1][1]
A1,1 A1,2 A2,1 A2,2
B1,1 B1,2 B2,1 B2,2
(a) GA CDAG for base case n = 2, using
Strassen's algorithm [35] (see Appendix B).
(b) Recursive construction of GA. Ai,j, Bi,j and
Ci,j denote the block-partition of A, B and C.
Figure 1: Blue vertices represent combinations of the input values from the factor matrices
A and B used as input values for the seven sub-problems; red vertices represent the output
of the seven sub-problems which are used to compute the values of the output matrix C.
Construction: GA can be obtained by using a recursive construction that mirrors the
recursive structure of the algorithm itself. Let P denote the entire matrix multiplication
problem computed by A. Consider the case for which, according to the instruction function
fA(P ), P is to be computed using an algorithm from the standard class. As we do not fix
a specific algorithm, we do not correspondingly have a fixed CDAG. The only feature of
interest for the analysis is that, in this case, the CDAG GA corresponds to the execution of
an algorithm from the standard class for input matrices of size n × n.
Consider instead the case for which, according to fA(P ), P is to be computed using
an algorithm from the fast class. In the base case for n = 2 the problem P is computed
without generating any further sub-problems. As an example, we present in Figure 1a the
base case for Strassen's original algorithm [35]. If n > 2, then fA(P ) specifies the divide
and conquer scheme to be used to generate the seven sub-problems P1, P2, . . . , P7, and the
instruction function for each of them. The sub-CDAGs of GA corresponding to each of
the seven sub-problems Pi, denoted as GAPi
are constructed according to fA(Pi), following
recursively the steps discussed previously. GA can then be constructed by composing the
seven sub-CDAGs GAPi
. n2 disjoint copies of an encoder sub-CDAG EncA (resp., EncB)
are used to connect the input vertices of G2n×2n, which correspond to the values of the
input matrix A (resp., B) to the appropriate input vertices of the seven sub-CDAGs GAPi
;
the output vertices of the sub-CDAGs GAPi
(which correspond to the outputs of the seven
sub-products) are connected to the appropriate output vertices of the entire GA CDAG
using n2 copies of the decoder sub-CDAG Dec. We present an example of such recursive
construction in Figure 1b.
Properties of GA: While the actual internal structure GA, and, in particular, the struc-
ture of encoder and decoder sub-CDAGs depends on the specific Strassen-like algorithm
being used by A, all versions share some properties of great importance. Let G(X, Y, E) de-
note an encoder CDAG for a fast multiplication algorithm 2×2 base case, with X (resp., Y )
denoting the set of input (resp., output) vertices, and E denoting the set of edges directed
from X to Y .
7
Lemma 1 (Lemma 3.3 [26]). Let G = (X, Y, E) denote an encoder graph for a fast matrix
multiplication algorithm with base case 2 × 2. There are no two vertices in Y with identical
neighbors sets.
While the correctness of this Lemma can be simply verified by inspection in the case of
Strassen's algorithm [35], Lemma 1 generalizes the statement to all encoders corresponding
to fast matrix multiplication algorithms with base case 2 × 2. From Lemma 1 we have:
Lemma 2. Let A ∈ H and let P1 and P2 be any two sub-problems generated by A with
input size greater than n0 × n0, such that P2 is not recursively generated while computing
P1 and vice versa. Then, the sub-CDAGs of GA corresponding, respectively, to P1 and to
P2 are vertex disjoint.
The following lemma, originally introduced for Strassen's algorithm in [11] and then
generalized for Strassen-like algorithms with base case 2×2 in in [26], captures a connectivity
property of encoder sub-CDAGs.
Lemma 3 (Lemma 3.1 [26]). Given an encoder CDAG for any Strassen-like algorithm
with base case 2 × 2, for any subset Y of its output vertices, there exists a subset X of
its input vertices, with min{Y , 1 + ⌈(Y − 1) /2⌉} ≤ X ≤ Y , such that there exist X
vertex-disjoint paths connecting the vertices in X to vertices in Y .
The proofs of Lemma 1 and Lemma 3 are based on an argument originally presented by
Hopcroft and Kerr [19]. We refer the reader to [26] for the proofs.
5 Maximal sub-problems and their properties
For an algorithm A ∈ H, let P ′ denote a sub-problem generated by A. In our presentation we
consider the entire matrix multiplication problem an improper sup-problem generated by A.
We refer as the ancestor sub-problems of P ′ as the sequence of sub-problems P ′0, P ′2, . . . , P ′i
generated by A such that P ′j+1 was recursively generated to compute P ′j for j = 0, 1, . . . , i−1,
and such that P was recursively generated to compute P ′i . Clearly, if P ′ is the entire
problem, then P ′ has no ancestors.
Towards studying the I/O complexity of algorithms in H we focus on the analysis of a
particular set of sub-problems.
Definition 2 (Maximal Sub-Problems (MSP)). Let A ∈ H be an algorithm used to multiply
matrices A, B ∈ Rn×n. If n ≤ 2√M we say that A does not generate any Maximal Sub-
Problem (MSP).
Let Pi be a sub-problem generated by A with input size ni × ni, with ni ≥ 2M , and such
that all its ancestors sub-problems are computed, according to A using algorithms from the
fast class. We say that:
• Pi is a Type 1 MSP of A if, according to A, is computed using an algorithm from
the standard class. If the entire problem is to be solved using an algorithm for the
standard class, we say that the entire problem is the unique (improper) Type 1 MSP
generated by A.
8
• Pi is a Type 2 MSP of A if, according to A, is computed by generating 7 sub-
problems according to the recursive scheme corresponding to an algorithm from the
fast (Strassen-like) class, and if the generated sub-problems have input size strictly
smaller than 2√M × 2√M . If the entire problem uses a recursive algorithm from the
fast class to generate 7 sub-problems with input size smaller than 2√M × 2√M , we
say that the entire problem is the unique, improper, Type 2 MSP generated by A.
In the following we denote as ν1 (resp., ν2) the number of Type 1 (resp., Type 2) MSPs
generated by A. Let Pi denote the i-th MSP generated by A and let GAPi
denote the
corresponding sub-CDAG of GA. We denote as Ai and Bi (resp., Ci) the input factor
matrices (resp., the output product matrix) of Pi.
Properties of MSPs and their corresponding sub-CDAGs: By Definition 2, we
have that for each pair of distinct MMSPs P1 and P2, P2 is not recursively generated by
A in order to compute P1 or vice versa. Hence, by Lemma 2, the sub-CDAGs of GA that
correspond each to one of the MSPs generated by A are vertex, disjoint.
In order to obtain our I/O lower bound for algorithms in H, we characterize properties
regarding the minimum dominator size of an arbitrary subset of Y and Z.
Lemma 4 (Proof in Appendix A.1). Let GA be the CDAG corresponding to an algorithm
A ∈ H which admits n1 Type 1 MSPs. For each Type 1 MSP Pi let Y i denote the set of
input vertices of the associated sub-CDAG GAPi
which correspond each to an entry of the
input matrices Ai and Bi. Further, we define Y = ∪ν1
Let Y ⊆ Y in GA such that Y ∩ Y i = ai/√bi, with ai, bi ∈ N, ai ≥ bi for i =
1, 2, . . . , ν1, and such that bi = 0 if and only if ai = 0.1 Any dominator set D of Y satisfies
D ≥ min{2M,Pν1
Lemma 5 (Proof in Appendix A.2). Let GA be the CDAG corresponding to an algorithm
A ∈ H which admits n2 Type 2 MSPs. Further let Z denote the set of vertices corresponding
to the entries of the output matrices of the n2 Type 2 MSPs. Given any subset Z ⊆ Z in
GA with Z ≤ 4M , any dominator set D of Z satisfies D ≥ Z/2.
i=1 ai/pPν1
i=1Y i.
i=1 bi}.
For each Type 1 MSP Pi generated by A, with input size2 ni × ni, we denote as Ti the
set of variables whose value correspond to the n3
i elementary products Ai[j][k]Bi[k][j] for
j, k = 0, 1, . . . , ni − 1. Further, we denote as Ti the set of vertices corresponding to the
variables in Ti, and we define T = ∪ν1
Lemma 6 (Proof in Appendix A.3). For any Type 1 MSPs generated by A consider T ′i ⊆ Ti.
Let Y (A)
i ⊆ Y i) denote a subset of the vertices corresponding to entries of
Ai (resp., Bi) which are multiplied in at least one of the elementary products in T ′i . Then
any dominator D of the vertices corresponding to T ′i with respect to the the vertices in Y i
is such that
i ⊆ Y i (resp., Y (B)
i=1Ti.
1Here we use as convention that 0/0 = 0.
2In general, different Type 1 MSP may have different input sizes
D ≥ max{Y′i ∩ Ai,Y′i ∩ Bi}.
9
6
I/O lower bounds for algorithm in H and UH (n0)
Theorem 7. Let A ∈ H be an algorithm to multiply two square matrices A, B ∈ Rn×n. If
run on a sequential machine with cache of size M and such that up to B memory words
stored in consecutive memory locations can be moved from cache to slow memory and vice
versa using a single memory operation, A's I/O complexity satisfies:
IOA (n, M, B) ≥ max{2n2, cT M−1/2, ν2M}B−1
(1)
for c = 0.38988157484, where T denotes the total number of internal elementary products
computed by the Type 1 MSPs generated by A and ν2 denotes the total number of Type 2
MSPs generated by A.
If run on P processors each equipped with a local memory of size M < n2 and where for
each I/O operation it is possible to move up to Bm words, A's I/O complexity satisfies:
IOA (n, M, Bm, P ) ≥ max{cT M−1/2, ν2M} (P Bm)−1
(2)
Proof. We prove the result in (1) (resp., (2)) for the case B = 1 (resp., Bm = 1). The
result then trivially generalizes for a generic B (resp., Bm). We first prove the result for
the sequential case in in (1). The bound for the parallel case in (2) will be obtained as a
simple generalization.
i=1 Ti.
The fact that IOA(n, M, 1) ≥ 2n2 follows trivially from the fact that as in our model
the input matrices A and B are initially stored in slow memory, it will necessary to move
the entire input to the cache at least once using at least 2n2 I/O operations. If A does not
generate any MSPs the statement in (1) is trivially verified. In the following, we assume
ν1 + ν2 ≥ 1.
Let GA denote the CDAG associated with algorithm A according to the construction in
Section 4. By definition, and from Lemma 2, the ν1 + ν2 sub-CDAGs of GA corresponding
each to one of the MSPs generated by A are vertex-disjoint. Hence, the Ti's are a partition
of T and T =Pν1
By Definition 2, the MSP generated by A have input (resp., output) matrices of size
greater or equal to 2√M × 2√M . Recall that we denote as Z the set of vertices which
correspond to the outputs of the ν2 Type 2 MSPs, we have Z ≥ 4M νl.
Let C be any computation schedule for the sequential execution of A using a cache of
size M . We partition C into non-overlapping segments C1,C2, . . . such that during each Cj
either (a) exactly M 3/2 distinct values 3 corresponding to vertices in T , denoted as T (j),
are explicitly computed (i.e., not loaded from slow memory), or (b) 4M distinct values
corresponding to vertices in Z (denoted as Z j) are evaluated for the first time. Clearly
there are at least max{T /M 3/2, ν2} such segments.
Below we show that the number gj of I/O operations executed during each Cj satisfies
gj ≥ cM for case (a) and gj ≥ M for case (b), from which the theorem follows.
Case (a): For each Type 1 MSP Pi let T (j)
i = T (j) ∩ Ti. As the ν1 sub-CDAGs
corresponding each to one of the Type 1 MSPs are vertex-disjoint, so are the sets Ti.
Hence, the T (j)
's constitute a partition of T (j). Let Ai and Bi (resp., Ci) denote the
i
3For simplicity of presentation, we assume M 3/2
∈ N+.
10
input matrices (resp., output matrix) of Pi with Ai, Bi, Ci ∈ Rni×ni, and let Ai and Bi
(resp., Ci) denote the set of the variables corresponding to the entries of Ai and Bi (resp.,
Ci). Further, we denote as Ti the set of values corresponding to the vertices in Ti. For
r, s = 0, 1, . . . , ni − 1, we say that Ci[r][s] is "active during Cj" if any of the elementary
multiplications Ai[r][k]Bi[k][s], for k = 0, 1, . . . , ni − 1, correspond to any of the vertices in
T (j)
. Further we say that a Ai[r][s] (resp., Bi[r][s]) is "accessed during Cj" if any of the
elementary multiplications Ai[r][s]Bi[s][k] (resp., Ai[k][r]Bi[r][s]), for k = 0, 1, . . . , ni − 1,
correspond to any of the vertices in T (j)
Our analysis makes use of the following property of standard matrix multiplication
.
i
i
algorithms:
Lemma 8 (Loomis-Whitney inequality [21, Lemma 2.2]). Let Y′i,A (resp., Y′i,B) denote
the set of vertices corresponding to the entries of Ai (resp., Bi) which are accessed during
Cj, and let Z′i denote the set of vertices corresponding to the entries of Ci which are active
during Cj. Then
T (j)
i
≤qY′i,AY′i,BZ′i.
(3)
Lemma 8 is a reworked version of a property originally presented by Irony et al. [21,
Lemma 2.2], which itself is a consequence of the Loomis-Whitney geometric theorem [25].
Let Ci[r][s] be active during Cj. In order to compute Ci[r][s] entirely during Cj (i.e.,
without using partial accumulation of the summation Pni−1
it will be
necessary to evaluate all the ni elementary products Ai[r][k]Bi[k][s], for k = 0, 1, . . . , ni− 1,
during Cj itself. Thus, at most ⌊T (j)
/ni⌋ entries of Ci[r][s] can be entirely computed during
Cj.
Let Ci[r][s] denote an entry of Ci which is active but not entirely computed during Cj.
There are two possible scenarios:
k=0 Ai[r][k]B[k][s]),
i
• Ci[r][s] is computed during Cj: The computation thus requires for a partial accumu-
lation of Pni−1
k=0 Ai[r][k]B[k][s] to have been previously computed and either held in
the cache at the beginning of Cj, or to moved to cache using a read I/O operation
during Cj;
• Ci[r][s] is not computed during Cj: As C is a parsimonious computation, the partial
accumulation of Pni−1
k=0 Ai[r][k]B[k][s] obtained from the elementary products com-
puted during Cj must either remain in the cache at the end of Cj, or be moved to slow
memory using a write I/O operation during Cj;
Let GAPi
denote the sub-CDAG of GA corresponding to the Type 1 MSP Pi.
In both cases, any partial accumulation either held in memory at the beginning (resp., end)
of Cj or read from slow memory to cache (resp., written from cache to slow memory) during
Cj is, by definition, not shared between multiple entries in Ci.
In the
following, we refer as D′i to the set of vertices of GAPi
corresponding to the values of such
partial accumulators. For each of the least D′i = max{0,Z′i − T (j)
/ni} entries of Ci
which are active but not entirely computed during Cj, either one of the entries of the cache
must be occupied at the beginning of Cj, or one I/O operation is executed during Cj. Let
i
11
i
ν1
ν1
Xi=1
Xi=1
D′ =
D′i =
i=1D′i. As, by Lemma 2, the sub-CDAGs corresponding to the ν1 Type 1 MSPs are
D′ = ∪ν1
vertex disjoint, so are the the sets D′i. Let Z′ =Pνl
max{0,Z′i − T (j)
i=1 Z′i. We have:
/ni} ≥ Z − T (j)/2√M ,
where the last passage follows from the fact that, by Definition 2, ni ≥ 2M .
From Lemma 8, the set of vertices Y′i,A (resp., Y′i,B) which correspond to entries of Ai
(resp., Bi) which are accessed during Cj satisfies Y′i,AY′i,B ≥ T (j)
2/Z′i. Hence, at least
Y′i,A +Y′i,B ≥ 2T (j)
/pZ′i entries from the input matrices of Pi are accessed during Cj.
Let Y denote the set of vertices corresponding to the entries of the input matrices Ai, Bi of
Pi. From Lemma 6 we have that there exists a set Y′i ⊆ Y i, withY′i ≥ max{Y′i,A,Y′i,B} ≥
T (j)
/pZ′i, such that the vertices in Y′i are connected by vertex disjoint pats to the vertices
in T (j)
. Let Y = ∪ν1
i=1Y′i. As, by Lemma 2, the sub-CDAGs corresponding to the νl Type
1 MSPs are vertex disjoint, so are the the sets Y′i for i = 1, 2, . . . , ν1. Hence
(4)
i
i
i
i
From Lemma 4 any dominator DY of Y , must be such that
ν1
ν1
i
T (j)
pZ′i
.
Y =
Xi=1
Y′i ≥
Xi=1
i=1 T (j)
DY ≥ min(cid:8)2M, Pν1
i=1pZ′i(cid:9) = min(cid:8)2M, T (j)
pZ′(cid:9).
Pν1
i
Hence, we can conclude that any dominator D′′ of T (j) must be such that
D′′ ≥ min(cid:8)2M,T (j)/pZ′(cid:9).
(5)
Consider the set D of vertices of GA corresponding to the at most M values stored in
the cache at the beginning of Cj and to the at most gj values loaded into the cache form
the slow memory (resp., written into the slow memory from the cache) during Cj by means
of a read (resp., write) I/O operation. Clearly, D ≤ M + gj.
In order for the M 3/2 values from T (j) to be computed during Cj there must be no path
connecting any vertex in T (j), and, hence, Y , to any input vertex of GA which does not
have at least one vertex in D, that is D has to admit a subset D′′ ⊆ D such that D′′ is a
dominator set of T (j). Note that, as the values corresponding to vertices in T (j) are actually
computed during CJ (i.e., not loaded from memory using a read I/O operation), D′′ does
not include vertices in T (j) itself. Further, as motivated in the previous discussion, D must
include all the vertices in the set D′ corresponding to values of partial accumulators of the
active output values of Type 1 MSPs during Cj.
By construction, D′ and D′′ are vertex disjoint. Hence, from (4) and (5) we have:
D ≥ D′ + D′′ ≥ Z′ − T (j)/2√M + min(cid:8)2M,T (j)/pZ′(cid:9).
As, by construction, T (j) = M 3/2, we have:
D > Z′ − M/2 + min(cid:8)2M, M 3/2/pZ′(cid:9).
12
(6)
By studying its derivative after opportunely accounting for the minimum, we have that (6)
is minimized for Z′ = 2−2/3M . Hence we have:
D > 2−2/3M + 21/3M 3/2 − M/2 =
1.38988157484M . Whence D ≤ M + gj, which implies gj ≥ D − M > 0.38988157484M ,
as stated above.
Case (b): In order for the 4M values from Z j to be computed during Cj there must
be no path connecting any vertex in Z j to any input vertex of GA which does not have
at least one vertex in Dj, that is Dj has to be a dominator set of Z j. From Lemma 5,
any dominator set D of any subset Z ⊆ Z with Z ≤ 4M satisfies D ≥ Z/2, whence
M + gi ≥ Di ≥ Z j/2 = 2M , which implies gj ≥ M as stated above. This concludes the
proof for the sequential case in (1).
The proof for the bound for the parallel model in (2), follows from the observation
that at least one of the P processors, denoted as P ∗, must compute at least T /P values
corresponding to vertices in T or Z/P values corresponding to vertices in Z (or both).
The bound follows by applying the same argument discussed for the sequential case to the
computation executed by P ∗.
Note that if A is such that the product is entirely computed using an algorithm from
the standard class (resp., a fast matrix multiplication algorithm), the bounds of Theorem 7
corresponds asymptotically to the results of [18] for the sequential case and [21] for the
parallel case (resp., the results in [11]).
The bound in (2) can be further generalized to a slightly different model in which each
of the P processors is equipped with a cache memory of size M and a slow memory of
unbounded size. In such case, the I/O complexity of the algorithms in H corresponds to the
total number of both messages received and the number of I/O operations used to move
data from cache to slow memory and vice versa (i.e., read and write) executed by at least
one of the P processors.
I/O lower bound for uniform, non stationary algorithms in UH (n0) : For the sub-
class of uniform, non stationary algorithms UH (n0) , given the values of n, M and n0 is
possible to compute a closed form expression for the values of ν1, ν2 and T .4 Then, by
applying Theorem 7 we have:
Theorem 9. Let A ∈ UH (n0) be an algorithm to multiply two square matrices A, B ∈
Rn×n. If run on a sequential machine with cache of size M and such that up to B memory
words stored in consecutive memory locations can be moved from cache to slow memory and
vice versa using a single memory operation, A's I/O complexity satisfies:
IOA (n, M, B) ≥ max{2n2,(cid:18)
n
max{n0, 2√M}(cid:19)log2 7(cid:18)max(cid:8)1,
4The constants terms in Theorem 9 assume that n, n0 and M are powers of two. If that not the case the
n0
2√M(cid:9)(cid:19)3
M}B−1
(7)
statement holds with minor adjustments.
13
If run on P processors each equipped with a local memory of size M < n2 and where for
each I/O operation it is possible to move up to Bm memory words, A's I/O complexity
satisfies:
IOA (n, M, Bm, P ) ≥(cid:18)
n
max{n0, 2√M}(cid:19)log2 7(cid:18)maxn1,
n0
2√Mo(cid:19)3 M
P Bm
.
(8)
Proof. The statement follows by bounding the values ν1,ν2 and T for A ∈ UH (n0) , and
by applying the general result from Theorem 7. In order to simplify the presentation, in
the following we assume that the values n, n0, M are powers of two. If that is not the case,
the theorem holds with some minor adjustments to the constant multiplicative factor.
H, at each of the i recursive levels A generates 7i sub-problems of size n/2i.
Let let i be the smallest value in N such that n/2i = max{n0, 2√M}. By definition of
• If n0 > 2√M , A generates
ν1 = 7i = 7log2 n/n0 =(cid:18) n
n0(cid:19)log2 7
Type 1 MSP each with input size n0 × n0. As, by Definition 2, the Type 1 MSP are
input dijoint we have:
T = ν1n0
• Otherwise, if n0 ≤ 2√M , A generates
3 =(cid:18) n
n0(cid:19)log2 7
n0
3.
ν2 = 7i = 7log2 n/2√M =(cid:18) n
2√M(cid:19)log2 7
Type 2 MSP each with input size 2√M × 2√M .
The statement then follows by applying the result in Theorem 7.
Theorem 9 extends the result by Scott [32] by expanding the class of hybrid matrix
multiplication algorithms being considered (e.g., it does not limit the class of standard
matrix multiplication to the divide and conquer algorithm based on block-partitioning),
and by removing the assumption that no intermediate value may be recomputed.
On the tightness of the bound: An opportune composition of the cache-optimal version
of the Strassen's algorithm [35] (as discussed in [3]) with the standard cache-optimal divide
and conquer algorithm for square matrix multiplication based on block-partitioning [13]
leads to a sequential hybrid algorithms in H (resp., UH (n0) ) whose I/O cost asymptotically
matches the I/O complexity lower bounds in Theorem 7 (1) (resp., Theorem 9 (7)).
Parallel algorithms in H (resp., UH (n0) ) asymptotically matching the I/O lower bounds
in for the parallel case in Theorem 7 (2) (resp., Theorem 9 (8)) can be obtained by composing
the communication avoiding version of Strassen's algorithm by Ballard et al. [3] with the
communication avoiding "2.5 " standard algorithm by Solomonik and Demmel [34].
14
Hence, the lower bounds in Theorem 7 and Theorem 9 are asymptotically tight and the
mentioned algorithms form H and UH (n0) whose I/O cost asymptotically match the lower
bounds are indeed I/O optimal.
Further, as the mentioned I/O optimal algorithms from H and UH (n0) do not recompute
any intermediate value 5, we can conclude that using recomputation may lead to at most a
constant factor reduction of the I/O cost of hybrid algorithms in H and UH (n0) .
Generalization to fast matrix multiplication model with base other than 2 × 2:
The general statement of Theorem 7 can be extended to by enriching H to include any
fast Strassen-like algorithm with base case other than 2 × 2 provided that the associated
encoder CDAG satisfies properties equivalent to those expressed by Lemma 2(i.e., the input
disjointedness of the sub-problems generated at each recursive step) and Lemma 3 (i.e., the
connectivity between input and output of the encoder CDAGs via vertex disjoint paths) for
the 2 × 2 base. If these properties hold, so does the general structure of Theorem 7, given
an opportune adjustment of the definition of maximal sub-problem.
7 Conclusion
This work introduced the first characterization of the I/O complexity of hybrid matrix mul-
tiplication algorithms combining fast Strassen-like algorithms with standard algorithms.
We established asymptotically tight lower bounds that hold even when recomputation is
allowed. The generality of the technique used the analysis makes it promising for the
analysis of other hybrid recursive algorithms, e.g., for hybrid algorithms for integer multi-
plication [12].
Our results contribute to the study of the effect of recomputation with respect to the I/O
complexity of CDAG algorithms. While we are far from a characterization of those CDAGs
for which recomputation is effective, this broad goal remains a fundamental challenge for
any attempt toward a general theory of the communication requirements of computations.
Acknowledgments
The author would like to thank Gianfranco Bilardi at the University of Padova for the
helpful suggestions and discussions.
References
[1] A. Aggarwal, B. Alpern, A. Chandra, and M Snir. A model for hierarchical memory.
In Proc. ACM STOC, pages 305 -- 314. ACM, 1987.
[2] Alok Aggarwal and S. Vitter, Jeffrey. The input/output complexity of sorting and
related problems. Commun. ACM, 31(9):1116 -- 1127, September 1988.
5We do not consider replication of the input used by the mentioned parallel algorithms as recomputation,
but rather as a repeated access to the input values.
15
[3] G. Ballard, J. Demmel, Olga H., B. Lipshitz, and O. Schwartz. Communication-optimal
parallel algorithm for Strassen's matrix multiplication. In Proc. ACM SPAA, pages
193 -- 204, 2012.
[4] G. Ballard, J. Demmel, O. Holtz, B. Lipshitz, and O. Schwartz. Brief announcement:
strong scaling of matrix multiplication algorithms and memory-independent commu-
nication lower bounds. In Proc. ACM SPAA, pages 77 -- 79. ACM, 2012.
[5] G. Ballard, J. Demmel, O. Holtz, B. Lipshitz, and O. Schwartz. Graph expansion
analysis for communication costs of fast rectangular matrix multiplication. In Design
and Analysis of Algorithms, pages 13 -- 36. Springer, 2012.
[6] G. Ballard, J. Demmel, O. Holtz, and O. Schwartz. Communication-optimal paral-
lel and sequential Cholesky decomposition. SIAM Journal on Scientific Computing,
32(6):3495 -- 3523, 2010.
[7] G. Ballard, J. Demmel, O. Holtz, and O. Schwartz. Minimizing communication in nu-
merical linear algebra. SIAM Journal on Matrix Analysis and Applications, 32(3):866 --
901, 2011.
[8] G. Ballard, J. Demmel, O. Holtz, and O. Schwartz. Graph expansion and communica-
tion costs of fast matrix multiplication. JACM, 59(6):32, 2012.
[9] G. Bilardi and F. Preparata. Processor-time trade offs under bounded speed message
propagation. Part 2: Lower Bounds. Theory of Computing Systems, 32(5):531 -- 559,
1999.
[10] G. Bilardi and F. P. Preparata. Horizons of parallel computation. Journal of Parallel
and Distributed Computing, 27(2):172 -- 182, 1995.
[11] Gianfranco Bilardi and Lorenzo De Stefani. The i/o complexity of strassen's matrix
multiplication with recomputation. In Workshop on Algorithms and Data Structures,
pages 181 -- 192. Springer, 2017.
[12] Gianfranco Bilardi and Lorenzo De Stefani. The I/O complexity of toom-cook integer
multiplication.
In Proceedings of the Thirtieth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2019, San Diego, California, USA, January 6-9, 2019,
pages 2034 -- 2052, 2019.
[13] L. E. Cannon. A cellular computer to implement the Kalman filter algorithm. Technical
report, DTIC Document, 1969.
[14] Lorenzo De Stefani. On space constrained computations. PhD thesis, University of
Padova, 2016.
[15] Fr´ed´eric Desprez and Fr´ed´eric Suter. Impact of mixed-parallelism on parallel implemen-
tations of the strassen and winograd matrix multiplication algorithms. Concurrency
and Computation: practice and experience, 16(8):771 -- 797, 2004.
16
[16] C. Douglas, M. Heroux, G. Slishman, and R. M. Smith. GEMMW: a portable level
3 BLAS Winograd variant of Strassen's matrix-matrix multiply algorithm. Journal of
Computational Physics, 110(1):1 -- 10, 1994.
[17] D. Y. Grigor'ev. Application of separability and independence notions for proving lower
bounds of circuit complexity. Zapiski Nauchnykh Seminarov POMI, 60:38 -- 48, 1976.
[18] J. Hong and H. Kung.
I/o complexity: The red-blue pebble game.
In Proc. ACM
STOC, pages 326 -- 333. ACM, 1981.
[19] John E Hopcroft and Leslie R Kerr. On minimizing the number of multiplications
necessary for matrix multiplication. SIAM Journal on Applied Mathematics, 20(1):30 --
36, 1971.
[20] Steven Huss-Lederman, Elaine M Jacobson, Jeremy R Johnson, Anna Tsao, and
Thomas Turnbull. Implementation of strassen's algorithm for matrix multiplication.
In Supercomputing'96: Proceedings of the 1996 ACM/IEEE Conference on Supercom-
puting, pages 32 -- 32. IEEE, 1996.
[21] D. Irony, S. Toledo, and A. Tiskin. Communication lower bounds for distributed-
Journal of Parallel and Distributed Computing,
memory matrix multiplication.
64(9):1017 -- 1026, 2004.
[22] R. Jacob and M. St´ockel. Fast output-sensitive matrix multiplication. In Proc. ESA,
pages 766 -- 778. Springer, 2015.
[23] S Lennart Johnsson. Minimizing the communication time for matrix multiplication on
multiprocessors. Parallel Computing, 19(11):1235 -- 1257, 1993.
[24] F. Le Gall. Powers of tensors and fast matrix multiplication. In Proc. ACM ISSAC,
pages 296 -- 303. ACM, 2014.
[25] L. H. Loomis and H. Whitney. An inequality related to the isoperimetric inequality.
Bull. Amer. Math. Soc., 55(10):961 -- 962, 10 1949.
[26] Roy Nissim and Oded Schwartz. Revisiting the i/o-complexity of fast matrix multi-
plication with recomputations. In Proceedings of the 33rd IEEE International Parallel
and Distributed Processing Symposium, pages 714 -- 716, 2019.
[27] R. Pagh and M. Stockel. The input/output complexity of sparse matrix multiplication.
In Proc. ESA, pages 750 -- 761. Springer, 2014.
[28] C. A. Patterson, M. Snir, and S. L. Graham. Getting Up to Speed:: The Future of
Supercomputing. National Academies Press, 2005.
[29] J. E. Savage. Extending the Hong-Kung model to memory hierarchies. In Computing
and Combinatorics, pages 270 -- 281. Springer, 1995.
[30] J. E. Savage. Models of Computation: Exploring the Power of Computing. Addison-
Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1st edition, 1997.
17
[31] J. Scott, O. Holtz, and O. Schwartz. Matrix multiplication I/O complexity by Path
Routing. In Proc. ACM SPAA, pages 35 -- 45, 2015.
[32] Jacob N Scott. An I/O-Complexity Lower Bound for All Recursive Matrix Multiplica-
tion Algorithms by Path-Routing. PhD thesis, UC Berkeley, 2015.
[33] M. Scquizzato and F. Silvestri. Communication lower bounds for distributed-memory
computations. arXiv preprint arXiv:1307.1805, 2013.
[34] Edgar Solomonik and James Demmel. Communication-optimal parallel 2.5 D matrix
multiplication and LU factorization algorithms. In European Conference on Parallel
Processing, pages 90 -- 109. Springer, 2011.
[35] V. Strassen. Gaussian elimination is not optimal. Numerische Mathematik, 13(4):354 --
356, 1969.
[36] Y. D. Burago V. A. Zalgaller, A. B. Sossinsky. The American Mathematical Monthly,
96(6):544 -- 546, 1989.
[37] Shmuel Winograd. On multiplication of 2× 2 matrices. Linear algebra and its appli-
cations, 4(4):381 -- 388, 1971.
18
A Proofs of technical lemmas
A.1 Proof of Lemma 4
The proof of Lemma 4 uses an analysis similar to that in [11](Lemma 6), albeit with several
important variations. Before delving into the details of the proof of Lemma 4 we introduce
the following technical lemma:
Lemma 10. Let ai, a2, . . . ai ∈ N (resp., bi, b2, . . . bi ∈ N) such that bj = 0 if and only if
ai = 0 and such that:
a1√b1 ≥
a2√b2 ≥ .... ≥
ai√bi
using the convention that 0/0 = 0. Then we have:
a1√b1
+
1
2
i
Xj=2
aj
pbj ≥ Pi
qPj
j=1 ai
j=1 bj
(9)
Proof. The proof is by induction on the value of i.
In the base case case i = 1, hence
a = a1 and b = b1. Thus the statement is trivially verified. We assume inductively that the
statement holds for i ≥ 1 and we show that it holds also for i + 1. Let
a1√b1 ≥
a2√b2 ≥ .... ≥
ai + 1
√bi + 1
be can therefore apply the inductive hypothesis to the first i elements and obtain:
1
2
i+1
Xj=2
aj
pbj ≥ Pi
qPi
j=1 aj
j=1 bj
+
1
2
.
ai+1
pbi+1
(10)
Let a′ = Pi
following we consider the case for which ai+1 > 0 and, by assumption, bi+1 > 0.
From (10), we have that the lemma is guaranteed to be verified if the following holds:
j=1 bj. If ai+1 = 0 the statement is trivially verified. In the
+
a1√b1
j=1 aj and b′ = Pi
a′
√b′
+
1
2
ai+1
pbi+1 ≥
a′ + ai+1
pb′ + bi+1
=
a′
pb′ + bi+1
+
ai+1
pb′ + bi+1
Clearly (11) is verified for b′ ≥ bi+1. Since, by assumption, b = b′ + bi+1, we can therefore
conclude that the statement is verified for b′ ≥ 3b/4.
In the following we consider the case for b′ < 3b/4. Let a′ = ax (resp., b′ = by), with
x ∈ (0, 1]. By assumption x, y 6= 0. Further, as we are considering the case such that
b′ < 3b/4, we have that y ∈ (0, 3/4). Since, by assumption, a = a′ + ai+1 and b = b′ + bi+1,
we have:
(11)
a′
√b′
+
1
2
ai+1
pbi+1
=
=
+
1
2
ax
√by
√b(cid:18) x
a
√y
a(1 − x)
pb(1 − y)
2√1 − y(cid:19)
1 − x
+
19
Therefore, the Lemma is clearly verified if the following holds:
x
√y
+
1 − x
√1 − y
=
2x√1 − y + (1 − x)√y
2√y√1 − y
≥ 1.
(12)
By multiplying both sides of the previous inequality by 2√y√y − 1 (which is clearly greater
than 0), we have that (12) holds if:
which in turn holds if:
2xp1 − y + (1 − x)√y ≥ 2√yp1 − y,
2p1 − y (x − √y) ≥ √y (x − 1) ,
(13)
(14)
Lemma follows.
As y ∈ (0, 3/4) we have 2√1 − y > 1 > √y and (cid:0)x − √y(cid:1) (x − 1). Hence (14) holds ant the
Recall that given an algorithm A ∈ H used to multiply input squared matrices A, B ∈
Rn×n, we denote as GA the corresponding CDAG constructed according to the description
in Section 4. Further, we denote as X the set of input vertices of GA. That is the set of
vertices corresponding each to the entries in the input matrices A and B. We now present
the proof of Lemma 4.
Proof of Lemma 4. The proof proceeds by induction on the number of Type 1 MSPs ν1
generated by A. In the base case ν1 = 1 and, by Definition 2 the entire problem is the
only, improper Type 1 MSP generated by A. Hence, the sets Y and X coincide and the
statement is trivially verified.
Assuming now inductively that the statement holds ν1 = k ≥ 1, we shall show it also
As ν1 > 1, we have that the algorithm executes at least one recursive step. Let P (j)
denote the 7 sub-problems generated at the first recursion step. We distinguish two cases:
(a) at least two of of the seven sub-problems P (j) generate each at least one Type 1 MSP;
(b) only one of the seven sub-problems generates all ν1 Type 1 MSPs. We first address
case (a), as case (b) will follow from a simple extension.
holds for ν1 = k + 1.
For case (a), let G(j), for j = 1, 2, . . . , 7 denote the seven sub-CDAGs of GA, each
corresponding to one of the seven sub-problems generated in the first recursive step of A
according to the chosen Strassen-like scheme as discussed in Section 3. By Definition 2 and
as, by assumption, A generates Type 1 MSPs, we have that each of these seven sub-problems
has input size greater or equal to 2√M × 2√M . Further, each of the seven sub-problems
P (j) generates at most ν1 − 1 Type 1 MSPs.
Let Y (j) (resp., Y (j)) denote the subsets of Y (resp., Y) in G(j), for j = 1, 2, . . . , 7.
That is Y (j) = Y ∩ Y (j), By Lemma 2 the G(j)'s have distinct input values and, hence,
are pairwise vertex-disjoint sub-CDAGs of GA. Thus, the Y (1), Y (2), . . . , Y (7) partition Y
(resp., Y (1),Y (2), . . . ,Y (7) partition Y) and P7
j=1 Y (j) = Y).
Let Y (j) = aj/pbj, we have Y =P7
j=1 aj/pbj.
By the inductive hypothesis, any dominator set D(j) of Y (j) with respect to the set K(j)
composed by the input vertices of G(j) must be such that D(j) ≥ min{2M, aj /√bj}. By
Definition 1 this implies that vertices in Y (j) can be connected to a subset K (j) ⊆ K(j)
j=1 Y (j) = Y (resp., P7
20
disjoint.
j=1 aj/qP7
of the input vertices of G(j) such that K (j) ≥ aj/√bj, using vertex-disjoint paths. Since
the sub-CDAGs G(j) are vertex disjoint, so are the paths connecting vertices in Y (j) to
vertices in K (j). In the following we show that it is indeed possible to extended at least
min{2M,P7
j=1 bj} of these paths to vertices in X while maintaining them vertex
According to the construction of GA as discussed in Section 4, vertices in X correspond-
ing to the entries of input matrix A (resp., B) are connected to vertices in K(1),K(2), . . . ,K(7)
(and, hence, K (1), K (2), . . . , K (7)) by means of n2 encoding sub-CDAGs EncA (resp., EncB).
None of these 2n2 encoding sub-CDAGs share any input or output vertices. No two
output vertices of the same encoder sub-CDAG belong to the same sub-CDAG G(j), for
j = 1, 2, . . . , 7. This fact ensures that for a single sub-CDAG G(j), for j = 1, 2, . . . , 7, it is
possible to connect all the vertices in K(j) (and, hence, K (j)) to a subset of the vertices in
X via vertex disjoint paths.
For each of the 2n2 encoder sub-CDAGs, let us consider the vector yl ∈ {0, 1}7 such that
yl[j] = 1 iff the corresponding j-th output vertex of the encoder, which is an input of G(j), is
in K (j). Therefore yl equals the number of output vertices of the l-th encoder sub-CDAG
which are in K. From Lemma 3, for each encoder sub-CDAG there exists a subset Xl ∈ X
of the input vertices of the l-th encoder sub-CDAG for which it is possible to connect each
vertex in Xl to a distinct output vertex of the l-th encoder sub-CDAG using vertex disjoint
paths, each constituted by a singular edge with min{yl, 1 + ⌈(yl − 1) /2⌉} ≤ Xl ≤ yl.
The number of vertex disjoint paths connecting vertices in X , to vertices in ∪7
j=1K (j) is
l=1 min{yl, 1 +⌈(yl − 1) /2⌉}, under the constraint that P2n2
therefore at least P2n2
l=1 yl[j] =
aj/pbj, for j = 1, 2, . . . , 7.
Let us assume w.l.o.g. that a1/√b1 ≥ a2/√b2 ≥ . . . ≥ a7/√b7. As previously stated,
it is possible to connect all vertices in K1 to vertices in X through vertex disjoint paths.
Consider now all possible dispositions of the vertices in ∪7
j=2K (j) over the outputs of the
2n2 encoder sub-CDAGs. Hence, if a1/√b1 ≥ 2M we have that there are therefore at least
M vertex disjoint paths connecting vertices in X to vertices in K1, and, thus, to vertices in
Y as desired. In the following we assume a1/√b1 < M .
Recall that the output vertices of an encoder sub-CDAG belong each to a different
sub-CDAG G(j). From Lemma 3 we have that for each encoder there exists a subset
Xl ⊂ X of the input vertices of the l-th encoder sub-CDAG, with Xl ≥ minnyl, 1 +
⌈(yl − 1) /2⌉o ≥ yl[1] +(cid:16)P7
j=2 yl[j](cid:17) /2, for which is possible to connect all vertices in
j=1K (j)
Xl to Xl distinct output vertices of the l-th encoder sub-CDAG which are in ∪7
using Xl vertex disjoint paths. As all the Enc sub-CDAGs are vertex-disjoint, we can add
their contributions so that the number of vertex-disjoint paths connecting vertices in X to
21
2
7
Xj=2
aj
pbj
j=1 aj
,
j=1 bj
K (j) =
1
a1√b1
≥ P7
qP7
vertices in ∪7
j=1K (j) is at least
K (1) +
1
2
7
Xj=2
where the last passage follows by applying Lemma 10. There are therefore at least a/√b
vertex disjoint paths connecting vertices in X to vertices in ∪7
j=1K (j), and, thus, to vertices
in Y as desired. This concludes the proof for case (a).
For case (b), only one of the seven sub-problems P (j) generates all ν1 Type 1 MSPs.
Without loss of generality, let P (1) denote such sub-problem and let GAP (1) be the corre-
sponding sub-CDAG. According to the construction of GA as discussed in Section 4, vertices
in X corresponding to the entries of input matrix A (resp., B) are connected to the input
vertices of GAP (1) , by means of n2 encoding sub-CDAGs EncA (resp., EncB). None of these
2n2 encoding sub-CDAGs share any input or output vertices. No two output vertices of
the same encoder sub-CDAG belong to the same sub-CDAG G(j), for j = 1, 2, . . . , 7. This
fact ensures that it is possible to connect all the input vertices of GAP (1) to a subset of the
vertices in X via vertex disjoint paths. The proof for case (b) then follows by recursively
applying the arguments in the Proof of Lemma 4 to GAP (1) .
A.2 Proof of Lemma 5
Given an algorithm A ∈ H used to multiply input squared matrices A, B ∈ Rn×n, let GA
denote the corresponding CDAG constructed according to the description in Section 4. In
the following we denote as X the set of input vertices of GA. That is the set of vertices
corresponding each to the entries in the input matrices A and B. Further, for each Type 2
MSP Pi we denote as Y i the set of input vertices of the sub-CDAG GAPi
associated with Pi.
That is the set of vertices corresponding each to the entries in the input matrices Ai and
Bi of Pi. Also, we define Y = ∪ν2
which is a heavily modified version of a result previously introduced in [11, Lemma 6].
Lemma 11. Consider an algorithm A ∈ H with input matrices A, B ∈ Rn×n which gener-
ates ν2 Type 2 MSPs. Given the corresponding CDAG GA, let Q be a set of internal (i.e.,
not input) vertices of its ν2 sub-CDAGs corresponding each to one of the generated Type 2
In order to simplify the presentation of the proof of Lemma 5 we first introduce Lemma 11
i=1Y i.
MSPs. For any Z ⊆ Z with Z ≥ 2Q there exist X ⊆ X with X ≥ 2pM (Z − 2Q)
such that each vertex in X is connected to some vertex in Z by a directed path with no
vertex in Q.
Proof. The proof proceeds by induction on the number ν2 of Type 2 MSPs generated by A.
In the base case ν2 = 1 and, by Definition 2 the entire problem is the only, improper, Type
2 MSP generated by A and, thus, the sets Y and X coincide. Consider now the following
lemma:
22
Lemma 12 ([11, Lemma 5]). Let Gn×n denote the CDAG corresponding to the execution of
an unspecified algorithm for the square matrix multiplication function with input matrices
of size n × n. Let O′ ⊆ O be a subset of its output vertices O. For any subset D of the
vertices of Gn×n with O′ ≥ 2D, there exists a set I′ ⊆ I of the input vertices I of Gn×n
with cardinality I′ ≥ 2npO′ − 2D, such that all vertices in I′ are connected to some
vertex in O′ by directed paths with no vertex in D.
As, by Definition 2, GA is a Gn×n CDAG, with n ≥ 2√M , we can conclude that the
statement of Lemma 11 is verified in the base case as a consequence of Lemma 12 and of
the fact that, as previously mentioned, Y and X coincide,
Lemma 12 was originally introduced in [11, Lemma 5] and is based on the analysis of
the Grigoriev's flow of the matrix multiplication function. For the sake of completeness we
present the proof in Appendix A.4.
Assuming now inductively that the statement holds for ν2 = k > 1, we shall show it
also holds for ν2 = k + 1. As ν2 > 1, we have that the algorithm executes at least one
recursive step. Let P (j) denote the 7 sub-problems generated at the first recursion step.
We distinguish two cases: (a) at least two of of the seven sub-problems P (j) generate each
at least one Type 2 MSP; (b) only one of the seven sub-problems generates all ν2 Type 2
MSPs. We first address case (a), as case (b) will follow from a simple extension.
j=1 Z (j) = Z and P7
j=1 δ(j) ≥ Z − 2Q.
For case (a), let G(j), for j = 1, 2, . . . , 7 denote the seven sub-CDAGs of GA, each
corresponding to one of the seven sub-problems generated in the first recursive step of A
according to the chosen Strassen-like scheme as discussed in Section 4. By Definition 2 and
as, by assumption, A generates Type 2 MSPs, we have that n/2 ≥ 2√M . Further, each of
the seven sub-problems P (j) generates at most ν2 − 1 Type 2 MSPs.
Let Z (j), Y (j) and Q(j) respectively denote the subsets of Z, Y and Q in G(j), for j =
1, 2, . . . , 7. By Lemma 2 the G(j)'s have distinct input values and, hence, are pairwise vertex-
disjoint sub-CDAGs of GA. Thus, Z1, Z2, . . . , Z7 partition Z, Y 1,Y 2, . . . ,Y 7 partition Y
and Q1, Q2, . . . , Q7 partition Q. This implies P7
j=1 Q(j) = Q. Let
δ(j) = max{0,Z (j) − 2Q(j)}, we have δ =P7
Applying the inductive hypothesis to each G(j), we have that there is a subset Y (j) ⊆ Y (j)
with Y (j) ≥ 4√M δ(j) such that vertices of Y (j) are connected to vertices in Z (j) via paths
with no vertex in Q(j). In the sequel the set Y referred to in the statement will be identified
as a suitable subset of ∪7
i=jY (j) so that property (b) will be automatically satisfied. Towards
property (a), we observe by the inductive hypothesis that vertices in Y (j) can be connected
to a subset K (j) of the input vertices of G(j) with K (j) = Y (j), using vertex-disjoint
paths. Since the sub-CDAGs G(j) are vertex disjoint, so are the paths connecting vertices
in Y (j) to vertices in K (j). It remains to show that at least 4pM (Z − 2Q) of these paths
can be extended to X while maintaining them vertex-disjoint.
According to the construction of GA as discussed in Section 4, vertices in X correspond-
ing to the entries of input matrix A (resp., B) are connected to vertices in K1, K2, . . . , K7
by means of n2 encoding sub-CDAGs EncA (resp., EncB). None of these 2n2 encoding
sub-CDAGs share any input or output vertices. No two output vertices of the same encoder
sub-CDAG belong to the same sub-CDAG G(j), for j = 1, 2, . . . , 7. This fact ensures that
for a single sub-CDAG G(j), for j = 1, 2, . . . , 7, it is possible to connect all the vertices in
23
Ki to a subset of the vertices in X via vertex disjoint paths.
For each of the 2n2 encoder sub-CDAGs, let us consider the vector yl ∈ {0, 1}7 such that
yl[j] = 1 iff the corresponding j-th output vertex of the encoder, which is an input of G(j), is
in K (j). Therefore yl equals the number of output vertices of the l-th encoder sub-CDAG
which are in K. From Lemma 3, for each encoder sub-CDAG there exists a subset Xl ∈ X
of the input vertices of the l-th encoder sub-CDAG for which it is possible to connect each
vertex in Xl to a distinct output vertex of the l-th encoder sub-CDAG using vertex disjoint
paths, each constituted by a singular edge with min{yl, 1 + ⌈(yl − 1) /2⌉} ≤ Xl ≤ yl.
j=1K (j) is
The number of vertex disjoint paths connecting vertices in X , to vertices in ∪7
therefore at least P2n2
l=1 min{yl, 1 +⌈(yl − 1) /2⌉}, under the constraint that P2n2
l=1 yl[j] =
4√M δ(j), for j = 1, 2, . . . , 7. Let us assume w.l.o.g.
that δ(1) ≥ δ(20 ≥ . . . ≥ δ(7). As
previously stated, it is possible to connect all vertices in K1 to vertices in X through vertex
disjoint paths. Consider now all possible dispositions of the vertices in ∪7
j=2K (j) over the
outputs of the 2n2 encoder sub-CDAGs. Recall that the output vertices of an encoder
sub-CDAG belong each to a different sub-CDAG G(j). From Lemma 3 we have that for
each encoder there exists a subset Xl ⊂ X of the input vertices of the l-th encoder sub-
CDAG, with Xl ≥ minnyl, 1 + ⌈(yl − 1) /2⌉o ≥ yl[1] +(cid:16)P7
j=2 yl[j](cid:17) /2, for which is
possible to connect all vertices in Xl to Xl distinct output vertices of the l-th encoder
j=1K (j) using Xl vertex disjoint paths. As all the Enc sub-
sub-CDAG which are in ∪7
CDAGs are vertex disjoint, we can add their contributions so that the number of vertex
Pj=2K (j) =
disjoint paths connecting vertices in X to vertices in ∪7
4√M √δ(1) + 1
Pj=2
2
Xj=2pδ(j)
4√M
pδ(1) +
.
As, by assumption, δ(1) ≥ . . . δ(7), we have: √δ(1)√δ(j) ≥ δ(j) for j = 2, . . . , 7. Thus:
4√M
δ(j) ≥(cid:16)4pM (Z − 2Q)(cid:17)2
√δ(j)!. Squaring this quantity leads to:
Xj=2pδ(j)
Xj=2pδ(j) +
δ(1) +pδ(1)
7
Xj=2pδ(j)
j=1K (j) is at least K1+ 1
2
pδ(1) +
= 16M
2
7
Xj=1
7
1
2
2
≥ 16M
7
7
1
2
7
.
There are therefore at least 4pM (Z − 2Q) vertex disjoint paths connecting vertices in
X to vertices in ∪7
j=2K (j) as desired. This concludes the proof for case (a).
For case (b), only one of the seven sub-problems P (j) generates all ν1 Type 1 MSPs.
Without loss of generality, let P (1) denote such sub-problem and let GAP (1) be the corre-
sponding sub-CDAG. According to the construction of GA as discussed in Section 4, vertices
in X corresponding to the entries of input matrix A (resp., B) are connected to the input
vertices of GAP (1) , by means of n2 encoding sub-CDAGs EncA (resp., EncB). None of these
2n2 encoding sub-CDAGs share any input or output vertices. No two output vertices of
24
7
2
1
2
the same encoder sub-CDAG belong to the same sub-CDAG G(j), for j = 1, 2, . . . , 7. This
fact ensures that it is possible to connect all the input vertices of GAP (1) to a subset of the
vertices in X via vertex disjoint paths. The proof for case (b) then follows by recursively
applying the arguments in the Proof of Lemma 11 to GAP (1) .
Lemma 11, provides the base for the proof of Lemma 5, which is itself a reworked
version a result from [11] (Lemma 7) modified in order to account for the hybrid nature of
algorithms being considered in this work.
Proof of Lemma 5. Suppose for contradiction that D is a dominator set for Z in GA such
that D ≤ 2M − 1. Let D′ ⊆ D be the subset of the vertices of D composed by vertices
which are not internal to the sub-CDAGs corresponding to the Type 2 MSPs generated, by
assumption, by A. From Lemma 11, with Q = D \ D′, there exist X ⊆ X and Y ⊆ Y with
X = Y ≥ 4pM (Z − 2 (D − D′)) such that vertices in X are connected to vertices in
Y by vertex-disjoint paths. Hence, each vertex in D′ can be on at most one of these paths.
Thus, there exists X′ ⊆ X and Y ′ ⊆ Y with X′ = Y ′ ≥ φ = 4pM (Z − 2 (D − D′))−
D′ paths from X′ to Y ′ with no vertex in D′. From Lemma 11, we also have that all vertices
in Y , and, hence, in Y ′, are connected to some vertex in Z by a path with no vertex in
D \ D′. Thus, there are at least φ paths connecting vertices in X′ ⊆ X to vertices in Z
with no vertex in D. We shall now show that the contradiction assumption D ≤ 2M − 1
implies φ > 0:
(cid:16)4pM (Z − 2 (D − D′))(cid:17)2
= 16M(cid:0)Z − 2(cid:0)D − D′(cid:1)(cid:1) ,
= 16M (Z − 2D) + 32MD′.
By D ≤ 2M − 1, we have Z − 2D > 4M − 2(M − 1) > 0. Furthermore, from D′ ⊆ D,
we have 32M > 2M − 1 > D ≥ D′. Therefore:
(cid:0)φ + D′(cid:1)2 =(cid:16)4pM (Z − 2 (D − D′))(cid:17)2
> D′2.
(15)
Again, D ≤ 2M − 1 implies M (Z − 2 (D − D′)) > 0. Hence, we can take the square
root on both sides of (15) and conclude that φ > 0. Therefore, for D ≤ 2M − 1 there are
at least φ > 0 paths connecting a global input vertex to a vertex in Z with no vertex in D,
contradicting the assumption that D is a dominator of Z.
i
i
. The proof for
follows an analogous argument. Let D be a dominator set for the
i
i
< Y (B)
A.3 Proof of Lemma 6
Proof of Lemma 6. Without loss of generality, let us assume Y (A)
the case Y (A)
set of vertices corresponding to T ′i with respect to Y (A)
to vertices in Y (B)
corresponding to a vertex in Y (A)
value a. The lemma follows combining statements (i) and (ii):
(i) There exists an assignment of the input variables of Pi corresponding to vertices in
Consider a possible assignment to the values of Bi such that all the values corresponding
are assigned value 1. Under such assignment, for every variable a ∈ Ai
at least one of the elementary products in T ′i assumes
≥ Y (B)
.
i
i
i
25
i
such that the output variables in T ′i assume at least RY (A)
Y i \Y (A)
all possible assignments of the variables corresponding to vertices in Y (A)
(ii) Since all paths form Y (A)
to the vertices corresponding to the variables in T ′i intercept
D, the values of the elementary products in T ′i are determined by the inputs in Y i \ Y (A)
,
which are fixed, and by the values of the vertices in D; hence the elementary products in
T ′i can take at most RD distinct values.
A.4 Proof of Lemma 12
distinct values under
.
i
i
i
i
Before presenting the proof of Lemma 12, we present the concept of "flow of a function"
which was originally introduced by Grigoriev [17] and then revised by Savage [30]. In this
work, we use the following version:
Definition 3 (Grigoriev's flow). A function f : Rp → Rq has a w (u, v) Grigoriev's flow if
for all subsets X1 (resp., Y1), of its p input (resp., q output) variables, with X1 ≥ u and
Y1 ≥ v, there is a sub-function h of f obtained by making some assignment to variables
of f not in X1 and discarding output variables not in Y1, such that h has at least Rw(u,v)
points in the image of its domain.
Grigoriev's flow is an inherent property of a function, agnostic to the algorithm used
to compute the function.
It provides a lower bound to the amount of information that
suitable sub-sets of outputs encode about suitable sub-sets of inputs. Since any information
about inputs that is encoded by outputs must be also be encoded by any dominator of those
outputs, we have the following lower bound on the size of dominator sets.
Lemma 13 ([11, Lemma 4]). Let G = (V, E) be a CDAG computing a function f : Rp →
Rq, defined on the ring R, with p input and q output variables. Let I (resp., O) denote
the set of input (resp., output) vertices of G. Further let IX (resp., OY ) denote the subset
of I (resp., O) which correspond to a subset X (resp., Y ) of the p input (resp., q output)
variables of f . For any O′ ⊆ OY and any I′ ⊆ IX , any dominator of O′ with respect to I′
satisfies D ≥ wfX,Y (I′,O′).
A lower bound on the Grigoriev's flow for the square matrix multiplication function
over the ring R was presented in [30, Theorem 10.5.1].
fn×n : R2n2 → Rn2
Lemma 14 (Grigoriev's flow of fn×n : R2n2 → Rn2
wn×n (u, v) Grigoriev's flow, where:
[30]). fn×n : R2n2 → Rn2
has a
wn×n (u, v) ≥
1
2 v − (cid:0)2n2 − u(cid:1)2
4n2
! , for 0 ≤ u ≤ 2n2, 0 ≤ v ≤ n2.
(16)
We can now state the proof of Lemma 12. Recall that we denote as Gn×n the CDAG cor-
responding to the execution of a unspecified algorithm for the square matrix multiplication
function with input matrices of size n × n.
Proof of Lemma 12. The statement follows by applying the results in Lemma 13 and Lemma 14
to the CDAG Gn×n. Let I′′ ⊆ I denote the set of all input vertices of Gn×n, such that all
26
paths connecting these vertices to the output vertices in O′ include at least a vertex in D
(i.e., I′′ is the largest subset of I with respect to whom D is a dominator set for O′). From
Lemmas 14 and 13 the following must hold:
D ≥ wn×n ≥
1
2 O′ − (cid:0)2n2 − I′′(cid:1)2
4n2
! .
(17)
Let I′ = I\I′′. By the definition of I′′, the vertices in I′ are exactly those that are connected
to vertices in O′ by directed paths with no vertex in D. Since I = 2n2, from (17) we have
I′2 ≥ 4n2 (O′ − 2D).
27
7
5
4
1
3
2
6
7
5
4
1
3
2
6
C1,1 C1,2 C2,1 C2,2
A1,1 A1,2 A2,1 A2,2
B1,1 B1,2 B2,1 B2,2
M7 M5 M4 M1 M3 M2 M6
(a) EncA
(b) EncB
(c) Dec
Figure 2: Encoder and Decoder sub-CDAGs corresponding to Strassen's original algorithm.
B Strassen's fast multiplication algorithm
Algorithm 1 Strassen's Matrix Multiplication
Input: matrices A, B
Output: matrix C
1: procedure StrassenMM(A,B)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
if n = 1 then
C = A · B
Decompose A and B into four equally sized block matrices as follows:
else
A =(cid:20) A1,1 A1,2
B2,1 B2,2 (cid:21)
A2,1 A2,2 (cid:21) , B =(cid:20) B1,1 B1,2
M1 = StrassenMM (A1,1 + A2,2, B1,1 + B2,2)
M2 = StrassenMM (A2,1 + A2,2, B1,1)
M3 = StrassenMM (A1,1, B1,2 − B2,2)
M4 = StrassenMM (A2,2, B2,1 − B1,1)
M5 = StrassenMM (A1,1 + A1,2, B2,2)
M6 = StrassenMM (A2,1 − A1,1, B1,1 + B1,2)
M7 = StrassenMM (A1,2 − A2,2, B2,1 + B2,2)
C1,1 = M1 + M4 − M5 + M7
C1,2 = M3 + M5
C2,1 = M2 + M4
C2,2 = M1 − M2 + M3 + M6
return C
The original version of Strassen's fast matrix multiplication [35] is reported in Algo-
rithm 1. We refer the reader to [37] for Winograd's variant, which reduces the number of
additions.
28
|
1904.09958 | 3 | 1904 | 2019-11-07T12:13:37 | Almost Optimal Testers for Concise Representations | [
"cs.DS"
] | We give improved and almost optimal testers for several classes of Boolean functions on $n$ inputs that have concise representation in the uniform and distribution-free model. Classes, such as $k$-junta, $k$-linear functions, $s$-term DNF, $s$-term monotone DNF, $r$-DNF, decision list, $r$-decision list, size-$s$ decision tree, size-$s$ Boolean formula, size-$s$ branching programs, $s$-sparse polynomials over the binary field and function with Fourier degree at most $d$. The method can be extended to several other classes of functions over any domain that can be approximated by functions that have a small number of relevant variables. | cs.DS | cs | Almost Optimal Testers for
Concise Representations
Nader H. Bshouty
Dept. of Computer Science
Technion, Haifa, 32000
November 11, 2019
Abstract
We give improved and almost optimal testers for several classes of Boolean functions on n
inputs that have concise representation in the uniform and distribution-free model. Classes,
such as k-Junta, k-Linear Function, s-Term DNF, s-Term Monotone DNF, r-DNF, Decision
List, r-Decision List, size-s Decision Tree, size-s Boolean Formula, size-s Branching Program,
s-Sparse Polynomial over the binary field and functions with Fourier Degree at most d.
The approach is new and combines ideas from Diakonikolas et al. [30], Bshouty [15], Goldreich
et al. [38], and learning theory. The method can be extended to several other classes of functions
over any domain that can be approximated by functions that have a small number of relevant
variables.
9
1
0
2
v
o
N
7
]
S
D
.
s
c
[
3
v
8
5
9
9
0
.
4
0
9
1
:
v
i
X
r
a
1
Contents
1 Inroduction
1.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 The Model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Our Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Testing Subclasses of k-Junta . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 Testing Classes that are Close to k-Junta . . . . . . . . . . . . . . . . . . . .
4
4
4
6
7
7
9
2 Preparing the Target for Accessing the Relevant Variables
10
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1 Preliminaries
2.2 Approximating the Target . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Testing the Relevant Sets
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Determining the Values of the Relevant Variables . . . . . . . . . . . . . . . . . . . . 15
3 Testing Subclasses of k-Junta
18
3.1 Testing the Closeness of f (xX ◦ 0X ) to F . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Testing the Closeness of F to C(Γ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Testing the Closeness of F to C(Γ) via Learning C(Γ) . . . . . . . . . . . . . . . . . 21
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 The First Tester
4 Results
26
4.1 Testing k-Junta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Testing k-Linear
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3 Testing k-Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Testing s-Term Monotone r-DNF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.5 Testing Size-s Decision Tree and Size s Branching Program . . . . . . . . . . . . . . 30
4.6 Functions with Fourier Degree at most d . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.7 Testing Length k Decision List
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.8 Testing s-Sparse Polynomial of Degree d . . . . . . . . . . . . . . . . . . . . . . . . . 31
5 Testing Classes that are Close to k-Junta
34
5.1 Removing Variables that only Appears in Large Size Terms . . . . . . . . . . . . . . 34
5.2 Testing s-term DNF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6 Results
37
6.1 Testing s-Term Monotone DNF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2 Testing Size-s Boolean Formula and Size-s Boolean Circuit
. . . . . . . . . . . . . . 38
6.3 Testing s-Sparse Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7 A General Method for Other Testers
41
7.1 Testing Decision List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
. . . . . . . . . . . . . . . . . . . 44
7.2 Testing r-DNF and r-Decision List for Constant r
2
8 Improvements and Further Results
45
8.1
Improved ApproxTarget Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.2 A General TestSets Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.3 A General RelVarValue Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.4
Improved Closef F Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.5 A Weaker ExQD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
8.6 An Improved Query Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.7 Some Improved Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
9 Appendix A
57
3
1
Inroduction
Property testing of Boolean function was first considered in the seminal works of Blum, Luby and
Rubinfeld [13] and Rubinfeld and Sudan [54] and has recently become a very active research area.
See for example, [2, 4, 5, 6, 9, 10, 15, 19, 20, 21, 22, 23, 25, 26, 30, 33, 37, 39, 42, 43, 47, 46, 50, 55]
and other works referenced in the surveys [36, 52, 53].
A Boolean function f : {0, 1}n → {0, 1} is said to be k-junta if it depends on at most k
coordinates. The class k-Junta is the class of all k-juntas. The class k-Junta has been of particular
interest to the computational learning theory community [11, 12, 17, 29, 40, 44, 48]. A problem
closely related to learning k-Junta is the problem of learning and testing subclasses C of k-Junta and
classes C of Boolean functions that can be approximated by k-juntas [10, 12, 31, 21, 30, 38, 39, 50]:
Given black-box query access to a Boolean function f . In learning, for f ∈ C, we need to learn,
with high probability, a hypothesis h that is ǫ-close to f . In testing, for any Boolean function f ,
we need to distinguish, with high probability, the case that f is in C versus the case that f is ǫ-far
from every function in C.
In the uniform-distribution property testing (and learning) the distance between Boolean func-
tions is measured with respect to the uniform distribution. In the distribution-free property test-
ing, [38], (and learning [56]) the distance between Boolean functions is measured with respect to
an arbitrary and unknown distribution D over {0, 1}n. In the distribution-free model, the testing
(and learning) algorithm is allowed (in addition to making black-box queries) to draw random
x ∈ {0, 1}n according to the distribution D. This model is studied in [15, 27, 32, 34, 41, 45].
1.1 Results
We give improved and almost optimal testers for several classes of Boolean functions on n inputs that
have concise representation in the uniform and distribution-free models. The classes studied here
are k-Junta, k-Linear Functions, k-Term, s-Term DNF, s-Term Monotone DNF, s-Term Monotone
r-DNF, r-DNF, Decision List, Length-k Decision List, r-Decision List, size-s Decision Tree, size-s
Branching Programs, size-s Boolean Formula, size-s-Boolean Circuit, s-Sparse Polynomials over
the binary field, s-Sparse Polynomials of Degree d and functions with Fourier Degree at most d.
In Table 1, we list all the previous results and our results in this paper. In the table, O(T )
stands for O(T · poly(log T )), U and D stand for uniform and distribution-free model, and Exp and
Poly stand for exponential and polynomial time.
It follows from the lower bounds of Saglam, [55], that our query complexity is almost optimal
(with log-factor) for the classes k-Junta, k-Linear, k-Term, s-Term DNF, s-Term Monotone DNF, r-
DNF (r constant), Decision List, r-Decision List (r constant), size-s Decision Tree, size-s Branching
Programs and size-s Boolean Formula. For more details on the previous results and the results in
this paper see Table 1 and Sections 4, 6 and 7.
1.2 Notations
In this subsection, we give some notations that we use throughout the paper.
Denote [n] = {1, 2, . . . , n}. For S ⊆ [n] and x = (x1, . . . , xn) we denote x(S) = {xii ∈ S}.
For X ⊂ [n] we denote by {0, 1}X the set of all binary strings of length X with coordinates
indexed by i ∈ X. For x ∈ {0, 1}n and X ⊆ [n] we write xX ∈ {0, 1}X to denote the projection
of x over coordinates in X. We denote by 1X and 0X the all-one and all-zero strings in {0, 1}X ,
4
Class of Functions
Model
#Queries
Time
Reference
s-Term Monotone DNF
s-Term Unate DNF
s-Term Monotone r-DNF
s-Term Unate r-DNF
s-Term DNF
r-DNF (Constant r)
Decision List
Length-k Decision List
r-DL (Constant r)
k-Linear
k-Term
size-s Decision Trees and
size-s Branching Programs
size-s Boolean Formulas
size-s Boolean Circuit
Functions with
Fourier Degree ≤ d
s-Sparse Polynomial
over F2 of Degree d
s-Sparse Polynomial
over F2
U
U
U
U
U
D
U
U
U
U
U
U
D
U
U
D
U
U
D
U
U
D
U
U
U
U
U
D
U
U
U
D
U
U
U
O(s2/ǫ)
O(s/ǫ2)
O(s/ǫ)
O(s/ǫ2)
O(s/ǫ)
O(s2r/ǫ)
O(s2/ǫ)
O(s/ǫ2)
O(s/ǫ)
O(1/ǫ)
O(1/ǫ2)
O(1/ǫ)
O(k2/ǫ)
O(1/ǫ)
O(k/ǫ)
O(k/ǫ)
O(1/ǫ)
O(1/ǫ)
O(k/ǫ)
O(s/ǫ2)
O(s/ǫ)
O(s2/ǫ)
O(s/ǫ2)
O(s/ǫ)
O(s2/ǫ2)
O(s2/ǫ)
O(22d/ǫ2)
Poly.
Exp.
Poly.
Exp.
Poly.
Poly.
Exp.
Exp.
Exp.
Poly.
Poly.
Poly.
Poly.
Poly.
Poly.
Poly.
Poly.
Poly.
Poly.
Exp.
Exp.
Exp.
Exp.
Exp.
Exp.
Exp.
Exp.
Poly.
poly(s/ǫ) + O(22d) Poly.
O(2d/ǫ + 22d)
[50]
[21]
This Paper
[21]
This Paper
This Paper
[30]
[21]
This Paper
This Paper
[30]
This Paper
This Paper
This Paper
[8, 13]
This Paper
[50]
This Paper
This Paper
[21]
This Paper
This Paper
[21]
This Paper
[21]
This Paper
[21]
This Paper
[1, 31]
O(s2/ǫ + 22d)
O(s/ǫ + s2d)
O(s2/ǫ + s2d)
O(s/ǫ2)
P oly(s/ǫ)
O(s2/ǫ)
This Paper
This Paper
Poly. This Paper+[1]
Poly.
Poly.
Exp.
Poly.
Poly.
[21]
[31]
This Paper
Figure 1: A table of the results. In the table, O(T ) stands for O(T · poly(log T )), U and D stand
for uniform and distribution-free model, and Exp and Poly stand for exponential and polynomial
time.
respectively. When we write xI = 0 we mean xI = 0I . For X1, X2 ⊆ [n] where X1 ∩ X2 = ∅ and
5
x ∈ {0, 1}X1 , y ∈ {0, 1}X2 we write x◦y to denote their concatenation, i.e., the string in {0, 1}X1∪X2
that agrees with x over coordinates in X1 and agrees with y over coordinates in X2. Notice that
x ◦ y = y ◦ x. When we write u = ◦w∈W w we mean that u is the concatenation of all the strings in
W . For X ⊆ [n] we denote X = [n]\X = {x ∈ [n]x 6∈ X}. We say that two strings x and y are
equal on I if xI = yI.
Given f, g : {0, 1}n → {0, 1} and a probability distribution D over {0, 1}n, we say that f
is ǫ-close to g with respect to D if Prx∈D[f (x) 6= g(x)] ≤ ǫ, where x ∈ D means x is chosen
from {0, 1}n according to the distribution D. We say that f is ǫ-far from g with respect to D if
Prx∈D[f (x) 6= g(x)] ≥ ǫ. For a class of Boolean functions C, we say that f is ǫ-far from every
function in C with respect to D if for every g ∈ C, f is ǫ-far from g with respect to D. We will use
U to denote the uniform distribution over {0, 1}n or over {0, 1}X when X in clear from the context.
For a Boolean function f and X ⊂ [n], we say that X is a relevant set of f if there are
a, b ∈ {0, 1}n such that f (a) 6= f (bX ◦ aX). We call the pair (a, b) (or just a when b = 0) a witness
of f for the relevant set X. When X = {i} then we say that xi is a relevant variable of f and a is
a witness of f for xi. Obviously, if X is relevant set of f then x(X) contains at least one relevant
variable of f .
where x is the negation of x.
We say that the Boolean function f : {0, 1}n → {0, 1} is a literal if f ∈ {x1, . . . , xn, x1, . . . , xn}
Let C be a class of Boolean functions f : {0, 1}n → {0, 1}. We say that C is closed under
variable projection if for every projection π : [n] → [n] and every f ∈ C, we have f (x(π)) ∈ C where
x(π) := (xπ(1),··· , xπ(n)). We say that C is closed under zero projection (resp. closed under one
projection) if for every f ∈ C and every i ∈ [n], f (0{i} ◦ x{i}) (resp. f (1{i} ◦ x{i}) ∈ C). We say it
is closed under zero-one projection if is closed under zero and one projection.
Throughout the paper, we assume that the class C is closed under variable and zero projection.
After section 3, we assume that it is also closed under one projection.
1.3 The Model
In this subsection, we define the testing and learning models.
In the testing model, we consider the problem of testing a class of Boolean function C in the
uniform and distribution-free testing models. In the distribution-free testing model (resp. uniform
model), the algorithm has access to a Boolean function f via a black-box that returns f (x) when a
string x is queried. We call this query membership query (MQf or just MQ). The algorithm also has
access to unknown distribution D (resp. uniform distribution) via an oracle that returns x ∈ {0, 1}n
chosen randomly according to the distribution D (resp. according to the uniform distribution). We
call this query example query (ExQD (resp. ExQ)).
testing algorithm) A for C is an algorithm
A distribution-free testing algorithm, [38], (resp.
that, given as input a distance parameter ǫ and the above two oracles to a Boolean function f ,
1. if f ∈ C then A outputs "accept" with probability at least 2/3.
2. if f is ǫ-far from every g ∈ C with respect to the distribution D (resp. uniform distribution)
then A outputs "reject" with probability at least 2/3.
We will also call A a tester (or ǫ-tester) for the class C and an algorithm for ǫ-testing C.
We say that A is one-sided if it always accepts when f ∈ C; otherwise, it is called two-sided
algorithm. The query complexity of A is the maximum number of queries A makes on any Boolean
function f .
6
In the learning models, C is a class of representations of Boolean functions rather than a
class of Boolean functions. Therefore, we may have two different representations in C that are
logically equivalent. In this paper, we assume that this representation is verifiable, that is, given a
representation g, one can decide in polynomial time on the length of this representation if g ∈ C.
A distribution-free proper learning algorithm (resp. proper learning algorithm under the uniform
distribution) A for C is an algorithm that, given as input an accuracy parameter ǫ, a confidence
parameter δ and an access to both MQf for the target function f ∈ C and ExQD, with unknown
D, (resp. ExQ or ExQU ), with probability at least 1 − δ, A returns h ∈ C that is ǫ-close to
f with respect to D (resp. with respect to the uniform distribution). This model is also called
proper PAC-learning with membership queries under any distribution (resp. under the uniform
distribution) [3, 56]. A proper exact learning algorithm [3] for C is an algorithm that given as input
a confidence parameter δ and an access to MQf for f ∈ C, with probability at least 1 − δ, returns
h ∈ C that is equivalent to f . The query complexity of A is the maximum number of queries A
makes on any Boolean function f ∈ C.
1.4 Our Techniques
In this section, we give a detailed overview of our techniques.
1.4.1 Testing Subclasses of k-Junta
For testing a subclass C of k-Junta that is closed under variable and zero projections, we use
TesterC in Figure 8. We first note that TesterC rejects if any procedure that it calls rejects.
First, TesterC calls the procedure ApproxTarget, in Figure 2. ApproxTarget partitions
the (indices of the) variables [n] into r = O(k2) disjoint sets X1, . . . , Xr. Since C ⊆ k−Junta it
follows that, with high probability (whp), if f ∈ C then different relevant variables of f fall into
different sets. Therefore, if f ∈ C, whp, every Xi contains at most one relevant variable of f . The
procedure then binary searches for enough relevant sets {Xi}i∈I such that, whp, for X = ∪i∈IXi,
h = f (xX ◦ 0X ) is (ǫ/3)-close to f with respect to D. If the procedure finds more than k relevant
If f ∈ C then the
sets of f then there are more than k relevant variables in f and it rejects.
procedure does not reject and, since C is closed under zero projection, h ∈ C. Since, whp, h is
(ǫ/3)-close to f with respect to D, it is enough to distinguish whether h is in C or (2ǫ/3)-far from
every function in C with respect to D. ApproxTarget also finds, for each relevant set Xi, i ∈ I,
a witness v(i) ∈ {0, 1}n of h for Xi. That is, for every i ∈ I, h(v(i)) 6= h(0Xi ◦ v(i)
). If f ∈ C, then
h ∈ C and, whp, for each i ∈ I, h(xXi ◦ v(i)
In the second stage, the tester calls the procedure TestSets, in Figure 4. TestSets verifies,
whp, that for every i ∈ I, h(xXi ◦ v(i)
) is (1/30)-close to some literal in {xτ (i), xτ (i)} for some
τ (i) ∈ Xi, with respect to the uniform distribution. If f ∈ C, then h ∈ C and, whp, for each
i ∈ I, h(xXi ◦ v(i)
) is a literal and therefore TestSets does not reject. Notice that if f ∈ C, then,
whp, Γ := {xτ (i)}i∈I are the relevant variables of h. This test does not give τ (i) but the fact that
h(xXi ◦ v(i)
) is close to xτ (i) or xτ (i) can be used to find the value of uτ (i) in every assignment
u ∈ {0, 1}n without knowing τ (i). The latter is done, whp, by the procedure RelVarValues. See
Figure 5. Both procedures make O(k) queries.
Recall that for ξ ∈ {0, 1}, ξX is the all ξ vector in {0, 1}X . Then the tester defines the Boolean
function F = h(0X ◦ ◦i∈I (xτ (i))Xi) on the variables {xτ (j)}j∈I , that is, the function F is obtained
) is a literal. ApproxTarget makes O(k/ǫ) queries.
Xi
Xi
Xi
Xi
Xi
7
by substituting in h for every i ∈ I and every xj ∈ x(Xi) the variable xτ (i). Since C ⊆ k-Junta and
C is closed under variable and zero projections, τ (i) ∈ Xi and, whp, Γ = {xτ (i)}i∈I are the relevant
variables of h we have:
• If the function f is in C then, whp, F = h ∈ C and F depends on all the variables in
Γ = {xτ (j)}j∈I .
If h is (2ǫ/3)-far from every function in C with respect to D then either h is (ǫ/3)-far from F with
respect to D or F is (ǫ/3)-far from every function in C(Γ) with respect to D where C(Γ) is the set
of all functions in C that depends on all the variables in Γ. Therefore,
• If the function f is ǫ-far from every function in C then, whp, either
1. h is (ǫ/3)-far from F with respect to D or
2. F is (ǫ/3)-far from every function in C(Γ) with respect to D.
Therefore, it remains to do two tests. The first is testing whether h = F given that h is either
(ǫ/3)-far from F with respect to D or h = F . The second is testing whether F ∈ C given that F
is either (ǫ/3)-far from every function in C(Γ) with respect to D or f ∈ C(Γ).
The former test, h = F , can be done, whp, by choosing O(1/ǫ) strings u ∈ {0, 1}n according to
the distribution D and testing whether F (u) = h(u). To compute F (u) we need to find {uτ (i)}i∈I ,
which can be done by the procedure RelVarValues. Therefore, each query to F requires one call
to the procedure RelVarValues that uses O(k) queries to f . Thus, the first test can be done using
O(k/ǫ) queries. This is done in the procedure Closef F in Figure 6.
Notice that, thus far, all the above procedures run in polynomial time and make O(k/ǫ) queries.
Testing whether F ∈ C can be done, whp, by choosing O((log C(Γ))/ǫ) strings u ∈ {0, 1}n
according to the distribution D and testing whether F (u) = g(u) for every g ∈ C(Γ). Notice here
that the time complexity is poly(C(Γ)) which is polynomial only when C(Γ) contains polynomial
number of functions.
If the distribution is uniform, we do not need to use RelVarValues to find {uτ (i)}i∈I because
when the distribution of u is uniform the distribution of {uτ (i)}i∈I is also uniform. Therefore we
can just test whether F (u) = g(u) for every g ∈ C(Γ) for uniform {uτ (i)}i∈I . Then computing F (u)
for random uniform string u can be done in one query to h. Thus, for the uniform distribution, the
algorithm makes O((log C(Γ))/ǫ) queries to f . This is the procedure CloseF CU in Figure 7.
If the distribution is unknown then each computation of F (u) for a random string u according to
the distribution D requires choosing u according to the distribution D, then extracting {uτ (i)}i∈I
from u and then substituting the values {uτ (i)}i∈I in F . This can be done by the procedure
RelVarValues using O(k) queries to h. Therefore, for unknown distribution the algorithm makes
O((k log C(Γ))/ǫ) queries to f . This is the procedures CloseF CD in Figure 7.
As we mentioned before the time complexity of CloseF CU and CloseF CD is polynomial only
if C(Γ) is polynomial. When C(Γ) is exponential, we solve the problem via learning theory. We
find a proper learning algorithm A for C(Γ). We run A to learn F . If the algorithm fails, runs
more time than it should, asks more queries than it should or outputs a hypothesis g 6∈ C then we
know that, whp, F 6∈ C(Γ). Otherwise, it outputs a function g ∈ C(Γ) and then, as above, we test
whether g = F given that g is (ǫ/3)-far from F or g = F .
Therefore, for the uniform distribution, if the proper learning algorithm for C makes m MQs
and q ExQs then the tester makes m + q + O(1/ǫ) queries. If the distribution is unknown, then the
tester makes m + O(kq + k/ǫ) queries.
8
1.4.2 Testing Classes that are Close to k-Junta
To understand the intuition behind the second technique, we demonstrate it for testing s-term
DNF.
The tester first runs the procedure ApproxC in Figure 11. This procedure is similar to the pro-
cedure ApproxTarget. It randomly uniformly partitions the variables to r = 4c2(c + 1)s log(s/ǫ)
disjoint sets X1, . . . , Xr and finds relevant sets {Xi}i∈I . Here c is a large constant. To find a new
relevant set, it chooses two random uniform strings u, v ∈ {0, 1}n and verifies if f (uX ◦ vX) 6= f (u)
where X is the union of the relevant sets that it has found thus far. If f (uX ◦ vX ) 6= f (u) then the
binary search finds a new relevant set.
In the binary search for a new relevant set, the procedure defines a set X ′ that is equal to
the union of half of the sets in {Xi}i6∈I . Then either f (uX∪X ′ ◦ vX ′) 6= f (u) or f (uX∪X ′ ◦ vX ′) 6=
f (uX ◦ vX). Then it recursively does the above until it finds a new relevant set Xℓ.
It is easy to show that if f is s-term DNF then, whp, for all the terms T in f of size at least
c2 log(s/ǫ), for all the random uniform strings u, v chosen in the algorithm and for all the strings w
generated in the binary search, T (uX ◦ vX ) = T (u) = T (w) = 0. Therefore, when f is s-term DNF,
the procedure, whp, runs as if there are no terms of size greater than c2 log(s/ǫ) in f . This shows
that, whp, each relevant set that the procedure finds contains at least one variable that belongs to
a term of size at most c2 log(s/ǫ) in f . Therefore, if f is s-term DNF, the procedure, whp, does not
generate more than c2s log(s/ǫ) relevant sets. If the procedure finds more than c2s log(s/ǫ) relevant
sets then, whp, f is not s-term DNF and therefore it rejects.
Let R be the set of all the variables that belong to the terms in f of size at most c2 log(s/ǫ). The
procedure returns h = f (xX ◦ wX) for random uniform w where X is the union of the relevant sets
X = ∪i∈IXi that is found by the procedure. If f is s-term DNF then since r = 4c2(c + 1)s log(s/ǫ)
and the number of relevant sets is at most c2s log(s/ǫ), whp, at least (1/2)c log(s/ǫ) variables in
each term of f that contains at least c log(s/ǫ) variables not in R falls outside X in the partition of
[n]. Therefore, for random uniform w, whp, terms T in f that contains at least c log(s/ǫ) variables
not in R satisfies T (xX ◦ wX ) = 0 and therefore, whp, are vanished in h = f (xX ◦ wX). Thus,
whp, h contains all the terms that contains variables in R and at most cs log(s/ǫ) variables not in
R. Therefore, whp, h contains at most c(c + 1)s log(s/ǫ) relevant variables. From this, and using
similar arguments as for the procedure ApproxTarget in the previous subsection, we prove that,
ApproxC makes at most O(s/ǫ) queries and
1. If f is s-term DNF then, whp, the procedure outputs X and w such that
• h = f (xX ◦ wX) is s-term DNF.
• The number of relevant variables in h = f (xX ◦ wX) is at most O(s log(s/ǫ)).
2. If f is ǫ-far from every s-term DNF then the procedure either rejects or outputs X and w
such that, whp, h = f (xX ◦ wX ) is (3ǫ/4)-far from every s-term DNF.
We can now run TesterC (with 3ǫ/4) on h from the previous subsection for testing C ∗ where C ∗ is
the set of s-term DNF with k = O(s log(s/ǫ)) relevant variables. All the procedures makes O(s/ǫ)
queries except CloseF CU that makes O(s2/ǫ) queries. This is because that the size of the class
C ∗(Γ) is 2 O(s2) and therefore CloseF CU makes O(s2/ǫ) queries. This gives a tester that makes
O(s2/ǫ) queries which is not optimal.
Instead, we consider the class C ′ of s-term DNF with O(s log(s/ǫ)) variables and terms of size
at most c log(s/ǫ) and show that, in CloseF CU , whp, all the terms T of size greater than c log(s/ǫ)
9
and all the random strings u chosen in the procedure satisfies T (u) = 0 and therefore it runs as
if the target function h has only terms of size at most c log(s/ǫ). This gives a tester that makes
O(s/ǫ) queries.
As in the previous section, all the procedures run in polynomial time except CloseF CU . For
some classes, we replace CloseF CU with polynomial time learning algorithms and obtains poly-
nomial time testers.
2 Preparing the Target for Accessing the Relevant Variables
In this Section we give the three procedures ApproxTarget, TestSets and RelVarValues.
2.1 Preliminaries
In this subsection, we give some known results that will be used in the sequel.
The following lemma is straightforward
Lemma 1. If {Xi}i∈[r] is a partition of [n] then for any Boolean function f the number of relevant
sets Xi of f is at most the number of relevant variables of f .
We will use the following folklore result that is formally proved in [45].
Lemma 2. Let {Xi}i∈[r] be a partition of [n]. Let f be a Boolean function and u, w ∈ {0, 1}n. If
f (u) 6= f (w) then a relevant set Xℓ of f with a string v ∈ {0, 1}n that satisfies f (v) 6= f (wXℓ ◦ vXℓ
)
can be found using ⌈log2 r⌉ queries.
The following is from [8]
Lemma 3. There exists a one-sided adaptive algorithm, UniformJunta(f, k, ǫ, δ), for ǫ-testing
k-junta that makes O(((k/ǫ) + k log k) log(1/δ)) queries and rejects f with probability at least 1 − δ
when it is ǫ-far from every k-junta with respect to the uniform distribution.
Moreover, it rejects only when it has found k + 1 pairwise disjoint relevant sets and a witness
of f for each one.
2.2 Approximating the Target
In this subsection we give the procedure ApproxTarget that returns (X = ∪i∈IXi, V = {v(i)}i∈I , I),
X ⊆ [n], V ⊆ {0, 1}n and I ⊆ [r] where, whp, each x(Xi), i ∈ I, contains at least one relevant
variable of h := f (xX ◦ 0X ) and exactly one if f is k-junta. Each v(i), i ∈ I, is a witness of
f (xX ◦ 0X ) for the relevant set Xi. Also, whp, f (xX ◦ 0X) is (ǫ/c)-close to the target with respect
to the distribution D.
Consider the procedure ApproxTarget in Figure 2. In steps 1-2 the procedure partitions the
set [n] into r = 2k2 disjoint sets X1, X2, . . . , Xr. In step 3 it defines the variables X, I, V and t(X).
At each iteration of the procedure, I contains the indices of some relevant sets of f (xX ◦ 0X ) where
X = ∪i∈IXi, i.e., each Xi, i ∈ I is relevant set of f (xX ◦ 0X ). The set V contains, for each i ∈ I, a
string v(i) ∈ {0, 1}n that satisfies f (v(i)
X\Xi ◦ 0Xi ◦ 0X ). That is, a witness of f (xX ◦ 0X )
for the relevant set Xi, i ∈ I.
The procedure in steps 4-19 tests if f (uX ◦ 0X ) = f (u) for at least c ln(15/k)/ǫ, independently
and at random, chosen u according to the distribution D. The variable t(X) counts the number
X ◦ 0X ) 6= f (v(i)
10
ApproxTarget(f,D, ǫ, c)
Input: Oracle that accesses a Boolean function f and
an oracle that draws x ∈ {0, 1}n according to the distribution D.
Output: Either "reject" or (X, V, I)
Partition [n] into r sets
1.
2. Choose uniformly at random a partition X1, X2, . . . , Xr of [n]
Set r = 2k2.
Choose u ∈ D.
t(X) ← t(X) + 1
If f (uX ◦ 0X ) 6= f (u) then
W ← ∅.
Binary Search to find a new relevant set from (u, uX ◦ 0X ) → ℓ;
Xℓ ◦ 0Xℓ );
and a string w(ℓ) ∈ {0, 1}n such that f (w(ℓ)) 6= f (w(ℓ)
Set X = ∅; I = ∅; V = ∅; t(X) = 0.
Find a close function and relevant sets
3.
4. Repeat M = ck ln(15k)/ǫ times
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
X ← X ∪ Xℓ; I ← I ∪ {ℓ}.
If I > k then Output("reject").
W = W ∪ {w(ℓ)}.
Choose w(r) ∈ W .
If f (w(r)
) then
X ◦ 0X ) 6= f (w(r)
X\Xr ◦ 0X∪Xr
W ← W\{w(r)}; v(r) ← w(r)
If W 6= ∅ then Goto 14
X ◦ 0X ; V ← V ∪ {v(r)};
X ◦ 0X ) 6= f (w(r)) then u ← w(r); Goto 9
Xr ◦ 0Xr ; Goto 9
Else If f (w(r)
Else u ← w(r)
16.
17.
18.
19.
t(X) = 0.
If t(X) = c ln(15k)/ǫ then Output(X, V, I).
Figure 2: A procedure that finds relevant sets {Xi}i∈I of f and a witness v(i) for each relevant
set Xi for h := f (xX ◦ 0X) where X = ∪i∈IXi. Also, whp, h is (ǫ/c)-close to the target.
of such u. If this happens then, whp, f (xX ◦ 0X ) is (ǫ/c)-close to f with respect to D and the
procedure returns (X, V, I). If not then f (uX ◦ 0X ) 6= f (u) for some u and then a new relevant set
is found. If the number of relevant sets is greater than k, it rejects. This is done in steps 8-18.
In steps 9-10, the procedure uses Lemma 2 to (binary) searches for a new relevant set. The search
gives an index ℓ of the new relevant set Xℓ and a witness w(ℓ) that satisfies f (w(ℓ)) 6= f (0Xℓ ◦ w(ℓ)
).
Then ℓ is added to I and X is extended to X ∪ Xℓ. The binary search gives a witness that Xℓ
is relevant set of f , but not a witness that it is relevant set of f (xX ◦ 0X ). This is why we need
steps 14-17. In those steps the procedure adds w(ℓ) to W . Then for each w(r) ∈ W (at the beginning
r = ℓ) it checks if w(r) is a witness of f (xX ◦ 0X ) for Xr. If it is then it adds it to V . If it isn't
Xℓ
11
then we show in the discussion below that a new relevant set can be found. The procedure rejects
when it finds more than k relevant sets.
If the procedure does not reject then it outputs (X, V, I) where I contains the indices of some
relevant sets of f (xX ◦ 0X ), X = ∪i∈IXi and the set V contains for each i ∈ I a string v(i) ∈ {0, 1}n
that is a witness of f (xX ◦ 0X ) for Xi, i.e., f (v(i)
X\Xi ◦ 0Xi ◦ 0X). We will also show
in Lemma 9 that, whp, PrD[f (xX ◦ 0X ) 6= f (x)] ≤ ǫ/c.
X ◦ 0X) 6= f (v(i)
We first prove
Lemma 4. Consider steps 1-2 in the ApproxTarget. If f is a k-junta then, with probability at
least 2/3, for each i ∈ [r], the set x(Xi) = {xjj ∈ Xi} contains at most one relevant variable of f .
Proof. Let xi1 and xi2 be two relevant variables in f . The probability that xi1 and xi2 are in the
same set is equal to 1/r. By the union bound, it follows that the probability that some relevant
variables xi1 and xi2, i1 6= i2, in f are in the same set is at most(cid:0)k
Sq
i=1 Xℓi
Xr
2(cid:1)/r ≤ 1/3.
X (j)
X (j) =Sj
X (j)\Xr
***********
***********
***********
i=j+1 Xℓi
00000000000
***********
00000000000
X (q)
000000
******
000000
*****
*****
*****
v(r)
w(r)
w(r)
X (j) ◦ 0
w(r)
Xr ◦ 0Xr
w(r)
X (j)\Xr ◦ 0
X (j) = v(r)
X (j)∪Xr
***********
00000
***********
******
***********
00000
00000000000
000000
Figure 3: The value of v(r), w(r), w(r)
any value.
X (j) ◦ 0
X (j) , w(r)
Xr ◦ 0Xr and w(r)
X (j)\Xr ◦ 0
X (j)∪Xr
where * indicates
Recall that after the binary search in step 9 the procedure has a witness w(ℓ) that satisfies
f (w(ℓ)) 6= f (w(ℓ)
Xℓ ◦ 0Xℓ ) that is not necessarily a witness of f (xX ◦ 0X) for Xℓ, i.e., does not
necessarily satisfies f (w(ℓ)
X\Xℓ ◦ 0Xℓ ◦ 0X ). This is why we first add w(ℓ) to W and
not to V . We next will show that an element w(r) in W is either a witness of f (xX ◦ 0X ) for Xℓ, in
which case we add it to V and remove it from W , or, this element generates another new relevant
set and then another witness of f is added to W .
X ◦ 0X ) 6= f (w(ℓ)
Suppose the variable ℓ in the procedure takes the values ℓ1, . . . , ℓq. Then Xℓ takes the values
Xℓ1, . . . , Xℓq and X takes the values X (i) where X (i) = X (i−1) ∪ Xℓi and X (0) = ∅. Notice that
X (0) ⊂ X (1) ⊂ ··· ⊂ X (q).
Suppose, at some iteration, the procedure chooses, in step 14, w(r) ∈ W where r = ℓi. By
step 10, f (w(r)) 6= f (w(r)
Xr ◦ 0Xr ). Suppose at this iteration X = X (j). Then r ≤ j, Xℓ1, . . . , Xℓj are
the relevant sets that are discovered so far and Xℓj+1, . . . , Xℓq ⊆ X (j). Since w(r) ∈ W , by step 11,
Xr ⊆ X (j). See the table in Figure 3.
)
then v(r) = w(r)
X (j)∪Xr
X (j) is added to the set V . This is the only step that adds an element to V .
X (j)) 6= f (w(r)
If in step 15, f (w(r)
X (j)\Xr ◦ 0
X (j) ◦ 0
X (j) ◦ 0
12
Since v(r) = w(r)
f (w(r)
X (j)\Xr ◦ 0
Therefore
X (j)∪Xr
X (j) ◦ 0
X (j) and X (j) ⊆ X (q) we have v(r)
) = f (v(r)
Xr ◦ 0Xr ).
X (q) = 0 and f (v(r)) = f (w(r)
X (j) ◦ 0
X (j)) 6=
Lemma 5. If the procedure outputs (X (q), V, I) then for every v(ℓ) ∈ V we have v(ℓ)
f (v(ℓ)) 6= f (v(ℓ)
Xℓ ◦ 0Xℓ ). That is, v(ℓ) ∈ V is a witness of f (xX (q) ◦ 0
X (q)) for Xℓ.
X (q) = 0 and
We now show that if, in step 15, f (w(r)
X ◦ 0X ) = f (w(r)
X\Xr ◦ 0X∪Xr
new relevant set.
) then the procedure finds a
Lemma 6. Consider step 15 in the procedure in the iteration where X = X (j). If w(r) is not a
witness of f (xX (j) ◦0
), then a new relevant
set is found.
X (j) ) for Xr, i.e., f (w(r)
X (j)) = f (w(r)
X (j)\Xr ◦0
X (j) ◦0
X (j)∪Xr
X (j) ◦ 0
X (j) ◦ 0
X (j)) or f (w(r)
Proof. See the table in Figure 3 throughout the proof. Since by step 10, f (w(r)) 6= f (w(r)
then either f (w(r)) 6= f (w(r)
f (w(r)
set in X (j). Step 9 finds a new relevant set because w(r) and w(r)
f (w(r)
Xr ◦ 0Xr ) then the procedure in step 17 assign u = w(r)
to step 9 to find a relevant set in X (j). Step 9 finds a new relevant set because w(r)
and w(r)
Xr ◦ 0Xr ),
If f (w(r)) 6=
X (j)) then the procedure in step 16 assign u = w(r) and goes to step 9 to find a relevant
X (j) are equal on X (j). If
Xr ◦ 0Xr and goes
X (j)\Xr ◦ 0
) 6= f (w(r)
) 6= f (w(r)
X (j)\Xr ◦ 0
Xr ◦ 0Xr ).
X (j)\Xr ◦ 0
X (j) ◦ 0
X (j)∪Xr
X (j)∪Xr
X (j)∪Xr
Xr ◦ 0Xr are equal on X (j).
Therefore, for every w(r) ∈ W the procedure either finds v(r) that satisfies the condition in
Lemma 5 or finds a new relevant set. If the number of relevant sets is greater than k, then the
procedure rejects. This is because each relevant set contains a relevant variable, and the relevant
sets are disjoint. So the function, in this case, is not k-junta and therefore not in C. If the number
of relevant sets is less than or equal to k, then the algorithm eventually finds, for each ℓ ∈ I, a
witness v(ℓ) of f (xX ◦ 0X ) for Xiℓ. This implies
Lemma 7. If ApproxTarget does not reject then it outputs (X = X (q), V = {v(ℓ1), . . . , v(ℓq)}, I =
{ℓ1, . . . , ℓq}) that satisfies
1. q = I ≤ k.
2. For every ℓ ∈ I, v(ℓ)
f (xX ◦ 0X ) for Xℓ .
X
= 0 and f (v(ℓ)) 6= f (0Xℓ ◦ v(ℓ)
Xℓ
). That is, v(ℓ) ∈ V is a witness of
3. Each x(Xℓ), ℓ ∈ I, contains at least one relevant variable of f (xX ◦ 0X ).
Lemma 8. If f is k-junta and each x(Xi) contains at most one relevant variable of f then
1. ApproxTarget outputs (X, V, I).
2. Each x(Xℓ), ℓ ∈ I, contains exactly one relevant variable in f (xX ◦ 0X ).
3. For every ℓ ∈ I, f (xXℓ ◦ v(ℓ)
) is a literal.
Xℓ
13
Proof. By 3 in Lemma 7, x(Xℓ), ℓ ∈ I, contains exactly one relevant variable. Thus, for every
ℓ ∈ I, f (xXℓ ◦ v(ℓ)
Since f contains at most k relevant variables, by Lemma 1, the number of relevant sets I is at
) is a literal.
most k. Therefore, ApproxTarget does not halt in step 12.
Xℓ
The following lemma shows that
Lemma 9. If ApproxTarget outputs (X, V, I) then I ≤ k and with probability at least 14/15
Pru∈D[f (uX ◦ 0X ) 6= f (u)] ≤ ǫ/c.
Proof. If I > k then, from step 12, ApproxTarget outputs "reject". Therefore, the probabil-
ity that ApproxTarget fails to output (X, V, I) with Pru∈D[f (uX ◦ 0X ) 6= f (u)] ≤ ǫ/c is the
probability that for some X (ℓ), Pru∈D[f (xX (ℓ) ◦ 0
X (ℓ)) = f (u)
for c ln(15k)/ǫ strings u chosen independently at random according to the distribution D. This
probability is at most
X (ℓ)) 6= f (x)] > ǫ/c and f (uX (ℓ) ◦ 0
k(cid:16)1 −
c
ǫ(cid:17)c ln(15k)/ǫ
1
15
.
≤
We now give the query complexity
Lemma 10. The procedure ApproxTarget makes O((k log k)/ǫ) queries.
Proof. The condition in step 7 requires two queries and is executed at most M = ck ln(15k)/ǫ times.
This is 2M = O((k log k)/ǫ) queries. Steps 9-17 are executed at most k + 1 times. This is because
each time it is executed, the value of I is increased by one, and when I = k + 1 the procedure
rejects. By Lemma 2, to find a new relevant set the procedure makes O(log r) = O(log k) queries.
This gives another O(k log k) queries. Therefore, the query complexity is O((k log k)/ǫ).
2.3 Testing the Relevant Sets
In this subsection we give the procedure TestSets that takes as an input (X, V = {v(ℓ1), . . . , v(ℓq)}, I =
{ℓ1, . . . , ℓq}) and tests if for all ℓ ∈ I, f (xXℓ ◦ v(ℓ)
) is (1/30)-close to some literal with respect to
Xℓ
the uniform distribution.
We first prove
Lemma 11. If f is k-junta and each x(Xi) contains at most one relevant variable of f then
TestSets returns "OK".
Proof. By Lemma 8, for every ℓ ∈ I, f (xXℓ ◦ v(ℓ)
Xℓ
) is a literal.
If TestSets rejects in step 3 then, by Lemma 3, for some Xℓ, ℓ ∈ I, f (xXℓ ◦ v(ℓ)
Xℓ
(literal or constant function) and therefore x(Xℓ) contains at least two relevant variables.
rejects in step 5, then f (bXℓ ◦ v(ℓ)
) and then f (xXℓ ◦ v(ℓ)
) = f (bXℓ ◦ v(ℓ)
Xℓ
Xℓ
Xℓ
) is not 1-Junta
If it
) is not a literal. In all cases
we get a contradiction.
In the following lemma we show that if TestSets returns "OK" then, whp, each f (xXℓ ◦ v(ℓ)
Xℓ
is close to a literal with respect to the uniform distribution.
)
14
TestSets(X, V, I)
Input: Oracle that accesses a Boolean function f and (X, V, I).
Output: Either "reject" or "OK"
1. For every ℓ ∈ I do
2.
3.
4.
5.
6. Return "OK"
If UniformJunta(f (xXℓ ◦ v(ℓ)
Choose b ∈ U
If f (bXℓ ◦ v(ℓ)
) = f (bXℓ ◦ v(ℓ)
then Output("reject")
Xℓ
Xℓ
Xℓ
), 1, 1/30, 1/15)="reject"
) then Output("reject")
Figure 4: A procedure that tests if for all ℓ ∈ I, f (xXℓ ◦ v(ℓ)
Xℓ
respect to the uniform distribution.
) is (1/30)-close to some literal with
Xℓ
Xℓ
) is (1/30)-far from every literal with respect to the
Lemma 12. If for some ℓ ∈ I, f (xXℓ ◦ v(ℓ)
uniform distribution then, with probability at least 1 − (1/15), TestSets rejects.
Proof. If f (xXℓ◦v(ℓ)
) is (1/30)-far from every literal with respect to the uniform distribution then it
is either (case 1) (1/30)-far from every 1-Junta (literal or constant) or (case 2) (1/30)-far from every
literal and (1/30)-close to 0-Junta. In case 1, by Lemma 3, with probability at least 1 − (1/15),
UniformJunta (f (xXℓ ◦ v(ℓ)
), 1, 1/30, 1/15) = "reject" and then the procedure rejects. In case 2,
if f (xXℓ ◦ v(ℓ)
is random uniform and for g(x) = f (xXℓ ◦ v(ℓ)
to 1. Suppose it is (1/30)-close to 0. Let b be a random uniform string chosen in steps 4. Then b
) is (1/30)-close to some 0-Junta then it is either (1/30)-close to 0 or (1/30)-close
) we have
Xℓ
Xℓ
Xℓ
Pr[The procedure does not reject] = Pr(cid:2)g(b) 6= g(b)(cid:3)
= Pr[g(b) = 1 ∧ g(b) = 0] + Pr[g(b) = 0 ∧ g(b) = 1]
≤ Pr[g(b) = 1] + Pr[g(b) = 1]
≤
1
15
.
Lemma 13. The procedure TestSets makes O(k) queries.
Proof. Steps 2 and 5 are executed I ≤ k times, and by Lemma 3, the total number of queries
made is O(1/(1/30) log(15))k + 2k = O(k).
2.4 Determining the Values of the Relevant Variables
In this subsection we give a procedure RelVarValue that for an input (w ∈ {0, 1}n, X, V, I, δ)
where (X, V, I) satisfies all the properties in the previous two subsections, the procedure, with
15
RelVarValues(w, X, V, I, δ)
Input: Oracle that accesses a Boolean function f , (X, V, I) and w ∈ {0, 1}n.
Output: Either "reject" or for every ℓ ∈ I, the value, zℓ = wτ (ℓ) where xτ (ℓ) is one of the
relevant variables of f (xX ◦ 0X ) in x(Xℓ)
For ξ ∈ {0, 1} set Yℓ,ξ = {j ∈ Xℓwj = ξ}.
Set Gℓ,0 = Gℓ,1 = 0;
Repeat h = ln(k/δ)/ ln(4/3) times
Choose b ∈ U ;
1. For every ℓ ∈ I do
2.
3.
4.
5.
6.
7.
8.
9.
10. Output("{zℓ}ℓ∈I ")
If f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
If f (bYℓ,1 ◦ bYℓ,0 ◦ v(ℓ)
Xℓ
Xℓ
) 6= f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) 6= f (bYℓ,1 ◦ bYℓ,0 ◦ v(ℓ)
Xℓ
Xℓ
) then Gℓ,0 ← Gℓ,0 + 1
) then Gℓ,1 ← Gℓ,1 + 1
If ({Gℓ,0, Gℓ,1} 6= {0, h}) then Output("reject")
If Gℓ,0 = h then zℓ ← 0 else zℓ ← 1
Figure 5: A procedure that takes as input (X, V, I) and a string w ∈ {0, 1}n and, with probability
at least 1 − δ, returns the values of wτ (i), i ∈ I, where f (xXi ◦ v(i)
) is (1/30)-close to one of the
literals in {xτ (i), xτ (i)} with respect to the uniform distribution.
Xi
probability at least 1 − δ, returns the values of wτ (i), i ∈ I, where f (xXi ◦ v(i)
) is (1/30)-close to
one of the literals in {xτ (i), xτ (i)} with respect to the uniform distribution. When f is k-junta and
each x(Xi) contains at most one relevant variable then {xτ (i)}i∈I is the set of the relevant variables
of f (xX ◦ 0X ) and wτ (i), i ∈ I are the values of the relevant variables. The procedure is in Figure 5.
Xi
We first prove
Lemma 14. If f is k-Junta and each x(Xi) contains at most one relevant variable of f then
RelVarValues outputs z such that zℓ = wτ (ℓ) where f (xXℓ ◦ 0Xℓ
Proof. Since Yℓ,0, Yℓ,1 is a partition of Xℓ, ℓ ∈ I and, by Lemma 8, x(Xℓ) contains exactly one
relevant variable xτ (ℓ) of f (xX ◦ 0X ), this variable is either in x(Yℓ,0) or in x(Yℓ,1) but not in both.
Suppose w.l.o.g. it is in x(Yℓ,0) and not in x(Yℓ,1). Then wτ (ℓ) = 0, f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) is a literal
and f (xYℓ,1 ◦ bYℓ,0 ◦ v(ℓ)
) 6=
f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
). Therefore, by steps 6-7 in the
procedure, Gℓ,0 = h and Gℓ,1 = 0 and the procedure does not output reject in step 8. Thus, by
step 9, zℓ = wτ (ℓ).
) is a constant function. This implies that for any b, f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) and f (bYℓ,1 ◦ bYℓ,0 ◦ v(ℓ)
) = f (bYℓ,1 ◦ bYℓ,0 ◦ v(ℓ)
) ∈ {xτ (ℓ), xτ (ℓ)}.
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
We now prove
Lemma 15. If for every ℓ ∈ I the function f (xXℓ ◦ v(ℓ)
) is (1/30)-close to a literal in {xτ (ℓ), ¯xτ (ℓ)}
with respect to the uniform distribution, where τ (ℓ) ∈ Xℓ, and in RelVarValues, for every ℓ ∈ I,
{Gℓ,0, Gℓ,1} = {0, h} then, with probability at least 1 − δ, we have: For every ℓ ∈ I, zℓ = wτ (ℓ).
Xℓ
16
Proof. Fix some ℓ. Suppose f (xXℓ ◦ v(ℓ)
) is (1/30)-close to xτ (ℓ) with respect to the uniform
distribution. The case when it is (1/30)-close to xτ (ℓ) is similar. Since Xℓ = Yℓ,0 ∪ Yℓ,1 and
Yℓ,0 ∩ Yℓ,1 = ∅ we have that τ (ℓ) ∈ Yℓ,0 or τ (ℓ) ∈ Yℓ,1, but not both. Suppose τ (ℓ) ∈ Yℓ,0. The case
where τ (ℓ) ∈ Yℓ,1 is similar. Define the random variable Z(xXℓ) = 1 if f (xXℓ ◦ v(ℓ)
) 6= xτ (ℓ) and
Z(xXℓ) = 0 otherwise. Then
Xℓ
Xℓ
ExXℓ
∈U [Z(xXℓ)] ≤
1
30
.
Therefore
and by Markov's bound
ExYℓ,1 ∈U ExYℓ,0 ∈U [Z(xYℓ,0 ◦ xYℓ,1)] ≤
1
30
PrxYℓ,1 ∈U(cid:20)ExYℓ,0 ∈U [Z(xYℓ,0 ◦ xYℓ,1)] ≥
2
15(cid:21) ≤
1
4
.
That is, for a random uniform string b ∈ {0, 1}n, with probability at least 3/4, f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
is (2/15)-close to xτ (ℓ) with respect to the uniform distribution. Now, given that f (xYℓ,0 ◦bYℓ,1 ◦v(ℓ)
)
Xℓ
is (2/15)-close to xτ (ℓ) with respect to the uniform distribution the probability that Gℓ,0 = 0 is the
probability that f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) for h random uniform strings b ∈ {0, 1}n.
Let b(1), . . . , b(h) be h random uniform strings in {0, 1}n, V (b) be the event f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) =
) and A the event that f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
to the uniform distribution. Let g(xYℓ,0) = f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) = f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) is (2/15)-close to xτ (ℓ) with respect
). Then
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
)
Pr[V (b)A] = Pr[g(bYℓ,0) = g(bYℓ,0)A]
= Pr[(g(bYℓ,0 ) = bτ (ℓ) ∧ g(bYℓ,0) = bτ (ℓ)) ∨ (g(bYℓ,0 ) = bτ (ℓ) ∧ g(bYℓ,0 ) = bτ (ℓ))A]
≤ Pr[g(bYℓ,0) 6= bτ (ℓ) ∨ g(bYℓ,0) 6= bτ (ℓ))A]
≤ Pr[g(bYℓ,0) 6= bτ (ℓ)A] + Pr[g(bYℓ,0 ) 6= bτ (ℓ))A] ≤
4
15
.
Since τ (ℓ) ∈ Yℓ,0, we have wτ (ℓ) = 0. Therefore, by step 9 and since τ (ℓ) ∈ Xℓ,
Pr[zℓ 6= wτ (ℓ)] = Pr[zℓ = 1]
= Pr[Gℓ,0 = 0 ∧ Gℓ,1 = h]
≤ Pr[Gℓ,0 = 0] = Pr[(∀j ∈ [h])V (b(j))]
= (Pr[V (b)])h ≤(cid:0)Pr[V (b)A] + Pr[A](cid:1)h
≤ (4/15 + 1/4)h ≤ (3/4)h
Therefore, the probability that zℓ 6= wτ (ℓ) for some ℓ ∈ I is at most k(3/4)h ≤ δ.
The following is obvious
Lemma 16. The procedure RelVarValues makes O(k log(k/δ)) queries.
17
3 Testing Subclasses of k-Junta
In this section, we give testers for subclasses of k-Junta that are closed under variable and zero
projections.
Our tester will start by running the two procedures ApproxTarget and TestSets and there-
fore, by Lemmas 4, 8 and 11, if f ∈ C (and therefore is k-junta) then, with probability at least 2/3,
both procedures do not reject and item 1 in the following Assumption happens. By Lemmas 7,
9, and 12, if f is ǫ-far from every function in C and both procedures do not reject then, with
probability at least 13/15, item 2 in the following Assumption happens. Obviously, the above two
probabilities can be changed to 1 − δ for any constant δ without changing the asymptotic query
complexity.
Assumption 17. Throughout this section we assume that there are X, q ≤ k, I = {ℓ1, . . . , ℓq}
and V = {v(ℓ1), . . . , v(ℓq)} such that: For every ℓ ∈ I, v(ℓ)
). That is,
v(ℓ) ∈ V is a witness of f (xX ◦ 0X) for Xℓ and
= 0 and f (v(ℓ)) 6= f (0Xℓ ◦ v(ℓ)
Xℓ
X
1. If f ∈ C (and therefore is k-junta)
• f (xX ◦ 0X ) ∈ C.
• Each x(Xℓ), ℓ ∈ I contains exactly one relevant variable.
• For every ℓ ∈ I, f (xXℓ ◦ v(ℓ)
) is a literal in {xτ (ℓ), xτ (ℓ)}.
Xℓ
2. If f is ǫ-far from every function in C then
• f (xX ◦ 0X) is (ǫ/3)-close to f with respect to D and therefore f (xX ◦ 0X ) is (2ǫ/3)-far
from every function in C with respect to D.
• For every ℓ ∈ I, f (xXℓ ◦ v(ℓ)
) is (1/30)-close to a literal in {xτ (ℓ), ¯xτ (ℓ)} with respect to
Xℓ
the uniform distribution.
We will also use the set of indices Γ := {τ (ℓ1), . . . , τ (ℓq)}. Notice that if f is k-junta then x(Γ) are
the relevant variables of f .
We remind the reader that for a projection π : X → X the string x(π) is defined as x(π)j = xπ(j)
for every j ∈ X. Define the projection πf,I : X → X that satisfies: For every ℓ ∈ I and every
j ∈ Xℓ, πf,I(j) = τ (ℓ). Define the function F (xΓ) = F (xτ (ℓ1), . . . , xτ (ℓq)) := f (x(πf,I) ◦ 0X ). That
is, F is the function that results from substituting in f (xX ◦ 0X) for every ℓ ∈ I and every xi,
i ∈ Xℓ, the variable xτ (ℓ). Note here that the tester does not know τ (ℓ1), . . . , τ (ℓq).
We now show how to query F by querying f
Lemma 18. For the function F we have
1. Given (y1, . . . , yq), computing F (y1, . . . , yq) can be done with one query to f .
2. Given x ∈ {0, 1}n and δ, there is an algorithm that makes O(k log(k/δ)) queries and, with
probability at least 1 − δ, either discovers that some Xi, i ∈ I contains at least two rel-
evant variables in f (and therefore, whp, f is not k-junta) and then rejects or computes
z = (xτ (ℓ1), . . . , xτ (ℓq)) and F (z).
18
Proof. 1 is immediate. To prove 2 we use Lemma 15. We run RelVarValues(x, X, V, I, δ).
If
it rejects then {Gℓ,0, Gℓ,1} 6= {0, h} for some ℓ ∈ I and therefore Gℓ,0, Gℓ,1 > 0. This implies
that for some b, b′ ∈ {0, 1}n, f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) 6=
f (b′
). Since Xℓ = Yℓ,0 ∪ Yℓ,1 and Yℓ,0 ∩ Yℓ,1 = ∅, the set x(Xℓ) contains at least two
relevant variables in f .
Yℓ,1 ◦ b′
If for every ℓ we have {Gℓ,0, Gℓ,1} = {0, h} then, by Lemma 15, with probability at least 1 − δ,
the procedure outputs z where for every ℓ, zℓ = xτ (ℓ). Then using 1 we compute F (z). Since by
Lemma 16, RelVarValue makes O(k log(k/δ)) queries, the result follows.
) 6= f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
Yℓ,0 ◦ v(ℓ)
Xℓ
Yℓ,0 ◦ v(ℓ)
Xℓ
Xℓ
) and f (b′
Yℓ,1 ◦ b′
Xℓ
We now give the key lemma for the first tester
Lemma 19. Let C ⊆ k−Junta be a class that is closed under variable and zero projections and f
be any Boolean function. Let F (xΓ) = f (x(πf,I ) ◦ 0X ) where Γ = {τ (ℓ)ℓ ∈ I} and C(Γ) be the set
of all functions in C that their relevant variables are x(Γ). If Assumption 17 is true, then
1. If f ∈ C then f (xX ◦ 0X ) = F ∈ C(Γ).
2. If f is ǫ-far from every function in C with respect to D then either
(a) f (xX ◦ 0X ) is (ǫ/3)-far from F with respect to D,
or
(b) F is (ǫ/3)-far from every function in C(Γ) with respect to D.
If f ∈ C, then since C is closed under variable and zero projection
Proof. We first prove 1.
f (xX ◦ 0X) ∈ C. We have f (xX ◦ 0X ) = f (◦ℓ∈I xXℓ ◦ 0X) and, by Assumption 17, every x(Xℓ),
ℓ ∈ I, contains exactly one relevant variable xτ (ℓ) of f (xX ◦ 0X ). Therefore, f (xX ◦ 0X ) is a function
that depends only on the variables xτ (ℓ), ℓ ∈ I. By the definition of x(πf,I ) and since τ (ℓ) ∈ Xℓ we
have x(πf,I )τ (ℓ) = xτ (ℓ) and therefore f (xX ◦ 0X) = f (x(πf,I ) ◦ 0X ) = F .
We now prove 2. Suppose, for the contrary, f (xX ◦ 0X) is (ǫ/3)-close to F with respect to D
and F is (ǫ/3)-close to some function g ∈ C(Γ) with respect to D. Then f (xX ◦ 0X ) is (2ǫ/3)-close
to g with respect to D. Since, by Assumption 17, f (xX ◦ 0X ) is (ǫ/3)-close to f with respect to D
we get that f is ǫ-close to g ∈ C with respect to D. A contradiction.
In the following two subsections we discuss how to test the closeness of f (xX ◦ 0X) to F and F
to C(Γ). We will assume all the procedures in the following subsections have access to X, V, I that
satisfies Assumption 17.
3.1 Testing the Closeness of f (xX ◦ 0X) to F
We now give the procedure Closef F that tests whether f (xX ◦0X ) is (ǫ/3)-far from F with respect
to D. See Figure 6.
Lemma 20. For any ǫ, a constant δ, and (X, V, I) that satisfies Assumption 17, procedure Closef F
makes O((k/ǫ) log(k/ǫ)) queries and
1. If f ∈ C then Closef F returns OK.
19
Closef F (f,D, ǫ, δ)
Input: Oracle that accesses a Boolean function f and D.
Output: Either "reject" or "OK"
1. Define F ≡ f (x(πf,I) ◦ 0X).
2. Repeat t = (3/ǫ) ln(2/δ) times
3.
4.
5.
6. Return "OK".
Choose u ∈ D.
z ←RelVarValue(u, X, V, I, δ/(2t)) .
If f (uX ◦ 0X) 6= F (z) then Output("reject")
Figure 6: A procedure that tests whether f (xX ◦ 0X ) is (ǫ/3)-far from F with respect to D.
2. If f (xX ◦ 0X ) is (ǫ/3)-far from F with respect to D then, with probability at least 1 − δ,
Closef F rejects.
Otherwise it rejects.
Assumption 17, z(i) = u(i)
Γ ) = f (u(i)
Γ and if F (u(i)
Γ ) = f (u(i)
Proof. Closef F draws t = (3/ǫ) ln(2/δ) random u(i) ∈ {0, 1}n, i = 1, . . . , t according to the
distribution D. It finds z(i) = u(i)
X ◦ 0X ) for all i then it returns "OK".
X ◦ 0X ) for every i. By Lemma 14 and
If f ∈ C then, by 1 in Lemma 19, F (u(i)
Suppose now f (xX◦0X ) is (ǫ/3)-far from F with respect to D. By 2 in Lemma 18, RelVarValue
makes O(k log((kt)/δ)) queries and computes F (u(i)
Γ ), i = 1, . . . , t, with failure probability at most
δ/2. Then the probability that it fails to reject is at most (1 − ǫ/3)t ≤ δ/2. This gives the result.
Γ for all i, and therefore Closef F returns OK.
Therefore, Closef F makes O((k/ǫ) log(k/ǫ)) queries and satisfies 1 and 2.
3.2 Testing the Closeness of F to C(Γ)
In this section, we give the procedures CloseF CD and CloseF CU that test whether F is (ǫ/3)-far
from every function in C(Γ) with respect to D and the uniform distribution, respectively. We prove
Lemma 21. For any ǫ and any constant δ and (X, V, I) that satisfies Assumption 17, the pro-
cedures CloseF CD and CloseF CU make O((k log C(Γ))/ǫ) and O((log C(Γ))/ǫ) queries to f ,
respectively, and
1. If f ∈ C then CloseF CD and CloseF CU output OK.
2. If F is (ǫ/3)-far from every function in C(Γ) with respect to D then, with probability at least
1 − δ, CloseF CD rejects.
3. If F is (ǫ/3)-far from every function in C(Γ) with respect to the uniform distribution, then
with probability at least 1 − δ, CloseF CU rejects.
Both procedures run in time poly(n,C(Γ), 1/ǫ).
20
CloseF CD(f,D, ǫ, δ)
Input: Oracles that access a Boolean function f and D.
Output: Either "reject" or "OK"
Choose u ∈ D.
z ←RelVarValue(u, X, V, I, 1/2) .
For every g ∈ C ∗
If C ∗ = ∅ then Output("Reject")
1. C ∗ ← C(Γ)
2. Repeat τ = (12/ǫ) ln(2C ∗/δ) times
3.
4.
5.
6.
7.
8. Return "OK"
-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -
CloseF CU (f, ǫ, δ)
Input: Oracle that accesses a Boolean function f .
Output: Either "reject" or "OK"
If g(z) 6= F (z) then C ∗ ← C ∗\{g}.
1. C ∗ ← C(Γ)
2. Repeat τ = (3/ǫ) ln(2C ∗/δ)) times
3.
4.
5.
6.
7. Return "OK"
Choose (z1, . . . , zq) ∈ U .
For every g ∈ C ∗
If C ∗ = ∅ then Output("Reject")
If g(z) 6= F (z) then C ∗ ← C ∗\{g}.
Figure 7: Two procedures that test whether F is (ǫ/3)-far from every function in C(Γ) with respect
to D and the uniform distribution, respectively.
Proof. The proof for CloseF CU is similar to the proof of Lemma 20 with union bound.
For CloseF CD, notice that it calls RelVarValue(u, X, V, I, 1/2), and therefore at each itera-
tion, with probability 1/2, z = uΓ. By Chernoff's bound ((9) in Lemma 69), with probability at
least 1 − δ/2, (3/ǫ) ln(2C ∗/δ) of the chosen us in the procedures satisfy z = uΓ. Then again, by
union bound, the result follows.
3.3 Testing the Closeness of F to C(Γ) via Learning C(Γ)
In this subsection, we show how proper learning implies testing the closeness of F to C(Γ). The
proofs are similar to the proof of Proposition 3.1.1 in [38].
Let (X, V, I) be as in Assumption 17 and q = I ≤ k. Let Y = {y1, . . . , yq} be a set of Boolean
variables and C(Y ) be the set of all functions in C that depend on all the variables of Y . Notice
that instead of using C(Y ) we could have used C({x1, . . . , xq}) but here we use the new Boolean
variables yi to avoid confusion with the variables xi of f .
Remark 22. In all the lemmas in this subsection and the following one, in addition to the fact
21
that F depends on all the variables of Y , the learning algorithms can also make use of (X, V, I) that
satisfies Assumption 17. This may help for some classes. For example1, if the target function is a
unate monotone function, then from the witnesses in V , we can know if F is positive or negative
unate in yi, for each variable yi.
The following is an immediate result that follows from the two procedures CloseF CD and
CloseF CU in the previous subsection
Lemma 23. If there is a polynomial time algorithm that given a set
Y = {(y(1), ξ1), . . . , (y(t), ξt)} ⊆ {0, 1}q × {0, 1}
decides whether there is a function F ∈ C(Y ) that is consistent with Y, i.e., F (y(i)) = ξi for all
i = 1, . . . , t, then there is a polynomial time algorithm BD (resp. BU ) that makes O((k log C(Γ))/ǫ)
queries (resp. O((log C(Γ))/ǫ) queries) to f and
1. If f ∈ C then BD and BU output OK.
2. If F is (ǫ/3)-far from every function in C(Γ) with respect to D then, with probability at least
1 − δ, BD rejects.
3. If F is (ǫ/3)-far from every function in C(Γ) with respect to the uniform distribution then,
with probability at least 1 − δ, BU rejects.
We now give the reduction from exact learning
Lemma 24. If there is a polynomial time algorithm A that, given as an input any constant δ,
properly exactly learns C(Y ) with confidence parameter δ and makes M (δ) membership queries to
F then there is a polynomial time algorithm B that, given as an input ǫ, any constant δ, makes
M (δ/3) + O((k/ǫ) log(1/ǫ)) (resp. M (δ/3) + O(1/ǫ)) queries to f and
1. If f ∈ C then, with probability at least 1 − δ, B outputs OK.
2. If F is (ǫ/3)-far from every function in C(Γ) with respect to D (resp. with respect to the
uniform distribution) then, with probability at least 1 − δ, B rejects.
Proof. Algorithm B runs A with confidence parameter δ/3 to learn F (y1, . . . , yq). By Lemma 18,
each membership query to F can be simulated by one membership query to f . If algorithm A
runs more than it should, asks more than M (δ/3) membership queries or outputs h 6∈ C(Y ) then
B rejects.
If A outputs h ∈ C(Y ) then the algorithm needs to distinguish whether h(xΓ) is equal to F (xΓ)
or (ǫ/3)-far from F (xΓ) with respect to the distribution D (resp. uniform distribution). When the
distribution is uniform, the algorithm chooses t = (3/ǫ) ln(3/δ) strings v(1), . . . , v(t) ∈ {0, 1}q and
if F (v(i)) = h(v(i)) for all i then it outputs "OK"; otherwise it rejects.
In the distribution-free model, B chooses t = (12/ǫ) ln(2/δ) strings u(i) ∈ {0, 1}n according to
the distribution D. Then runs RelValValue(u(i), X, V, I, 1/2) to find, with probability 1/2 the
value of u(i)
Γ ) for all i, then it outputs "OK"; otherwise it rejects.
The analysis and correctness of the algorithm are the same as in the above proofs and Propo-
Γ , i = 1, . . . , t. If F (u(i)
Γ ) = h(u(i)
sition 3.1.1 in [38].
1See the definition of unate in Subsection 4.4
22
We now give the reduction from learning from MQ and ExQD
Lemma 25. If there is a polynomial time algorithm A that, given as an input a constant δ, any
ǫ, learns C(Y ) with respect to the distribution D (resp. uniform distribution), with confident δ,
accuracy ǫ, makes M (ǫ, δ) M Q to F and Q(ǫ, δ) ExQD (resp. ExQU ) queries to F then there is a
polynomial time algorithm BD (resp. BU ) that asks
queries (resp.
O(cid:18)M (ǫ/12, δ/3) + kQ(ǫ/12, δ/3) log(kQ(ǫ/12, δ/3)) +
ǫ(cid:19)
O(cid:18)M (ǫ/12, δ/3) + Q(ǫ/12, δ/3) +
1
k
ǫ
log
1
ǫ(cid:19)
queries) to f and
1. If f ∈ C then with probability at least 1 − δ, BD and BU output OK.
2. If F is (ǫ/3)-far from every function in C(Γ) with respect to D then, with probability at least
1 − δ, BD rejects.
3. If F is (ǫ/3)-far from every function in C(Γ) with respect to the uniform distribution then,
with probability at least 1 − δ, BU rejects.
Proof. Algorithm B runs A with confidence parameter δ/3 and accuracy ǫ/12. By Lemma 18,
every membership query to F (y) can be simulated with one membership query to f . Every ExQD′
(resp. ExQ) for the induced distribution D′ of D on the coordinates Γ, can be simulated with one
ExQD and k log(3kQ(ǫ/12, δ/3)/δ) membership queries (resp. one ExQ) with failure probability
δ/(3Q(ǫ/12, δ/3)), and therefore, with failure probability δ/3 for all the ExQD′ queries asked in the
learning algorithm.
If algorithm A runs more than it should, asks more than Q(ǫ/12, δ/3) ExQD′, asks more than
If A outputs h ∈ C(Y ) then, with
M (ǫ/12, δ/3) MQ or outputs h 6∈ C(Y ) then BD rejects.
probability at least 1 − (2δ/3),
1. If F ∈ C(Γ) then F is (ǫ/12)-close to h
2. If F is (ǫ/3)-far from every function in C(Γ) then F is (ǫ/4)-far from h.
Now, using Chernoff's bound in Lemma 69, algorithm B, can estimate the distance between F
and h with accuracy ǫ/24 and confidence δ/6 using O((log(1/δ))/ǫ) strings chosen according to
the distribution D′. This can be done using O((log(1/δ))/ǫ) queries in the uniform model and
O((k/ǫ)(log(1/δ) log(1/(ǫδ))) with confidence δ/6 in the distribution-free model.
Define the oracle WExQD (Weak ExQD) that returns with probability 1/2 a x ∈ {0, 1}n ac-
cording to the distribution D and with probability 1/2 an arbitrary x ∈ {0, 1}n. In some of the
learning algorithms given in the sequel, the algorithms still work if we replace the oracle ExQD with
WExQD. In that case, we can save the factor of log(3kQ(ǫ/12, δ/3)/δ) in the query complexity of
Lemma 25 in the distribution-free setting. We will discuss this in Section 8.
23
3.4 The First Tester
We are now ready to give the first tester.
Consider the tester TesterC in Figure 8. Note that the tester rejects if any one of the procedures
called by the tester rejects. We prove
TesterC(f,D, ǫ)
Input: Oracle that accesses a Boolean function f and D.
Output: If any one of the procedures reject
then "reject" or "accept"
(X, V, I) ←ApproxTarget(f,D, ǫ, 1/3).
1.
2. TestSets(X, V, I).
3. Define F ≡ f (x(πf,I ) ◦ 0X )
4. Closef F (f,D, ǫ, 1/15)
For any distribution
5. CloseF CD(f,D, ǫ, 1/15)
6. Return "accept"
For the uniform distribution
5. CloseF CU (f, ǫ, 1/15)
6. Return "accept"
Figure 8: A tester for subclasses C of k-Junta
Theorem 26. Let C ⊆ k−Junta that is closed under zero and variable projections. Then
1. There is a poly(C(Γ), n, 1/ǫ) time two-sided adaptive algorithm, TesterC, for ǫ-testing C
that makes O((1/ǫ)(k + log C(Γ))) queries. That is
(a) If f ∈ C then, with probability at least 2/3, TesterC accepts.
(b) If f is ǫ-far from every function in C with respect to the uniform distribution then, with
probability at least 2/3, TesterC rejects.
2. There is a poly(C(Γ), n, 1/ǫ) time two-sided distribution-free adaptive algorithm, TesterC,
for ǫ-testing C that makes O((k/ǫ) log(2C(Γ))) queries. That is
(a) If f ∈ C then, with probability at least 2/3, TesterC accepts.
(b) If f is ǫ-far from every function in C with respect to the distribution D then, with
probability at least 2/3, TesterC rejects.
Proof. We prove 1a and 2a. Let f ∈ C. Consider step 1 in TesterC. By Lemma 4 and 8, with
probability at least 2/3, ApproxTarget outputs (X, V, I) that satisfies Assumption 17. Now with
this assumption we have: By Lemma 11, TestSets in step 2 does not reject. By Lemma 20,
24
Closef F in step 4 does not reject. By Lemma 21, CloseF CD and CloseF CU in step 5 do not
reject. Therefore, with probability at least 2/3 the tester accepts.
We now prove 1b and 2b. Suppose f is ǫ-far from every function in C with respect to D. The
proof for the uniform distribution is similar. If in step 1 ApproxTarget outputs (X, V, I) then by
Lemma 9, with probability at least 14/15, f (xX ◦ 0X ) is (ǫ/3)-close to f . If in step 2 TestSets
does not reject then, by Lemma 12, with probability at least 14/15, for all ℓ ∈ I, f (xXℓ ◦ v(ℓ)
) is
(1/30)-close to a literal {xτ (ℓ), xτ (ℓ)}. Therefore with probability at least 13/15, Assumption 17 is
true. Then by Lemma 19, either f (xX ◦ 0X ) is (ǫ/3)-far from F with respect to D, or F is (ǫ/3)-far
from every function in C(Γ) with respect to D. If f (xX ◦ 0X ) is (ǫ/3)-far from F with respect to D
then by Lemma 20, with probability at least 14/15, Closef F rejects. If F is (ǫ/3)-far from every
function in C(Γ) with respect to D then by Lemma 21, with probability at least 14/15, CloseF C
rejects. By the union bound the probability that the tester rejects is at least 2/3.
Xℓ
The query complexity follows from Lemmas 10, 13, 20 and 21.
If we replace CloseF CD and CloseF CU with the testers in Lemma 23, 24 and 25, we get the
following results
Theorem 27. If there is a polynomial time algorithm that given a set
Y = {(y(1), ξ1), . . . , (y(t), ξt)} ⊆ {0, 1}q × {0, 1}
decides whether there is a function F ∈ C(Y ) that is consistent with Y, then
1. There is a polynomial time two-sided adaptive algorithm for ǫ-testing C that makes O((1/ǫ)(k+
log C(Γ))) queries.
2. There is a polynomial time two-sided distribution-free adaptive algorithm for ǫ-testing C that
makes O((k/ǫ) log(2C(Γ))) queries.
Theorem 28. If there is a polynomial time algorithm A that, given as an input any constant δ,
properly exactly learns C(Y ) with confidence parameter δ and makes M (δ) membership queries then
1. There is a polynomial time two-sided adaptive algorithm for ǫ-testing C that makes M (1/24)+
O(k/ǫ) queries.
2. There is a polynomial time two-sided distribution-free adaptive algorithm for ǫ-testing C that
makes M (1/24) + O(k/ǫ) queries.
Theorem 29. If there is a polynomial time algorithm A that, given as an input a constant δ and
any ǫ, learns C(Y ), with confident δ, accuracy ǫ, makes M (ǫ, δ) M Q and Q(ǫ, δ) ExQU (resp.
ExQD) then
1. There is a polynomial time two-sided adaptive algorithm for ǫ-testing C that makes
O(cid:18)M (ǫ/12, 1/24) + Q(ǫ/12, δ/3) +
k
ǫ(cid:19)
queries.
25
2. There is a polynomial time two-sided distribution-free adaptive algorithm for ǫ-testing C that
makes
queries.
O(cid:18)M (ǫ/12, 1/24) + kQ(ǫ/12, 1/24) +
k
ǫ(cid:19)
Finally, one trivial but useful result is
Theorem 30. If there is a polynomial time algorithm A that, given as an input any constant δ and
any ǫ makes M (ǫ, δ) MQs and Q(ǫ, δ) ExQU (resp. ExDD) and distinguish between F ∈ C(Y ) and
F ǫ-far from every function in C(Y ) with respect to the uniform distribution (resp. with respect to
the distribution D) then
1. There is a polynomial time two-sided adaptive algorithm for ǫ-testing C that makes
O(cid:18)M (ǫ/12, 1/24) + Q(ǫ/12, 1/24) +
k
ǫ(cid:19)
queries.
2. There is a polynomial time two-sided distribution-free adaptive algorithm for ǫ-testing C that
makes
queries.
4 Results
O(cid:18)M (ǫ/12, 1/24) + kQ(ǫ/12, 1/24) +
k
ǫ(cid:19)
In this section we define the classes and give the results for the classes k-Junta, k-Linear, k-Term,
s-Term Monotone r-DNF, size-s Decision Tree, size-s Branching Program, Functions with Fourier
Degree at most d, Length-k Decision List and s-Sparse Polynomial of Degree d.
We will use words that are capitalized for classes and non-capitalized words for functions. For
example, k-Junta is the class of all k-juntas.
4.1 Testing k-Junta
For k-Junta in uniform distribution framework, Ficher et al. [33] introduced the junta testing
problem and gave a non-adaptive algorithm that makes O(k2)/ǫ queries. Blais in [7] gave a non-
adaptive algorithm that makes O(k3/2)/ǫ queries and in [8] an adaptive algorithm that makes
O(k log k + k/ǫ) queries. On the lower bounds side, Fisher et al. [33] gave an Ω(√k) lower bound for
non-adaptive testing. Chockler and Gutfreund [28] gave an Ω(k) lower bound for adaptive testing
and, recently, Saglam in [55] improved this lower bound to Ω(k log k). For the non-adaptive testing
Chen et al. [24] gave the lower bound Ω(k3/2)/ǫ.
For testing k-junta in the distribution-free model, Chen et al. [45] gave a one-sided adaptive
algorithm that makes O(k2)/ǫ queries and proved a lower bound Ω(2k/3) for any non-adaptive
algorithm. The result of Halevy and Kushilevitz in [41] gives a one-sided non-adaptive algorithm
that makes O(2k/ǫ) queries. The adaptive Ω(k log k) uniform-distribution lower bound from [55]
26
trivially extends to the distribution-free model. Bshouty [15] gave a two-sided adaptive algorithm
that makes O(1/ǫ)k log k queries.
Our algorithm in this paper gives
Theorem 31. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing k-Junta that makes O(k/ǫ) queries.
Proof. We use Theorem 30. Since every F (Y ) is in k-Junta(Y ), the algorithm A always accepts.
Therefore, we have M = Q = 0, and the algorithm makes O(k/ǫ) queries.
4.2 Testing k-Linear
Blum et al.
The function is linear if it is a sum (over the binary field F2) of variables. The class Linear is the
class of all linear functions. The class k-Linear is Linear∩k-Junta. That is, the class of functions
that are the sum of at most k variables.
[13] showed that there is an algorithm for testing Linear under the uniform dis-
tribution that makes O(1/ǫ) queries. For testing k-Linear under the uniform distribution, Fisher,
et al. [33] gave a tester that makes O(k2/ǫ) queries. They also gave the lower bound Ω(√k) for
non-adaptive algorithms. Goldreich [35], proved the lower bound Ω(k) for non-adaptive algorithms
and Ω(√k) for adaptive algorithms. Then Blais et al. [9] proved the lower bound Ω(k) for adaptive
algorithms. Blais and Kane, in [10], gave the lower bound k − o(k) for adaptive algorithms and
2k − o(k) for non-adaptive algorithms.
Testing k-Linear can be done by first testing if the function is k-Junta and then testing if
it is Linear. Therefore, there is an adaptive algorithm for ǫ-testing k-Linear under the uniform
distribution that makes O(k/ǫ) queries.
In this paper we prove
Theorem 32. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing k-Linear that makes O(k/ǫ) queries.
Proof. We use Theorem 29. Here C(Y ) = {y1 + ··· + yq} contains one function and therefore the
learning algorithm just outputs y1 + ··· + yq. Therefore M = Q = 0 and the result follows.
4.3 Testing k-Term
A term (or monomial) is a conjunction of literals and Term is the class of all terms. A k-term is a
term with at most k literals and k-Term is the class of all k-terms.
In the uniform distribution model, Pernas et al. [50], gave a tester for k-terms that makes
O(1/ǫ) queries in the uniform model. We give the same result in the next section. In this paper we
prove the following result for the distribution-free model. When k = n, better results can be found
in [34, 32].
Theorem 33. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing k-Term that makes O(k/ǫ) queries.
i = xi and x1
1 ∧ ··· ∧ yξq
q ξ ∈ {0, 1}q} contains 2q
Proof. Recall that x0
functions. We use Theorem 29 with Remark 22. Since V contains witnesses for each variable it
follows that ξi are known. Just take any string a that satisfies F (a) = 1 and then ξi = ai. Therefore
M = Q = 0 and the result follows.
i = xi. Here C(Y ) = {yξ1
27
4.4 Testing s-Term Monotone r-DNF
A DNF is a disjunction of terms. An r-DNF is a disjunction of r-terms. The class s-Term r-DNF
is the class of all r-DNFs with at most s terms. The class s-Term Monotone r-DNF is the class of
all r-DNFs with at most s terms with no negated variables. A DNF f is called unate DNF if there
is ξ ∈ {0, 1}n such that f (xξ1
n ) is monotone DNF. If ξi = 0 then we say that f is positive
unate in xi; otherwise we say that f is negative unate in xi. Similarly, one can define the classes
Unate DNF, Unate s-DNF etc.
1 , . . . , xξn
We first give a learning algorithm for s-Term Monotone r-DNF. The algorithm is in Figure 9.
In the algorithm, we use P1/r for the probability distribution over the strings b ∈ {0, 1}n where
each coordinate bi is chosen randomly and independently to be 1 with probability 1 − 1/r and
0 with probability 1/r. For two strings x, y ∈ {0, 1}n we denote x ∗ y = (x1y1, . . . , xnyn) where
xiyi = xi ∧ yi. The procedure FindMinterm(f, a) flips bits that are one in a to zero as long as
f (a) = 1.
LearnMonotone(f,D, ǫ, δ, s, r)
Input: Oracle that accesses a Boolean function f
that is s-term monotone r-DNF and D.
Output: h that is s-term monotone r-DNF
Choose a ∈ D.
If f (a) = 1 and h(a) = 0 then
1. h ← 0.
2. Repeat 4(s/ǫ) log(1/δ) times.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12. Output h
h ← h ∨Qai=1 xi
t ← 0
While t ≤ α := 4r ln(2ns/δ) and wt(a) > r do
t ← t + 1; If t = α + 1 Output "fail"
Choose y ∈ P1/r
If f (a ∗ y) = 1 then a ← a ∗ y
a ←FindMinterm(f, a)
Figure 9: A learning algorithm for s-Term Monotone r-DNF
We now show
Lemma 34. If the target function f is s-term monotone r-DNF then for any constant δ, algorithm
LearnMonotone asks O(s/ǫ) ExQD and O(sr log(ns)) MQ and, with probability at least 1 − δ,
learns an s-term monotone r-DNF, h, that satisfies PrD[h 6= f ] ≤ ǫ.
Proof. We first show that if in the mth iteration of the algorithm (steps 3-11) the function h
contains ℓ terms of f and f (a) = 1 and h(a) = 0 then, with probability at least 1 − δ/(2s), steps
5 to 11 adds to h a new term of f . This implies that in the (m + 1)th iteration h contains ℓ + 1
terms of f . Then, since the number of terms of f is at most s, with probability at least 1− δ/2, all
28
the terms in h are terms in f . We then show that, with probability at least 1 − δ/2, the procedure
outputs h that satisfies PrD[h 6= f ] ≤ ǫ.
First notice that if f (a) = 1 and h(a) = 0 then for every y ∈ {0, 1}n, h(a ∗ y) = 0. This follows
from the fact that h is monotone and a ∗ y ≤ a. Therefore if a receives the values a(1), . . . , a(τ ) in
the While loop then f (a(i)) = 1 and h(a(i)) = 0 for all i = 1, . . . , τ . We also have a(i+1) = a(i) if
f (a(i)∗y) = 0 and a(i+1) = a(i)∗y if f (a(i)∗y) = 1. Consider the random variable Wi = wt(a(i))− r.
We will now compute E[Wi+1Wi]. Since f (a(i)) = 1 and h(a(i)) = 0, there is a term T in f that is
not in h that satisfies T (a(i)) = 1. Suppose T = xj1xj2 ··· xjr′ , r′ ≤ r. Then a(i)
jr′ = 1.
= ··· = a(i)
Consider another r − r′ entries in a(i) that are equal to 1, a(i)
= 1. Such entries
exist because of the condition wt(a) > r of the While command. Note that wt(a(i)) = Wi + r.
Let jr+1, . . . , jr+Wi be the other entries of a(i) that are equal to 1. Let A be the event that, for
the y ∈ P1/r chosen at this stage, yj1 = ··· = yjr = 1. Notice that if event A happens then
T (a(i+1)) = f (a(i+1)) = 1 and a(i+1) = a(i) ∗ y. Then
≤ E[Wi+1Wi, A](cid:18)1 −
= Wi(cid:18)1 −
= Wi(cid:18)1 −
E[Wi+1Wi] = E[Wi+1Wi, A]Pr[A] + E[Wi+1Wi, ¯A]Pr[ ¯A]
+ Wi(cid:18)1 −(cid:18)1 −
1
r(cid:19)r(cid:19)
4r(cid:19) .
+ Wi(cid:18)1 −(cid:18)1 −
r(cid:19)r(cid:19) ≤ Wi(cid:18)1 −
r(cid:19)r+1
r(cid:18)1 −
= ··· = a(i)
r(cid:19)r(cid:19)
1
r(cid:19)r
j1
jr
(1)
(2)
1
1
jr′+1
1
1
1
The inequality in (1) follows from the fact that Wi+1 ≤ Wi and (2) follows from the fact that the
expected number of ones in yjr+1ajr+1, . . . , yjr+Wi
Therefore E[Wi] ≤ n(1 − 1/(4r))i. The probability that the algorithm fails is the probability
is (1 − 1/r)Wi.
that t = 4r ln(2ns/δ). By Markov's Bound, Lemma 67, this is bounded by
ajr+Wi
Pr[wt(a(t)) > r] = Pr[Wt > 1] ≤ E[Wt] ≤ n(cid:18)1 −
1
4r(cid:19)t
δ
2s
.
≤
This completes the first part of the proof.
Now we show that, with probability at least 1 − δ/2, the procedure outputs h that satisfies
PrD[f 6= h] ≤ ǫ. Let h(i) be the function h at iteration i = 1, 2, . . . , w. Since h(1) =⇒ h(2) =⇒
··· =⇒ h(w) = h =⇒ f , if PrD[f 6= h] > ǫ then PrD[f 6= h(i)] > ǫ for all i. Therefore, the
probability that PrD[f 6= h] > ǫ is less than the probability that for v = 4s log(1/δ)/ǫ strings
a(1), . . . , a(v) chosen independently at random according to the distribution D, less than s of them
satisfies gi(a(i)) 6= f (a(i)) for Boolean functions gi that satisfy PrD[gi 6= f ] ≥ ǫ. By Chernoff's
bound, Lemma 69, this probability is less than δ/2.
The algorithm asks at most 4s log(1/δ)/ǫ = O(s/ǫ) ExQD and at most s · 4r ln(2ns/δ) =
O(sr log(ns)) MQ.
Now we show
Theorem 35. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing s-Term Monotone r-DNF that makes O(rs2/ǫ) queries.
29
Proof. The number of relevant variables in any s-term monotone r-DNF is at most q ≤ k =
sr. By Lemma 34, C(Y ) can be learned with constant confidence δ and accuracy ǫ in M =
O(sr log(qs)) = O(sr) MQ and O(s/ǫ) ExQD. By Theorem 29, there is a distribution-free tester
for s-Term Monotone r-DNF that makes O(s2r/ǫ) queries.
Theorem 36. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing s-Term Unate r-DNF that makes O(rs2/ǫ) queries.
Proof. The set of witnesses tells us, for every variable xi, if f is positive unate in xi or negative
) = 0 then f is positive unate in xτ (ℓ) and if f (v(ℓ)) = 0,
unate.
) = 1 then f is negative unate in xτ (ℓ). Then the result immediately follows from
If f (v(ℓ)) = 1, f (0Xℓ ◦ v(ℓ)
Xℓ
f (0Xℓ ◦ v(ℓ)
Xℓ
Theorem 35.
4.5 Testing Size-s Decision Tree and Size s Branching Program
A decision tree is a rooted binary tree in which each internal node is labeled with a variable xi
and has two children. Each leaf is labeled with an output from {0, 1}. A decision tree computes a
Boolean function in an obvious way: given an input x ∈ {0, 1}n, the value of the function on x is
the output in the leaf reached by starting at the root and going left or right at each internal node
according to whether the variable's value in x is 0 or 1, respectively. The size of a decision tree is
the number of leaves of the tree. The class size-s Decision Tree is the class of all decision trees of
size s.
A branching program is a rooted directed acyclic graph with two sink nodes labeled 0 and 1.
As in the decision tree, each internal node is labeled with a variable xi and has two children. The
two edges to the children are labeled with 0 and 1. Given an input x, the value of the branching
program on x is the label of the sink node that is reached as described above. The size of a
branching program is the number of nodes in the graph. The class size-s Branching Program is the
class of all Branching Program of size s.
Diakonikolas et al. [30], gave a tester for size-s Decision Tree and size s Branching Program
under the uniform distribution that makes O(s4/ǫ2) queries. Chakraborty et al. [21] improved the
query complexity to O(s/ǫ2). In this paper we prove
Theorem 37. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing size-s Decision
Tree and size-s Branching Program that makes O(s/ǫ) queries.
There is a two-sided distribution-free adaptive algorithm for ǫ-testing size-s Decision Tree and
size s Branching Program that makes O(s2/ǫ) queries.
Proof. For decision tree, C(Y ) contains the decision trees with q = Y ≤ k = s relevant variables.
It is shown in [30] that C(Y ) ≤ (8s)s. For branching programs C(Y ) ≤ (s + 1)3s. Now by
Theorem 26 the result follows.
4.6 Functions with Fourier Degree at most d
For convenience here we take the Boolean functions to be f : {−1, 1}n → {−1, 1}. Then every
Boolean function has a unique Fourier representation f (x) =PS⊆[n]
fSQi∈S xi where fS are the
Fourier coefficients of f . The Fourier degree of f is the largest d = S with fS 6= 0.
30
Let C be the class of all Boolean functions over {−1, 1}n with Fourier degree at most d. Nisan
and Szegedy, [49], proved that any Boolean function with Fourier degree d must have at most
k := d2d relevant variables. Diakinikolas et al. [30], show that every nonzero Fourier coefficient
of f ∈ C is an integer multiple of 1/2d−1. Since PS⊆[n]
f 2
S = 1, there are at most 22d−2 nonzero
Fourier coefficients in f ∈ C.
Diakonikolas et al. [30], gave an exponential time tester for Boolean functions with Fourier
degree at most d under the uniform distribution that makes O(26d/ǫ2) queries. Chakraborty et
al. [21] improved the query complexity to O(22d/ǫ2). In this paper we prove
Theorem 38. For any ǫ > 0, there is a poly(2d, n) time two-sided distribution-free adaptive al-
gorithm for ǫ-testing for the class of Boolean functions with Fourier degree at most d that makes
O(22d + 2d/ǫ) queries.
Proof. Bshouty gives in [16] an exact learning algorithm for such class2 that asks M = O(22d log n)
membership queries for any constant confidence parameter δ. Now since q = Y ≤ k = d2d, by
Theorem 28 the result follows.
4.7 Testing Length k Decision List
A decision list is a sequence f = (xi1, ξ1, a1), . . . , (xis, ξs, as) for any s where ξi, ai ∈ {0, 1}. This
sequence represents the following function: f (x) := If xi1 = ξ1 then output(a1) else if xi2 = ξ2 then
output(a2) else if ··· else if xis = ξs then output(as). Length-k decision list is a decision list with
s ≤ k. The class Decision List is the class of all decision lists and the class Length-k Decision List
is the class of all length-k decision lists.
It is known that this class is learnable under any distribution with O((k log n + log(1/δ))/ǫ)
ExQD, [14, 51]. This implies
Theorem 39. For any ǫ > 0, there is a polynomial time two-sided distribution-free adaptive algo-
rithm for ǫ-testing Length-k Decision List that makes O(k2/ǫ) queries.
Proof. The result follows from Theorem 29.
4.8 Testing s-Sparse Polynomial of Degree d
A polynomial (over the field F2) is a sum (in the binary field F2) of monotone terms. An s-sparse
polynomial is a sum of at most s monotone terms. We say that the polynomial f is of degree d
if its terms are monotone d-terms. The class s-Sparse Polynomial of Degree d is the class of all
s-sparse polynomials of degree d. The class Polynomial of Degree d is the class of all polynomials
of degree d.
In the uniform distribution model, Diakonikolas et al. [30], gave the first testing algorithm for the
class s-Sparse Polynomial that runs in exponential time and makes O(s4/ǫ2) queries. Chakraborty
et al. [21] improved the query complexity to O(s/ǫ2). Diakonikolas et al. gave in [31] the first
polynomial time testing algorithm that makes poly(s, 1/ǫ) queries. In [1], Alon et al. gave a testing
algorithm for Polynomial of Degree d that makes O(1/ǫ + d22d) queries. They also show the lower
bound Ω(1/ǫ + 2d) for the number of queries. Combining those results we get a polynomial time
2The class in [16] is the class of decision trees of depth d but the analysis is the same for the class of functions
with Fourier degree at most d
31
testing algorithm for s-Sparse Polynomial of Degree d that makes poly(s, 1/ǫ) + O(22d) queries.
Just run the Alon et al. algorithm in [1] and then run Diakonikolas et al. algorithm in [31] and
accept if both algorithms accept.
Here we prove the following Theorem.
Theorem 40. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing s-Sparse Poly-
nomial of Degree d that makes O(s/ǫ + s2d) queries.
For any ǫ > 0, there is a two-sided distribution-free adaptive algorithm for ǫ-testing s-Sparse
Polynomial of Degree d that makes O(s2/ǫ + s2d) queries.
We first give a learning algorithm LearnPolynomial for s-sparse polynomial of degree d. See
Figure 10.
LearnPolynomial(f,D, ǫ, δ, s, d)
Input: Oracle that accesses an s-sparse polynomial f
of degree d and D.
Output: An s-sparse polynomial of degree d, h, or "fail"
1. h ← 0, t(h) ← 0.
2. Repeat (s/ǫ) ln(3s/δ) times.
Choose a ∈ D.
3.
t(h) ← t(h) + 1.
4.
5.
If (f + h)(a) = 1 then
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
If wt(a) > d then "fail"
M ← Find a monotone term in (f + h)(a ∗ x)
h ← h + M
t(h) ← 0.
If t(h) = (1/ǫ) ln(3s/δ) then Output h
m ← 0;
While m ≤ α := 16 · 2d(2 ln(s/δ) + ln n) and wt(a) > d do
m ← m + 1;
Choose y ∈ U
If (f + h)(a ∗ y) = 1 then a ← a ∗ y
Figure 10: A learning algorithm for s-sparse polynomial of degree d
Lemma 41. Let f be an s-sparse polynomial of degree d. For any constant δ, algorithm Learn-
Polynomial asks O((s/ǫ) log s) ExQD and O(s2d log(ns)) MQ and, with probability at least 1 − δ,
learns an s-sparse polynomial of degree d, h, that satisfies PrD[h 6= f ] ≤ ǫ.
Proof. Suppose f =PM ∈F M , where F are the set of monotone d-terms of f and F = s′ ≤ s.
Suppose, at some stage of the algorithm h =PM ∈F ′ M where F ′ ⊂ F . Then f + h =PM ∈F \F ′ M .
Notice that F ′ is the set of terms of f that is found by the algorithm up to this stage and F\F ′ is
the set of terms that is yet to be found. Since the number of terms of f is at most s, all we need
to show is that:
32
1. Each time the algorithm executes steps 6-14, with probability at least 1 − δ/(2s), it finds a
term of f + h, and therefore, a new term of f .
2. Assuming 1., the algorithm, with probability at least 1− δ/2, outputs an s-sparse polynomial
of degree d, h that satisfies PrD[f 6= h] ≤ ǫ.
Then, by the union bound, the success probability of the algorithm is at least 1 − δ and the result
follows.
We first prove 1. Let g = f + h. Suppose that the algorithm finds a string a such that g(a) = 1.
Then a satisfies at least one term in g. Let M ′ ∈ F\F ′ be one of them and let d′ ≤ d be the degree
of M ′ . Then g(a ∗ x) =PM ∈F \F ′,M (a)=1 M contains M ′ and therefore g(a ∗ x) is not zero.
We first show that the probability that after α1 := 16 · 2d ln(sn/δ) iterations of steps 7-10, the
weight of a does not drop below 24d is less than δ/(4s). Then we show that, if the weight of a is less
than or equal to 24d then the probability that after α2 := α − α1 more iterations of steps 7-10 the
weight of a does not drop below d + 1 is less than δ/(4s). If these two facts are true then after the
algorithm finishes executing the While command, with probability at least 1 − δ/(2s), the weight
of a is less than d + 1.
It is known that for any non-zero polynomial H of degree at most d, PrU [H(x) = 1] ≥ 1/2d, [18].
Since g(a ∗ x) is of degree at most d, for a random uniform string y we get Pr[g(a ∗ y) = 1] ≥ 1/2d.
Now suppose wt(a) ≥ 24d. By Chernoff's bound, Lemma 69, the probability that wt(a ∗ y) >
(3/4)wt(a) is at most e−wt(a)/24 ≤ 2−d−1. Therefore, by the union bound,
1
Pr[g(a ∗ y) = 1 and wt(a ∗ y) ≤ (3/4)wt(a)] ≥ 1 −(cid:18)1 −
2d+1(cid:19) ≥
1
2d+1 .
1
2d +
(3)
The probability that after α1 = 16 · 2d ln(sn/δ) iterations of steps 7-10, the weight of a does not
drop below 24d is less than the probability that for α1 = 16· 2d ln(sn/δ) random uniform strings y,
less than log(n)/ log(4/3) of them satisfies g(a ∗ y) = 1 and wt(a ∗ y) ≤ (3/4)wt(a) given that a
satisfies wt(a) ≥ 24d and g(a ∗ x) 6= 0. By (3) and Chernoff's bound this probability is less than
δ/(4s).
We now show that if wt(a) < 24d, then after α2 = α − α1 = 16 · 2d ln(s/δ) iterations of steps 7-
10, with probability at least 1 − δ/(4s), the weight of a drops below d + 1. Take a that satisfies
d + 1 ≤ wt(a) < 24d. Then
Pr[g(a ∗ y) = 1 and wt(a ∗ y) < wt(a)] ≥ Pr[g(a ∗ y) = 1] − Pr[wt(a ∗ y) = wt(a)]
1
2d −
1
2d+1 =
1
2d+1 .
≥
Then as before, with an additional α2 iterations of steps 7-10, with probability at least 1 − δ/(4s),
the weight of a drops below d + 1.
Once the weight of a is less or equal to d, the algorithm finds in step 12 a monotone term in
g(a ∗ x) by building a truth table of g(a ∗ x) using at most 2d queries and learning one of its terms.
This term is in g because all the terms of g(a ∗ x) are terms of g(x).
The proof that the algorithm, with probability at least 1 − δ/2, outputs h such that PrD[f 6=
h] ≤ ǫ is identical to the proof in Lemma 9 for the output of ApproxTarget.
We are now ready to prove Theorem 40.
Proof. By Lemma 41, Theorem 29 and since n = Y = sd, Q = (s/ǫ) log s and M = s2d log(sd)
the result follows.
33
5 Testing Classes that are Close to k-Junta
In this section, we show the result for s-term DNF in the uniform distribution model. Then in the
following section, we show how to extend it to other classes.
The main idea is the following. We first run the procedure ApproxC in Figure 11 that finds
X ⊂ [n] and w ∈ {0, 1}n such that, with high probability,
1. The projection xX ◦ wX removes variables from f that appear only in terms of f of size at
least c log(s/ǫ) for some large constant c.
2. h = f (xX ◦ wX) is (ǫ/8)-close to f .
From (1) we conclude that the terms of size at most c log(s/ǫ) in f contain all the variables of h.
Since the number of terms in f is at most s the number of variables that remain in h is at most
k := cs log(s/ǫ). From (2) we conclude that if f is ǫ-far from every s-term DNF then h is (7ǫ/8)-
far from every s-term DNF and therefore h is (7ǫ/8)-far from every s-term DNF with at most
k variables. Therefore, it is enough to distinguish whether h is an s-term DNF with at most k
variables or (7ǫ/8)-far from every s-term DNF with at most k variables. This can be done by the
algorithm TesterC in the previous section
Note that removing variables that only appears in large size terms does not necessarily remove
large terms in f . Therefore, h may still contain large terms even after running ApproxTarget in
TesterC. To handle large terms, we can use any learning algorithm that learns h with accuracy
ǫ/12 and use Theorem 29.
This gives a tester for s-term DNF that makes O(s2/ǫ) queries, which is not optimal. This
is because the number of s-term DNF with at most k variables is m := 2O(ks) (and therefore the
number of queries in TesterC is at least O((log m)/ǫ) = O(s2/ǫ)). To get an optimal query tester,
we do the following. We build a tester that uses only random uniform queries for the class s-term
DNF with at most k variables and terms of size at most r = c′ log(s/ǫ) where c′ is a large constant
and show that this tester, with high probability, works well for h. The reason for that is that when
the algorithm uses random uniform queries, with high probability, all the terms of h that are of size
greater than r are zero for every query. Since the number of s-term DNF with at most k variables
and terms of size at most r is at most m = 2O(rs log k) the number of queries in TesterC is at most
O(k/ǫ + (log m)/ǫ) = O(s/ǫ).
In the next subsection, we give the procedure ApproxC that removes variables that only appear
in large size terms, and in Subsection 5.2, we give the tester for s-Term DNF. Then in Section 6,
we extend the above to other classes.
5.1 Removing Variables that only Appears in Large Size Terms
We explain our technique by proving the result for s-term DNF.
We remind the reader that for a term T , the size of T is the number of variables that are in it.
For a variable x and ξ ∈ {0, 1}, xξ = x if ξ = 0 and xξ = x if ξ = 1. For a term T = xc1
i1 ∧ ··· ∧ xcv
we denote by Va(T ) = {xi1, . . . , xiv}, the set of variables that appears in T . For a set of terms T
we denote Va(T ) = ∪T ∈T Va(T ). Here λ > 1 is any constant and we use c = O(log λ) to denote a
large constant.
iv
Consider the procedure ApproxC in Figure 11. We will prove the following two Lemmas
34
Algorithm ApproxC(f, ǫ, λ)
Input: Oracle that accesses a Boolean function f and
Output: Either "X ⊆ [n], w ∈ {0, 1}n" or "reject"
Partition [n] into r sets
1.
2. Choose uniformly at random a partition X1, X2, . . . , Xr of [n]
Set m = c log(s/ǫ); r = 8ms.
Choose u, v ∈ U .
t(X) ← t(X) + 1
If f (uX ◦ vX ) 6= f (u) then
Find a close function and relevant sets
Set X = ∅; I = ∅; t(X) = 0; k = 3ms
3.
4. Repeat M = 100λk ln(100k)/ǫ times
5.
6.
7.
8.
9.
10.
11.
12.
13.
Binary Search to find a new relevant set Xℓ; X ← X ∪ Xℓ; I ← I ∪ {ℓ}.
If I > k then output "reject" and halt.
t(X) = 0.
If t(X) = 100λ ln(100k)/ǫ then
Choose a random uniform w;
Output(X, w, f (xX ◦ wX )).
Figure 11: A procedure that removes variables from f that only appears in large size terms.
Lemma 42. Let f be an s-term DNF. ApproxC makes O(s/ǫ) queries and, with probability at
least 9/10, outputs X and w such that
1. f (xX ◦ wX ) is s-term DNF.
2. The number of relevant variables in f (xX ◦ wX) is at most 3cs log(s/ǫ) = O(s log(s/ǫ)).
Lemma 43. Let f be ǫ-far from every s-term DNF. ApproxC makes O(s/ǫ) queries and either
rejects, or with probability at least 9/10, outputs X and w such that f (xX ◦ wX) is (1 − 1/λ)ǫ-far
from every s-term DNF.
We first prove Lemma 42
Proof. Consider an s-term DNF, f = T1 ∨ T2 ∨ ··· ∨ Ts′, where s′ ≤ s. Let T = {T1, . . . , Ts′}. Let
T1 = {T ∈ T : Va(T ) ≤ m} be the set of terms in T of size at most m := c log(s/ǫ) and let R1 =
Va(T1) be the set of variables that appear in the terms in T1. Let T2 = {T ∈ T :
Va(T )\R1 ≤ m}
be the set of terms T ∈ T that contain at most m variables not in R1. Let R2 = R1 ∪ Va(T2) be the
variables in R1 and of the terms in T2. Let T3 = {T ∈ T : Va(T )\R1 > m} be the set terms T ∈ T
that contain more than m variables that are not in R1. Let T4 = {T ∈ T :
Va(T )\R2 ≥ m} be
the set of terms in T ∈ T that contain at least m variables not in R2. Let R3 = R2 ∪ Va(T3\T4) be
the set of variables in R2 and of the terms in T3\T4. Then
R1 ≤ ms,R2 ≤ 2ms,R3 ≤ 3ms and T4 ⊆ T3.
35
In steps 1-2, procedure ApproxC uniformly at random partitions [n] into r = 8ms sets
X1, . . . , Xr. Suppose that the variables in R2 are distributed to the sets Xj1, . . . , Xjq , q ≤ R2 ≤
2ms.
For each T ∈ T4, the expected number of variables in Va(T )\R2 that are not distributed to
one of the sets Xj1, . . . , Xjq is greater or equal to (1 − q/r) m = (3/4)m. By Hoeffding's bound,
Lemma 70, and the union bound, the probability that it is greater than m/2 in every term T ∈ T4
is at least 1 − s · exp(−m/8) ≥ 99/100.
In steps 5-7 procedure ApproxC are repeated M times and it finds relevant sets using two
random uniform strings u and v. If f (uX ◦ vX) 6= f (u) then a new relevant set is found. Consider
any T ∈ T3. The size of T is at least m. Suppose T = xξ1
am′ , m′ ≥ m. For random
uniform u, v ∈ {0, 1}n, the probability that there is no j ∈ [m′] such that uaj = (uX ◦ vX)aj = ξj
is at most (3/4)m. The probability that this happens for at least one T ∈ T3 and at least one of
the M randomly uniformly chosen u and v in the procedure is at most (3/4)msM ≤ 1/100. Notice
that if uaj = (uX ◦ vX )aj = ξj then T (u) = T (uX ◦ vX) = 0 and T (w) = 0 for every string w
in the binary search that is made to find a new relevant set. Therefore, with probability at least
99/100, the procedure runs as if f contains no terms in T3. Let f ′ = ∨T ∈T \T3T . With probability
at least 99/100 the procedure runs as if f = f ′. The number of relevant variables in f ′ is at most
R2 ≤ 2ms and all those variables are distributed to Xj1, . . . , Xjq . Therefore, with probability at
least 99/100, the procedure generates at most 2ms < k relevant sets and therefore, by step 9, it
does not reject and those relevant sets are from Xj1, . . . , Xjq .
a1 ∧ ··· ∧ xξm′
The output of the procedure is X ⊆ Xj1 ∪ ··· ∪ Xjq and a random uniform w. We now show
that with high probability f (xX ◦ wX) contains at most k = 3ms relevant variables.
We have shown above that with probability at least 99/100 every term T ∈ T4 contains at least
m/2 variables that are not distributed to Xj1, . . . , Xjq . Therefore, for a fixed term T ∈ T4 and for
a random uniform w, the probability that T (xX ◦ wX) = 0 is at least 1− (1/2)m/2. The probability
that T (xX ◦ wX) = 0 for every T ∈ T4 is at least 1 − s(1/2)m/2 ≥ 99/100. Therefore, when we
randomly uniformly choose w ∈ {0, 1}n, with probability at least 99/100, the function f (xX ◦ wX )
does not contain terms from T4. Thus, with probability at least 99/100, f (xX ◦ wX) contains at
most R3 ≤ 3ms variables.
Now as in the proof of Lemma 10, the query complexity is 2M + k log r = O(s/ǫ).
We now prove Lemma 43.
Proof. As in the proof of Lemma 9, the probability that the algorithm fails to output X that
satisfies Prx,y∈U [f (xX ◦ yX) = f (x)] ≤ ǫ/(100λ) is at most
k(cid:16)1 −
=
1
100
.
ǫ
100λ(cid:17)(100λ ln(100/k))/ǫ
If Prx,y∈U [f (xX ◦ yX) = f (x)] ≤ ǫ/(100λ) then, by Markov's inequality, Lemma 67, for a random
uniform w, with probability at least 99/100, Prx∈U [f (xX ◦ wX) = f (x)] ≤ ǫ/λ.
Now as in the proof of Lemma 10, the query complexity is 2M + k log r = O(s/ǫ).
In the next subsection, we give the result for s-term DNF, and then we show how to use the
above technique to other classes.
36
5.2 Testing s-term DNF
We have shown in Lemma 42 and 43 that, using O(s/ǫ) queries, the problem of testing s-Term DNF
can be reduced to the problem of testing s-Term DNF with k = O(s log(s/ǫ)) relevant variables.
We then can use TesterC for the latter problem. This gives a tester for s-term DNF that makes
O(s2/ǫ). This is because the number of s-term DNF with k = O(s log(s/ǫ)) relevant variables is
2 O(s2). We will now show how to slightly change the tester TesterC and get one that makes O(s/ǫ)
queries.
In TesterC the procedures ApproxTarget, TestSet and Closef F make O(k log k), O(k) and
O((k/ǫ) log(k/ǫ)) queries, respectively. This is O(s/ǫ) queries. Therefore, the only procedure that
makes O(s2/ǫ) queries in TesterC is CloseF CU . So we will change this procedure.
Let h be an s-term DNF. Notice that in step 3 in CloseF CU , for a random uniform z =
(z1, . . . , zq) ∈ {0, 1}q , the probability that z satisfies a term T in h of size at least c log(s/ǫ) (i.e.,
T (z) = 1), is at most (ǫ/s)c. Therefore, if in CloseF CU we define C ∗ to be the class of all s-
term DNF with terms of size at most c log(s/ǫ), then the probability that for at least one of the
τ = (3/ǫ) log(C ∗/δ) random uniform z = (z1, . . . , zq) and at least one of the terms T in h of size
at least c log(s/ǫ), T satisfies z, is at most sτ (ǫ/s)c ≤ 1/100. Therefore, with probability at least
99/100 the algorithm runs as if h is s-term DNF with terms of size at most c log(s/ǫ) and then
accept. Since log C ∗ = O(s) we get
Theorem 44. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing s-Term DNF
that makes O(s/ǫ) queries.
For completeness we wrote the tester. See TesterApproxC in Figure 12.
In the tester
C({y1, . . . , yq}, c log(s/ǫ)) is the class of all s-term DNFs with terms of size at most c log(s/ǫ)
over q variables.
6 Results
In this section, we extend the technique used in the previous section to other classes.
6.1 Testing s-Term Monotone DNF
We first use the algorithm LearnMonotone in Figure 9 to show
Lemma 45. Let f : {0, 1}n → {0, 1} be an s-term Monotone DNF. For constant δ, algorithm
LearnMonotone(f, U, ǫ/2, δ/2, s, 2(log(s/ǫ) + log(1/δ))) asks O(s/ǫ) ExQU and O(s(log n + log s)·
log(s/ǫ)) MQ and, with probability at least 1 − δ, learns an s-term monotone DNF h that satisfies
PrU [h 6= f ] ≤ ǫ.
Proof. Let f be the function f without the terms of size greater than 2(log(s/ǫ) + log(1/δ)). Then
PrU [f 6= f ] ≤ s2−2(log(s/ǫ)+log(1/δ))) ≤
ǫ
2
.
In the algorithm LearnMonotone the probability that one of the assignments in step 3 (that
is, a where a ∈ U ) satisfies one of the terms in f of size greater than 2(log(s/ǫ) + log(1/δ)) is less
than
(4s/ǫ)(log(1/δ))s2−2(log(s/ǫ)+log(1/δ)) ≤
37
δ
2
.
TesterApproxC(f,D, ǫ)
Input: Oracle that access a Boolean function f and D.
Output: Either "reject" or "accept"
(X, w) ←ApproxC(f, ǫ, 1/6).
(X, V, I) ←ApproxTarget(h, U, ǫ, 1/6).
1.
2. h := f (xX ◦ wX).
3.
4. TestSets(X, V, I).
5. Closef F (f, U, ǫ, 1/15)
6. C ∗ ← C({y1, . . . , yq}, c log(s/ǫ)) where q = V
Test closeness to C ∗
7. Repeat τ = (3/ǫ) log(30C ∗) times
8.
9.
10.
11.
12. Return "accept"
Choose (z1, . . . , zq) ∈ U .
For every g ∈ C ∗
If g(z) 6= F (z) then C ∗ ← C ∗\{g}.
If C ∗ = ∅ then "reject"
Figure 12: A procedure for testing s-Term DNF
Also for a monotone term T , if T (a) = 0 then for any y, T (a ∗ y) = 0. Therefore, with probability
at least 1− δ/2, the algorithm runs as if f is f (which is s-term monotone (2(log(s/ǫ) + log(1/δ)))-
DNF). By Lemma 34, if the target is f then, with probability at least 1 − δ/2, LearnMonotone
outputs h that is (ǫ/2)-close to f . Since f is (ǫ/2)-close to f and h is (ǫ/2)-close to f we have that
h is ǫ-close to f . This happens with probability at least 1 − δ.
The number of queries follows from Lemma 34.
We now prove
Theorem 46. For any ǫ > 0, there is a polynomial time two-sided adaptive algorithm for ǫ-testing
s-Term Monotone DNF that makes O(s/ǫ) queries.
Proof. We first run ApproxC and get an s-term monotone DNF h with O(s log(s/ǫ)) variables
that is (ǫ/6)-close to f . We then use Theorem 29 with Lemma 45.
We also have
Theorem 47. For any ǫ > 0, there is a polynomial time two-sided adaptive algorithm for ǫ-testing
s-Term Unate DNF that makes O(s/ǫ) queries.
Proof. The proof is similar to the proof of Theorem 36.
6.2 Testing Size-s Boolean Formula and Size-s Boolean Circuit
A Boolean formula is a rooted tree in which each internal node has arbitrarily many children and
is labeled with AND or OR. Each leaf is labeled with a Boolean variable xi or its negation ¯xi. The
38
size of a Boolean formula is the number of AND/OR gates it contains. The class size-s Boolean
Formula is the class of all Boolean formulas of size at most s.
A Boolean circuit is a rooted directed acyclic graph with internal nodes labeled with an AND,
OR or NOT gate. Each AND/OR gate is allowed to have arbitrarily many descendants. Each
directed path from the root ends in one of the nodes x1, x2, . . . , xn, 0, 1.
The same analysis that we did for s-term DNF also applies to size-s Boolean formulas and size-s
Boolean circuit. Analogous to the size of terms, we take the number of distinct literals a gate has.
If the gate is labeled with AND (respectively, OR) and the number of distinct literals it has is more
than c log(s/ǫ), then we replace the gate with a node labeled with 0 (respectively, 1) and remove
all the edges to its children.
Therefore we have
Lemma 48. Lemma 42 and 43 are also true for size-s Boolean formulas and size-s Boolean circuit.
We now prove
Theorem 49. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing size-s Boolean
Formula that makes O(s/ǫ) queries.
Proof. Similar to testing s-term DNF, we can ignore gates that have more than c log(s/ǫ) distinct
literals. Just replace it with 0 if its label is AND and with 1 if it is OR.
The number of Boolean formulas of size s that have at most c log(s/ǫ) distinct literal in each
gate and at most k = O(s log(s/ǫ)) variables is 2 O(s), [30]. The rest of the proof goes along with
the proof of testing s-term DNF.
The number of Boolean circuits of size s that have at most c log(s/ǫ) distinct literals in each
gate and at most k = O(s log(s/ǫ)) variables is 2 O(s2), [30]. Then similar to the above proof one
can show
Theorem 50. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing size-s Boolean
Circuit that makes O(s2/ǫ) queries.
6.3 Testing s-Sparse Polynomial
In the literature, the first testing algorithm for the class s-Sparse Polynomial runs in exponential
time [30] and makes O(s4/ǫ2) queries. Chakraborty et al., [21], then gave another exponential time
algorithm that makes O(s/ǫ2) queries. Diakonikolas et al. gave in [31] the first polynomial time
testing algorithm that makes poly(s, 1/ǫ) queries. Here we prove
Theorem 51. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing s-Sparse Poly-
nomial that makes O(s2/ǫ) queries.
We have shown in Lemma 42 and 43 that the problem of testing s-Term DNF can be reduced to
the problem of testing s-Term DNF with k = O(s log(s/ǫ)) relevant variables. The same reduction
and analysis show that the problem of testing s-sparse polynomials can be reduced to the problem
of testing s-sparse polynomials with k = O(s log(s/ǫ)) relevant variables. The reduction makes
O(s/ǫ) queries. Thus
Lemma 52. Lemma 42 and 43 are also true for s-Sparse Polynomials.
39
LearnPolyUnif(f, ǫ, δ, s)
Input: Oracle that accesses a Boolean function f that is s-sparse polynomial.
Output: h that is s-sparse polynomial
1. h ← 0, t(h) ← 0, w ← 0.
2. Repeat (s/ǫ) ln(3s/δ) log(3s/δ) times.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
h ← h +Qai=1 xi
If t(h) = (1/ǫ) ln(3s/δ) then Output h
Choose a ∈ U .
t(h) ← t(h) + 1.
If (f + h)(a) = 1 then
m ← 0, w ← w + 1
If w = log(3s/δ) then Output h
While m ≤ α := 16 · (8s/ǫ)(2 ln(s/δ) + ln n) and wt(a) > log(s/ǫ) + 3 do
m ← m + 1;
Choose y ∈ U
If (f + h)(a ∗ y) = 1 then a ← a ∗ y
If wt(a) > log(s/ǫ) + 3 then Goto 16
w ← 0.
a ← Find a monotone term in (f + h)(a ∗ x)
t(h) ← 0.
Figure 13: A learning algorithm for s-sparse polynomial under the uniform distribution
We can then use TesterC for the latter problem. Therefore, all we need to do in this section
is to find a learning algorithm for the class of s-sparse polynomials with k = O(s log(s/ǫ)) relevant
variables.
Consider the algorithm LearnPolyUnif in Figure 13. We prove
Lemma 53. Let f be a s-sparse polynomial with n variables. For constant δ, algorithm Learn-
PolyUnif asks O(s/ǫ) ExQU and O((s2/ǫ) log n) MQ and, with probability at least 1− δ, learns an
s-sparse polynomial h that satisfies PrU [h 6= f ] ≤ ǫ.
Proof. Let f = M1 + M2 + ··· + Ms′, s′ ≤ s, where deg(M1) ≤ ··· ≤ deg(Ms′′) ≤ log(s/ǫ) + 3 <
deg(Ms′′+1) ≤ ··· ≤ deg(Ms′ ). Let f1 = M1 + ··· + Ms′′ and f2 = Ms′′+1 + ··· + Ms′. Then
f = f1 + f2 and PrU [f2 = 1] ≤ s2− log(s/ǫ)−3 = ǫ/8. If f (a) 6= f1(a) then f2(a) = 1 and therefore
PrU [f 6= f1] ≤ PrU [f2 = 1] ≤ ǫ/8 and PrU [f = f1] ≥ 1 − ǫ/8. In fact, if A(x) is the event that
Ti(x) = 0 for all i > s′′ then PrU [A] ≥ 1 − ǫ/8.
The algorithm LearnPolyUnif is similar to the algorithm LearnPolynomial with d = log(s/ǫ)+
3 with the changes that is described below.
First notice that in the algorithm the hypothesis h contains only terms from f1, that is, terms
of size at most log(s/ǫ) + 3. This is because in step 12 the algorithm skips the command that adds
a term to h when wt(a) ≥ log(s/ǫ) + 3. Suppose at some stage of the algorithm, h =Pi∈B⊆[s′′] Mi
contains some terms of f1 and let g = f + h = Pi∈[s′]\B Mi. Suppose PrU [f 6= h] ≥ ǫ. Then
PrU [g = 1] ≥ ǫ. We want to compute the probability that the algorithm finds a term in f1 + h =
40
Pi∈[s′′]\B Mi and not in f2. That is, the probability that it finds a term of f of degree at most
log(s/ǫ) + 3. We first have
Pra∈U [(f1 + h)(a) = 1 ∧ A(f + h)(a) = 1] = Pra∈U [g(a) = 1 ∧ Ag(a) = 1]
=
≥
Pra∈U [g(a) = 1 ∧ A]
Pra∈U [g(a) = 1]
Pra∈U [g(a) = 1] − Pra∈U [A]
Pra∈U [g(a) = 1]
≥ 1 −
ǫ/8
ǫ ≥
7
8
.
That is, if PrU [f 6= h] ≥ ǫ and (f + h)(a) = 1 then with probability at least 7/8, (f1 + h)(a) = 1
and for every term T in f2, T (a) = 0. If the event A(a) happens then for any y and for every term
T in f2, T (a ∗ y) = 0, and therefore, for such a, the algorithm runs as if the target is f1.
have three cases:
When the algorithm reaches step 5 and finds a string a ∈ {0, 1}n such that (f + h)(a) = 1, we
1. The event B ≡ [ ((f1 + h)(a) = 1 and A(a)] happens.
2. PrU [f 6= h] ≥ ǫ and B.
3. PrU [f 6= h] < ǫ and B.
Case 1. Notice that steps 8-11 are identical to steps 7-10 in LearnPolynomial in Figure 10 with
d = log(s/ǫ) + 3. Therefore, as in the proof of Lemma 41, with probability at least 1 − δ/3 every
assignment a that satisfies B gives a term in f1 that is not in h.
Case 2. This case can happen with probability at most 1/8. So the probability that it happens
log(3s/δ) consecutive times is at most δ/(3s). The probability that it does happen for some of the
at most s different hypothesis h generated in the algorithm is at most δ/3. Notice that w counts
the number of consecutive times that this case happens and step 7 outputs h when it does happen
log(3s/δ) consecutive times. Therefore, with probability at most δ/3 the algorithm halts in step 7
and output h that satisfies Pr[f 6= h] ≥ ǫ. This is also the reason that the algorithm repeats the
search for a in step 2, (s/ǫ) ln(3s/δ) log(3s/δ) times which is log(3s/δ) times more than in algorithm
LearnPolynomial.
When this case happens, the algorithm either ends up with a string a of weight that is greater
than ℓ := log(s/ǫ) + 3 and then it ignores this string, or, it ends up with a string of weight less
than or equal to ℓ and then, Step 14 finds a new term in f1. This is because that a string of weight
less than or equal to ℓ cannot satisfy a term of degree more than ℓ.
Case 3. This case cannot happen more than log(3s/δ) consecutive times because if it does step 7
outputs h which is a good hypothesis.
7 A General Method for Other Testers
In this section, we generalize the method we have used in the previous section and then prove some
more results
We define the distribution D[p] to be over Qn
i=1{0, 1, xi} where each coordinate i is chosen to
be xi with probability p, 0 with probability (1 − p)/2 and 1 with probability (1 − p)/2. We will
denote by f the size of f in C which is the length of the representation of the function f in C.
41
Algorithm ApproxGeneralC(f, ǫ, δ)
Input: Oracle that accesses a Boolean function f
Output: Either "X ⊆ [n], w ∈ {0, 1}n" or "reject"
Partition [n] into r sets
1.
2. Choose uniformly at random a partition X1, X2, . . . , Xr of [n]
Set r = kc+1.
Set X = ∅; I = ∅; t(X) = 0
Choose u, v ∈ U .
t(X) ← t(X) + 1
If f (uX ◦ vX) 6= f (u) then
Find a close function and relevant sets
3.
4. Repeat M = (4c1k/(δǫ)) ln(4k/δ) times
5.
6.
7.
8.
9.
10.
11.
12.
13.
Find a new relevant set Xℓ; X ← X ∪ Xℓ; I ← I ∪ {ℓ}.
If I > k then Output "reject"
t(X) = 0.
If t(X) = (4c1/(δǫ)) ln(4k/δ) then
Choose a random uniform w;
Output(X, w).
Figure 14: An algorithm that removes variables from f that have a small influence.
Consider the algorithm ApproxGeneralC in Figure 14. We start with the following result
Lemma 54. Let δ < 1/2, c1 ≥ 1, λ > 1 and c ≥ 1 be any constants, k := k(ǫ, δ,f) be an
integer and M = (4c1k/(δǫ)) ln(4k/δ). Let C be a class of functions where for every f ∈ C there is
h ∈ (C ∩ k-Junta) with relevant variables x(Y ) = {xix ∈ Y }, Y ⊆ [n], and h′ ∈ (C ∩ (λk)-Junta)
that satisfy the following:
1. Prz∈D[1/2][f (z) 6= h(z)] ≤ δ/(4M ).
2. Pry∈D[1/kc][f (xY ◦ yY ) 6= h′(y)] ≤ δ/4.
The algorithm ApproxGeneralC makes O(k/ǫ) queries and,
1. If f ∈ C then, with probability at least 1− δ, the algorithm does not reject and outputs X and
w such that f (xX ◦ wX) ∈ C has at most λk relevant variables.
2. For any f , if the algorithm does not reject then, with probability at least 1 − δ, Pr[f (x) 6=
f (xX ◦ wX )] ≤ ǫ/c1.
Proof. Let f ∈ C. Let h ∈ C ∩ k−Junta and h′ ∈ (C ∩ (λk)-Junta) be functions that satisfies 1
and 2. The algorithm in step 5 chooses two random uniform strings u and v. Define z = (z1, . . . , zn)
such that zi = 0 if ui = vi = 0, zi = 1 if ui = vi = 1, and zi = xi if ui 6= vi. Since u and v are chosen
uniformly at random we have that z ∈ D[1/2] and therefore, with probability at least 1 − δ/(4M ),
42
f (z) = h(z). If f (z) = h(z) then f (u) = h(u) and f (uX ◦ vX) = h(uX ◦ vX ) for any X ⊆ [n].
In Step 8 ApproxGeneral does a binary search to find a new relevant set. In the binary search
it queries strings a that satisfy ai = zi = ui = vi for all i that satisfies zi ∈ {0, 1}. Therefore,
f (a) = h(a) for all the strings a generated in the binary search for finding a relevant set. Therefore,
with probability at least 1− δ/(4M ), f (a) = h(a) for all the queries used in one iteration and, with
probability at least 1 − δ/4, f (a) = h(a) for all the queries used in the algorithm. That is, with
probability at least 1 − δ/4, the algorithm runs as if the target is h.
Therefore, if f ∈ C, then with probability at least 1−δ/4, each one of the relevant sets discovered
in the algorithm contains at least one relevant variable of h. Then since h ∈ k-Junta, the algorithm
does not reject in Step 9, that is, I ≤ k.
Now we show that, if f ∈ C then with probability at least 1− δ/4, f (xX ◦ wX ) contains at most
λk relevant variables. Consider the partition in steps 1-2 in the algorithm and let Xi1, . . . , Xik′ ,
k′ ≤ k, be the sets where the indices of the relevant variables x(Y ) of h are distributed. Let
X ′ = Xi1 ∪ ··· ∪ Xik′ . Notice that for a random uniform w ∈ {0, 1}n, (xX ′ ◦ wX ′) = (xY ◦ yY )
where y ∈ D[k′/kc+1]. That is, given that the relevant variables of h are distributed to k′ different
sets that their union is X ′, the probability distributions of (xX ′ ◦ wX ′) and of (xY ◦ yY ) are
identical. Choosing a string in D[k′/kc+1] can be done by first choosing a string b in D[1/kc]
and then substitute in each variable xi, i 6∈ Y , in b, 0, 1 or xi with probability 1/2 − k′/(2k),
1/2− k′/(2k) or k′/k, respectively. Therefore, since Pry∈D[1/kc][f (xY ◦ yY ) 6= h′(y)] ≤ δ/4, we have
Pry∈D[k′/kc+1][f (xY ◦ yY ) 6= h′(y)] ≤ δ/4 and, thus, with probability at least 1 − δ/4, we have that
f (xX ′ ◦ wX ′) = h′(x). In particular, with probability at least 1 − δ/4, f (xX ′ ◦ wX ′) has at most λk
relevant variables. Since X ⊆ X ′ we also have with the same probability f (xX ◦ wX ) has at most
λk relevant variables. This completes the proof of 1.
Now let f be any boolean function. If the algorithm does not reject then I ≤ k. Since for
the final X, f (uX ◦ vX) 6= f (u) for (4c1/(δǫ)) ln(4k/δ) random uniform u and v, we have that the
probability that the algorithm fails to output X that satisfies Prx,y∈U [f (xX◦yX ) 6= f (x)] ≤ δǫ/(4c1)
is at most
δǫ
4c1(cid:19)(4c1/(δǫ)) ln(4k/δ)
k(cid:18)1 −
=
δ
4
.
If Prx,y∈U [f (xX ◦ yX) 6= f (x)] ≤ δǫ/(4c1) then, by Markov's inequality, for a random uniform w,
with probability at least 1 − δ/4, Prx∈U [f (xX ◦ wX ) 6= f (x)] ≤ ǫ/c1. This completes the proof
of 2.
In the next subsections, we give some results that follow from Lemma 54.
7.1 Testing Decision List
In [30], Diakonikolas et al. gave a polynomial time tester for Decision List that makes O(1/ǫ2)
queries. In this paper, we give a polynomial time tester that makes O(1/ǫ) queries.
We show
Theorem 55. For any ǫ > 0, there is a two-sided adaptive algorithm for ǫ-testing Decision List
that makes O(1/ǫ) queries.
Proof. Let f = (xi1, ξ1, a1), . . . , (xis, ξs, as) be any decision list. We first use Lemma 54.
Define k = min(s, c′ log(1/(ǫδ))) for some large constant c′ and h = (xi1, ξ1, a1), . . . , (xik , ξk, ak).
For the distribution D[1/2], the probability that f (z) = h(z) is 1 when k = s and at least 1 −
43
(3/4)k ≥ 1 − 1/(3M ) where M = (4c1k/(δǫ)) ln(4k/δ). For the distribution D[1/k2] and Y =
{i1, . . . , ik}, the probability that f (xY ◦yY ) has more than 2k relevant variables is less than (3/4)k ≤
δ/3.
The query complexity of ApproxGeneral is O(k/ǫ) = O(1/ǫ). Therefore all we need to do to
get the result is to give a tester for decision list of size k = O(log(1/ǫ)) that makes O(1/ǫ) queries.
The learnability of this class with O(1/ǫ) ExQs follows from [51, 14].
7.2 Testing r-DNF and r-Decision List for Constant r
An r-decision list is a sequence f = (T1, ξ1, a1), . . . , (Ts, ξs, as) for any s where ξi, ai ∈ {0, 1} and Ti
are r-terms. This sequence represents the following function: f (x) :="If T1 = ξ1 then output(a1)
else if T2 = ξ2 then output(a2) else if ··· else if Ts = ξs then output(as)". The class r-Decision
List is the class of all r-decision lists and the class of Length-s r-Decision List is the class of all
r-decision lists f = (T1, ξ1, a1), . . . , (Tm, ξm, am) with m ≤ s.
In this subsection, we show
Theorem 56. Let r be any constant. For any ǫ > 0, there is a two-sided adaptive algorithm for
ǫ-testing r-Decision List and r-DNF that makes O(1/ǫ) queries.
It is known that the class Length-s r-Decision List is learnable under any distribution in time
O(nr) using O((sr log n + log(1/δ))/ǫ) ExQD, [14, 51]. Also we may assume that T1, . . . , Ts are
distinct and therefore s is less thanPr
constant r, it is enough to prove that r-Decision List and r-DNF satisfies 1 and 2 in Lemma 54
with k = poly(log(1/ǫ)).
i(cid:1)2i , the number of terms of size at most r. Thus, for
i=1(cid:0)n
We now prove the result for r-Decision List when r is constant. The same analysis shows that
the result is also true for r-DNF.
Consider an r-decision list f = (T1, ξ1, a1)··· (Ts, ξs, as). If ξi = 0 then we change (Ti, 0, ai) to
. Therefore we may
, 1, ai) where Ti = xc1
i1 ··· xcℓ
the equivalent expression (x
assume that ξj = 1 for all j. In that case we just write f = (T1, a1)··· (Ts, as).
(Tiℓ , aiℓ) such that 1 ≤ i1 < i2 < ··· < iℓ ≤ s.
For an r-decision list f = (T1, a1)··· (Ts, as), a sublist of f is an r-decision list g = (Ti1, ai1)···
We first prove
, 1, ai)··· (x
1−ci1
i1
1−ciℓ
iℓ
iℓ
Lemma 57. Let r be a constant. For any r-decision list f there is a kr := O(logr(1/ǫ))-length
r-decision list h that is a sublist of f and satisfies 1 and 2 in Lemma 54.
Proof. We show that it satisfies 1 in Lemma 54. The proof that it also satisfies 2 is similar.
We give a stronger result as long as r is constant. We prove by induction that for any r-
decision list f there is a kr = O(logr(1/ǫ))-length r-decision list h that is a sublist of f and satisfies
Prz∈D[1/2][f (z) 6= h(z)] ≤ poly(ǫ).
The proof is by induction on r. For r = 1, the result follows from the proof of Theorem 55 in
the previous subsection. Assume the result is true for r-decision list. We now show the result for
(r + 1)-decision list.
Let c be a large constant. Let f = (T1, a1)··· (Ts, as) be (r + 1)-decision list. Let s1 = 1 and
Ts1, . . . , Tsw be a sequence of terms such that s1 < s2 < ··· < sw ≤ s and for every i, si is the
minimal integer that is greater than si−1 such that the variables in Tsi do not appear in any one of
44
the terms Ts1, Ts2, . . . , Tsi−1. Define h0 = (T1, a1)(T2, a2)··· (Tsw′ , asw′ ) if w′ := c2r+1 ln(1/ǫ) ≤ w
and h0 = f if w′ > w. Then
Prz∈D[1/2][f (z) 6= h0(z)] ≤ Pr[T1(z) = 0 ∧ T2(z) = 0 ∧ ··· ∧ Tsw′ (z) = 0]
≤ Pr[Ts1(z) = 0 ∧ Ts2(z) = 0 ∧ ··· ∧ Tsw′ (z) = 0]
≤ (cid:18)1 −
1
2r+1(cid:19)w′
≤ poly(ǫ).
Let S = {xj1, . . . , xjt} be the set of the variables in Ts1, . . . , Tsw′ . Then t ≤ (r + 1)w′ = c(r +
1)2r+1 ln(1/ǫ) and every term Ti in h0 contains at least one variable in S. Consider all the terms that
contains the variable xj1, Ti1 = xj1T ′
, i1 < i2 < ··· < iℓ. Consider
the r-decision list g := (T ′
, aℓ). By the induction hypothesis there is an r-
decision list g′ that is a sublist of g of length at most kr = O(logr(1/ǫ)) such that Prz∈D[1/2][g(z) 6=
g′(z)] ≤ poly(ǫ). Let h1 be h0 without all the terms (Tiw , aiw ) that correspond to the terms (T ′
iw , aiw )
that do not occur in g′. It is easy to see that Prz∈D[1/2][h0(z) 6= h1(z)] ≤ poly(ǫ). We do the same
for all the other variables xj2, . . . , xjt of S and get a sequence of r-decision lists h2, h3, . . . , ht that
satisfies Prz∈D[1/2][hw(z) 6= hw+1(z)] ≤ poly(ǫ). Therefore Prz∈D[1/2][f (z) 6= ht(z)] ≤ t · poly(ǫ) =
poly(ǫ) and the length of ht is at most tkr = O(logr+1(1/ǫ)) = kr+1.
i1, Ti2 = xj1T ′
i2, a2)··· (T ′
i2, . . . , Tiℓ = xj1T ′
iℓ
i1, a1)(T ′
iℓ
8
Improvements and Further Results
In this section, we go over the previous results and show how some modifications can give improved
results. The improvements are in the logarithmic part of the query complexity. We concentrate
on the first tester that test subclasses of k-Junta. Similar improvements can also be made for the
second and third testers.
8.1
Improved ApproxTarget Procedure
The procedure ApproxTarget makes O((k log k)/ǫ) queries. We show that the number of queries
can be reduced to O(k/ǫ + k log k). We call the new procedure IApproxTarget
The procedure IApproxTarget repeats steps 5-18 in ApproxTarget M = 16ck/ǫ times and
does not have step 19, and therefore does not need the variable t(X). It also saves all the values
of X. Suppose it does not reject and X takes the values X (1), X (2), . . . , X (k′) where k′ ≤ k and let
X (i)) for i ∈ [k′]. In the case where the procedure does not reject we add another
h(i) = f (xX (i) ◦ 0
procedure that estimates Prx∈D[h(i)(x) 6= f (x)] for all i ∈ [k′], with constant multiplicative error
and outputs h(i) with the least estimated value. This can be done with N = (96c/ǫ) ln(60k) queries
to ExQD.
We first prove that, with probability at least 29/30, there is i ∈ [k′] such that Prx∈D[h(i)(x) 6=
f (x)] ≤ ǫ/(2c). For suppose there is no such i: Then all the functions h(i) are (ǫ/(2c))-far from
f (x) and for M = 16ck/ǫ random u ∈ D only k′ ≤ k of them witness that h(i)(x) 6= f (x) for some i.
By Chernoff's bound this can happen with probability at most e−(7/8)2·8k/2 ≤ e−3k ≤ 1/30.
Now to distinguish between those that are (ǫ/(2c))-close to f and those that are (ǫ/c)-far from
f , by Chernoff's bound, we need to estimate all Prx∈D[h(i)(x) 6= f (x)], i ∈ [k′], with confidence
probability 29/30 and accuracy ǫ/(4c). For this we need N ≤ (96c/ǫ) ln(60k) queries to ExQD.
45
We've proved
Lemma 58. The IApproxTarget makes O(k/ǫ + k log k) queries.
Obviously, if the h(ℓ) gets the least estimated value then the procedure removes from V (and
I) all the witnesses of X (i), i > ℓ and returns (X (ℓ), V, I).
8.2 A General TestSets Procedure
In this subsection, we give a small change in the procedure TestSets and show in the sequel
how this change improves the query complexity. The general TestSets GTestSet has another
), 1, β, 1/15) in step 2. The
parameter β < 1/30 as an input and calls UniformJunta(f (xXℓ ◦ v(ℓ)
other commands are the same. The value of β will depend on the class being learned.
Xℓ
The same proofs as for TestSet give
Lemma 59. Let β < 1/30. The procedure GTestSet satisfies
1. If for some ℓ ∈ I, f (xXℓ ◦ v(ℓ)
Xℓ
bution then, with probability at least 14/15, GTestSet rejects.
) is β-far from every literal with respect to the uniform distri-
2. GTestSet makes O(k/β) queries.
8.3 A General RelVarValue Procedure
In this, subsection we give a small change in the procedure RelVarValue and show in the sequel
how this change improves the query complexity. The general RelVarValue GRelVarValue(w, X, V, I, δ, β),
that has another parameter β < 1/30 as an input, repeats steps 5-7
times. The other commands are the same. The value of β will depend on the class being learned.
h = max(1, (log(k/δ))/ log(1/(3pβ)))
We give the proof of the following
Lemma 60. If for every ℓ ∈ I the function f (xXℓ ◦ v(ℓ)
) is β-close to a literal in {xτ (ℓ), ¯xτ (ℓ)} with
respect to the uniform distribution, where τ (ℓ) ∈ Xℓ, and {Gℓ,0, Gℓ,1} = {0, h} then, with probability
at least 1 − δ, we have: For every ℓ ∈ I, zℓ = wτ (ℓ).
Proof. We follow the proof of Lemma 15 with the same notations and the following changes in the
proof. We have
Xℓ
Therefore
and by Markov's bound
ExXℓ
∈U [Z(xXℓ)] ≤ β.
ExYℓ,1 ∈U ExYℓ,0 ∈U [Z(xYℓ,0 ◦ xYℓ,1)] ≤ β
PrxYℓ,1 ∈UhExYℓ,0 ∈U [Z(xYℓ,0 ◦ xYℓ,1)] ≥pβi ≤pβ.
That is, for a random uniform string b ∈ {0, 1}n, with probability at least 1−√β, f (xYℓ,0◦bYℓ,1◦v(ℓ)
is √β-close to xτ (ℓ) with respect to the uniform distribution. Now, given that f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
Xℓ
Xℓ
)
)
46
is √β-close to xτ (ℓ) with respect to the uniform distribution the probability that Gℓ,0 = 0 is the
probability that f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) for h random uniform strings b ∈ {0, 1}n.
Let b(1), . . . , b(h) be h random uniform strings in {0, 1}n, V (b) be the event f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
) =
) is √β-close to xτ (ℓ) with respect to
) and A the event that f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
Xℓ
the uniform distribution. Let g(xYℓ,0) = f (xYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
). Then
) = f (bYℓ,0 ◦ bYℓ,1 ◦ v(ℓ)
Xℓ
Xℓ
Xℓ
Xℓ
Xℓ
Pr[V (b)A] ≤ Pr[g(bYℓ,0) 6= bτ (ℓ)A] + Pr[g(bYℓ,0 ) 6= bτ (ℓ))A] ≤ 2pβ.
Since τ (ℓ) ∈ Yℓ,0, we have wτ (ℓ) = 0. Therefore, by step 9 and since τ (ℓ) ∈ Xℓ,
Pr[zℓ 6= wτ (ℓ)] ≤(cid:0)Pr[V (b)A] + Pr[A](cid:1)h
≤ (3pβ)h
Therefore, the probability that zℓ 6= wτ (ℓ) for some ℓ ∈ I is at most k(3√β)h ≤ δ.
The following is obvious
Lemma 61. The procedure GRelVarValues makes O(k + k(log(k/δ))/ log(1/β))) queries.
8.4
Improved Closef F Procedure
The procedure Closef F makes O((k/ǫ) log(k/ǫ)) queries. We show that the number of queries can
be reduced to O((k/ǫ) + (k/ǫ) log(k/ǫ)/ log(1/β)) queries in the distribution-free model and
O(k log(1/ǫ) + k log(1/ǫ) log(k log(1/ǫ))/ log(1/β) + 1/ǫ)
queries in the uniform distribution model. We call the new procedures IClosef F and IUClosef F .
Notice that if you choose β = ǫ, the query complexity of TestSets is O(k/ǫ) and the query
complexity of Closef F is O((k/ǫ)(1 + (log k)/ log(1/ǫ))) queries in the distribution-free model and
O(k log k + k log log(1/ǫ) + 1/ǫ) queries in the uniform distribution model. The former complexity
is O(k/ǫ) when ǫ = 1/kΩ(1).
For IClosef F we just take the procedure Closef F and instead of calling RelVarValue it
calls GRelVarValues with the parameter β.
Lemma 62. For any constant δ, procedure IClosef F makes O((k/ǫ) + (k/ǫ)(log(k/ǫ))/ log(1/β))
queries.
When the distribution is uniform, rather than choosing random uniform and independent strings
u we choose random uniform pairwise independent strings. We choose t = O(log(1/ǫ)) random uni-
form and independent strings W = {w(1), . . . , w(t)} and generate random uniform pairwise inde-
pendent strings from the linear space that are spanned by W . It is enough to run GRelVarValues
for the strings in W to find w(1)
i=1 ηiw(i) in the linear
space we have uΓ =Pt
i=1 ηiw(i)
Γ because for any other string u =Pt
Γ , . . . , w(t)
Γ . The algorithm is in Figure 15.
Lemma 63. For any constant δ and (X, V, I) that satisfies Assumption 17, procedure IUClosef F
makes O(k log(1/ǫ) + (k log(1/ǫ) log(k log(1/ǫ))/ log(1/β) + 1/ǫ) queries and
1. If f ∈ C then IUClosef F outputs OK.
47
IUClosef F (f, ǫ, δ, β)
Input: Oracle that accesses a Boolean function f .
Output: Either "reject" or "OK"
1. Define F ≡ f (x(πf,I) ◦ 0X).
2. Choose t = log(12/(ǫδ)) i.i.d. strings w(1), . . . , w(t) ∈ U
3. For i = 1 to t.
4.
5. Choose arbitrary m = 6/(ǫδ) non-zero strings ξ(1), . . . , ξ(m) ∈ {0, 1}t
6. For i = 1 to m.
7.
8.
9. Return "OK"
ζ (i) ←GRelVarValue(w(i), X, V, I, δ/(2t), β) .
Set u ←Pt
j=1 ξ(i)
If f (uX ◦ 0X ) 6= F (z) then "reject"
j w(j); z ←Pt
j=1 ξ(i)
j ζ (j);
Figure 15: A procedure that tests whether f (xX ◦ 0X) is (ǫ/3)-far from F with respect to the
uniform distribution.
2. If f (xX ◦ 0X) is (ǫ/3)-far from F then, with probability at least 1 − δ, IUClosef F rejects.
j=1 ξ(i)
j=1 ξ(i)
Proof. The proof of 1 is the same as in the proof of Lemma 20. We now prove 2.
j w(j) and z(i) = Pt
Let u(i) = Pt
j ζ (j) for i ∈ [m]. By Lemma 60 and 61, with
probability at least 1 − δ/2, ζ (i) = w(i)
Γ for all i ∈ [t] . The map x → xΓ is a linear map and
therefore, with probability at least 1 − δ/2, u(i)
Γ = z(i) for all i ∈ [m]. Define, for i ∈ [m], the event
X ◦ 0X ) 6= F (z(i)) and 0 otherwise. Let A = Pi∈[m] Ai. Then ǫ′ := E[Ai] ≥ ǫ/3,
Ai = 1 if f (u(i)
µ := E[A] = Pi∈[m] ǫ′ = ǫ′m ≥ ǫm/3 = 2/δ and Var[Ai] = ǫ′ − ǫ′2. Since Ai are pairwise
independent, Var[A] =Pi∈[m] Var[Ai] = µ(1 − ǫ′). By (7) in Lemma 68 we have
Pr[Algorithm outputs "OK"] = Pr[A = 0] ≤ Pr[A − µ ≥ µ] ≤
Var[A]
µ2 ≤
1
µ ≤
δ
2
.
Therefore, if f (xX ◦ 0X ) is (ǫ/3)-far from F then, with probability at least 1 − δ, IUClosef F
rejects.
The query complexity follows from Lemma 61.
8.5 A Weaker ExQD
In this subsection, we define a weak ExQD, WExQD, and show that learning with WExQD and
MQ gives a tester for F (y) using a better query complexity. We then show that in many of the
learning algorithms in this paper, ExQD can be replaced with WExQD.
We define the WExQD that returns with probability 1/2 a string drawn according to distribution
D and with probability 1/2 an arbitrary string (chosen by an adversary). We now show that in
some of the learning algorithms in this paper one can replace ExQD with WExQD
We first show
48
Lemma 64. For constant δ, there is a polynomial time learning algorithm for s-Term Monotone
r-DNF that makes O(s/ǫ) WExQD and O(sr log(ns)) MQ and, with probability at least 1−δ, learns
an s-term monotone r-DNF h that satisfies PrD[h 6= f ] ≤ ǫ.
Proof. Consider the algorithm LearnMonotone in Figure 9 with WExQD.
If steps 2-11 are
repeated 16(s/ǫ) log(1/δ) times then with probability at least 1 − δ, more than 4(s/ǫ) log(1/δ) of
the examples received by the WExQD are according to the distribution D. The rest of the proof
follows directly from the proof of Lemma 34.
Consider now algorithm LearnPolynomial in Figure 10 that learns s-sparse polynomial over
F2 of Degree d. We can use the technique in Subsection 8.1, to get an algorithm that learns this
class with O(s/ǫ) ExQD and O(2d log(ns)) MQ. Then as in the proof of Lemma 45 the ExQD can
be replaced with WExQD. Therefore
Lemma 65. For constant δ, there is a polynomial time learning algorithm for s-Sparse Polynomial
of Degree d that makes O(s/ǫ) WExQD and O(2ds log(ns)) MQ and, with probability at least 1− δ,
learns an s-sparse polynomial of degree d, h, that satisfies PrD[h 6= f ] ≤ ǫ.
8.6 An Improved Query Complexity
We first show
If there is a polynomial time algorithm A that, given as an
Theorem 66. Let C ⊆ k-Junta.
input a constant δ, any ǫ and (X, V, I) that satisfies Assumption 17, learns C(Y ) with respect the
distribution D (resp. uniform distribution), with confident δ, accuracy ǫ, makes M (ǫ, δ) M Q to F
and Q(ǫ, δ) ExQD (resp. WExDD, ExDU ) queries to F then there is a polynomial time tester TD
(resp. TW D, TU ) for C that makes
O k
ǫ
+ k log k + M ′ + kQ′ + k ·
1
ǫ log k + Q′ log(kQ′)
ǫ + log log k + log Q′!
log 1
queries (resp.
O k
ǫ
+ k log k + M ′ + kQ′ + k log k ·
ǫ + log log k + log Q′! ,
1
ǫ + Q′
log 1
queries) where Q′ = Q(ǫ/12, 1/24), M ′ = M (ǫ/12, 1/24) and
O(cid:18)k
ǫ
+ k log k + M ′ + Q′(cid:19)
1. If f ∈ C then, with probability at least 2/3, TD, TW D and TU accepts.
2. If f is ǫ-far from every function in C with respect to D then, with probability at least 2/3, TD
(resp. TW D) rejects.
3. If f is ǫ-far from every function in C with respect to the uniform distribution then, with
probability at least 2/3, TU rejects.
49
Proof. We run TesterC(f,D, ǫ) in Figure 8 with the following changes:
In step 1 we run the
procedure IApproxTarget(f,D, ǫ, 1/3). By Lemma 58, it uses O(k/ǫ + k log k) queries. In step 2
we run the procedure GTestSet(X, V, I, β). By Lemma 59, it uses k/β queries. The value of
In step 4, we will run the procedure IClosef F (f,D, ǫ, 1/15, β) for
β will be determined later.
any distribution D and IUClosef F (f, ǫ, 1/15, β) for the uniform distribution. By Lemma 62,
IClosef F makes O(k/ǫ + (k/ǫ)(log(k/ǫ))/ log(1/β)) queries. By Lemma 63, IUClosef F makes
O(k log(1/ǫ) + k log(1/ǫ) log(k log(1/ǫ))/ log(1/β) + 1/ǫ)
queries.
By Lemma 24 and 61, to distinguish between F ∈ C(Y ) and F ǫ-far from every function in
C(Y ) we need to ask M ′ MQ and Q′′ = Q′ + O(1/ǫ) ExQD queries to f . For that we need to run
GRelVarValue(w, X, V, I, 1/(24Q′′ ), β), Q′′ times. This takes O(kQ′′ + kQ′′(log(kQ′′))/ log(1/β))
queries. To make Q′′ WExQD, by Lemma 61, we need to run GRelVarValue(w, X, V, I, 1/2, β),
Q′′ times. This takes O(kQ′ + kQ′(log k)/ log(1/β)) queries. For the uniform distribution we just
need Q′ queries. For MQ we need M ′ queries.
Now for the distribution-free model we choose (resp. with WExQD)
β =
(resp.
ǫ log k
1
ǫ log k
log2(cid:0) 1
log2(cid:0) 1
1
ǫ + Q′′ log(kQ′′)
ǫ + Q′′ log(kQ′′)(cid:1)
ǫ + Q′ log k(cid:1)
ǫ + Q′ log k
ǫ log k
ǫ log k
β =
) and get the result. For the uniform distribution model we choose β = ǫ and get the result.
We note here that when Q′ = 0 then we define log 0 = 0.
8.7 Some Improved Results
In this subsection, we give some of the results. See the Table in Figure 16.
For k-Junta we have M ′ = Q′ = 0 and then by Theorem 66, we get the query complexity
O(k/ǫ + k log k) for the uniform model and
O k
ǫ
+ k log k +
k
ǫ log k
ǫ + log log k!
log 1
for the distribution-free model. Notice that, the latter query complexity is O(k/ǫ + k log k) when
1/ǫ = O(log log k) or 1/ǫ = kΩ(1).
The same result follows for the classes k-Linear and k-Term.
For s-Term Monotone r-DNF and s-Unate r-DNF, by Lemma 45, there is an algorithm that
learns F ∈ C(Y ) with k = sr variables with M ′ = O(sr log(sr)) MQ and Q′ = O(s/ǫ) WExQD.
This gives the result in the Table.
For Length-k Decision List, M ′ = 0 MQ and Q′ = O((k/ǫ) log k) WExQD. By Theorem 66,
we get the query complexity O(k(log k)/ǫ) for the uniform model and O(k2(log k)/ǫ) for the
distribution-free model.
It is an open problem whether the class of Length-k Decision List is
50
Class of Functions
s-Term Monotone r-DNF
s-Term Unate r-DNF
k-Junta, k-Term
and k-Linear
Length k Decision List
size-s Decision Trees and
size-s Branching Programs
F. With Fourier Degree ≤ d
Distri-
bution
U
D
U
D
U
D
U
D
D
#Queries
Complexity O(·)
sr
ǫ + sr log(sr)
sr log r + s2r
ǫ + s2r
k
ǫ + k log k
ǫ
log r
log s+log log r+log 1
ǫ
k
ǫ + k log k +
k
ǫ log k
ǫ +log log k
log 1
k log k
ǫ
k2 log k
ǫ
s log s
ǫ
s2 log s
ǫ
d2d
ǫ + d522d
Figure 16: A table of the results. In the table, U and D stand for uniform and distribution-free
model.
learnable with O(k/ǫ) ExQD. If it is, then the query complexity can be reduced to O(k/ǫ + k log k)
for the uniform model O(k2/ǫ) for the distribution-free model.
For Functions with Fourier Degree ≤ d, it is known from [16], Lemma 21, that this class is
learnable with O(d322d log2 n) queries. Therefore, k = d2d, M ′ = O(d522d) and Q′ = 0. By
Theorem 66, we get a query complexity O(d2d/ǫ + d522d) in the distribution-free model.
References
[1] Noga Alon, Tali Kaufman, Michael Krivelevich, Simon Litsyn, and Dana Ron. Test-
ing low-degree polynomials over GF(2).
In Approximation, Randomization, and Combi-
natorial Optimization: Algorithms and Techniques, 6th International Workshop on Ap-
proximation Algorithms for Combinatorial Optimization Problems, APPROX 2003 and
7th International Workshop on Randomization and Approximation Techniques in Com-
puter Science, RANDOM 2003, Princeton, NJ, USA, August 24-26, 2003, Proceed-
ings, pages 188 -- 199, 2003.
URL: https://doi.org/10.1007/978-3-540-45198-3_17,
doi:10.1007/978-3-540-45198-3\_17.
[2] Noga Alon, Tali Kaufman, Michael Krivelevich, Simon Litsyn, and Dana Ron. Test-
IEEE Trans. Information Theory, 51(11):4032 -- 4039, 2005. URL:
ing reed-muller codes.
https://doi.org/10.1109/TIT.2005.856958, doi:10.1109/TIT.2005.856958.
[3] Dana Angluin. Queries and concept learning. Machine Learning, 2(4):319 -- 342, 1987.
[4] Roksana Baleshzar, Meiram Murzabulatov, Ramesh Krishnan S. Pallavoor, and Sofya
Raskhodnikova. Testing unateness of real-valued functions. CoRR, abs/1608.07652, 2016.
URL: http://arxiv.org/abs/1608.07652, arXiv:1608.07652.
[5] Aleksandrs Belovs and Eric Blais. A polynomial
lower bound for testing monotonicity.
In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing,
51
STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 1021 -- 1032, 2016. URL:
https://doi.org/10.1145/2897518.2897567, doi:10.1145/2897518.2897567.
[6] Arnab Bhattacharyya, Swastik Kopparty, Grant Schoenebeck, Madhu Sudan, and David Zuck-
In Property Testing - Current Research and
erman. Optimal testing of reed-muller codes.
Surveys, pages 269 -- 275. 2010. URL: https://doi.org/10.1007/978-3-642-16367-8_19,
doi:10.1007/978-3-642-16367-8\_19.
Improved bounds
and Combinatorial Optimization. Algorithms
[7] Eric Blais.
domization
International Workshop,
RANDOM 2008, Boston, MA, USA, August
317 -- 330,
doi:10.1007/978-3-540-85363-3\_26.
In Approximation, Ran-
and Techniques,
11th
International Workshop,
pages
https://doi.org/10.1007/978-3-540-85363-3_26,
2008. Proceedings,
APPROX 2008,
juntas.
testing
25-27,
URL:
2008.
12th
and
for
[8] Eric Blais. Testing juntas nearly optimally.
In Proceedings of the 41st Annual ACM
Symposium on Theory of Computing, STOC 2009, Bethesda, MD, USA, May 31 -
June 2, 2009, pages 151 -- 158, 2009. URL: https://doi.org/10.1145/1536414.1536437,
doi:10.1145/1536414.1536437.
[9] Eric Blais, Joshua Brody, and Kevin Matulef. Property testing lower bounds via commu-
nication complexity. In Proceedings of the 26th Annual IEEE Conference on Computational
Complexity, CCC 2011, San Jose, California, USA, June 8-10, 2011, pages 210 -- 220, 2011.
URL: https://doi.org/10.1109/CCC.2011.31, doi:10.1109/CCC.2011.31.
[10] Eric Blais and Daniel M. Kane.
Tight bounds for testing k-linearity.
In Approx-
imation, Randomization,
and Tech-
niques - 15th International Workshop, APPROX 2012, and 16th International Work-
shop, RANDOM 2012, Cambridge, MA, USA, August 15-17,
2012. Proceedings,
pages
https://doi.org/10.1007/978-3-642-32512-0_37,
doi:10.1007/978-3-642-32512-0\_37.
and Combinatorial Optimization. Algorithms
435 -- 446,
URL:
2012.
[11] Avrim Blum. Learning a function of r relevant variables. In Computational Learning The-
ory and Kernel Machines, 16th Annual Conference on Computational Learning Theory and
7th Kernel Workshop, COLT/Kernel 2003, Washington, DC, USA, August 24-27, 2003, Pro-
ceedings, pages 731 -- 733, 2003. URL: https://doi.org/10.1007/978-3-540-45167-9_54,
doi:10.1007/978-3-540-45167-9\_54.
[12] Avrim
Blum
examples
and
in
and
271,
doi:10.1016/S0004-3702(97)00063-5.
1997.
Pat
machine
URL:
Langley.
Selection
features
97(1-2):245 --
https://doi.org/10.1016/S0004-3702(97)00063-5,
learning.
relevant
Intell.,
Artif.
of
[13] Manuel Blum, Michael Luby,
with applications
595,
doi:10.1016/0022-0000(93)90044-W.
URL:
1993.
to numerical problems.
and Ronitt Rubinfeld.
Self-testing/correcting
47(3):549 --
https://doi.org/10.1016/0022-0000(93)90044-W,
J. Comput. Syst. Sci.,
52
[14] Anselm
Blumer,
Andrzej
Ehrenfeucht,
fred K. Warmuth.
380,
doi:10.1016/0020-0190(87)90114-1.
Occam's
URL:
1987.
Man-
24(6):377 --
https://doi.org/10.1016/0020-0190(87)90114-1,
David
Inf. Process.
Haussler,
razor.
Lett.,
and
[15] Nader H. Bshouty. Almost optimal distribution-free junta testing. CoRR, abs/1901.00717,
2018.
[16] Nader H. Bshouty. Exact learning from an honest teacher that answers membership queries.
Theor. Comput. Sci., 733:4 -- 43, 2018. URL: https://doi.org/10.1016/j.tcs.2018.04.034,
doi:10.1016/j.tcs.2018.04.034.
[17] Nader H. Bshouty and Areej Costa. Exact learning of juntas from membership queries. Theor.
Comput. Sci., 742:82 -- 97, 2018. URL: https://doi.org/10.1016/j.tcs.2017.12.032,
doi:10.1016/j.tcs.2017.12.032.
[18] Nader H. Bshouty and Yishay Mansour.
trees and multivariate polynomials.
https://doi.org/10.1137/S009753979732058X, doi:10.1137/S009753979732058X.
for decision
SIAM J. Comput., 31(6):1909 -- 1925, 2002. URL:
Simple learning algorithms
[19] Deeparnab Chakrabarty and C. Seshadhri.
functions over the hypercube.
STOC'13, Palo Alto, CA, USA, June 1-4, 2013, pages 411 -- 418,
https://doi.org/10.1145/2488608.2488660, doi:10.1145/2488608.2488660.
A o(n) monotonicity tester for boolean
In Symposium on Theory of Computing Conference,
URL:
2013.
[20] Deeparnab Chakrabarty and C. Seshadhri. A O(n) non-adaptive tester for unateness. CoRR,
abs/1608.06980, 2016. URL: http://arxiv.org/abs/1608.06980, arXiv:1608.06980.
[21] Sourav Chakraborty, David Garc´ıa-Soriano, and Arie Matsliah. Efficient sample extrac-
tors for juntas with applications.
In Automata, Languages and Programming - 38th In-
ternational Colloquium, ICALP 2011, Zurich, Switzerland, July 4-8, 2011, Proceedings,
Part I, pages 545 -- 556, 2011. URL: https://doi.org/10.1007/978-3-642-22006-7_46,
doi:10.1007/978-3-642-22006-7\_46.
[22] Xi Chen, Anindya De, Rocco A. Servedio, and Li-Yang Tan. Boolean function monotonic-
ity testing requires (almost) n1/2 non-adaptive queries. In Proceedings of the Forty-Seventh
Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA,
June 14-17, 2015, pages 519 -- 528, 2015. URL: https://doi.org/10.1145/2746539.2746570,
doi:10.1145/2746539.2746570.
[23] Xi Chen, Rocco A. Servedio, and Li-Yang Tan. New algorithms and lower bounds for mono-
tonicity testing. CoRR, abs/1412.5655, 2014. URL: http://arxiv.org/abs/1412.5655,
arXiv:1412.5655.
[24] Xi Chen, Rocco A. Servedio, Li-Yang Tan, Erik Waingarten, and Jinyu Xie.
Settling
the query complexity of non-adaptive junta testing.
In 32nd Computational Complex-
ity Conference, CCC 2017, July 6-9, 2017, Riga, Latvia, pages 26:1 -- 26:19, 2017. URL:
https://doi.org/10.4230/LIPIcs.CCC.2017.26, doi:10.4230/LIPIcs.CCC.2017.26.
53
[25] Xi Chen, Erik Waingarten, and Jinyu Xie.
Beyond talagrand functions: new lower
bounds for testing monotonicity and unateness.
In Proceedings of the 49th Annual ACM
SIGACT Symposium on Theory of Computing, STOC 2017, Montreal, QC, Canada, June
19-23, 2017, pages 523 -- 536, 2017. URL: https://doi.org/10.1145/3055399.3055461,
doi:10.1145/3055399.3055461.
[26] Xi Chen, Erik Waingarten, and Jinyu Xie.
adaptive queries.
In 58th IEEE Annual Symposium on Foundations of Computer Sci-
ence, FOCS 2017, Berkeley, CA, USA, October 15-17, 2017, pages 868 -- 879, 2017. URL:
https://doi.org/10.1109/FOCS.2017.85, doi:10.1109/FOCS.2017.85.
Boolean unateness testing with eO(n3/4)
[27] Xi Chen and Jinyu Xie. Tight bounds for the distribution-free testing of monotone con-
junctions. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 54 -- 71, 2016. URL:
https://doi.org/10.1137/1.9781611974331.ch5, doi:10.1137/1.9781611974331.ch5.
[28] Hana Chockler and Dan Gutfreund.
Inf. Pro-
cess. Lett., 90(6):301 -- 305, 2004. URL: https://doi.org/10.1016/j.ipl.2004.01.023,
doi:10.1016/j.ipl.2004.01.023.
A lower bound for testing juntas.
[29] Peter Damaschke.
Ma-
chine Learning, 41(2):197 -- 215, 2000. URL: https://doi.org/10.1023/A:1007616604496,
doi:10.1023/A:1007616604496.
Adaptive versus nonadaptive attribute-efficient
learning.
[30] Ilias Diakonikolas, Homin K. Lee, Kevin Matulef, Krzysztof Onak, Ronitt Rubinfeld, Rocco A.
Servedio, and Andrew Wan. Testing for concise representations. In 48th Annual IEEE Sym-
posium on Foundations of Computer Science (FOCS 2007), October 20-23, 2007, Providence,
RI, USA, Proceedings, pages 549 -- 558, 2007. URL: https://doi.org/10.1109/FOCS.2007.32,
doi:10.1109/FOCS.2007.32.
[31] Ilias Diakonikolas, Homin K. Lee, Kevin Matulef, Rocco A. Servedio, and Andrew Wan.
Efficiently testing sparse GF (2) polynomials. Algorithmica, 61(3):580 -- 605, 2011. URL:
https://doi.org/10.1007/s00453-010-9426-9, doi:10.1007/s00453-010-9426-9.
[32] Elya Dolev and Dana Ron.
linear number of queries.
https://doi.org/10.4086/toc.2011.v007a011, doi:10.4086/toc.2011.v007a011.
Theory of Computing,
2011.
Distribution-free testing for monomials with a sub-
URL:
7(1):155 -- 176,
[33] Eldar Fischer, Guy Kindler, Dana Ron, Shmuel Safra, and Alex Samorodnitsky. Test-
ing juntas.
In 43rd Symposium on Foundations of Computer Science (FOCS 2002), 16-
19 November 2002, Vancouver, BC, Canada, Proceedings, pages 103 -- 112, 2002. URL:
https://doi.org/10.1109/SFCS.2002.1181887, doi:10.1109/SFCS.2002.1181887.
[34] Dana Glasner
and Rocco A. Servedio.
Distribution-free
testing
for basic boolean functions.
https://doi.org/10.4086/toc.2009.v005a010, doi:10.4086/toc.2009.v005a010.
Theory of Computing,
5(1):191 -- 216,
2009.
lower bound
URL:
[35] Oded Goldreich.
On testing computability by small width obdds.
imation, Randomization,
niques,
13th
and Combinatorial Optimization. Algorithms
International Workshop, APPROX 2010,
and
14th
In Approx-
and Tech-
International
54
Workshop, RANDOM 2010, Barcelona, Spain, September 1-3,
pages
doi:10.1007/978-3-642-15369-3\_43.
2010. Proceedings,
https://doi.org/10.1007/978-3-642-15369-3_43,
574 -- 587,
URL:
2010.
[36] Oded Goldreich,
editor.
ume
https://doi.org/10.1007/978-3-642-16367-8, doi:10.1007/978-3-642-16367-8.
of Lecture Notes
Property Testing - Current Research and Surveys, vol-
URL:
in Computer Science.
Springer,
2010.
6390
[37] Oded Goldreich, Shafi Goldwasser, Eric Lehman, Dana Ron, and Alex Samorod-
URL:
nitsky.
https://doi.org/10.1007/s004930070011, doi:10.1007/s004930070011.
Testing monotonicity.
Combinatorica,
20(3):301 -- 337,
2000.
[38] Oded Goldreich, Shafi Goldwasser, and Dana Ron.
nection to learning and approximation.
J. ACM, 45(4):653 -- 750,
https://doi.org/10.1145/285055.285060, doi:10.1145/285055.285060.
Property testing and its con-
URL:
1998.
[39] Parikshit Gopalan, Ryan O'Donnell, Rocco A. Servedio, Amir Shpilka, and Karl Wimmer.
Testing fourier dimensionality and sparsity. SIAM J. Comput., 40(4):1075 -- 1100, 2011. URL:
https://doi.org/10.1137/100785429, doi:10.1137/100785429.
[40] David Guijarro, Jun Tarui, and Tatsuie Tsukiji. Finding relevant variables in PAC model
with membership queries.
In Algorithmic Learning Theory, 10th International Confer-
ence, ALT '99, Tokyo, Japan, December 6-8, 1999, Proceedings, page 313, 1999. URL:
https://doi.org/10.1007/3-540-46769-6_26, doi:10.1007/3-540-46769-6\_26.
[41] Shirley Halevy and Eyal Kushilevitz.
J. Comput.,
doi:10.1137/050645804.
37(4):1107 -- 1138,
2007.
Distribution-free property-testing.
SIAM
URL: https://doi.org/10.1137/050645804,
[42] Subhash Khot, Dor Minzer, and Muli Safra. On monotonicity testing and boolean isoperi-
metric type theorems. In IEEE 56th Annual Symposium on Foundations of Computer Sci-
ence, FOCS 2015, Berkeley, CA, USA, 17-20 October, 2015, pages 52 -- 58, 2015. URL:
https://doi.org/10.1109/FOCS.2015.13, doi:10.1109/FOCS.2015.13.
[43] Subhash Khot and Igor Shinkar. An eO(n) queries adaptive tester for unateness. CoRR,
abs/1608.02451, 2016. URL: http://arxiv.org/abs/1608.02451, arXiv:1608.02451.
[44] Richard J. Lipton, Evangelos Markakis, Aranyak Mehta, and Nisheeth K. Vishnoi.
On the fourier spectrum of symmetric boolean functions with applications to learn-
ing symmetric juntas.
In 20th Annual IEEE Conference on Computational Complex-
ity (CCC 2005), 11-15 June 2005, San Jose, CA, USA, pages 112 -- 119, 2005. URL:
https://doi.org/10.1109/CCC.2005.19, doi:10.1109/CCC.2005.19.
[45] Zhengyang Liu, Xi Chen, Rocco A. Servedio, Ying Sheng, and Jinyu Xie. Distribution-free
In Proceedings of the 50th Annual ACM SIGACT Symposium on Theory of
junta testing.
Computing, STOC 2018, Los Angeles, CA, USA, June 25-29, 2018, pages 749 -- 759, 2018.
URL: https://doi.org/10.1145/3188745.3188842, doi:10.1145/3188745.3188842.
55
[46] Kevin Matulef, Ryan O'Donnell, Ronitt Rubinfeld, and Rocco A. Servedio. Testing ±1-
weight halfspace.
In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques, 12th International Workshop, APPROX 2009, and 13th In-
ternational Workshop, RANDOM 2009, Berkeley, CA, USA, August 21-23, 2009. Pro-
ceedings, pages 646 -- 657, 2009. URL: https://doi.org/10.1007/978-3-642-03685-9_48,
doi:10.1007/978-3-642-03685-9\_48.
[47] Kevin Matulef, Ryan O'Donnell, Ronitt Rubinfeld, and Rocco A. Servedio. Testing halfspaces.
SIAM J. Comput., 39(5):2004 -- 2047, 2010. URL: https://doi.org/10.1137/070707890,
doi:10.1137/070707890.
[48] Elchanan Mossel, Ryan O'Donnell,
and Rocco A. Servedio.
of k relevant variables.
https://doi.org/10.1016/j.jcss.2004.04.002, doi:10.1016/j.jcss.2004.04.002.
J. Comput. Syst. Sci.,
69(3):421 -- 434,
Learning functions
URL:
2004.
[49] Noam Nisan and Mario Szegedy. On the degree of boolean functions as real poly-
nomials.
In Proceedings of the 24th Annual ACM Symposium on Theory of Comput-
ing, May 4-6, 1992, Victoria, British Columbia, Canada, pages 462 -- 467, 1992. URL:
https://doi.org/10.1145/129712.129757, doi:10.1145/129712.129757.
[50] Michal
boolean
SIAM J. Discrete Math.,
http://epubs.siam.org/sam-bin/dbq/article/40744.
Parnas,
formulae.
Dana Ron,
and Alex
Samorodnitsky.
Testing
16(1):20 -- 46,
2002.
basic
URL:
[51] Ronald L. Rivest. Learning decision lists. Machine Learning, 2(3):229 -- 246, 1987. URL:
https://doi.org/10.1007/BF00058680, doi:10.1007/BF00058680.
[52] Dana Ron. Property testing: A learning theory perspective. Foundations and Trends
in Machine Learning, 1(3):307 -- 402, 2008. URL: https://doi.org/10.1561/2200000004,
doi:10.1561/2200000004.
[53] Dana Ron.
Algorithmic and analysis
techniques
dations and Trends
https://doi.org/10.1561/0400000029, doi:10.1561/0400000029.
in Theoretical Computer Science,
in property testing.
2009.
5(2):73 -- 205,
Foun-
URL:
[54] Ronitt Rubinfeld and Madhu Sudan.
applications
https://doi.org/10.1137/S0097539793255151, doi:10.1137/S0097539793255151.
SIAM J. Comput., 25(2):252 -- 271, 1996.
to program testing.
Robust characterizations of polynomials with
URL:
[55] Mert Saglam. Near log-convexity of measured heat in (discrete) time and consequences. In 59th
IEEE Annual Symposium on Foundations of Computer Science, FOCS 2018, Paris, France,
October 7-9, 2018, pages 967 -- 978, 2018. URL: https://doi.org/10.1109/FOCS.2018.00095,
doi:10.1109/FOCS.2018.00095.
[56] Leslie G. Valiant. A theory of the learnable. Commun. ACM, 27(11):1134 -- 1142, 1984. URL:
https://doi.org/10.1145/1968.1972, doi:10.1145/1968.1972.
56
9 Appendix A
In this Appendix, we give some bounds that are used in the paper
Lemma 67. Markov's Bound. Let X ≥ 0 be a random variable with finite expected value
µ = E[X]. Then for any real numbers κ, K > 0,
E[X]
Pr(X ≥ κ) ≤
κ
Pr(X ≥ KE[X]) ≤
1
K
.
.
(4)
(5)
Lemma 68. Chebyshev's Bound. Let X be a random variable with finite expected value µ =
E[X] and finite non-zero variance Var[X] = E[X 2] − E[X]2. Then for any real numbers κ, K > 0,
Pr(X − µ ≥ κpVar[X]) ≤
Pr(X − µ ≥ K) ≤
1
κ2 .
Var[X]
.
K 2
(6)
(7)
Lemma 69. Chernoff 's Bound. Let X1, . . . , Xm are independent random variables taking values
i=1 Xi denote their sum and let µ = E[X] denote the sum's expected value.
in {0, 1}. Let X =Pm
Then
(8)
(9)
Pr[X > (1 + λ)µ] ≤(cid:18)
eλ
(1 + λ)(1+λ)(cid:19)µ
≤ e− λ2µ
2+λ ≤(e− λ2µ
e− λµ
3
3
if 0 < λ ≤ 1
if λ > 1
.
For 0 ≤ λ ≤ 1 we have
Pr[X < (1 − λ)µ] ≤(cid:18)
e−λ
(1 − λ)(1−λ)(cid:19)µ
≤ e− λ2µ
2 .
Lemma 70. Hoeffding's Bound. Let X1, . . . , Xm are independent random variables taking values
i=1 Xi denote their sum and let µ = E[X] denote the sum's expected value.
in {0, 1}. Let X =Pm
Then for 0 ≤ λ ≤ 1 we have
and
Pr[X > µ + λm] ≤ e−2λ2m
Pr[X < µ − λm] ≤ e−2λ2m
(10)
(11)
57
|
1507.01490 | 1 | 1507 | 2015-07-06T14:58:24 | Fast and Simple Computation of Top-k Closeness Centralities | [
"cs.DS"
] | Closeness is an important centrality measure widely used in the analysis of real-world complex networks. In particular, the problem of selecting the k most central nodes with respect to this measure has been deeply analyzed in the last decade. However, even for not very large networks, this problem is computationally intractable in practice: indeed, Abboud et al have recently shown that its complexity is strictly related to the complexity of the All-Pairs Shortest Path (in short, APSP) problem, for which no subcubic "combinatorial" algorithm is known. In this paper, we propose a new algorithm for selecting the k most closeness central nodes in a graph. In practice, this algorithm significantly improves over the APSP approach, even though its worst-case time complexity is the same. For example, the algorithm is able to compute the top k nodes in few dozens of seconds even when applied to real-world networks with millions of nodes and edges. We will also experimentally prove that our algorithm drastically outperforms the most recently designed algorithm, proposed by Olsen et al. Finally, we apply the new algorithm to the computation of the most central actors in the IMDB collaboration network, where two actors are linked if they played together in a movie. | cs.DS | cs | Fast and Simple Computation
of Top-k Closeness Centralities
IMT Institute for Advanced Studies Lucca, Italy
Michele Borassi
Pierluigi Crescenzi
Dipartimento di Ingegneria dell’Informazione, Università di Firenze
Andrea Marino
Dipartimento di Informatica, Università di Pisa
5
1
0
2
l
u
J
6
]
S
D
.
s
c
[
1
v
0
9
4
1
0
.
7
0
5
1
:
v
i
X
r
a
ABSTRACT
Closeness is an important centrality measure widely used in
the analysis of real-world complex networks. In particular,
the problem of selecting the k most central nodes with re-
spect to this measure has been deeply analyzed in the last
decade. However, even for not very large networks, this
problem is computationally intractable in practice: indeed,
Abboud et al have recently shown that its complexity is
strictly related to the complexity of the All-Pairs Shortest
Path (in short, APSP) problem, for which no subcubic “com-
binatorial” algorithm is known. In this paper, we propose
a new algorithm for selecting the k most closeness central
nodes in a graph. In practice, this algorithm significantly im-
proves over the APSP approach, even though its worst-case
time complexity is the same. For example, the algorithm is
able to compute the top k nodes in few dozens of seconds
even when applied to real-world networks with millions of
nodes and edges. We will also experimentally prove that
our algorithm drastically outperforms the most recently de-
signed algorithm, proposed by Olsen et al. Finally, we apply
the new algorithm to the computation of the most central
actors in the IMDB collaboration network, where two actors
are linked if they played together in a movie.
Categories and Subject Descriptors
G.2.2 [Discrete Mathematics]: Graph Theory—graph al-
gorithms; H.2.8 [Database Management]: Database Ap-
plications—data mining; I.1.2 [Computing Methodolo-
gies]: Algorithms—analysis of algorithms
General Terms
Design, Algorithms, Performance
INTRODUCTION
1.
The problem of identifying the most central nodes in a net-
work is a fundamental question that has been asked many
times in a plethora of research areas, such as biology, com-
puter science, sociology, and psychology. Because of the im-
portance of this question, several centrality measures have
been introduced in the literature (for a recent survey, see [3]).
Closeness is certainly one of the oldest and of the most
widely used among these measures [2]. Informally, the close-
ness of a node v is the inverse of the expected distance be-
tween v and a random node w, and it somehow estimates the
efficiency of a node in spreading information to other nodes:
the larger the closeness, the most “influential” the node is.
Formally, given a (directed strongly) connected graph G =
(V, E), the closeness of a vertex v is defined as c(v) = n−1
f (v) ,
where n = V , f (v) = Pw∈V d(v, w) is the farness of v,
and d(v, w) is the distance between the two vertices v and
w (that is, the number of edges in a shortest path from
v to w). If G is not (strongly) connected, the definition is
more complicated, because d(v, w) is defined only for vertices
reachable from v. Let R(v) be the set of these vertices,
and let r(v) denote its cardinality (note that v ∈ R(v) by
definition). In the literature [15, 21, 17], the most common
generalization is
c(v) =
r(v) − 1
r(v) − 1
f (v)
n − 1
=
(r(v) − 1)2
(n − 1)f (v)
where f (v) = Pw∈R(v) d(v, w).
In order to compute the k vertices with largest closeness,
the textbook algorithm computes c(v) for each v and re-
turns the k largest found values. The main bottleneck of
this approach is the computation of d(v, w), for each pair
of vertices v and w (that is, solving the All-Pairs Shortest
Paths or APSP problem). This can be done in two ways:
either by using fast matrix multiplication, in time O(n2.373)
[22], or by performing a breadth-first search (in short, BFS)
from each vertex v ∈ V , in time O(mn), where m = E.
Usually, the BFS approach is preferred because the other
approach contains big constants hidden in the O notation,
and because real-world networks are usually sparse, that is,
m is not much bigger than n. However, also this approach
is too time-consuming if the input graph is very big (with
millions of nodes and hundreds of millions of edges). Our
algorithm heavily improves the BFS approach by “cutting
the BFSes”, through an efficient pruning procedure, which
allows us to either compute the closeness of a node v or stop
the corresponding BFS as soon as we are sure that the close-
ness of v is not among the k largest ones. The worst-case
Name
Graphs
Degree functions
Distance function
Reachability set function
Neighborhood functions
Ball functions
Farness functions
Closeness functions
Table 1: notations used throughout the paper.
Symbol
Definition
G = (V, E)
n
G = (V, E )
deg(v)
outdeg(v)
d(v, w)
R(v)
r(v)
α(v)
ω(v)
Γd(v)
γd(v)
γd+1(v)
Nd(v)
nd(v)
f (v)
fd(v)
fd(v, x)
c(v)
cd(v)
Graph with node/vertex set V and edge/arc set E
V
Weighted DAG of strongly connected components (see Section 1.3)
Degree of a node in an undirected graph
Out-degree of a node in a directed graph
Number of edges in a shortest path from v to w
Set of nodes reachable from v (by definition, v ∈ R(v))
R(v)
Lower bound on r(v), that is, α(v) ≤ r(v) (see Section 2.4)
Upper bound on r(v), that is, r(v) ≤ ω(v) (see Section 2.4)
Set of nodes at distance d from v, that is, {w ∈ V : d(v, w) = d}
Number of nodes at distance d from v, that is, Γd(v)
Upper bound on γd+1(v), defined as Pu∈Γd(v) deg(u) if the graph is
undirected, Pu∈Γd(v) outdeg(u) otherwise (clearly, γd+1(v) ≤ γd+1(v))
Set of nodes at distance at most d from v, that is, {w ∈ V : d(v, w) ≤ d}
Number of nodes at distance at most d from v, that is, Nd(v)
Farness of node v, that is, Pw∈R(v) d(v, w)
Farness of node v up to distance d, that is, Pw∈Nd (v) d(v, w)
Lower bound function on farness of node v, that is, fd(v) − γd+1(v) +
(d + 2)(x − nd(v)) (see Lemma 1)
Closeness of node v, tha is, (r(v)−1)2
Upper bound on closeness of node v, that is,
lary 1)
(see Corol-
(n−1)f (v)
(r(v)−1)2
(n−1) fd(v,r(v))
time complexity of this algorithm is the same as the one of
the textbook algorithm: indeed, under reasonable complex-
ity assumptions, this worst-case bound cannot be improved.
In particular, it was proved in [1] that the complexity of
finding the most central vertex is at least the complexity
of the APSP: faster combinatorial algorithms for this latter
problem would imply unexpected breakthrough in complex-
ity theory [23]. However, in practice, our algorithm heavily
improves the APSP approach, as shown by our experiments.
Finally, we will apply our algorithm to an interesting and
prototypical case study, that is, the IMDB actor collabora-
tion network:
in a little more than half an hour, we have
been able to identify the evolution of the set of the 10 most
central actors, by analysing snapshots of this network taken
every 5 years, starting from 1940 and up to 2014.
1.1 Related Work
Closeness is a “traditional” definition of centrality, and con-
sequently it was not “designed with scalability in mind”, as
stated in [12]. Also in [5], it is said that closeness central-
ity can “identify influential nodes”, but it is “incapable to
be applied in large-scale networks due to the computational
complexity”. The simplest solution considered was to de-
fine different measures, that might be related to closeness
centrality [12].
A different line of research has tried to develop more efficient
algorithms for this problem. As already said, due to theoret-
ical bounds [4, 1], the worst-case complexity cannot be im-
proved in general, unless widely believed conjectures prove
to be false. Nevertheless, it is possible to design approxima-
tion algorithms: the simplest approach samples the distance
between a node v and l other nodes w, and return the aver-
age of all values d(v, w) found [10]. The time-complexity is
O(lm) to approximate the centrality of all nodes. More re-
fined approximation algorithms are provided in [7, 6], based
on the concept of All-Distance Sketch, that is, a procedure
that computes the distances between v and O(log n) other
nodes, chosen in order to provide good estimates of the close-
ness centrality of v. Even if these approximation algorithms
work quite well, they are not suited to the ranking of nodes,
because the difference between the closeness centrality of
two nodes might be very small. Nevertheless, they were
used in [16], where the sampling technique developed in [10]
was used to actually compute the top k vertices: the re-
sult is not exact, but it is exact with high probability. The
authors proved that the time-complexity of their algorithm
3 log n), under the rather strong assumption that
is O(mn
closeness centralities are uniformly distributed between 0
and D, where D is the maximum distance between two nodes
(in the worst case, the time-complexity of this algorithm is
O(mn)).
2
Other approaches have tried to develop incremental algo-
rithms, that might be more suited to real-world networks
analyses. For instance, in [14], the authors develop heuristics
to determine the k most central vertices in a varying environ-
ment. A different work addressed the problem of updating
centralities after edge insertion or deletion [18]: for instance,
it is shown that it is possible to update the closeness central-
ity of 1.2 million authors in the DBLP-coauthorship network
460 times faster than recomputing it from scratch.
Finally, some works have tried to exploit properties of real-
world networks in order to find more efficient algorithms.
In [13], the authors develop a heuristic to compute the k
most central vertices according to different measures. The
basic idea is to identify central nodes according to a sim-
ple centrality measure (for instance, degree of nodes), and
then to inspect a small set of central nodes according to this
measure, hoping it will contain the top k vertices according
to the “complex” measure. Another approach [17] tried to
exploit the properties of real-world networks in order to de-
velop exact algorithms with worst-case complexity O(mn),
but performing much better in practice. As far as we know,
this is the only exact algorithm that is able to efficiently
compute the k most central vertices in networks with up to
1 million nodes.
However, despite this huge amount of research, the ma-
jor graph libraries still implement the textbook algorithm:
among them, Boost Graph Library [11], Sagemath [9],
igraph [20], and NetworkX [19]. This is due to the fact
that the only efficient algorithm published until now for top
k closeness centrality is [17], which was published only 1
year ago, and it is quite complicated, because it is based on
several other algorithms.
1.2 Our Results
As we said before, in this paper we present a new and very
simple algorithm for the exact computation of top k close-
ness central vertices. We show that we drastically improve
both the probabilistic approach [16], and the best algorithm
available until now [17]. We have computed for the first
time the 10 most central nodes in networks with millions of
nodes and hundreds of millions of edges, in very little time.
A significant example is the IMDB actor network (1 797 446
nodes and 72 880 156 edges): we have computed the 10 most
central actors in less than 10 minutes. Moreover, in our
DBLP co-authorship network (which should be quite sim-
ilar to the network used in [18]), our performance is more
than 6 000 times better than the performance of the text-
book algorithm:
if only the most central node is needed,
we can recompute it from scratch more than 10 times faster
than performing their update. Finally, our approach is not
only very efficient, but it is also very easy to code, making it
a very good candidate to be implemented in existing graph
libraries.
In addition to these definitions,
1.3 Preliminary Definitions
We assume the reader to be familiar with the basic notions
of graph theory (see, for example, [8]): all the notations
and definitions used throughout this paper are summarised
in Table 1.
let us pre-
cisely define the weighted directed acyclic graph G = (V, E )
of strongly connected components (in short, SCCs) corre-
sponding to a directed graph G = (V, E). In this graph, V
is the set of SCCs of G, and, for any two SCCs C, D ∈ V,
(C, D) ∈ E if and only if there is an arc in E from a node in
C to the a node in D. For each SCC C ∈ V, the weight w(C)
of C is equal to C, that is, the number of nodes in the SCC
C. Note that, for each node v ∈ C, r(v) = PD∈R(C) w(D),
where R(C) denotes the set of SCCs that are reachable from
C in G, and r(C) denotes its cardinality.
1.4 Structure of the Paper
In Section 2, we will explain how our algorithm works, and
we will prove its correctness. Section 3 will experimentally
prove that our algorithm outperforms the best available al-
gorithms, by performing several tests on a dataset of real-
world networks. Section 4 applies the new algorithm to the
analysis of the 10 most central actors in the IMDB actor
network.
2. THE ALGORITHM
2.1 Overview
As we already said, the textbook algorithm for comput-
ing the k vertices with largest closeness performs a BFS
from each vertex v, computes its closeness c(v), and, finally,
returns the k vertices with biggest c(v) values. Similarly,
our algorithm (see Algorithm 1) sets c(v) = BFSCut(v, xk),
where xk is the k-th biggest closeness value found until now
(xk = 0 if we have not processed at least k vertices).
If
BFSCut(v, xk) = 0, it means that v is not one of k most cen-
tral vertices, otherwise c(v) is the actual closeness of v. This
means that, at the end, the k vertices with biggest closeness
values are again the k most central vertices.
In order to
speed-up the function BFSCut(v, xk), we want xk to be as
big as possible, and consequently we need to process central
vertices as soon as possible. To this purpose, following the
idea of [14], we process vertices in decreasing order of degree.
Algorithm 1: overview of the new algorithm. The function
Kth(c) returns the k-th biggest element of c and the function
TopK(c) returns the k biggest elements.
1 PreProcessing (G); // see Section 2.4
2 c(v) ← 0 for each v;
3 xk ← 0;
4 for v ∈ V in decreasing order of degree do
5
6
7
c(v) ← BFSCut (v, xk);
if c(v) 6= 0 then
xk ← Kth(c);
8 return TopK(c);
As we will see in the next sections, Algorithm 1 needs linear
time in order to execute the preprocessing. Moreover, it
requires time O(n log n) to sort vertices, O(log k) to perform
function Kth, and O(1) to perform function TopK, by using
a priority queue containing at each step the k most central
vertices. Since all other operations need time O(1), the total
running time is O(m+n log n+n log k+T ) = O(m+n log n+
T )), where T is the time needed to perform the function
BFSCut n times.
Before explaining in details how the function BFSCut op-
erates,
let us note that we can easily parallelise the for
loop in Algorithm 1, by giving each vertex to a different
thread. The parallelisation is almost complete, because the
only “non-parallel” part deals with the shared variable xk,
which must be updated in a synchronised way. In any case,
this shared variable does not affect performances. First of
all, the time needed for the update is very small, compared
to the time of BFSes. Moreover, in principle, a thread might
use an “old” value of xk, and consequently lose performance:
we will experimentally show that the effect of this positive
race condition is negligible (Section 3.2).
2.2 An Upper Bound on the Closeness of
Lemma 2. For each vertex v and for each d ≥ 0,
Nodes
The goal of this section is to define an upper bound cv() on
the closeness of a node v, which has to be updated whenever,
for any d ≥ 0, all nodes in Γd(v) has been reached by the
BFS starting from v (that is, whenever the exploration of
the d-th level of the BFS tree is finished). More precisely,
this upper bound is obtained by first proving a lower bound
on the farness of v, as shown in the following lemma.
Lemma 1. For each vertex v and for each d ≥ 0,
f (v) ≥ fd(v, r(v))
(see Table 1 for the definition of fd(v, r(v))).
Proof. From the definitions of Table 1, it follows that
f (v) ≥ fd(v) + (d + 1)γd+1(v) + (d + 2)(r(v) − nd+1(v)).
Since nd+1(v) = γd+1(v) + nd(v),
f (v) ≥ fd(v) − γd+1(v) + (d + 2)(r(v) − nd(v)).
Finally, since γd+1(v) is an upper bound on γd+1(v),
f (v) ≥ fd(v) − γd+1(v) + (d + 2)(r(v) − nd(v)) = fd(v, r(v))
1
c(v)
≥ λd(v) := (n − 1) min(cid:18) fd(v, α(v))
(α(v) − 1)2
,
fd(v, ω(v))
(ω(v) − 1)2(cid:19) .
Proof. From Lemma 1, if we denote a = d + 2 and b =
γd+1(v) + a(nd(v) − 1) − fd(v),
f (v) ≥ fd(v) − γd+1(v) + a(r(v) − nd(v))
= a(r(v) − 1) + fd(v) − γd+1(v) − a(nd(v) − 1)
= a(r(v) − 1) − b.
Note that a > 0 because d > 0, and b > 0 because
fd(v) = Xw∈Nd(v)
d(v, w) ≤ d(nd(v) − 1) < a(nd(v) − 1)
≥ (n − 1) a(r(v)−1)−b
x3
x2 . The derivative g′(x) = −ax+2b
if w = v, then
where the first inequality holds because,
d(v, w) = 0, and if w ∈ Nd(v), then d(v, w) ≤ d. Hence,
1
(r(v)−1)2 . Let us consider the function
cv
g(x) = ax−b
is positive
for 0 < x < 2b
a : this means that
2b
a is a local maximum, and there are no local minima for
x > 0. Consequently, in each closed interval [x1, x2] where
x1 and x2 are positive, the minimum of g(x) is reached in
x1 or x2. Since 0 < α(v) − 1 ≤ r(v) − 1 ≤ ω(v) − 1,
a and negative for x > 2b
g(r(v) − 1) ≥ min(g(α(v) − 1), g(ω(v) − 1))
and the lemma follows.
and the conclusion follows.
Corollary 1. For each vertex v and for each d ≥ 0,
c(v) ≤ cd(v), where cd(v) is defined in Table 1.
2.3 Computing the Upper Bound
Apart from r(v), all quantities necessary to compute
fd(v, r(v)) (and, hence, to compute the upper bound of
Lemma 1) are available as soon as all vertices in Nd(v) are
visited by a BFS. Note that, if the graph G is (strongly)
connected, then r(v) = n is also available. Moreover, if the
graph G is undirected (but not necessarily connected), we
can compute r(v) for each vertex v in linear time, at the
beginning of the algorithm (by simply computing the con-
nected components of G).1 It thus remain to deal with the
case in which G is directed and not strongly connected.
In this case, let us assume, for now, that we know a lower
(respectively, upper) bound α(v) (respectively, ω(v)) on r(v)
(see also Table 1): without loss of generality we can assume
that α(v) > 1. The next lemma shows that, instead of ex-
amining all possible values of r(v) between α(v) and ω(v), it
is sufficient to examine only the two extremes of this inter-
val. In order to apply this idea, however, we have to work
1
with the inverse
c(v) of the closeness of v, because otherwise
the denominator might vanish or even be negative (due to
the fact that we are both upper bounding γd+1(v) and lower
bounding r(v)). In other words, the lemma will provide us
a lower bound λd(v) on 1
c(v) (so that, if λd(v) is negative,
then λd(v) ≤ 1
1Note that if G is undirected, Lemma 1 still holds if we rede-
c(v) trivially holds).
fine, for any d ≥ 1, γd+1(v) = Pu∈Γd(v)(deg(u)−1), because
at least one edge from each vertex u in Γd(v) connects u to
a node in Γd−1(v).
2.4 Computing α(v) and ω(v)
It now remains to compute α(v) and β(v) (in the case of
a directed graph which is not strongly connected). This
can be done during the preprocessing phase of our algo-
rithm as follows. Let G = (V, E ) be the weighted directed
acyclic graph of SCCs, as defined in Section 1.3. We al-
ready observed that, if v and w are in the same SCC, then
of SCCs that are reachable from C in G. This means that
we simply need to compute a lower (respectively, upper)
r(v) = r(w) = PD∈R(C) w(D), where R(C) denotes the set
bound α(C) (respectively, ω(C)) on PD∈R(C) w(D), for ev-
ery SCC C. To this aim, we first compute a topological sort
{C1, . . . , Cl} of V (that is, if (Ci, Cj) ∈ E , then i < j). Suc-
cessively, we use a dynamic programming approach, and, by
starting from Cl, we process the SCCs in reverse topological
order, and we set
and
α(C) = w(C) + max
(C,D)∈E
α(D)
ω(C) = w(C) + X(C,D)∈E
ω(D).
Note that processing the SCCs in reverse topological order-
ing ensures that the values α(D) and ω(D) on the right
hand side of these equalities are available when we process
the SCC C. Clearly, the complexity of computing α(C) and
ω(C), for each SCC C, is linear in the size of G, which in
turn is smaller than G.
Observe that the bounds obtained through this simple ap-
proach can be improved by using some “tricks”. First of all,
when the biggest SCC C is processed, we do not use the dy-
namic programming approach and we can exactly compute
PD∈R( C) w(D) by simply performing a BFS starting from
any node in C. This way, not only α( C) and ω( C) are exact,
but also α(C) and ω(C) are improved for each SCC C from
which it is possible to reach C. Finally, in order to compute
the upper bounds for the SCCs that are able to reach C, we
can run the dynamic programming algorithm on the graph
obtained from G by removing all components reachable from
C, and we can then add PD∈R( C) w(D).
2.5 The BFSCut Function
We are now ready to define the function BFSCut(v, x), which
returns the closeness c(v) of vertex v if c(v) > x, 0 otherwise.
To do so, this function performs a BFS starting from v, and
during the BFS it updates the upper bound cd(v) ≥ c(v)
(the update is done whenever all nodes in Γd(v) has been
reached): as soon as cd(v) ≤ x, we know that c(v) ≤ cd(v) ≤
x, and we return 0. If this situation never occurs, at the end
of the visit we have clearly computed c(v).
Algorithm 2 is the pseudo-code of the function BFSCut when
implemented for strongly connected graphs (recall that, in
this case, r(v) = n): this code can be easily adapted to the
case of undirected graphs (see the beginning of Section 2.3)
and to the case of directed (not necessarily strongly con-
nected) graphs (see Lemma 2).
Algorithm 2: the BFSCut(v, x) function in the case of
strongly connected directed graphs.
1 Create queue Q;
2 Q.enqueue(v);
3 Mark v as visited;
4 d ← 0; f ← 0; γ ← 0; nd ← 0;
5 while Q is not empty do
6
7
8
9
10
11
12
13
14
15
16
17
18
19
u ← Q.dequeue();
if d(v, u) > d then
f ← f − γ + (d + 2)(n − nd);
c ← (n−1)
if c ≤ x then
return 0;
f
;
d ← d + 1;
f ← f + d(v, u);
γ ← γ + outdeg(u);
nd ← nd + 1;
for w in adjacency list of u do
if w is not visited then
Q.enqueue(w);
Mark w as visited;
20 return n−1
f ;
section, we will
real-world networks,
3. EXPERIMENTAL RESULTS
test our algorithm on sev-
In this
to show its per-
in order
eral
exper-
formances.
SNAP
iments were
(nexus.igraph.org),
(snap.stanford.edu/),
LAW
LASAGNE
(law.di.unimi.it), and IMDB (www.imdb.com).
Our
tests have been performed on a server running an Intel(R)
the networks used in our
datasets
(piluc.dsi.unifi.it/lasagne),
from the
collected
All
NEXUS
Xeon(R) CPU E5-4607 0, 2.20GHz, with 48 cores, 250GB
RAM, running Ubuntu 14.04 2 LTS; our code has been writ-
ten in Java 1.7, and it is available at tinyurl.com/kelujv7.
3.1 Comparison with the State of the Art
In order to compare the performance of our algorithm with
state of the art approaches, we have selected 21 networks
whose number of nodes ranges between 1 000 and 30 000
nodes.
We have compared our algorithm Bcm with our implemen-
tations of the best available algorithms for top-k closeness
centrality.2 The first one [17] is based on a pruning technique
and on ∆-BFS, a method to reuse information collected dur-
ing a BFS from a node to speed up a BFS from one of its
in-neighbors; we will denote this algorithm as Olh. The
second one provides top-k closeness centralities with high
probability [16].
It is based on performing BFSes from a
random sample of nodes to estimate the closeness central-
ity of all the other nodes, then computing the exact cen-
trality of all the nodes whose estimate is big enough. Note
that this algorithm requires the input graph to be (strongly)
connected: for this reason, differently from the other algo-
rithms, we have run this algorithm on the largest (strongly)
connected component. We will denote this latter algorithm
as Ocl.
In order to perform a fair comparison we have considered
the improvement factor mvis
, where mvis is the number of
mtot
arcs visited during the algorithm and mtot is the number of
arcs visited by the textbook algorithm, which is based on all-
pair shortest-path. It is worth observing that mtot is m · n if
the graph is (strongly) connected, but it might be smaller in
general. The improvement factor does not consider the pre-
processing time, which is negligible, and it is closely related
to the running-time, since all the three algorithms are based
on performing BFSes. Furthermore, it does not depend on
the implementation and on the machine used for the algo-
rithm. In the particular case of Olh, we have just counted
the arcs visited in BFS and ∆-BFS, without considering all
the operations done in the pruning phases (see [17]).
Our results are summarized in Table 2, where we report the
arithmetic mean and the geometric mean of all the improve-
ment factors (among all the graphs).
In our opinion, the
geometric mean is more significant, because the arithmetic
mean is highly influenced by the maximum value.
Table 2: comparison of the improvement factors of the new
algorithm (Bcm), of the algorithm in [17] (Olh), and the
algorithm in [16] (Ocl). Values are computed for k = 1 and
k = 10, and they are averaged over all graphs in the dataset.
Arithmetic Mean
Geometric Mean
k Alg
Bcm
1 Olh
Ocl
Bcm
10 Olh
Ocl
Dir
Both
Undir
Undir
Both
Dir
4.5% 2.8% 3.46% 1.3% 0.8% 0.9%
31.6% 35.6% 15.8% 21.5%
43.5% 24.2%
55.6% 67.3% 42.8% 50.8%
72.3% 45.4%
8.6% 6.3% 2.7% 3.8%
14.1% 5.3%
31.6% 35.6% 15.8% 21.6%
43.6% 24.2%
80.7% 59.3%
67.5% 78.4% 57.9% 65.0%
2Note that the source code of our competitors is not avail-
able.
Table 3: detailed comparison of the improvement factor of the three algorithms with respect to the all-pair-shortest-path
algorithm, with k = 1 and k = 10.
Directed Networks
Network
polblogs
p2p-Gnutella08
wiki-Vote
p2p-Gnutella09
p2p-Gnutella06
freeassoc
p2p-Gnutella04
as-caida20071105
Network
Homo
HC-BIOGRID
Mus musculus
Caenorhabditis elegans
ca-GrQc
advogato
hprd pp
ca-HepTh
Drosophila melanogaster
oregon1 010526
oregon2 010526
GoogleNw
dip20090126 MAX
Nodes
1224
6301
7115
8114
8717
10617
10876
26475
Nodes
1027
4039
4610
4723
5242
7418
9465
9877
10625
11174
11461
15763
19928
Edges
19022
20777
103689
26013
31525
72172
39994
106762
Edges
1166
10321
5747
9842
14484
42892
37039
25973
40781
23409
32730
148585
41202
k = 10
k = 1
Olh
Ocl
Olh
Bcm
Bcm
3.052% 41.131% 88.323%
8.491% 41.321% 91.992%
4.592% 53.535% 87.229% 23.646% 53.626% 92.350%
0.068% 25.205% 40.069%
0.825% 25.226% 62.262%
7.458% 55.754% 86.867% 18.649% 55.940% 90.248%
0.808% 52.615% 77.768% 18.432% 52.831% 88.884%
17.315% 58.204% 85.831% 20.640% 57.954% 87.300%
2.575% 56.788% 84.128% 21.754% 56.813% 89.961%
0.036%
4.740% 42.955%
4.740% 27.985%
0.100%
Ocl
Undirected Networks
k = 1
Olh
k = 10
Ocl
Ocl
Olh
Bcm
Bcm
5.259% 82.794% 82.956% 14.121% 82.794% 88.076%
8.928% 19.112% 72.070%
5.914% 19.112% 65.672%
7.535% 66.507%
5.135%
7.535% 55.004%
1.352%
1.161%
9.489% 45.623%
1.749%
9.489% 58.521%
5.115% 13.815% 62.523%
3.472% 13.815% 55.099%
0.891% 82.757% 61.688%
0.427% 82.757% 41.364%
2.079% 15.827% 54.300%
0.219% 15.827% 44.084%
3.630% 15.474% 52.196%
2.796% 15.474% 46.257%
1.991% 18.347% 46.847%
1.454% 18.347% 40.513%
0.058%
4.937% 28.221%
0.233%
4.937% 49.966%
5.848% 40.102%
0.269%
5.848% 23.780%
0.090%
0.007%
7.377% 33.501%
4.438%
7.377% 75.516%
14.610% 31.627% 27.727% 20.097% 31.673% 42.901%
More detailed results are available in Table 3.
In the case k = 1 (respectively, k = 10), the geometric
mean of the improvement factor of Bcm is 23 (resp. 6) times
smaller than Olh and 54 times smaller than Ocl (resp. 17).
Moreover we highlight that the new algorithm outperforms
all the competitors in each single graph, both with k = 1
and with k = 10.
We have also tested our algorithm on the three unweighted
graphs analyzed in [17], respectively called Web, Wiki, and
DBLP. By using a single thread implementation of Bcm, in
the Web graph (resp. DBLP) we computed the top-10 nodes
in 10 minutes (resp. 10 minutes) on the whole graph, having
875 713 nodes (resp. 1 305 444), while Olh needed about 25
minutes (resp. 4 hours) for a subgraph of 400 000 nodes. The
most striking result deals with Wiki, where Bcm needed 30
seconds for the whole graph having 2 394 385 nodes instead
of about 15 minutes on a subgraph with 1 million nodes.
Using multiple threads our performances are even better, as
we will show in Section 3.3.
3.2 Real-World Large Networks
In this section, we will run our algorithm on a bigger dataset,
composed by 25 directed and 15 undirected networks, with
up to 7 414 768 nodes and 191 606 827 edges. Once again, we
will consider the number of visited arcs by Bcm, i.e. mvis,
but this time we will analyze the performance ratio mvis
mn
instead of the improvement factor. Indeed, due to the large
size of these networks, the textbook algorithm did not finish
in a reasonable time.
It is worth observing that we have been able to compute
for the first time the k most central nodes of networks with
millions of nodes and hundreds of millions of arcs, with k = 1
and k = 10. The detailed results are shown in Table 5, where
for each network we have reported the performance ratio,
Table 4: the arithmetic and geometric mean of the perfor-
mance ratios (percentage).
Arithmetic Mean
Geometric Mean
k
1
10
Dir
Undir
Both
2.89% 1.14% 2.24% 0.36% 0.12% 0.24%
3.82% 1.83% 3.07% 0.89% 0.84% 0.87%
Undir
Both
Dir
both for k = 1 and k = 10. A summary of these results is
provided by Table 4, providing the arithmetic and geometric
means.
First of all, we note that the values obtained are impres-
sive: the geometric mean is always below 1%, and for k = 1
it is even smaller. The arithmetic mean is slightly bigger,
mainly because of amazon product-co-purchasing networks,
two web networks and one collaboration network, where the
performance ratio is quite high. Most of the other networks
have a very low performance ratio: with k = 1, 65% of the
networks are below 1%, and 32.5% of the networks are be-
low 0.1%. With k = 10, 52.5% of the networks are below
1% and 12.5% are below 0.1%.
We also outline that in some cases the performance ratio is
even smaller: a striking example is com-Orkut, where our
algorithm for k = 1 is more than 40 000 times faster than
the textbook algorithm, whose performance is m · n, because
the graph is connected.
3.3 Multi-Thread Experiments
In this section, we will test the performance of a parallel
version of Bcm (see Section 2.1).
In particular, we have
considered the ratio between the time needed to compute
the most central node with one thread and with x threads,
where x ∈ {1, 2, 4, 8, 16}. This ratio is plotted in Figure 1 for
k = 1 (for k = 10 very similar results hold). Ideally, the ratio
should be very close to the number of threads; however, due
Table 5: performance ratio of the new algorithm.
Directed Networks
Performance Ratio (%)
Network
cit-HepTh
cit-HepPh
p2p-Gnutella31
soc-Epinions1
soc-sign-Slashdot081106
soc-Slashdot0811
twitter-combined
soc-sign-Slashdot090216
soc-sign-Slashdot090221
soc-Slashdot0902
gplus-combined
amazon0302
email-EuAll
web-Stanford
web-NotreDame
amazon0312
amazon0601
amazon0505
web-BerkStan
web-Google
in-2004
soc-pokec-relationships
wiki-Talk
indochina-2004
Nodes
27770
34546
62586
75879
77350
77360
81306
81867
82140
82168
107614
262111
265214
281903
325729
400727
403394
410236
685230
875713
1382870
1632803
2394385
7414768
Edges
352768
421534
147892
508837
516575
828161
1768135
545671
549202
870161
13673453
1234877
418956
2312497
1469679
3200440
3387388
3356824
7600595
5105039
16539643
30622564
5021410
191606827
k = 1
2.31996
1.21083
0.73753
0.16844
0.59535
0.01627
0.76594
0.5375
0.54833
0.01662
0.36511
10.16028
0.00192
6.55454
0.04592
8.17931
7.80459
7.81823
21.0286
0.13662
1.81649
0.00411
0.00029
1.341
k = 10
4.65612
1.64227
2.51074
1.92346
0.65012
0.48842
1.03332
0.58774
0.5995
0.76048
0.3896
11.90729
0.00764
10.67736
0.62945
9.40879
9.33853
9.11571
23.70427
0.23239
2.51671
0.02257
0.00247
2.662
Undirected Networks
Performance Ratio (%)
Network
ca-HepPh
ca-AstroPh
ca-CondMat
dblp-conf2015-net-bigcomp
email-Enron
loc-gowalla-edges
com-dblp.ungraph
com-amazon.ungraph
com-youtube.ungraph
dblp22015-net-bigcomp
as-skitter
com-orkut.ungraph
Nodes
12008
18772
23133
31951
36692
196591
317080
334863
1134890
1305444
1696415
3072441
Edges
118489
198050
93439
95084
183831
950327
1049866
925872
2987624
6108712
11095298
117185083
k = 1
9.41901
1.35832
0.23165
2.46188
0.10452
0.00342
0.20647
2.55046
0.04487
0.01618
0.54523
0.00241
k = 10
9.57862
3.2326
1.07725
3.42476
0.28912
3.04066
0.312
3.08037
0.60811
0.0542
0.61078
0.38956
to memory access, the actual ratio is smaller. For graphs
where the performance ratio is small,
like in the case of
wiki-Talk (see Table 5 in the appendix), the running time
is mostly consumed by the preprocessing phase (even if it is
O(n log n)); in these cases, we observe that there seems to
be no room for parallelization improvements. On the other
hand, when the computation is more time consuming, like
in the case of web-google or as-skitter, the parallelization
is very efficient and close to the optimum.
We have also tested if the positive race condition in the
update of xk affects performances (see Section 2.1), by con-
sidering the performance ratio with a varying number of
threads. In Table 6, we report how much the performance
ratio increases if the algorithm is run with 16 threads in-
stead of 1 (more formally, if pi is the performance ratio with
i threads, we report p16−p1
). For instance, a value of 100%
means that the number of visited arcs has doubled: ideally,
this should result in a factor 8 speedup, instead of a factor
16 (that is, the number of threads). We observe that these
values are very small, especially for k = 10, where all values
except one are below 5%. This means that, ideally, instead
of a factor 16 improvement, we obtain a factor 15.24. The
only case where this is not verified is wiki-Talk, where in
any case the performance ratio is very small (see Table 5).
p1
Table 6: the increase in performance ratio from 1 to 16
threads.
Network
MathSciNet
com-dblp
ydata-v1-0
wiki-Talk
web-Google
com-youtube
dblp22015
as-skitter
in-2004
soc-pokec-relationships
imdb
k = 1
3.88%
1.32%
1.40%
k = 10
0.18%
0.74%
0.24%
192.84% 49.40%
0.95%
0.10%
0.47%
0.18%
0.10%
4.51%
4.03%
1.70%
1.91%
2.54%
0.14%
0.09%
17.71%
5.61%
IMDB CASE STUDY
4.
In this section, we will apply the new algorithm Bcm
to analyze the IMDB graph, where nodes are actors,
and two actors are connected if they played together in
a movie (TV-series are ignored).
The data collected
come from the website http://www.imdb.com:
in line with
http://oracleofbacon.org, we decided to exclude some
genres from our database: awards-shows, documentaries,
game-shows, news, realities and talk-shows. We analyzed
t
n
e
m
e
v
o
r
p
m
i
e
c
n
a
m
r
o
f
r
e
P
t
n
e
m
e
v
o
r
p
m
i
e
c
n
a
m
r
o
f
r
e
P
15
10
5
0
15
10
5
0
0
0
web-google
in-2004
soc-pokec-relationships
wiki-Talk
10
5
Number of threads
(a) Directed networks.
15
com-dblp
MathSciNet
ydata-v1_0
imdb
com-youtube
dblp22015
as-skitter
10
5
Number of threads
(b) Undirected networks.
15
Figure 1: the benefits of parallelization.
snapshots of the actor graph, taken every 5 years from 1940
to 2010, and 2014. The total time needed to perform the
computation was 37 minutes with 30 threads, and the re-
sults are reported in Table 7.
The Algorithm. The results outline that the performance
ratio decreased drastically while the graph size increased:
this suggests that the performances of our algorithm increase
with the input size. This is even more visible from Figure 2,
where we have plotted the inverse of the performance ratio
with respect to the number of nodes. It is clear that the plot
is very close to a line, especially if we exclude the last two
values. This means that the performance ratio is close to
1
cn , where c is the slope of the line in the plot, and the total
running time is well approximated by 1
cn O(mn) = O(m).
This means that, in practice, for the IMDB actor graph, our
algorithm is linear in the input size.
·104
8
6
4
2
0
o
i
t
a
r
e
c
n
a
m
r
o
f
r
e
p
f
o
e
s
r
e
v
n
I
0
0.5
1
1.5
Millions of nodes
Figure 2: growth of performance ratio with respect to the
number of nodes.
The Results. In 2014, the most central actor is Michael
Madsen, whose career spans 25 years and more than 170
films. Among his most famous appearances, he played as
Jimmy Lennox in Thelma & Louise (Ridley Scott, 1991),
as Glen Greenwood in Free Willy (Simon Wincer, 1993), as
Bob in Sin City (Frank Miller, Robert Rodriguez, Quentin
Tarantino), and as Deadly Viper Budd in Kill Bill (Quentin
Tarantino, 2003-2004). It is worth noting that he played in
movies of very different kinds, and consequently he could
“reach” many actors in a small amount of steps. The sec-
ond is Danny Trejo, whose most famous movies are Heat
(Michael Mann, 1995), where he played as Trejo, Machete
(Ethan Maniquis, Robert Rodriguez, 2010) and Machete
Kills (Robert Rodriguez, 2013), where he played as Mathete.
The third “actor” is not really an actor: he is the German
dictator Adolf Hitler: he was also the most central actor in
2005 and 2010, and he was in the top-10 since 1990. This a
consequence of his appearances in several archive footages,
that were re-used in several movies (he counts 775 credits,
even if most of them are in documentaries or TV-shows,
that were eliminated). Among the movies he is credited
in, we find Zelig (Woody Allen, 1983), and The Imitation
Game (Morten Tyldum, 2014): obviously, in both movies,
he played himself.
Among the other most central actors, we find many people
who played a lot of movies, and most of them are quite im-
portant actors. However, this ranking does not discriminate
between important roles and marginal roles:
for instance,
the actress Bess Flowers is not widely known, because she
rarely played significant roles, but she appeared in over 700
movies in her 41 years career, and for this reason she was
the most central for 30 years, between 1950 and 1980. Fi-
nally, it is worth noting that we never find Kevin Bacon in
the top 10, even if he became famous for the “Six Degrees
of Kevin Bacon” game http://oracleofbacon.org, where
the player receives an actor x, and he has to find a path of
length at most 6 from x to Kevin Bacon in the actor graph.
Kevin Bacon was chosen as the goal because he played in
several movies, and he was thought to be one of the most
central actors: this work shows that, actually, he is quite
[15] Nan Lin. Foundations of social research. McGraw-Hill,
1976.
[16] Kazuya Okamoto, Wei Chen, and XY Li. Ranking of
closeness centrality for large-scale social networks.
Frontiers in Algorithmics, 5059:186–195, 2008.
[17] Paul W. Olsen, Alan G. Labouseur, and Jeong-Hyon
Hwang. Efficient top-k closeness centrality search.
Proceedings of the IEEE 30th International Conference
on Data Engineering (ICDE), pages 196–207, 2014.
[18] Ahmet E. Sar A´syuce, Kamer Kaya, Erik Saule, and
Umit V. Catalyurek. Incremental algorithms for
closeness centrality. In IEEE International Conference
on BigData, pages 118–122, 2013.
[19] Jeremy G. Siek, Lie Quan Lee, and Andrew
Lumsdaine. Boost Graph Library: User Guide and
Reference Manual, The. Pearson Education, 2001.
[20] William Stein and David Joyner. Sage: System for
algebra and geometry experimentation. SIGSAM
Bulletin, 39(2):61–64, 2005.
[21] Stanley Wasserman and Katherine Faust. Social
Network Analysis: Methods and Applications.
Structural Analysis in the Social Sciences. Cambridge
University Press, 1994.
[22] Virginia V. Williams. Multiplying Matrices Faster
Than Coppersmith-Winograd. In Proceedings of the
Forty-fourth Annual ACM Symposium on Theory of
Computing, STOC ’12, pages 887–898, New York, NY,
USA, 2012. ACM.
[23] Virginia V. Williams and Ryan Williams. Subcubic
Equivalences between Path, Matrix and Triangle
Problems. pages 645–654. Ieee, October 2010.
far from being in the top 10. Indeed, his closeness central-
ity is 0.336, while the most central actor, Michael Madsen,
has centrality 0.354, and the 10th actor, Christopher Lee,
has centrality 0.350. We have run again the algorithm with
k = 100, and we have seen that the 100th actor is Rip Torn,
with closeness centrality 0.341.
5. REFERENCES
[1] Amir Abboud, Fabrizio Grandoni, and Virginia V.
Williams. Subcubic equivalences between graph
centrality problems, APSP and diameter. pages
1681–1697, 2015.
[2] Alez Bavelas. Communication patterns in
task-oriented groups. Journal of the Acoustical Society
of America, 22:725, 1950.
[3] Paolo Boldi and Sebastiano Vigna. Axioms for
Centrality. Internet Mathematics, 10(3-4):222–262,
2014.
[4] Michele Borassi, Pierluigi Crescenzi, and Michel
Habib. Into the Square On the Complexity of
Quadratic-Time Solvable Problems. arXiv:1407.4972,
2014.
[5] Duanbing Chen, Linyuan Lu, Ming-Sheng Shang,
Yi-Cheng Zhang, and Tao Zhou. Identifying influential
nodes in complex networks. Physica A: Statistical
Mechanics and its Applications, 391(4):1777–1787,
February 2012.
[6] Edith Cohen. All-distances sketches, revisited: HIP
estimators for massive graphs analysis. In Proceedings
of the 33rd ACM SIGMOD-SIGACT-SIGART
symposium on Principles of database systems, pages
88–99. ACM, 2014.
[7] Edith Cohen, Daniel Delling, Thomas Pajor, and
RF Werneck. Computing classic closeness centrality,
at scale. In Proceedings of the second ACM conference
on Online social networks, pages 1–13. ACM, 2014.
[8] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and
C. Stein. Introduction to Algorithms (3rd edition).
MIT Press, 2009.
[9] G´abor Cs´ardi and Tam´as Nepusz. The igraph software
package for complex network research. InterJournal,
Complex Systems, 2006.
[10] David Eppstein and Joseph Wang. Fast
Approximation of Centrality. J. Graph Algorithms
Appl., pages 1–2, 2004.
[11] Aric A. Hagberg, Daniel A. Schult, and Pieter J.
Swart. Exploring network structure, dynamics, and
function using NetworkX. In Proceedings of the 7th
Python in Science Conference (SciPy), number SciPy,
pages 11–15, 2008.
[12] U Kang, Spiros Papadimitriou, Jimeng Sun, and Tong
Hanghang. Centralities in large networks: Algorithms
and observations. In SDM, pages 119–130, 2011.
[13] Erwan Le Merrer, Nicolas Le Scouarnec, and Gilles
Tr´edan. Heuristical Top-k: Fast Estimation of
Centralities in Complex Networks. Information
Processing Letters, 114:432–436, 2014.
[14] Yeon-sup Lim, Daniel S. Menasch´e, Bruno Ribeiro,
Don Towsley, and Prithwish Basu. Online estimating
the k central nodes of a network. In Proceedings of the
2011 IEEE Network Science Workshop, number
January 1993, pages 118–122, 2011.
Year
Nodes
Edges
Perf. ratio
1st
2nd
3rd
4th
5th
6th
7th
8th
9th
10th
Year
Nodes
Edges
Perf. ratio
1st
2nd
3rd
4th
5th
6th
7th
8th
9th
10th
Year
Nodes
Edges
Perf. ratio
1st
2nd
3rd
4th
5th
6th
7th
8th
9th
10th
Year
Nodes
Edges
Perf. ratio
1st
2nd
3rd
4th
5th
6th
7th
8th
9th
10th
Table 7: detailed results of the IMDB actor graph.
1940
69 011
3 417 144
5.62%
Semels, Harry (I)
Corrado, Gino
Steers, Larry
Bracey, Sidney
Lucas, Wilfred
White, Leo (I)
Martell, Alphonse
Conti, Albert (I)
Flowers, Bess
Sedan, Rolfe
1960
146 253
11 197 509
2.21%
Flowers, Bess
Harris, Sam (II)
Farnum, Franklyn
Miller, Harold (I)
Chefe, Jack
Holmes, Stuart
Steers, Larry
Par`ıs, Manuel
1945
83 068
5 160 584
4.43%
Corrado, Gino
Steers, Larry
Flowers, Bess
Semels, Harry (I)
White, Leo (I)
Mortimer, Edmund
Boteler, Wade
Phelps, Lee (I)
Ring, Cyril
Bracey, Sidney
1965
174 826
12 649 114
1.60%
Flowers, Bess
Harris, Sam (II)
Farnum, Franklyn
Miller, Harold (I)
Holmes, Stuart
Sayre, Jeffrey
Chefe, Jack
Par`ıs, Manuel
O’Brien, William H.
O’Brien, William H.
Sayre, Jeffrey
Stevens, Bert (I)
1950
97 824
6 793 184
4.03%
Flowers, Bess
Steers, Larry
Corrado, Gino
Harris, Sam (II)
Semels, Harry (I)
Davis, George (I)
Magrill, George
Phelps, Lee (I)
Ring, Cyril
Moorhouse, Bert
1970
210 527
14 209 908
1.14%
1955
120 430
8 674 159
2.91%
Flowers, Bess
Harris, Sam (II)
Steers, Larry
Corrado, Gino
Miller, Harold (I)
Farnum, Franklyn
Magrill, George
Conaty, James
Davis, George (I)
Cording, Harry
1975
257 896
16 080 065
0.83%
Flowers, Bess
Harris, Sam (II)
Tamiroff, Akim
Farnum, Franklyn
Miller, Harold (I)
Sayre, Jeffrey
Quinn, Anthony (I)
O’Brien, William H.
Holmes, Stuart
Stevens, Bert (I)
Flowers, Bess
Harris, Sam (II)
Tamiroff, Akim
Welles, Orson
Sayre, Jeffrey
Miller, Harold (I)
Farnum, Franklyn
Kemp, Kenner G.
Quinn, Anthony (I)
O’Brien, William H.
1980
310 278
18 252 462
0.62%
Flowers, Bess
Harris, Sam (II)
Welles, Orson
Sayre, Jeffrey
Quinn, Anthony (I)
Tamiroff, Akim
Miller, Harold (I)
Kemp, Kenner G.
Farnum, Franklyn
Niven, David (I)
2000
681 358
33 564 142
0.22%
Lee, Christopher (I)
Hitler, Adolf
Pleasence, Donald
1985
375 322
20 970 510
0.45%
Welles, Orson
Flowers, Bess
Harris, Sam (II)
1990
463 078
24 573 288
0.34%
Welles, Orson
Carradine, John
Flowers, Bess
Quinn, Anthony (I)
Lee, Christopher (I)
Sayre, Jeffrey
Carradine, John
Kemp, Kenner G.
Miller, Harold (I)
Niven, David (I)
Tamiroff, Akim
Harris, Sam (II)
Quinn, Anthony (I)
Pleasence, Donald
Sayre, Jeffrey
Tovey, Arthur
Hitler, Adolf
1995
557 373
28 542 684
0.26%
Lee, Christopher (I)
Welles, Orson
Quinn, Anthony (I)
Pleasence, Donald
Hitler, Adolf
Carradine, John
Flowers, Bess
Mitchum, Robert
Harris, Sam (II)
Sayre, Jeffrey
2005
880 032
41 079 259
0.18%
2010
1 237 879
53 625 608
0.19%
2014
1 797 446
72 880 156
0.14%
Hitler, Adolf
Hitler, Adolf
Madsen, Michael (I)
Lee, Christopher (I)
Lee, Christopher (I)
Steiger, Rod
Welles, Orson
Sutherland, Donald (I)
Quinn, Anthony (I)
Steiger, Rod
Carradine, John
Sutherland, Donald (I)
Mitchum, Robert
Connery, Sean
Pleasence, Donald
Hopper, Dennis
Keitel, Harvey (I)
von Sydow, Max (I)
Caine, Michael (I)
Sheen, Martin
Hopper, Dennis
Keitel, Harvey (I)
Carradine, David
Sutherland, Donald (I)
Dafoe, Willem
Caine, Michael (I)
Sheen, Martin
Kier, Udo
Trejo, Danny
Hitler, Adolf
Roberts, Eric (I)
De Niro, Robert
Dafoe, Willem
Jackson, Samuel L.
Keitel, Harvey (I)
Carradine, David
Lee, Christopher (I)
|
1710.06339 | 1 | 1710 | 2017-10-17T15:14:44 | Understanding the Correlation Gap for Matchings | [
"cs.DS"
] | Given a set of vertices $V$ with $|V| = n$, a weight vector $w \in (\mathbb{R}^+ \cup \{ 0 \})^{\binom{V}{2}}$, and a probability vector $x \in [0, 1]^{\binom{V}{2}}$ in the matching polytope, we study the quantity $\frac{E_{G}[ \nu_w(G)]}{\sum_{(u, v) \in \binom{V}{2}} w_{u, v} x_{u, v}}$ where $G$ is a random graph where each edge $e$ with weight $w_e$ appears with probability $x_e$ independently, and let $\nu_w(G)$ denotes the weight of the maximum matching of $G$. This quantity is closely related to correlation gap and contention resolution schemes, which are important tools in the design of approximation algorithms, algorithmic game theory, and stochastic optimization.
We provide lower bounds for the above quantity for general and bipartite graphs, and for weighted and unweighted settings. he best known upper bound is $0.54$ by Karp and Sipser, and the best lower bound is $0.4$. We show that it is at least $0.47$ for unweighted bipartite graphs, at least $0.45$ for weighted bipartite graphs, and at lea st $0.43$ for weighted general graphs. To achieve our results, we construct local distribution schemes on the dual which may be of independent interest. | cs.DS | cs |
Understanding the Correlation Gap for Matchings
Guru Guruganesh∗
Euiwoong Lee†
Computer Science Department
Carnegie Mellon University
Pittsburgh, PA 15213.
Given a set of vertices V with V = n, a weight vector w ∈ (R+ ∪ {0})(V
2 ), and a probability
vector x ∈ [0, 1](V
2 ) in the matching polytope, we study the quantity
Abstract
EG[νw(G)]
∑
(u,v)∈(V
2 ) wu,vxu,v
where G is a random graph where each edge e with weight we appears with probability xe in-
dependently, and let νw(G) denotes the weight of the maximum matching of G. This quantity
is closely related to correlation gap and contention resolution schemes, which are important
tools in the design of approximation algorithms, algorithmic game theory, and stochastic opti-
mization.
We provide lower bounds for the above quantity for general and bipartite graphs, and for
weighted and unweighted settings. The best known upper bound is 0.54 by Karp and Sipser,
and the best lower bound is 0.4. We show that it is at least 0.47 for unweighted bipartite graphs,
at least 0.45 for weighted bipartite graphs, and at least 0.43 for weighted general graphs. To
achieve our results, we construct local distribution schemes on the dual which may be of inde-
pendent interest.
1 Introduction
2 ) and the weight vector w ∈ (R+ ∪ {0})(V
We study the size (weight) of the maximum matching of a random graph sampled from various
random graph models. Let V be the set of vertices with V = n. Given the probability vector x ∈
[0, 1](V
n,w,x be the distribution of random graphs
with n vertices such that each pair e ∈ (V
2 ) becomes an edge with probability xe independently. If
it becomes an edge, its weight is we. For bipartite graphs, let V1 and V2 be the set of left and right
vertices with V1 = V2 = n. Given the probability vector x ∈ [0, 1]V1 ×V2 and the weight vector
w ∈ (R+ ∪ {0})V1 ×V2, let DB
n,w,x be the distribution of random bipartite graphs with 2n vertices
such that each pair e ∈ V1 × V2 becomes an edge with probability xe independently. If it becomes
an edge, its weight is we. We use DB
2 ), let DG
n,x (resp. DG
n,x) for the unweighted case (w = (1, 1, . . . , 1)).
We focus on the case when the probability vector x is in the matching polytope of the complete
(bipartite) graph. Recall that for bipartite graphs, x ∈ [0, 1]V1 ×V2 is in the matching polytope if
∗Supported in part by Anupam Gupta's NSF awards CCF-1319811 and CCF-1536002. [email protected]
†Supported by a Samsung Fellowship and Venkat Guruswami's NSF CCF-1526092. [email protected]
1
each v ∈ V1 ∪ V2 satisfies ∑u xu,v 6 1. For general graphs, x ∈ [0, 1](V
2 ) is in the matching polytope
if each v ∈ V satisfies ∑u xu,v 6 1 and each odd set S ⊆ V satisfies ∑{u,v}⊆S xu,v 6 ⌊(S − 1)/2⌋.1
Given a weighted graph G, let νw(G) be the weight of the maximum weight matching of G.
If G is unweighted, ν(G) denotes the cardinality of the maximum matching of G. For any x ∈
[0, 1](V
2 ) wu,vxu,v, simply because
the probability that (u, v) is included in the maximum matching is at most xu,v. The analogous
statement also holds for bipartite graphs.
2 ) and w ∈ (R+ ∪ {0})(V
[νw(G)] 6 ∑(u,v)∈(V
2 ), we have E
G∼DG
n,w,x
If x is in the matching polytope2, we can prove that EG[νw(G)] > κ · ∑ wu,vxu,v for some con-
stant 0 < κ < 1 . For the general graph model, κ is known to be at least (1 − 1/e)2 ∼ 0.40 for
every w [6]. For the bipartite graph model, κ is known to be at least 0.4 for every w [5]. Karp
and Sipser [11] showed an upper bound of 0.54 for both bipartite and general graphs, by demon-
strating it for the unweighted models where every edge appears with equal probability. Our main
results are the following improved lower bounds on κ. Our first theorem concerns the unweighted
bipartite model.
Theorem 1.1. Let V1 = V2 = n and x ∈ [0, 1]V1 ×V2 be in the matching polytope of the complete
bipartite graph on V1 ∪ V2. Then
E
G∼DB
n,x
[ν(G)]
∑(u,v)∈V1×V2 xu,v
> 0.476.
(1)
We also obtain a slightly weaker result on the weighted bipartite model.
Theorem 1.2. Let V1 = V2 = n and x ∈ [0, 1]V1 ×V2 be in the matching polytope of the complete
bipartite graph on V1 ∪ V2. Then for any w ∈ (R+ ∪ {0})V1 ×V2,
E
G∼DB
n,w,x
[νw(G)]
∑(u,v)∈V1×V2 wu,vxu,v
> (cid:18)1 −
3
2e(cid:19) > 0.4481.
Finally, we prove an improved bound on the weighted general graph model.
Theorem 1.3. Let V = n and x ∈ [0, 1](V
Then for any w ∈ (R+ ∪ {0})(V
2 ),
2 ) be in the matching polytope of the complete graph on V1 ∪ V2.
E
n,w,x
G∼DG
∑(u,v)∈(V
[νw(G)]
2 ) wu,vxu,v
>
e2 − 1
2e2
> 0.4323.
1.1 Applications and Related Work.
Contention Resolution Schemes and Correlation Gap. Our work is inspired by and related to
the rounding algorithms studied in approximation algorithms. Given a downward-closed family
I ⊆ 2E defined over a ground-set E and a submodular function f : 2E → R+, Chekuri et al. [5]
considered the problem of finding maxS∈I f (S) and introduced contention resolution schemes (CR
schemes) to obtain improved approximation algorithms for numerous problems. Let PI be the
1Our result for general graphs, Theorem 1.3 holds even when x satisfies the first type of constraints.
2If x is not in the matching polytope, one can construct examples where κ = Ω(n).
2
convex combination of all incidence vectors {1S}S∈I . A c-CR scheme π for x ∈ PI is a procedure
that, when R is a random subset of E with e ∈ R independently with probability xe, returns
π(R) ⊆ R such that π(R) ∈ I with probability 1 and Pr[e ∈ π(R)] > c for all e ∈ E.
To construct a CR scheme, they introduced the notion of correlation gap of a polytope, inspired
by [1].3 Formally, the correlation gap of I is defined as
κ(I) := inf
x∈PI , w>0
ER∼Dx [maxS⊆R,S∈I ∑e∈S we]
∑e∈E xewe
,
(2)
where Dx is the distribution where each element e appears in R with probability xe independently.
It is easy to see that the existence of c-CR scheme for all x ∈ PI implies κ(I) > c. Chekuri et al. [5]
proved the converse that every x ∈ PI admits a κ(I)-CR scheme.
By setting E to be the set of all possible edges of a complete (bipartite) graph, and I to be
the set of all matchings of a complete graph, our Theorem 1.2 and Theorem 1.3 for weighted
bipartite graphs and weighted general graphs imply that there exist 0.4481-CR scheme and 0.4323-
CR scheme for bipartite matching polytopes and general matching polytopes respectively. Note
that these lower bounds hold when E′ is the set of edges and I ′ is a matching polytope of an
arbitrary graph G′, since
κ(I) = inf
x∈PI , w>0
ER∼Dx [maxS⊆R,S∈I ∑e∈S we]
∑e∈E xewe
6
inf
xE′ ∈PI′ , wE′ =0
ER∼Dx [maxS⊆R,S∈I ∑e∈S we]
∑e∈E xewe
= κ(I ′).
Maximum Matching of Random Graphs. The study of maximum matchings in random graphs
has a long history. It was pioneered by the work of Erdos and R´enyi [7, 8], where they proved that
a random graph Gn,p has a perfect matching with high probability when p = Ω( ln n
n ). The case for
sparse graphs was investigated by Karp and Sipser [11] who gave an accurate estimate of ν(G) for
Gn,p where p = c
n−1 for some constant c > 0.
After these two pioneering results, subsequent work has addressed two aspects. The Karp-
Sipser algorithm is a simple randomized greedy algorithm, and the first line of works extend
the range of models where this algorithm (or its variants) returns an almost maximum matching.
Aronson et al. [2] and Chebolu et al. [4] augmented the Karp-Sipser algorithm to achieve tighter
results in the standard Gn,p model. Bohman and Frieze [3] considered a new model where a graph
is drawn uniformly at random from the collection of graphs with a fixed degree sequence and
gave a sufficient condition where the Karp-Sipser algorithm finds an almost perfect matching.
The second line of work is based on the following observation: the standard Gn,p model,
p = Ω( ln n
n ) is required to have a perfect matching, because otherwise there will be an isolated
vertex. This naturally led to the question of finding a natural and sparser random graph model
with a perfect matching. The considered models include a random regular graph, and a Gn,p with
prescribed minimal degree. We refer the reader to the work of Frieze and Pittel [10] and Frieze [9]
and references therein.
1.2 Organization
Our main technical contribution is lower bounding correlation gaps via local distribution schemes
for dual variables, which are used to prove Theorem 1.1 and Theorem 1.2 for unweighted and
3[1] defined the correlation gap of a set function f
: 2E → R+. Our results apply to this definition too when f
denotes the weight of the maximum matching.
3
weighted bipartite graphs. We present this framework in Section 2 and prove our bounds for
unweighted bipartite graphs (Section 3) and weighted bipartite graphs (Section 4). Our result for
weighted general graphs is presented in Section 5.
2 Techinques for Bipartite Graphs
Let V = V1 ∪ V2 be the set of vertices with V1 = V2 = n, E := V1 × V2. Fix w ∈ (R+ ∪ {0})E and
x ∈ [0, 1]E in the bipartite matching polytope of (V, E).
Our proofs for Theorem 1.1 and 1.2 for bipartite graphs follow the following general frame-
work. Let G = (V, E(G)) be a sampled from the distribution where each potential edge e ∈ E
appears with probability xe independently (recall that E = V1 × V2 is the set of all potential edges
and E(G) is the edges of one sample G). Let y(G) ∈ (R+ ∪ {0})V be an optimal fractional vertex
cover such that for every e = (u, v) ∈ E(G), yu(G) + yv(G) > we. By K onig-Egerv´ary theorem,
ky(G)k1 = ν(G).
Given G, consider the situation where initially each vertex v has mass yv(G), and each potential
edge has mass ye(G) = 0 (we slightly abuse notation and consider y(G) ∈ (R+ ∪ {0})V∪E). We
construct local distribution schemes FG : (V ∪ E) × (V ∪ E) → R where FG(a, b) indicates the amount
of mass sent from a to b. We require that FG(a, a) = 0, but we allow FG(a, b) 6= −FG(b, a) for a 6= b
(the net flow from a to b in this case is FG(a, b) − FG(b, a)). Let t(G) ∈ RV∪E denote the mass of
each vertex and edge after the distribution.
ta(G) := ya(G) + ∑
FG(b, a) − ∑
FG(a, b).
b∈V∪E
b∈V∪E
We choose FG so that it ensures tv(G) > 0 for every v ∈ V. This implies
∑
e∈E
te(G) 6 ∑
ta(G) = ∑
a∈V∪E
a∈V∪E
ya(G) = ∑
v∈V
yv(G) = ν(G).
Therefore, if we prove that for each potential edge e ∈ E
EG[te(G)] > α · wexe,
(3)
for some α > 0, it implies that
EG[ν(G)] > α · ∑
e∈E
EG[te(G)] > α · ∑
e∈E
wexe.
For weighted and unweighted cases, we construct different local distribution schemes {FG}G that
prove (3) with different values of α.
Weighted Bipartite Graphs. Given a sample G = (V, E(G)) and a fractional vertex cover y ∈
(R+ ∪ {0})V, our FG(v, e) = yv(G)/ degG(v) if e ∈ E(G) is an edge incident on v ∈ V, and 0
otherwise. Intuitively, each vertex v distributes its mass yv(G) evenly to its incident edges in G.
This clearly satisfies tv(G) > 0 for every v ∈ V, and for each e = (u, v) ∈ E, we use the following
4
approximation:
EG[te(G)] = Pr[e ∈ G] · EG (cid:20)te(G)e ∈ G(cid:21)
e ∈ G(cid:21)
1
yv(G)
degG(v)
+
= xe EG (cid:20) yu(G)
degG(u)
> xe EG (cid:20)(yu(G) + yv(G))
> xewe EG (cid:20)
1
max(degG(u), degG(v))
e ∈ G(cid:21).
max(degG(u), degG(v))
e ∈ G(cid:21)
Therefore, to prove Theorem 1.2, it suffices to prove that for every potential edge e ∈ E,
E
n,w,x (cid:20)
G∼DB
1
max(degG(u), degG(v))
e ∈ G(cid:21) > 0.4481,
when G is sampled from DB
n,w,x with x in the matching polytope. Experimentally trying several
extreme cases indicates that the worst case for e = (u, v) ∈ E happens when xe = ε for very small
ε, u has only one other edge eu with xeu = 1 − ε, and v is incident on n − 1 edges ev1, . . . , evn−1 with
] as n grows,
xevi
1
n−1 ). Section 4 formally proves that this
where YU is drawn from a binomial distribution B(n − 1,
is indeed the worst case.
max(degG(u),degG (v)) e ∈ G] converges to E[
n−1. As ε approaches to 0, EG[
= 1−ε
1+YU
1
1
Unweighted Bipartite Graphs. One simple but important observation is that in the above ex-
]xe, e is an edge with very small xe = ε, and it is adjacent to a
ample where EG[te(G)] ≈ E[
large edge eu with xeu = 1 − ε. From the persepctive of xeu, the expected number of adjacent edges
is at most 2ε, so EG[teu (G)] ≈ xeu ≈ 1. Since eu gets much more than what it needs (E[teu ] > 0.476
suffices to prove Theorem 1.1), it is natural to take some value from teu(G) to increase te(G).
1+YU
1
Formally, given G = (V, E(G)), our new local distribution scheme FG : (V ∪ E) × (V ∪ E) → R
is defined as follows. Let c be an universal constant that will be determined later.
FG(a, b) =
ya(G)
degG (a)
cx2
a xb
0
if a ∈ V, b ∈ E(G), a ∈ b
if a 6= b ∈ E, a ∩ b 6= ∅
otherwise.
(4)
Intuitively, on top of the old local distribution scheme for weighted graphs, each edge e pays
cx2
e x f to every adjacent edge f with probability 1 (this quantity does not depend on G). Because
this term quadratically depends on the x value of the sender, this payment penalizes edges with
large x values to help edges with small x values. For a fixed edge e = (u, v) ∈ E with fixed xe = ε,
Theorem 3.1 shows that the worst case is when both u and v have n − 1 other edges of whose x
values are equal to 1−ε
n . Finally, Lemma 3.2 shows that E[te] > 0.476 for every ε ∈ (0, 1], proving
Theorem 1.1.
3 Unweighted Bipartite Graphs
We prove Theorem 1.1 for unweighted bipartite graphs. Given G = (V, E(G)), consider the local
distribution scheme FG : (V ∪ E) × (V ∪ E) → R given in (4). This implies that the mass after this
5
new distribution scheme for an edge e = (u, v) is given by
te(G) = αe(G) + ∑
c(xex2
f − x2
e x f ) + ∑
c(x2
gxe − x2
e xg),
f ∈E\{e}: f ∋u
g∈E\{e}:g∋v
where αe(G) := yu(G)/ degG(u) + yv(G)/ degG(v) denotes the mass after the old distribution
scheme used for weighted bipartite graphs. We define βe(x) to be the following.
βe(x) := E
G∼DB
n,x
= E
G∼DB
n,x
[te(G)]
[αe(G)] + ∑
c(xex2
f − x2
e x f ) + ∑
c(x2
gxe − x2
e xg)
f ∈E\{e}: f ∋u
g∈E\{e}:g∋v
To prove Theorem 1.1, it suffices to prove that βe(x) > 0.476xe for each e. Fix e = (u, v).
Let eu1, . . . , eun−1 be n − 1 other edges incident on u and ev1, . . . , evn−1 be n − 1 other edges inci-
max(degG(u),degG(v)) e ∈ G] as before. Define
dent on v. E
F(x0, y1, . . . , yn−1, z1, . . . , zn−1) by
[αe(G)] is lower bounded by xe EG[
G∼DB
n,x
1
F(x0, y1, . . . , yn−1, z1, . . . , zn−1) := x0 E[
1
1 + max(Y, Z)
] +
n−1
∑
i=1
c(x0y2
i − x2
0yi)
+
n−1
∑
i=1
c(x0z2
i − x2
0zi),
, . . . , xeun−1
where Y := Y1 + · · · + Yn−1 and Z := Z1 + · · · + Zn−1 and each Yi (resp. Zi) is an indepen-
dent Bernoulli random variable with E[Yi] = yi (resp. E[Zi] = zi). By construction, βe(x) >
F(xe, xeu1
, the following theorem
shows that F is minimized when xeu1
Theorem 3.1. For x0, y1, . . . , ym, z1, . . . , zm ∈ [0, 1] where ys := ∑m
1 − x0,
and ∑n−1
i=1 xvui
= · · · = xevn−1
i=1 yi 6 1 − x0 and zs := ∑m
). Given fixed ∑n−1
i=1 xeui
and xev1
= · · · = xeun−1
, . . . , xevn−1
i=1 zi 6
, xev1
.
F(x0, y1, . . . , ym, z1, . . . , zm) > F(x0,
ys
m
, . . . ,
ys
m
,
zs
m
, . . . ,
zs
m
).
Proof. Without loss of generality, assume y1 > . . . > ym. We will show that if y1 > ym,
∂F
∂ym
−
∂F
∂y1
6 0.
(5)
This implies that as long as y1 > ym, decreasing y1 and increasing ym by the same amount will
never increase F while maintaining y1 + · · · + ym = ys, so F is minimized when y1 = · · · = ym =
ys
m . The same argument for z1, . . . , zm will prove the theorem.
Let Y := Y1 + · · · + Ym and Z := Z1 + · · · + Zm, where each Yi (resp. Zi) is an independent
Bernoulli random variable with E[Yi] = yi (resp. E[Zi] = zi). To prove (5), we first compute
6
∂ E[
1
1+max(Y,Z) ]
∂ym
−
∂ E[
1
1+max(Y,Z) ]
∂y1
. Let Y′ := Y2 + · · · + Ym−1. We decompose E[
1
1+max(Y,Z) ] as follows.
E[
1
]
1 + max(Y, Z)
m
∑
m
∑
i=0
m
∑
j=0(cid:18) Pr[Y = i] · Pr[Z = j] ·
i=0(cid:18) Pr[Y′ = i] · Pr[Z 6 i](cid:0)
i=0(cid:18) Pr[Y′ = i] · Pr[Z = i + 1](cid:0)
m
∑
=
=
+
1
1 + max(i, j)(cid:19)
(1 − y1)(1 − ym)
1 + i
+
y1(1 − ym) + (1 − y1)ym
2 + i
+
y1ym
3 + i(cid:1)(cid:19)
1 − y1ym
2 + i
+
y1ym
3 + i(cid:1)(cid:19) +
m
∑
i=0
Pr[Y′ = i] · Pr[Z > i + 2] ·
1
3 + i
Therefore, the directional derivative can be written as
(
∂
∂ym
−
) E[
1
1 + max(Y, Z)
]
∂
∂y1
m
∑
1
1 + i
m
∑
i=0(cid:18) Pr[Y′ = i] · Pr[Z 6 i](cid:0)
i=0(cid:18) Pr[Y′ = i] · Pr[Z = i + 1](cid:0) −
i=0(cid:18) Pr[Y′ = i] · Pr[Z 6 i](cid:0)
i=0(cid:18) Pr[Y′ = i] · Pr[Z 6 i](cid:0)
m
∑
m
∑
1 + i
1 + i
1
1
−
−
−
2
2 + i
+
1
2 + i
+
2
2 + i
2
2 + i
+
+
1
1
3 + i(cid:1)(cid:19)
3 + i(cid:1)(cid:19)
3 + i(cid:1)(cid:19)
3 + i(cid:1)(cid:19)
1
1
=(y1 − ym)
+(y1 − ym)
6(y1 − ym)
6(y1 − ym)
6
y1 − ym
3
,
where the last inequality follows from the fact that
1
1 + i
(cid:0)
−
2
2 + i
+
1
3 + i(cid:1) =
2
(1 + i)(2 + i)(3 + i)
6
1
3
.
Finally,
(
−
∂
∂
∂y1
∂ym
∂
∂
∂y1
∂ym
xe(y1 − ym)
−
=(
6
3
)F
)(xe E[
1
1 + max(Y, Z)
− 2cxe(y1 − ym) = 0.
] + cxey2
1 − cx2
e y1 + cxey2
m − cx2
e ym)
By taking c = 1
6 .
Therefore, for any e ∈ E, βe(x) > F(xe, ys
n−1 , . . . , ys
n−1,
zs
n−1, . . . ,
zs
n−1 ) for some ys 6 1 − xe and
7
zs 6 1 − xe. Let
G(xe, ys, zs) :=F(xe,
ys
n − 1
, . . . ,
ys
n − 1
,
zs
n − 1
, . . . ,
] + (n − 1)c(xe(
)2 − x2
e (
ys
n − 1
))
) − xe) + cxezs((
zs
n − 1
) − xe)
zs
)
n − 1
ys
n − 1
))
zs
n − 1
ys
n − 1
1
=xe E[
1 + max(Y, Z)
zs
+ (n − 1)c(xe(
)2 − x2
e (
n − 1
=xe E[
>xe E[
1
1 + max(Y, Z)
1
1 + max(Y, Z)
] + cxeys((
] − 2cx2
e
where Y ∼ Binomial(n − 1, ys
mized when ys = zs = 1 − xe. Finally, let
n−1 ), Z ∼ Binomial(n − 1,
zs
n−1 ). Note that the final quantity is mini-
Hn−1(xe) := xe E[
1
1 + max(Y, Z)
] − 2cx2
e ,
where Y, Z ∼ Binomial(n − 1, 1−xe
n−1 ).
Lemma 3.2. For any m ∈ N and xe ∈ [0, 1], Hm(xe) > 0.476xe.
Proof. Since the binomial distribution is approximated by the Poisson distribution in the limit, we
1+max(Y,Z) ] − x2/3 (we
use this to ease the calculation. Let Y, Z ∼ Poisson(1 − x). Let H(x) := x E[
substitute c = 1/6 into the earlier equation). In particular, we write the expectation in full to get
1
E[
1
1 + max(Y, Z)
] =
∞
∑
k=0
∞
∑
j=0
1
1 + max(j, k)
e−2(1−x) (1 − x)j+k
j!k!
=
1
e2(1−x)
∞
k=0(cid:16)
∑
k
∑
j=0
1
1 + max(j, k − j)
1
j!(k − j)!(cid:17)(1 − x)k
Let Pt(x) denote the above sum truncated at k = t. I.e.
Pt(x) :=
1
e2(1−x)
t
k=0(cid:16)
∑
k
∑
j=0
1
1 + max(j, k − j)
1
j!(k − j)!(cid:17)(1 − x)k
This is a degree t-polynomial in (1 − x) with a normalizing factor of e−2(1−x) and note that E[
Pt(x) for any t ∈ N.
1
1+max(Y,Z) ] >
Truncating this polynomial with t = 15, we can see that this has a minimum value of 0.476
1+max(Y,Z) ] − x/3 > P15(x) − x/3. In the interval
for all values of x ∈ [0, 1]. we can see that E[
x ∈ [0, 1], this function achieves its minimum at x = 0 achieving a minimum of 0.476.
1
4 Weighted Bipartite Graphs
We prove Theorem 1.2 for weighted bipartite graphs. As explained in Section 2, it suffices to prove
that for each e = (u, v) ∈ E,
E
n,w,x (cid:20)
G∼DB
1
max(degG(u), degG(v))
e ∈ G(cid:21) > 0.4481.
8
Fix e = (u, v) and assume V = {v, v1, . . . , vn−1} ∪ {u, u1, . . . , un−1}. Let Y = degG(u) − 1 and
Z = degG(v) − 1. Given e ∈ G, Y and Z can be represented as Y = ∑n−1
i=1 Zi,
where Yi indicates where (u, vi) ∈ E(G) and Zi indicates where (v, ui) ∈ E(G). This construction
ensures that
i=1 Yi and Z = ∑n−1
EG(cid:20)
1
max(degG(u), degG(v))
e ∈ G(cid:21) = EY,Z(cid:20)
1
1 + max(Y, Z)(cid:21).
Note that Y1, . . . , Yn−1, Z1, . . . , Zn−1 are mutually independent, and E[Y], E[Z] 6 1. By mono-
tonicity, assuming E[Y] = E[Z] = 1 never increases the lower bound. The following theorem
shows that the worst case happens when one of Y, Z is consistently 1 and the other is drawn from
Binomial(n − 1,
Theorem 4.1. Let Y = Y1 + · · · + Ym and Z = Z1 + · · · + Zm, where Y1, . . . , Ym, Z1, . . . , Zm are mutu-
ally independent Bernoulli random variables with E[Y] = E[Z] = 1. Then,
1
n−1 ).
E(cid:20)
where YU is drawn from Binomial(m, 1
Proof. We decompose E[
1
1
1 + max(Y, Z)(cid:21) > E(cid:20)
m ).
1
1 + YU(cid:21),
1+max(Y,Z) ] as follows.
E(cid:20)
1
1 + max(Y, Z)(cid:21) =
=
=
=
m
∑
i=0
m
∑
i=0
m
∑
i=0
m
∑
i=0
m
∑
j=0(cid:18) Pr[Y = i] · Pr[Z = j] ·
Pr[Y = i](cid:20)(cid:0)
Pr[Z = j](cid:1) ·
i
∑
j=0
Pr[Y = i] ·
Pr[Y = i] ·
1
1 + i
1
1 + i
−
−
m
∑
i=0
m
∑
j=1
1
1 + max(i, j)(cid:19)
1
1 + i
+(cid:0)
Pr[Y = i](cid:20) m
∑
j=i+1
Pr[Z = j](cid:20) j−1
∑
i=0
m
∑
j=i+1
Pr[Z = j] ·
1
1 + j(cid:1)(cid:21)
Pr[Z = j](cid:0)
Pr[Y = i](cid:0)
1
1 + i
1
1 + i
−
−
1
1 + j(cid:1)(cid:21)
1 + j(cid:1)(cid:21).
1
Let tj := ∑
j−1
i=0 Pr[Y = i] ·(cid:0) 1
1+i − 1
tj
j .
>
Lemma 4.2. For all j > 3, t2
2
1+j(cid:1). We prove the following facts about tj's.
Proof. Fix j > 3. By the definition of t2 and tj,
t2
2
−
tj
j
=
=
1
2(cid:18) Pr[Y = 0](1 −
1
3
) + Pr[Y = 1](
1
2
−
1
3
)(cid:19)(cid:19) −
1
j(cid:18) j−1
∑
i=0
1
3
Pr[Y = 0] +
1
12
Pr[Y = 1] −
= (
1
3
−
1
1 + j
) Pr[Y = 0] + (
1
12
−
1
j(cid:18) j−1
∑
i=0
j − 1
2j(j + 2)
Pr[Y = i](cid:0)
1
1 + i
−
) Pr[Y = 1]
Pr[Y = i](cid:0)
1 + j(cid:1)(cid:19)
1
1
1 + i
−
1
1 + j(cid:1)(cid:19)
−
1
j(cid:18) j−1
∑
i=2
> (cid:18) 1
3
−
Pr[Y = i](cid:0)
j−1
∑
i=2(cid:0)
1
j
−
1
1 + j
1
1 + i
−
1
1 + i
−
1
1 + j(cid:1)(cid:19)
1 + j(cid:1)(cid:19) Pr[Y = 0] + (
1
1
12
−
j − 1
2j(j + 2)
) Pr[Y = 1],
9
where the inequality follows from Pr[Y = 0] > Pr[Y = i] for i > 2. To prove t2
to prove that 1
1+i − 1
for j > 3. The former can be proved as
1+j(cid:1) > 0, and 1
i=2(cid:0) 1
12 − j−1
3 − 1
1+j − 1
2j(j+2)
j ∑
j−1
2 − tj
j
> 0, it suffices
> 0. It is easy to verify the latter
1
3
1
3
−
+
+
=
>
1
3
1
3
2
3j
=(cid:0)
=
−
−
1
1 + i
j(1 + j)
1 + j
j − 2
1
1 + j
−
1
j
j−1
∑
i=2(cid:0)
−(cid:0)
−(cid:0)
3j (cid:1) +(cid:0)
1
1
j − 2
j(1 + j)
j − 2
1 + j
j − 2
j(1 + j)
+
+
1
−
1 + j(cid:1)
j−1
1
∑
1 + i(cid:1)
j
i=2
j − 2
1
3j (cid:1)
1 + j(cid:1)
1
−
2
j(1 + j)
> 0,
where the first inequality follows from 1
1+i
j > 3.
6 1
3 for i > 2 and the last inequality follows from
We prove the theorem by considering the following two cases.
Case 1: 2 Pr[Y = 0] > Pr[Y = 1] or 2 Pr[Z = 0] > Pr[Z = 1]. Without loss of generality, assume
that 2 Pr[Y = 0] > Pr[Y = 1]. It is equivalent to
Pr[Y = 0] >
2
3
Pr[Y = 0] +
1
6
Pr[Y = 1]
⇔ t1 >
t2
2
.
By Lemma 4.2, it implies that t1 >
tj
j for all j > 2. Then, since E[Z] = ∑m
j=1 j · Pr[Z = j] = 1,
E(cid:20)
1
1 + max(Y, Z)(cid:21) =
>
=
m
∑
i=0
m
∑
i=0
m
∑
i=0
Pr[Y = i] ·
Pr[Y = i] ·
Pr[Y = i] ·
1
1 + i
1
1 + i
1
1 + i
−
m
∑
j=1
Pr[Z = j]tj
j · Pr[Z = j]
− t1
m
∑
j=1
− t1
= E[
1
1 + max(Y, 1)
].
The following lemma proves the theorem in the case t1 > t2
2 .
Lemma 4.3. E[
1
1+max(Y,1) ] > E[
1
1+max(YU,1) ].
Proof. Note that Y = Y1 + · · · + Ym, and each Yi is a Bernoulli random variable. Let yi := E[Yi].
Without loss of generality, assume y1 > . . . > ym. We will show that if y1 > ym,
∂ E[
1
1+max(Y,1) ]
∂ym
−
∂ E[
1
1+max(Y,1) ]
∂y1
6 0.
(6)
10
This implies that as long as y1 > ym, decreasing y1 and increasing ym by the same amount will
1+max(Y,1) ] while maintaining y1 + · · · + ym = 1, so the expectation is minimized
never increase E[
when y1 = · · · = ym, or Y = YU. Consider the following decomposition of E[
1
1
1+max(X,Y) ].
EY(cid:20)
1
1 + max(1, Y)(cid:21) = Pr[Y = 0] ·
1
2
+
m
∑
i=1
Pr[Y = i] ·
1
1 + i
=
=
=
1
2
1
2
1
2
(1 −
m
∑
i=1
Pr[Y = i]) +
m
∑
i=1
Pr[Y = i] ·
1
1 + i
−
−
m
∑
i=2
m
∑
i=2
Pr[Y = i] · (
Pr[Y > i] · (
1
2
1
i
−
−
1
1 + i
1
1 + i
)
).
To prove (6), it suffices to prove that for all i > 2,
∂ Pr[Y > i]
∂ym
−
∂ Pr[Y > i]
∂y1
> 0.
Let Y′ = Y2 + · · · + Ym−1, and fix i > 3.
Pr[Y > i] = Pr[Y′ = i − 2]y1ym + Pr[Y′ = i − 1](cid:0)y1(1 − ym) + (1 − y1)ym + y1ym)
+ Pr[Y′ > i]
∂ Pr[Y > i]
∂y1
= Pr[Y′ = i − 2]ym + Pr[Y′ = i − 1](cid:0)1 − ym(cid:1)
Therefore,
∂ Pr[Y > i]
∂ym
−
∂ Pr[Y > i]
∂y1
= Pr[Y′ = i − 2](y1 − ym) + Pr[Y′ = i − 1](cid:0)ym − y1(cid:1)
= (y1 − ym)(cid:0) Pr[Y′ = i − 2] + Pr[Y′ = i − 1](cid:1).
Finally, it remains to show that Pr[Y′ = j] > Pr[Y′ = j + 1] for all j > 0. The case j = 0 is true since
Pr[Y′ = 0] = ∏m−1
k=2 (1 − yk) and
Pr[Y′ = 1] =
m−1
∑
k=2
Pr[Y′ = 0] ·
yk
1 − yk
6
m−1
∑
k=2
Pr[Y′ = 0]
yk
1 − y2
=
Pr[Y′ = 0]
1 − y2
m−1
∑
i=2
yk 6 Pr[Y′ = 0],
where the last line follows from ∑m−1
k=2 yi 6 1 − y1 6 1 − y2 since y1 is the biggest element. The case
j > 1 follows from the fact the sequence (Pr[Y′ = j])j has one mode or two consecutive modes, and
at least one of them occurs at j = 0 (E[Y′] < 1 implies Pr[Y′ = 0] > Pr[Y′ = j] for all j > 2).
Case 2: 2 Pr[Y = 0] 6 Pr[Y = 1] and 2 Pr[Z = 0] 6 Pr[Z = 1]. Since ∑m
E[Z] = ∑m
2 Pr[Z = 0] 6 Pr[Z = 1], it implies
i=1 i · Pr[Z = i] = 1, we have Pr[Z = 0] = ∑m
i=0 Pr[Z = i] = 1 and
i=2(i − 1) Pr[Z = i]. Together with the fact
1 − Pr[Z = 1] = Pr[Z = 0] +
m
∑
i=2
Pr[Z = i] 6 2 Pr[Z = 0] < Pr[Z = 1],
11
so Pr[Z = 1] > 1
2 . Finally,
1 + max(Y, Z)(cid:21) =
1
E(cid:20)
=
>
=
>
m
∑
i=0
m
∑
i=0
m
∑
i=0
m
∑
i=0
m
∑
i=0
Pr[Y = i] ·
1
1 + i
−
m
∑
j=1
Pr[Z = j] · tj
Pr[Y = i] ·
Pr[Y = i] ·
Pr[Y = i] ·
1
1 + i
1
1 + i
1
1 + i
− Pr[Z = 1] · t1 −
− Pr[Z = 1] · t1 −
m
∑
j=2
m
∑
j=2
Pr[Z = j] · tj
j · Pr[Z = j] ·
t2
2
− Pr[Z = 1] · t1 −
t2
2
(1 − Pr[Z = 1])
Pr[Y = i] ·
1
1 + i
−
t1
2
−
t2
4
= E(cid:20)
1
1 + max(Y, YH)(cid:21),
where YH is drawn from Binomial(2, 1
second inequality follows from Pr[Z = 1] > 0.5 and t1 6 t2
2 .
2 ). The first inequality follows from Lemma 4.2, and the
1
1+max(Y,YH) ] >
Since YH satisfies 2 Pr[YH = 0] = Pr[YH = 1], the analysis for Case 1 shows that E[
1+max(1,YU) ].
1
E[
The following lemma finishes the proof of Theorem 1.2.
Lemma 4.4. For any m ∈ N, if Y ∼ Binomial(m, 1
m ),
E(cid:20)
1
1 + max(1, Y)(cid:21) > 0.4481
Proof. Since the binomial distribution is approximated by the Poisson distribution in the limit, we
use this to ease the calculation. Let Y ∼ Poisson(1).
E(cid:20)
1
1 + max(1, Y)(cid:21) =
1
k + 1
Pr[Y′ = k] +
1
2
Pr[Y < 2]
∞
∑
k=2
∞
∑
k=2
1
=
=
1
k!
e(cid:0)
k + 1
∞
∑
k=0
5
2
> 0.4481
= (e −
)
1
e
1
1
+
1
2
(
1
e
+
1
e
k! · e
1
2(cid:1) +
− 1 − 1 −
+
1
e
)
1
2
(
1
e
+
1
e
)
5 General Graphs
In this section, we prove Theorem 1.3 for weighted general graphs. Our proof methods here
closely follow that of Lemma 4.9 of Chekuri et al. [5] that lower bounds the correlation gap for
monotone submodular functions by 1 − 1/e. The only difference is that Lemma 5.1 holds for
matching with a weaker guarantee (if ν was a monotone submodular function, Lemma 5.1 would
hold with 2ν(G) replaced by ν(G)).
12
Proof. Fix weights w ∈ (R+ ∪ {0})E. Define F : [0, 1] → (R+ ∪ {0}) as F(x) := E
Now, fix x ∈ [0, 1]E in the matching polytope. We will show F(x) > 0.43 ∑e∈E wexe.
G∼DG
n,w,x
[ν(G)].
Consider the function φ(t) := F(tx) for t ∈ [0, 1].
For each e ∈ E,
∂F
∂xe(cid:12)(cid:12)(cid:12)(cid:12)tx
dφ
dt
= x · ∇F(tx) = ∑
e∈E
xe
∂F
∂xe(cid:12)(cid:12)(cid:12)(cid:12)tx
∂ E
=
G∼DG
n,w,tx
∂xe
[ν(G)]
(cid:12)(cid:12)(cid:12)(cid:12)tx
= E
= E
G∼DG
n,w,tx
G∼DG
n,w,tx
[ν(G)e ∈ G] − E
G∼DG
n,w,tx
[ν(G)e /∈ G]
[ν(G ∪ {e}) − ν(G \ {e})],
where G ∪ {e} (resp. G \ {e}) denotes the graph (V, E(G) ∪ {e}) (resp. (V, E(G) \ {e}).
Lemma 5.1. For any fixed graph G with weights {we} and any point x in the matching polytope,
∑
e∈E
xe(cid:0)ν(G ∪ {e}) − ν(G \ {e})(cid:1) + 2ν(G) > ∑
e∈E
xewe.
Proof. Let M ⊆ E(G) be a maximum weight matching of G. Note that
∑
e∈E
> ∑
e∈E
> ∑
e∈E
xe(cid:0)ν(G ∪ {e}) − ν(G \ {e})(cid:1) + 2ν(G)
xe(cid:0)ν(G ∪ {e}) − ν(G)(cid:1) + 2 ∑
xe(cid:0)ν(G ∪ {e}) − ν(G)(cid:1) + ∑
e∈E:e∼ f
f ∈M
f ∈M
w f
∑
xew f
(7)
(8)
where f ∼ e indicates that two edges f and e share an endpoint. To prove the lemma, it suffices
to show that for each e ∈ E, the coefficient of of xe in (8) is at least we. We consider the following
cases.
• If M ∪ {e} is a matching, ν(G ∪ {e}) > ν(G) + we and ν(G \ {e}) 6 ν(G), so ν(G ∪ {e}) −
ν(G \ {e}) > we.
• If e intersects exactly one edge f ∈ M, the coefficient of xe is ν(G ∪ {e}) − ν(G) + w f . If
w f > we, it is at least we. If w f < we, M ∪ {e} \ { f } is a matching of weight ν(G) + we − w f .
It implies that e /∈ E(G) and ν(G ∪ {e}) − ν(G) > we − w f , so ν(G ∪ {e}) − ν(G) + w f > we.
• If e intersects two edges f , g ∈ M, the coefficient of xe is ν(G ∪ {e}) − ν(G) + w f + wg. If
w f + wg > we, it is at least we. If w f + wg < we, M ∪ {e} \ { f , g} is a matching of weight
ν(G) + we − w f − wg. It implies that e /∈ E(G) and ν(G ∪ {e}) − ν(G) > we − w f − wg, so
ν(G ∪ {e}) − ν(G) + w f + wg > we.
13
Combining (7) and Lemma 5.1,
dφ
dt
= ∑
e∈E
= ∑
e∈E
> ∑
e∈E
= ∑
e∈E
xe
E
∂F
∂xe(cid:12)(cid:12)(cid:12)(cid:12)tx
G∼DG
n,w,tx
[ν(G ∪ e) − ν(G \ e)]
xewe − 2 E
G∼DG
n,w,tx
[ν(G)]
xewe − 2φ(t).
which implies that,
Since φ(0) = 0,
d
dt
(e2tφ(t)) = 2e2tφ(t) + e2t dφ
dt
> e2t ∑
e∈E
xewe.
e2φ(1) > ∑
e∈E
xewe Z 1
0
e2tdt =
e2 − 1
2
xewe,
∑
e∈E
which proves the theorem.
References
[1] S. Agrawal, Y. Ding, A. Saberi, and Y. Ye. Price of correlations in stochastic optimization.
Operations Research, 60(1):150 -- 162, 2012. 3
[2] J. Aronson, A. Frieze, and B. G. Pittel. Maximum matchings in sparse random graphs: Karp-
sipser revisited. Random Structures and Algorithms, 12(2):111 -- 177, 1998. 3
[3] T. Bohman and A. Frieze. Karp -- sipser on random graphs with a fixed degree sequence.
Combinatorics, Probability and Computing, 20(05):721 -- 741, 2011. 3
[4] P. Chebolu, A. Frieze, and P. Melsted. Finding a maximum matching in a sparse random
graph in o(n) expected time. J. ACM, 57(4):24:1 -- 24:27, May 2010. 3
[5] C. Chekuri, J. Vondrk, and R. Zenklusen. Submodular function maximization via the multi-
linear relaxation and contention resolution schemes. SIAM Journal on Computing, 43(6):1831 --
1879, 2014. Preliminary version in STOC'11. 2, 3, 12
[6] M. Cygan, F. Grandoni, and M. Mastrolilli. How to sell hyperedges: the hypermatching
assignment problem. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 342 -- 351. Society for Industrial and Applied Mathematics, 2013. 2
[7] P. Erdos and A. R´enyi. On the existence of a factor of degree one of a connected random
graph. Acta Mathematica Hungarica, 17(3-4):359 -- 368, 1966. 3
[8] P. Erdos and A. R´enyi. On random matrices ii. Studia Sci. Math. Hungar., 3:459 -- 464, 1968. 3
[9] A. Frieze. Perfect matchings in random bipartite graphs with minimal degree at least 2.
Random Structures and Algorithms, 26(3):319 -- 358, 2005. 3
14
[10] A. Frieze and B. Pittel. Perfect matchings in random graphs with prescribed minimal degree.
In Mathematics and Computer Science III, pages 95 -- 132. Springer, 2004. 3
[11] R. M. Karp and M. Sipser. Maximum matching in sparse random graphs. In Foundations of
Computer Science, 1981. SFCS'81. 22nd Annual Symposium on, pages 364 -- 375. IEEE, 1981. 2, 3
15
|
1802.06992 | 1 | 1802 | 2018-02-20T07:23:28 | Sublinear Algorithms for MAXCUT and Correlation Clustering | [
"cs.DS"
] | We study sublinear algorithms for two fundamental graph problems, MAXCUT and correlation clustering. Our focus is on constructing core-sets as well as developing streaming algorithms for these problems. Constant space algorithms are known for dense graphs for these problems, while $\Omega(n)$ lower bounds exist (in the streaming setting) for sparse graphs.
Our goal in this paper is to bridge the gap between these extremes. Our first result is to construct core-sets of size $\tilde{O}(n^{1-\delta})$ for both the problems, on graphs with average degree $n^{\delta}$ (for any $\delta >0$). This turns out to be optimal, under the exponential time hypothesis (ETH). Our core-set analysis is based on studying random-induced sub-problems of optimization problems. To the best of our knowledge, all the known results in our parameter range rely crucially on near-regularity assumptions. We avoid these by using a biased sampling approach, which we analyze using recent results on concentration of quadratic functions. We then show that our construction yields a 2-pass streaming $(1+\epsilon)$-approximation for both problems; the algorithm uses $\tilde{O}(n^{1-\delta})$ space, for graphs of average degree $n^\delta$. | cs.DS | cs |
Sublinear Algorithms for MAXCUT and Correlation Clustering ∗
Aditya Bhaskara1, Samira Daruki1,2, and Suresh Venkatasubramanian1
1School of Computing, University of Utah
2Expedia Research
Abstract
We study sublinear algorithms for two fundamental graph problems, MAXCUT and correla-
tion clustering. Our focus is on constructing core-sets as well as developing streaming algorithms
for these problems. Constant space algorithms are known for dense graphs for these problems,
while Ω(n) lower bounds exist (in the streaming setting) for sparse graphs.
Our goal in this paper is to bridge the gap between these extremes. Our first result is to
construct core-sets of size eO(n1−δ) for both the problems, on graphs with average degree nδ (for
any δ > 0). This turns out to be optimal, under the exponential time hypothesis (ETH). Our
core-set analysis is based on studying random-induced sub-problems of optimization problems.
To the best of our knowledge, all the known results in our parameter range rely crucially on
near-regularity assumptions. We avoid these by using a biased sampling approach, which we
analyze using recent results on concentration of quadratic functions. We then show that our
construction yields a 2-pass streaming (1 + ε)-approximation for both problems; the algorithm
uses eO(n1−δ) space, for graphs of average degree nδ.
Introduction
1
Sublinear algorithms are a powerful tool for dealing with large data problems. The range of
questions that can be answered accurately using sublinear (or even polylogarithmic) space or time
is enormous, and the underlying techniques of sketching, streaming, sampling and core-sets have
been proven to be a rich toolkit.
When dealing with large graphs, the sublinear paradigm has yielded many powerful results. For
many NP-hard problems on graphs, classic results from property testing [22, 7] imply extremely
efficient sublinear approximations. In the case of dense graphs, these results (and indeed older ones
of [10, 18]) provide constant time/space algorithms. More recently, graph sketching techniques
have been used to obtain efficient approximation algorithms for cut problems on graphs [2, 3] in a
streaming setting. These algorithms use space that is nearly linear in n (the number of vertices)
and are sublinear in the number of edges as long as E = ω(n) (this is called the "semi-streaming"
setting).
By way of lower bounds, recent results have improved our understanding of the limits of sketch-
ing and streaming. In a sequence of results [24, 25, 27], it was shown that for problems like matching
and MaxCut in a streaming setting, Ω(n) space is necessary in order to obtain any approximation
better than a factor 2 in one round. (Note that a factor 2 is trivial by simply counting edges.)
∗This research was supported in part by National Science Foundation under grants IIS-1251049, IIS-1633724
1
Furthermore, Andoni et al. [9] showed that any sketch for all the cuts in a graph must have size
Ω(n).
While these lower bounds show that O(n) space is the best possible for approximating problems
like MaxCut in general, the constructions used in these bounds are quite specialized. In particular,
the graphs involved are sparse, i.e., have Θ(n) edges. Meanwhile, as we mentioned above, if a graph
is dense (Ω(n2) edges), random sampling is known to give O(1) space and time algorithms. The
question we study in this paper is if there is a middle ground: can we get truly sublinear (i.e., o(n))
algorithms for natural graph problems in between (easy) dense graphs and (hard) sparse graphs?
Our main contribution is to answer this in the affirmative. As long as a graph has average
degree nδ for some δ > 0, truly sub-linear space (1 + ǫ) approximation algorithms are possible for
problems such as MaxCut and correlation clustering. Note that we consider the max-agreement
version of correlation clustering (see Section 2) Indeed, we show that a biased sample of vertices
forms a "core-set" for these problems. A core-set for an optimization problem (see [1]), is a subset
of the input with the property that a solution to the subset provides an approximation to the
solution on the entire input.
Our arguments rely on understanding the following fundamental question: given a graph G, is
the induced subgraph on a random subset of vertices a core-set for problems such as MaxCut?
This question of sub-sampling and its effect on the value of an optimization problem is well studied.
Results from property testing imply that a uniformly random sample of constant size suffices for
many problems on dense graphs. [18, 6] generalized these results to the case of arbitrary k-CSPs.
More recently, [12], extending a result in [16], studied the setting closest to ours. For graphs, their
results imply that when the maximum and minimum degrees are both Θ(nδ), then a random induced
subgraph with eO(n1−δ) acts as a core-set for problems such as MaxCut. Moreover, they showed
that for certain lifted relaxations, subsampling does not preserve the value of the objective. Finally,
using more modern techniques, [33] showed that the cut norm of a matrix (a quantity related to
the MaxCut) is preserved up to a constant under random sampling, improving on [18, 6]. While
powerful, we will see that these results are not general enough for our setting. Thus we propose
a new, conceptually simple technique to analyze sub-sampling, and present it in the context of
MaxCut and correlation clustering.
1.1 Our Results
As outlined above, our main result is to show that there exist core-sets of size eO(n1−δ) for MaxCut
and correlation clustering for graphs with Ω(n1+δ) edges (where 0 < δ ≤ 1). This then leads to
a two-pass streaming algorithm for MaxCut and correlation clustering on such graphs, that uses
eO(n1−δ) space and produces a 1 + ε approximation.
This dependence of the core-set size on δ is optimal up to logarithmic factors, by a result of [17].
Specifically, [17] showed that any (1+ε) approximation algorithm for MaxCut on graphs of average
degree nδ must have running time 2Ω(n1−δ ), assuming the exponential time hypothesis (ETH). Since
a core-set of size o(n1−δ) would trivially allow such an algorithm (we can perform exhaustive search
over the core-set), our construction is optimal up to a logarithmic factor, assuming ETH.
Our streaming algorithm for correlation clustering can be viewed as improving the semi-streaming
(space O(n)) result of Ahn et al. [4], while using an additional pass over the data. Also, in the
context of the lower bound of Andoni et al. [9], our result for MaxCut can be interpreted as saying
that while a sketch that approximately maintains all cuts in a graph requires an Ω(n) size, one
that preserves the MaxCut can be significantly smaller, when the graph has a polynomial average
degree.
2
At a technical level, we analyze the effect of sampling on the value of the MaxCut and corre-
lation clustering objectives. As outlined above, several techniques are known for such an analysis,
but we give a new and conceptually simple framework that (a) allows one to analyze non-uniform
sampling for the first time, and (b) gets over the assumptions of near-regularity (crucial for [16, 12])
and density (as in [18, 6]). We expect the ideas from our analysis to be applicable to other settings
as well, especially ones for which the 'linearization' framework of [10] is applicable.
The formal statement of results, an outline of our techniques and a comparison with earlier
works are presented in Section 4.
1.2 Related Work
MaxCut and correlation clustering are both extremely well-studied problems, and thus we will
only mention the results most relevant to our work.
Dense graphs. A graph is said to be dense if its average degree is Ω(n). Starting with the
work of Arora et al. [10], many NP hard optimization problems have been shown to admit a PTAS
when the instances are dense. Indeed, a small random induced subgraph is known to be a core-set
for problems such as MaxCut, and indeed all k-CSPs [22, 6, 18, 31]. The work of [10] relies on an
elegant linearization procedure, while [18, 6] give a different (and more unified) approach based on
"cut approximations" of a natural tensor associated with a CSP.
Polynomial density. The focus of our work is on graphs that are in between sparse (constant
average degree) and dense graphs. These are graphs whose density (i.e., average degree) is nδ, for
some 0 < δ < 1. Fotakis et al.
[17] extended the approach of [10] to this setting, and obtained
best possible, under the exponential time hypothesis (ETH). By way of core-sets, in their celebrated
work on the optimality of the Goemans-Williamson rounding, Feige and Schechtman [16] showed
(1 + ε) approximation algorithms with run-time exp(eO(n1−δ)). They also showed that it was the
that a random sample of eO(n1−δ) is a core-set for MaxCut, if the graphs are almost regular and
have an average degree nδ. This was extended to other CSPs by [12]. These arguments seem to use
near-regularity in a crucial way, and are based on restricting the number of possible 'candidates'
for the maximum cut.
In the streaming setting, there are several
Streaming algorithms and lower bounds.
algorithms [2, 29, 20, 3, 21, 28] that produce cut or spectral sparsifiers with O( n
ǫ2 ) edges using
O( n
ǫ2 ) space. Such algorithms preserves every cut within (1 + ǫ)-factor (and therefore also preserve
the max cut). Andoni et al. [9] showed that such a space complexity is essential; in fact, [9] show
that any sketch for all the cuts in a graph must have bit complexity Ω( n
ǫ2 ) (not necessarily streaming
ones). However, this does not rule out the possibility of being able to find a maximum cut in much
smaller space.
For MaxCut, Kapralov et al.
[26] and independently Kogan et al.
[30] proved that any
streaming algorithm that can approximate the MaxCut value to a factor better than 2 requires
O(√n) space, even if the edges are presented in random order. For adversarial orders, they showed
that for any ǫ > 0, a one-pass (1 + ǫ)-approximation to the max cut value must use n1−O(ǫ) space.
Very recently, Kapralov et al. [27] went further, showing that there exists an ǫ∗ > 0 such that every
randomized single-pass streaming algorithm that yields a (1 + ǫ∗)-approximation to the MAXCUT
size must use Ω(n) space.
[11] and has
Correlation clustering. Correlation clustering was formulated by Bansal et al.
been studied extensively. There are two common variants of the problem – maximizing agreement
and minimizing disagreement. While these are equivalent for exact optimization (their sum is a
constant), they look very different under an approximation lens. Maximizing agreement typically
3
admits constant factor approximations, but minimizing disagreement is much harder. In this paper,
we focus on the maximum-agreement variant of correlation clustering and in particular we focus
on (1 + ǫ)-approximations. Here, Ailon and Karnin [5] presented an approximation scheme with
sublinear query complexity (which also yields a semi-streaming algorithm) for dense instances of
correlation clustering. Giotis and Guruswami [19] described a sampling based algorithm combined
with a greedy strategy which guarantees a solution within (ǫn2) additive error. (Their work is
similar to the technique of Mathieu and Schudy [31].) Most recently, Ahn et al. [4] gave a single-
pass semi-streaming algorithm for max-agreement. For bounded weights, they provide an (1 + ǫ)-
approximation streaming algorithm and for graphs with arbitrary weights, they present a 0.766(1−
ǫ)-approximation algorithm. Both algorithms require (nǫ−2) space. The key idea in their approach
was to adapt multiplicative-weight-update methods for solving the natural SDPs for correlation
clustering in a streaming setting using linear sketching techniques.
2 Definitions
Definition 2.1 (MaxCut). Let G = (V, E, w) be a graph with weights w : E → R+. Let (A, B)
be a partition of V and let w(A, B) denote the sum of weights of edges between A and B. Then
MaxCut(G) = max(A,B) partition of V w(A, B).
For ease of exposition, we will assume that the input graph for MaxCut is unweighted. Our
techniques apply as long as all the weights are O(1). Also, we denote by ∆ the average degree, i.e.,
ij where for every edge ij we have c+
ij, c−
ij and for each vertex, di =Pi∈Γ(j) ηij. We will also assume that all
Moving now to correlation clustering, let G = (V, E, c+, c−) be a graph with edge weights c+
ij
ij ≥ 0 and only one of them is nonzero. For every edge
Pi,j wij/V .
and c−
ij ∈ E, we define ηij = c+
the weights are bounded by an absolute constant in magnitude (for simplicity, we assume it is 1).
We define the "average degree" ∆ (used in the statements that follow) of a correlation clustering
instance to be (Pi di)/n.
same cluster and 0 otherwise. The MAX-AGREE score of this clustering is given by Pij c+
Pij c−
Note that the objective value can be simplified toPij c−
denotes the sumPij c−
Definition 2.2 (MAX-AGREE correlation clustering). Given G = (V, E, c+, c−) as above, consider
a partition of V into clusters C1, C2, . . . , Ck, and let χij be an indicator that is 1 if i an j are in the
ijχij +
ij(1 − χij). The goal is to find a partition maximizing this score. The maximum value of the
ij + ηijχij = C − +Pij ηijχij, where C −
We will also frequently use concentration bounds, which we state next.
score over all partitions of V will be denoted by CC(G).
ij − c−
ij.
3 Preliminaries
We will frequently appeal to Bernstein's inequality for concentration of linear forms of random
variables. For completeness, we state it here.
Theorem 3.1 (Bernstein's inequality[15]). Let the random variables X1,··· , Xn be independent
i be the variance of X.
Then, for any t > 0,
with Xi − E[Xi] ≤ b for each i ∈ [n]. Let X =Pi Xi and let σ2 =Pi σ2
Pr[X − E[X] > t] ≤ exp(−
4
t2
2σ2(1 + bt/3σ2)
)
A slightly more non-standard concentration inequality we use is from Boucheron, Massart and
Lugosi [13]. It can be viewed as an exponential version of the classic Efron-Stein lemma.
Theorem 3.2 ([13]). Assume that Y1,··· , Yn are random variables, and Y n
1 is the vector of these
n random variables. Let Z = f (Y1,··· , Yn), where f : χn → R is a measurable function. De-
fine Z (i) = f (Y1,··· , Yi−1, Y ′
n denote the independent copies of
Y1,··· , Yn. Then, for all θ > 0 and λ ∈ (0, 1
θ ),
i , Yi+1,··· , Yn), where Y1,··· , Y ′
log E[eλ(Z−E[Z])] ≤
log E[e
λL+
θ
],
λθ
1 − λθ
where L+ is the random variable defined as
L+ = E[
nXi=1
2
(Z − Z (i))
1Z>Z (i)Y n
1 ].
4 Technical overview
We now present an outline of our main ideas. Suppose we have a graph G = (V, E). First, we define
a procedure vertex sample. This takes as input probabilities pi for every vertex, and produces a
random weighted induced subgraph.
Sample a set S′ of vertices by selecting each vertex vi
Procedure vertex sample ({pi}i∈V ).
with probability pi independently. Define H to be the induced subgraph of G on the vertex set S′.
For i, j ∈ S′, define wij = 1
Intuitively, the edge weights are chosen so that the total number of edges remains the same,
in expectation. Next, we define the notion of an importance score for vertices. Let di denote the
degree of vertex i.
pipj∆2 .1
Definition 4.1. The importance score hi of a vertex i is defined as hi = min{1, max{di,ǫ∆}
αǫ is an appropriately chosen parameter (for MaxCut, we set it to
clustering, we set it to
}, where
C log n , and for correlation
C log n , where C is an absolute constant).
ǫ8
∆2αǫ
ǫ4
The main result is now the following:
Theorem 4.2 (Core-set). Let G = (V, E) have an average degree ∆. Suppose we apply vertex
sample with probabilities pi ∈ [hi, 2hi] to obtain a weighted graph H. Then H has eO( n
∆ ) vertices and
the quantities MaxCut(H) and CC(H) are within a (1 + ǫ) factor of the corresponding quantities
MaxCut(G) and CC(G), w.p. at least 1 − 1
n2 .
While the number of vertices output by the vertex sample procedure is small, we would like a
core-set of small "total size". This is ensured by the following.
Procedure edge sample (H). Given a weighted graph H with total edge weight W , sample
ε2W ), to obtain a graph H ′.
each edge e ∈ E(H) independently with probability pe := min(1, 8S ′we
Now, assign a weight we/pe to the edge e in H ′.
The procedure samples roughly S′/ε2 edges, with probability proportional to the edge weights.
The graph is then re-weighted in order to preserve the total edge weight in expectation, yielding:
1In correlation clustering, we have edge weights to start with, so the weight in H will be wij · c+
ij (or c−
ij ).
5
Theorem 4.3 (Sparse core-set). Let G be a graph n vertices and average degree ∆ = nδ. Let H ′
be the graph obtained by first applying vertex sample and then applying edge sample. Then
H ′ is a ǫ-core-set for MaxCut and CC, having size eO( n
∆ ) = eO(n1−δ).
We then show how to implement the above procedures in a streaming setting. This gives:
Theorem 4.4 (Streaming algorithm). Let G be a graph on n vertices and average degree ∆ = nδ,
whose edges arrive in a streaming fashion in adversarial order. There is a two-pass streaming algo-
rithm with space complexity eO( n
and CC(G).
∆ ) = eO(n1−δ) for computing a (1+ǫ)-approximation to MaxCut(G)
Of these, Theorem 4.2 is technically the most challenging. Theorem 4.3 follows via standard
edge sampling methods akin to those in [2] (which show that w.h.p., every cut size is preserved). It
is presented in Section 7, for completeness. The streaming algorithm, and a proof of Theorem 4.4,
are presented in Section 8. In the following section, we give an outline of the proof of Theorem 4.2.
4.1 Proof of the sampling result (theorem 4.2): an outline
In this outline we will restrict ourselves to the case of MaxCut as it illustrates our main ideas.
Let G be a graph as in the statement of the theorem, and let H be the output of the procedure
vertex sample.
Showing that MaxCut(H) is at least MaxCut(G) up to an εn∆ additive term is easy. We
simply look at the projection of the maximum cut in G to H (see, for instance, [16]). Thus, the
challenge is to show that a sub-sample cannot have a significantly larger cut, w.h.p. The natural
approach of showing that every cut in G is preserved does not work as 2n cuts is too many for the
purposes of a union bound.
There are two known ways to overcome this. The first approach is the one used in [22, 16]
and [12]. These works essentially show that in a graph of average degree ∆, we need to consider
only roughly 2n/∆ cuts for the union bound. If all the degrees are roughly ∆, then one can show that
all these cuts are indeed preserved, w.h.p. There are two limitations of this argument. First, for
i , where p is the sampling probability) can be large,
and we cannot take a union bound over exp(n/∆) cuts. Second, the argument is combinatorial,
and it seems difficult to generalize this to analyze non-uniform sampling.
non-regular graphs, the variance (roughlyPi pd2
The second approach is via cut decompositions, developed in [18, 6]. Here, the adjacency matrix
A is decomposed into poly(1/ε) rank-1 matrices, plus a matrix that has a small cut norm. It turns
out that solving many quadratic optimization problems (including MaxCut) on A is equivalent
(up to an additive εn∆) to solving them over the sum of rank-1 terms (call this A′). Now, the
adjacency matrix of H is an induced square sub-matrix of A, and since we care only about A′
(which has a simple structure), [6] could show that MaxCut(H) ≤ MaxCut(G) + εn2, w.h.p. To
the best of our knowledge, such a result is not known in the "polynomial density" regime (though
the cut decomposition still exists).
Our technique. We consider a new approach. While inspired by ideas from the works above,
it also allows us to reason about non-uniform sampling in the polynomial density regime. Our
starting point is the result of Arora et al. [10], which gives a method to estimate the MaxCut
using a collection of linear programs (which are, in turn, derived using a sample of size n/∆). Now,
by a double sampling trick (which is also used in the approaches above), it turns out that showing
a sharp concentration bound for the value of an induced sub-program of an LP as above, implies
Theorem 4.2. As it goes via a linear programming and not a combinatorial argument, analyzing
non-uniform sampling turns out to be quite direct. Let us now elaborate on this high level plan.
6
Induced sub-programs.
First, we point out that an analysis of induced sub-programs is
also an essential idea in the work of [6]. The main difference is that in their setting, only the
variables are sub-sampled (and the number of constraints remains the same).
In our LPs, the
constraints correspond to the vertices, and thus there are fewer constraints in the sampled LP. This
makes it harder to control the value of the objective. At a technical level, while a duality-based
argument using Chernoff bounds for linear forms suffices in the setting of [6], we need the more
recent machinery on concentration of quadratic functions.
We start by discussing the estimation technique of [10].
Estimation with Linear Programs. The rough idea is to start with the natural quadratic
program for MaxCut: maxP(i,j)∈E xi(1 − xj), subject to xi ∈ {0, 1}.2 This is then "linearized"
using a seed set of vertices sampled from G. We refer to Section 5 for details. For now, Est(G) is
a procedure that takes a graph G and a set of probabilities {γi}i∈V (G), samples a seed set using γ,
and produces an estimate of MaxCut (G).
Now, suppose we have a graph G and a sample H. We can imagine running Est(G) and
Est(H) to obtain good estimates of the respective MaxCut values. But now suppose that in both
cases, we could use precisely the same seed set. Then, it turns out that the LPs used in Est(H)
would be 'induced' sub-programs (in a sense we will detail in Section 6) of those used in Est(H),
and thus proving Theorem 4.2 reduces to showing a sufficiently strong concentration inequality for
sub-programs.
The key step above was the ability to use same seed set in the Est procedures. This can be
formalized as follows.
Double sampling. Consider the following two strategies for sampling a pair of subsets (S, S′)
of a universe [n] (here, qv ≤ pv for all v):
• Strategy A: choose S′ ⊆ [n], by including every v w.p. pv, independently; then for v ∈ S′,
include them in S w.p. qv/pv, independently.
• Strategy B: pick S ⊆ [n], by including every v w.p. qv; then iterate over [n] once again,
placing v ∈ S′ with a probability equal to 1 if v ∈ S, and p∗
v if v 6∈ S.
Lemma 4.5. Suppose p∗
strategies A and B are identical.
v = pv(1 − qv
pv
)/(1 − pv). Then the distribution on pairs (S, S′) obtained by
The proof is by a direct calculation, which we state it here.
Proof. Let us examine strategy A. It is clear that the distribution over S is precisely the same as
the one obtained by strategy B, since in both the cases, every v is included in S independently of
the other v, with probability precisely qv. Now, to understand the joint distribution (S, S′), we
need to consider the conditional distribution of S′ given S. Firstly, note that in both strategies,
S ⊆ S′, i.e., Pr[v ∈ S′v ∈ S] = 1. Next, we can write Prstrategy A[v ∈ S′v 6∈ S] as
Prstrategy A[v ∈ S′ ∧ v 6∈ S]
Prstrategy A[v 6∈ S]
=
pv
pv(1 − qv
1 − pv
)
.
Noting that PrstrategyB[v ∈ S′v 6∈ S] = p∗
2This is a valid formulation, because for every xi 6= xj that is an edge contributes 1 to the objective, and xi = xj
v (by definition) concludes the proof.
contribute 0.
7
To show the theorem, we use pv as in the statement of the theorem,
Proof of Theorem 4.2.
and set q to be the uniform distribution qv = 16 log n
ε2∆ . The proof now proceeds as follows. Let S′ be
a set sampled using the probabilities pv. These form the vertex set of H. Now, the procedure Est
on H (with sampling probabilities qv/pv) samples the set S (as in strategy A). By the guarantee of
the estimation procedure (Corollary 5.1.2), we have MaxCut(H) ≈ Est(H), w.h.p. Next, consider
the procedure Est on G with sampling probabilities qv. Again, by the guarantee of the estimation
procedure (Corollary 5.1.1), we have MaxCut(G) ≈ Est(G), w.h.p.
Now, we wish to show that Est(G) ≈ Est(H). By the equivalence of the sampling strategies,
we can now take the strategy B view above. This allows us to assume that the Est procedures
use the same S, and that we pick S′ after picking S. This reduces our goal to one of analyzing the
value of a random induced sub-program of an LP, as mentioned earlier. The details of this step
are technically the most involved, and are presented in Section 6. This completes the proof of the
theorem. (Note that the statement also includes a bound on the number of vertices of H. This
follows immediately from the choice of pv.)
5 Estimation via linear programming
We now present the estimation procedure Est used in our proof. It is an extension of [10] to the
case of weighted graphs and non-uniform sampling probabilities.
Let H = (V, E, w) be a weighted, undirected graph with edge weights wij, and let γ : V →
[0, 1] denote sampling probabilities. The starting point is the quadratic program for MaxCut:
max Pij∈E wijxi(1 − xj), subject to xi ∈ {0, 1}. The objective can be re-written asPi∈V xi(di −
Pj∈Γ(i) wijxj), where di is the weighted degree, Pj∈Γ(i) wij. The key idea now is to "guess" the
value of ρi :=Pj∈Γ(i) wijxj, by using a seed set of vertices. Given a guess, the idea is to solve the
following linear program, which we denote by LPρ(V ).
xi(di − ρi) − si − ti
maximize Xi
subject to ρi − ti ≤ Xj∈Γ(i)
0 ≤ xi ≤ 1,
si, ti ≥ 0.
wijxj ≤ ρi + si
The variables are xi, si, ti. Note that if we fix the xi, the optimal si, ti will satisfy si + ti =
ρi −Pj∈Γ(i) wijxj. Also, note that if we have a perfect guess for ρi's (coming from the MaxCut),
i ∈ V is included w.p. γi independently. For every partition (A, S \ A) of S, set ρi =Pj∈Γ(i)∩A
the objective can be made ≥ MaxCut(H).
Estimation procedure. The procedure Est is the following: first sample a set S ⊆ V where each
wij
,
γj
and solve LPρ(V ) (in what follows, we denote this LP by LP γ
A,S\A(V ), as this makes the partition
and the sampling probabilities clear). Return the maximum of the objective values.
Our result here is a sufficient condition for having Est(H) ≈ MaxCut(H).
Theorem 5.1. Let H be a weighted graph on n vertices, with edge weights wij that add up to W .
Suppose the sampling probabilities γi satisfy the condition
Then, we have Est(H, γ) ∈ MaxCut(H) ± εW , with probability at least 1 − 1/n2 (where the
probability is over the random choice of S).
for all i, j.
(1)
wij ≤
W ε2
8 log n
γiγjPu γu
8
The proof of the Theorem consists of claims showing the upper and lower bound separately.
Claim 1. The estimate is not too small. I.e., w.h.p. over the choice of S, there exists a cut
(A, S \ A) of S such that LPA,S\A(H) ≥ MaxCut(H) − εW .
The estimate is not much larger than an optimal cut. Formally, for any feasible
Claim 2.
solution to the LP (and indeed any values ρi), there is a cut in H of value at least the LP objective.
Proof of Claim 1. Let (AH , V \ AH ) be the max cut in the full graph H. Now consider a sample
=
S, and let (A, S \ A) be its projection onto S. For any vertex i, recall that ρi =Pj∈Γ(i)∩A
Pj∈Γ(i)∩AH
, where Yj is the indicator for j ∈ S. Thus
wij
γj
wij
γj
Yj
E[ρi] = Xj∈Γ(i)∩AH
γj
wij
γj
= Xj∈Γ(i)∩AH
wij.
We will use Bernstein's inequality to bound the deviation in ρi from its mean. To this end, note
that the variance can be bounded as
(1 − γj)w2
ij
γj
.
Var[ρi] = Xj∈Γ(i)∩AH
j ≤ Xj∈Γ(i)
In what follows, let us write di =Pj∈Γ(i) wij and fi = W γi
on the wij implies that wij
Now, using Bernstein's inequality (Theorem 3.1),
γj(1 − γj)
Pu γu
w2
ij
γ2
γj ≤ ε2
Pr[ρi − E[ρi] > t] ≤ exp −
. Then, for every j, our assumption
8 log n fi. Thus, summing over j, we can bound the variance by ε2difi
8 log n .
t2
ε2difi
4 log n + 2t
3
8 log n! .
ε2fi
(2)
Setting t = ǫ(di + fi), and simplifying, we have that the probability above is < exp(−4 log n) = 1
n4 .
Thus, we can take a union bound over all i ∈ V , and conclude that w.p. ≥ 1 − 1
n3 ,
≤ ε(di + fi)
for all i ∈ V .
(3)
Q(x) :=Xi
xi(cid:0)di − Xj∈Γ(i)
wijxj(cid:1).
9
ρi − Xj∈Γ(i)∩AH
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
wij(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
xi(di − ρi) − ε(di + fi) ≥Xi
Xi
For any S that satisfies the above, consider the solution x that sets xi = 1 for i ∈ AH and 0
(eq. (3)). Thus the LP objective can be lower bounded as
otherwise. We can choose si + ti = ρi −Pj∈Γ(i) wijxj ≤ ε(di + fi), by the above reasoning
xi(di − Xj∈Γ(i)
wijxj) − 2ε(di + fi).
of the claim.
This is precisely MaxCut(G) − 2εPi(di + fi) ≥ MaxCut(G) − 4εW . This completes the proof
Proof of Claim 2. Suppose we have a feasible solution x to the LP, of objective value Pi xi(di −
ρi)−Pi ρi −Pj∈Γ(i) wijxj, and we wish to move to a cut of at least this value. To this end, define
the quadratic form
The first observation is that for any x ∈ [0, 1]n, and any real numbers ρi, we have
Q(x) ≥Xi
xi(di − ρi) −Xi
ρi − Xj∈Γ(i)
wijxj.
This is true simply because Q(x) =Pi xi(cid:0)di − ρi(cid:1) + xi(cid:0)ρi −Pj∈Γ(i) wijxj(cid:1), and the fact that the
second term is at least −ρi −Pj∈Γ(i) wijxj, as xi ∈ [0, 1].
Next, note that the maximum of the form Q(x) over [0, 1]n has to occur at a boundary point,
since for any fixing of variables other than a given xi, the form reduces to a linear function of
xi, which attains maximum at one of the boundaries. Using this observation repeatedly lets us
conclude that there is a y ∈ {0, 1}n such that Q(y) ≥ Q(x). Since any such y corresponds to a cut,
and Q(y) corresponds to the cut value, the claim follows.3
Finally, to show Theorem 4.2 (as outlined in Section 4.1), we need to apply Theorem 5.1 with
specific values for γ and wij. Here we state two related corollaries to Theorem 5.1 that imply good
estimates for the MaxCut.
Corollary 5.1.1. Let H in the framework be the original graph G, and let γi = 16 log n
Then the condition wij ≤ ε2
ǫW , w.p. ≥ 1 − n−2.
ε2∆ for all i.
holds for all i, j, and therefore Est(G, γ) ∈ MaxCut(G) ±
8 log n · W γiγj
Pu γu
The proof is immediate (with a slack of 2), as wij = 1, W = n∆, and all γu are equal.
Corollary 5.1.2. Let H be the weighted sampled graph obtained from vertex sample, and let γi =
16 log n
. Then the condition (1) holds w.p. ≥ 1− n−3, and therefore Est(H, γ) ∈ MaxCut(H)±
ε2∆
1
pi
ǫW w.p. ≥ 1 − n−2.
Proof. In this case, we have wij = 1
pipj∆2 . Thus, simplifying the condition, we need to show that
1
pipj∆2 ≤
2W
pipj∆
1
Pu∈H
.
1
pu
Now, for H sampled via probabilities pi, we have (in expectation) W = n
A straightforward application of Bernstein's inequality yields that W ≥ n
w.p. at least 1 − n−3. This completes the proof.
∆ , andPu∈H
2∆ and Pu∈H
1
pu
= n.
pu ≤ 2n,
1
6 Random induced linear programs
We will now show that the Est on H has approximately the same value as the estimate on G (with
appropriate γ values). First, note that Est(G) is maxA⊆S LP γ
A,S\A(G), where γi = qi. To write the
. For the graph
H, the estimation procedure uses an identical program, but the sampling probabilities are now αi :=
. Also,
pipj∆2 .
LP, we need the constants ρi, defined by the partition (A, S\A) as ρi :=Pj∈Γ(i)∩A
qi/pi, and the estimates ρ, which we now denote by eρi, are defined by eρi :=Pj∈Γ(i)∩A
pi∆2 . The degrees are now edi :=Pj∈Γ(i)∩S ′ wij =Pj∈Γ(i)∩S ′
by the way we defined wij,eρi = ρi
Our aim in this section is to show the following:
The two LPs are shown in Figure 1.
pjwij
1
qj
qj
1
3We note that the proof in [10] used randomized rounding to conclude this claim, but this argument is simpler;
also, later papers such as [17] used such arguments for derandomization.
10
max Xi∈G
s.t. Xj∈Γ(i)
− Xj∈Γ(i)
[xi(di − ρi) − (si + ti)]
xj ≤ ρi + si, ∀i ∈ [n]
xj ≤ −ρi + ti, ∀i ∈ [n]
max Xi∈S ′
s.t. Xj∈Γ(i)∩S ′
− Xj∈Γ(i)∩S ′
[xi(edi −eρi) − (si+ti)]
wijxj ≤eρi + si, ∀i ∈ S′
wijxj ≤ −eρi + ti, ∀i ∈ S′
si, ti ≥ 0 ∀i ∈ S′
0 ≤ xi ≤ 1 ∀i ∈ [n]
0 ≤ xi ≤ 1,
(a) The LP on the full graph
(b) The sampled LP
Figure 1: The two LPs.
Theorem 6.1. Let G be an input graph, and let (S, S′) be sampled as described in Section 4.1.
Then, with probability ≥ 1 − 1
max
A⊆S
A,S\A(G) ≥ ∆2 · max
A,S\A(H) − εn∆.
n2 , we have
LP γ
LP α
A⊆S
Proof outline. To prove the theorem, the idea is to take the "strategy B" viewpoint of sampling
(S, S′), i.e., fix S, and sample S′ using the probabilities p∗. Then, we only need to understand
the behavior of an "induced sub-program" sampled with the probabilities p∗. This is done by
considering the duals of the LPs, and constructing a feasible solution to the induced dual whose
cost is not much larger than the dual of the full program, w.h.p. This implies the result, by linear
programming duality.
Let us thus start by understanding the dual of LP γ
note that for any given z, the optimal choice of ui is max{0, di − ρi −Pj∈Γ(i) zj}; thus we can think
of the dual solution as being the vector z. The optimal ui may thus be bounded by 2di, a fact that
we will use later. Next, we write down the dual of the induced program, LP α
A,S\A(H), as shown in
Figure 2b.
A,S\A(G) given A, shown in Figure 2a. We
minimize Xi∈G
ui + Xj∈Γ(i)
ui + ρizi
s.t.
zj ≥ di − ρi ∀i ∈ V
ui ≥ 0, −1 ≤ zi ≤ 1 ∀i ∈ V
(a) The dual of LP γ
A,S\A(G)
minimize Xi∈S ′
ui + Xj∈Γ(i)∩S ′
s.t.
[ui +eρizi]
wij zj ≥ edi −eρi ∀i ∈ S′
ui ≥ 0, −1 ≤ zi ≤ 1 ∀i ∈ S′.
(b) The dual of
LP α
A,S\A(H).
the induced program
Figure 2: The dual LPs
Following the outline above, we will construct a feasible solution to LP (2b), whose cost is close
to the optimal dual solution to LP (2a). The construction we consider is very simple: if z is the
optimal dual solution to (2a), we set zi = zi for i ∈ S′ as the candidate solution to (2b). This is
clearly feasible, and thus we only need to compare the solution costs. The dual objective values
11
are as follows
DualG =Xi∈V
ρizi + max{0, di − ρi − Xj∈Γ(i)
DualH ≤Xi∈S ′eρizi + max{0, edi −eρi − Xj∈Γ(i)∩S ′
zj}
wijzj}
(4)
(5)
Note that there is a ≤ in (5), as zi = zi is simply one feasible solution to the dual (which is a
minimization program). Next, our goal is to prove that w.p. at least 1 − 1
n2 ,
max
A⊆S
DualH ≤
1
∆2 · max
A⊆S
DualG +
εn
∆
.
Note that here, the probability is over the choice of S′ given S (as we are taking view-B of the
sampling). The first step in proving the above is to move to a slight variant of the quantity DualH,
which is motivated by the fact that Pr[Yi = 1] is not quite pi, but p∗
i (as we have conditioned on S).
ij.
ij := 1
j ∆2 . So also, let d∗
i p∗
p∗
ρi
pi∆2 ), and w∗
i := ρi
p∗
i :=Pj∈Γ(i) Yjw∗
(6)
Let us defineeρ∗
Then, define
Dual∗
i ∆2 (recall that eρi is
H :=Xi∈S ′eρ∗
i zi + max{0, ed∗
H ≤ εn
2∆ .
i − Xj∈Γ(i)∩S ′
i −eρ∗
w∗
ijzj}.
A straightforward lemma, which we use here, is the following. Here we bound the difference
between the "corrected" dual we used to analyze, and the value we need for the main theorem.
Specifically, we bound DualH − Dual∗
Lemma 6.2. Let (S, S′) be sampled as in Section 4.1. Then w.p. at least 1 − 1
all z ∈ [−1, 1]n and for all partitions (A, S \ A) of S,4 DualH − Dual∗
H ≤ εn
2∆ .
Proof. To prove the lemma, it suffices to prove that w.p. ≥ 1 − 1
n4 ,
Yjwij − w∗
n4 , we have that for
ij ≤
εn
2∆
(7)
.
Xi
Yieρi −eρ∗
i +Xi
Yi Xj∈Γ(i)
n4 over the choice of (S, S′), we have:
This is simply by using the fact that zi are always in [−1, 1]. Before showing this, we introduce
some notation and make some simple observations. First, denote by Y the indicator vector for S′
and by X the indicator for S.
Observation 6.3. With probability ≥ 1 − 1
1. For all i ∈ V ,Pj∈Γ(i)
qj ≤ 2(di + ε∆).
2. For all i ∈ V ,Pj∈Γ(i),j6∈S
3. Pi
≤ 2ε2n∆.
4. Pi6∈S
≤ 2n∆.
4Note that the partition defines the ρi.
j ≤ 2(di + ε∆).
Xi(di+ε∆)
Yi(di+ε∆)
Yj
p∗
Xj
p∗
i
pi
12
All the inequalities are simple consequences of Bernstein's inequality (and our choice of parame-
i , qi), and we thus skip the proofs. Next, note that as an immediate consequence of part-1,
ters pi, p∗
we have
ρi ≤ 2(di + ε∆),
for all partitions (A, S \ A) of S.
Also, note that from the definitions of the quantities (and the fact qi/pi ≤ ε2), we have
Now, we are ready to show (7). The first term can be bounded as follows:
(8)
(9)
(10)
∀i 6∈ S,
ε2
p∗
i
1
p∗
1
pi −
(cid:12)(cid:12)(cid:12)(cid:12)
i(cid:12)(cid:12)(cid:12)(cid:12) =Xi∈S
i(cid:12)(cid:12)(cid:12)(cid:12) ≤
∆2(cid:12)(cid:12)(cid:12)(cid:12)
ρi
Xi
i =Xi
Yieρi −eρ∗
Yiρi
∆2 (cid:12)(cid:12)(cid:12)(cid:12)
1
pi −
1
p∗
1
pi − 1(cid:12)(cid:12)(cid:12)(cid:12) +Xi6∈S
Yiρi
∆2 (cid:12)(cid:12)(cid:12)(cid:12)
1
pi −
1
p∗
i(cid:12)(cid:12)(cid:12)(cid:12) .
Using (8) and part-3 of the observation, the first term can be bounded by O(ε2n/∆). For the
second term, using (9) together with part-4 of the observation gives a bound of O(ε2n/∆). Thus
the RHS above is at most εn
16∆ , as we may assume ε is small enough.
Now, consider the second term in (7). When i 6∈ S and j 6∈ S, we have wij − w∗
ij being "small".
We can easily bound by 2ε2w∗
ij, using
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
p∗
1
p∗
i p∗
In the last steps, we used (9) and the fact that p∗
1
pipj −
1
pipj −
1
i pj −
p∗
i pj(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
For i ∈ S or j ∈ S, we can simply bound wij − w∗
ij by 2wij. Thus we can bound the second
+Xi6∈S
pi∆2 Xj∈Γ(i)
Yi
j(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤
i ∆2 Xj∈Γ(i)\S
j(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12)
4Xi∈S
i ≤ pi for i 6∈ S.
ε2
j ≤
i p∗
p∗
term in (7) as
2ε2
p∗
i p∗
j
ε2
p∗
i pj
1
p∗
i p∗
ε2Yj
pj
Yj
pj
p∗
+
1
.
.
The second term has only a sum over j not in S – this is why have an extra 2 factor for the first term.
Yj
.
pj
j for j 6∈ S, we have
Now, consider the first term. The inner summation can be written asPj∈Γ(i)∩S
Using parts 1 and 2 of the observation, together with pj ≥ qj/αε, and pj ≥ p∗
Pj∈Γ(i)
Let us thus consider the second term. Again using part 2 along with pj ≥ p∗
j for j 6∈ S, we can
bound the inner sum by 2ε2(di + ε∆). Then we can appeal to part-4 of the observation to obtain
the final claim.
pj ≤ 4αε(di + ε∆). Then, using part-3 gives the desired bound on the first term.
+Pj∈Γ(i)\S
1
pj
Yj
This ends up bounding the second term of (7), thus completing the proof of the lemma.
Thus our goal is to show the following:
Lemma 6.4. Let S satisfy the conditions (a) S ≤ 20n log n
2(di + ε∆). Then, w.p. ≥ 1 − 1
n4 over the choice of S′ given S, we have
ε2∆ , and (b) for all i ∈ V ,Pj∈Γ(i)∩S
1
qj ≤
max
A⊆S
Dual∗
H ≤
1
∆2 · max
A⊆S
DualG +
εn
2∆
.
The condition (b) on S is a technical one that lets us bound ρi in the proofs.
Given Lemma 6.2 and 6.4, it is easy to prove the Theorem 6.1 as follows:
13
Proof of Theorem 6.1. The conditions we assumed on S in Lemma 6.4 hold w.p. at least 1 − 1
n4
(via a simple application of Bernstein's inequality). Thus the conclusion of the lemma holds w.p.
at least 1− 2
∆2 maxA DualG + εn
w.p. at least 1 − 3
n4 . Combining this with Lemma 6.2, we have that maxA DualH ≤ 1
n4 . The theorem then follows via LP duality.
∆
It thus suffices to prove Lemma 6.4. The main step is to show a concentration bound on a
quadratic function that is not quite a quadratic form. This turns out to be quite technical, and we
discuss it in the following sections.
6.1 Proof of Lemma 6.4
Let Yi = 1i∈S ′. For convenience, let us denote the max{} terms in equations (4) and (6) by ui and
u∗
i , respectively. Now,
Dual∗
H −
1
∆2
DualG =Xi (cid:18)Yieρ∗
i zi −
1
∆2 · ρizi(cid:19) +(cid:18)Yi u∗
i −
1
∆2 · ui(cid:19) .
(11)
We view the RHS as two summations (shown by the parentheses), and bound them separately.
The first is relatively easy. Recall that by definition, eρ∗
i ∆2(cid:0)Yi − p∗
i(cid:1). The expectation of this quantity is 0. We will apply Bernstein's inequality to
bound its magnitude. For this, note that the variance is at most (using zi ≤ 1)
i = ρi
i ∆2 . Thus the first term is equal to
p∗
Pi
ρizi
p∗
Xi
ρ2
i )2∆4 p∗
i
(p∗
i (1 − p∗
i ) ≤Xi
4(di + ε∆)2(1 − p∗
i )
p∗
i ∆4
.
The condition on S gives the bound on ρi that was used above. Next, we note that unless p∗
we have p∗
4αε·(di+ε∆)
i ≥ (di+ε∆)
αε∆2 . Thus the variance is bounded byPi
∆2
≤ 8αεn
∆ . Next,
i = 1,
(Again, this is because we can ignore terms with p∗
Thus, by Bernstein's inequality,
i = 1, and for the rest, we have a lower bound.)
≤ 2αε.
i
max
p∗
i ∆2
ρizi
p∗
2(di + ε∆)
i ∆2(cid:12)(cid:12)(cid:12)(cid:12) ≤
(cid:12)(cid:12)(cid:12)(cid:12)
i(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ t] ≤ exp −
i ∆2(cid:0)Yi − p∗
ρizi
p∗
Pr[(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi
t2
∆ + 2tαε! .
16αεn
Setting t = εn/4∆, the bound simplifies to exp(− ε2n
αε and our size bound on S, this is < exp(−S)/n4.
to be important) of splitting it into two terms by adding a "hybrid" term, as follows:
The second term of (11) requires most of the work. We start with the trick (which turns out
), for a constant C. Thus, by our choice of
C∆αε
Xi
Yi u∗
i −
1
∆2 · ui =Xi (cid:18)Yi u∗
i − Yi
ui
i ∆2(cid:19) +Xi (cid:18)Yi
p∗
ui
i ∆2 −
p∗
1
∆2 · ui(cid:19) .
The second term will again be bounded using Bernstein's inequality (in which we use our
earlier observation that ui = O(di)). This gives an upper bound of εn/8∆, with probability
1 − exp(−S)/n4. We omit the easy details.
14
Let us focus on the first term. We now use the simple observation that max{0, A}−max{0, B} ≤
Showing a concentration bound for such a quadratic function will be subject of the rest of the
i , it cancels out. Now, writing cj = 1 − zj (which now ∈ [0, 2]) and using
A − B, to bound it by
w∗
ijzj −
i , we can bound the above by
Yi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)ed∗
Xi
i − Xj∈Γ(i)∩S ′
i −eρ∗
By the definition of eρ∗
the definition of ed∗
Yi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i ∆2 cj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Yi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Γ(i)
Xi
f (Y ) := f (Y1, . . . , Yn) :=Xi
section. Let us define
=Xi
Xj∈Γ(i)
ijcj −
1
p∗
Yjw∗
.
zj(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
p∗
i ∆2(cid:0)di − ρi − Xj∈Γ(i)
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
w∗
ijcj(Yj − p∗
Yi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Γ(i)
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(using w∗
ij =
1
j ∆2 )
p∗
i p∗
w∗
ijcj(Yj − p∗
.
(12)
We wish to show that Pr[f > εn
∆ ] ≤ exp(−S). Unfortunately, this is not true – there are
counter-examples (in which the neighborhoods of vertices have significant overlaps) for which it is
not possible to obtain a tail bound better than exp(−n/∆2), roughly speaking. To remedy this, we
resort to a trick developed in [22, 16]. The key idea is to condition on the event that vertices have
a small weighted degree into the set S′, and obtain a stronger tail bound.
"Good" conditioning. We say that a choice of Y 's is good if for all i ∈ V , we have
Xj∈Γ(i)
w∗
ijYj ≤
ε∆ + 2di
p∗
i ∆2
.
The first lemma is the following.
Lemma 6.5. Let H be the weighted graph on S′ obtained by our algorithm. For any vertex i ∈ V ,
we have
Yj
p∗
j(cid:17). The term in the paren-
ε , together with
j ≤ αε∆
1
p∗
Pr Xj∈Γ(i)
w∗
ijYj >
ε∆ + 2di
p∗
1
n6 .
i ∆2 <
i ∆2(cid:16)Pj∈Γ(i)
ijYj = 1
p∗
thesis has expectation precisely di. Thus, applying Bernstein using maxj
Proof. Fix some i ∈ V , and consider Pj∈Γ(i) w∗
Pj∈Γ(i)
j )2 ≤ di maxj
j (1−p∗
p∗
j )
(p∗
, we have
1
p∗
j
Yj
p∗
j
Pr(cid:2) Xj∈Γ(i)∩VH
exp(cid:18)−
> di + t(cid:3) ≤ exp(cid:18)−
(2di + ε∆)αε∆(cid:19) ≤ exp(cid:18)−
ε(di + ε∆)2
εt2
(di + t)αε∆(cid:19) .
ε2
2αε(cid:19) <
1
n6 .
By setting t = (di + ε∆), the RHS above can be bounded by
This completes the proof, using our choice of αε.
15
Conditioning on the Y being good, we show the following concentration theorem.
Theorem 6.6. Let Yi's be independent random variables, that are 1 w.p. p∗
let f (Y ) be defined as in (12). Then we have
i and 0 otherwise, and
Pr(cid:2)f (Y ) ≥
εn
8∆ (cid:12)(cid:12) Y is good(cid:3) ≤
1
n5 · e−20n log n/ε2
.
We observe that the theorem implies Lemma 6.4. This is because by the Theorem and the
, for any A ⊆ S. Thus
n5 . Since the
H − DualG ≤ εn/∆ Y good] ≥ 1 − exp(−S)
H − maxA DualG ≤ εn/∆ Y good] ≥ 1 − 1
preceeding discussions, Pr[Dual∗
by union bound over A, Pr[maxA Dual∗
probability of the good event is at least 1 − 1
6.2 Concentration bound for quadratic functions
n5 (by Lemma 6.5), the desired conclusion follows.
n5
To conclude our proof, it suffices to show Theorem 6.6. To bound the quadratic function f , we
bound the moment generating function (MGF), E[eλf good]. This is done via a decoupling
argument, a standard tool for dealing with quadratic functions. While decoupling is immediate for
'standard' quadratic forms, the proof also works for our f (which has additional absolute values).
The rest of the proof has the following outline.
Proof outline. The main challenge is the computation of the MGF under conditioning (which
introduces dependencies among the Yi, albeit mild ones). The decoupling allows us to partition
vertices into two sets, and only consider edges that go across the sets. We then show that it suffices
to bound the MGF under a "weakened" notion of conditioning (a property we call δ-good). Under
this condition, all the vertices in one of the sets of the partition become independent, thus allowing
a bound on the moment - in terms of quantities that depend on the variables on the other set of
the partition. Finally, we appeal to a strong concentration bound of Boucheron et al. [13] to obtain
an overall bound, completing the proof.
We now expand the proof outline above.
Decoupling. Consider independent Bernoulli random variables δi that take values 0 and 1 w.p.
1/2 each, and consider the function
fδ :=Xi
δiYi Xj∈Γ(i)
(1 − δj)w∗
ijcj(Yj − p∗
j )
Using the fact that E[g(x)] ≥ E[g(x)] for any function g, and defining Eδ as the expectation with
respect to the δi's, we have
(1 − δj)w∗
ijcj(Yj − p∗
j ) =Xi
1
2 · YiEδ Xj∈Γ(i)
(1 − δj)w∗
ijcj(Yj − p∗
j )
1
Eδfδ = EδXi
≥Xi
=Xi
δiYi Xj∈Γ(i)
2 · Yi Xj∈Γ(i)
4 · Yi Xj∈Γ(i)
1
Eδ(1 − δj)w∗
ijcj(Yj − p∗
j )
ijcj(Yj − p∗
w∗
j ) =
1
4
f
We used the fact that i never appears in the summation term involving Yi to obtain the first
equality. Next, using Jensen's inequality, we have:
EY [eλf good] ≤ EY [e4λEδ fδ good] ≤ EY,δ[e4λfδ good]
16
where EY,δ means the expectation with respect to both random variables Y and δ. Now, the
interpretation of fδ is simply the following. Consider the partitioning (V +, V −) of V defined by
V + = {i ∈ [n] : δi = 1} and V − = {i ∈ [n] : δi = 0}, then
ijcj(Yj − p∗
w∗
Yi(cid:12)(cid:12) Xj∈Γ(i)∩V −
For convenience, define Ri = Pj∈Γ(i)∩V − w∗
j ), for i ∈ V +. Thus we can write
fδ = Pi∈V + YiRi. The condition that Y is good now gives us a bound on Ri. For any cj (it is
important to note that the good condition does not involve the constants cj, as those depend on
the LP solution; all we know is that 0 ≤ cj ≤ 2), we have
ij(Yj + p∗
fδ = Xi∈V +
ijcj(Yj − p∗
j )(cid:12)(cid:12).
2w∗
Ri ≤(cid:12)(cid:12) Xj∈Γ(i)∩V −
j )(cid:12)(cid:12)
2di
i ∆2 ≤
p∗
Now, the quantity we wish to bound can be written as
2(ε∆ + 2di)
p∗
i ∆2
≤
+
2ε∆ + 6di
.
p∗
i ∆2
[eλf good] ≤ E
Y,δ
E
Y
[e4λfδ good] ≤ E
δ
E
Y −
E
Y +
[e4λ Pi∈V + YiRi good].
(13)
The key advantage that decoupling gives us is that we can now integrate over Yi ∈ V +, i.e., evaluate
the innermost expectation, for any given choice of {Yi : i ∈ V −} (which define the Ri). The problem
with doing this in our case is that the good condition introduces dependencies on the Yi, for i ∈ V +.
Fortunately, weakening conditioning does not hurt much in computing expectations. This is
captured by the following simple lemma.
Lemma 6.7. Let Ω be a space with a probability measure µ. Let Q1 and Q2 be any two events such
that Q1 ⊂ Q2, and let Z : Ω 7→ R+ be a non-negative random variable. Then,
E[ZQ1] ≤
E[ZQ2]
Pr[Q1]
.
Proof. Let Ω1 (resp. Ω2) be the subset of Ω in which Q1 (resp. Q2) is satisfied. By hypothesis,
Ω1 ⊆ Ω2. Now by the definition of conditional expectation, and the non-negativity of Z, we have
E[XQ1] =
Z(x)µ(x)dx ≤
1
µ(Q1)Zx∈Ω1
1
µ(Q1)Zx∈Ω2
Z(x)µ(x)dx =
µ(Q2)
µ(Q1)
E[ZQ2].
Since µ(Q2) ≤ 1, the conclusion follows.
Weaker good property.
The next crucial notion we define is a property "δ-good". Given
a δ ∈ {0, 1}n (and corresponding partition (V +, V −)), a set of random variables Y is said to be
δ-good if for all i ∈ V +, we have
Xj∈Γ(i)∩V −
w∗
ijYj ≤
ε∆ + 2di
p∗
i ∆2
.
(14)
We make two observations. First, the good property implies the δ-good property, for any
choice of δ. Second, and more crucial to our proof, conditioning on δ-good does not introduce any
17
dependencies on the variables {Yi : i ∈ V +}. Now, continuing from (13), and using the fact that
the good condition holds with probability > 1/2, we have
E
δ
E
Y −
E
Y +
[e4λ Pi∈V + YiRi good] ≤ E
δ
E
Y −
E
Y +
[2e4λ Pi∈V + YiRi δ-good].
E
Now for any 0/1 choices for variables Y −, the Ri's get fixed for every i ∈ V +, and we can bound
Y +[e4λ Pi∈V + YiRi Ri] easily.
Lemma 6.8. Let Yi be independent random 0/1 variables taking value 1 w.p. p∗
given, for i ∈ V +. Suppose λ > 0 satisfies λRi ≤ 1 for all i. Then
i R2
i Ri+λ2p∗
i .
i , and let Ri be
E
Y +[eλ Pi∈V + YiRi] ≤ ePi∈V + λp∗
Proof. Since the lemma only deals with i ∈ V +, we drop the subscript for the summations and
expectations. One simple fact we use is that for a random variable Z with Z ≤ 1,
Using this, and the independence of Yi together with Y 2
E[eZ ] ≤ E[1 + Z + Z 2] ≤ eE[Z]+E[Z 2].
E[eλ P YiRi] =Y E[eλYiRi] ≤Y eE[λYiRi]+E[λ2R2
i = Yi,
i Yi].
As E[Yi] = p∗
i , this completes the proof of the lemma.
Using the lemma, replacing λ with 4λ yields the following
E
δ
E
Y −
E
Y +
[e4λ Pi∈V + YiRi δ-good] ≤ E
δ
The second term in the summation is already small enough. I.e., using (14)
i Ri+16λ2p∗
i R2
i δ-good(cid:3).
E
Y −(cid:2)ePi∈V + 4λp∗
i ≤ 2λ2αε
n
∆
.
λ2p∗
i R2
Xi∈V +
(15)
(16)
While the bound on Ri can be used to bound the first term, it turns out that this is not good
enough. We thus need a more involved argument. Thus the focus is now to bound
E
δ
E
Y −(cid:2)eλg(Y ) δ-good(cid:3), where g(Y ) := Xi∈V +
Xj∈Γ(i)∩V −
w∗
ijcj(Yj − p∗
.
p∗
i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Outline: concentration bound for g. To deduce a concentration bound for g, we first remove
the conditioning (again appealing to Lemma 6.7). This then gives us independence for the Yj, for
j ∈ V −. We can then appeal to the fact that changing a Yj only changes g by a small amount,
to argue concentration. However, the standard "bounded differences" concentration bound ([13],
Theorem ...) will not suffice for our purpose, and we need more sophisticated results [13] (restated
as Theorem 3.2).
To use the same notation as the theorem, define Z = g(Y ), where we only consider Yr, r ∈ V −.
Now, for any such r, consider Z − Z (r). Since Z (r) is obtained by replacing Yr by an independent
Y ′
r and re-computing g, we can see that the only terms i which could possibly be affected are
i ∈ Γ(r) ∩ V +. Further, we can bound the difference Z − Z (r) by
i w∗
2p∗
ir,
Z − Z (r) ≤ Yr − Y ′
r Xi∈Γ(r)∩V +
18
where we have used cr ≤ 2. The summation can be bounded by 2dr
Then, to use the theorem, we need
r ∆2 . Denote this quantity by θr.
p∗
EY ′
rZ − Z (r)2 ≤ Yr − Y ′
r2θ2
r ≤ Yr − Y ′
rθ2
r ≤ Yrθ2
r + p∗
rθ2
r .
We used the fact that Yr − Y ′
r ∈ {0, 1}. Now applying Theorem 3.2 by setting θ = 1
V −[eλ(g−E[g])] ≤ E
λ2
2 Pr∈V − Yrθ2
V −[e
rθ2
r ]
r +p∗
E
2λ , we have:
Once again, since λθr will turn out to be < 1, we can use the bound E[eλ2Yrθ2
conclude that
r /2] ≤ eλ2θ2
r p∗
r , and
E
V −[eλ(g−E[g])] ≤ e2λ2 Pr∈V − p∗
rθ2
r ≤ e4λ2αε
n
∆ .
The last inequality is due to a reasoning similar to earlier.
We are nearly done. The only step that remains for proving Theorem 6.6 is to obtain a bound
on E[g]. For this, we need to bound, for any i, the term
By Cauchy-Schwartz and the fact that Yj are independent, we have
w∗
ijcj(Yj − p∗
j )].
E[Ri] = E[ Xj∈Γ(i)∩V −
i ] = Xj∈Γ(i)∩V −
1 − p∗
i )2∆4 Xj∈Γ(i)
j ≤
p∗
(w∗
(p∗
≤
1
j
E[Ri]2 ≤ E[R2
ij)2cjp∗
j (1 − p∗
j )
αεdi
i )2∆3 .
ε(p∗
j ≤ αε∆/ε. Thus we have
d1/2
i ≤(cid:16) αε
ε∆3(cid:17)1/2
n∆1/2 ≤
n
∆(cid:16) αε
ε (cid:17)1/2
.
To complete the bound, we end up setting λ = ε
αε
. For this value of λ, we must ensure that
ε
αε
(dr + ε∆)
r∆2 ≤ 1,
p∗
which is indeed true.
7 Sparse Core-set for Max-Cut
In Theorem 4.2, we have shown that there is a core-set (i.e., a smaller weighted graph with the
same MAXCUT value) with a small number of vertices. We now show that the number of edges
can also be made small (roughly n/∆). This will prove Theorem 4.3.
We start with a lemma about sampling edges in graphs with edge weights ≤ 1. (We note that
essentially the same lemma is used in several works on sparsifiers for cuts.)
19
p∗
j )/p∗
E[Xi∈V +
We have used the fact that (1 − p∗
iRi] ≤(cid:16) αε
4 · εn
∆ .
≤ exp(cid:16)λ
ε∆3(cid:17)1/2Xi
From our choice of αε, this is < 1
εn
4∆
+ λ2αε
n
∆(cid:17) .
Putting everything together, we get that the desired moment
Lemma 7.1. Consider the weighted graph H resulted from vertex sample, defined on set of
vertices S′ in which wij < 1. If we apply edge sample on H, then the resulted graph H ′ have
with probability at least 1 − 1
n2 .
MaxCut(H ′) ± (1 + ǫ)MaxCut(H)
Let us first see that the lemma gives us Theorem 4.3.
Proof of Theorem 4.3. We only need to verify the bound on the number of edges. As every edge
is sampled with probability pij = min(1, 8wij
ε2 ), and since the total edge weight is normalized to be
S′, we have that the expected number of edges is ≤ 8S ′
n4 , this is at most
ε2 , and w.p. at least 1 − 1
16S ′
ε2
, completing the proof.
7.1 Proof of Edge Sampling
We now prove Lemma 7.1.
Proof. Recall that in edge sample algorithm we first rescale the edge weights so that they sum
up to S′, and then sample each edge ij in the resulted graph (also denoted H, as we can assume it
to
to be a pre-processing) with probability pij = min(wij
obtain the graph H ′. Define the indicator variable Xij for each edge eij in graph H, where Xij = 1
if the corresponding edge eij is selected by edge sample and Xij = 0 otherwise.
ǫ2 , 1) and reweigh the edges to w′
ij = wij
pij
8
Our goal is to show that all cuts are preserved, w.h.p. Consider any cut (A, B) in H. Call the
A,B.
set of all the edges on this cut as CA,B and the set of in the cut on the sampled graph H ′ as C ′
wij and w′(C ′
w′
ij. Then we have,
A,B)] = E[X w′
ijXij] =X w′
ij
Pr(Xij = 1) =X pijw′
ij =X wij = w(CA,B).
We will now apply Bernstein's inequality to bound the deviation. For this, the variance is first
bounded as follows.
Var[X w′
ijXij] =X w′
ij
2
Var(Xij) ≤X w2
ij
pij
(1 − pij) ≤
ǫ2
8 X wij =
ǫ2
8
w(CA,B)
We used the inequality that unless pij = 1 (in which case the term drops out), we have wij/pij ≤
ε2/8.
By the observation on wij/pij above, we can use Bernstein's inequality 3.1 with b = ε2/8, to
Set w(CA,B) =Peij ∈CA,B
E[w′(C ′
A,B) =Peij ∈C ′
A,B
obtain
Pr[X w′
ijXij − w(CA,B) ≥ t] ≤ exp −
t2
ε2w(CA,B )
4
8 ! .
+ tε2
Setting t = εW , where W is the sum of all the edge weights (which is equal to S′ after the
pre-processing), the bound above simplifies to exp(−2S′), and thus we can take a union bound
over all cuts. This completes the proof.
8 A 2-pass streaming algorithm
We now show how our main core set result can be used to design a streaming algorithm for MaxCut.
The algorithm works in two passes: the first pass builds a core-set S of size eO(n/∆) as prescribed
by Theorem 4.2 and the second pass builds the induced weighted graph G[S] and computes its max
cut. This algorithm works under edge insertion/deletion.
20
8.1 Pass 1: Building a core set
To construct S, Theorem 4.2 states that each vertex must be sampled with probability pi, where
pi ≥ hi and hi = min(1, max(di,ǫ∆)
) is the importance score of a vertex. As the goal is to only choose
a small number of vertices, we will also make sure that pi ≤ 2hi. The challenge here is two-fold:
we need to sample (roughly) proportional to the degree di, which can only be computed after the
stream has passed, and we also need the actual value of pi (or a close enough estimate of it) in
order to correctly reweight edges in the second pass.
∆2αǫ
The degree di of a vertex i is the "count" of the number of times i appears in the edge stream.
To sample with probability proportional to di we will therefore make use of streaming algorithms
for ℓ1-sampling [32, 8, 23]. We borrow some notation from [8].
Definition 8.1. Let ρ > 0, f ∈ [1, 2]. A (ρ, f )-approximator to τ > 0 is a quantity τ such that
τ /f − ρ ≤ τ ≤ f τ + ρ
Lemma 8.2 ([23] (rephrased from [8])). Given a vector x ∈ Rn and parameters ǫ, δ > 0, c > 0
there exists an algorithm A that uses space O(log(1/ǫ)ǫ−1 log2 n log(1/δ)) and generates a pair (i, v)
from a distribution Dx on [1 . . . n] such that with probability 1 − δ
• Dx(i) is a ( 1
• v is a (0, 1 + ǫ)-approximator to xi
nc , 1 + ǫ)-approximator to xi/kxk1
where c is a fixed constant.
We will also need to maintain heavy hitters: all vertices of degree at least ∆2 (up to constants).
To do this, we will make use of the standard CountMin sketch [14]. For completeness, we state
its properties here.
Lemma 8.3 ([14]). Fix parameters k, δ > 0. Then given a stream of m updates to a vector x ∈ Rn
there is a sketch CM of size O(k log δ−1(log m + log n)) and a reconstruction procedure f : [n] → R
such that with probability 1 − δ, for any xi, xi − f (i) ≤ kxk1/k
Outline. We will have a collection of r, roughly n/∆ ℓ1-samplers. These samplers will together
give a good estimate ((1 + ε)-approximation) of the importance hi for all the vertices that have a
small degree (which we define to be < αε∆2). Then, we use the CM sketch to maintain the degrees
of all the 'high degree' vertices, i.e., those with degrees ≥ αε∆2. Taken together, we obtain the
desired sampling in the first pass.
Definition 8.4. Given two sets of pairs of numbers S, S′, let S∪maxS′ = {(x, max(x′,y)∈S∪S ′,x′=x y)}
Lemma 8.5. Let S = {(i, vi)} be the set returned by Algorithm 1. Then
∆(cid:1).
• S has size eO(cid:0) n
• Each i ∈ [n] is sampled with probability pi that is (0, 1 + ǫ) approximated by vi and that
(n−c, 1 + ǫ)-approximates hi.
Proof sketch. Consider any vertex i with di ≥ ∆2αǫ. By Lemma 8.3, such a vertex will report a
count of at least f (i) = (1− ζ)∆2αǫ and thus is guaranteed to be included in Sh. Its reported score
vi = 1 satisfies the requirement of the Lemma. Secondly, consider any vertex with degree di < ǫ∆.
For such a vertex, hi = ǫ/∆αǫ and thus it is included in Sl with the desired probability and vi.
Finally, consider a vertex i with ǫ∆ ≤ di < ∆2αǫ. The probability that none of the ℓ1-samplers
yield i is (1− di/n∆)r, and since di/n∆ ≪ 1, this can be approximated as (1− rdi/n∆). Thus, the
probability of seeing i is rdi/n∆ = di/∆2αǫ as desired.
Corollary 8.5.1. For each (i, vi) ∈ S, hi ≤ vi ≤ 2hi.
21
Algorithm 1 Given n and average degree ∆
Initialize Sl, Sm, Sh ← ∅, and ζ = αε.
Sample elements from [1 . . . n] each with probability ǫ/∆αǫ. For each sampled i add (i, ǫ/∆αǫ)
to Sl.
Fix ζ > 0. Initialize a CountMin sketch CM with size parameter k = n/∆ζ 2. Let f be the
associated reconstruction procedure.
Initialize r = O(n/∆αǫ) copies A1 . . . Ar of the algorithm A from Lemma 8.2.
for each stream update (i, w) (a vertex to which current edge is incident, and weight) do
Update each Aj, 1 ≤ j ≤ r.
Update CM.
end for
for j = 1 to r do
Sample (i, v) from Aj. Sm = Sm ∪max {(i, v)}
end for
Sh = {(i, 1) f (i) ≥ (1 − ζ)∆2αǫ}
return Sl ∪max Sm ∪max Sh
8.2 Pass 2: Building the induced weighted graph
The first pass produces a set S of eO(n/∆) vertices together with estimates vi for their importance
score hi. If we weight each edge ij in G[S] by wij = 1/vivj∆2, Theorem 4.2, along with Corol-
lary 8.5.1 guarantee that a MaxCut in the resulting weighted graph is a good approximation of
the true max cut.
Thus, knowing S, constructing the re-weighted G[S] in the second pass is trivial if we had space
proportional to the number of edges in G[S]. Unfortunately this can be quadratic in S, so our
goal is to implement the edge sampling of Theorem 4.3 in the second pass. This is done as follows:
we maintain a set of edges E′. Every time we encounter an edge ij with i, j ∈ S, we check to see if
it is already in E′. If not, we toss a coin and with probability pij = min(1, wij log n/ǫ2) we insert
(i, j, wij /pij) into E′. By Lemma 7.1, the size of E′ is eO(n/∆), and the resulting graph yields a
(1 + ǫ) approximation to the MaxCut.
9 Correlation Clustering
Our argument for correlation clustering parallels the one for MaxCut. The MAX-AGREE variant
of correlation clustering, while not a CSP (as the number of clusters is arbitrary), almost behaves
as one.
We start with two simple observations. The first is that we can restrict the number of clusters
to 1/ε, for the purposes of a (1 + ε) approximation (Lemma 9.1) . Next, observe that the optimum
objective value is at least max{C +, C −} ≥ n∆/2. This is simply because placing all the vertices
in a single cluster gives a value C +, while placing them all in different clusters gives C −. Thus, it
suffices to focus on additive approximation of εn∆.
Lemma 9.1. Let C be the optimal clustering, and let OPT be its max-agree cost. Then there exists
a clustering C′ that has cost ≥ (1 − O(ε))OP T , and has at most 1/ε clusters.
The lemma is folklore in the correlation clustering literature.
Proof. Let A1, A2, . . . , Ak be the clusters in the optimal clustering C. Now, suppose we randomly
color the clusters with t = 1/ε colors, i.e., each cluster Ai is colored with a random color in [t].
22
We then merge all the clusters of a given color into one cluster, thus obtaining the clustering C′.
Clearly, C′ has at most t colors.
Now, we observe that if u, v ∈ Ai to begin with, then u, v are still in the same cluster in C′.
But if u ∈ Ai and v ∈ Aj, and the colors of Ai and Aj are the same, then u and v are no longer
separated in C′. Let us use this to see what happens to the objective. Let χuv be an indicator for
u, v being in the same cluster in the optimal clustering C, and let χ′
uv be a similar indicator in C′.
The original objective is
C − +Xij
ηijχij.
From the above reasoning, if χuv = 1, then χ′
uv] = 1/t, i.e., there
is a probability precisely 1/t = ε that the clusters containing u, v now get the same color. Thus,
we can write the expected value of the new objective as
uv = 1. Also, if χuv = 0, E[χ′
C − +Xij
ηijχij + εXij
ηij(1 − χij).
Let us denote S =Pij ηijχij and S′ =Pij ηij(1 − χij). Then by definition, we have S + S′ =
Pij ηij = C + − C −. Now, to show that we have a (1 − ε) approximation, we need to show that
C − + S + εS′ ≥ (1 − ε)(C − + S). This simplifies to C − + S + S′ ≥ 0, or equivalently, C + ≥ 0,
which is clearly true.
This completes the proof.
Once we fix the number of clusters k, we can write correlation clustering as a quadratic program
in a natural way: for each vertex i, have k variables xiℓ, which is supposed to indicate if i is given
the label ℓ. We then have the constraint thatPℓ xiℓ = 1 for all i. The objective function then has
a clean form:
Xij
[
kXℓ=1
xiℓ(1 − xjℓ)c−
ij + xiℓxjℓc+
ij] =Xij Xℓ
xiℓc−
xiℓ(ρiℓ + d−
i ),
ij + xiℓxjℓηij =Xi,ℓ
i =Pj c−
ij.
where xiℓ = 1 iff vertex i ∈ Cℓ, and ρiℓ =Pj∈Γ(i) xjℓηij and d−
Note the similarity with the program for MaxCut. We will show that the framework from
Section 4.1 carries over with minor changes. The details of the new Est procedure can be found
in Section 9.1 (it requires one key change: we now need to consider k-partitions of the seed set in
order to find ρ). The duality based proof is slightly more involved; however we can use the same
rough outline. The proof is presented in Section 9.2.
9.1 LP Estimation Procedure for Correlation Clustering
Let H = (V, E, c) be a weighted, undirected graph with edge weights c+
ij as before, and let
γ : V → [0, 1] denote sampling probabilities. We define EstC(H, γ) to be the output of the following
randomized algorithm: sample a set S by including each vertex i in it w.p. γi independently; next,
for each partition (A1,··· , Ak) of S, solve the LP defined below, and output the largest objective
value.
ij and c−
23
LPA1,··· ,Ak (V ) is the following linear program. (As before, we use constants ρiℓ :=Pj∈Γ(i)∩Aℓ
ηij
γj
.)
xiℓ(ρiℓ + d−
maximize Xi
subject to ρiℓ − tiℓ ≤ Xj∈Γ(i)
i ) − (siℓ + tiℓ)
ηijxjℓ ≤ ρiℓ + siℓ ∀i, ℓ
Xℓ
xiℓ = 1 ∀i ∈ [n]
siℓ, tiℓ ≥ 0 ∀i, ℓ
Once again, the best choice of siℓ, tiℓ for each pair i, ℓ are so that siℓ + tiℓ = ρiℓ −Pj∈Γ(i) ηijxjℓ.
We now show a result analogous to the one earlier – that under appropriate conditions, EstC
is approximately equal to the optimal correlation clustering objective value.
Theorem 9.2. Let H be a weighted graph on n vertices with edge weights c+
Suppose the sampling probabilities γi satisfy the condition
ij, c−
ij that add up to W .
wij ≤
W ε2
8k2 log n
γiγjPu γu
for all i, j.
(17)
Then, we have EstC(H, γ) ∈ CC(H)± εW , with probability at least 1− 1/n2 (where the probability
is over the random choice of S).
Note that the only difference is the k2 term in (17).
Proof. As before, the proof follows from two complementary claims.
Claim 1. W.h.p. over the choice of S, there exists a partitioning (A1,··· , Ak) of S such that
LPA1,··· ,Ak (H) ≥ CC(H) − εW .
Claim 2. Consider any feasible solution to the LP above (for some values ρiℓ, siℓ, tiℓ). There exists
a partitioning in H of objective value at least the LP objective.
The proof of Claim 1 mimics the proof in the case of MAXCUT. We use Bernstein's inequality
for every ℓ ∈ [k] with deviation being bounded by ε(di + ∆)/k in each term. This is why we need
an extra k2 term in the denominator of (17). We omit the details.
Proof of Claim 2. Suppose we have a feasible solution x to the LP of objective valuePi,ℓ xiℓ(d−
ρiℓ) −Pi,ℓ ρiℓ −Pj∈Γ(i) ηijxjℓ, and we wish to move to a partitioning of at least this value. To
this end, define the quadratic form
i +
Q(x) :=Xi,ℓ
xiℓ(cid:0)d−
i + Xj∈Γ(i)
ηijxjℓ(cid:1).
The first observation is that for any x ∈ [0, 1]nk, and any real numbers ρi, we have
xiℓ(d−
Q(x) ≥Xi,ℓ
ρiℓ − Xj∈Γ(i)
This is true simply because Q(x) =Pi,ℓ xiℓ(cid:0)d−
the second term is at least −ρiℓ −Pj∈Γ(i) ηijxjℓ, as xi ∈ [0, 1].
i + ρiℓ) −Xi,ℓ
ηijxjℓ.
i + ρiℓ(cid:1) − xiℓ(cid:0)ρiℓ −Pj∈Γ(i) ηijxjℓ(cid:1), and the fact that
Next, note that the maximum of the form Q(x) over [0, 1]nk has to occur at a boundary point,
since for any fixing of variables other than the ith group of variables xi1,··· , xik for a given i, the
24
form reduces to a linear function of xiℓ, 1 ≤ ℓ ≤ k, which attains maximum at one of the boundaries
when subject to the constraint Pℓ xiℓ = 1. Using this observation repeatedly for i ∈ [n] lets us
conclude that there is a y ∈ {0, 1}nk such that Q(y) ≥ Q(x). Since any such y corresponds to a
partitioning, and Q(y) corresponds to its objective value, the claim follows.
This completes the proof of Theorem 9.2.
As in the case of MAXCUT, we can show that the estimation procedure can be used for
estimating both in the original graph (with uniform probabilities qi), and with the graph H, with
sampling probabilities qi/pi. We thus skip stating these claims formally.
9.2
Induced Linear Programs for Correlation Clustering
We next need to prove that the EstC procedures have approximately the same values on G and H
(with appropriate γ's). To show this, we consider a sample (S, S′) drawn as before, and show that
max
(A1,··· ,Ak):S
LP γ
A1,··· ,Ak
(V ) ≥ ∆2 max
(A1,··· ,Ak):S
LP α
A1,··· ,Ak(S′) − εn∆,
(18)
where γi = qi and αi = qi/pi. As before, we consider the duals of the two programs. This is the
main place in which our correlation clustering analysis differs from the one for MAXCUT. The dual
is as follows
subject to ui + Xj∈Γ(i)
minimize Xi
ηjℓzjℓ ≥ d−
ui +Xi,ℓ
i + ρiℓ ∀i, ℓ
ρiℓziℓ
−1 ≤ ziℓ ≤ 1 ∀i ∈ [n], ℓ ∈ [k]
The difference now is that for any vector z, the optimal choice of ui
i + ρiℓ −
Pj∈Γ(i) ηjℓzjℓ}. This is now a maximum of k terms, as opposed to the max of 0 and one other term
in the case of MAXCUT. But once again, we can think of the dual solution as being the vector z,
and we again have ui ≤ 2di. The dual of LP α
the vertices in S′) as the solution to the dual on H. The objective values are now as follows.
As we did earlier, we take a solution z to the dual of the LP on G, and use the same values (for
(H) can be written down similarly.
A1,··· ,Ak
is maxℓ{d−
DualG =Xi,ℓ
ρiℓziℓ +Xi
DualH ≤ Xi∈S ′,ℓeρiℓziℓ +Xi
i + ρiℓ − Xj∈Γ(i)
ℓ {d−
max
i +eρiℓ − Xj∈Γ(i)∩S ′
ℓ { ed−
max
ηijzjℓ}
pi∆2 . We now show that w.p. at least 1 − 1
n4 ,
εn
∆
1
∆2 max
DualH ≤
DualG +
(A1,...,Ak):S
(A1,...,Ak):S
max
.
wijηijzjℓ}
Here also, we have eρiℓ = ρiℓ
(19)
(20)
(21)
We can use the same trick as in the MAXCUT case, and move to Dual∗
iℓ := ρiℓ
show that w.h.p. (assuming S satisfies conditions analogous to Lemma 6.4),
H, in which we use
ij as before. The proof of Lemma 6.2 applies verbaitm. Thus it suffices to
pi∆2 , weights w∗
eρ∗
max
(A1,...,Ak):S
Dual∗
H ≤
1
∆2 max
(A1,...,Ak):S
DualG +
εn
2∆
.
25
Consider the expression
Dual∗
H −
1
∆2
DualG =Xi Xℓ
(Yieρ∗
1
∆2 · ρiℓziℓ)! +(cid:18)Yi u∗
i −
1
∆2 · ui(cid:19) .
(22)
iℓziℓ −
We view this as two summations (shown by the parentheses), and bound them separately. The
first is relatively easy. We observe that by definition,
iℓ = Xj∈Γ(i)∩Cℓ
eρ∗
w∗
ijηijp∗
j
qj
=
ηij
p∗
i ∆2 Xj∈Γ(i)∩Cℓ
1
qi
=
This implies that we can write the first term asPiℓ
via Bernstein's inequality.
ρiℓziℓ
p∗
i ∆2(cid:0)Yi − p∗
(23)
1
i ∆2 · ρiℓ.
p∗
i(cid:1). This will then be bounded
For bounding the second term, we start with the trick of splitting it into two terms by adding
a "hybrid" term, as follows:
Xi
Yi u∗
i −
1
∆2 · ui =Xi (cid:18)Yi u∗
i − Yi
ui
i ∆2(cid:19) +Xi (cid:18)Yi
p∗
ui
i ∆2 −
p∗
1
∆2 · ui(cid:19) .
The second term can again be bounded using Bernstein's inequality, using the fact that ui =
O(di). Let us thus consider the first term. We can appeal to the fact max{P1, . . . , Pk} −
ηijw∗
ij(1 − zjℓ)(Yj − p∗
(24)
The nice thing now is that we can appeal (in a black-box manner, using the boundedness of η
and z) to the concentration bound for quadratic functions in Section 6.2, with ε replaced by ε/k,
to conclude the desired concentration inequality. For this goal, we again use a similar conditioning
as for MaxCut, which we state it here.
"Good" conditioning. We say that a choice of Y 's is good if for all i ∈ V , we have
Xj∈Γ(i)
w∗
ijηijYj ≤
ε∆ + 2di
p∗
i ∆2
26
max{Q1, . . . , Qk} ≤Pi Pi − Qi, to bound it by
Xi
YiXℓ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ed∗−
i +eρ∗
ij = 1
p∗
i p∗
iℓ − Xj∈Γ(i)∩S ′
w∗
ijηijzjℓ −
1
p∗
j ∆2 we can bound the above by
Using (23) and w∗
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Γ(i):ηij <0
YiXℓ
Xi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
YiXℓ
=Xi
Xj∈Γ(i):ηij <0
Yjηijw∗
ij(1 − zjℓ) − ηij
p∗
ηijw∗
ij(1 − zjℓ)(Yj − p∗
i ∆2 cjℓ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+Xi
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+Xi
This leads to a sum over ℓ of terms of the form
Xi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Γ(i):ηij 6=0
.
ηijzjℓ(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ηijw∗
ijzjℓ(Yj − p∗
ηijw∗
ijzjℓ(Yj − p∗
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i + ρiℓ − Xj∈Γ(i)
i ∆2(cid:0)d−
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj∈Γ(i):ηij >0
Xj∈Γ(i):ηij >0
j )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
YiXℓ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
YiXℓ
.
Lemma 9.3. Let H be the weighted graph obtained after sampling with probabilities p∗
vertex i ∈ V , we have
i . For any
Pr Xj∈Γ(i)
w∗
ijηijYj >
p∗
ε∆ + 2di
1
n4 .
i ∆2 <
i ∆2(cid:16)Pj∈Γ(i)
ijηijYj = 1
p∗
parenthesis has expectation precisely di. Thus, applying Bernstein using maxj
Proof. Fix some i ∈ V , and consider Pj∈Γ(i) w∗
withPj∈Γ(i)
≤ di maxj
j (1−p∗
j )
p∗2
j
, we have
ηij p∗
1
p∗
j
Yj ηij
p∗
j (cid:17). The term in the
ε , together
j ≤ αε∆
1
p∗
By setting t = (di + ε∆
k ), the RHS above can be bounded by
Pr(cid:2) Xj∈Γ(i)∩VH
ηijYj
p∗
j
exp −
ε(di + ε∆
(2di + ε∆
> di + t(cid:3) ≤ exp(cid:18)−
k )αε∆! ≤ exp(cid:18)−
k )2
εt2
(di + t)αε∆(cid:19) .
ε2
2αε(cid:19) <
1
n4 .
This completes the proof, using our choice of αε.
Conditioning on the Y being good, we can obtain the concentration bound for the quadratic
function (24).
This lets us take a union bound over all possible partitions of S (of which there are at most kn),
to obtain (21), which then completes the proof of the main result (Theorem 4.2), for correlation
clustering.
Acknowledgement
We thank Michael Kapralov for helpful discussions as we embarked upon this project.
References
[1] Pankaj K Agarwal, Sariel Har-Peled, and Kasturi R Varadarajan. Geometric approximation
via coresets. Combinatorial and computational geometry, 52:1–30, 2005.
[2] Kook Jin Ahn and Sudipto Guha. Graph sparsification in the semi-streaming model. In In-
ternational Colloquium on Automata, Languages, and Programming, pages 328–338. Springer,
2009.
[3] Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Graph sketches: sparsification, spanners,
and subgraphs. In Proceedings of the 31st symposium on Principles of Database Systems, pages
5–14. ACM, 2012.
[4] KookJin Ahn, Graham Cormode, Sudipto Guha, Andrew McGregor, and Anthony Wirth.
In International Conference on Machine Learning,
Correlation clustering in data streams.
pages 2237–2246, 2015.
27
[5] Nir Ailon and Zohar Karnin. A note on: No need to choose: How to get both a ptas and
sublinear query complexity. arXiv preprint arXiv:1204.6588, 2012.
[6] Noga Alon, W Fernandez De La Vega, Ravi Kannan, and Marek Karpinski. Random sampling
and approximation of max-csps. Journal of computer and system sciences, 67(2):212–243,
2003.
[7] Noga Alon, Eldar Fischer, Ilan Newman, and Asaf Shapira. A combinatorial characterization of
the testable graph properties: it's all about regularity. SIAM Journal on Computing, 39(1):143–
167, 2009.
[8] Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Streaming algorithms via pre-
cision sampling. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Sym-
posium on, pages 363–372. IEEE, 2011.
[9] Alexandr Andoni, Robert Krauthgamer, and David P Woodruff. The sketching complexity of
graph cuts. arXiv preprint arXiv:1403.7058, 2014.
[10] Sanjeev Arora, David Karger, and Marek Karpinski. Polynomial time approximation schemes
for dense instances of np-hard problems. In Proceedings of the twenty-seventh annual ACM
symposium on Theory of computing, pages 284–293. ACM, 1995.
[11] Nikhil Bansal, Avrim Blum, and Shuchi Chawla. Correlation clustering. Machine Learning,
56(1-3):89–113, 2004.
[12] Boaz Barak, Moritz Hardt, Thomas Holenstein, and David Steurer.
Subsampling
mathematical relaxations and average-case complexity.
the Twenty-
second Annual ACM-SIAM Symposium on Discrete Algorithms, SODA '11, pages 512–531,
Philadelphia, PA, USA, 2011. Society for Industrial and Applied Mathematics. URL:
http://dl.acm.org/citation.cfm?id=2133036.2133077.
In Proceedings of
[13] St´ephane Boucheron, G´abor Lugosi, and Pascal Massart. Concentration inequalities using the
entropy method. Annals of Probability, pages 1583–1614, 2003.
[14] Graham Cormode and Shan Muthukrishnan. An improved data stream summary: the count-
min sketch and its applications. Journal of Algorithms, 55(1):58–75, 2005.
[15] Devdatt P Dubhashi and Alessandro Panconesi. Concentration of measure for the analysis of
randomized algorithms. Cambridge University Press, 2009.
[16] Uriel Feige and Gideon Schechtman. On the optimality of the random hyperplane rounding
technique for max cut. Random Structures & Algorithms, 20(3):403–440, 2002.
[17] Dimitris Fotakis, Michael Lampis, and Vangelis Th Paschos. Sub-exponential approximation
schemes for csps: from dense to almost sparse. arXiv preprint arXiv:1507.04391, 2015.
[18] Alan Frieze and Ravi Kannan. The regularity lemma and approximation schemes for dense
problems. In Foundations of Computer Science, 1996. Proceedings., 37th Annual Symposium
on, pages 12–20. IEEE, 1996.
[19] Ioannis Giotis and Venkatesan Guruswami. Correlation clustering with a fixed number of clus-
ters. In Proceedings of the seventeenth annual ACM-SIAM symposium on Discrete algorithm,
pages 1167–1176. Society for Industrial and Applied Mathematics, 2006.
28
[20] Ashish Goel, Michael Kapralov, and Sanjeev Khanna. Graph sparsification via refinement
sampling. arXiv preprint arXiv:1004.4915, 2010.
[21] Ashish Goel, Michael Kapralov, and Ian Post. Single pass sparsification in the streaming model
with edge deletions. arXiv preprint arXiv:1203.4900, 2012.
[22] Oded Goldreich, Shari Goldwasser, and Dana Ron. Property testing and its connection to
learning and approximation. Journal of the ACM (JACM), 45(4):653–750, 1998.
[23] Hossein Jowhari, Mert Saglam, and G´abor Tardos. Tight bounds for lp samplers, finding
duplicates in streams, and related problems. In Proceedings of the thirtieth ACM SIGMOD-
SIGACT-SIGART symposium on Principles of database systems, pages 49–58. ACM, 2011.
[24] Michael Kapralov. Better bounds for matchings in the streaming model.
In SODA, pages
1679–1697. SIAM, 2013.
[25] Michael Kapralov, Sanjeev Khanna, and Madhu Sudan. Approximating matching size from
random streams. In SODA, pages 734–751. SIAM, 2014.
[26] Michael Kapralov, Sanjeev Khanna, and Madhu Sudan. Streaming lower bounds for approxi-
mating max-cut. In SODA, pages 1263–1282. SIAM, 2015.
[27] Michael Kapralov, Sanjeev Khanna, Madhu Sudan, and Ameya Velingker.
(1+ ω (1))-
approximation to max-cut requires linear space. In Proceedings of the Twenty-Eighth Annual
ACM-SIAM Symposium on Discrete Algorithms, pages 1703–1722. SIAM, 2017.
[28] Michael Kapralov, Yin Tat Lee, Cameron Musco, Christopher Musco, and Aaron Sidford.
Single pass spectral sparsification in dynamic streams. In Foundations of Computer Science
(FOCS), 2014 IEEE 55th Annual Symposium on, pages 561–570. IEEE, 2014.
[29] Jonathan A Kelner and Alex Levin. Spectral sparsification in the semi-streaming setting.
Theory of Computing Systems, 53(2):243–262, 2013.
[30] Dmitry Kogan and Robert Krauthgamer. Sketching cuts in graphs and hypergraphs. In Proc.
ITCS, pages 367–376. ACM, 2015.
[31] Claire Mathieu and Warren Schudy. Yet another algorithm for dense max cut: go greedy. In
Proceedings of the nineteenth annual ACM-SIAM symposium on Discrete algorithms, pages
176–182. Society for Industrial and Applied Mathematics, 2008.
[32] Morteza Monemizadeh and David P Woodruff. 1-pass relative-error lp-sampling with applica-
tions. In Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms,
pages 1143–1160. SIAM, 2010.
[33] Mark Rudelson and Roman Vershynin.
Sampling from large matrices:
J. ACM, 54(4), July 2007.
proach through geometric functional analysis.
http://doi.acm.org/10.1145/1255443.1255449, doi:10.1145/1255443.1255449.
An ap-
URL:
29
|
1606.00996 | 1 | 1606 | 2016-06-03T07:51:56 | A Minimal Variance Estimator for the Cardinality of Big Data Set Intersection | [
"cs.DS",
"cs.DB"
] | In recent years there has been a growing interest in developing "streaming algorithms" for efficient processing and querying of continuous data streams. These algorithms seek to provide accurate results while minimizing the required storage and the processing time, at the price of a small inaccuracy in their output. A fundamental query of interest is the intersection size of two big data streams. This problem arises in many different application areas, such as network monitoring, database systems, data integration and information retrieval. In this paper we develop a new algorithm for this problem, based on the Maximum Likelihood (ML) method. We show that this algorithm outperforms all known schemes and that it asymptotically achieves the optimal variance. | cs.DS | cs |
A Minimal Variance Estimator for the Cardinality of
Big Data Set Intersection
Reuven Cohen Liran Katzir Aviv Yehezkel
Department of Computer Science
Technion
Haifa 32000, Israel
Abstract
In recent years there has been a growing interest in developing "streaming algo-
rithms" for efficient processing and querying of continuous data streams. These algo-
rithms seek to provide accurate results while minimizing the required storage and the
processing time, at the price of a small inaccuracy in their output. A fundamental
query of interest is the intersection size of two big data streams. This problem arises
in many different application areas, such as network monitoring, database systems,
data integration and information retrieval. In this paper we develop a new algorithm
for this problem, based on the Maximum Likelihood (ML) method. We show that
this algorithm outperforms all known schemes and that it asymptotically achieves the
optimal variance.
1
Introduction
Classical processing algorithms for database management systems usually require several
passes over (static) data sets in order to produce an accurate answer to a user query. How-
ever, for a wide range of application domains, the data set is very large and is updated on
a continuous basis, making this approach impractical. For this reason, there is a growing
interest in developing "streaming algorithms" for efficient processing and querying of con-
tinuous data streams in data stream management systems (DSMSs). These algorithms seek
to provide accurate results while minimizing both the required storage and the processing
time per stream element, at the price of a small inaccuracy in their output [2, 7, 11, 21].
Streaming algorithms for DSMSs typically summarize the data stream using a small sketch,
and use probabilistic techniques in order to provide approximate answers to user queries.
Such big data streams appear in a wide variety of computer science applications. They are
common, for example, in computer networks, where detailed usage statistics (such as the
source IP addresses of packets) from different parts of the network need to be continuously
collected and analyzed for various security and management tasks.
A fundamental query of interest is the intersection size of two big data streams. Consider
two streams of elements, A and B, taken from two sets A and B respectively. Suppose that
each element may appear more than once in each stream. Let n =(cid:12)(cid:12)A∩ B(cid:12)(cid:12). For example, for
A = a, b, c, d, a, b and B = a, a, c, c, we get n = 2. Finding n is a problem that arises in many
different application areas such as network monitoring, database systems, data integration
and information retrieval [4, 5].
on-line detection of denial of service attacks.
As an application example, ai ∈ A and bj ∈ B could be streams of IP packets passing
through two routers, R1 and R2. In this case, A ∩ B represents the number of flows for-
warded by both routers. Thus, \A ∩ B can be used for security and traffic monitoring, e.g.,
As another example, ai ∈ A and bj ∈ B could be sets of sub-strings found within two
text documents. In this case, A ∩ B represents the number of sub-strings shared by both
documents, which can be viewed as the similarity of the documents. Among many other
applications, this allows plagiarism detection and pruning of near-duplicate search results in
search engines.
One can find the exact value of n by computing the intersection set C in the following
way. For every element bi ∈ B, compare bi to every aj ∈ A and ck ∈ C. If bi /∈ C and bi ∈ A,
add bi to C. After all the elements are treated, return the number of elements in C. This
naive approach does not scale if storage is limited or if the sets are very large. In these cases,
the following estimation problem should be solved. Given two streams of elements (with
repetitions) A = a1, a2, . . . , ap, and B = b1, b2, . . . , bq, such that A and B are the respective
where m ≪ n.
\ρ(A,B)
A ∩ B can be estimated in a straightforward manner using the inclusion-exclusion prin-
sets of the two streams, and n = A ∩ B, find an estimatebn of n using only m storage units,
ciple: \A ∩ B = cA+cB− \A ∪ B, taking advantage of the fact that estimating A, B and
A ∪ B is relatively easy. However, we will later show that this scheme produces inaccurate
results. In [2], it is proposed to estimate the Jaccard similarity, defined as ρ(A, B) = A∩B
A∪B .
The idea is to estimate A ∪ B, and then to extract A ∩ B. A third scheme, proposed in
[11], suggests that \A ∩ B =
ρ(A,B)+1 (cA +cB). With respect to the above three estimation
schemes, the main contributions of this paper are as follows:
1. For the first time, we present a complete analysis of the statistical performance (bias
and variance) of the above three schemes.
2. We find the optimal (minimum) variance of any unbiased set intersection estimator.
3. We present and analyze a new unbiased estimator, based on the Maximum Likelihood
(ML) method, which outperforms the above three schemes.
The rest of the paper is organized as follows. Section 2 discusses previous work and
presents the three previously known schemes. Section 3 presents our new Maximum Like-
lihood (ML) estimator. It also shows that the new scheme achieves optimal variance and
that it outperforms the three known schemes. Section 4 analyzes the statistical performance
(bias and variance) of the three known schemes. Section 5 presents simulation results con-
firming that the new ML estimator outperforms the three known schemes. Finally, Section
6 concludes the paper.
2 Related Work and Previous Schemes
The database research community has extensively explored the problem of data cleaning: de-
tecting and removing errors and inconsistencies from data to improve the quality of databases
[23].
Identifying which fields share similar values, identifying join paths, estimating join
directions and sizes, and detecting inclusion dependencies are well-studied aspects of this
problem [1, 8, 11, 17, 22]. For example, in [11] the authors present several methods for
finding related database fields. Their main idea is to hash the values of each field and keep
a small sketch that contains the minimal hash values for each. Then, the Jaccard similarity
is used to measure similarities between fields. In [1], the authors study the related problem
of detecting inclusion dependencies, i.e., pairs of fields A and B such that A ⊆ B, and they
present an efficient way to test all field pairs in parallel.
All the above problems are closely related to the "cardinality estimation problem" dis-
cussed in this paper. This problem has received a great deal of attention in the past decade
thanks to the growing number of important real-time "big data" applications, such as es-
timating the propagation rate of viruses, detecting DDoS attacks [14, 15], and measuring
general properties of network traffic [21].
Many works address the cardinality estimation problem [7, 9, 12, 16, 20, 21] and propose
statistical algorithms for solving it. These algorithms are usually limited to performing only
one pass on the received packets and using a fixed small amount of memory. A common
approach is to hash every element into a low-dimensional data sketch, which can be viewed
as a uniformly distributed random variable. Then, one of the following schemes is often used
to estimate the number of distinct elements in the set:
1. Order-statistics based estimators: In this family of schemes, the identities of the small-
est (or largest) k elements are remembered for the considered set. These values are
then used for estimating the total number of distinct elements [2, 9, 16, 20]. The family
of estimators with k = 1 (where the minimal/maximal identity is remembered) is also
known as min/max sketches.
2. Bit-pattern based estimators: In this family of schemes, the highest position of the
leftmost 1-bit in the binary representation of the identity of each element is remembered
and then used for the estimation [7, 12].
If only one hash function is used, the schemes estimate the value of n with an infinite
variance. To bound the variance, both schemes repeat the above procedures for m different
hash functions and use their combined statistics for the estimation1.
A comprehensive overview of different cardinality estimation techniques is given in [7, 21].
State-of-the art cardinality estimators have a standard error of about 1/√m, where m is the
number of storage units [6]. The best known cardinality estimator is the HyperLogLog
algorithm [12], which belongs to the family of min/max sketches and has a standard error
of 1.04/ √m, [12]. For instance, this algorithm estimates the cardinality of a set with 109
elements with a standard error of 2% using m = 2, 048 storage units.
Cardinality estimation algorithms can be used for estimating the cardinality of set inter-
section. As mentioned in Section 1, a straightforward technique is to estimate the intersection
1Stochastic averaging can be used to reduce the number of hash functions from m to only two [13].
using the following inclusion-exclusion principle:
\A ∩ B = cA + cB − \A ∪ B.
This method will be referred to as Scheme-1. Other algorithms first estimate the Jaccard
similarity ρ(A, B) = A∩B
A∪B , and then use some algebraic manipulation on it [2, 11]. Specif-
ically, the scheme proposed in [2] estimates both the Jaccard similarity and A ∪ B, and
then uses
\A ∩ B = \ρ(A, B) · \A ∪ B.
This scheme will be referred to as Scheme-2. The third scheme discussed in this paper,
referred to as Scheme-3, is presented in [11]. It estimates the Jaccard similarity, A, and
B, and then uses
\A ∩ B =
\ρ(A, B)
ρ(A, B) + 1
(cA + cB).
The above equation is obtained by substituting the Jaccard similarity definition into ρ(A,B)
which yields that
ρ(A,B)+1 ,
ρ(A, B)
ρ(A, B) + 1
=
A ∩ B
A ∩ B + A ∪ B
= A ∩ B
A + B
.
In [1, 18], the set intersection estimation problem is solved using smaller sample sets.
In
While these techniques are simple and unbiased, they are inaccurate for small sets.
addition, they are sensitive to the arrival order of the data, and to the repetition pattern.
3 A New Maximum Likelihood Scheme with Optimal
Variance
In this section we present a new unbiased estimator for the set intersection estimation prob-
lem. Because this estimator is based on the Maximum Likelihood (ML) method, it achieves
optimal variance and outperforms the three known schemes.
Maximum-Likelihood estimation (ML) is a method for estimating the parameters of a
statistical model. For example, suppose we are interested in the height distribution of a
given population, but are unable to measure the height of every single person. Assuming
that the heights are Gaussian distributed with some unknown mean and variance, the mean
and variance can be estimated using ML and only a small sample of the overall popula-
tion. In general, for a given set of data samples and an underlying statistical model, ML
finds the values of the model parameters that maximize the likelihood function, namely, the
"agreement" of the selected model with the given sample.
In the new scheme, we first find the (probability density) likelihood function of the set
intersection estimation problem, L(xA = s, xB = t; θ); namely, given θ = (a, b, n) as the
problem parameters2, this is the probability density of s to be the maximal hash value for
A and t to be the maximal hash value for B. Then, we look for values of the problem
parameters that maximize the likelihood function.
Table 1 shows some of the notations we use for the rest of the paper.
notation
value
A ∩ B
A ∪ B
n
u
a
b
α
β
A
B
A \ B
B \ A
Table 1: Notations
3.1 The Likelihood Function of the Set Intersection Estimation
Problem
B = maxb
A and xk
We first find the likelihood function for one hash function hk, namely, L(xA = s, xB = t; θ)k,
and then generalize it for all hash functions. Recall that for the kth hash function, xk
A =
i=1 {hk(ai)} and xk
maxa
j=1 {hk(bj)}. To simplify the notation, we shall omit the
superscript k for xk
We use PDFU (w) to denote the probability density function (PDF) of a uniformly dis-
tributed random variable U(0, 1) at w. Denote the elements in A∩B as {z1, z2, . . . , zn}. Thus,
the elements in A and the elements in B can be written as {x1, x2, . . . , xα, z1, z2, . . . , zn} and
{y1, y2, . . . , yβ, z1, z2, . . . , zn} respectively (see Table 1).
We now divide the likelihood function according to the three possible relations between
xA and xB: xA = xB, xA > xB and xA < xB.
B, and use xA and xB respectively.
Case 1: xA = xB
When xA = xB = s holds, the element with the maximal hash value must belong to
A ∩ B. The likelihood function of θ given this outcome is
L(xA = xB = s; θ) =
=
PDFU (s) · Pr(cid:0)x(A∪B)\{zi } < s(cid:1)
su−1 = n · su−1.
nXi=1
nXi=1
(1)
This equality holds because there are n possible elements in A ∩ B whose hash value
can be the maximum in A ∪ B, and because PDFU (s) = 1.
2The set intersection estimation problem has six identifying parameters (n, u, a, b, α, β), only three of
which are needed to derive the others.
Case 2: xA < xB
In order to have xA < xB, where xA = s and xB = t, the maximal hash value in B
must also be in B \ A, and its value must be t. The likelihood function of θ in this
case is
L(xB\A = t; θ) =
βXj=1
PDFU (t) · Pr(cid:0)x(B\A)\{yj } < t(cid:1) = β · tβ−1.
In addition, the maximal hash value in A must be s. The probability density for this
is
L(xA = s; θ) =Xe∈A
PDFU (s) · Pr(cid:0)xA\{e} < s(cid:1) = asa−1.
Thus,
L(xA < xB, xA = s, xB = t; θ) = asa−1 · βtβ−1.
(2)
Case 3: xA > xB
This case is symmetrical to the previous case. Thus, the likelihood function of θ in
this case is
L(xA > xB, xA = s, xB = t; θ) = αsα−1btb−1.
Thus, the likelihood function for set intersection is
L(xA = s, xB = t; θ)k =
nsu−1
xA = xB
βtβ−1asa−1 xA < xB
αsα−1btb−1
xA > xB .
We now use the following indicator variables:
1. I1 = 1 if xA = xB, and I1 = 0 otherwise,
2. I2 = 1 if xA < xB, and I2 = 0 otherwise,
3. I3 = 1 if xA > xB, and I3 = 0 otherwise,
to obtain that
L(xA = s, xB = t; θ)k = (nsu−1)I1 · (βtβ−1asa−1)I2 · (αsα−1btb−1)I3.
(3)
Eq. (3) states the likelihood function for one hash function. To generalize this equation
to all m hash functions, denote S = (s1, s2, . . . , sm) and T = (t1, t2, . . . , tm) as the m-
dimensional vectors of s and t for each hash function.
Corollary 1
The likelihood function for the set intersection estimation problem, for all m hash functions,
satisfies:
L(xA = S, xB = T ; θ) =
mYk=1
(n(sk)u−1)I1,k · (β(tk)β−1a(sk)a−1)I2,k · (α(sk)α−1b(tk)b−1)I3,k,
where I1,k is the value of I1 for hash function k, and the same holds for I2,k and I3,k.
(cid:4)
It is usually easier to deal with the log of a likelihood function than with the likelihood
function itself. Because the logarithm is a monotonically increasing function, its maximum
value is obtained at the same point as the maximum of the function itself. In our case,
log L(xA = S, xB = T ; θ) = log
L(xk
A = sk, xk
B = tk; θ)k
mYk=1
mXk=1
mXk=1
mXk=1
=
=
+
log L(xk
A = sk, xk
B = tk; θ)k
I1,k · log (n · (sk)u−1) +
mXk=1
I2,k · log (β · (tk)β−1 · a · (sk)a−1)
I3,k · log (α · (sk)α−1 · b · (tk)b−1).
(4)
3.2 The New Scheme
We use Corollary 1 in order to find θ = (a, b, n) that maximizes Eq. (4). Let g(a, b, n) and
H(a, b, n) be the gradient and the Hessian matrix of the log-likelihood function. Namely, g
is the vector whose components are the partial derivatives of the log-likelihood function for
the problem parameters θ = (a, b, n):
g(a, b, n) = (
∂ log L
∂a
,
∂ log L
∂b
,
∂ log L
∂n
),
and H is the matrix of the second-order partial derivatives of the log-likelihood function
Hi,j =
∂2
∂θi∂θj
log L , 1 ≤ i, j ≤ 3 ,
where θ1 = a, θ2 = b and θ3 = n.
(5)
(6)
The new scheme finds the maximal value of the log-likelihood function, i.e., the root
of its gradient g(a, b, n). Practically, this is done using iterations of the Newton-Raphson
method on the gradient and Hessian matrix g and H. Starting from an initial estimation
bθ0 = (ba0,bb0, bn0), the Newton-Raphson method implies that a better estimation is bθ1 =
bθ0 − H −1(bθ0) · g(bθ0). The process is repeated, namely,
(7)
until a sufficiently accurate estimation is reached. This idea is summarized in the following
algorithm.
dθl+1 =bθl − H −1(bθl) · g(bθl),
k=1, where xk
Algorithm 1
(A Maximum Likelihood scheme for the set intersection estimation problem)
A and xk
B
are the maximal hash values of A and B respectively for the kth hash function, and returns
an estimate of their set intersection cardinality.
The scheme gets as an input the sketches of the sets(cid:8)xk
A(cid:9)m
k=1 and(cid:8)xk
B(cid:9)m
1) Estimate a0 =ba, b0 =bb andbu using any cardinality estimation algorithm, such as [12].
2) Estimate the Jaccard similarity bρ from the given sketches of A and B.
n0 =bρ ·bu as an initial value of n (see Scheme-2 in Section 1), and a0, b0 as an initial
4) Return bn.
When we implemented Algorithm 1, we discovered that 3 Newton-Raphson iterations are
enough for the algorithm to converge.
3) Find the maximum of the likelihood function L (Eq.
values of a and b respectively.
(4)) as explained above; use
3.3 The Optimal Variance of the New Estimator
The new estimator proposed in this section is based on Maximum Likelihood and thus it
asymptotically achieves optimal variance [24]. We use the Cramer-Rao bound to compute
this optimal variance.
The Cramer-Rao bound states that the inverse of the Fisher information matrix is a lower
bound on the variance of any unbiased estimator of θ [24]. The Fisher information matrix
Fi,j is a way of measuring the amount of information that a random variable X carries about
an unknown parameter θ upon which the probability of X depends. It is defined as:
Fi,j = −E(cid:20) ∂2
∂θi∂θj
log L(cid:21) .
(8)
We now use the log-likelihood function (Eq. (4)) to derive this matrix for the set inter-
section estimation problem:
F(a, b, n) = m ·
β
u · 1
u·α
a2 + 1
0
−1
u·α
0
b2 + 1
u·β
α
u · 1
−1
u·β
−1
u·α
−1
u·β
1
u·n + 1
u·β + 1
u·α
,
(9)
where each term is derived due to algebraic manipulations and derivatives of the log-
likelihood function. Note that the expected values of the indicator variables I1,k, I2,k and I3,k
are required to derive the matrix. For I1,k we get:
E[I1,k] = Pr(cid:0)xk
n
u
, 1 ≤ k ≤ m .
A = xk
B(cid:1) =
The first equality is due to the definition of I1,k, and the second is due to Eq. (11). The
following are obtained in the same way for every 1 ≤ k ≤ m:
u .
A > xk
A < xk
B(cid:1) = β
B(cid:1) = α
1. E[I2,k] = Pr(cid:0)xk
2. E[I3,k] = Pr(cid:0)xk
Let \A ∩ B be an unbiased estimator for the set intersection estimation problem. Then,
according to the Cramer-Rao bound, Varh \A ∩ Bi ≥ (F−1)3,3, where (F−1)3,3 is the term in
place [3, 3] in the inverse Fisher information matrix. Finally, from the computation of the
term (F−1)3,3, we can obtain the following corollary:
u .
Corollary 2
Varh \A ∩ Bi ≥ (F−1)3,3, where,
(F−1)3,3 = n·u
m ·
α·n(a2+αβ)+β·n(b2+αβ)+(a2+αβ)(b2+αβ).
(b2+αβ)(a2+αβ)
4 An Analysis of the Three Schemes From Section 1
In this section we will analyze the statistical performance (bias and variance) of the three
schemes discussed in Section 1 for set intersection estimation.
4.1 Preliminaries
4.1.1 Jaccard Similarity
Recall that the Jaccard similarity is defined as: ρ(A, B) = A∩B
A∪B , where A and B are two
finite sets. Its value ranges between 0, when the two sets are completely different, and 1,
when the sets are identical. An efficient and accurate estimate of ρ can be computed as
follows [3]. Each item in A and B is hashed into (0, 1), and the maximal value of each
set is considered as a sketch that represents the whole set. To improve accuracy, m hash
functions are used3, and the sketch of each set is a vector of m maximal values. Given a set
A = {a1, a2, . . . , ap} and m different hash functions h1, h2, . . . , hm, the maximal hash value
for the jth hash function can be formally expressed as:
xj
A =
p
i=1 {hj(ai)} , 1 ≤ j ≤ m,
max
and the sketch of A is:
A, x2
X(A) =(cid:8)x1
\ρ(A, B) = Pm
A, . . . , xm
A(cid:9) .
A==xj
B
j=1 Ixj
m
X(B) is computed in the same way. Then, the two sketches are used to estimate the Jaccard
similarity of A and B:
,
(10)
where the indicator function Ixj
notation, for the rest of the paper we use ρ to indicate ρ(A, B).
A==xj
B
is 1 if xj
A = xj
B, and 0 otherwise. To shorten our
3Stochastic averaging can be used to reduce the number of hash functions from m to only two [13].
Lemma 1
m ρ(1 − ρ)(cid:1).
Proof:
Consider the j-th hash function. According to [3],
In Eq. (10),bρ is a normally distributed random variable with mean ρ and variance 1
i.e., bρ → N(cid:0)ρ, 1
Pr(cid:0)xj
B(cid:1) = A ∩ B
A ∪ B
A = xj
.
mρ(1−ρ);
(11)
A = xj
binomially distributed, and can be asymptotically approximated to normal distribution as
The intuition is to consider the hash function hj and define m(S), for every set S, to be
the element in S with the maximum hash value of hj, i.e., hj(m(S)) = xj
S. Then, we get
m(A) = m(B) only when m(A ∪ B) lies also in their intersection A ∩ B. The probability of
this is the Jaccard ratio ρ, and therefore Pr(cid:0)xj
From Eqs. (10) and (11) follows thatbρ is a sum of m Bernoulli variables. Therefore, it is
→ N(cid:0)ρ, 1
m → ∞; namely, bρ =
4.1.2 The Cardinality Estimation Problem
Algorithms for estimating the cardinality of set intersection use estimations of A, B, and
A ∪ B. These estimations can be found using well-known algorithms for the following
cardinality estimation problem:
m ρ(1 − ρ)(cid:1).
B(cid:1) = ρ.
Pm
l=1 I
x
j
A
=x
j
B
m
Instance: A stream of elements x1, x2, . . . , xs with repetitions. Let c be the number of
different elements, namely c = {x1, x2, . . . , xs}.
Objective: Find an estimatebc of c using only m storage units, where m ≪ c.
For the rest of the paper we consider the HyperLogLog algorithm [12] for solving the
above problem. As indicated in Section 2, this algorithm has a very small standard error, of
about 1.04/ √m where m is the number of storage units. The pseudo-code of this algorithm
is as follows:
Algorithm 2 The HyperLogLog algorithm for the cardinality estimation problem
1. Initialize m registers: C1, C2, . . . , Cm to 0.
2. For each input element xi do:
(a) Let ρ = ⌊− log2 (h1(xi))⌋ be the leftmost 1-bit position of the hashed value.
(b) Let j = h2(xi) be the bucket for this element.
(c) Cj ← max{Cj, ρ}.
3. To estimate the value of n do:
(a) Z ← (Pm
j=1 2−Cj )−1 is the harmonic mean of 2Cj .
(b) return αmm2Z, where
.
Lemma 2
du(cid:1)−1
αm =(cid:0)mR ∞
The following lemma summarizes the statistical performance of Algorithm 2:
the estimate computed by the algorithm, and m is the number of storage units used by the
algorithm. When Algorithm 2 is used with two sets A and B, the following holds:
0 (cid:0)log2(cid:0) 2+u
1+u(cid:1)(cid:1)m
For Algorithm 2,bc → N(cid:16)c, c2
m(cid:17), where c is the actual cardinality of the considered set,bc is
m ! ,
cA → N A , A2
cB → N B , B2
m ! ,
m ! .
and \A ∪ B → N A ∪ B , A ∪ B2
(12)
The proof is given in [12].
Let us also recall three general lemmas, not related to set intersection cardinality esti-
mation. The first lemma, known as the Delta Method, allows us to compute the probability
distribution for a function of an asymptotically normal estimator using the estimator's vari-
ance:
Lemma 3 (Delta Method)
Let θ1, θ2, . . . , θm be a sequence of m random variables such that for every integer i √i(θi −
θ) → N (0, σ2), where θ and σ2 are finite valued constants. Then, for every integer i and for
every function g for which g′(θ) exists and g′(θ) 6= 0, the following holds:
√i(g(θi) − g(θ)) → N(cid:16)0, σ2g′(θ)2(cid:17) .
A proof is given in [24].
The next lemma shows how to compute the probability distribution of a random variable
that is a product of two normally distributed random variables whose covariance is 0:
Lemma 4 (Product distribution)
Let X and Y be two random variables satisfying X → N (µx, σ2
that Cov [X, Y ] = 0. Then, the product X · Y asymptotically satisfies the following:
x) and Y → N(cid:0)µy, σ2
y(cid:1), such
X · Y → N(cid:0)µxµy, µ2
A proof is given in [24].
yσ2
x + µ2
xσ2
y(cid:1) .
The final lemma states the distribution of the maximal hash value. Let us first recall
the beta distribution. Beta (α, β) is defined over the interval (0, 1) and has the following
probability and cumulative density functions (PDF and CDF respectively):
f (x) =
Γ (α + β)
Γ (β) Γ (α)
xα−1(1 − x)β−1
F (x) =Z x
0
Γ (α + β)
Γ (β) Γ (α)
xα−1(1 − x)β−1dx,
where Γ(z) is the gamma function, defined asR ∞
0 e−ttz−1dt. Using integration by parts, the
gamma function can be shown to satisfy Γ(z + 1) = z · Γ(z). Combining this with Γ(1) = 1
yields that Γ(n) = (n − 1)! holds for every integer n. Two other known beta identities are
[19]:
Beta (1, 1) ∼ U(0, 1)
and
Beta (α, β) ∼ 1 − Beta (β, α) .
The following lemma presents some key properties of the beta distribution, which we will
use in the analysis.
Lemma 5
Let x1, x2, . . . , xn be independent RVs, where xi ∼ U(0, 1). Then,
(a) X = maxn
i=1 xi ∼ Beta (n, 1).
(b) X satisfies the following
(1) E[X] = n
n+1; and
(2) Var [X] =
n
(n+1)2(n+2) .
A proof for (a) is given in [10]; the other equalities follow the beta distribution of X.
4.2 Analysis of Scheme-1
Scheme-1 estimates the cardinality of A ∩ B using the inclusion-exclusion principle:
\A ∩ B = cA + cB − \A ∪ B.
performance.
Let bn1 be the estimator found by Scheme-1. The following theorem summarizes its statistical
n → N(cid:16)1,
(cid:17), where m is the number of
Theorem 3
cn1
storage units, and Z satisfies:
mn2 (u2 − a2 − b2) − 2a·b
m·u·n + 2u·(a2(b2+αβ)+b2(a2+αβ))
m·Z·n
1
Z = α · n(a2 + αβ) + β · n(b2 + αβ) + (a2 + αβ)(b2 + αβ).
Proof:
For the expectation, from Lemma 2 follows that:
E[bn1] = E[ba] + Ehbbi − E[bu] =
= a + b − u = n.
The first equality is due to the definition of Scheme-1 and the expectation properties, and
the second equality is due to Lemma 2. Thus, the estimator is unbiased. For the variance,
Lemma 7 in the Appendix proves that Cov [a, b] = n·a·b
m·u , and Lemma 8 in the Appendix
m·u − u·n·b2(a2+αβ)
proves that Cov [a, u] = a2
m + n·a·b
,
where Z = α · n(a2 + αβ) + β · n(b2 + αβ) + (a2 + αβ)(b2 + αβ). We get that:
m·u − u·n·a2(b2+αβ)
and Cov [b, u] = b2
m + n·a·b
m·Z
m·Z
Var [bn1] = Var [ba] + Varhbbi + Var [bu] + 2 Covhba,bbi − 2 Cov [ba,bu] − 2 Covhbb,bui =
=
+
a2
m
− 2 ·
− 2 ·
1
=
m
+
b2
m
u2
+ 2 ·
m
a2
n · a · b
m − 2 ·
m · u
b2
n · a · b
m − 2 ·
m · u
(u2 − a2 − b2) −
n · a · b
m · u
+ 2 ·
+ 2 ·
2n · a · b
m · u
u · n · a2(b2 + αβ)
u · n · b2(a2 + αβ)
m · Z
m · Z
2u · n · (a2(b2 + αβ) + b2(a2 + αβ))
=
+
m · Z
.
(13)
The first equality is due to variance properties and because a, b and u are dependent
(Cov [a, b], Cov [a, u] and Cov [b, u] are all 6= 0). The second equality is due to Lemma 2,
Lemma 7 and Lemma 8 (both are in the Appendix). The third equality is due to algebraic
manipulations. Finally, after dividing by n2 we get the result.
4.3 Analysis of Scheme-2
Scheme-2 estimates the cardinality of A ∩ B by estimating the Jaccard similarity ρ and
A ∪ B [2]:
\A ∩ B =bρ · \A ∪ B.
performance.
Theorem 4
cn2
mρ(cid:17), where m is the number of storage units.
Let bn2 be the estimator found by Scheme-2. The following theorem summarizes its statistical
n → N(cid:16)1, 1
From the definition of Scheme-2, bn2 =bρ ·bu. From Lemma 1 and Lemma 2 follows that:
1. bρ → N(cid:0)ρ, 1
mρ(1 − ρ)(cid:1).
Proof:
Therefore, the estimator is unbiased. For the variance, applying again Lemma 4 yields:
Applying Lemma 4 for the expectation yields:
2. bu → N(cid:16)u, u2
m(cid:17).
E[bn2] = E[bρ ·bu] = ρ · u = n.
Var [bn2] = u2 ·
1
ρ(1 − ρ) + ρ2 ·
m
1
(u · n − n2 + n2)
m
n2
,
mρ
=
=
u2
m
where all the equalities are due to Lemma 4 and algebraic manipulations. Finally, after
dividing by n2, we get cn2
mρ(cid:17).
n → N(cid:16)1, 1
4.4 Analysis of Scheme-3
Scheme-3 estimates the cardinality of A∩ B by estimating the Jaccard similarity ρ, A, and
B [11]:
\A ∩ B = dρ
ρ + 1
(cA + cB).
Lemma 6
Let bn3 be the estimator found by Scheme-3. We will use the following lemma in the analysis.
ba +bb → N(cid:0)a + b, 1
m(a2 + b2 + 2abρ)(cid:1), where m is the number of storage units.
Proof:
For the expectation,
The first equality is due to expectation properties, and the second is due to Lemma 2. For
the variance, Lemma 7 in the Appendix proves that Cov [a, b] = n·a·b
m·u . Thus,
Ehba +bbi = E[ba] + Ehbbi = a + b.
Varhba +bbi =
= Var [ba] + Varhbbi + 2 Covhba,bbi =
+
=
b2
m
+ 2 ·
n · a · b
m · u
(a2 + b2 + 2abρ).
=
=
a2
m
1
m
(14)
(15)
(16)
The first equality is due to variance properties and because a and b are dependent (Cov [a, b] 6=
0). The second equality is due to Lemma 2 and Lemma 7, and the third equality is due to
algebraic manipulations and the Jaccard similarity definition (ρ = n
u ). Combining Eqs. 15
and 16 yields that
1
m
(17)
(18)
Theorem 5
cn3
Proof:
m (1 + 2ab
u(a+b) + (α + β)
u2
Applying Lemma 3 on ρ
ρ+1 yields:
1
m
ρ(1 − ρ)(cid:19) .
(a2 + b2 + 2abρ)(cid:19) .
n(a+b)2 )(cid:17), where m is the number of storage units.
ρ+1 (ba +bb). From Lemma 1 follows that:
ba +bb → N(cid:18)a + b,
The following theorem summarizes the statistical performance of bn3.
n → N(cid:16)1, 1
From the definition of Scheme-3, bn3 =dρ
bρ → N(cid:18)ρ,
ρ + 1 → N(cid:18) ρ
dρ
E[bn3] = E(cid:20) dρ
Var [bn3] =
A simple comparison yields that Var [bn2] > Var [bn3], i.e., Scheme-3 outperforms Scheme-
2. However, a similar comparison between Scheme-1 and Scheme-3 cannot be done, because
neither one is always better than the other.
Therefore, the estimator is unbiased. For the variance, applying again Lemma 4 yields:
ρ + 1 · [a + b(cid:21) =
Applying Lemma 4 for Eqs. (18) and (17) yields:
n → N(cid:16)1, 1
ρ
ρ + 1 · (a + b) = n.
Finally, after dividing by n2, we get: cn3
1
(1 + ρ)4(cid:19) .
1
m
(n2 +
2abn2
u(a + b)
,
1
m
ρ(1 − ρ)
ρ + 1
+ (α + β)
u2n
(a + b)2 ).
m(1 + 2ab
u(a+b) + (α + β)
u2
n(a+b)2 )(cid:17).
5 Simulation Results
In this section we examine the performance of our new ML estimator and show that it indeed
outperforms the three known schemes. We implemented all four schemes, and simulated two
sets, A and B, whose cardinalities are as follows:
1. A = a = 106;
4
0
0
0
.
s
a
b
i
2
0
0
0
.
0
0
0
0
.
m=1,000
m=10,000
0.0
0.2
0.4
0.6
0.8
1.0
Figure 1: The bias of the new ML scheme for f = 1 and for different α values, for m = 1, 000
and m = 10, 000
2. B = a · f , where f > 0;
3. A ∩ B = a · α, where 0 ≤ α ≤ 1.
We estimate \A ∩ B for each of the four schemes, for f ∈ {1, 5, 10}, and for α ∈
{0, 0.01, 0.02, . . . , 0.98, 0.99, 1}. We repeat the test for 10, 000 different sets. Thus, for each of
the four schemes, and for each f and α values, we get a vector of 10, 000 different estimations.
Then, for each f and α values, we compute the variance and bias of this vector, and view the
result as the variance and bias of the estimator (for the specific f and α values). Each such
of estimations for a specific scheme and for specific f and α values. Let µ = 1
the mean of vf,α. The bias and variance of vf,α are computed as follows:
computation is represented by one point in the graph. Let vf,α = (bn1, . . . ,bn104) be the vector
i=1bni, be
Bias(vf,α) =(cid:12)(cid:12)(cid:12)(cid:12)
104P104
and
Figure 1 presents the bias of the ML estimator for f = 1, different α values, and two
values of m: m = 1, 000 and m = 10, 000 (recall that m is the number of hash values used
small for all α and m values. We got very similar results for bigger f values as well.
for the estimations of cA, cB and the Jaccard similaritybρ). We can see that the bias is very
n(cid:3)) of our ML estimator for different f
Figure 2 presents the normalized variance (Var(cid:2) bn
and α values, and for m ∈ {100, 500, 1000, 10000}. As expected, the normalized variance
1
n
(µ − n)(cid:12)(cid:12)(cid:12)(cid:12)
104Xi=1
(bni − µ)2.
Var [vf,α] =
1
104
a
e
c
n
a
i
r
a
v
d
e
z
i
l
a
m
r
o
n
8
0
0
.
4
0
.
0
0
0
0
.
m=100
m=500
m=1,000
m=10,000
e
c
n
a
i
r
a
v
d
e
z
i
l
a
m
r
o
n
0
3
.
0
0
2
0
.
0
1
0
.
0
0
0
.
m=100
m=500
m=1,000
m=10,000
e
c
n
a
i
r
a
v
d
e
z
i
l
a
m
r
o
n
5
0
.
4
0
.
3
0
.
2
0
.
1
0
.
0
0
.
m=100
m=500
m=1,000
m=10,000
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
1.0
(a) f = 1
(b) f = 5
(c) f = 10
Figure 2: The normalized variance of the new ML scheme for different values of f , α and m
decreases as the number of hash values (m) increases, or as α increases. Overall, the nor-
malized variance is very small for all values of α, f and m, indicating that the new scheme
is very precise.
After showing that the new ML scheme indeed yields good results, we now compare
its performance to that of Schemes 1-3. When comparing the statistical performance of
two algorithms, it is common to look at their MSE (mean squared error) or RMSE, where
MSE = (Bias(bθ))2 + Varhbθi, and RMSE = r(Bias(bθ))2 + Varhbθi.
all the estimators are unbiased, we compare only their variance. We define the "relative
variance improvement" of the new ML scheme over each scheme as
In our case, because
Varhbθii − VarhdθMLi
,
Varhbθii
where θi is the estimator of the ith scheme, and θML is the new ML estimator.
Figure 3 presents the simulation results for two values of m: m = 10, 000 (upper graphs)
and m = 1, 000 (lower graphs). We can see that the new ML estimator outperforms
the three schemes for all values of α and f , and for both values of m. This
improvement varies between 100% to a few percent.
6 Conclusion
In this paper we studied the problem of estimating the number of distinct elements in the set
intersection of two streams. We presented a complete analysis of the statistical performance
(bias and variance) of three previously known schemes. We then computed the likelihood
function of the problem and used it to present a new estimator, based on the ML method.
We also found the optimal variance of any unbiased set intersection estimator, which is
asymptotically achieved by our new ML scheme. We can conclude that our new scheme
outperforms the three known schemes, significantly improves the variance (precision) of the
estimator, and yields better results than the three previously known schemes.
a
a
a
t
n
e
m
e
v
o
r
p
m
i
%
t
n
e
m
e
v
o
r
p
m
i
%
0
0
1
0
8
0
6
0
4
0
2
0
0
0
1
0
8
0
6
0
4
0
2
0
f=10
f=5
f=1
t
n
e
m
e
v
o
r
p
m
i
%
0
8
0
6
0
4
0
2
0
f=10
f=5
f=1
t
n
e
m
e
v
o
r
p
m
i
%
0
8
0
6
0
4
0
2
0
f=10
f=5
f=1
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
f=10
f=5
f=1
t
n
e
m
e
v
o
r
p
m
i
%
0
8
0
6
0
4
0
2
0
f=10
f=5
f=1
t
n
e
m
e
v
o
r
p
m
i
%
0
8
0
6
0
4
0
2
0
f=10
f=5
f=1
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
(a) Scheme-1
(b) Scheme-2
(c) Scheme-3
Figure 3: The percentage of variance improvement of the new ML scheme over each of the
other schemes for different values of f and α; the upper graphs are for m = 10, 000 and the
lower graphs are for m = 1, 000
a
a
a
a
a
a
References
[1] J. Bauckmann, U. Leser, F. Naumann, and V. Tietz. Efficiently detecting inclusion
dependencies. In ICDE, pages 1448–1450, 2007.
[2] K. S. Beyer, P. J. Haas, B. Reinwald, Y. Sismanis, and R. Gemulla. On synopses for
In SIGMOD Conference, pages
distinct-value estimation under multiset operations.
199–210, 2007.
[3] A. Z. Broder. On the resemblance and containment of documents. In IEEE Compression
and Complexity of Sequences 1997, pages 21–29.
[4] A. Z. Broder. Identifying and filtering near-duplicate documents. In CPM, pages 1–10,
2000.
[5] A. Z. Broder, S. C. Glassman, M. S. Manasse, and G. Zweig. Syntactic clustering of
the web. Computer Networks, 29(8-13):1157–1166, 1997.
[6] P. Chassaing and L. G´erin. Efficient estimation of the cardinality of large data sets. In
Proceedings of the 4th Colloquium on Mathematics and Computer Science, volume AG
of Discrete Mathematics & Theoretical Computer Science Proceedings, pages 419–422,
2006.
[7] P. Clifford and I. A. Cosma. A statistical analysis of probabilistic counting algorithms.
Scandinavian Journal of Statistics, 2011.
[8] C. Clifton, E. Housman, and A. Rosenthal. Experience with a combined approach to
attribute-matching across heterogeneous databases. In DS-7, pages 428–451, 1997.
[9] E. Cohen and H. Kaplan. Tighter estimation using bottom k sketches. PVLDB,
1(1):213–224, 2008.
[10] R. Cohen, A. Yehezkel, and L. Katzir. A Unified Scheme for Generalizing Cardinality
Estimators to Sum Aggregation. Inf. Process. Lett., 2014.
[11] T. Dasu, T. Johnson, S. Muthukrishnan, and V. Shkapenyuk. Mining database struc-
ture; or, how to build a data quality browser. In SIGMOD Conference, pages 240–251,
2002.
[12] P. Flajolet, ´E. Fusy, O. Gandouet, and F. Meunier. Hyperloglog: the analysis of a
near-optimal cardinality estimation algorithm. In Analysis of Algorithms (AofA) 2007.
DMTCS.
[13] P. Flajolet and G. N. Martin. Probabilistic counting algorithms for data base applica-
tions. J. Comput. Syst. Sci., 31:182–209, Sep. 1985.
[14] ´E. Fusy and F. Giroire. Estimating the number of active flows in a data stream over
a sliding window. In D. Panario and R. Sedgewick, editors, ANALCO, pages 223–231.
SIAM, 2007.
[15] S. Ganguly, M. N. Garofalakis, R. Rastogi, and K. K. Sabnani. Streaming algorithms
In ICDCS, page 4. IEEE Computer
for robust, real-time detection of ddos attacks.
Society, 2007.
[16] F. Giroire. Order statistics and estimating cardinalities of massive data sets. Discrete
Applied Mathematics, 157:406–427, 2009.
[17] M. A. Hern´andez and S. J. Stolfo. Real-world data is dirty: Data cleansing and the
merge/purge problem. Data Min. Knowl. Discov., 2(1):9–37, 1998.
[18] H. Kohler. Estimating set intersection using small samples.
In ACSC, pages 71–78,
2010.
[19] K. Krishnamoorthy. Handbook of Statistical Distributions with Applications. Chapman
& Hall/CRC Press, Boca Raton, FL, 2006.
[20] J. Lumbroso. An optimal cardinality estimation algorithm based on order statistics and
its full analysis. In Analysis of Algorithms (AofA) 2010. DMTCS.
[21] A. Metwally, D. Agrawal, and A. E. Abbadi. Why go logarithmic if we can go linear?:
Towards effective distinct counting of search traffic. In Proceedings of the 11th Interna-
tional Conference on Extending Database Technology: Advances in Database Technology,
EDBT '08, pages 618–629.
[22] A. E. Monge. Matching algorithms within a duplicate detection system. IEEE Data
Eng. Bull., 23(4):14–20, 2000.
[23] E. Rahm and H. H. Do. Data cleaning: Problems and current approaches. IEEE Data
Eng. Bull., 23(4):3–13, 2000.
[24] J. Shao. Mathematical Statistics. Springer, 2003.
Appendix
Lemma 7
The covariance of cA and cB satisfies Covhba,bbi = n·a·b
functions used for the estimation ofba andbb.
m·u , where m is the number of hash
Proof:
Denote xA and xB as the maximal hash values for A and B respectively. Using our notation
for I1, I2 and I3 from Section 3.1, and applying the linearity of expectation, we obtain:
E[xA · xB] = E[(xA · xB) · (I1 + I2 + I3)]
= E[(xA · xB) · I1] + E[(xA · xB) · I2] + E[(xA · xB) · I3] .
(19)
We consider each term separately:
1) xA = xB: From Eq. (1),
E[(xA · xB) · I1] =ZxA=xB
n · xu+1
A dx =
n
u + 2
.
2) xA < xB: From Eq. (2),
E[(xA · xB) · I2] =ZZxA>xB
B · a · xa
A dxA dxB =Z 1
0
β · xβ
β
.
u + 2
a
a + 1 ·
=
3) xB > xA:
This case is symmetrical to the second case. We get that
E[(xA · xB) · I3] =
b
b + 1 ·
α
u + 2
.
Substituting the three terms into Eq. (19) yields that
E[xA · xB] =
=
n
u + 2
+
a
a + 1 ·
β
u + 2
+
b
b + 1 ·
α
u + 2
=
ab · (u + 2) + n
(u + 2)(a + 1)(b + 1)
.
Using the covariance definition, we obtain:
β · xβ
B dxB ·Z xB
0
a · xa
A dxA =
(20)
Cov [xA, xB] = E[xA · xB] − E[xA] · E[xB] =
ab · (u + 2) + n
(u + 2)(a + 1)(b + 1) −
a
a + 1 ·
b
b + 1
=
n
=
(u + 2)(a + 1)(b + 1)
.
(21)
The first equality is due to the covariance definition, and the second equality is due to Eq.
(20) and Lemma 5(b).
Eq. (21) states the covariance for one hash function. We can generalize it for all hash
B be the maximal hash values for the ith hash function, for A and
A and xi
functions. Let xi
B respectively. Then
A, xj
Cov(cid:2)xi
B(cid:3) =(0
n
(u+2)(a+1)(b+1)
i 6= j
i = j .
(22)
We are now ready to compute Covhba,bbi. To this end, we first need to choose the cardinality
estimator that we will use to estimateba andbb. For simplicity, we use the estimator from [7]:
(23)
m
,
ba =
Pm
i=1 (1 − xi
A)
where m hash functions are used (symmetrically forbb). Denoting X =P (1 − xi
X (Y is symmetrically defined forbb). For each i, xi
rewrite the estimator as ba = m
maximum of a uniformly distributed variables, and thus
a
A), we can
A is the
E(cid:2)xi
A(cid:3) =
a + 1
Both equalities follow the beta distribution of xi
that
(a + 1)2(a + 2)
.
a
, and Var(cid:2)xi
A(cid:3) =
A)i = m · (1 − E(cid:2)xi
A(cid:3)) = m ·
A)i = m · Var(cid:2)xi
A(cid:3) =
and the variance
E[X] = EhX (1 − xi
Var [X] = VarhX (1 − xi
1
a + 1
=
m
a + 1
,
ma
(a + 1)2(a + 2)
.
A (Lemma 5(b)). From Eq. (24) it follows
(24)
(25)
(26)
From Eqs. (25) and (26) we can conclude that X → N(cid:16) m
(a+1)2(a+2)(cid:17), and symmetrically
X (symmetrically forbb). Applying
for Y . Recall that the estimators can be rewritten asba = m
the Delta Method (multivariate version) for the estimator's vector (ba,bb) = (m/X, m/Y )
yields
a+1,
ma
m/Y (cid:19) → N(cid:18)(cid:18) a
b (cid:19) , g′ · Σ · g′T(cid:19) ,
where g(X, Y ) = (m/X, m/Y ) is the function used in the Delta Method, g′ is its partial-
derivatives matrix and Σ is the variance matrix:
(cid:18) ba
bb (cid:19) =(cid:18) m/X
Σ =(cid:18) Var [X]
Cov [X, Y ]
Cov [X, Y ]
Var [Y ] (cid:19) .
Using covariance properties and Eq. (22), we obtain:
Cov [X, Y ] = CovhX 1 − xi
Ai =
A,X 1 − xj
mXi=1
Cov(cid:2)xi
A, xi
B(cid:3) =
mn
(u + 2)(a + 1)(b + 1)
.
(27)
Substituting the terms from Eqs. (26) and (27) in Σ yields
mn
(a+1)(b+1)(u+2)
(b+1)2(b+2) ! .
mb
(a+1)(b+1)(u+2)
ma
mn
(a+1)2(a+2)
Σ =
bb (cid:19) → N (cid:18) a
(cid:18) ba
Computing g′ · Σ · g′T yields the final distribution of the estimators:
b (cid:19) , a(a+1)2
m(a+2)
m(u+2)
n(a+1)(b+1)
n(a+1)(b+1)
m(b+2) !! .
m(u+2)
b(b+1)2
Finally, according to the Delta Method, Covhba,bbi is the term in place [1, 2] in the matrix.
Lemma 8
The covariance of cA and \A ∪ B satisfies Cov [ba,bu] = a2
the number of storage units, and Z satisfies:
Z = α · n(a2 + αβ) + β · n(b2 + αβ) + (a2 + αβ)(b2 + αβ).
m + n·a·b
m·u − u·n·a2(b2+αβ)
m·Z
, where m is
Proof:
According to Fisher information matrix properties, \Cov [ba,bn] = (F−1)1,3 is a Maximum
Likelihood estimator for Cov [ba,bn], where (F−1)1,3 is the term in place [1, 3] in the inverse
Fisher information matrix. Computing the term (F−1)1,3 yields that
Therefore,
u · n · a2(b2 + αβ)
.
m · Z
Cov [ba,bn] =
Cov [ba,bu] = Covhba,ba +bb −bni = Cov [ba,ba] + Covhba,bbi − Cov [ba,bn] =
u · n · a2(b2 + αβ)
+
=
.
a2
m
n · a · b
m · u −
m · Z
(28)
The first equality is due to the inclusion-exclusion principle, and the second is due to covari-
ance properties. The third equality is due to covariance properties, Lemma 2, Lemma 7 and
Eq. (28). Similarly, we can obtain the covariance of cB and \A ∪ B:
u · n · b2(a2 + αβ)
+
.
n · a · b
m · u −
m · Z
Covhbb,bui =
b2
m
|
1504.02671 | 1 | 1504 | 2015-04-10T13:23:27 | Longest Common Extensions in Sublinear Space | [
"cs.DS"
] | The longest common extension problem (LCE problem) is to construct a data structure for an input string $T$ of length $n$ that supports LCE$(i,j)$ queries. Such a query returns the length of the longest common prefix of the suffixes starting at positions $i$ and $j$ in $T$. This classic problem has a well-known solution that uses $O(n)$ space and $O(1)$ query time. In this paper we show that for any trade-off parameter $1 \leq \tau \leq n$, the problem can be solved in $O(\frac{n}{\tau})$ space and $O(\tau)$ query time. This significantly improves the previously best known time-space trade-offs, and almost matches the best known time-space product lower bound. | cs.DS | cs |
Longest Common Extensions in Sublinear Space
Philip Bille1 ⋆, Inge Li Gørtz1 ⋆, Mathias Baek Tejs Knudsen2 ⋆⋆,
Moshe Lewenstein3 ⋆ ⋆ ⋆, and Hjalte Wedel Vildhøj1
1 Technical University of Denmark, DTU Compute
2 Department of Computer Science, University of Copenhagen
3 Bar Ilan University
Abstract. The longest common extension problem (LCE problem) is to
construct a data structure for an input string T of length n that supports
LCE(i, j) queries. Such a query returns the length of the longest common
prefix of the suffixes starting at positions i and j in T . This classic
problem has a well-known solution that uses O(n) space and O(1) query
time. In this paper we show that for any trade-off parameter 1 ≤ τ ≤ n,
the problem can be solved in O( n
τ ) space and O(τ ) query time. This
significantly improves the previously best known time-space trade-offs,
and almost matches the best known time-space product lower bound.
1
Introduction
Given a string T , the longest common extension of suffix i and j, denoted
LCE(i, j), is the length of the longest common prefix of the suffixes of T starting
at position i and j. The longest common extension problem (LCE problem) is
to preprocess T into a compact data structure supporting fast longest common
extension queries.
The LCE problem is a basic primitive that appears as a central subproblem in
a wide range of string matching problems such as approximate string matching
and its variations [1, 4, 13, 15, 21], computing exact or approximate repetitions
[6, 14, 18], and computing palindromes [12, 19]. In many cases the LCE problem
is the computational bottleneck.
Here we study the time-space trade-offs for the LCE problem, that is, the
space used by the preprocessed data structure vs. the worst-case time used by
LCE queries. The input string is given in read-only memory and is not counted
in the space complexity. Throughout the paper we use ℓ as a shorthand for
LCE(i, j). The standard trade-offs are as follows: At one extreme we can store
⋆ Supported by the Danish Research Council and the Danish Research Council under
the Sapere Aude Program (DFF 4005-00267)
⋆⋆ Research partly supported by Mikkel Thorup's Advanced Grant from the Danish
Council for Independent Research under the Sapere Aude research career programme
and the FNU project AlgoDisc - Discrete Mathematics, Algorithms, and Data Struc-
tures.
⋆ ⋆ ⋆ This research was supported by a Grant from the GIF, the German-Israeli Founda-
tion for Scientific Research and Development, and by a BSF grant 2010437
a suffix tree combined with an efficient nearest common ancestor (NCA) data
structure [7,22]. This solution uses O(n) space and supports LCE queries in O(1)
time. At the other extreme we do not store any data structure and instead answer
queries simply by comparing characters from left-to-right in T . This solution uses
O(1) space and answers an LCE(i, j) query in O(ℓ) = O(n) time. Recently, Bille
et al. [2] presented a number of results. For a trade-off parameter τ , they gave: 1)
a deterministic solution with O( n
τ ) space and O(τ 2) query time, 2) a randomized
Monte Carlo solution with O( n
τ ) space and O(τ log( ℓ
τ )) query
time, where all queries are correct with high probability, and 3) a randomized Las
Vegas solution with the same bounds as 2) but where all queries are guaranteed
to be correct. Bille et al. [2] also gave a lower bound showing that any data
structure for the LCE problem must have a time-space product of Ω(n) bits.
τ )) = O(τ log( n
Our Results Let τ be a trade-off parameter. We present four new solutions with
the following improved bounds. Unless otherwise noted the space bound is the
number of words on a standard RAM with logarithmic word size, not including
the input string, which is given in read-only memory.
-- A deterministic solution with O(n/τ ) space and O(τ log2(n/τ )) query time.
-- A randomized Monte Carlo solution with O(n/τ ) space and O(τ ) query time,
-- A randomized Las Vegas solution with O(n/τ ) space and O(τ ) query time.
-- A derandomized version of the Monte Carlo solution with O(n/τ ) space and
such that all queries are correct with high probability.
O(τ ) query time.
Hence, we obtain the first trade-off for the LCE problem with a linear time-space
product in the full range from constant to linear space. This almost matches the
Data Structure
Preprocessing
Trade-off
Space Query Correct Space
Time
range
Reference
1
n
ℓ
1
always
always
1
n
-
-
Store nothing
Suffix tree + NCA
1
n
n2
τ
n2
n2+ε
n
c
i
t
s
i
n
i
m
r
e
t
e
D
τ log ℓ
τ 2
τ
τ
τ
1
τ log2 n
e
t
n
o
M
o
l
r
a
C
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
n
τ
τ w.h.p.
w.h.p.
always
τ always
always
1 ≤ τ ≤ √n [2]
log n ≤ τ ≤ n this paper, Sec. 2
1 ≤ τ ≤ n this paper, Sec. 4
1 ≤ τ ≤ n
1 ≤ τ ≤ n this paper, Sec. 3
1 ≤ τ ≤ n
1 ≤ τ ≤ n this paper, Sec. 3.5
Table 1. Overview of solutions for the LCE problem. Here ℓ = LCE(i, j), ε > 0 is an
arbitrarily small constant and w.h.p. (with high probability) means with probability
at least 1 − n−c for an arbitrarily large constant c. The data structure is correct if it
answers all LCE queries correctly.
n(τ + log n) w.h.p.
n log n
τ
τ log ℓ
τ
always
always
n3/2
w.h.p.
[2]
[2]
s
a
L
s
a
g
e
V
time-space product lower bound of Ω(n) bits, and improves the best deterministic
upper bound by a factor of τ , and the best randomized bound by a factor log( n
τ ).
See the columns marked Data Structure in Table 1 for a complete overview.
While our main focus is the space and query time complexity, we also provide
efficient preprocessing algorithms for building the data structures, supporting
independent trade-offs between the preprocessing time and preprocessing space.
See the columns marked Preprocessing in Table 1.
To achieve our results we develop several new techniques and specialized data
structures which are likely of independent interest. For instance, in our determin-
istic solution we develop a novel recursive decomposition of LCE queries and for
the randomized solution we develop a new sampling technique for Karp-Rabin
fingerprints that allow fast LCE queries. We also give a general technique for
efficiently derandomizing algorithms that rely on "few" or "short" Karp-Rabin
fingerprints, and apply the technique to derandomize our Monte Carlo algo-
rithm. To the best of our knowledge, this is the first derandomization technique
for Karp-Rabin fingerprints.
Preliminaries We assume an integer alphabet, i.e., T is chosen from some al-
phabet Σ = {0, . . . , nc} for some constant c, so every character of T fits in O(1)
words. For integers a ≤ b, [a, b] denotes the range {a, a + 1, . . . , b} and we define
[n] = [0, n− 1]. For a string S = S[1]S[2] . . . S[S] and positions 1 ≤ i ≤ j ≤ S,
S[i...j] = S[i]S[i + 1]··· S[j] is a substring of length j − i + 1, S[i...] = S[i,S]
is the ith suffix of S, and S[...i] = S[1, i] is the ith prefix of S.
2 Deterministic Trade-Off
τ log n
τ ) space and O(τ log n
Here we describe a completely deterministic trade-off for the LCE problem with
O( n
τ ) query time for any τ ∈ [1, n]. Substituting
τ = τ / log(n/τ ), we obtain the bounds reflected in Table 1 for τ ∈ [1/ log n, n].
A key component in this solution is the following observation that allows us
to reduce an LCE(i, j) query on T to another query LCE(i′, j′) where i′ and j′
are both indices in either the first or second half of T .
Observation 1 Let i, j and j′ be indices of T , and suppose that LCE(j′, j) ≥
LCE(i, j). Then LCE(i, j) = min(LCE(i, j′), LCE(j′, j)).
We apply Observation 1 recursively to bring the indices of the initial query within
distance τ in O(log(n/τ )) rounds. We show how to implement each round with
a data structure using O(n/τ ) space and O(τ ) time. This leads to a solution
using O( n
τ ) query time. Finally in Section 2.4, we
show how to efficiently solve the LCE problem for indicies within distance τ in
O(n/τ ) space and O(τ ) time by exploiting periodicity properties of LCEs.
τ ) space and O(τ log n
τ log n
2.1 The Data Structure
We will store several data structures, each responsible for a specific subinterval
I = [a, b] ⊆ [1, n] of positions of the input string T . Let Ileft = [a, (a + b)/2],
Iright = ((a + b)/2, b], and I = b − a + 1. The task of the data structure for
I will be to reduce an LCE(i, j) query where i, j ∈ I to one where both indices
belong to either Ileft or Iright.
The data structure stores information for O(I/τ ) suffixes of T that start in
Iright. More specifically, we store information for the suffixes starting at positions
b − kτ ∈ Iright, k = 0, 1, . . . , (I/2)/τ . We call these the sampled positions of
Iright. See Figure 1 for an illustration.
a
j ′
k
a+b
2
τ
b
Ileft
Iright
Fig. 1. Illustration of the contents of the data structure for the interval I = [a, b]. The
black dots are the sampled positions in Iright, and each such position has a pointer to
an index j ′
k ∈ Ileft.
For every sampled position b − kτ ∈ Iright, k = 0, 1, . . . , (I/2)/τ , we store
the index j′
k of the suffix starting in Ileft that achieves the maximum LCE value
with the suffix starting at the sampled position, i.e., T [b − kτ...] (ties broken
arbitrarily). Along with j′
k, we also store the value of the LCE between suffix
T [j′
k and Lk are defined as follows,
k...] and T [b − kτ...]. Formally, j′
LCE(h, b − kτ )
j′
k = argmax
h ∈ Ileft
and Lk = LCE(j′
k, b − kτ ) .
Building the structure We construct the above data structure for the interval
[1, n], and build it recursively for [1, n/2] and (n/2, n], stopping when the length
of the interval becomes smaller than τ .
2.2 Answering a Query
We now describe how to reduce a query LCE(i, j) where i, j ∈ I to one where
both indices are in either Ileft or Iright. Suppose without loss of generality that
i ∈ Ileft and j ∈ Iright. We start by comparing δ < τ pairs of characters of T ,
starting with T [i] = T [j], until 1) we encounter a mismatch, 2) both positions
are in Iright or 3) we reach a sampled position in Iright. It suffices to describe the
last case, in which T [i, i + δ] = T [j, j + δ], i + δ ∈ Ileft and j + δ = b − kτ ∈ Iright
for some k. Then by Observation 1, we have that
LCE(i, j) = δ + LCE(i + δ, j + δ)
= δ + min(LCE(i + δ, j′
= δ + min(LCE(i + δ, j′
k), LCE(j′
k), Lk) .
k, b − kτ ))
Thus, we have reduced the original query to computing the query LCE(i + δ, j′
k)
in which both indices are in Ileft.
2.3 Analysis
Each round takes O(τ ) time and halves the upper bound for i−j, which initially
is n. Thus, after O(τ log(n/τ )) time, the initial LCE query has been reduced to
one where i − j ≤ τ . At each of the O(log(n/τ )) levels, the number of sampled
positions is (n/2)/τ , so the total space used is O((n/τ ) log(n/τ )).
2.4 Queries with Nearby Indices
We now describe the data structure used to answer a query LCE(i, j) when
i − j ≤ τ . We first give some necessary definitions and properties of periodic
strings. We say that the integer 1 ≤ p ≤ S is a period of a string S if any
two characters that are p positions apart in S match, i.e., S[i] = S[i + p] for all
positions i s.t. 1 ≤ i < i + p ≤ S. The following is a well-known property of
periods.
Lemma 1 (Fine and Wilf [5]). If a string S has periods a and b and S ≥
a + b − gcd(a, b), then gcd(a, b) is also a period of S.
The period of S is the smallest period of S and we denote it by per(S). If
per(S) ≤ S/2, we say S is periodic. A periodic string S might have many
periods smaller than S/2, however it follows from the above lemma that
Corollary 1. All periods smaller than S/2 are multiples of per(S).
The Data Structure Let Tk = T [kτ...(k + 2)τ − 1] denote the substring of
length 2τ starting at position kτ in T , k = 0, 1, . . . , n/τ . For the strings Tk that
are periodic, let pk = per(Tk) be the period. For every periodic Tk, the data
structure stores the length ℓk of the maximum substring starting at position kτ ,
which has period pk. Nothing is stored if Tk is aperiodic.
Answering a Query We may assume without loss of generality that i = kτ , for
some integer k. If not, then we check whether T [i + δ] = T [j + δ] until i + δ = kτ .
Hence, assume that i = kτ and j = i + d for some 0 < d ≤ τ . In O(τ ) time, we
first check whether T [i + δ] = T [j + δ] for all δ ∈ [0, 2τ ]. If we find a mismatch
we are done, and otherwise we return LCE(i, j) = ℓk − d.
Correctness If a mismatch is found when checking that T [i + δ] = T [j + δ]
for all δ ∈ [0, 2τ ], the answer is clearly correct. Otherwise, we have established
that d ≤ τ is a period of Tk, so Tk is periodic and d is a multiple of pk (by
Corollary 1). Consequently, T [i + δ] = T [j + δ] for all δ s.t. d + δ ≤ ℓk, and thus
LCE(i, j) = ℓk − d.
2.5 Preprocessing
The preprocessing of the data structure is split into the preprocessing of the
recursive data structure for queries where i − j > τ and the preprocessing of
the data structure for nearby indices, i.e., i − j ≤ τ .
If we allow O(n) space during the preprocessing phase then both can be
solved in O(n) time. If we insist on O(n/τ ) space during preprocessing then the
times are not as good.
Recursive Data Structure Say O(n) space is allowed during preprocessing.
Generate a 2D range searching environment with a point (x(l), y(l)) for every
text index l ∈ [1, n]. Set x(l) = l and y(l) equal to the rank of suffix T [l...] in the
lexicographical sort of the suffixes. To compute j′
k ∈ Ileft = [a, (a+ b)/2] that has
the maximal LCE value with a sampled position i = b−kτ ∈ Iright = ((a+b)/2, b]
we require two 3-sided range successor queries, see detailed definitions in [17].
The first is a range [a, (a + b)/2]× (−∞, y(i)− 1] which returns the point within
the range with y-coordinate closest to, and less-than, y(i). The latter is a range
[a, (a + b)/2]× [y(i) + 1,∞) with point with y-coordinate closest to, and greater-
than, y(i) returned.
Let l′ and l′′ be the x-coordinates, i.e., text indices, of the points returned by
the queries. By definition both are in [a, (a + b)/2]. Among suffixes starting in
[a, (a + b)/2], T [l′...] has the largest rank in the lexicographic sort which is less
than l. Likewise, T [l′′...] has the smallest rank in the lexicographic sort which is
greater than l. Hence, j′
k is equal to either l′ or l′′, and the larger of LCE(l, l′)
and LCE(l, l′′) determines which one.
The LCE computation can be implemented with standard suffix data struc-
tures in O(n) space and O(1) query time for an overall O(n) time. 3-sided range
successor queries can be preprocessed in O(n log n) space and time and queries
can be answered in O(log log n) time, see [11, 16] for an overall O(n log log n)
time and O(n log n) space.
If one desires to keep the space limited to O(n/τ ) space then generate a
sparse suffix array for the (evenly spaced) suffixes starting at kτ in Iright, k =
0, 1, . . . , (I/2)/τ , see [9]. Now for every suffix in Ileft find its location in the
sparse suffix array. From this data one can compute the desired answer. The
time to compute the sparse suffix array is O(n) and the space is O(n/τ ). The
time to search a suffix in Ileft is O(n + log(I/τ )) using the classical suffix search
in the suffix array [20]. The time per I is O(In). Overall levels this yields
O(n · n + n · (n/2) + n · (n/4) + ··· + n · 1) = O(n2) time.
The Nearby Indices Data Structure If Tk is periodic then we need to find its
period pk and find the length ℓk of the maximum substring starting at position
kτ , which has period pk.
Finding period pk, if it exists, in Tk can be done in O(τ ) time and constant
space using the algorithm of [3]. The overall time is O(n) and the overall space
is O(n/τ ) for the pk's.
To compute ℓk we check whether the period pk of Tk extends to Tk+2. That
is whether pk is the period of T [kτ...(k + 4)τ − 1]. The following is why we do
so.
Lemma 2. Suppose Tk and Tk+2 are periodic with periods pk and pk+2, respec-
tively. If Tk · Tk+2 is periodic with a period of length at most τ then pk is the
period and pk+2 is a rotation of pk.
Proof. To reach a contradition, suppose that pk is not the period of Tk · Tk+2,
i.e., it has period p′ of length at most τ . Since Tk is of size 2τ , p′ must also be a
period of Tk. However, pk and p′ are both prefixes of Tk and, hence, must have
different lengths. By Corollary 1 p′ must be a multiple of pk and, hence, not
a period. The fact that pk+2 is a rotation of pk can be deduced similarly.
⊓⊔
By induction it follows that Tk · Tk+2 · Tk+4 · . . . · Tk+2M , M ≥ 1, is periodic
with pk if each concatenated pair of Tk+2i · Tk+2(i+1) is periodic with a period at
most τ . This suggests the following scheme. Generate a binary array of length
n/τ where location j is 1 iff Tk+2(j−1) · Tk+2j has a period of length of at most
τ . In all places where there is a 0, i.e., Tk · Tk+2 does not have a period of
length at most τ and Tk does have period pk, we have that ℓk ∈ [2τ, 4τ − 1]. We
directly compute ℓk by extending pk in Tk · Tk+2 as far as possible. Now, using
the described array and the ℓk ∈ [2τ, 4τ − 1], in a single sweep for the even kτ 's
(and a seperate one for the odd ones) from right to left we can compute all ℓk's.
It is easy to verify that the overall time is O(n) and the space is O(n/τ ).
3 Randomized Trade-Offs
In this section we describe a randomized LCE data structure using O(n/τ ) space
with O(τ + log ℓ
τ ) query time. In Section 3.6 we describe another O(n/τ )-space
LCE data structure that either answers an LCE query in constant time, or
provides a certificate that ℓ ≤ τ 2. Combining the two data structures, shows
that the LCE problem can be solved in O(n/τ ) space and O(τ ) time.
The randomization comes from our use of Karp-Rabin fingerprints [10] for
comparing substrings of T for equality. Before describing the data structure,
we start by briefly recapping the most important definitions and properties of
Karp-Rabin fingerprints.
3.1 Karp-Rabin Fingerprints
For a prime p and x ∈ [p] the Karp-Rabin fingerprint [10], denoted φp,x(T [i...j]),
of the substring T [i...j] is defined as
φp,x(T [i...j]) = X
i≤k≤j
T [k]xk−i mod p .
If T [i...j] = T [i′...j′] then clearly φp,x(T [i...j]) = φp,x(T [i′...j′]). In the Monte
Carlo and the Las Vegas algorithms we present we will choose p such that p =
Θ(n4+c) for some constant c > 0 and x uniformly from [p]\ {0}. In this case a
simple union bound shows that the converse is also true with high probability,
i.e., φ is collision-free on all substring pairs of T with probability at least 1−n−c.
Storing a fingerprint requires O(1) space. When p, x are clear from the context
we write φ = φp,x.
For shorthand we write f (i) = φ(T [1, i]), i ∈ [1, n] for the fingerprint of the
ith prefix of T . Assuming that we store the exponent xi mod p along with the
fingerprint f (i), the following two properties of fingerprints are well-known and
easy to show.
Lemma 3. 1) Given f (i), the fingerprint f (i ± a) for some integer a, can be
computed in O(a) time. 2) Given fingerprints f (i) and f (j), the fingerprint
φ(T [i..j]) can be computed in O(1) time.
In particular this implies that for a fixed length l, the fingerprint of all substrings
of length l of T can be enumerated in O(n) time using a sliding window.
3.2 Overview
The main idea in our solution is to binary search for the LCE(i, j) value using
Karp-Rabin fingerprints. Suppose for instance that φ(T [i, i + M ]) 6= φ(T [j, j +
M ]) for some integer M , then we know that LCE(i, j) ≤ M , and thus we can
find the true LCE(i, j) value by comparing log(M ) additional pair of fingerprints.
The challenge is to obtain the fingerprints quickly when we are only allowed to
use O(n/τ ) space. We will partition the input string T into n/τ blocks each of
length τ . Within each block we sample a number of equally spaced positions.
The data structure consists of the fingerprints of the prefixes of T that ends
at the sampled positions, i.e., we store f (i) for all sampled positions i. In total
we sample O(n/τ ) positions. If we just sampled a single position in each block
(similar to the approach in [2]), we could compute the fingerprint of any substring
in O(τ ) time (see Lemma 3), and the above binary search algorithm would take
time O(τ log n) time. We present a new sampling technique that only samples an
additional O(n/τ ) positions, while improving the query time to O(τ + log(ℓ/τ )).
Preliminary definitions We partition the input string T into n/τ blocks of τ
positions, and by block k we refer to the positions [kτ, kτ + τ ), for k ∈ [n/τ ].
We assume without loss of generality that n and τ are both powers of two.
Every position q ∈ [1, n] can be represented as a bit string of length lg n. Let
q ∈ [1, n] and consider the binary representation of q. We define the leftmost
lg(n/τ ) bits and rightmost lg(τ ) bits to be the head, denoted h(q) and the tail,
denoted t(q), respectively. A position is block aligned if t(q) = 0. The significance
of q, denoted s(q), is the number of trailing zeros in h(q). Note that the τ
positions in any fixed block k ∈ [n/τ ] all have the same head, and thus also the
same significance, which we denote by µk. See Figure 2.
h(q)
t(q)
q
0 1 1 0 0 1 0 0 0 0 0 1 1 1 0 1 0 1 1
s(q)
Fig. 2. Example of the definitions for the position q = 205035 in a string of length
n = 219 with block length τ = 28. Here h(q) is the first lg(n/τ ) = 11 bits, and t(q) is
the last lg(τ ) = 8 bits in the binary representation of q. The significance is s(q) = 5.
3.3 The Monte Carlo Data Structure
The data structure consists of the values f (i), i ∈ S, for a specific set of sampled
positions S ⊆ [1, n], along with the information necessary in order to look up
the values in constant time. We now explain how to construct the set S. In
block k ∈ [n/τ ] we will sample bk = min(cid:8)2⌊µk/2⌋, τ(cid:9) evenly spaced positions,
where µk is the significance of the positions in block k, i.e., µk = s(kτ ). More
precisely, in block k we sample the positions Bk = {kτ + jτ /bk j ∈ [bk]}, and
let S = ∪k∈[n/τ ] Bk. See Figure 3.
We now bound the size of S. The significance of a block is at most lg(n/τ ),
and there are exactly 2lg(n/τ )−µ blocks with significance µ, so
n/τ −1
lg(n/τ )
S =
X
k=0
bk ≤
X
µ=0
2lg(n/τ )−µ2⌊µ/2⌋ ≤
n
τ
∞
X
µ=0
2−µ/2 = (cid:16)2 + √2(cid:17) n
τ
= O (cid:16) n
τ (cid:17) .
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
8.0
9.0
10.0
11.0
12.0
13.0
14.0
15.0
4
0
1
0
2
0
1
0
3
0
1
0
2
0
1
0
4.0
1.0
1.0
1.0
2.0
1.0
1.0
1.0
2.0
1.0
1.0
1.0
2.0
1.0
1.0
1.0
k
µk
bk
T
Fig. 3. Illustration of a string T partitioned into 16 blocks each of length τ . The
significance µk for the positions in each block k ∈ [n/τ ] is shown, as well as the bk
values. The block dots are the sampled positions S.
3.4 Answering a query
We now describe how to answer an LCE(i, j) query. We will assume that i is
block aligned, i.e., i = kτ for some k ∈ [n/τ ]. Note that we can always obtain this
situation in O(τ ) time by initially comparing at most τ − 1 pairs of characters
of the input string directly.
Algorithm 1 shows the query algorithm. It performs an exponential search to
locate the block in which the first mismatch occurs, after which it scans the block
directly to locate the mismatch. The search is performed by calls to check(i, j, c),
which computes and compares φ(T [i...i + c]) and φ(T [j...j + c]). In other words,
assuming that φ is collision-free, check(i, j, c) returns true if LCE(i, j) ≥ c and
false otherwise.
⊲ Compute an interval such that ℓ ∈ [ℓ, 2ℓ].
(i, j, ℓ) ← (i + 2µτ, j + 2µτ, ℓ + 2µτ )
if s(j) > µ then
ℓ ← 0
µ ← 0
while check(i, j, 2µτ ) do
Algorithm 1 Computing the answer to a query LCE(i, j)
1: procedure LCE(i, j)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
µ ← µ − 1
(i, j, ℓ) ← (i + 1, j + 1, ℓ + 1)
(i, j, ℓ) ← (i + 2µ−1τ, j + 2µ−1τ, ℓ + 2µ−1τ )
return ℓ
µ ← µ + 1
while µ > 0 do
if check(i, j, 2µ−1τ ) then
⊲ Identify the block in which the first mismatch occurs
while T [i] = T [j] do ⊲ Scan the final block left to right to find the mismatch
Analysis We now prove that Algorithm 1 correctly computes ℓ = LCE(i, j) in
O(τ + log(ℓ/τ )) time. The algorithm is correct assuming that check(i, j, 2µτ )
always returns the correct answer, which will be the case if φ is collision-free.
The following is the key lemma we need to bound the time complexity.
Lemma 4. Throughout Algorithm 1 it holds that ℓ ≥ (2µ − 1) τ , s(j) ≥ µ, and
µ is increased in at least every second iteration of the first while-loop.
Proof. We first prove that s(j) ≥ µ. The claim holds initially. In the first loop j
is changed to j + 2µτ , and s(j + 2µτ ) ≥ min{s(j), s(2µτ )} = min{s(j), µ} = µ,
where the last equality follows from the induction hypothesis s(j) ≥ µ. Moreover,
µ is only incremented when s(j) > µ. In the second loop j is changed to j+2µ−1τ ,
which under the assumption that s(j) ≥ µ, has significance s(j + 2µ−1τ ) = µ− 1.
Hence the invariant is restored when µ is decremented at line 11.
Now consider an iteration of the first loop where µ is not incremented, i.e.,
is even, and hence s(j+2µτ ) > µ,
s(j) = µ. Then j
so µ will be incremented in the next iteration of the loop.
2µτ is an odd integer, i.e. j+2µτ
In order to prove that ℓ ≥ (2µ − 1) τ we will prove that ℓ ≥ (2µ − 1) τ in the
first loop. This is trivial by induction using the observation that (2µ − 1) τ +
2µτ = (cid:0)2µ+1 − 1(cid:1) τ .
⊓⊔
Since ℓ ≥ (2µ − 1)τ and µ is increased at least in every second iteration of the
first loop and decreased in every iteration of the second loop, it follows that
2µτ
there are O(log(ℓ/τ )) iterations of the two first loops. The last loop takes O(τ )
time. It remains to prove that the time to evaluate the O(log(ℓ/τ )) calls to
check(i, j, 2µτ ) sums to O(τ + log(ℓ/τ )).
Evaluating check(i, j, 2µτ ) requires computing φ(T [i...i+2µτ ]) and φ(T [j...j+
2µτ ]). The first fingerprint can be computed in constant time because i and
i + 2µτ are always block aligned (see Lemma 3). The time to compute the second
fingerprint depends on how far j and j + 2µτ each are from a sampled position,
which in turn depends inversely on the significance of the block containing those
positions. By Lemma 4, µ is always a lower bound on the significance of j, which
implies that µ also lower bounds the significance of j + 2µτ , and thus by the way
we sample positions, neither will have distance more than τ /2⌊µ/2⌋ to a sam-
pled position in S. Finally, note that by the way µ is increased and decreased,
check(i, j, 2µτ ) is called at most three times for any fixed value of µ. Hence, the
total time to compute all necessary fingerprints can be bounded as
O
lg(ℓ/τ )
X
µ=0
1 + τ /2⌊µ/2⌋
= O(τ + log(ℓ/τ )) .
3.5 The Las Vegas Data Structure
We now describe an O(cid:0)n3/2(cid:1)-time and O(n/τ )-space algorithm for verifiying
that φ is collision-free on all pairs of substrings of T that the query algorithm
compares. If a collision is found we pick a new φ and try again. With high
probability we can find a collision-free φ in a constant number of trials, so we
obtain the claimed Las Vegas data structure.
If τ ≤ √n we use the verification algorithm of Bille et al. [2], using O(nτ +
n log n) time and O(n/τ ) space. Otherwise, we use the simple O(n2/τ )-time and
O(n/τ )-space algorithm described below.
Recall that all fingerprint comparisions in our algorithm are of the form
φ(cid:0)T [kτ...kτ + 2lτ − 1](cid:1) ?= φ(cid:0)T [j...j + 2lτ − 1](cid:1)
for some k ∈ [n/τ ], j ∈ [n], l ∈ [log(n/τ )]. The algorithm checks each l ∈
[log(n/τ )] separately. For a fixed l it stores the fingerprints φ(T [kτ...kτ + 2lτ ])
for all k ∈ [n/τ ] in a hash table H. This can be done in O(n) time and O(n/τ )
space. For every j ∈ [n] the algorithm then checks whether φ(cid:0)T [j...j + 2lτ ](cid:1) ∈ H,
and if so, it verifies that the underlying two substrings are in fact the same by
comparing them character by character in O(2lτ ) time. By maintaining the fin-
gerprint inside a sliding window of length 2lτ , the verification time for a fixed l
becomes O(n2lτ ), i.e., O(n2/τ ) time for all l ∈ [log(n/τ )].
3.6 Queries with Long LCEs
In this section we describe an O( n
τ ) space data structure that in constant time
either correctly computes LCE(i, j) or determines that LCE(i, j) ≤ τ 2. The data
structure can be constructed in O(n log n
τ ) time by a Monte Carlo or Las Vegas
algorithm.
The Data Structure Let Sτ ⊆ [1, n] called the sampled positions of T (to
be defined below), and consider the sets A and B of suffixes of T and T R,
respectively.
A = {T [i...] i ∈ Sτ}
, B = {T [...i]R i ∈ Sτ} .
We store a data structure for A and B, that allows us to perform constant time
longest common extension queries on any pair of suffixes in A or any pair in
B. This can be achieved by well-known techniques, e.g., storing a sparse suffix
tree for A and B, equipped with a nearest common ancestor data structure. To
define Sτ , let Dτ = {0, 1, . . . , τ} ∪ {2τ, . . . , (τ − 1)τ}, then
Sτ = {1 ≤ i ≤ n i mod τ 2 ∈ Dτ} .
(1)
Answering a Query To answer an LCE query, we need the following def-
initions. For i, j ∈ Sτ let LCER(i, j) denote the longest common prefix of
T [...i]R ∈ B and T [...j]R ∈ B. Moreover, for i, j ∈ [n], we define the function
δ(i, j) = (cid:0)((i − j) mod τ ) − i(cid:1) mod τ 2 .
(2)
We will write δ instead of δ(i, j) when i and j are clear from the context.
The following lemma gives the key property that allows us to answer a query.
Lemma 5. For any i, j ∈ [n − τ 2], it holds that i + δ, j + δ ∈ Sτ .
Proof. Direct calculation shows that (i + δ) mod τ 2 ≤ τ , and that (j + δ)
mod τ = 0, and thus by definition both i + δ and j + δ are in Sτ .
To answer a query LCE(i, j), we first verify that i, j ∈ [n − τ 2] and that
LCER(i+δ, j+δ) ≥ δ. If this is not the case, we have established that LCE(i, j) ≤
δ < τ 2, and we stop. Otherwise, we return δ + LCE(i + δ, j + δ) − 1.
Analysis To prove the correctness, suppose i, j ∈ [n − τ 2] (if not clearly
LCE(i, j) < τ 2) then we have that i + δ, j + δ ∈ Sτ (Lemma 5). If LCER(i + δ, j +
δ) ≥ δ it holds that T [i...i + δ] = T [j...j + δ] so the algorithm correctly computes
LCE(i, j) as δ + 1 + LCE(i + δ, j + δ). Conversely, if LCER(i + δ, j + δ) < δ,
T [i...i + δ] 6= T [j...j + δ] it follows that LCE(i, j) < δ < τ 2.
Query time is O(1), since computing δ, LCER(i+δ, j +δ) and LCE(i+δ, j +δ)
all takes constant time. Storing the data structures for A and B takes space
O(A + B) = O(Sτ) = O( n
τ ). For the preprocessing stage, we can use recent
algorithms by I et al. [8] for constructing the sparse suffix tree for A and B
in O( n
τ ) time
(correct w.h.p.), and a Las Vegas algorithm using O( n
τ ) space. They provide a Monte Carlo algorithm using O(n log n
τ ) time (w.h.p.).
4 Derandomizing the Monte Carlo Data Structure
Here we give a general technique for derandomizing Karp-Rabin fingerprints,
and apply it to our Monte Carlo algorithm. The main result is that for any con-
stant ε > 0, the data structure can be constructed completely deterministically
in O(n2+ε) time using O(n/τ ) space. Thus, compared to the probabilistic pre-
processing of the Las Vegas structure using O(n3/2) time with high probability,
it is relatively cheap to derandomize the data structure completely.
Our derandomizing technique is stated in the following lemma.
Lemma 6. Let A, L ⊂ {1, 2, . . . , n} be a set of positions and lengths respectively
such that max(L) = nΩ(1). For every ε ∈ (0, 1), there exist a fingerprinting
function φ that can be evaluated in O (cid:0) 1
ε(cid:1) time and has the property that for all
a ∈ A, l ∈ L, i ∈ {1, 2, . . . , n}:
φ(T [a...a + (l− 1)]) = φ(T [i...i + (l− 1)]) ⇐⇒ T [a...a + (l− 1)] = T [i...i + (l− 1)]
S max(L)L(cid:17) time,
We can find such a φ using O (cid:0) S
for any value of S ∈ [1,A].
Proof. We let p be a prime contained in the interval [max(L)nε, 2 max(L)nε].
The idea is to choose φ to be φ = (φp,x1 , . . . , φp,xk ) , k = O(1/ε), where φp,xi is
the Karp-Rabin fingerprint with parameters p and xi.
Let Σ be the alphabet containing the characters of T . If p ≤ Σ we note that
since p = nΩ(1) and Σ = nO(1) we can split each character into O(1) characters
from a smaller alphabet Σ′ satisfying p > Σ′. So wlog. assume p > Σ from
now on.
ε (cid:1) space and O (cid:16) n1+ε log n
A
ε2
We let S be the set defined by:
S = {(T [a . . . a + (l − 1)], T [i . . . i + (l − 1)]) a ∈ A, l ∈ L, i ∈ {1, 2, . . . , n}}
Then we have to find φ such that φ(u) = φ(v) ⇐⇒ u = v for all (u, v) ∈ S.
We will choose φp,x1 , φp,x2, . . . , φp,xk successively. For any 1 ≤ l ≤ k we let
φ≤l = (φp,x1, . . . , φp,xl).
For any fingerprinting function f we denote by B(f ) the number of pairs
(u, v) ∈ S such that f (u) = f (v). We let B(id) denote the number of pairs
(u, v) ∈ S such that u = v. Hence B(f ) ≥ B(id) for every fingerprinting function
f , and f satisfies the requirement iff B(f ) = B(id).
Say that we have chosen φ≤l for some l < k. We now compare B(φ≤l+1) =
B((φp,xl+1 , φ≤l)) to B(φ≤l) when xl+1 is chosen uniformly at random from [p].
For any pair (u, v) ∈ S such that u 6= v and φ≤l(u) = φ≤l(v) we see that, if xl+1 is
chosen randomly independently then the probability that φp,xl+1(u) = φp,xl+1(v)
is at most max{u,v}
p
≤ n−ε. Hence
Exl+1(B(φ≤l+1) − B(id)) ≤
B(φ≤l) − B(id)
nε
,
and thus there exists xl+1 ∈ {0, 1, . . . , p − 1} such that
B(φ≤l+1) − B(id)) ≤
B(φ≤l) − B(id)
nε
.
Now consider the algorithm where we construct φ successively by choosing xl+1
such that B(φ≤l+1) is as small as possible. Then4
B(φ≤k−1) − B(id)
nε
B(1) − B(id)
.
nkε
≤ . . . ≤
ε L n log n A
B(φ) − B(id) = B(φ≤k) − B(id) ≤
Since B(1) ≤ n3 and B(φ) − B(id) is an integer it suffices to choose k = (cid:6) 4
order to conclude that B(φ) = B(id).
ε(cid:7) in
We just need to argue that for given x1, . . . , xl we can find B(φ≤l) in space
S (cid:17). First we show how to do it in the case S =
ε S(cid:1) and time O (cid:16) 1
O (cid:0) 1
A. To this end we will count the number of collisions for all (u, v) ∈ S such that
u = v = l ∈ L for some fixed l ∈ L. First we compute φ≤l(T [a . . . a+(l−1]) for
all a ∈ A and store them in a multiset. Using a sliding window the fingerprints
can be computed and inserted in time O (ln log n) Now for each position i ∈ [1, n]
we compute φ≤l(T [i . . . i + (l− 1]) and count the number of times the fingerprint
appears in the multiset. Using a sliding window this can be done in O (ln log n)
time as well. Doing this for all l ∈ L gives the number of collisions.
S (cid:17) and partition
A into A1, . . . , Ar consisting of ≤ S elements each. For each j = 1, 2, . . . , r we
define Sj as
Now assume that S < A. Then let r = l A
S m = O (cid:16) A
Sj = {(T [a . . . a + (l − 1)], T [i . . . i + (l − 1)]) a ∈ Aj, l ∈ L, i ∈ {1, 2, . . . , n}}
Then S1, . . . , Sr is a partition of S. For each j = 1, 2, . . . , r we can find the
collisions in among the elements in Sj in the manner described above. Doing
this for each j and adding the results is sufficient to get the desired space and
⊓⊔
time complexity.
Corollary 2. For any τ ∈ [1, n], the LCE problem can be solved by a deter-
ministic data structure with O(n/τ ) space usage and O(τ ) query time. The data
structure can be constructed in O(n2+ε) time using O(n/τ ) space.
Proof. We use the lemma with A = {kτ k ∈ [n/τ ]}, L = {2lτ l ∈ [log(n/τ )]},
S = A = n/τ and a suitable small constant ε > 0.
References
1. A. Amir, M. Lewenstein, and E. Porat. Faster algorithms for string matching with
k mismatches. J. Algorithms, 50(2):257 -- 275, 2004.
4 B(1) is equal to S since 1 is the constant function
2. P. Bille, I. L. Gørtz, B. Sach, and H. W. Vildhøj. Time-space trade-offs for longest
common extensions. J. of Discrete Algorithms, 25:42 -- 50, 2014.
3. D. Breslauer, R. Grossi, and F. Mignosi. Simple real-time constant-space string
matching. Theor. Comput. Sci., 483:2 -- 9, 2013.
4. R. Cole and R. Hariharan. Approximate String Matching: A Simpler Faster Algo-
rithm. SIAM J. Comput., 31(6):1761 -- 1782, 2002.
5. N. J. Fine and H. S. Wilf. Uniqueness Theorems for Periodic Functions. Proc.
AMS, 16(1):109 -- 114, 1965.
6. D. Gusfield and J. Stoye. Linear time algorithms for finding and representing all
the tandem repeats in a string. J. Comput. Syst. Sci., 69:525 -- 546, 2004.
7. D. Harel and R. E. Tarjan. Fast algorithms for finding nearest common ancestors.
SIAM J. Comput., 13(2):338 -- 355, 1984.
8. T. I, J. Karkkainen, and D. Kempa. Faster Sparse Suffix Sorting. In Proc. 31st
STACS, volume 25, pages 386 -- 396, Dagstuhl, Germany, 2014.
9. J. Karkkainen and E. Ukkonen. Sparse suffix trees. In Computing and Combina-
torics, Second Annual International Conference, COCOON '96, Hong Kong, June
17-19, 1996, Proceedings, pages 219 -- 230, 1996.
10. R. M. Karp and M. O. Rabin. Efficient randomized pattern-matching algorithms.
IBM J. Res. Dev., 31(2):249 -- 260, 1987.
11. O. Keller, T. Kopelowitz, and M. Lewenstein. Range non-overlapping indexing and
successive list indexing. In Proc. of Workshop on Algorithms and Data Structures
(WADS), pages 625 -- 636, 2007.
12. R. Kolpakov and G. Kucherov. Searching for gapped palindromes. In Proc. 19th
CPM, LNCS, volume 5029, pages 18 -- 30, 2008.
13. G. M. Landau, E. W. Myers, and J. P. Schmidt. Incremental string comparison.
SIAM J. Comput., 27(2):557 -- 582, 1998.
14. G. M. Landau and J. P. Schmidt. An Algorithm for Approximate Tandem Repeats.
J. Comput. Biol., 8(1):1 -- 18, 2001.
15. G. M. Landau and U. Vishkin. Fast Parallel and Serial Approximate String Match-
ing. J. Algorithms, 10:157 -- 169, 1989.
16. H.-P. Lenhof and M. H. M. Smid. Using persistent data structures for adding
range restrictions to searching problems. Theoretical Informatics and Applications
(ITA), 28(1):25 -- 49, 1994.
17. M. Lewenstein. Orthogonal range searching for text indexing. In Space-Efficient
Data Structures, Streams, and Algorithms, pages 267 -- 302, 2013.
18. M. G. Main and R. J. Lorentz. An O (n log n) algorithm for finding all repetitions
in a string. J. Algorithms, 5(3):422 -- 432, 1984.
19. G. Manacher. A New Linear-Time "On-Line" Algorithm for Finding the Smallest
Initial Palindrome of a String. J. ACM, 22(3):346 -- 351, 1975.
20. U. Manber and E. W. Myers. Suffix arrays: A new method for on-line string
searches. SIAM J. Comput., 22(5):935 -- 948, 1993.
21. E. W. Myers. An O(N D) difference algorithm and its variations. Algorithmica,
1(2):251 -- 266, 1986.
22. P. Weiner. Linear pattern matching algorithms.
In Proc. 14th FOCS (SWAT),
pages 1 -- 11, 1973.
|
1207.0933 | 1 | 1207 | 2012-07-04T10:02:15 | Optimal Cuts and Bisections on the Real Line in Polynomial Time | [
"cs.DS",
"cs.CC",
"cs.DM"
] | The exact complexity of geometric cuts and bisections is the longstanding open problem including even the dimension one. In this paper, we resolve this problem for dimension one (the real line) by designing an exact polynomial time algorithm. Our results depend on a new technique of dealing with metric equalities and their connection to dynamic programming. The method of our solution could be also of independent interest. | cs.DS | cs |
Optimal Cuts and Bisections on the Real Line
in Polynomial Time
Marek Karpinski∗
Andrzej Lingas†
Dzmitry Sledneu‡
Abstract
The exact complexity of geometric cuts and bisections is the longstanding open
problem including even the dimension one. In this paper, we resolve this problem for
dimension one (the real line) by designing an exact polynomial time algorithm. Our
results depend on a new technique of dealing with metric equalities and their connection
to dynamic programming. The method of our solution could be also of independent
interest.
1
Introduction
The metric MAX-CUT, MAX-BISECTION, MIN-BISECTION and other Partitioning prob-
lems were all proved to have polynomial time approximation schemes (PTAS) [5, 3, 1, 4, 2, 6].
The above problems are known to be NP-hard in exact setting. The status of those problems
for geometric (thus including Euclidean) metrics and this even for dimension one was widely
open.
In this paper we resolve the status of those problems for just dimension one by giving a
polynomial time algorithm. Our solution, somewhat surprisingly, involves certain new ideas
for applying dynamic programming which could be also of independent interest.
2 Preliminaries and General Setting
We shall define our dynamic programming method in terms of generalized subproblems on
finite multisets of reals generalizing slightly a geometric metric setting.
For a partition of a finite multiset P of reals into two multisets P1 and P2, the value of
the cut is the total length of all intervals on the real line that have one endpoint in P1 and
the other one in P2.
∗Research supported partly by DFG grants and the Hausdorff Center grant EXC59-1. Department of
Computer Science, University of Bonn. Email: [email protected]
†Research supported in part by VR grant 621-2008-4649. Department of Computer Science, Lund Uni-
versity. Email: [email protected]
‡Centre for Mathematical Sciences, Lund University. Email: [email protected].
1
The MAX-CUT problem for P will be now to find a partition of P into two multisets that
maximizes the value of the cut. If P = n and the two multisets are additionally required to
be of cardinality k and n− k, respectively, then we obtain the (k, n− k) MAX-PARTITION
problem for P. In particular, if n is even and k = n/2 then we have the MAX-BISECTION
problem. Next, if we replace the requirement of maximization with that of minimization
then we obtain the (k, n − k) MIN-PARTITION and MIN-BISECTION problems for P ,
respectively.
In this paper we study geometric instances of the above problems in dimension one (the
real line) which could be rephrased as the problems of partitioning arbitrary finite metric
spaces of this dimension.
3 The Algorithm
The global idea behind our algorithm is as follows. We guess how many of the copies of
the rightmost real are respectively in the first and second set and move them to the next
to the rightmost real. We also guess how many copies of the reals in the remaining part of
the multiset are in the first and second set respectively. Having this information, we can
compute exactly the difference between the value of optimal solution for the whole input
multiset and that for the multiset resulting from the movement under the guessed partition
proportions. The difference can be easily evaluated due to the triangle equality that holds
on the real line. Now, we solve the problem for the transformed multiset whose elements are
copies of the shrunk set of reals recursively under the guessed partition proportions, after
additionally guessing the proportions of the partition for the original copies of the next to
the right real. Eventually, we end up with the trivial multiset where all elements are copies
of the leftmost real. We eliminate guesses by an enumeration of all possibilities and choosing
the best one. Next, we solve the resulting subproblems in bottom up fashion instead the top
down to obtain a polynomial time solution.
Consider a finite multiset P of reals. We assume that P = n and that P consists of
copies of l ≤ n distinct reals. For i = 1, ..., l, let xi denote the i-th smallest real whose copy
is in P, and let Pi denote the sub-multiset of P consisting of all elements of P which are
copies of reals in {x1, ..., xi}. For convention, we assume P0 = ∅.
We shall consider a family of generalized subproblems Si(p, q, r, t), where i ∈ {1, ..., l},
p, q, r, t are integers in {0, ..., n} such that p + q = Pi−1 and r + t = n − Pi−1. The
subproblem Si(p, q, r, t) is to find for a multiset that is the union of Pi−1 with r + t copies of
xi a partition into two multisets such that p elements of Pi and r copies of xi form the first
set and the value of the cut is maximized. The value of such a maximum cut is denoted by
M AXCU T (Si(p, q, r, t)).
Lemma 1. For i ≥ 2, M AXCU T (Si(p, q, r, t)) = (xi − xi−1)(pt + qr)+
+ maxr0≤p∧t0≤q∧r0+t0=Pi−1−Pi−2 M AXCU T (Si−1(p − r0, q − t0, r0 + r, t0 + t)).
Proof. For 1 ≤ r0 ≤ p ∧ 1 ≤ t0 ≤ q, consider an optimal solution to Si−1(p − r0, q − t0, r0 +
r, t0 + t), where r0 + t0 = Pi−1 − Pi−2. Let us move r copies of the real xi−1 in the first
2
Figure 1: Si−1(p − r0, q − t0, r0 + r, t0 + t).
Figure 2: Si(p, q, r, t).
set of the solution to the real xi and t copies of xi−1 in the second set of the solution to the
same real xi. We obtain a feasible solution to Si(p, q, r, t) whose cut value is (xi − xi−1)(pt +
qr) + M AXCU T (Si−1(p− r0, q− t0, r0 + r, t0 + t)). It follows that M AXCU T (Si(p, q, r, t)) ≥
(xi − xi−1)(pt + qr)+
+ maxr0≤p∧t0≤q∧r0+t0=Pi−1−Pi−2 M AXCU T (Si−1(p − r0, q − t0, r0 + r, t0 + t)).
Contrary, consider an optimal solution to Si(p, q, r, t). Suppose that r0 copies of xi−1 are
in the first set of the solution and t0 copies of xi−1 are in the second set of the solution.
Note that r0 + t0 = Pi−1 − Pi−2 holds. Let us move r copies of xi in the first set of the
solution to the real xi−1 and t copies of xi in the second set of the solution to the same real
xi−1. We obtain a feasible solution to Si−1(p − r0, q − t0, r0 + r, t0 + t) whose cut value is
M AXCU T (Si(p, q, r, t)) − (xi − xi−1)(pt + qr). It follows that M AXCU T (Si(p, q, r, t)) ≤
(xi − xi−1)(pt + qr)+
+ maxr0≤p∧t0≤q∧r0+t0=Pi−1−Pi−2 M AXCU T (Si−1(p − r0, q − t0, r0 + r, t0 + t)).
Theorem 1. The geometric MAX-CUT problem on the real line as well as the geometric
MAX-BISECTION problem on the real line are solvable in O(n4) time.
Proof. First, we shall show that the subproblems Si(p, q, r, t), where i ∈ {1, ..., l}, p, q, r, t
are integers in {0, ..., n} such that p + q = Pi−1 and r + t = n−Pi−1 are solvable in O(n4)
time.
Since p + q = Pi−1, there are O(n) choices for the parameters p, q, and similarly since
r + t = n − Pi−1, there are O(n) choices for the parameters r, t. It follows that the total
number of considered subproblems is O(n3).
We can compute the values of M AXCU T (Si(p, q, r, t)) in bottom up fashion in increasing
i order. If i = 1, then p = 0 and q = 0 and M AXCU T (Si(p, q, r, t)) = 0 holds trivially. For
i ≥ 2, we apply Lemma 1 in order to compute the value of M AXCU T (Si(p, q, r, t)) in O(n)
3
x1...xi−2xi−1xiq−t0p−r0t0r0trq−t0p−r0t0r0trx1...xi−2xi−1xi time by r0 + t0 = Pi−1 − Pi−2. The corresponding optimal solutions can be obtained by
backtracking. The upper bound O(n4) follows.
Let xl be the largest real whose copy is in P, and let m be the number of copies of xl in
P. The optimal solution to the geometric MAX-CUT problem for P can be found among the
optimal solutions to the O(n2) subproblems Sk(p, q, r, t), where p + q = Pk−1 and r + t = m.
Furthermore, the optimal solution to the geometric (k, n− k) MIN-PARTITION problem for
P can be found among the optimal solutions to the O(n) subproblems Sk(p, q, r, t), where
p + q = Pk−1, r + t = m, p + r = k and q + t = n − k.
To solve the (k, n− k) MIN-PARTITION problem on the real line, we consider an analo-
gous family of generalized subproblems Ui(p, q, r, t), where i ∈ {1, ..., l}, p, q, r, t are integers
in {0, ..., n} such that p + q = Pi−1 and r + t = n − Pi−1. The subproblem Ui(p, q, r, t) is
to find for a multiset that is the union of Pi−1 with r + t copies of xi a partition into two
multisets such that p elements of Pi and r copies of xi form the first set and the value of the
cut is minimized. The value of such a minimum cut is denoted by M IN CU T (Si(p, q, r, t)).
Analogously, we obtain the following counterparts of Lemma 1 and Theorem 1 for (k, n−
k) MIN-PARTITION and MIN-BISECTION.
Lemma 2. For i ≥ 2, M IN CU T (Ui(p, q, r, t)) =
(xi−xi−1)(pt+qr)+minr0≤p∧t0≤q∧r0+t0=Pi−1−Pi−2 M IN CU T (Ui−1(p−r0, q−t0, r0+r, t0+t)).
Theorem 2. The geometric (k, n − k) MIN-PARTITION problem on the real line, in par-
ticular the geometric MIN-BISECTION problem on the real line, are analogously solvable by
dynamic programming in O(n4) time.
4 Final Remarks
It remains an open problem whether our method can be generalized to higher dimensions or
those problems turn out to be inherently hard. At stake is the exact computational status
of other geometric problems for which our knowledge is very limited at the moment.
5 Acknowledgments
We thank Uri Feige, Ravi Kannan and Christos Levcopoulos for a number of interesting
discussions on the subject of this paper.
References
[1] N. Alon, W.F. de la Vega, R. Kannan and M. Karpinski. Random Sampling and Approximation
of MAX-CSPs. J. Computer and System Sciences 67 (2003), pp. 212 -- 243.
[2] W.F. de la Vega, R. Kannan, M. Karpinski and S. Vempala. Tensor Decomposition and
Approximation Schemes for Constraint Satisfaction Problems. Proc. 37th ACM STOC (2005),
pp. 747 -- 754.
4
[3] W.F. de la Vega and M. Karpinski. Polynomial Time Approximation of Dense Weighted
Instances of MAX-CUT. Random Structures and Algorithms 16 (2000), pp. 314 -- 332.
[4] W.F. de la Vega, M. Karpinski and C. Kenyon. Approximation Schemes for Metric Bisection
and Partitioning. Proc. 15th ACM-SIAM SODA (2004), pp. 506 -- 515.
[5] W.F. de la Vega and C. Kenyon. A Randomized Approximation Scheme for Metric MAX-
CUT. Proc. Proc. 39th IEEE FOCS (1998), pp. 468 -- 471; journal version in J. Computer and
System Sciences 63 (2001), pp. 531 -- 541.
[6] R. Kannan. Spectral Methods for Matrices and Tensors. Proc. 42nd ACM STOC (2010),
pp. 1 -- 12.
5
|
1112.3506 | 3 | 1112 | 2013-11-06T16:02:31 | Max-Cut Parameterized Above the Edwards-Erd\H{o}s Bound | [
"cs.DS",
"cs.CC",
"cs.DM"
] | We study the boundary of tractability for the Max-Cut problem in graphs. Our main result shows that Max-Cut above the Edwards-Erd\H{o}s bound is fixed-parameter tractable: we give an algorithm that for any connected graph with n vertices and m edges finds a cut of size m/2 + (n-1)/4 + k in time 2^O(k)n^4, or decides that no such cut exists. This answers a long-standing open question from parameterized complexity that has been posed several times over the past 15 years. Our algorithm is asymptotically optimal, under the Exponential Time Hypothesis, and is strengthened by a polynomial-time computable kernel of polynomial size. | cs.DS | cs |
Max-Cut Parameterized Above the Edwards-Erdos Bound
Robert Crowston∗
Mark Jones†
Matthias Mnich‡
We study the boundary of tractability for the Max-Cut problem in graphs. Our main result shows
that Max-Cut parameterized above the Edwards-Erdos bound is fixed-parameter tractable: we give
an algorithm that for any connected graph with n vertices and m edges finds a cut of size
Abstract
4
in time 2O(k) · n4, or decides that no such cut exists.
number of times over the past 15 years.
n − 1
+
m
2
+ k
This answers a long-standing open question from parameterized complexity that has been posed a
Our algorithm is asymptotically optimal, under the Exponential Time Hypothesis, and is strength-
ened by a polynomial-time computable kernel of polynomial size.
1
Introduction
The study of cuts in graphs is a fundamental area in theoretical computer science, graph theory, and
polyhedral combinatorics, dating back to the 1960s. A cut of a graph is an edge-induced bipartite
subgraph, and its size is the number of edges it contains. Finding cuts of maximum size in a given graph
was one of Karp's famous 21 NP-complete problems [25]. Since then, the Max-Cut problem has received
considerable attention in the areas of approximation algorithms, random graph theory, combinatorics,
parameterized complexity, and others; see the survey [36].
As a fundamental NP-complete problem, the computational complexity of Max-Cut has been inten-
sively scrutinized. We continue this line of research and explore the boundary between tractability and
hardness, guided by the question: Is there a dichotomy of computational complexity of Max-Cut with
respect to the size of the maximum cut?
This question was already studied by Erdos [17] in the 1960s, who gave a randomized polynomial-time
algorithm that in any n-vertex graph with m edges finds a cut of size at least m/2. Erdos [17, 18] also
(erroneously) conjectured that the value m/2 can be raised to m/2 + εm for some ε > 0; only much later
it was shown [22, 34] that finding cuts of size m/2 + εm is NP-hard for every ε > 0. Furthermore, the
∗Royal Holloway - University of London, United Kingdom. [email protected]
†Royal Holloway - University of London, United Kingdom. [email protected]
‡Cluster of Excellence, Saarbrucken, Germany. [email protected]
1
Max-Cut Gain problem -- maximize the gain compared to a random solution that cuts m/2 edges -- does
not allow constant approximation [26] under the Unique Games Conjecture, and the best one can hope
for is to cut a 1/2 + Ω(ε/ log(1/ε)) fraction of edges in graphs in which the optimum is 1/2 + ε [8].
However, the lower bound m/2 can be increased, by a sublinear function: Edwards [15, 16] in 1973
proved that in connected graphs a cut of size
m/2 + (n − 1)/4,
(1)
always exists. Thus, any graph with n vertices, m edges and t components has a cut of size at least
m/2 + (n − t)/4. The lower bound (1) is famously known as the Edwards-Erdos bound, and it is tight
for cliques of every odd order n. The bound has been proved several times [6, 9, 19, 34, 35], with some
proofs yielding polynomial-time algorithms to attain it.
As (1) is tight for infinitely many non-isomorphic graphs, and finding maximum cuts is NP-hard,
raising the lower bound (1) even further requires a new approach: a fixed-parameter algorithm, that
for any connected graph with n vertices and m edges, and integer k ∈ N, finds a cut of size at least
m/2 + (n − 1)/4 + k (if such exists) in time f (k) · nc, where f is an arbitrary function dependent only
on k and c is an absolute constant independent of k. The point here is to confine the combinatorial
explosion to the (small) parameter k. But at first sight, it seems not even clear how to find a cut of size
m/2 + (n − 1)/4 + k in time nf (k), for any function f . We refer to the books by Downey and Fellows [14]
and Flum and Grohe [20] for background on parameterized complexity.
In 1997, Mahajan and Raman [29] gave a fixed-parameter algorithm for the variant of this problem
with Erdos' lower bound m/2, and showed how to decide existence of a cut of size m/2 + k in time
2O(k) · nO(1). Their result was strengthened by Bollob´as and Scott [6] who replaced m/2 by the stronger
bound
(2)
m/2 +
1
8 (cid:0)√8m + 1 − 1(cid:1) .
For unweighted graphs, this bound is weaker than (1). It remained an open question ([9, 21, 29, 30, 38])
whether this result could be strengthened further by replacing (2) with the stronger bound (1).
1.1 Main Results
We settle the computational complexity of Max-Cut above the Edwards-Erdos bound (1).
Theorem 1. There is an algorithm that computes, for any connected graph G with n vertices and m
edges and any integer k ∈ N, in time 2O(k) · n4 a cut of G with size at least m/2 + (n − 1)/4 + k, or
decides that no such cut exists.
Theorem 1 answers a question posed several times over the past 15 years [9, 21, 29, 30, 38]. In particu-
lar, instances with k = O(log m) can be solved efficiently, which thus extends the classical polynomial-time
algorithms [6, 9, 19, 34] that compute a cut of size (1).
The running time of our algorithm is likely to be optimal, as the following theorem shows.
Theorem 2. No algorithm can find cuts of size m/2 + (n− 1)/4 + k in time 2o(k)· nO(1) given a connected
graph with n vertices and m edges, and integer k ∈ N, unless the Exponential Time Hypothesis fails.
2
The Exponential Time Hypothesis was introduced by Impagliazzo and Paturi [24], and states that
n-variable 3-CNF-SAT formulas cannot be solved in time subexponential in n.
Fixed-parameter tractability of Max-Cut above Edwards-Erdos bound m/2 + (n − 1)/4 implies the
existence of a so-called kernelization, which is an algorithm that efficiently compresses any instance
(G, k) into an equivalent instance (G′, k′), the kernel, whose size g(k) = G′ + k′ itself depends on k
only. Alas, the size g(k) of the kernel for many fixed-parameter tractable problems is enormous, and in
particular many fixed-parameter tractable problems do not admit kernels of size polynomial in k unless
coNP ⊆ NP/poly [5]. We prove the following.
Theorem 3. There is a polynomial-time algorithm that, for any integer k ∈ N, compresses any connected
graph G = (V, E) with integer k ∈ N to a connected graph G′ = (V ′, E′) of order O(k5), and produces an
integer k′ ≤ k, such that G has a cut of size E/2 + (V − 1)/4 + k if and only if G′ has a cut of size
E′/2 + (V ′ − 1)/4 + k′, for some k′ ≤ k.
Note that Theorems 1 and 3 only hold for unweighted graphs; determining the parameterized com-
plexity of the weighted versions remains open.
1.2 Our Techniques and Related Work
A number of standard approaches that have been developed for "above-guarantee" parameterizations of
other problems are unavailable for this problem; hence, our algorithm differs significantly from others
in the area. The most common approach is to use probabilistic analysis of a random variable whose
expected value corresponds to a solution matching the guarantee. However, there is no simple randomized
procedure known giving a cut of size m/2 + (n− 1)/4. Another approach is to make use of approximation
algorithms that give a factor-c approximation, when the problem is parameterized above the bound c· n;
here, n is the maximum value of the objective function; yet there is no such approximation algorithm for
the Max-Cut problem.
Instead, we make use of "one-way" reduction rules. Unlike standard reduction rules, our rules do not
produce an equivalent instance, but merely preserve "no"-instances. If they produce a "yes"-instance we
know the original instance is also a "yes"-instance, but if they do not, then this fact allows us to show
the original instance has a relatively simple structure, which we can then use to solve the problem. We
have not seen this approach used in parameterized algorithmics before, and it may prove useful in solving
other problems where useful two-way reduction rules are hard to find.
Our results are based on algorithmic as well as combinatorial arguments. To prove Theorem 1, we
use one-way reduction rules to reduce the problem to one on graphs which are close to being in a special
class of graphs we call "clique-forests", on which we show how to solve the problem efficiently. Theorem 2
follows from a straightforward parameterized reduction. Theorem 3 is proven by a careful analysis of
random cuts via the probabilistic method.
For variants of Max-Cut, the "boundary of tractability" above guaranteed values has also been
investigated in the setting of parameterized complexity. For instance, in Max-Bisection we seek a cut
such that the number of vertices in both sides of the bipartition is as equal as possible, the tight lower
3
bound on the bisection size is only m/2; fixed-parameter tractability of Max-Bisection above m/2 was
recently shown by Gutin and Yeo [21] and Mnich and Zenklusen [33].
2 Preliminaries
With the exception of the term "clique-forest" (see below), we use standard graph theory terminology
and notation. Given a graph G, let V (G) be the vertices of G and let E(G) be the edges of G. For
disjoint sets S, T ⊆ V (G), let E(S, T ) denote the set of edges in G with one vertex in S and one vertex
in T . For S ⊆ V (G), let G[S] denote the subgraph induced by the vertices of S, and let G − S denote
the graph G[V (G) \ S]. We say that G has a cut of size t if there exists a set S ⊆ V (G) such that
E(S, V (G) \ S) = t. The graph G is connected if any two of its vertices are connected by a path, and
it is 2-connected if G − {v} is connected for every v ∈ V (G). A vertex v is a cut-vertex for G if G is
connected and G − {v} is disconnected.
We study the following formulation of Max-Cut parameterized above Edwards-Erdos bound:
Max-Cut above Edwards-Erdos (Max-Cut-AEE)
Instance: A connected graph G with n = V (G) vertices and m = E(G) edges, and an integer
k ∈ N.
Parameter : k.
Question: Does G have a cut of size at least m
2 + n−1
4 + k
4 ?
We ask for a cut of size m
4 + k, so that we may treat k
as an integer at all times. Note that this does not affect the existence of a fixed-parameter algorithm or
polynomial-size kernel.
4 , rather than the more usual m
2 + n−1
4 + k
2 + n−1
An assignment or coloring on G is a function α : V (G) → {red, blue}, and an edge is cut or satisfied
by α if one of its vertices is colored red and the other vertex is colored blue. Note that a graph has a cut
of size t if and only if it has an assignment that satisfies at least t edges. A partial assignment on G is a
function α : X → {red, blue}, where X is a subset of V (G).
A block in G is a maximal 2-connected subgraph of G. It is well known that the blocks of a graph can
be found in linear time, and that for a connected graph, if two blocks intersect then their intersection
is a cut-vertex. It follows that any cycle is contained in a single block. We say a block is a leaf-block if
it has at most one vertex which is contained in another block. Note that every non-empty graph has a
leaf-block.
We define a class of graphs, clique-forests, as follows. A clique is a clique-forest, as is an empty graph.
The disjoint union of any clique-forests is a clique-forest. Finally, a graph formed from a clique forest by
identifying two vertices from separate components is also a clique-forest. Note that the clique-forests are
exactly the graphs for which every block is a clique. Such graphs are sometimes called block graphs, but
this term has contradictory definitions in the literature1 and we will not use it in this paper.
1e.g. the definition in by Bandelt and Mulder [4] agrees with our definition of clique-forest, but the definition in Chapter
3 of Diestel's book [13] has that every block graph is a tree.
4
To arrive at a realistic analysis of the required computational effort, throughout our model of com-
putation is the random access machine with the restriction that arithmetic operations are considered
unit-time only for constant-size integers; in this model, two b-bit integers can be added, subtracted, and
compared in O(b) time.
3 Fixed-Parameter Algorithm for Max-Cut above the Edwards-Erdos
Bound
In this section, we prove Theorem 1. Central to this proof is the following lemma, which also forms the
basis of our kernel in Theorem 3.
Lemma 4. Given a connected graph G with n vertices and m edges and an integer k, we can in polynomial
time decide that either G has a cut of size at least m
4 , or find a set S of at most 3k vertices in
G such that G − S is a clique-forest.
2 + n−1
4 + k
Given a set S as described in Lemma 4, after guessing a coloring of S we reduce Max-Cut-AEE
to a related problem on clique-forests, which we show is polynomial time solvable using Lemma 8. As
there are at most 23k possible colorings of S, we get an algorithm for Max-Cut-AEE with the required
running time.
To prove Lemma 4, we use a set of "one-way" reduction rules. These rules produce an instance (G′, k′)
such that if (G′, k′) is a "yes"-instance then (G, k) is also a "yes"-instance; this is shown in Lemma 5.
The converse does not necessarily hold; if (G′, k′) is a "no"-instance then (G, k) may be a "yes"- or
"no"-instance. We show in Lemma 6 that (G′, k′) will be an instance that is trivially easy to solve (one
with no edges), so if (G′, k′) is a "yes"-instance then we are done. Otherwise, the reduction rules mark a
set of vertices in G, and we will show using Lemma 7 that if S is the set of marked vertices, then S ≤ 3k
and G − S is a clique-forest, as required.
Rule 1:
Apply to a connected graph G with v ∈ V (G), X ⊆ V (G) such that X is a connected
component of G − {v} and X ∪ {v} is a clique.
All vertices in X.
Nothing.
Remove:
Mark:
Parameter: Reduce k by 1 if X is odd, otherwise leave k the same.
Rule 2:
Apply to a connected graph G reduced by Rule 1 with v ∈ V (G), X ⊆ V (G) such that
X is a connected component of G − {v} and X is a clique.
All vertices in X.
v.
Remove:
Mark:
Parameter: Reduce k by 2.
5
Apply to a connected graph G with a, b, c ∈ V (G) such that {a, b},{b, c} ∈
E(G),{a, c} /∈ E(G), and G − {a, b, c} is connected.
The vertices a, b, c.
a, b, c.
Rule 3:
Rule 4:
Remove:
Mark:
Parameter: Reduce k by 1.
Apply to a connected graph G with x, y ∈ V (G) such that {x, y} /∈ E(G), G − {x, y}
has two connected components, X and Y , and X ∪ {x} and X ∪ {y} are cliques.
Vertices {x, y} ∪ X.
x, y.
Remove:
Mark:
Parameter: Reduce k by 1.
These rules can be applied exhaustively in polynomial time, as each rule reduces the number of vertices
in G, and for each rule we can check for any applications of that rule by trying every set of at most three
vertices in V (G) and examining the components of the graph when those vertices are removed.
Lemma 5. Let (G, k) and (G′, k′) be instances of Max-Cut-AEE such that (G′, k′) is reduced from
(G, k) by an application of Rules 1, 2, 3 and 4. Then G′ is connected, and if (G′, k′) is a "yes"-instance
of Max-Cut-AEE then so is (G, k).
Proof. First, we show that G′ is connected. For Rules 1 and 2, observe that for s, t ∈ V (G) \ X, no path
between s and t passes through X, so G − X is connected. For Rule 3, the conditions explicitly state
that we only apply the rule if the resulting graph is connected. For Rule 4, observe that we remove all
vertices except those in Y , and Y forms a connected component.
Second, we prove separately for each rule the following claim, in which n′ denotes the number of
vertices and m′ the number of edges removed by the rule.
Any assignment to the vertices of G′ can be extended to an
assignment on G that cuts an additional m′
2 + n′
4 + k−k′
4
edges.
(⋆)
Rule 1: Since v is the only vertex connecting X to the rest of the graph, any assignment to G′ can be
extended to one which is optimal on X ∪{v}. (Let α be an optimal coloring of G[X ∪ {v}], and let α′ be
the α with all colors reversed. Both α and α′ are optimal colorings of G[X ∪ {v}], and one of these will
agree with the coloring we are given on G′ since the only overlap is v.) Therefore it suffices to show that
X + {v} has a cut of size m′
, since the edges we
remove form a clique including v, and all vertices in the clique except v are removed.
4 . Observe that n′ = X and m′ = X(X+1)
4 + k−k′
2 + n′
2
If X is even then the maximum cut of the clique X ∪ {v} has size
+ X
4
+ 1(cid:19) = X(X + 2)
= X(X + 1)
2 (cid:18)X
X
2
4
4
=
m′
2
+
n′
4
,
which is what we require as k is unchanged in this case.
6
If X is odd then the maximum cut of the clique X ∪ {v} has size
m′
2
= X(X + 2)
(X + 1)
(X + 1)
1
4
+
=
2
2
4
+
n′
4
+
1
4
,
which is what we require as we reduce k by 1 in this case.
Rule 2: Order the vertices of X as x1, x2, . . . , xn′ such that there exists an r so that xj is adjacent
to v for all j ≤ r, and xj is not adjacent to v for all j > r. Since G is connected and reduced by Rule
1, r ≥ 2 (if there is only one vertex x in X adjacent to v then Rule 1 applies), and since X is a clique
but X ∪ {v} is not (otherwise Rule 1 applies), there exists x ∈ X not adjacent to v, and so r ≤ n′ − 1.
2 m, and otherwise xj is the same
Color the vertices of X such that xj is the opposite color to v if j ≤ l n′
2 mo.
color as v. Observe that the total number of satisfied edges incident with X is l n′
Since m′ = n′(n′−1)
+ r, this means that the total number of satisfied edges incident with X is
2 k + minnr,l n′
2 mj n′
2
m′
2
+
n′2
4
n′
4 −
+(cid:24) n′
2 (cid:25)(cid:22) n′
,(cid:24) n′
2 (cid:23) + min(cid:26) r
2
2 when X is even, and at least m′
2 + n′
4 + 1
2 .
2 + n′
r
2 (cid:25) −
4 + 3
2(cid:27) ,
4 when X is odd. Hence, the
4 + 1
2 + n′
which is at least m′
number of satisfied edges incident with X is at least m′
Rule 3: Observe that n′ = 3 and m′ = 2 + E(G′,{a, b, c}). Consider two colorings α, α′ of {a, b, c}:
α(a) = α(c) = red, α(b) = blue, and α′(a) = α′(c) = blue, α′(b) = red. Both these colorings satisfy edges
{a, b} and {b, c}, and at least one of them will satisfy at least half the edges between {a, b, c} and G′.
Therefore, the number of satisfied edges incident with {a, b, c} is at least 2 + E(G′,{a,b,c})
4 + 1
4 .
Rule 4: Given an assignment to G′, color x, y either both red or both blue, such that at least half of
the edges between G′ and {x, y} are satisfied. Assume, without loss of generality, that x and y are both
colored red. Recall that X = G − (G′ ∪ {x, y}). Let ¯n = X and let ¯m be the number of edges incident
with vertices in X. Observe that ¯m = ¯n(¯n−1)
2 vertices with red. Then the total number
4 + ¯n + 1 = ¯m
4 + 3
4 .
2 − 1 vertices with red. Then the total
2 + 1) = ¯n2
4 + 1.
Observe that m′ = E(G′,{x, y}) + ¯m and n′ = ¯n + 2, and hence the total number of removed edges
2 + 1 of the vertices in X with blue and ¯n
2 − 1) + 2( ¯n
4 = m′
4 + 3
2 of the vertices in X with blue and ¯n−1
2 = ¯n(¯n−1)
4 − 1 + ¯n + 2 = ¯m
4 + 1
4 .
of satisfied edges incident with X is ¯n+1
2
number of satisfied edges incident with X is ( ¯n
satisfied is at least E(G′,{x,y})
If ¯n is odd, color ¯n+1
If ¯n is even, color ¯n
4 + 1 = ¯m
4 + ¯n−1
2 + 2 ¯n+1
2 + n′−2
2 + 2¯n.
2 + 1)( ¯n
2 + ¯n−1
4 = m′
2 + n′
2 + n′
· ¯n−1
2 + ¯n
2 + ¯n
2 + ¯n
4 + 3
= m′
+ ¯m
2
2
This concludes the proof of the claim (⋆).
We now know that any assignment on G′ can be extended to an assignment on G that cuts an
4 , then G has a cut of
4 . Therefore, if (G′, k′) is a "yes"-instance then
edges. Hence, if G′ has a cut of size E(G′)
2 + n′
2 + n−n′−1
4 + k−k′
4 + k′
2 + V (G′)−1
4 + k−k′
2 + n−1
4 = m
4 + m′
2 + n′
4 + k
+ k′
4
4
additional m′
size m−m′
so is (G, k).
Lemma 6. To any connected graph G with at least one edge, at least one of Rules 1 -- 4 applies.
7
Proof. If G is not 2-connected, then let X be a leaf-block in G with cut-vertex r. Otherwise, let X = G
and let r be any vertex in G. We say that a set of vertices {a, b, c} in G is an induced P3 if G[{a, b, c}]
is a path with 3 vertices (i.e. ab, bc ∈ E(G), ac /∈ E(G)). For any set S of vertices in X, let R(r, S)
denote the set of vertices reachable from r in G − S (if r ∈ S then R(r, S) = ∅) . For a, b, c ∈ X, we
say {a, b, c} is an important P3 if {a, b, c} is an induced P3 and there is no induced P3 {a′, b′, c′} ⊆ X
such that R(r,{a, b, c}) ⊂ R(r,{a′, b′, c′}). (The concept of an important P3 is similar to the concept of
"important separators" introduced by Marx [31].)
Observe that if there is no induced P3 in X then X is a clique, and Rule 1 applies. Therefore we may
assume there is an important P3 in X. Let {a, b, c} be an important P3 with ab, bc ∈ E(G), ac /∈ E(G). If
r ∈ {a, b, c} then X − r must have no induced P3 and so X − r is a clique, and Rule 2 applies. Therefore
we may assume r /∈ {a, b, c}.
If Y = ∅ then G − {a, b, c} is connected and Rule 3 ap-
plies. Therefore we may assume Y 6= ∅. Since X is 2-connected, at least two of a, b, c are adjacent
to R(r,{a, b, c}). In particular, one of a and c is adjacent to R(r,{a, b, c}). Without loss of generality,
assume that c is adjacent to R(r,{a, b, c}).
Let Y = G − (R(r,{a, b, c}) ∪ {a, b, c}).
We now prove three properties of Y , which will be used to show that Rule 4 applies:
(P1) For any x ∈ Y , ax ∈ E(G) if and only if bx ∈ E(G).
For suppose there exists x ∈ Y which is adjacent to a but not b. Then {x, a, b} is an induced P3,
and R(r,{x, a, b}) ⊇ R(r,{a, b, c})∪{c}, a contradiction as {a, b, c} is an important P3. Similarly, if
x is adjacent to b but not a, then {a, b, x} is an induced P3 and R(r,{a, b, x}) ⊇ R(r,{a, b, c})∪{c}.
(P2) For each s ∈ {a, b, c} and any x, y ∈ Y , if sx ∈ E(G) and xy ∈ E(G), then sy ∈ E(G).
For suppose not; then {s, x, y} is an induced P3, and since at least one vertex t ∈ {a, b, c}\s is
adjacent to R(r,{a, b, c}), we have that R(r,{s, x, y}) ⊆ R(r,{a, b, c}) ∪ {t}, a contradiction as
{a, b, c} is an important P3.
(P3) For each s ∈ {a, b, c} and any x, y ∈ Y , if sx ∈ E(G) and sy ∈ E(G), then xy ∈ E(G).
For suppose not. Then {x, s, y} is an induced P3, and as with Property (P2) we have a contradiction.
There are now two cases to consider. First, consider the case when a is adjacent to R(r,{a, b, c}).
Then, by an argument similar to that used for proving Property (P1), we may show that for any x ∈ Y ,
cx ∈ E(G) if and only if bx ∈ E(G). This together with Property (P1) implies that a, b, c have exactly
the same neighbors in Y . Then by Properties (P2) and (P3) we have that Y is a clique and every vertex
in Y is adjacent to each of a, b, c. If b is adjacent to R(r,{a, b, c}) then for any x ∈ Y , {a, x, c} is an
induced P3 and R(r,{a, x, c}) ⊇ R(r,{a, b, c}) ∪ {b}, a contradiction as {a, b, c} is an important P3. So b
is not adjacent to R(r,{a, b, c}). Then Y ∪ {b} and R(r,{a, b, c}) are the two components of G − {a, c},
and (Y ∪ {b}) ∪ {a} and (Y ∪ {b}) ∪ {c} are cliques. Therefore, Rule 4 applies.
Second, consider the case when a is not adjacent to R(r,{a, b, c}). Then b must be adjacent to
R(r,{a, b, c}). Furthermore, as X is 2-connected, there must be a path from a to c in G − b, and the
intermediate vertices in this path must be in Y . By Property (P2), there exists x ∈ Y adjacent to a and
c. Then {a, x, c} is an induced P3 and R(r,{a, x, c}) ⊇ R(r,{a, b, c}) ∪ {b}, a contradiction as {a, b, c} is
an important P3.
8
Lemma 7. Let (G′, k′) be an instance derived from (G, k) by exhaustively applying Rules 1 -- 4, and let S
be the set of vertices marked during the construction of (G′, k′). Then G − S is a clique-forest.
Proof. By Lemma 6, G′ contains no edges, and is therefore a clique-forest. (In fact by Lemma 5, G′ is
also connected, and therefore consists of a single vertex.) Let G′′ be a graph derived from G by a single
application of Rule 1, 2, 3 or 4. By induction on the length of the reduction from (G, k) to (G′, k′), it is
enough to show that if G′′ − S is a clique-forest, then G − S is a clique-forest.
If G′′ is derived from G by an application of Rule 1, observe that G − S can be formed from G′′ − S
by adding a disjoint clique and identifying one of its vertices with the vertex v in G′′ (unless v is in S, in
which case G − S is just formed by adding a disjoint clique), where v is the vertex referred to in Rule 1.
Therefore, G − S is a clique-forest. For Rule 3, observe that G − S = G′′ − S.
For Rules 2 and 4, observe that G− (G′ ∪ S) is a clique, and that S disconnects G− (G′ ∪ S) from G′.
Therefore, G−S can be formed from G′−S by adding a disjoint clique, and so G−S is a clique-forest.
Putting Lemmas 5 and 7 together, we can now prove Lemma 4.
of Lemma 4. Let (G′, k′) be an instance derived from (G, k) by exhaustively applying Rules 1 -- 4, and let
S be the set of vertices marked during the construction of (G′, k′). By Lemma 7, G− S is a clique-forest.
Therefore, if S < 3k we are done. It remains to show that if S ≥ 3k then G has an assignment that
satisfies at least m
So suppose that S ≥ 3k. Observe that every time k is reduced, at most three vertices are marked.
Therefore since at least 3k vertices are marked, we have k′ ≤ 0. But since the Edwards-Erdos bound
holds for all connected graphs, G′ is a "yes"-instance. Therefore, by Lemma 5, G is a "yes"-instance with
parameter k, as required.
2 + n−1
4 + k
4 edges.
In what follows, we assume we are given a set S such that S ≤ 3k and G − S is a clique-forest, as
given by Lemma 4. Note that we should not assume that (G, k) is reduced by any of Rules 1 -- 4. These
rules were only used to produce the separate instance (G′, k′) and S. We may now forget about (G′, k′)
and Rules 1 -- 4.
We now show that, for a given assignment to S, we can efficiently find an optimal extension to G− S.
For this, we consider the following generalisation of Max-Cut where each vertex has an associated weight
for each part of the partition. These weights may be taken as an indication of how much we would like
the vertex to appear in each part. For convenience, here we will think of an assignment as a function
from V to {0, 1} rather than {red, blue}.
Max-Cut-with-Weighted-Vertices
Instance: A graph G with weight functions w0 : V (G) → N0 and w1 : V (G) → N0,
and an integer t ∈ N.
Question: Does there exist an assignment f : V → {0, 1} such that
Pxy∈E f (x) − f (y) +Pf (x)=0 w0(x) +Pf (x)=1 w1(x) ≥ t?
Now Max-Cut is the special case of Max-Cut-with-Weighted-Vertices in which w0(x) =
w1(x) = 0 for all x ∈ V (G).
9
Lemma 8. Max-Cut-with-Weighted-Vertices is solvable in polynomial time when G is a clique-
forest.
0, w′
Proof. We provide a polynomial-time transformation that replaces an instance (G, w0, w1, t) with an
1, t′) such that G′ has fewer vertices than G. By applying the transformation
equivalent instance (G′, w′
at most V (G) times to get a trivial instance, we have a polynomial-time algorithm to decide (G, w0, w1, t).
We may assume that G is connected, as otherwise we can handle each component of G separately.
Let X ∪ {r} be the vertices of a leaf-block in G, with r a cut-vertex of G (unless G consists of a single
block, in which case let r be an arbitrary vertex and X = V (G) − {r}). Recall that by definition of a
clique-forest, X ∪ {r} is a clique. For each possible assignment to r, we will in polynomial time calculate
the optimal extension to the vertices in X. (This optimal extension depends only on the assignment to r,
since no other vertices are adjacent to vertices in X.) We can then remove all the vertices in X, and
change the values of w0(r) and w1(r) to reflect the optimal extension for each assignment.
Suppose we assign r the value 1. Let ε(x) = w1(x) − w0(x) for each x ∈ X. Now arrange the vertices
of X in order x1, x2, . . . xn′ (where n′ = X), such that if i < j then ε(xi) ≥ ε(xj). Observe that there is
an optimal assignment for which xi is assigned 1 for every i ≤ t, and xi is assigned 0 for every i > t, for
some 0 ≤ t ≤ n′. (Consider an assignment for which f (xi) = 0 and f (xj) = 1, for i < j, and observe that
switching the assignments of xi and xj will increase Pf (x)=0 w0(x) + Pf (x)=1 w1(x) by ε(xi) − ε(xj).)
Therefore we only need to try n′ + 1 different assignments to the vertices in X in order to find the optimal
coloring when f (r) = 1. Let A be the value of this optimal assignment (over X ∪ {r}).
By a similar method we can find the optimal assignment when r is assigned 0. Let the number of
satisfied edges in this coloring be B. Now remove the vertices in X and incident edges, and let w1(r) = A,
and let w0(r) = B.
We are ready to prove Theorem 1, and show that Max-Cut-AEE is fixed-parameter tractable.
2 + n−1
4 + k
of Theorem 1. By Lemma 4, we can in polynomial time decide that either G has an assignment that
satisfies at least m
4 edges, or find a set S of at most 3k vertices in G such that -G− S is a clique-
forest. So assume we have found such an S. Then we transform our instance into at most 23k instances
of Max-Cut-with-Weighted-Vertices, such that the answer to our original instance is "yes" if and
only if the answer to at least one of the instances of Max-Cut-with-Weighted-Vertices is "yes", and
in each Max-Cut-with-Weighted-Vertices instance the graph is a clique-forest. As each of these
instances can be solved in polynomial time by Lemma 8, we have a fixed-parameter tractable algorithm.
For every possible assignment to the vertices in S, we construct an instance of Max-Cut-with-
Weighted-Vertices as follows. For every vertex x ∈ G − S, let w0(x) equal the number of vertices
in S adjacent to x which are colored blue, and let w1(x) equal the number of vertices in S adjacent to x
which are colored red. Then remove the vertices of S from G. By Lemma 4, the resulting graph G′ is a
clique-forest. Let m′ be the number of edges in G − S, let n′ be the number of vertices in G − S, and
let p be the number of edges within S satisfied by the assignment to S. Then for an assignment to the
vertices of G − S, the total number of satisfied edges in G would be exactly Pxy∈E(G−S) f (x) − f (y) +
p + Pf (x)=0 w0(x) + Pf (x)=1 w1(x), where f : V (G) \ S → {0, 1} is such that f (x) = 0 if x is colored
red, and f (x) = 1 if x is colored blue. Thus, the assignment to S can be extended to one that cuts at
10
2 + n−1
least m
"yes"-instance with t = m
4 + k
4 edges in G if and only if the instance of Max-Cut-with-Weighted-Vertices is a
2 + n−1
4 + k
4 − p.
4 Algorithmic Lower Bounds
We now prove Theorem 2, by a reduction from the Max-Cut problem with parameter the size k of the
cut. By a result of Cai and Juedes [7], the Max-Cut problem with parameter the size k of the cut cannot
be solved in time 2o(k)·nO(1) unless the Exponential Time Hypothesis fails. Note that by deleting isolated
vertices and identifying together one vertex from each component, we may assume the input graph of
an instance of Max Cut is connected. Given the graph G with m edges and n vertices and the integer
k ∈ N, let k′ = k − ⌈m/2 + (n − 1)/4⌉. That is, k′ is the integer such that G has a cut of size at least
k if and only if G has a cut of size at least m/2 + (n − 1)/4 + k′. Then use a hypothetical algorithm to
solve Max-Cut AEE in time 2o(k′) · nO(1) on instance (G, k′), return the answer of the algorithm for the
pair (G, k) and the Max-Cut problem. Thus, since k′ ≤ k, an algorithm of time complexity 2o(k) · nO(1)
for Max-Cut AEE forces the Exponential Time Hypothesis to fail [14]. This completes the proof of
Theorem 2.
5 Polynomial Kernel for Max-Cut Above Edwards-Erdos
In this section, we prove Theorem 3. By Lemma 4, in polynomial time we can either decide that (G, k)
is a "yes"-instance, or find a set S of vertices in G such that S < 3k and G − S is a clique-forest. In
what follows we assume that we have found such a set S.
Let n∗ be the number of blocks of G − S. Let C1, . . . , Cn∗ be the blocks of G − S, and recall that by
definition each block is a clique. Let J be the set of vertices in G − S that appear in two or more block.
For each block Ci, let Ai = Ci \ J. Note that {A1, . . . , An∗ , J} is a partition of the vertices of G − S.
Let L ⊆ {1, . . . , n∗} be the set of all i for which Ci is a leaf-block. Let I ⊆ [n∗] be the set of all i for
which Ci ∩ J ≥ 3. It can be seen that L ≥ I + 2. (Indeed, consider the bipartite graph with vertex
sets J and [n∗], with an edge between v ∈ J and i ∈ [n∗] if v ∈ Ci. Then this is a forest, with all leaves
being leaf-blocks in G − S, and for each i ∈ [n∗] the degree of i is Ci ∩ J. It is well known that in a tree
the number of vertices of degree at least three is at most the number of leaves −2, from which the claim
follows.)
We first apply Reduction Rules 5 -- 8. Note that unlike Rules 1 -- 4, these are traditional "two-way"
reduction rules, and once we have shown they are valid, we may assume the instance is reduced by these
rules. The correctness of these rules is given by Lemma 9.
After applying these reduction rules, we will show that either we have a Yes-instance, or the number
of vertices in G − S is bounded. Observe that to bound the number of vertices in G − S, it is enough
to bound n∗ and Ai for each i. In Lemma 10, we bound L. As L ≥ I + 2, in order to bound n∗ it
remains to bound the number of i for which Ci ∩ J ≥ 3. This is done in Lemma 11. Finally in Lemma
12 we bound Ai for each i. These results are combined in Lemma 13.
11
Rule 5:
Apply if there exists a vertex x ∈ G − S and a set of vertices X ⊆ V (G) \ S such that
X > 1, X ∪ {x} is a clique and X is a connected component of G − (S ∪ {x}), and no
vertex in X is adjacent to any vertex in S.
All vertices in X.
Remove:
Parameter: Reduce k by 1 if X is odd, otherwise leave k the same.
Rule 6:
Apply if there exists vertices s ∈ S, x ∈ G − S and a set of vertices X ⊆ V (G) \ S
such that X ∪ {x} is a clique, X ∪ {s} is a clique, X is a connected component of
G − (S ∪ {x}), and s is the only vertex in S adjacent to X.
All but one vertex of X.
Remove:
Parameter: Reduce k by 1 if X is even, otherwise leave k the same.
Rule 7:
Apply if there exist blocks X, Y of G − S such that X and Y are odd, with vertices
x ∈ X, y ∈ Y,{z} = X ∩ Y , such that x, z are the only vertices in X adjacent to a
vertex in G − X, and y, z are the only vertices in Y adjacent to a vertex in G − Y .
All vertices in (X ∪ Y ) − {x, y, z}.
New vertices u, v and edges such that {x, y, z, u, v} is a clique
Remove:
Add:
Parameter: No change.
Rule 8:
Apply if for any block Ci in G − S, there exists X ⊆ Ai such that X > Ai+J+S
and for all x, y ∈ X, x and y have exactly the same neighbors in S.
Two arbitrary vertices from X.
2
Remove:
Parameter: No change.
Lemma 9. Let (G, k) and (G′, k′) be instances of Max-Cut-AEE, and S a set of vertices, such that
G − S is a clique-forest, and (G′, k′) is reduced from (G, k) by an application of Rule 5, 6, 7 or 8. Then
G′ is connected, G′ − S is a clique-forest, and (G′, k′) is a "yes"-instance if and only if (G, k) is a
"yes"-instance.
of Lemma 9. It is easy to observe that each of the rules preserves connectedness, and that G′ − S is a
clique-forest.
We now show for each rule that (G′, k′) is a "yes"-instance if and only if (G, k) is a "yes"-instance.
Rule 5: Let n′ be the number of vertices and m′ the number of edges removed. Note that m′ = n′(n′+1)
.
Observe that whatever we assign to the rest of the graph, we can always find an assignment to X that
2 + n′+1
,
4 edges
edges in G′, and if X is even, we can satisfy
2 + n−n′−1
satisfies the largest possible number of edges within X ∪{x}. If X is odd this is (n′+1)(n′+1)
and if X is even this is n′(n′+2)
in G if and only if we can satisfy m−m′
2 + n−1
m
Rule 6: Let n′ be the number of vertices and m′ the number of edges removed. Note that n′ = X − 1
and m′ = n′(n′+1)
4 edges in G if and only if we can satisfy m−m′
+ 2n′ = n′(n′+5)
= m′
4 + k
2 + n−1
2 + n′
2 + n−n′−1
4 . Therefore if X is odd, we can satisfy m
4 edges in G′.
4 + k
+ k−1
4
4
4 + k
= m′
2
4
4
4
2
.
2
12
First consider the case when x and s are adjacent. Observe that whatever x and s are assigned,
it is possible to find an assignment to X that satisfies the maximum possible number of edges within
X ∪ {s, x}. This is (n′+2)(n′+4)
if X is even.
Furthermore note that in G′, whatever x and s are assigned, we will be able to satisfy 2 edges between x, s
and the reamining part of X. Therefore if X is odd, we can satisfy m
4 edges in G if and only if
we can satisfy m−m′
4 edges in G′, and if X is even, we can satisfy
2 + n−n′−1
2 + n−1
4 + k−1
4 + k
m
edges in G′.
4 + k
4 edges in G if and only if we can satisfy m−m′
if X is odd, and (n′+3)(n′+3)
2 + n−1
4 + 2 = m−m′
4 + 2 = m−m′
2 + n−n′−1
2 + n−n′−1
2 + n−n′−1
4 + k−9
4 + k−8
2 + n′+9
2 + n′+8
4 + k
= m′
= m′
4
4
4
4
4
4
4
.
(n′
4
4
4
4
4
4
4
(n′+2)(n′+4)
1 − 1)
+
4 + k
1 + 1)(n′
2 + 1)(n′
1 and m′
2 + n′+8
2 + n′+9
2 + n′+4
2 + n′+5
4 + k−1
− 1 = m′
− 1 = m′
− 1 = m′
G′ we will be able to satisfy 2 edges between x, s and the remaining part of X.
the maximum number of edges within X ∪ {x, s} we can satisfy is (n′+2)(n′+4)
odd, and (n′+3)(n′+3)
between x, s and the remaining part of X.
Second consider the case when x and s are not adjacent.Observe that if x and s are colored differently,
if X is
if X is even. Furthemore in G′ we will be able to satisfy 1 edge
If x and s are colored the same, the maximum number of edges within X ∪ {x, s} we can satisfy is
if X is even. Furthemore in
Observe that whether or not x and s are colored the same, in G we can satisfy t more edges between
if X
x, s and X than in G′, where t = m−m′
4 if X is odd, and t = m−m′
is even.
Rule 7: Let n′ = X ∪ Y and let m′ be the number of edges within X ∪ Y . Let n′
2 = Y ,
and let m′
2 be the number of edges within X and Y , respectively. Observe that whatever x, y
and z are assigned, we can always satisfy the maximal possible number of edges in within X ∪ Y , which
is
if X is odd, and (n′+3)(n′+3)
1 = X and n′
2 + n−n′−1
2 + n−n′−1
− 1 = m′
3 = 5 = {x, y, z, u, v} and let m′
. Thus, the amount we gain above the Edwards-Erdos bound remains the same.
2 − 1)
3 = 10 be the number of edges within {x, y, z, u, v}. Then in G′,
whatever x, y and z are assigned, the maximum number of edges within {x, y, z, u, v} we can satisfy is
6 = m′
Rule 8: Let n′ = Ai. For any assignment to the vertices in S ∪ J, and for each x ∈ Ai, let εR(x) be
the number of neighbors of x in S ∪ J which are assigned red, and let εB(x) be the number of neighbors
of x in S ∪ J which are assigned blue. Let ε(x) = εB(x) − εR(x). Let the vertices of Ai be numbered
x1, x2, . . . , xn′ such that ε(x1) ≥ ε(x2) ≥ . . . ≥ ε(xn′). Observe that the optimal assignment to Ai will
be one in which xj is assigned red for j ≤ n′+r
2 , and blue otherwise, for some integer r. Observe that the
optimal value of r will be between −(J + S) and (J + S). Indeed, if r > J + S, then switching
one of the vertices from red to blue will gain at least J + S edges within Ai, and lose at most J + S
edges between Ai and J ∪ S. (A similar argument holds when r < −J − S.)
Now let y, z be the two vertices in X removed by the rule. Since y, z have exactly the same neighbors
in S ∪ J, and since X > Ai+J+S
, we may assume that y = xj, z = xj′, for some j, j′ such that
j′ − j > J ∪ S ≥ r, and hence y is assigned red and z is assigned blue. Then note that if we remove y
and z, we lose m′′ = 2N (y) ∩ (S ∪ J) + 2(n′ − 2) + 1 edges and n′′ = 2 vertices. Of the edges removed,
exactly N (y) ∩ (S ∪ J) + (n′ − 2) + 1 = m′′
4 were satisfied. Thus, the amount we gain over the
n′
2 − 1
4
n′
1 − 1
4
3
2 + n′
3−1
4
n′ − 1
4
2 + n′′
=
+
+
=
+
m′
2
m′
1
2
m′
2
2
(n′
4
Let n′
2
+
4
13
Edwards-Erdos bound remains the same. Note that this happens whatever the assignment to S ∪ J, and
that we may assume without loss of generality that y and z are colored differently for any assignment
to S ∪ J.
We now assume that G is reduced by Rules 5, 6, 7 and 8.
Lemma 10. If L ≥ 4S2 + 2S + 2k − 2 then (G, k) is a "yes"-instance.
Proof. Suppose L ≥ 4S2 +2S+2k−2. For a leaf block Ci, let Ei be set of edges within G[Ai] together
with the set of edges between Ai and S, and let EL = Si∈L Ei. Let VL be the total set of vertices in
Si∈L Ai, and observe that VL = Pi∈L Ai. Let mi = Ei and ni = Ai.
assignment α on S such that χα(EL) ≥ EL
which each vertex is assigned red or blue with equal probability. For any vertex x ∈ G − S, let εα
the number of neighbors of x which are assigned red, let εα
are assigned blue, and let εα(x) = εα
x1, . . . , xni be an ordering of the vertices of Ai such that εα(x1) ≥ εα(x2) ≥ . . . ≥ εα(xni).
2 + ni−1
the total number of satisfied edges in Ei is equal to
For a partial assignment α on L and a set of edges E, let χα(E) be the maximum possible number of
edges we can satisfy in E, given α. We will show via the probabilistic method that there exists a partial
4 . Consider a random assignment α on S, in
R(x) be
B(x) be the number of neighbors of x which
R(x). Now consider χα(Ei) for some i where i ∈ L. Let
First, suppose that Ai is odd. Observe that if there exists x ∈ Ai with εα(x) ≥ r, then χα(Ei) ≥
and blue otherwise. Then
2 . Indeed, suppose εα(x) = r′ > r. Then color xj red if j ≤ ni+1
B(x) − εα
2 + VL
4 + L
4 + r
mi
2
ni + 1
2
ni − 1
2
·
+
ni +1
2
Xj=1
εα
B(xj) +
ni
Xj= ni+3
2
εα
R(xj)
=
ni(ni − 1)
4
+
1
2
ni +1
2
Xj=1
(εα
B(xj) + εα
R(xj)) +
1
2
ni
Xj= ni+3
2
(εα
B(xj) + εα
R(xj))
+
ni − 1
4
+
1
2
ni+1
2
Xj=1
(εα
B(xj) − εα
R(xj)) +
1
2
ni
Xj= ni+3
2
(εα
R(xj) − εα
B(xj))
=
mi
2
+
ni − 1
2
+
1
2
ni +1
2
Xj=1
εα(xj) −
ni
Xj= ni+3
2
εα(xj)
.
4 + 1
2 + ni−1
Observe that this is at least mi
2 . A similar argument applies when εα(xj) =
−r′ < −r. Therefore, if there exists x ∈ Ai with an odd number of neighbors in S, then εα(x) ≥ 1 and
4 + 1
so χα(Ei) ≥ mi
4 .
If there exists x ∈ Ai with a non-zero even number of neighbors in S, then observe that either εα(x) = 0
or εα(x) ≥ 2. Furthermore, the probability that εα(x) = 0 is at most 1
2 , since given an assignment to
4 . This is true whatever α assigns to S, and therefore Eα[χα(Ei)] ≥ mi
2 εα(x1) ≥ mi
2 + ni
2 + ni
2 + ni−1
4 + r
4 + 1
14
2 · 2
4 + 1
4 + 1
2 + ni
2 + ni
2 + ni
4 + 1
2 = mi
2 + ni−1
2 · 0 + 1
4 + 1
4 .
Second, suppose that Ai is even. Observe that if there exist x, y ∈ Ai with εα(x) > εα(y), then
εα(xj) ≥ 1. So color xj with red if j ≤ ni
all but one of the neighbors in S of x, at most one of the possible assignments to the remaining neighbor
will lead to εα(x) being 0. Therefore, Eα[χα(Ei)] ≥ mi
4 . We know
that one of the above two cases must hold, since otherwise no vertex in Ai has any neighbors in S, and
Rule 5 applies. Therefore, if Ai is odd, Eα[χα(Ei)] ≥ mi
j=1 εα(xj) −Pni
χα(Ei) ≥ mi
and blue otherwise. Then the total number of satisfied edges in Ei is equal to
Xj= ni
Also observe that if εα(x) ≥ 2 for all x ∈ Ai, then χα(Ei) ≥ mi
εα(xj)
≥
2 . Indeed, then P
Indeed, suppose
2 + 1 and blue otherwise. Then the total
εα(x) = r ≥ 2 for all x ∈ Ai. Then color xj with red if j ≤ ni
number of satisfied edges in Ei is equal to
εα(xj) −
Xj= ni
εα
R(xj) =
εα
B(xj) +
2 + ni
4 + 1.
ni
2 ·
Xj=1
Xj=1
mi
2
mi
2
ni
4
ni
4
j= ni
2 +1
+
.
+
+
ni
2
+
ni
2 +1
ni
2 +1
ni
2
1
2
+
1
2
ni
2
ni
2
2
(cid:16) ni
2
+ 1(cid:17)(cid:16) ni
2 − 1(cid:17) +
ni
2 +1
Xj=1
εα
B(xj) +
ni
Xj= ni
2 +2
εα
R(xj) =
=
≥
εα(xj) −
ni
Xj= ni
2 +2
εα(xj)
ni
2 +1
Xj=1
2r
1
2
1
2
mi
2
mi
2
mi
2
+
ni
4 − 1 +
+
+
ni
4 − 1 +
ni
+ 1 .
4
A similar argument applies when εα(x) = −r ≤ −2 for all x ∈ Ai. Finally observe that if εα(x) = 0 or 1
for all x ∈ Ai, then by coloring xj with red if j ≤ ni
4 . Therefore,
if there exist x, y ∈ Ai such that x has a neighbor in S which is not adjacent to y, then the probability
that εα(x) = εα(y) is at most 1
Otherwise, all vertices in Ai have the same neighbors in S. There must be at least two vertices in S
which are adjacent to the vertices in Ai, as otherwise Rule 5 or 6 would apply. Then the probability that
εα(x) = 0 for every x ∈ Ai is at most 1
4 . (Consider any
assignment to all but two of the neighbors of x in S, and observe that of the four possible assignments to
the remaining two vertices, at most two will lead to εα(x) = 0, and at least one will lead to εα(x) ≥ 2.)
Therefore, Eα[χα(Ei)] ≥ mi
2 and blue otherwise, χα(Ei) ≥ mi
2 + ni
2 , and the probability that εα(x) ≥ 2 is at least 1
2 and so Eα[χα(Ei)] ≥ mi
4 · 1 = mi
2 · 0 + 1
2 · 0 + 1
2 + ni
2 = mi
2 + ni
2 + ni
2 + ni
4 + 1
4 .
4 + 1
4 .
4 + 1
4 + 1
2 · 1
By linearity of expectation and the fact that χα(EL) = Pi∈L χα(Ei), it holds
+ VL
4
4(cid:19) = EL
(cid:18) mi
Eα[χα(EL)] ≥ Xi∈L
E[χα(Ei)] ≥ Xi∈L
ni
4
+
+
2
1
2
+ L
4
.
4 + L
Now let t′ be the number of components in G − (S ∪ VL), and observe that t′ ≤ L
Therefore, there exists a partial assignment α on S such that χα(EL) ≥ EL
4 , as required.
2 . Indeed, if C is
a component in G − S then C − VL is connected if not empty. Furthermore, every component in G − S
2 + VL
15
Let C1, . . . ,Ct′ be the components of G − (S ∪ VL). For each j ∈ {1, . . . , t′} let n′
contains at least one leaf-block, and if a component in G− S contains only one leaf-block, then the whole
component is a single leaf block, and so C − VL is empty. Therefore, every component in G − (S ∪ VL)
has at least two leaf-blocks.
j be the number of
vertices in Cj and m′
j the number of edges incident with vertices in Cj (including edges between Cj and
4 + L
S∪VL). Now let S∪VL be colored such that the number of satisfied edges in EL is at least EL
4 .
Note that this might mean that none of the edges in G[S] are satisfied, but that there are at most S2
j−1
of these. Now for each j ∈ {1, . . . , t′} let Cj be colored so that at least half the edges in G[Cj] plus
4
are satisfied, which can be done as this is the Edwards-Erdos bound. By reversing all the colors in Cj if
needed, we can ensure that at least half the edges between Cj and the rest of the graph are satisfied, and
therefore we can ensure at least
of the edges incident with Cj are satisfied.
We now have a complete assignment of colors to vertices, which satisfies at least
2 + VL
m′
j
2 +
j−1
4
n′
n′
t′
t′
t′
+
(cid:18) m′
j
2
n′
j − 1
4 (cid:19) = EL
EL
2
+ VL
4
+ L
4
+
Xj=1
Xj=1
+ L
8
edges. Since L ≥ 4S2 + 2S + 2k − 2, the right-hand side of (3) is is at least
n − 1
Xj=1
2
m − S2
4S2 + 2S + 2k − 2
(4S2 + 2S)
n − S
≥
+
+
+
=
+
+
+
m′
j
2
+ VL
4
2
m
2
+
n
4 −
8
8
4
n′
j
4
+ L − t′
4
(3)
(4)
k
4
.
4
m
2
Recall that n∗ is the number of blocks in G − S.
Lemma 11. If n∗ ≥ 4S2 + 2S + 4L + 2k − 7 then (G, k) is a "yes"-instance.
Proof. We will first produce a partial assignment on G − S which satisfies E(G−S)
edges,
where t is the number of connected components of G− S. Together with this we produce a set of vertices
R ⊆ V \ S such that we can change the color of any vertex in R without changing the number of satisfied
edges in G − S.
Observe that by choosing the right order on the blocks of G − S, it is possible to color all the blocks
in such a way that each time we come to color a block, it contains at most one vertex that has already
been colored. So color the vertices of each block Ci such that if Ci is even then half the vertices of Ci
are colored red and half are colored blue, and if Ci is odd then Ci+1
of the vertices are colored red and
are colored blue. Furthermore, if Ci is odd and Ai contains a vertex which is adjacent to S, then
we ensure that at least one such vertex x is colored red, and we add x to R.
+ V (G−S)−t
Ci−1
2
2
4
2
Observe that the number of satisfied edges in a component Cj of G − S is E(G[Cj])
, where
ej is the number of even sized blocks in Cj. Let e be the number of blocks in G− S with an even number
+ Cj+ej−1
4
2
16
of vertices. Therefore, the total number of satisfied edges within G − S is E(G−S)
. Note
also that t ≤ L. Observe that if we change the color of any vertex in R, the number of satisfied edges
within Cj is unchanged.
First, suppose that R ≥ 2S2+S+L−e+k−1
; we will show that we have a "yes"-instance. Color all
the vertices in S red, or all blue, whichever satisfies the most edges in E(S, G− S). If this satisfies at least
edges in E(S, G − S), then we are done, as the total number of satisfied
+ 2S2+S+L−e+k−1
+ V (G−S)+e−t
E(S,G−S)
4
2
2
2
4
edges is at least
E(G − S)
2
+ E(S, G − S)
2
m − S2
+ V (G − S) + e − t
e − L
n − S
+
4
2S2 + S + L − e + k − 1
+
2S2 + S + L − e + k − 1
4
≥
Otherwise, the number of satisfied edges in E(S, G − S) is between E(S,G−S)
≥
+
+
4
2
4
4
2
2S2+S+L−e+k−1
m
2
+
n − 1
4
+
k
4
.
and E(S,G−S)
2
+
2
2
4
. Now change the colors of all the vertices in R from red to blue, and recall that this
does not affect the number of satisfied edges within G − S. If the vertices of S were all colored red, then
+ 2S2+S+L−e+k−1
the number of satisfied edges in E(S, G− S) is increased by R to at least E(S,G−S)
,
and so again we are done. Otherwise, the vertices of S were all colored blue and we lose R satisfied
edges, so the number of satisfied edges in E(S, G − S) is at most E(S,G−S)
. But
then by changing the color of S from blue to red, we satisfy at least E(S,G−S)
in E(S, G − S), and we are done.
− (2S2+S+L−e+k−1)
+ 2S2+S+L−e+k−1
So now we may assume that R < 2S2+S+L−e+k−1
. Let M ⊆ [n∗] be the set of all i for which
either Ci is odd and Ai contains a vertex adjacent to S, or Ci is even. Observe that M = R + e <
2R + e < 2S2 + S + L + k − 1. Observe furthermore that L ⊆ M , as if there is a leaf-block Ci
with Ci odd and no vertex in Ai adjacent to S, we have an application of Rule 5. It remains to bound
[n∗] \ M. Recall that I ⊆ [n∗] is the set of all i for which Ci ∩ J ≥ 3, and that L ≥ I + 2. Therefore
we have that M ∪ I < 2S2 + S + L + L − 2 + k − 1 = 2S2 + S + 2L + k − 3. Finally, for
any i /∈ M ∪ I, it must be the case that Ci is odd, Ci ∩ J = 2, and Ai has no vertices adjacent to S.
But then if Ci, Cj share a vertex for any i, j /∈ M ∪ I, we have an application of Rule 7. Therefore for
i /∈ M ∪ I, Ci provides an edge between Cj and Cj′, for some j, j′ ∈ M ∪ I. From the fact that there are
no cycles not contained in blocks, it can be seen that [n∗] \ (M ∪ I) ≤ M ∪ I − 1. Thus, we have that
n∗ ≤ 2M ∪ I − 1 < 4S2 + 2S + 4L + 2k − 7.
edges
4
4
2
2
2
Lemma 12. If Ai ≥ 2S3 + 5S2 + (L + k − 3)S − 2L − J − 2k for some i ∈ {1, . . . , n∗} then
(G, k) is a "yes"-instance.
Proof. Let i be such that Ai ≥ 2S3 + 5S2 + (L + k − 3)S − 2L − J − 2k. Let n′ = Ai and
let m′ be the number of edges within Ai and between Ai and S. Observe first that if we can find an
assignment to S ∪ Ai that satisfies at least m′
of these edges, then we have a
"yes"-instance: indeed, the component of G − S containing Ai is still connected in G − (S ∪ Ai) unless it
4 + 2S2+S+L+k
2 + n′−1
4
17
is empty. Therefore, the number of components in G− (S ∪ Ai) is no more than in G− S and is therefore
at most t ≤ L. Let C1,C2, . . . ,Ct be the components of G− (S ∪ Ai). Color each component Cj optimally
so that at least E(G[Cj])
of the edges within Cj are satisfied, and then reverse the colors of Cj
if necessary to ensure that at least half the edges between Cj and S ∪ Ai are satisfied. Then the total
number of satisfied edges is at least
+ Cj −1
2
4
m − E(G[S ∪ Ai])
≥
+
2
m − m′ − S2
2
n − S ∪ Ai − t
m′
2
n − n′ − S − L
+
+
4
4
n′ − 1
+
+
m′ + S2
4
+
2
2S2 + S + L + k
4
+
n′ + S + L + k − 1
4
=
m
2
+
n − 1
4
+
k
4
,
and so we have a "yes"-instance.
4
2 + n′−1
4 + 2S2+S+L+k
R(x) be the number of neighbors of x which are assigned red, let εα
We now show that if n′ ≥ 2S3 + 5S2 + (L + k − 3)S − 2L − J − 2k then we can find a partial
assignment to S ∪ Ai that satisfies at least m′
of the edges within G[Ai] and
between Ai and S, completing the proof. For a given partial assignment α on S, and for any vertex
x ∈ Ai, let εα
B(x) be the number of
neighbors of x which are assigned blue, and let εα(x) = εα
R(x). Let x1, . . . , xl be the vertices of
Ai, ordered so that εα(x1) ≥ εα(x2) ≥ . . . ≥ εα(xl) (the ordering will depend on α). Suppose for some
r ∈ {1, . . . , n∗} we assign xj red for j ≤ n′+r
2 vertices in Ai blue. Then
2 (cid:18)P
εα(xj)(cid:19). Suppose there
the number of satisfied edges is m′
4 + 1
exist X1, X2 ⊆ Ai with X1,X2 ≥ 2S2+S+L+k
and such that εα(x) > εα(y) for all x ∈ X1, y ∈ X2.
Observe that by setting r = 0 (if n′ is even) or r ∈ {±1} (if n′ is odd), we get that
2 , and the assign the remaining n′−r
4 − r2
j=1 εα(xj) −Pn′
B(x) − εα
2 + n′
j= n′+r+2
n′ +r
2
2
2
1
2
n′+r
2
Xj=1
εα(xj) −
n′
Xj= n′+r+2
2
εα(xj)
≥
2S2 + S + L + k
4
,
and so we are done.
It remains to show that there is some partial assignment α on S such that X1, X2 exist. For the sake
of contradiction, suppose this were not the case. Then for any assignment α there exist a set of at least
n′− 2S2−S−L− k vertices x in Ai for which εα(x) is the same. Consider first the partial assignment
α for which every vertex in S is assigned red. Then for every x ∈ Ai, εα(x) = N (x) ∩ S. It follows that
N (x)∩ S is the same for at least n′ − 2S2 −S−L− k vertices x in Ai. Let A′
i be this set of vertices.
Now for each vertex s ∈ S, consider the partial assignment α for which s is assigned blue and every other
vertex in S is assigned red. Let X1 be the set of vertices in A′
i which are adjacent to s, and let X2 be
the set of vertices in A′
i which are not adjacent to s. Then εα(x) = εα(y) if x, y ∈ X1 or x, y ∈ X2,
and εα(x) = εα(y) − 2 if x ∈ X1, y ∈ X2. Therefore by our assumption either X1 < 2S2+S+L+k
or
X2 < 2S2+S+L+k
vertices in A′
i
which are adjacent to s, or at most 2S2+S+L+k
i which are not adjacent to s. Therefore,
. It follows that for each s ∈ S there are either at most 2S2+S+L+k
vertices in A′
2
2
2
2
18
there is a set of at least
A′
≥ n′ − 2S2 − S − L − k − S
vertices in Ai which are all adjacent to exactly the same vertices in S. If
i − S
2
(2S2 + S + L + k)
2S2 + S + L + k
2
2
then we would have an application of Rule 8. Therefore,
2S2 − S − L − k + S
2S2 + S + L + k
≥ Ai + J + S
2
Ai ≤ 4S2 − 2S − 2L − 2k + S(2S2 + S + L + k) − J − S
= 2S3 + 5S2 + (L + k − 3)S − 2L − J − 2k .
To complete the proof of Theorem 3, we prove Lemma 13, from which Theorem 3 follows.
Lemma 13. For a connected graph G that is reduced by Rules 5, 6, 7 and 8 and satisfies V (G) >
29160k5 + 6480k4 − 8532k3 − 492k2 + 731k − 80, the pair (G, k) is a "yes"-instance.
Proof. Observe that J ≤ n∗ − 1. By the preceding three lemmas, we may assume that
1. L < 4S2 + 2S + 2k − 2, and
2. n∗ < 4S2 + 2S + 4L + 2k − 7, and
3. Ai < 2S3 + 5S2 + (L + k − 3)S − 2L − J − 2k for any i ∈ {1, . . . , n∗},
for otherwise we have a "yes"-instance. Furthermore, we know that S < 3k. Putting everything together,
we have that
This in turn implies that
L < 4S2 + 2S + 2k − 2 < 36k2 + 8k − 2 .
and so J < 180k2 + 40k − 16. For every i ∈ {1, . . . , n∗}, we have
n∗ < 4L + 4S2 + 2S + 2k − 7 < 180k2 + 40k − 15,
Ai < 2S3 + 5S2 + (S − 2)L + (k − 3)S − J − 2k
< 54k3 + 45k2 + (3k − 2)(36k2 + 8k − 2) + 3k2 − 9k − J − 2k
≤ 162k3 − 33k + 4 .
(Here we assume that S ≥ 2 and and k ≥ 3, as otherwise the problem can be solved in polynomial
time.) Observe that V = Si∈n∗ Ai ∪ J ∪ S. Therefore, the number of vertices in G is at most
(180k2 + 40k − 16)(162k3 − 33k + 4) + (180k2 + 40k − 16) + 3k
= 29160k5 + 6480k4 − 8532k3 − 600k2 + 688k − 64 + 108k2 + 43k − 16
= 29160k5 + 6480k4 − 8532k3 − 492k2 + 731k − 80 .
19
6 Discussion and Open Problems
We showed fixed-parameter tractability of Max-Cut parameterized above the Edwards-Erdos bound
m/2 + (n− 1)/4, and thereby resolved an open question from [9, 21, 29, 30, 38]. Furthermore, we showed
that the problem has a kernel with O(k5) vertices. We have not attempted to optimize running time
or kernel size, and indeed we conjecture that Max-Cut has a kernel with O(k3) vertices and the edge
version admits a linear kernel. Our conjecture was recently answered in the affirmative, by Crowston et
al. [12].
It remains an open problem whether the weighted version of Max-Cut above the Edwards-Erdos
bound is fixed-parameter tractable; our conjecture is that this problem is also fixed-parameter tractable
and admits a polynomial kernel.
Since an extended abstract of this paper appeared [11], it was shown [10, 32] that the techniques
developed in this paper are strong enough to show fixed-parameter tractability of the Maximum Acyclic
Subdigraph problem in oriented graphs parameterized above tight lower bound; this solves another open
question by Raman and Saurabh [37] in the field of parameterized complexity.
Acknowledgment. We thank Tobias Friedrich and Gregory Gutin for help with the presentation of
the results. Part of this research has been supported by an International Joint Grant from the Royal
Society.
References
[1] Alon, N.: Bipartite subgraphs. Combinatorica 16(3), 301 -- 311 (1996)
[2] Alon, N., Halperin, E.: Bipartite subgraphs of integer weighted graphs. Discrete Mathematics 181, 19 -- 29 (1998)
[3] Andersen, L.D., Grant, D.D., Linial, N.: Extremal k-colourable subgraphs. Ars Combinatoria 16, 259 -- 270 (1983)
[4] Bandelt, H.J., Mulder, H.M.: Distance-hereditary graphs. Journal of Combinatorial Theory, Series B 41(2), 182 -- 208
(1986)
[5] Bodlaender, H.L., Downey, R.G., Fellows, M.R., Hermelin, D.: On problems without polynomial kernels. Journal of
Computer and System Sciences 75(8), 423 -- 434 (2009)
[6] Bollob´as, B., Scott, A.: Better bounds for Max Cut. In: Contemporary Combinatorics, Bolyai Society Mathematical
Studies, vol. 10, pp. 185 -- 246 (2002)
[7] Cai, L., Juedes, D.: On the existence of subexponential parameterized algorithms. Journal of Computer and System
Sciences 67(4), 789 -- 807 (2003)
[8] Charikar, M., Wirth, A.: Maximizing quadratic programs: Extending Grothendieck's inequality. In: Proc. FOCS 2004,
pp. 54 -- 60 (2004)
[9] Crowston, R., Fellows, M., Gutin, G., Jones, M., Rosamond, F., Thomass´e, S., Yeo, A.: Simultaneously Satisfying
Linear Equations Over F2: MaxLin2 and Max-r-Lin2 Parameterized Above Average. In: IARCS Annual Conference
on Foundations of Software Technology and Theoretical Computer Science (FSTTCS 2011), Leibniz International
Proceedings in Informatics, vol. 13, pp. 229 -- 240 (2011)
20
[10] Crowston, R., Gutin, G., Jones, M.: Directed acyclic subgraph problem parameterized above Poljak-Turz´ık bound. In:
IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science (FSTTCS 2012),
Leibniz International Proceedings in Informatics, vol. 18, pp. 400-411 (2012).
[11] Crowston, R., Jones, M., Mnich, M.: Max-cut parameterized above the Edwards-Erdos bound. In: Automata, Lan-
guages, and Programming, Lecture Notes in Computer Science, vol. 7391, pp. 242 -- 253 (2012)
[12] Crowston, R., Gutin, G., Jones, M., Muciaccia, G.: Maximum Balanced Subgraph Problem Parameterized above Lower
Bound In: Computing and Combinatorics, Lecture Notes in Computer Science, vol. 7936, pp. 434 -- 445 (2013)
[13] Diestel, R.: Graph Theory. Graduate Texts in Mathematics. Springer (2010)
[14] Downey, R.G., Fellows, M.R.: Parameterized complexity. Monographs in Computer Science (1999)
[15] Edwards, C.S.: Some extremal properties of bipartite subgraphs. Canadian Journal of Mathematics 25, 475 -- 485 (1973)
[16] Edwards, C.S.: An improved lower bound for the number of edges in a largest bipartite subgraph. In: Recent Advances
in Graph Theory, pp. 167 -- 181 (1975)
[17] Erdos, P.: On some extremal problems in graph theory. Israel Journal of Mathemathics 3, 113 -- 116 (1965)
[18] Erdos, P.: On even subgraphs of graphs. Matematikai Lapok 18, 283 -- 288 (1967)
[19] Erdos, P., Gy´arf´as, A., Kohayakawa, Y.: The size of the largest bipartite subgraphs. Discrete Mathematics 177, 267 -- 271
(1997)
[20] Flum, J., Grohe, M.: Parameterized complexity theory. Texts in Theoretical Computer Science. An EATCS Series.
Springer-Verlag, Berlin (2006)
[21] Gutin, G., Yeo, A.: Note on maximal bisection above tight lower bound.
Information Processing Letters 110(21),
966 -- 969 (2010)
[22] Haglin, D.J., Venkatesan, S.M.: Approximation and intractability results for the maximum cut problem and its variants.
IEEE Transactions on Computers 40(1), 110 -- 113 (1991)
[23] Hofmeister, T., Lefmann, H.: A combinatorial design approach to Max Cut. Random Structures and Algorithms 9,
163 -- 175 (1996)
[24] Impagliazzo, R., Paturi, R.: On the complexity of k-SAT. Journal of Computer and System Sciences 62(2), 367 -- 375
(2001)
[25] Karp, R.M.: Reducibility among combinatorial problems. In: Complexity of computer computations (Proceed Sympo-
sium, IBM Thomas J. Watson Research Center, Yorktown Heights, N.Y., 1972), pp. 85 -- 103 (1972)
[26] Khot, S., O'Donnell, R.: SDP gaps and UGC-hardness for Max-Cut-Gain. Theory of Computing 5, 83 -- 117 (2009)
[27] Lehel, J., Tuza, Zs.: Triangle-free partial graphs and edge covering theorems. Discrete Mathematics 39(1), 59 -- 65 (1982)
[28] Locke, S.C.: Maximum k-colorable subgraphs. Journal of Graph Theory 6(2), 123 -- 132 (1982)
[29] Mahajan, M., Raman, V.: Parameterizing above guaranteed values: MaxSat and MaxCut. Tech. Rep. TR97-033,
Electronic Colloquium on Computational Complexity (1997). http://eccc.hpi-web.de/report/1997/033/
[30] Mahajan, M., Raman, V., Sikdar, S.: Parameterizing above or below guaranteed values. Journal of Computer and
System Sciences 75(2), 137 -- 153 (2009)
21
[31] Marx, D.: Parameterized graph separation problems. Theoretical Computer Science 351(3), 394 -- 406 (2006)
[32] Mnich, M., Philip, G., Saurabh, S., Such´y, O.: Beyond Max-Cut: λ-extendible properties parameterized above the
Poljak-Turz´ık bound. In: IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer
Science (FSTTCS 2012), Leibniz International Proceedings in Informatics, vol. 18, pp. 412 -- 423 (2012).
[33] Mnich, M., Zenklusen, R.: Bisections above tight lower bounds. In: Graph-Theoretic Concepts in Computer Science,
Lecture Notes in Computer Science, vol. 7551, pp. 184 -- 193 (2012).
[34] Ngo. c, N.V., Tuza, Zs.: Linear-time approximation algorithms for the max cut problem. Combinatorics, Probability
and Computing 2(2), 201 -- 210 (1993)
[35] Poljak, S., Turz´ık, D.: A polynomial algorithm for constructing a large bipartite subgraph, with an application to a
satisfiability problem. Canadian Journal of Mathematics 34(3), 519 -- 524 (1982)
[36] Poljak, S., Tuza, Z.: Maximum cuts and large bipartite subgraphs. In: Combinatorial optimization (New Brunswick,
NJ, 1992 -- 1993), DIMACS Series in Discrete Mathematics and Theoretical Computer Science, vol. 20, pp. 181 -- 244
(1995)
[37] Raman, V., Saurabh, S.: Parameterized algorithms for feedback set problems and their duals in tournaments. Theo-
retical Computer Science 351(3), 446 -- 458 (2006)
[38] Sikdar, S.: Parameterizing from the Extremes: Feasible Parameterizations of some NP-optimization problems. Ph.D.
thesis, The Institute of Mathematical Sciences, Chennai, India (2010)
22
|
1509.05494 | 1 | 1509 | 2015-09-18T03:08:55 | FPTAS for Hardcore and Ising Models on Hypergraphs | [
"cs.DS"
] | Hardcore and Ising models are two most important families of two state spin systems in statistic physics. Partition function of spin systems is the center concept in statistic physics which connects microscopic particles and their interactions with their macroscopic and statistical properties of materials such as energy, entropy, ferromagnetism, etc. If each local interaction of the system involves only two particles, the system can be described by a graph. In this case, fully polynomial-time approximation scheme (FPTAS) for computing the partition function of both hardcore and anti-ferromagnetic Ising model was designed up to the uniqueness condition of the system. These result are the best possible since approximately computing the partition function beyond this threshold is NP-hard. In this paper, we generalize these results to general physics systems, where each local interaction may involves multiple particles. Such systems are described by hypergraphs. For hardcore model, we also provide FPTAS up to the uniqueness condition, and for anti-ferromagnetic Ising model, we obtain FPTAS where a slightly stronger condition holds. | cs.DS | cs | FPTAS for Hardcore and Ising Models on Hypergraphs
Pinyan Lu∗
Kuan Yang†
Chihao Zhang‡
Abstract
Hardcore and Ising models are two most important families of two state spin systems in
statistic physics. Partition function of spin systems is the center concept in statistic physics
which connects microscopic particles and their interactions with their macroscopic and statistical
properties of materials such as energy, entropy, ferromagnetism, etc. If each local interaction of
the system involves only two particles, the system can be described by a graph. In this case,
fully polynomial-time approximation scheme (FPTAS) for computing the partition function of
both hardcore and anti-ferromagnetic Ising model was designed up to the uniqueness condition
of the system. These result are the best possible since approximately computing the partition
function beyond this threshold is NP-hard. In this paper, we generalize these results to general
physics systems, where each local interaction may involves multiple particles. Such systems are
described by hypergraphs. For hardcore model, we also provide FPTAS up to the uniqueness
condition, and for anti-ferromagnetic Ising model, we obtain FPTAS where a slightly stronger
condition holds.
5
1
0
2
p
e
S
8
1
]
S
D
.
s
c
[
1
v
4
9
4
5
0
.
9
0
5
1
:
v
i
X
r
a
∗Microsoft Research. [email protected]
†Shanghai Jiao Tong University. [email protected]
‡Shanghai Jiao Tong University. [email protected]
1
Introduction
In recent couple of years, there are remarkable progress on designing approximate counting algo-
rithms based on correlation decay approach [Wei06, BG08, GK12, RST+11, LLY12, SST12, LLY13,
LY13, LLL14, LWZ14, LLZ14, LL15b, LL15a]. Unlike the previous major approximate counting
approach that based on random sampling such as Markov Chain Monte Carlo (MCMC) (see for
examples [JS97, JS93, JSV04, GJ11, DJV02, Jer95, Vig99, DFJ02, DG00, LV97]), correlation decay
based approach provides deterministic fully polynomial-time approximation scheme (FPTAS). New
FPTASes were designed for a number of interesting combinatorial counting problems and computing
partition functions for statistic physics systems, where partition function is a weighted counting
function from the computational point of view. One most successful example is the algorithm
for anti-ferromagnetic two-spin systems [LLY12, SST12, LLY13], including counting independent
sets [Wei06]. The correlation decay based FPTAS is beyond the best known MCMC based FPRAS
and achieves the boundary of approximability [SS12, GSV12].
In this paper, we generalize these results of anti-ferromagnetic two-spin systems to hypergraphs.
For physics point of view, this corresponds to spin systems with higher order interactions, where
each local interaction involves more than two particles. There are two main ingredients for the
original algorithms and analysis on normal graphs (we will use the term normal graph for a graph
to emphasize that it is not hypergraphs): (1) the construction of the self-avoiding walk tree by
Weitz [Wei06], which transform a general graph to a tree; (2) correlation decay proof for the tree,
which enables one to truncate the tree to get a good approximation in polynomial time. However,
the construction of the self-avoiding walk tree cannot be extended to hypergraphs, which is the
main obstacle for the generalization.
The most related previous work is counting independent sets for hypergraphs by Liu and
Lu [LL15b]. They established a computation tree with a two-layers recursive function instead
of the self-avoiding walk tree and provided a FPTAS to count the number of independent sets for
hypergraphs with maximum degree of 5, extending the algorithm for normal graph with the same
degree bound. Their proof was significantly more complicated than the previous one due to the
complication of the two-layers recursive function. In particular, the "right" degree bound for the
problem is a real number between 5 and 6 if one allow fraction degree in some sense. This integer
gap provides some room of flexibility and enables them to do some case-by-case numerical argument
to complete the proof. However, the parameters for the anti-ferromagnetic two-spin systems on
hypergraphs are real numbers. To get a sharp threshold, we do not have any room for numerical
approximation.
1.1 Our results
We study two most important anti-ferromagnetic two-spin systems on hypergraphs: the hardcore
model and the anti-ferromagnetic Ising model. The formal definitions of these two models can be
found in Section 2.
Our first result is an FPTAS to compute the partition function of hypergraph hardcore model.
Theorem 1. For hardcore model with a constant activity parameter of λ, there is an FPTAS to
compute the partition function for hypergraphs with maximum degree ∆ ≥ 2 if λ < (∆−1)∆−1
(∆−2)∆ .
This bound is exactly the uniqueness threshold for the hardcore on normal graphs. Thus,
it is tight since normal graphs are special cases of hypergraphs. To approximately compute the
partition function beyond this threshold is NP-hard.
In particular, The FPTAS in [LL15b] for
counting the number of independent sets for hypergraphs with maximum degree of 5 can be viewed
1
as a special case of our result with parameters ∆ = 5, λ = 1, which satisfies the above uniqueness
condition. Another interesting special case is when ∆ = 2. This is not an interesting case for normal
graphs since a normal graph with maximum degree of 2 is simply a disjoint union of paths and
cycles, whose partition function can be computed exactly. However, the problem becomes more
complicated on hypergraphs:
it can be interpreted as counting weighted edge covers on normal
graphs by viewing vertices of degree two as edges and hyperedges as vertices. The exact counting
of this problem is known to be #P-complete and an FPTAS was found recently [LLZ14]. In our
model, the uniqueness bound (∆−1)∆−1
(∆−2)∆ is infinite for ∆ = 2 and as a result we give an FPTAS for
counting weighted edge covers for any constant edge weight λ. This gives an alternative proof for
the main result in [LLZ14].
Our second result is on computing the partition function of anti-ferromagnetic Ising model.
Theorem 2. For Ising model with interaction parameter 0 < β < 1 and external field λ, there
is an FPTAS to compute the partition function for hypergraphs with maximum degree ∆ if β ≥
1 −
2e−1/2∆+3
2
.
The tight uniqueness bound for anti-ferromagnetic Ising model on normal graphs is β ≥ 1 − 2
∆ .
So, our bound is in the same asymptotic order but a bit worse in the constant coefficient as
2e−1/2 ≈ 1.213 > 1. Moreover, our result can apply beyond Ising model to a larger family of
anti-ferromagnetic two-spin systems on hypergraphs.
1.2 Our techniques
We also use the correlation decay approach. Although the framework of this method is standard,
in many work along this line of research, new tools and techniques are developed to make this
relatively new approach more powerful and widely applicable. This is indeed the case for the
current paper as well. We summarize the new techniques we introduced here.
For hardcore model, we replace the numerical case-by-case analysis by a monotone argument
with respect to the edge size of the hypergraph which shows that the normal graphs with edge size
of 2 is indeed the worst cases. This gives a tight bound for hardcore model.
To handle hypergraph with unbounded edge sizes, we need to prove that the decay rate is much
smaller for edges of larger size. Such effect is called computationally efficient correlation decay,
which has been used in many previous works to obtain FPTASes for systems with unbounded
degrees or edge sizes. In all those works, one sets a threshold for the parameter and proves different
types of bounds for large and small ones separately. Such artificial separation gets a discontinuous
bound which adds some complications in the proof and usually ends with a case-by-case discussion.
In particular, this separation is not compatible with the above monotone argument. To overcome
this, we propose a new uniform and smooth treatment for this by modifying the decay rate by
a polynomial function of the edge size. After this modification, we only need to prove one single
bound which automatically provides computationally efficient correlation. We believe that this idea
is important and may find applications in other related problems.
For the Ising model, the main difficulty is to get a computation tree as a replacement of the
self-avoiding walk tree. We proposed one, which also works for general anti-ferromagnetic two-
spin systems on hypergraphs. However, unlike the case of the hardcore model, the computation
tree is not of perfect efficiency and this is the main reason that the bound we achieve in Theo-
rem 2 is not tight. To get the computationally efficient correlation decay, we also use the above
mentioned uniform and smooth treatment. We also extend our result beyond Ising to a family of
anti-ferromagnetic two-spin systems on hypergraphs.
2
1.3 Discussion and open problems
One obvious open question is to close the gap for Ising model, or more generally extend our work
to anti-ferromagnetic two-spin systems on hypergraphs with better parameters. However, it seems
that it is impossible to obtain a tight result in these models using the computation tree proposed
in this paper, due to its imperfectness. How to overcome this is an important open question.
In particular, MCMC based approach does show that larger edge size helps:
Even for the hardcore model, our result is tight only for the family of all hypergraphs since the
normal graphs are special cases. From both physics and combinatorics point of view, it would be
very interesting to study the family of w-uniform hypergraphs where each hyperedge is of the same
size w. By our monotone argument, it is plausible to conjecture that one can get better bound
for larger w.
for
hypergraph independent set with maximum degree of ∆ and minimum edge size w, an FPRAS for
w ≥ 2∆ + 1 was shown in [BDK08]. However, their result is not tight. Can we get a tight bounds in
terms of ∆ and w by correlation decay approach? The high level idea sounds promising, but there
is an obstacle to prove such result by our computation tree. To construct the computation tree,
we need to construct modified instances. In these modified instances, the size of a hyperedge may
decrease to as low as 2. Therefore, even if we start with w-uniform hypergraphs or hypergraphs
with minimum edge size of w, we may need to handle the worst case of normal graphs during the
analysis. How to avoid this effect is a major open question whose solution may have applications
in many other problems.
The fact that larger hyperedge size only makes the problem easier is not universally true for
approximation counting. One interesting example is counting hypergraph matchings. FPTAS for
counting 3D matchings of hypergraphs with maximum degree 4 is given in [LL15b], and extension
to weighted setting are studied in [YZ15]. In particular, a uniqueness condition in this setting is
defined in [YZ15], and it is a very interesting open question whether this uniqueness condition is
also the transition boundary for approximability.
2 Preliminaries
A hypergraph G(V,E) consists of a vertex set V and a set of hyperedges E ⊆ 2V . For every
hyperedge e ∈ E and vertex v ∈ V , we use e − v to denote e \ {v} and use e + v to denote e ∪ {v}.
2.1 Hypergraph hardcore model
The hardcore model is parameterized by the activity parameter λ > 0. Let G(V,E) be a hypergraph.
An independent set of G is a vertex set I ⊆ V such that e 6⊆ I for every hyperedge e ∈ E. We use
I(G) to denote the set of independent sets of G. The weight of an independent set I is defined as
w(I) , λI. We let Z(G) denote the partition function of G(V,E) in the hardcore model, which is
defined as
Z(G) , XI∈I(G)
w(I).
The weight of independent sets induces a Gibbs measure on G. For every I ∈ I, we use
PrG [I] ,
w(I)
Z(G)
3
to denote the probability of obtaining I if we sample according to the Gibbs measure. For every
v ∈ V , we use
PrG [v ∈ I] , XI∈I(G)
v∈I
PrG [I]
to denote the marginal probability of v.
2.2 Hypergraph two state spin model
Now we give a formal definition to hypergraph two state spin systems. This model is parameterized
by the external field λ > 0. An instance of the model is a labeled hypergraph G(V,E, (β, γ)) where
β, γ : E → R are two labeling functions that assign each edge e ∈ E two reals β(e), γ(e). A
configuration on G is an assignment σ : V → {0, 1} whose weight w(σ) is defined as
w(σ) , Ye∈E
w(e, σ) Yv∈V
w(v, σ)
where for a hyperedge e = {v1, . . . , vw}
β(e)
γ(e)
1
w(e, σ) ,
and for a vertex v,
if σ(v1) = σ(v2) = ··· = σ(vw) = 0
if σ(v1) = σ(v2) = ··· = σ(vw) = 1
otherwise
w(v, σ) ,(λ if σ(v) = 1
otherwise.
1
The partition function of the instance is given by
Z(G) = Xσ∈{0,1}V
w(σ).
Similarly, the weight of configurations induces a Gibbs measure on G. For every σ ∈ {0, 1}V , we
use
PrG [σ] ,
w(I)
Z(G)
to denote the probability of σ in the measure. For every v ∈ V , we use
PrG [σ(v) = 1] , Xσ∈{0,1}V
σ(v)=1
PrG [σ]
to denote the marginal probability of v.
The anti-ferromagnetic Ising model is the special case that β , β(e) = γ(e) ≤ 1 for all e ∈ E.
In this model, we call β the interaction parameter of the model. The hardcore model introduced
in previous section is the special case that β(e) = 1 and γ(e) = 0 for all e ∈ E.
the more general two state spin system and establish the following theorem:
We give the whole proof to Theorem 2 in appendix. More precisely, we design an FPTAS for
4
Theorem 3. Consider a class of two state spin system with external field λ such that each instance
2e−1/2∆+3 ≤ β(e), γ(e) ≤ 1 where ∆ is the maximum degree
G(V,E, (β, γ)) in the class satisfies 1−
of G. There exists an FPTAS to compute the partition function for every instance in the class.
2
Theorem 2 then follows since it is a special case of Theorem 3.
Actually, the main idea of FPTAS design and proof for this model is similar to the idea we use to
solve hypergraph hardcore model. However, the details of recursion function design and techniques
for proof of correlation decay property are pretty different from that in hypergraph hardcore model,
so we put the whole section in appendix.
3 Hypergraph Hardcore Model
3.1 Recursion for computing marginal probability
We first fix some notations on graph modification specific to hypergraph independent set. Let
G(V,E) be a hypergraph.
• For every v ∈ V , we denote G − v , (V \ {v} ,E′) where E′ , {e \ {v} e ∈ E}.
• For every e ∈ E, we denote G − e , (V,E \ {e}).
• Let x be a vertex or an edge and y be a vertex or an edge, we denote G− x− y , (G− x)− y.
• Let S = {v1, . . . , vk} ⊆ V , we denote G − S , G − v1 − v2 ··· − vk.
• Let F = {e1, . . . , ek} ⊆ E, we denote G − F , G − e1 − e2 ··· − ek.
Let G(V,E) be a hypergraph and v ∈ V be an arbitrary vertex with degree d. Let {e1, . . . , ed}
be the set of hyperedges incident to v and for every i ∈ [d], ei = {v} ∪ {vij j ∈ [wi]} consists of
wi + 1 vertices.
We first define a graph G′(V ′,E′), which is the graph obtained from G by replacing v by d
copies of itself and each ei contains a distinct copy. Formally, V ′ , (V \ {v}) ∪ {v1, . . . , vd},
E′ , {e ∈ E v 6∈ e} ∪ {ei − v + vi i ∈ [d]}.
For every i ∈ [d] and j ∈ [wi], we define a hypergraph Gij(Vij,Eij):
Gij , G′ − {vk i ≤ k ≤ d} − {ek 1 ≤ k ≤ i} − {vik 1 ≤ k < j} .
Let Rv = PrG[v∈I]
PrG[v6∈I] and Rij =
PrGij [vij∈I]
PrGij [vij6∈I] . We can compute Rv by following recursion:
Lemma 4.
Rv = λ
d
Yi=1
The proof of this lemma is postponed to appendix.
1 −
wi
Yj=1
Rij
1 + Rij
.
(1)
The Uniqueness Condition Let the underlying graph be an infinite d-ary tree, then the recur-
sion (1) becomes
uniqueness of the Gibbs measure is that (cid:12)(cid:12)(cid:12)
1 + x(cid:19)d
.
fλ,d(x) = λ(cid:18) 1
f′λ,d(x)(cid:12)(cid:12)(cid:12)
5
Let x be the positive fixed-point of fλ,d(x), i.e., x > 0 and fλ,d(x) = x. The condition on λ for the
< 1. The following proposition is well-known.
< 1.
f′λ,d(x)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
Proposition 5. Let λc =
dd
(d−1)d+1 , then (cid:12)(cid:12)(cid:12)
f′λc,d(x)(cid:12)(cid:12)(cid:12)
= 1 and for every 0 < λ < λc, it holds that
3.2 The algorithm to compute marginal probability
Let G(V,E) be a hypergraph with maximum degree ∆ and v ∈ V be an arbitrary vertex with
degree d. Define Gij, Rv, Rij as in Section 3.1. Then the recursion (1) gives a way to compute the
marginal probability PrG [v ∈ I] exactly. However, an exact evaluation of the recursion requires a
computation tree with exponential size. Thus we introduce the following truncated version of the
recursion, with respect to constants c > 0 and 0 < α < 1.
R(G, v, L) =
λQd
λQd
i=1(cid:16)1 −Qwi
i=1(cid:16)1 −Qwi
λ
j=1
j=1
R(Gij ,vij ,L)
1+R(Gij ,vij ,L)(cid:17)
1+R(Gij ,vij ,L−⌊1+c log1/α wi⌋)(cid:17) if d < ∆ and L > 0
R(Gij ,vij ,L−⌊1+c log1/α wi⌋)
if d = ∆
otherwise.
The recursion can be directly used to compute R(G, v, L) for any given L and it induces a
truncated computation tree (with height L in some special metric). It is worth noting that, the
case that d = ∆ can only happen at the root of the computation tree, since in each smaller instance,
the degree of vij is decreased by at least one.
We claim that R(G, v, L) is a good estimate of Rv with a suitable choice of c and α, for those
(λ, ∆) in the uniqueness region.
Lemma 6. Let G(V,E) be a hypergraph with maximum degree ∆ ≥ 2. Let v ∈ V be a vertex with
degree d and let λ < λc = (∆−1)∆−1
(∆−2)∆ be the activity parameter. There exist constants C > 0 (more
precisely, C = 6λ√1 + λ) and α ∈ (0, 1) such that
R(G, v, L) − Rv ≤ C · αmax{0,L}
for every L.
The whole proof of this lemma is postponed to the next section.
Proof of Theorem 1. Assuming Lemma 6, the proof of Theorem 1 is routine and we put the proof
into appendix.
3.3 Correlation decay
In this section, we establish Lemma 6. We first prove some technical lemmas.
Suppose f : Dd → R is a d-ary function where D ⊆ R is a convex set, let φ : R → R be an
increasing differentiable function and Φ(x) , φ′(x). The following proposition is a consequence of
the mean value theorem:
Proposition 7. For every x = (x1, . . . , xd), x = (x1, . . . , xd) ∈ Dd, it holds that
Φ(x) φ(f (x)) − φ(f (x)) for some x ∈ D;
1. f (x) − f (x) = 1
2. φ(f (x)) − φ(f (x)) ≤Pd
i=1
Φ(f )
Φ(xi)(cid:12)(cid:12)(cid:12)
∂f (x)
∂xi (cid:12)(cid:12)(cid:12) · φ(xi) − φ(xi) for some x = (x1, . . . , xd) ∈ Dd.
6
xij
1√x(1+x)
Lemma 8. Let ∆ ≥ 2 be a constant integer and λ < λc = (∆−1)∆−1
(∆−2)∆ be a constant real. Let d < ∆
1+xij(cid:17) be a (cid:16)Pd
i=1 wi(cid:17)-ary function.
and w1, . . . , wd > 0 be integers and f = λQd
. Let c < minn log(1+λ)−log λ
λ (cid:1) − 1o be a positive number. There
Let Φ(x) =
exists a constant α < 1 depending on λ and d (but not depending on wi for all i ∈ [d]) such that
Φ(xab)(cid:12)(cid:12)(cid:12)(cid:12)
i=1(cid:16)1 −Qwi
log(cid:0) 1+λ
∂xab (cid:12)(cid:12)(cid:12)(cid:12)
≤ α < 1
Xb=1
Xa=1
for every x = (xij)i∈[d],j∈[wi] where each xij ∈ [0, λ]
, 2λ+1
∂f (x)
Φ(f )
wc
a
2+4λ
j=1
d
wa
2
The lemma bounds the amortized decay rate, which is the key to the proof of correlation decay.
In previous works, the amortized decay rate is defined as
d
Xa=1
wa
Xb=1
,
Φ(f )
Φ(xab)(cid:12)(cid:12)(cid:12)(cid:12)
∂f (x)
∂xab (cid:12)(cid:12)(cid:12)(cid:12)
without the wc
a factor. Then one need to give a constant α < 1 bound for small wa and a sub
constant bound for large wa. With this modification, we only need to prove a single bound as
above.
Notice that we require c to be a positive constant, so it is necessary to verify that 2λ+1
1 > 0 for every λ > 0. To see this, let h(λ) , 2λ+1
log(cid:0) 1+λ
λ (cid:1)−
λ (cid:1) − 1, then we can compute that
1 + 2λ
2λ + 2λ2 ,
2
2
log(cid:0) 1+λ
λ (cid:19) −
h′(λ) = log(cid:18)1 + λ
2λ2(1 + λ)2 .
h′′(λ) =
1
Since h′′(λ) > 0 for every λ, h′(λ) is increasing. Along with the fact that limλ→∞ h′(λ) = 0, we
have h′(λ) < 0 for every λ > 0. This implies that h(λ) is decreasing. Also note that
lim
λ→∞
h(λ) = lim
λ→∞
log (cid:18)1 +
1
λ(cid:19)λ(cid:18)1 +
1
λ(cid:19)1/2! − 1 = 0.
It holds that h(λ) > 0 for every λ > 0. Thus a positive c satisfying c < h(λ) exists for every λ > 0.
Proof of Lemma 8. To simplify the notation, we first let tij = xij
1+xij
, then for every i ∈ [d] and
1 −
wi
Yj=1
tij
= f ·
(1 − tab)2
tab
7
j=1 taj
· Qwa
1 −Qwa
j=1 taj
.
j ∈ [wi], it holds that tij ∈h0, λ
1+λi and
f = λ
For every a ∈ [d] and b ∈ [wi], we have
= λ(1 − tab)2 Yj∈[wa]
∂f
taj · Yi∈[d]
i6=a
j6=b
(cid:12)(cid:12)(cid:12)(cid:12)
∂xab(cid:12)(cid:12)(cid:12)(cid:12)
d
Yi=1
1 −
wi
Yj=1
tij
.
Thus
d
wa
wa
d
wa
∂f
wc
a
d
wc
Φ(f )
1 + f
=s f
1 + f
wc
aQwa
1 −Qwa
j=1 taj
j=1 taj
Xb=1
Xa=1
Xa=1
1 − tab√tab
Xb=1
Let t = (tij)i∈[d],j∈[wi], define
h(t) ,s f
=vuuut
∂xab(cid:12)(cid:12)(cid:12)(cid:12)
Φ(xab)(cid:12)(cid:12)(cid:12)(cid:12)
aQwa
Xb=1
Xa=1
1 −Qwa
j=1 tij(cid:17)
i=1(cid:16)1 −Qwi
λQd
j=1 tij(cid:17)
i=1(cid:16)1 −Qwi
1 + λQd
For every t = (tij)i∈[d],j∈[wi] where each tij ∈ [0, λ
1+λ ], define a tuple t = (tij)i∈[d],j∈wi such that for
every i ∈ [d],
λ (cid:1)wi−1
Qwi
aQwa
1 −Qwa
tij =((cid:0) 1+λ
if j = 1
otherwise.
1 − tab√tab
1 − tab√tab
j=1 taj
taj
j=1 taj
taj
Xa=1
Xb=1
k=1 tik
λ
1+λ
d
wc
wa
.
.
j
j
We claim that h(t) ≤ h(t). To see this, first note that for every i ∈ [d], Qwi
sufficient to prove that for every i ∈ [d]
Xj=1
1 − tij√tij ≤
Xj=1
wi
wi
.
1 − tij
qtij
j=1 tij = Qwi
j=1
tij, it is
This is a consequence of the Karamata's inequality by noticing that the function 1−ex
√ex is convex.
We rename ti1 to ti and it is sufficient to upper bound
g(t, w) ,vuuuuut
where ti ∈h0, λ
d
i=1(cid:18)1 −(cid:16) λ
i=1(cid:18)1 −(cid:16) λ
ti(cid:19)
1+λ(cid:17)wi−1
λQd
Xi=1
ti(cid:19) ·
1+λ(cid:17)wi−1
1 + λQd
1+λi and wi ∈ Z+ for every i ∈ [d].
For every i ∈ [d], we let zi , 1 −(cid:16) λ
For every fixed z = (z1, . . . , zd), we can write (2) as
1+λ(cid:17)wi−1
The argument so far is similar to the proof in [LL15b]. In the following, we prove a monotonicity
property of each wi and thus avoid the heavy numerical analysis in [LL15b] and allow us to obtain
a tight result.
wc
1+λ(cid:17)wi−1
i (cid:16) λ
1+λ(cid:17)wi−1
1 −(cid:16) λ
ti
ti
·(cid:18) 1 − ti√ti
+
√λ + λ2(cid:19) (2)
(wi − 1)
ti and thus equivalently ti = (1 − zi)(cid:0) 1+λ
zi (cid:18) 1 − ti√ti
1 − zi
√λ + λ2(cid:19) wc
wi − 1
i .
+
.
λ (cid:1)wi−1
(3)
(4)
i=1 zi
gz(w) =s λQd
1 + λQd
i=1 zi
d
Xi=1
We show that gz(w) is monotonically decreasing with wi for every i ∈ [d].
Denote Ti , 1−ti√ti
+ (wi−1)
√λ+λ2 , then
∂gz(w)
∂wi
=s λQd
1 + λQd
i=1 zi
i=1 zi ·
zi (cid:18) ∂Ti
1 − zi
∂wi
8
wc
i + cwc−1
i Ti(cid:19) .
The partial derivative (4) is negative for a suitable choice of c:
1 − zi
zi
1 − zi
zi
1 − zi
zi
1 − zi
zi
=
=
=
Denote
i
1
1
2
wc
∂zi
+ t−3/2
t′i(t−1/2
i + cwc−1
i Ti(cid:19)
·(cid:18) ∂Ti
·(cid:18)(cid:18)−
√λ + λ2(cid:19) wc
log(cid:18) 1 + λ
i (cid:18)(cid:18)−
λ (cid:19) (t1/2
i + t−1/2
· wc−1
i (cid:18) (c + 1)wi − c
√λ + λ2 −(cid:18)t1/2
i (cid:18)1
· wc−1
) +
1
2
2
i
i
) +
wi log(cid:18) 1 + λ
i + cwc−1
1
+
√λ + λ2(cid:19)(cid:19)
i (cid:18) 1 − ti√ti
(wi − 1)
√λ + λ2(cid:19) wi + c(cid:18)t−1/2
− t1/2
wi log(cid:18) 1 + λ
λ (cid:19) + c(cid:19) + t−1/2
(cid:18)1
i +
2
i
i
√λ + λ2(cid:19)(cid:19)
(wi − 1)
λ (cid:19) − c(cid:19)(cid:19)(cid:19)
p(t, w) ,
√λ + λ2 −(cid:18)t1/2(cid:18) 1
(c + 1)w − c
2
Since c ≤ log(1+λ)−log λ
2+4λ
, the term
w log(cid:18) 1 + λ
λ (cid:19) + c(cid:19) + t−1/2(cid:18) 1
2
w log(cid:18)1 + λ
λ (cid:19) − c(cid:19)(cid:19)
t1/2(cid:18) 1
2
w log(cid:18) 1 + λ
λ (cid:19) + c(cid:19) + t−1/2(cid:18) 1
2
w log(cid:18)1 + λ
λ (cid:19) − c(cid:19)
achieves its minimum at t = λ
1+λ . Thus
Moreover, c < 2λ+1
1 + λ
p(t, w) ≤ p(cid:18) λ
log(cid:0) 1+λ
, w(cid:19) =(cid:18) λ
λ (cid:1) − 1 implies that c+1
1 + λ(cid:19)1/2(cid:18) c + 1
λ −
2λ log(cid:0) 1+λ
λ < 2λ+1
2λ + 1
2λ
2
leads to p(cid:16) λ
in p(cid:16) λ
1+λ , 1(cid:17) < 0.
In all, we choose a positive constant c < minn log(1+λ)−log λ
1+λ , w(cid:17) ≤ p(cid:16) λ
1+λ , 1(cid:17) < 0.
2+4λ
λ (cid:19)(cid:19) w.
log(cid:18) 1 + λ
λ (cid:1) holds, which consequently
λ (cid:1) − 1o, and this results
log(cid:0) 1+λ
, 2λ+1
2
In light of the monotonicity of wi's, for every fixed z, gz(w) achieves its maximum when w = 1.
Thus
g(t, w) =
max
t∈[0, λ
1+λ ]d
max
z=(z1,...,zd)
∀i∈[d],zi∈h1−( λ
1+λ )wi −1
gz(w) ≤
,1i
max
z=(z1,...,zd)
∀i∈[d],zi∈h1−( λ
1+λ )wi −1
gz(1) ≤ max
z∈[0,1]d
gz(1).
,1i
Actually, the case that all wi's are 1 corresponds to counting weighted independent sets on
normal graphs and arguments to bound gz(1) can be found in [LLY13]. For the sake of completeness,
we give a proof of gz(1) ≤ α < 1 (Lemma 9) in appendix.
Lemma 9. Let ∆ > 1, be a constant. Assume λ < λc = (∆−1)∆−1
for some constant α < 1, gz(1) ≤ α < 1 where z = (z1, . . . , zd) ∈ [0, 1]d.
(∆−2)∆ be a constant and d < ∆. Then
We are now ready to prove the main lemma.
Proof of Lemma 6. Let Φ(x) =
and φ(x) = R Φ(x) dx = 2 sinh−1(√x). We first apply
induction on ℓ , max{0, L} to show that, if d < ∆, then φ(Rv) − φ(R(G, v, L)) ≤ 2√λαL for
1√x(1+x)
some constant α < 1.
9
If ℓ = 0, note that Rv ∈ [0, λ], thus φ(R(G, v, L)) − φ(Rv) ≤ 2√λ. We now assume L = ℓ > 0
and the lemma holds for smaller ℓ. For every i ∈ [d] and j ∈ [wi], we denote xij = Rij and
xij = R(Gij, vij, L − ⌊1 + c log1/α wi⌋). Let x = (xij)i∈[d],j∈[wi], x = (xij)i∈[d],j∈[wi]. Let f =
1−xij(cid:17) , then it follows from Proposition 7 that for some x = (xij)i∈[d],j∈[wi] with
i=1(cid:16)1 −Qwi
λQd
each xij ∈ [0, λ]
j=1
xij
φ(Rv) − φ(R(G, v, L)) ≤
· φ(xij) − φ(xij)
d
wi
(♠)
d
wi
Φ(f )
∂f (x)
Φ(xij)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
Xi=1
≤ 2√λ
Xj=1
Xi=1
≤ 2√λαL.
∂xij (cid:12)(cid:12)(cid:12)(cid:12)
Φ(xij)(cid:12)(cid:12)(cid:12)(cid:12)
Φ(f )
(♥)
∂f (x)
∂xij (cid:12)(cid:12)(cid:12)(cid:12)
αL−⌊1+c log1/α wi⌋
(♠) follows from the induction hypothesis and (♥) is due to Lemma 8.
The case that d = ∆ can only happen at the root of our computational tree. Following the
arguments in the proofs of 8, 9 and the bound in (5), it is easy to see that a universal constant
upper bound for the error contraction exists, i.e.,
∆
wi
Xi=1
Xj=1
Φ(f )
Φ(xij)(cid:12)(cid:12)(cid:12)(cid:12)
∂f (x)
∂xij (cid:12)(cid:12)(cid:12)(cid:12)
Thus φ(Rv) − φ(R(G, v, L)) ≤ 6√λαL for every v.
Then the lemma follows from Proposition 7, since
< max
z∈[0,1]s ∆2(∆ − 1)∆−1z∆(1 − z)
(∆ − 2)∆ + (∆ − 1)∆−1z∆ < 3.
Rv − R(G, v, L) =
1
Φ(x) · φ(Rv) − φ(R(G, v, L))
≤ 6λ√1 + λ · αL
for some x ∈ [0, λ]
References
[BDK08] Magnus Bordewich, Martin Dyer, and Marek Karpinski. Path coupling using stopping
times and counting independent sets and colorings in hypergraphs. Random Structures
& Algorithms, 32(3):375 -- 399, 2008.
[BG08]
Antar Bandyopadhyay and David Gamarnik. Counting without sampling: Asymptotics
of the log-partition function for certain statistical physics models. Random Structures
& Algorithms, 33(4):452 -- 479, 2008.
[DFJ02] Martin E. Dyer, Alan M. Frieze, and Mark Jerrum. On counting independent sets in
sparse graphs. SIAM Jounal on Computing, 31(5):1527 -- 1541, 2002.
[DG00] Martin E. Dyer and Catherine S. Greenhill. On Markov Chains for independent sets.
Journal of Algorithms, 35(1):17 -- 49, 2000.
[DJV02] Martin Dyer, Mark Jerrum, and Eric Vigoda. Rapidly mixing Markov chains for disman-
tleable constraint graphs. In Randomization and Approximation Techniques in Computer
Science, pages 68 -- 77. Springer, 2002.
10
[GJ11]
Leslie Ann Goldberg and Mark Jerrum. A polynomial-time algorithm for estimating the
partition function of the ferromagnetic ising model on a regular matroid. In Proceedings
of ICALP, pages 521 -- 532, 2011.
[GK12]
David Gamarnik and Dmitriy Katz. Correlation decay and deterministic FPTAS for
counting colorings of a graph. Journal of Discrete Algorithms, 12:29 -- 47, 2012.
[GSV12] A. Galanis, D. Stefankovic, and E. Vigoda. Inapproximability of the partition function
for the antiferromagnetic ising and hard-core models. Arxiv preprint arXiv:1203.2226,
2012.
[Jer95] Mark Jerrum. A very simple algorithm for estimating the number of k-colorings of a
low-degree graph. Random Structures & Algorithms, 7(2):157 -- 166, 1995.
[JS93]
[JS97]
Mark Jerrum and Alistair Sinclair. Polynomial-time approximation algorithms for the
ising model. SIAM Journal on Computing, 22(5):1087 -- 1116, 1993.
Mark Jerrum and Alistair Sinclair. The Markov chain Monte Carlo method: an approach
to approximate counting and integration, pages 482 -- 520. PWS Publishing Co., Boston,
MA, USA, 1997.
[JSV04] Mark Jerrum, Alistair Sinclair, and Eric Vigoda. A polynomial-time approximation
algorithm for the permanent of a matrix with nonnegative entries. Journal of the ACM,
51:671 -- 697, July 2004.
[LL15a]
Jingcheng Liu and Pinyan Lu. FPTAS for #BIS with degree bounds on one side. In
Proceedings of STOC (to appear), 2015.
[LL15b]
Jingcheng Liu and Pinyan Lu. FPTAS for counting monotone CNF. In Proceedings of
the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA, pages
1531 -- 1548, 2015.
[LLL14] Chengyu Lin, Jingcheng Liu, and Pinyan Lu. A simple FPTAS for counting edge cov-
In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete
ers.
Algorithms, pages 341 -- 348, 2014.
[LLY12]
Liang Li, Pinyan Lu, and Yitong Yin. Approximate counting via correlation decay in
spin systems. In Proceedings of SODA, pages 922 -- 940, 2012.
[LLY13]
Liang Li, Pinyan Lu, and Yitong Yin. Correlation decay up to uniqueness in spin
systems. In Proceedings of SODA, pages 67 -- 84, 2013.
[LLZ14]
Jingcheng Liu, Pinyan Lu, and Chihao Zhang. FPTAS for counting weighted edge
covers. In Algorithms - ESA 2014 - 22th Annual European Symposium, Wroclaw, Poland,
September 8-10, 2014. Proceedings, pages 654 -- 665, 2014.
[LV97]
Michael Luby and Eric Vigoda. Approximately counting up to four (extended abstract).
In Proceedings of STOC, pages 682 -- 687, 1997.
[LWZ14] Pinyan Lu, Menghui Wang, and Chihao Zhang. FPTAS for weighted fibonacci gates and
its applications. In ICALP (1), pages 787 -- 799, 2014.
11
[LY13]
Pinyan Lu and Yitong Yin. Improved FPTAS for multi-spin systems. In Approximation,
Randomization, and Combinatorial Optimization. Algorithms and Techniques - 16th In-
ternational Workshop, APPROX 2013, and 17th International Workshop, RANDOM
2013, Berkeley, CA, USA, August 21-23, 2013. Proceedings, pages 639 -- 654, 2013.
[RST+11] Ricardo Restrepo, Jinwoo Shin, Prasad Tetali, Eric Vigoda, and Linji Yang. Improved
mixing condition on the grid for counting and sampling independent sets. In Proceedings
of FOCS, pages 140 -- 149, 2011.
[SS12]
Allan Sly and Nike Sun. The computational hardness of counting in two-spin models on
d-regular graphs. In Proceedings of FOCS, pages 361 -- 369, 2012.
[SST12] Alistair Sinclair, Piyush Srivastava, and Marc Thurley. Approximation algorithms for
two-state anti-ferromagnetic spin systems on bounded degree graphs. In Proceedings of
SODA, pages 941 -- 953, 2012.
[Vig99]
Eric Vigoda. Improved bounds for sampling coloring. In Proceedings of FOCS, pages
51 -- 59, 1999.
[Wei06]
Dror Weitz. Counting independent sets up to the tree threshold.
STOC, pages 140 -- 149, 2006.
In Proceedings of
[YZ15]
Yitong Yin and Jinman Zhao. Counting hypergraph matchings up to uniqueness thresh-
old. arXiv preprint arXiv:1503.05812, 2015.
A Proof of Theorem 1
Proof. The input of the FPTAS is an instance G(V,E) and an accuracy parameter 0 < ε < 1/2.
Assume V = {v1, . . . , vn}. Note that I = ∅ is an independent set of G with w(I) = 1. Therefore
Z(G) = 1/PrG [I] = PrG" n
^i=1
vi 6∈ I#!−1
For every 1 ≤ i ≤ n, we define a graph Gi(Vi, Ei):
n
=
Yi=1
vi 6∈ I (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
PrG
−1
.
i−1
^j=1
vj 6∈ I
Gi−1 incident to vi.
• G1 , G;
• For every i ≥ 2, Gi , Gi−1 − vi−1 − E′ where E′ , {e ∈ Ei−1 vi−1 ∈ e} consists of edges in
j=1 vj 6∈ Ii = PrGi [vi 6∈ I] for every 1 ≤ i ≤ n.
It is straightforward to verify that PrGhvi 6∈ I (cid:12)(cid:12)(cid:12)Vi−1
Thus,
n
Z(G) =
(PrGi [vi 6∈ I])−1 =
(1 + Ri) ,
Yi=1
n
Yi=1
where Ri , PrGi [vi∈I]
L =
log(2Cnε−1)
log α−1
PrGi [vi6∈I] . Let C and α be constants in Lemma 6. We compute R(Gi, vi, L) with
for every 1 ≤ i ≤ n, then
Ri − R(Gi, vi, L) ≤
ε
2n
.
12
This implies
ε
2n ≤
1 −
1 + Ri
1 + R(G, vi, L) ≤ 1 +
ε
2n
.
i=1 (1 + R(Gi, vi, L))−1 be our estimate of the partition function, then it holds that
Let Z =Qn
e−ε ≤
Z(G)
Z ≤ eε.
It remains to bound the running time of our algorithm. Let T (L) denote the maximum running
time of computing R(G, v, L) (over all choices of d ≤ ∆ and arbitrary wi). Then by the definition
of R(G, v, L), for every L > 0,
T (L) ≤
d
wi
Xi=1
Xj=1
T (L − ⌊1 + c log1/α wi⌋) + O(n).
ε(cid:1)O(log ∆) for our choice of L. Thus our algorithm is an
It is easy to verify that T (L) = n∆O(L) =(cid:0) n
FPTAS for computing Z(G).
B Proof of Lemma 4
Proof. By the definition of Rv, we have
Rv =
PrG [v ∈ I]
PrG [v 6∈ I]
= λ ·
PrG′hVd
PrG′hVd
i=1 vi ∈ Ii
i=1 vi 6∈ Ii
= λ ·
d
Yi=1
PrG′hvi ∈ I ∧Vi−1
PrG′hvi 6∈ I ∧Vi−1
j=1 vj 6∈ I ∧Vd
j=1 vj 6∈ I ∧Vd
j=i+1 vj ∈ Ii
j=i+1 vj ∈ Ii
For every i ∈ [d], define Gi , G′ − {vk i < k ≤ d} − {ek 1 ≤ k < i}, we have
PrG′hvi ∈ I ∧Vi−1
j=1 vj 6∈ I ∧Vd
PrG′hvi 6∈ I ∧Vi−1
j=1 vj 6∈ I ∧Vd
This is because fixing vj ∈ I is equivalent to removing vj from the graph and fixing vj 6∈ I is
equivalent to removing all edges incident to vj from the graph.
PrG′hvi ∈ I (cid:12)(cid:12)(cid:12)Vi−1
PrG′hvi 6∈ I (cid:12)(cid:12)(cid:12)Vi−1
j=1 vi 6∈ I ∧Vd
j=1 vj 6∈ I ∧Vd
j=i+1 vj ∈ Ii
j=i+1 vj ∈ Ii
j=i+1 vj ∈ Ii
j=i+1 vj ∈ Ii
=
=
PrGi [vi ∈ I]
PrGi [vi 6∈ I]
.
Since ei is the unique hyperedge in Gi that contains vi, we have
PrGi [vi ∈ I]
PrGi [vi 6∈ I]
wi
= 1 − PrGi−vi−ei
^j=1
vij ∈ I
= 1 −
wi
Yj=1
PrGij [vij ∈ I] = 1 −
Rij
1 + Rij
.
wi
Yj=1
C Proof of Lemma 9
Proof. Let λ′c
,
dd
(d−1)d+1 be the uniqueness threshold for the d-ary tree. Then λ < λc ≤ λ′c.
Plugging w = 1 into (3), we have
gz(1) =s λQd
1 + λQd
i=1 zi
i=1 zi ·
13
√1 − zi.
d
Xi=1
Let z =(cid:16)Qd
i=1 zi(cid:17)
1
d , it follows from Jensen's inequality that
1 + λzd < ds λ′czd(1 − z)
1 + λ′czd
(5)
g(t, 1) ≤ ds λzd(1 − z)
1+x(cid:17)d
1 + λ′czd (cid:19)′
(cid:18) λ′czd(1 − z)
λ′czd−1
= −
Recall that fλ,d(x) = λ(cid:16) 1
that dq λ′
czd(1−z)
1+λ′
is
. Let x be the positive fixed-point of fλ′
czd achieves its maximum when z = z. The derivative of λ′
c,d(x) and z = 1
1+x . We show
czd with respect to z
czd(1−z)
1+λ′
(1 + λ′czd)2 (cid:16)z + λ′czd+1 − d(1 − z)(cid:17) .
dd
Since λ′c =
verify that fλ′
(d−1)d+1 , the above achieves maximum at z = d−1
z , then it is easy to
c,d(x) = x, which implies z = z because of the uniqueness of the positive fixed-point.
d . If we let x = 1−z
Therefore, we have for some α < 1,
gz(1) ≤ α < ds λ′c zd(1 − z)
1 + λ′c zd =(cid:12)(cid:12)(cid:12)
f′λ′
c,d(x)(cid:12)(cid:12)(cid:12)
= 1.
D FPTAS for Hypergraph Ising Model
D.1 The algorithm and recursion for computing marginal probability
In this section, we give the recursion function and design an algorithm to compute marginal prob-
ability in hypergraph Ising Model. Then we prove that the algorithm is indeed an FPTAS for any
instance G(V,E, (β, γ)) if 1 −
2e−1/2∆+3 ≤ β(e), γ(e) ≤ 1.
2
Graph Operations We first fix some notations on graph modification specific to our hypergraph
two state spin model. Let G(V,E, (β, γ)) be an instance. The first groups of operations are about
vertex removal and edge removal, which is similar to the case of hardcore model:
• For every v ∈ V , we denote G − v , (V \ {v} ,E′, (β′, γ′)) where E′ , {e − v e ∈ E} and
β′(e − v) = β(e), γ′(e − v) = γ(e) for every e ∈ E..
• For every e ∈ E, we denote G − e , (V,E \ {e} , (β, γ)).
• Let x be a vertex or an edge and y be a vertex or an edge, we denote G− x− y , (G− x)− y.
• Let S = {v1, . . . , vk} ⊆ V , we denote G − S , G − v1 − v2 ··· − vk.
• Let F = {e1, . . . , ek} ⊆ E, we denote G − F , G − e1 − e2 ··· − ek.
The second group of operations is about pinning the value of a vertex, whose effect is to change
β(e) or γ(e) for edge e incident to it.
14
• Let v ∈ V be a vertex, we denote Gv=0 , (V \{v} ,E′, (β′, γ′)e∈E ′) where E′ , {e − v e ∈ E},
β′(e − v) , β(e) for every e ∈ E and
γ′(e − v) ,(γ(e)
1
if v 6∈ e
otherwise.
This operation is to pin the value of v to 0.
• Similarly, for a vertex v ∈ V , we denote Gv=1 , (V \{v} ,E′, (β′, γ′)) where E′ , {e \ {v} e ∈ E},
γ′(e − v) , γ(e) for every e ∈ E and
β′(e − v) ,(β(e)
1
if v 6∈ e
otherwise.
This operation is to pin the value of v to 1.
• Let u, v ∈ V be two vertices and i, j ∈ {0, 1}. We denote Gu=i,v=j , (Gu=i) v=j and this
notation generalizes to more vertices.
• Let S = {v1, . . . , vk} ⊆ V , we use GS=0 (resp. GS=1) to denote Gv1=0,v2=0,...,vk=0 and
Gv1=1,v2=1,...,vk=1.
Let G(V,E, (β, γ)) be an instance and v ∈ V be an arbitrary vertex with degree d. Let E(v) =
{e1, . . . , ed} be the set of edges incident to v. For every i ∈ [d], we assume ei = {v}∪{vij j ∈ [wi]}
consists of wi + 1 vertices where wi ≥ 0.
We define a new instance G′(V ′,E′, (β′e, γ′e)) which is obtained from G by replacing v by d
copies of itself and each ei contains a distinct copy. Formally, V ′ , (V \ {v}) ∪ {v1, . . . , vd} and
E′ , (E \ E(v)) ∪ {ei − v + vi i ∈ [d]}; β′(e) = β(e), γ′(e) = γ(e) for every e ∈ E \ E(v) and
β′(ei − v + vi) = β(ei), γ′(ei − v + vi) = γ(ei) for every i ∈ [d].
ij)(cid:17) and
For every i ∈ [d] and j ∈ [wi], we define two smaller instances G0
ij(cid:16)V 1
ij,E 1
• G0
• G1
, (((G′V<i=0)V>i=1) − ei)V i
, (((G′V<i=0)V>i=1) − ei)V i
ij(cid:16)V 0
ij)(cid:17):
ij, (β1
ij, γ1
ij, (β0
ij , γ0
ij,E 0
<j =1,
<j =0;
G1
ij
ij
where V<i , {vk k < i}, V>i , {vk k > i}, V i
<j
, {vik k < j}.
We now define the ratio of the probability on instances defined above.
• Let Rv , PrG[σ(v)=1]
PrG[σ(v)=0] ;
• For every i ∈ [d] and j ∈ [wi], let R0
ij
,
Pr
Pr
G0
ij
G0
ij
[σ(vij )=1]
[σ(vij )=0] , R1
ij
,
Pr
Pr
G1
ij
G1
ij
[σ(vij )=1]
[σ(vij )=0] .
It follows from our definition that for every i ∈ [d], R0
We can compute Rv by the following recursion:
i1 = R1
i1.
Lemma 10.
Rv = λ ·
d
Yi=1
1 − (1 − γ(ei)) R0
1 − (1 − β(ei))
i1
1+R0
1+R0
1
i1 Qwi
i1 Qwi
j=2
j=2
15
R1
ij
1+R1
ij
1
1+R0
ij
.
(6)
Proof. By the definition of Rv, we have
Rv =
PrG [σ(v) = 1]
PrG [σ(v) = 0]
= λ·
PrG′hVd
PrG′hVd
i=1 σ(vi) = 1i
i=1 σ(vi) = 0i
d
Yi=1
= λ·
For every i ∈ [d], define Gi , (G′V<i=0)V>i=1, then
j=i+1 σ(vj) = 1i
PrG′hσ(vi) = 1 ∧Vi−1
j=i+1 σ(vj) = 1i
PrG′hσ(vi) = 0 ∧Vi−1
j=1 σ(vj) = 0 ∧Vd
j=1 σ(vj) = 0 ∧Vd
PrG′hσ(vi) = 1 ∧Vi−1
PrG′hσ(vi) = 0 ∧Vi−1
PrG′hσ(vi) = 1(cid:12)(cid:12)(cid:12)Vi−1
PrG′hσ(vi) = 0(cid:12)(cid:12)(cid:12)Vi−1
PrGi [σ(vi) = 1]
PrGi [σ(vi) = 0]
.
=
=
j=1 σ(vj) = 0 ∧Vd
j=1 σ(vj) = 0 ∧Vd
j=i+1 σ(vj) = 1i
j=i+1 σ(vj) = 1i
.
j=1 σ(vj) = 0 ∧Vd
j=1 σ(vj) = 0 ∧Vd
j=i+1 σ(vj ) = 1i
j=i+1 σ(vj ) = 1i
Since ei is the unique edge in Gi that contains vi, we have
PrGi [σ(vi) = 1]
PrGi [σ(vi) = 0]
=
=
=
=
j=1 σ(vij) = 1i(cid:17)
j=1 σ(vij ) = 0i(cid:17)
j=2 PrGi−vi−eihσ(vij) = 1(cid:12)(cid:12)(cid:12)Vj−1
j=2 PrGi−vi−eihσ(vij) = 0(cid:12)(cid:12)(cid:12)Vj−1
[σ(vij) = 1]
ij
[σ(vij) = 0]
j=1 σ(vij) = 1i +(cid:16)1 − PrGi−vi−eihVwi
j=1 σ(vij) = 0i +(cid:16)1 − PrGi−vi−eihVwi
γ(ei)PrGi−vi−eihVwi
β(ei)PrGi−vi−eihVwi
1 − (1 − γ(ei))PrGi−vi−ei [σ(vi1) = 1] ·Qwi
1 − (1 − β(ei))PrGi−vi−ei [σ(vi1) = 0] ·Qwi
1 − (1 − γ(ei))PrG1
1 − (1 − β(ei))PrG0
1 − (1 − γ(ei)) R0
1 − (1 − β(ei))
[σ(vi1) = 1] ·Qwi
[σ(vi1) = 0] ·Qwi
i1 Qwi
i1 Qwi
R1
ij
1+R1
ij
i1
1+R0
1+R0
ij
1+R0
j=2
j=2
i1
i1
1
1
.
j=2 PrG1
j=2 PrG0
ij
k=1 σ(vk) = 1i
k=1 σ(vk) = 0i
The last equality is due to the fact that R0
i1 = R1
i1.
The algorithm description is similar to Section 3.2. Let G(V,E, (β, γ)) be an instance of our
two state spin model with maximum degree ∆, and v ∈ V be an arbitrary vertex with degree d.
Let E(v) = {e1, . . . , ed} denote the set of edges incident to v. Define G0
ij as in
Section D. Then the recursion (6) gives a way to compute the marginal probability PrG [σ(v) = 1]
exactly. However, an exact evaluation of the recursion requires a computation tree with exponential
size. Thus we introduce the following truncated version of the recursion, with respect to constants
c > 0 and 0 < α < 1. Let
ij, Rv, R0
ij, G1
ij, R1
fG(r) = λ ·
d
Yi=1
1 − (1 − γ(ei)) r0
1 − (1 − β(ei))
1
i1
1+r0
1+r0
r1
ij
1+r1
ij
1
1+r0
ij
,
j=2
j=2
i1 Qwi
i1 Qwi
where r , ((r0
ij)1≤j≤wi, (r1
ij )2≤j≤wi)i∈[d] with
r0
ij = R(G0
r1
ij = R(G1
ij , vij, L − ⌊1 + c log1/α wi⌋)
ij , vij, L − ⌊1 + c log1/α wi⌋).
16
We can describe our truncated recursion as follows:
R(G, v, L) =(fG(r)
λ
if L > 0
otherwise.
The recursion can be directly used to compute R(G, v, L) for any given L and it induces a
truncated computation tree (with height L in some special metric).
We claim that R(G, v, L) is a good estimate of Rv for a suitable choice of c and α, for those
2
2e−1/2∆+3 ≤ β(e), γ(e) ≤ 1 for all e ∈ E.
(β, γ, ∆) satisfying that 1 −
Lemma 11. Let G(V,E, (β, γ)) be an instance of our generalized two state spin system model with
2e−1/2∆+3 ≤ β(e), γ(e) ≤ 1 for all
maximum degree ∆ ≥ 2. Let v ∈ V be a vertex and assume 1 −
e ∈ E. There exist constants C > 0 and 0 < α < 1 such that
2
for every L.
R(G, v, L) − Rv ≤ C · αmax{0,L}
The Lemma 11 will be proved in the next section. The proof of Theorem 3 with this lemma is
almost identical to the proof of Theorem 1 and we omit it here.
D.2 Correlation decay
Again, the key is to bound the amortized decay rate as in the following lemma.
Lemma 12. Let ∆ ≥ 2 be a constant integer. Suppose d ≤ ∆ is a constant integer, λ ∈ (0, 1), βc =
are constant real numbers and γ1, γ2, . . . γd, β1, β2, . . . , βd ∈ [βc, 1]. Let w1, . . . , wd > 0
1−
be integers and
2e−1/2∆+3
2
d
i1
1+r0
1 − (1 − γi) r0
1 − (1 − βi)
1
1+r0
r1
ij
1+r1
ij
1
1+r0
ij
j=2
j=2
i1 Qwi
i1 Qwi
x . There exist constants α < 1 and c > 0 depending on
Φ(f )
Φ(r1
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
].
≤ α < 1
∂f (r)
∂r1
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
c , λβ−∆
c
for all 1 ≤ i ≤ d and 2 ≤ j ≤ wi. The it holds
d
f (r) = λ
be a (cid:16)Pd
γ and ∆ (but not depending on wi) such that
Yi=1
i=1 wi(cid:17)-ary function. Let Φ(x) = 1
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=2
for every r = (rij)i∈[d],j∈[wi] where each rij ∈ [λβ∆
, zij , r0
Proof. We first let xi , r0
ij
i1
1+r0
1+r0
i1
ij
that xi, yij, zij ∈h λβ∆
λ
i
∂f (r)
∂r0
Φ(f )
Φ(r0
Xj=1
Xi=1
1+λβ∆
c
λ+β∆
wc
+
wi
c
,
wi
, yij , r1
ij
1+r1
ij
c i. To simplify the notation, we denote
Ai , 1 − (1 − γi)xi
yij
wi
Yj=2
Bi , 1 − (1 − βi)(1 − xi)
wi
Yj=2
(1 − zij)
17
where Ai ∈ [γi, 1] ⊆ [βc, 1], Bi ∈ [βi, 1] ⊆ [βc, 1].
Then we can write f as
d
f = λ
Ai
Bi
.
Yi=1
We can directly compute the partial derivatives of f , which yields the following for all i ∈ [d] and
2 ≤ j ≤ wi:
(1 − xi) ((1 − Ai)(1 − Bi) − (1 − xi)(1 − Ai) − xi(1 − Bi))
xiB2
i
Ak
Bk
xiAiBi
=
λ Yk∈[d]
k6=i
(1 − xi) ((1 − Ai)(1 − Bi) − (1 − xi)(1 − Ai) − xi(1 − Bi))
= f ·
=
λ Yk∈[d]
=
λ Yk∈[d]
= f · −(1 − yij)2(1 − Ai)
= f · −(1 − Bi)(1 − zij)
· −(1 − yij)2(1 − Ai)
· −Ai(1 − Bi)(1 − zij)
Ak
Bk
Ak
Bk
yijBi
yijAi
B2
i
k6=i
k6=i
Bi
;
.
;
wi
wi
+
Φ(f )
Φ(r1
Φ(f )
Φ(r0
∂f (r)
∂r0
∂f (r)
∂r1
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 − xi (cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
i
−(1 − Ai)(1 − Bi) + (1 − xi)(1 − Ai) + xi(1 − Bi)
Xj=2
Xj=2
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1 − yi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂f (r)
∂r0
∂f (r)
∂r1
Xj=2
xi
yi
+
+
wi
wi
zi
1 − zi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
AiBi
j=2 yij(cid:17)1/(wi−1)
j=2 1 − zij(cid:17)1/(wi−1)
, zi , 1 −(cid:16)Qwi
Ai = 1 − (1 − γi)xiywi−1
Bi = 1 − (1 − βi)(1 − xi)(1 − zi)wi−1,
,
i
∂f (r)
∂r0
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=2
wi
wi
1 − Ai
1 − Bi
+
Ai
(1 − yij) +
Xj=2
for every i ∈ [d], then it holds that
Bi
zij
·
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
i1 (cid:12)(cid:12)(cid:12)(cid:12)
∂f
∂r0
i1
∂f
∂r1
ij
∂f
∂r0
ij
Thus,
d
d
1
=
wc
wc
f (r)
Xi=1
i
Xi=1
i
Xi=1
Let yi ,(cid:16)Qwi
wc
=
d
wi
Xj=2
(1 − yij) ≤ (wi − 1)(1 − yi)
wi
Xj=2
zij ≤ (wi − 1)zi.
18
d
wi
wi
+
wc
wc
≤
AiBi
1 − Ai
−(1 − Ai)(1 − Bi) + (1 − xi)(1 − Ai) + xi(1 − Bi)
Since Ai, Bi ∈ [βc, 1] for every i ∈ [d], we have
Xi=1
Xi=1
Xi=1
1 − βc
β2
c
i
i (cid:18) (1 − xi)(1 − Ai) + xi(1 − Bi) − (1 − Ai)(1 − Bi)
i (1 − xi)(1 − Ai) + xi(1 − Bi)
Xi=1
i (cid:16)(1 − xi)xi(ywi−1
+ (1 − βc)(wi − 1) xiywi−1
+ (1 − zi)wi−1) + βc(wi − 1)(xiywi−1
(1 − yij) +
+ (wi − 1)(cid:18) (1 − Ai)(1 − yi)
zij
Xj=2
(cid:19)(cid:19)
(1 − Bi)zi
!!
(1 − yi) + (1 − xi)(1 − zi)wi−1zi)(cid:17)
Ai
(1 − xi)(1 − zi)wi−1zi
(1 − yi)
Ai
1 − Bi
Xj=2
AiBi
AiBi
≤
≤
wc
wc
Bi
Bi
Bi
Ai
+
+
d
d
i
i
d
i
We now assume Lemma 13, thus there exist constants α < 1, c > 0 depending on λ and ∆ such
that
+ (1 − zi)wi−1) + βc(wi − 1)(xiywi−1
i
(1 − yi) + (1 − xi)(1 − zi)wi−1zi)(cid:17) ≤ αβce
1
2βc −1
−(1 − Ai)(1 − Bi) + (1 − xi)(1 − Ai) + xi(1 − Bi)
AiBi
+
1 − Ai
Ai
wi
Xj=2
(1 − yij) +
1 − Bi
Bi
wi
Xj=2
zij
i
wc
i (cid:16)(1 − xi)xi(ywi−1
for every i ∈ [d]. Thus,
i
Xi=1
1 − βc
· d · α · eβ−1
≤
wc
d
βc
c /2−1
2d
2αd
2e−1/2d + 3 ·
2e−1/2d + 3
2e−1/2d + 1 · e(1−2/(2e−1/2d+3))−1/2−1
2e−1/2d + 1 · e(1−2/(2e−1/2d+3))−1/2−1
≤
= α ·
≤ α.
The last inequality is due to the fact that h(d) ,
function on d and limd→∞ h(d) = 1.
and every r = (rij)i∈[d],j∈[wi] with each rij ∈ [λβ∆
Xj=2
∂f (r)
∂r0
Φ(f )
Φ(r0
Xi=1
Xj=1
wc
+
wi
wi
d
i
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
c , λβ−∆
c
], it holds
Φ(f )
Φ(r1
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂f (r)
∂r1
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ α < 1
2e−1/2d+1 · e(1−2/(2e−1/2d+3))−1/2−1 is an increasing
Combining all above, there exist constants α < 1 and c > 0 such that for all w1, w2, . . . wd > 0
2d
It remains to prove Lemma 13:
Lemma 13. For all β < 1 and 0 < δ < 1/2, there exist constants α < 1 and c > 0 such that
wc(cid:0)(1 − x)x(yw−1 + (1 − z)w−1) + β(w − 1)xyw−1(1 − y) + β(w − 1)(1 − x)(1 − z)w−1z(cid:1) ≤ αβe
for all w ∈ N and all x, y, z ∈ [δ, 1 − δ].
1
2β −1
19
Proof. For a fixed w, let
Vw(x, y, z) , (1 − x)x(yw−1 + (1 − z)w−1) + β(w − 1)xyw−1(1 − y) + β(w − 1)(1 − x)(1 − z)w−1z.
Then Vw(x, y, z) achieves maximum value when
y = yc , β(w − 1) + 1 − x
βw
= 1 −
1 − (1 − x)/β
w
,
z = zc , 1 −
β(w − 1) + x
βw
=
1 − x/β
w
.
w
, Vw(x, y, z) is an increasing function on y.
If y < 1 − 1−(1−x)/β
decreasing function on z.
Let w , maxn(cid:6) 1
ln(1−δ)mo + 1. If w > w, the monotonicity of Vw(x, y, z) implies that the
function achieves it maximum when the y and z are at the boundary, i.e., y = 1 − z = 1 − δ, for
every fixed x.
e1/(2β)−1 !1/2
We now analyze the case w ≤ w and the case w > w separately. Define α , (cid:16)1+ 1/(2β)−1
δ(cid:7) ,l− 1
w , Vw(x, y, z) is a
If z > 1−x/β
(cid:17) w
.
w
It is easy to verify that α < 1 since (1 + x/n)n < ex for every x > 0.
• (If w ≤ w.) We plug y = yc and z = zc into Vw(x, y, z):
c +(1−zc)w−1)+β(w−1)xyw−1
Vw(x, y, z) ≤ (1−x)x(yw−1
The derivative of the right hand side with respect to x is
(β(w − 1) + 1 − x)w−1(1 + β(w − 1) − (w + 1)x) − (β(w − 1) + x)w−1(1 + β(w − 1) − (w + 1)(1 − x))
(1−yc)+β(w−1)(1−x)(1−zc)w−1zc.
c
(βw)w
.
It is easy to see that the function above is zero when x = 1/2 and positive for smaller x,
negative for larger x.
Thus, when x = 1/2, y = yc, z = zc, Vw(x, y, z) achieves its maximum value
Vw 1
2
, 1 −
1 − 1
w
Let c1 , log w
1
2β
αe
1− 1
2β
1−
,
−1
2β
2β
2β
2β
≤ β 1 −
w !w
1 − 1
1 − 1
w ! = β 1 −
w ! w
1 − 1
w ! w = − log w α > 0, then for all w ≤ w and c′ ∈ (0, c1],
w ! w
1 − 1
w ! w
1 − 1
α 1 −
=
2β
2β
β
wc′
Vw(x, y, z) ≤ wc1Vw(x, y, z) ≤ wc1β 1 −
< βe
1
2β −1.
= αβe
1
2β −1.
• (If w > w.) Let c2 , − w log(1 − δ) − 1 > 0. Then
Vw(x, y, z) ≤ Vw(x, 1 − δ, δ) = 2x(1 − x)(1 − δ)w−1 + β(w − 1)δ(1 − δ)w−1
The above achieves its maximum when x = 1/2, i.e., Vw(x, y, z) ≤ Vw(cid:0) 1
1 + 2β(w − 1)δ
wc2 · Vw(x, y, z) ≤ wc2 · Vw(cid:18) 1
2
, 1 − δ, δ(cid:19) = wc2 ·
2
2 , 1 − δ, δ(cid:1). Therefore
(1 − δ)w−1.
20
We now prove that for all c′ ∈ (0, c2], g(w) , wc′ 1+2β(w−1)δ
when w > w. Since c′ ≤ c2 = − w log(1 − δ) − 1 ≤ −(w − 1) log(1 − δ) − 1 for all w > w,
∂g
∂w
(1 − δ)w−1 is decreasing on w
=
2
2
(cid:0)c′ + w log(1 − δ)(cid:1) + βδwc′−1(1 − δ)w−1(cid:0)w + (w − 1)(c′ + w log(1 − δ))(cid:1)
(log(1 − δ) − 1) + βδwc′−1(1 − δ)w−1 (w + (w − 1)(log(1 − δ) − 1))
+ βδwc′−1(1 − δ)w−1(1 + (w − 1) log(1 − δ))
− βδwc′−1(1 − δ)w−1c′)
wc′−1(1 − δ)w−1
wc′−1(1 − δ)w−1
wc′−1(1 − δ)w−1
wc′−1(1 − δ)w−1
2
2
2
≤
≤ −
≤ −
< 0
Thus, g(w) is decreasing on w when w > w. In light of this, wc′ · Vw(x, y, z) can be upper
bounded by wc′
In all, we have
· V w(cid:0) 1
wc′
2 , 1 − δ, δ(cid:1) for all c′ ∈ (0, c2].
· Vw(cid:18) 1
· Vw(x, y, z) ≤ wc′
· V w(cid:18) 1
≤ wc′
≤ wc1 · V w 1
, 1 − δ, δ(cid:19)
, 1 − δ, δ(cid:19)
1 − 1
, 1 −
w
2β −1 wc′−c1
2
2
2
1
2β
,
2β
1 − 1
w ! wc′−c1
= αβe
for all w > w and c′ ∈ (0, c2].
Let c = min{c1, c2}, then we have wcVw(x, y, z) ≤ wc1Vw(x, y, z) ≤ αβe
wcVw(x, y, z) ≤ wc1V w(x, y, z) wc′−c1 ≤ αβe
wc(cid:0)(1 − x)x(yw−1 + (1 − z)w−1) + β(w − 1)xyw−1(1 − y) + β(w − 1)(1 − x)(1 − z)w−1z(cid:1) ≤ αβe
for all w ∈ N and all x, y, z ∈ [δ, 1 − δ].
2β −1 if w > w, i.e.,
2β −1 if w ≤ w and
1
1
2β −1
1
We are now ready to prove Lemma 11.
Proof of Lemma 11. Let Φ(x) = 1
If ℓ = 0, note that fv ∈ [λβ∆
x and φ(x) = R Φ(x) dx = ln x. We apply induction on ℓ ,
max{0, L} to show that, if d ≤ ∆, then φ(fv) − φ(f (G, v, L)) ≤ 4√eαL for some constant α < 1.
], thus φ(f (G, v, L)) − φ(fv) ≤ −2∆ ln βc. Since ln(1 − x) ≤
2x for all x ∈(cid:0)0, 1
2(cid:1),
φ(f (G, v, L)) − φ(fv) ≤ −2∆ ln βc ≤ 4∆
c , λβ−∆
2√e ∆ + 3 ≤ 4√e.
2
c
We now assume L = ℓ > 0 and the lemma holds for smaller ℓ. For every i ∈ [d] and j ∈ [wi],
ij, vij, L − ⌊1 +
ij)2≤j≤wi)i∈[d]. Let f (r) =
ij = R(G0
ij)2≤j≤wi)i∈[d], r = ((r0
we denote r0
c log1/α wi⌋). Let r = ((r0
ij , vij, L − ⌊1 + c log1/α wi⌋), r1
ij and r0
ij)1≤j≤wi, (r1
ij)1≤j≤wi, (r1
ij = R(G1
ij = R1
ij = R0
ij, r1
21
i=1
1−(1−γi)
1−(1−βi)
where each r0
λQd
r0
i1
1+r0
1
j=2
i1 Qwi
i1 Qwi
ij ∈ [λβ∆
j=2
1+r0
ij, r1
r1
ij
1+r1
ij
1
1+r0
ij
c , λβ−∆
c
],
φ(Rv) − φ(R(G, v, L)) ≤
d
wi
Xi=1
Xj=1
≤ 4√e
(♠)
Xi=1
≤ 4√eαL.
(♥)
d
, then it follows from Proposition 7 that for some r = ((r0
ij)1≤j≤wi, (r1
ij)2≤j≤wi)i∈[d]
ij) − φ(r0
Φ(f )
Φ(r0
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=1
wi
∂f (r)
∂r0
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
·(cid:12)(cid:12)φ(r0
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
∂f (r)
∂r0
Φ(f )
Φ(r0
+
wi
Xj=2
ij ) − φ(r1
ij)(cid:12)(cid:12)
wi
Φ(f )
Φ(r1
∂f (r)
∂r1
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xj=2
·(cid:12)(cid:12)φ(r1
ij (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
αL−⌊1+c log1/α wi⌋
∂f (r)
∂r1
ij)(cid:12)(cid:12) +
ij)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Φ(f )
Φ(r1
(♠) follows from the induction hypothesis and (♥) is due to Lemma 12 for all d ≤ ∆.
Then the lemma follows from Proposition 7, since
Rv − R(G, v, L) =
for some x ∈ [λβ∆
c , λβ−∆
c
]
1
Φ(x) · φ(Rv) − φ(R(G, v, L))
≤ e√eλ−1 · 4√eαL
2 +√eαL.
= 4λ−1e
c =(cid:16)1 −
2√e(cid:17)−∆
√e
∆+ 3
1
≤ e√e.
The inequality above is due to β−∆
22
|
1704.04362 | 4 | 1704 | 2018-09-03T18:33:35 | Fast Randomized Algorithms for t-Product Based Tensor Operations and Decompositions with Applications to Imaging Data | [
"cs.DS"
] | Tensors of order three or higher have found applications in diverse fields, including image and signal processing, data mining, biomedical engineering and link analysis, to name a few. In many applications that involve for example time series or other ordered data, the corresponding tensor has a distinguishing orientation that exhibits a low tubal structure. This has motivated the introduction of the tubal rank and the corresponding tubal singular value decomposition in the literature. In this work, we develop randomized algorithms for many common tensor operations, including tensor low-rank approximation and decomposition, together with tensor multiplication. The proposed tubal focused algorithms employ a small number of lateral and/or horizontal slices of the underlying 3-rd order tensor, that come with {\em relative error guarantees} for the quality of the obtained solutions. The performance of the proposed algorithms is illustrated on diverse imaging applications, including mass spectrometry data and image and video recovery from incomplete and noisy data. The results show both good computational speed-up vis-a-vis conventional completion algorithms and good accuracy. | cs.DS | cs | FAST RANDOMIZED ALGORITHMS FOR T-PRODUCT BASED TENSOR
OPERATIONS AND DECOMPOSITIONS WITH APPLICATIONS TO
IMAGING DATA∗
DAVOUD ATAEE TARZANAGH† AND GEORGE MICHAILIDIS‡
Abstract. Tensors of order three or higher have found applications in diverse fields, including image and sig-
nal processing, data mining, biomedical engineering and link analysis, to name a few. In many applications that
involve for example time series or other ordered data, the corresponding tensor has a distinguishing orientation
that exhibits a low tubal structure. This has motivated the introduction of the tubal rank and the corresponding
tubal singular value decomposition in the literature. In this work, we develop randomized algorithms for many
common tensor operations, including tensor low-rank approximation and decomposition, together with tensor
multiplication. The proposed tubal focused algorithms employ a small number of lateral and/or horizontal slices
of the underlying 3-rd order tensor, that come with relative error guarantees for the quality of the obtained
solutions. The performance of the proposed algorithms is illustrated on diverse imaging applications, including
mass spectrometry data and image and video recovery from incomplete and noisy data. The results show both
good computational speed-up vis-a-vis conventional completion algorithms and good accuracy.
Key words. tensor slices selection, circulant algebra, low rank decomposition, nuclear norm minimization
AMS subject classifications. 15A69, 15A23, 68W20, 65F30
1. Introduction. Tensors are multi-dimensional arrays that have been used in diverse
fields of applications, including chemometrics [45], psychometrics [30], image/video and signal
processing [24, 25, 35, 44, 54] and link analysis [29]. They have also been the object of intense
mathematical study (see for example the review paper by Kolda and Bader [28] and references
therein).
Analogously to the matrix case, a number of tensor decompositions have been proposed in
the literature, briefly described next for the case of 3-way tensors. Let Z ∈ Rn1×n2×n3. Then,
there exists a factorization, called a "Tucker decomposition"of Z, of the form
R1(cid:88)
R2(cid:88)
R3(cid:88)
(cid:16)
(cid:17)
(1)
Z =
gr1r2r3
u(1)
r1
◦ u(2)
r2
◦ u(3)
r3
= G ×1 U (1) ×2 U (2) ×3 U (3),
r1=1
r2=1
r3=1
where G ∈ RR1×R2×R3 is the core tensor associated with this decomposition, U (i) ∈ Rni×Ri is
the i-th factor matrix and ×i is the mode−i product for i = 1, 2, 3. The 3-tuple (R1, R2, R3)
is called the Tucker rank or multilinear rank of tensor Z [8, 22]. The conventional Tucker
decomposition corresponds to the orthonormal Tucker decomposition, which is also known as
the higher-order SVD (HOSVD). De Lathauwer et al. [10] proposed an algorithm to compute a
HOSVD decomposition. Soon afterwards they proposed the higher-order orthogonal iteration
(HOOI) [11] to provide an inexact Tucker decomposition. The CP decomposition of a tensor
is another important notion of tensor-decomposition, which leads to the definition of CP rank.
The CP model can be considered as a special case of the Tucker model with a superdiagonal
core tensor. Further, the CP rank of a tensor equals that of its Tucker core [21].
1.1. Third order tensor as operator on matrices. While the Tucker-based factoriza-
tion may be sufficient for many applications, in this paper we consider an entirely different tensor
decomposition based on circulant algebra [26]. In this factorization, a tensor in Rn1×n2×n3 is
viewed as a n1 × n2 matrix of "tubes", also known as elements of the ring Rn3 where addition
∗
R01 GM114029-03 to GM.
Funding: This work was supported in part by NSF grants CCF 1540093, IIS 1632730, and NIH grant 5
†Department of Mathematics & UF Informatics Institute, University of Florida, Gainesville, FL,
‡Department of Statistics & UF Informatics
Institute, University of Florida, Gainesville, FL,
([email protected]).
([email protected]).
1
8
1
0
2
p
e
S
3
]
S
D
.
s
c
[
4
v
2
6
3
4
0
.
4
0
7
1
:
v
i
X
r
a
2
D. A. TARZANAGH AND G. MICHAILIDIS
is defined as vector addition and multiplication as circular convolution. This "matrix-of-tubes"
viewpoint leads to definitions of a new multiplication for tensors ("tubal multiplication"), a
new rank for tensors ("tubal rank"), and a new notion of a singular value decomposition (SVD)
for tensors ("tubal SVD"). The tubal SVD (t-SVD) provides the "best" tubal rank-r approxi-
mation to Z, as measured with respect to any unitary invariant tensor norm.
A limitation of the t-SVD decomposition is that it depends directly on the orientation of
the tensor, whereas the CP and Tucker decompositions are not. This suggests that the lat-
ter decompositions are well suited for data applications where the tensor's orientation is not
critical - e.g. chemometrics [45] and/or psychometrics [30]. However, in applications involving
time series or other ordered data, the orientation of the tensor is fixed. Examples include,
but not limited to, computed tomography (CT) [44], facial recognition [18] and video compres-
sion [54], where the tensor decomposition is dependent on the third dimension. Analogous to
compressing a two-dimensional image using the matrix SVD (a classic linear algebra example,
with detailed writeup in [23]), the t-SVD decomposition can be used to compress several im-
ages taken over time (e.g. successive frames from a video). Since such images do not change
substantially from frame to frame, we expect tubal compression strategies to provide better
results than performing a matrix SVD on each separate image frame. The former consider
the tensor as a whole object, rather than as a collection of distinct images [27, 26, 54, 44, 35].
Further, t-SVD is essentially based on a group theoretic approach, where the multidimensional
structure is unraveled by constructing group-rings along the tensor fibers 1. The advantage of
such an approach over existing ones is that the resulting algebra and corresponding analysis
enjoys many similar properties to matrix algebra and analysis. For example, it is shown in [20]
that recovering a 3-way tensor of length n and Tucker rank (r, r, r) from random measurements
requires O(rn2 log2(n)) observations under a matrix coherence condition on all mode-n unfold-
ings. However, the number of samples needed for exact recovery of a 3-way tensor of length
n and tensor tubal-rank r is O(rn2 log(n2)) under a weaker tensor coherence assumption [53].
Further, consider the decomposition X = L + E, where L ∈ Rn1×n2×n3 is low rank and E is
sparse. Let n(1) = max(n1, n2) and n(2) = min(n1, n2). The work in [35] shows that for tensor
L with coherence parameter µ, the recovery is guaranteed with high probability for the tubal
rank of order n(2)n3/(µ(log n(1)n3)2) and a number of nonzero entries in E of order O(n1n2n3).
Hence, under the same coherence condition (see, Definitions 14 and 15), the tubal robust tensor
factorization problem perfectly recovers the low-rank and sparse components of the underlying
tensor.
A shortcoming of these three classical decompositions is their brittleness with respect to
severely corrupted or outlier data entries. To that end, a number of approaches have been
developed in the literature to recover a low-rank tensor representation from data subject to
noise and corrupted entries. We focus on two instances of the problem based on the t-SVD
algorithm: (i) noisy tensor completion, i.e., recovering a low-rank tensor from a small subset
of noisy entries, and (ii) noisy robust tensor factorization, i.e., recovering a low-rank tensor
from corrupted data by noise and/or outliers of arbitrary magnitude [35, 54]. These two classes
of tensor factorization problems have attracted significant interest in the research community
[2, 6, 44, 54]. In particular, convex formulations of noisy tensor factorization have been shown to
exhibit strong theoretical recovery guarantees and a number of algorithms has been developed
for solving them [54, 44, 35].
tages, also exhibits a number of drawbacks listed below:
It is frequently mentioned that (noisy) tensor factorization, despite its numerous advan-
• The available methods [4, 27, 26, 54, 44, 35] are inherently sequential and all rely on the
repeated and costly computation of t-SVD factors, Discrete Fourier Transform (DFT)
and its inverse (IDFT), that limit their scalability.
• The basis tensor vectors resulting from t-SVD have little concrete meaning, which
makes it challenging for practitioners to interpret the obtained results. For instance,
1We consider the group rings constructed out of cyclic groups, resulting in an algebra of circulants. However,
the results presented in this paper hold true for the general group-ring construction.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
3
√
the vector [(1/2) age - (1/
2) height + (1/2)income], being one of the significant
uncorrelated factors from a data set of people's features is not easily interpretable (see
discussion in [15]). Kuruvilla et al. in [31] have also claimed: "it would be interesting
to try to find basis vectors for all experiment vectors, using actual experiment vectors
and not artificial bases that offer little insight".
• The t-SVD decomposition for sparse tensors does not preserve sparsity in general, which
for large size tensors leads to excessive computations and storage requirements. Hence,
it is important to compute low-rank tensor factorizations that preserve such structural
properties of the original data tensor.
1.2. Main Contributions. In this paper, we study scalable randomized tensor multipli-
cation (rt-product algorithm) and tensor factorization (rt-project) operations and extend the
matrix CX and CUR type decompositions [12, 13, 15, 36] to third order tensors using a circulant
algebra embedding. To that end, we develop a basic algorithm (t-CX), together with a more
general one (t-CUR) based on tensor slice selection that come with relative error guarantees.
Finally, we propose a new tensor nuclear norm minimization method, called CUR t-NN, which
solves the noisy tensor factorization problem using a small number of lateral and/or horizontal
slices of the underlying tensor. Specifically, CUR t-NN uses an adaptive technique to sample
slices of the tensor based on a leverage score for them and subsequently solves a convex op-
timization problem followed by a projection step to recover a low rank approximation to the
underlying tensor. Advantages of CUR t-NN include:
• Contrary to nuclear norm minimization based approaches, which minimize the sum of
the singular values of the underling tensor, our approach only minimizes the sum of
singular values of the set of sampled slices corresponding to the largest leverage scores.
Thus, we obtain a more accurate and robust approximation to the rank function.
• Using subspace sampling, we obtain provable relative-error recovery guarantees for
tubal product based tensor factorization. This technique is likely to be useful for other
tensor approximation and data analysis problems.
• The proposed algorithm for noisy tensor factorization tasks has polynomial time com-
plexity.
1.3. Related work on randomized tensor factorization. Note that there has been
prior work on the topic of randomized methods for tensor low rank approximations/decomposi-
tions. For example, recent work includes [37, 5]. Our methods are different from these studies,
since we rely on the t-product construct in [4, 26, 27] to provide tubal tensor multiplication and
low rank approximation. In [37], the proposed tensor-CUR algorithm employs a CUR matrix
approximation to one of the unfolding matrix modes (the distinguished mode) providing an
approximation based on few tube fibers and few slices. Hence, that algorithm achieves an ad-
ditive error guarantee in terms of the flattened (unfolded) tensor, rather than the original one.
However, the algorithms presented in this work offer relative error guarantees for the obtained
approximations, in terms of the original low tubal rank tensor. Further, we propose a new
tensor nuclear norm minimization method, which solves the noisy tensor factorization problem
in the fully and partially observed setting using a small number of lateral and/or horizontal
slices of the underlying tensor. It is worth mentioning that randomized algorithms were used
to efficiently solve the robust matrix PCA problem using small sketches constructed from ran-
dom linear measurements of the low rank matrix [43, 55, 34, 36, 39]. Our proposed nuclear
norm minimization approach is different from these studies, since we relay on slice selection
and projection (mainly t-CUR factorization). Further, our proposed algorithm uses an adap-
tive sampling strategy and provides high probability recovery guarantees of the exact low rank
tensor approximation under a weaker coherence assumption.
The remainder of the paper is organized as follows. In Section 2, we review some relevant
mathematical concepts including the tensor circulant algebra, basic definitions and theorems of
tensors, and the t-SVD decomposition based on the t-product concept. In Section 3, we provide
the randomized tensor decompositions and provide their relative error guarantees and introduce
4
D. A. TARZANAGH AND G. MICHAILIDIS
the concept of slice selection and projection. In Section 4, we introduce, evaluate and analyze
our proposed algorithm (CUR t-NN) for large scale noisy tensor decomposition. Experimental
results are presented in Section 5 and some concluding remarks are drawn in Section 6.
The detailed proofs of the main results established are delegated to the Appendix.
2. Mathematical preliminaries. Next, we introduce key definitions and concepts used
in subsequent developments.
Tensor indexing. We denote tensors by boldface Euler script letters, e.g., X, matrices by
boldface capital letters, e.g., X, vectors by lowercase letters, e.g., x. The order of a
tensor is the number of dimensions (also refereed to as ways or modes). In this work,
we focus on 3-way tensors.
Fibers and slices [51]. A fiber of tensor X is a one-dimensional array defined by fixing two of
the indices. Specifically, X:jk is the (j,k)-th column fiber, Xi:k is the (i,k)-th row fiber,
and Xij: is the (i,j)-th tube fiber. A slice of tensor X is a two-dimensional array defined
by fixing one index only. Specifically, Xi:: is the i-th horizontal slice, X:j: is the j-th
lateral slice, and X::k is the k-th frontal slice. For convenience, X::k is written as Xk.
The vectorization of X is denoted by vec(X). For a 3-way tensor X ∈ Rn1×n2×n3, we
denote its (i, j, k)-th entry as Xijk.
(a) horizontal slices
(b) lateral slices
(c) frontal slices
Fig. 1: Slices of an n1 × n2 × n3 tensor X.
Norms. We denote the (cid:96)1 norm as (cid:107)X(cid:107)1 := (cid:80)
maxijk xijk, and the Frobenius norm as (cid:107)X(cid:107)F :=
reduce to the vector or matrix norms if X is a vector or a matrix.
ijk xijk, the infinity norm as (cid:107)X(cid:107)∞ :=
ijk xijk2. The above norms
Operators. For X ∈ Rn1×n2×n3, and using the Matlab commands fft,ifft, we denote by
X the result of applying the Discrete Fourier Transform (DFT) on X along the 3-rd
dimension, i.e., X = fft(X, [], 3). Analogously, one can also compute X from X via
the Inverse Discrete Fourier Transform (IDFT), using ifft( X, [], 3). In particular, we
denote by X the block diagonal matrix with each blockcorresponding to the frontal
slice X::k of X. i.e.,
(cid:113)(cid:80)
X::1
X::2
. . .
X::n3
.
X = bdiagk∈[n3]( X::k) :=
2.1. Tensor basics. Next, we review relevant mathematical concepts including the tensor
SVD (t-SVD), basic definitions and operations and other technical results of tensors [27, 44, 53],
that are used throughout the paper.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
5
Definition 1. (Tensor product). Given two tensors Z ∈ Rn1×n2×n3 and X ∈ Rn2×n4×n5,
the t-product Z ∗ X is the n1 × n4 × n3 tensor,
(2)
where
and
C = Z ∗ X = fold (circ (Z)) .unfold (X) ,
,
Z::n3
Z::1
...
Z::n3−1
. . . Z::2
. . . Z::3
...
. . .
. . . Z::1
circ (Z) :=
Z::1
Z::2
...
Z::n3
,
X::1
X::2
...
X::n3
unfold (X) :=
fold (unfold (X)) = X.
Because the circular convolution of two tube fibers can be computed by the DFT, the
t-product can be alternatively computed in the Fourier domain, as shown in Algorithm 1.
Algorithm 1 t-product C = Z ∗ X in the Fourier domain
1: Input: Z ∈ Rn1×n2×n3; X ∈ Rn2×n4×n3
2: Z ← fft(Z, [], 3);
3: X ← fft(X, [], 3);
4: for k = 1, . . . , n3 do
X::k;
5:
6: end for
7: C ← ifft(C, [], 3);
8: return C
C::k = Z::k
Definition 2. (Conjugate transpose). The conjugate transpose of a tensor X ∈ Rn1×n2×n3
is tensor X∗ ∈ Rn2×n1×n3 obtained by conjugate transposing each of the frontal slices and then
reversing the order of transposed frontal slices 2 through n3.
Definition 3. (Identity tensor). The identity tensor I ∈ Rn×n×n3 is the tensor whose
first frontal slice is the n × n identity matrix, and whose other frontal slices are all zeros.
Definition 4. (Orthogonal tensor). A tensor Q ∈ Rn×n×n3 is orthogonal if it satisfies
Q∗ ∗ Q = Q ∗ Q∗
= I.
Definition 5. (F-diagonal Tensor). A tensor is called f−diagonal if each of its frontal
slices is a diagonal matrix.
Theorem 6. (t-SVD). Let X ∈ Rn1×n2×n3. Then, it can be factored as
(3)
where U ∈ Rn1×n2×n3, V ∈ Rn1×n2×n3 are orthogonal and Σ ∈ Rn1×n2×n3 is a f-diagonal
tensor.
X = U ∗ Σ ∗ VT ,
Note that t-SVD can be efficiently computed based on the matrix SVD in the Fourier
domain. This is based on the key property that the block circulant matrix can be mapped to
a block diagonal matrix in the Fourier domain, i.e.
(4)
where Fn3 denotes the n3 × n3 DFT matrix and ⊗ denotes the Kronecker product.
n3
(Fn3 ⊗ In1) · circ (X) · (F −1
⊗ In2 ) = X,
6
D. A. TARZANAGH AND G. MICHAILIDIS
Definition 7. (Tensor multi and tubal rank). The tensor multi-rank of X ∈ Rn1×n2×n3
is a vector υ ∈ Rn3 with its i − th entry being the rank of the i-th frontal slice of X, ri =
rank( Xi::). The tensor tubal rank, denoted by r = rankt(X), is defined as the number of nonzero
singular tubes of Σ, where Σ is obtained from the t-SVD of X = U ∗ Σ ∗ VT . That is,
(5)
r = card{i : Σii: (cid:54)= 0} = max
i
ri,
where card denotes the cardinality of a set.
The tensor tubal rank shares some properties of the matrix rank, e.g.
rankt(X ∗ Z) ≤ min(rankt(X), rankt(Z)).
Many tensor completion and decomposition techniques for video and seismic noise reduction
rely on a low-rank factorization of a time-frequency transform. Further, certain energy methods
broadly used in image processing, e.g., PDEs [1] and belief propagation techniques mainly focus
on local relationships. The basic assumption is that the missing entries depend primarily on
their neighbors. Hence, the further apart two pixels are, the smaller their dependance is.
However, for video and time series of images the value of the missing entry also depends on
entries which are relatively far away in the time/sequence dimension. Thus, it is necessary to
develop a tool to directly capture such global information in the data. Using (5), r0 is the
rank of the "mean image" across the video sequence. Meanwhile, r1 is the rank of the next
frequency's content across frames, etc. Under a smoothness assumption that captures global
information at given pixels across time, the frontal slices (after FFT) for bigger i have smaller
singular values.
Definition 8. (Tensor nuclear norm). The tensor nuclear norm of a tensor X ∈
Rn1×n2×n3, denoted by (cid:107)X(cid:107)(cid:126), is defined as the average of the nuclear norm of all frontal slices
of X, i.e.
(6)
(cid:107)X(cid:107)(cid:126) :=
1
n3
(cid:107) X::k(cid:107)∗.
n3(cid:88)
k=1
The above tensor nuclear norm is defined in the Fourier domain. It is closely related to the
nuclear norm of the block circulant matrix in the original domain. Indeed,
1
n3
(cid:107) X(cid:107)∗
(cid:107) X::k(cid:107)∗ =
n3(cid:88)
(cid:107)(Fn3 ⊗ In1 ) · circ (X) · (F −1
(cid:107)circ (X)(cid:107)∗.
k=1
n3
(cid:107)X(cid:107)(cid:126) =
=
=
1
n3
1
n3
1
n3
⊗ In2 )(cid:107)∗
The above relationship gives an equivalent definition of the tensor nuclear norm in the
original domain. Thus, the tensor nuclear norm is the nuclear norm (with a factor 1/n3) of a
new matricization (block circulant matrix) of a tensor.
Definition 9. (Tensor spectral norm). The tensor spectral norm of X ∈ Rn×n×n3 ,
denoted as (cid:107)X(cid:107), is defined as
(cid:107)X(cid:107) := (cid:107) X(cid:107)2 = (cid:107)(Fn3 ⊗ In1) · circ (X) · (F −1
⊗ In2)(cid:107)2,
n3
where (cid:107) · (cid:107)2 denotes the spectral norm of a matrix.
Definition 10. (Inverse of tensor). The inverse of X ∈ Rn×n×n3, denoted by X−1
satisfies
(7)
where I is the identity tensor of size n × n × n3.
X−1 ∗ X = X ∗ X−1 = I,
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
7
Definition 11. (Standard tensor basis). The lateral basis ei, is of size n1 × 1 × n3
with only one entry equal to 1 and the remaining equal to zero, in which the nonzero entry 1 will
(cid:62)
only appear at the first frontal slice of ei. Normally its transpose e
is called the horizontal
i
basis. The other standard tensor basis is called tube basis ei, and corresponds to a tensor of
size 1× 1× n3 with one entry equal to 1 and the rest equal to 0. Figure 2 illustrates these bases.
Fig. 2: The lateral basis e3 and the tube basis e5. The black cubes are 1, gray and white cubes
are 0. The white cubes stand for the potential entries that could be 1.
2.2. Linear algebra with tensors: free submodules. The set of complex numbers C
with standard scalar addition and multiplication is a field and Cn3 forms a vector space over
this field. However, as pointed out in [25, 27], the set of tubes C1×1×n3 equipped with the tensor
product form a ring with unity. A module over a ring can be thought of as a generalization
of a vector space over a field, where the corresponding scalars are the elements of the ring. In
linear algebra over a ring, the analog of a subspace is a free submodule. Our algorithm relies
on submodules, and the following theorem.
Theorem 12. The set of slices
(8)
forms a free module over the set of tubes C1×1×n3.
Υ := {X:j: X:j: ∈ Cn1×1×n3,
j ∈ [n2]},
Proof. A detailed proof is provided in [4].
Using Theorem 12, Υ has a basis so that any element of it can be written as a "t-linear
combination" of elements of a set of basis slices. A "t-linear combination", is defined as a sum
of slices multiplied, based on the t-product, by coefficients from C1×1×n3.
Definition 13. (Slice-wise linear independence) The slices in a subset Λ = {X:1:, . . . , X:n:}
of Υ are said to be linearly dependent, if there exists a finite number of distinct slices X:1:, . . . , X:m:
in Λ, and tubes C11:, . . . , Cmm:, not all zero, such that
where O denotes the lateral slice comprising of all zeros.
The slices in a subset Λ = {X:1:, . . . , X:n:} of Υ are said to be linearly independent if the
can only be satisfied by all zero tubes Cjj:, j = 1, . . . , n.
j=1
Based on the computable t-SVD, the tensor nuclear norm [44] is used to replace the tubal
rank for low-rank tensor recovery (from incomplete/corrupted tensors) by solving the following
(9)
equation
m(cid:88)
j=1
n(cid:88)
X:j: ∗ Cjj: = O,
X:j: ∗ Cjj: = O,
8
D. A. TARZANAGH AND G. MICHAILIDIS
convex program,
(10)
where (cid:107)L(cid:107)(cid:126) denotes the tensor nuclear norm and
min
s.t.
L
(cid:107)L(cid:107)(cid:126)
(cid:107)PΩ(X − L)(cid:107) ≤ ∆,
(PΩ(X)ij = Xij,
if
(i, j) ∈ Ω and (PΩ(X))ij = 0
otherwise.
Similarly to the matrix completion problem, recovery of tensor X from its observed entries
is essentially infeasible if the large majority of the entries are equal to zero [7]. For the tensor
completion case, it is the case that if tensor X only has a few entries which are not equal to
zero, in its t-SVD U ∗ S ∗ V(cid:62)
= X, the singular tensors U and V will be highly concentrated.
Indeed, not all tensors can be recovered from data sets with missing entries and/or large outliers.
Hence, in analogy to the main idea in matrix completion [7], tensor slices U(:, i, :) and V(:, i, :
), i = 1, 2, ..., r, need to be sufficiently spread out, which in turn implies that they should be
uncorrelated with the standard tensor basis. Our analysis in Section 4 will focus on noisy tensor
completion based on a robust factorization algorithm whose estimation/recovery guarantees are
expressed in terms of the coherence of the target low-rank tensor L. It establishes that lower
values of tensor coherence provide better recovery results. In addition, we propose a randomized
approximation algorithm, whose guarantees are also related to the notion of tensor coherence.
The following three notions of tensor coherence are defined next.
Definition 14. (Tensor µ0-coherence). Let V ∈ Rn×r×n3 contain orthonormal lateral
basis with r ≤ n. Then, the µ0-coherence of V is given by:
nn3
nn3
µ0(V) :=
max
1≤i≤n
r
(cid:107)V(cid:62) ∗ei(cid:107)2
F =
(cid:107)Vi::(cid:107)2
F ,
max
1≤i≤n
r
where ei is standard lateral basis.
Definition 15. (Tensor µ1-coherence). For a tensor L ∈ Rn1×n2×n3 assume that
rankt(L) = r. Then, the µ1-coherence of L is defined as
r
where L has the skinny t-SVD L = U ∗ S ∗ V(cid:62)
.
µ1(L) :=
n1n2n2
3
(cid:107)U ∗ VT(cid:107)2∞,
Definition 16. (Tensor (µ, r)-coherence). For any µ > 0, we call a tensor L (µ, r)-
coherent if
rankt(L) = r,
max{µ0(U), µ0(V)} ≤ µ,
µ1(L) ≤ µ.
Note that the standard tensor coherence condition is much weaker than the matrix weak co-
herence one for each frontal slice of X [53]. Hence, in the analysis of the proposed randomized
algorithms, we will use the standard tensor coherence condition.
3. Approximate Tensor Low Rank Decompositions. We develop extensions of the
matrix CX and CUR decompositions [13, 15] to third-order tensors. The proposed algorithms
result in computing low-rank tensor approximations that are explicitly expressed in terms of a
small number of slices of the input tensor. We start by introducing a basic computational tool
-a randomized tensor multiplication procedure- that is used in subsequent developments.
3.1. Approximate tensor multiplication. Next, we present a randomized tensor-product
and provide key results on the quality of the resulting approximation. Given two tensors A
and B, using Definition 1, the t-product may be written as follows:
(11)
A ∗ B =
A:t: ∗ Bt::,
n2(cid:88)
t=1
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
9
where ∗ denotes the tensor product.
It can be easily shown that the left hand side of (11) is equivalent to the block multiplication
and then summation in the Fourier domain. When tensor multiplication is formulated as (11),
we can develop a randomized algorithm to approximate the product A ∗ B.
Algorithm 2 takes as input two tensors A ∈ Rn1×n2×n3 and B ∈ Rn2×n4×n3 , a positive
integer c ≤ n2, and a probability distribution {pi}n2
i=1 over [n2]. It returns as output two tensors
C and R, where the lateral slices of C correspond to a small number of sampled and rescaled
slices of A, and similarly the horizontal slices of R constitute a small number of sampled and
rescaled slices of B. Specifically, consider at most c lateral slices (in expectation) of A selected,
with the i-th lateral slice of A in C be chosen with probability pi = min{1, cpi}. Then, define
the sampling tensor S ∈ Rn2×n2×n3 to be binary where Sii1 = 1 if the i-th slice is selected
and Sij2:n3 = 0 otherwise. Define the rescaling tensor D ∈ Rn2×c×n3 to be the tensor with
cpj, if i − 1 of the previous slices have been selected and Dij2:n3 = 0 otherwise.
Dij1 = 1/
Analogously, the rt-product samples and rescales the corresponding horizontal slices of tensor
B. In both cases, C = A ∗ S ∗ D is an n1 × c × n3 tensor consisting of sampled and rescaled
copies of the lateral slices of A, and R = (S ∗ D)(cid:62) ∗ B = D ∗ S(cid:62) ∗ B is a c × n4 × n3 tensor
comprising of sampled and rescaled copies of the horizontal slices of B. For the case of n3 = 1,
the algorithm selects column-row pairs in Algorithm 2 and is identical to the algorithm in [13].
√
Algorithm 2 (rt-product), a fast Monte-Carlo algorithm for approximate tensor multiplica-
tion.
1: Input: A ∈ Rn1×n2×n3, B ∈ Rn2×n4×n3, pi ≥ 0, i ∈ [n2] s.t. (cid:80)
i∈[n2] pi = 1, positive
integer c ≤ n2.
2: Initialize S ∈ Rn2×n2×n3 and D ∈ Rn2×c×n3 to the all zeros tensors;
3: t = 1;
4: for i = 1, . . . , n2 do
5:
6:
7:
8:
Pick i with probability min{1, cpi};
if i is picked then
Sit1 = 1, Sit2:n3 = 0;
Dtt1 = 1/ min{1,
t = t + 1;
cpi}, Dtt2:n3 = 0;
√
9:
end if
10:
11: end for
12: C = A ∗ S ∗ D, R = D ∗ S(cid:62) ∗ B;
13: C ← fft(C, [], 3),
14: for k = 1, . . . , n3 do
R::k;
15:
16: end for
17: Z ← ifft( Z, [], 3);
18: return Z
Z::k = C::k
R ← fft(R, [], 3);
3.2. Running time of rt-product. The rt-product is computationally efficient, has small
memory requirements, and is well suited for large scale problems.
Indeed, the rt-product
algorithm can be implemented without storing tensors A ∈ Rn1×n2×n3 and B ∈ Rn2×n4×n3
and uses O(max(n1, n4)cn3 log(n3)) flops to transform the input to the Fourier domain and
O(c(n1 + n2 + n4)n3) flops to construct C ∈ Rn1×c×n3 and R ∈ Rc×n4×n3 , where
(12)
A ∗ B ∼ C ∗ R.
It is worth mentioning that we do not require storing the sampling and rescaling tensors S and
D in our implementation.
Next, we provide the main result on the quality of the approximation obtained by Algo-
rithm 2 that specifies the assumptions under which (12) holds. The most interesting of these
10
D. A. TARZANAGH AND G. MICHAILIDIS
assumptions is that the sampling probabilities used to randomly sample the lateral slices of A
and the corresponding horizontal slices of B are non-uniform and depend on the product of
the norms of the lateral slices of A and/or the corresponding horizontal slices of B. Following
[13], we consider two examples of nonuniform sampling probabilities:
i. If we would like to use information from both tensors A and B, we consider sampling
probabilities {pi}n2
i=1 such that
(13)
pi ≥ β
(cid:80)n2
(cid:107)A:i:(cid:107)F(cid:107)Bi::(cid:107)F
i=1 (cid:107)A:i:(cid:107)F(cid:107)Bi::(cid:107)F
,
β ∈ (0, 1].
ii. If only information on A is easily available, then we use sampling probabilities {pi}n2
i=1 such
that
(14)
pi ≥ β
(cid:107)A:i:(cid:107)2
(cid:107)A(cid:107)2
F
F
,
β ∈ (0, 1].
The following theorem gives the main result for Algorithm 2, and generalizes Theorem 6 in
[15] to third order tensors .
Theorem 17. Suppose A ∈ Rn1×n2×n3, B ∈ Rn2×n4×n5, and c ≤ n2. In Algorithm 2, if
the sampling probabilities {pi}n2
i=1 satisfy (13) or (14), then the following holds
(15)
E[(cid:107)A ∗ B − C ∗ R(cid:107)F ] ≤ 1√
βc
(cid:107)A(cid:107)F(cid:107)B(cid:107)F .
The following lemma establishes a bound for tensor multiplication with respect to the spectral
norm with improved sampling complexity.
Lemma 18. Given a tensor A ∈ Rn1×n2×n3, choose c ≥ 48r log(4r/(βδ))/(β2). Let C ∈
be the corresponding subtensors of A. If sampling probabilities {pi}n2
Rn1×c×n3 and R = C(cid:62)
satisfy (14), then with probability at least 1 − δ, we have
i=1
(16)
where kπ = k ∈ [n3] is the index of the tensor's frontal slice with the maximum spectral norm
(cid:107) A::k
(cid:62)
::k − C::k
(cid:62)
::k(cid:107)2.
A
C
≤ /2(cid:107) A::kπ(cid:107)2
2,
3.3. Slice-based tensor CX decomposition. Next, using the concept of free submod-
ules and Definition 13, we introduce the novel notion of slice selection and projection before
formulating a tensor CX decomposition.
Definition 19. Let C ∈ Rn1×c×n3 and X ∈ Rn1×n2×n3. Then, the tensor project (t-
project) operator for X, ΠC(X), is an n1 × n2 × n3 tensor obtained as
(17)
where ∗ denotes the t-product, ΠC(X) is the projection of X onto the subspace spanned by the
lateral slices of C, and C†
is the Moore-Penrose generalized inverse of tensor C [26].
ΠC(X) := C ∗ C† ∗ X,
Because the circulant convolution of two tube fibers can be computed by the DFT, the
t-project can be alternatively computed in the Fourier domain, as shown in Algorithm 3.
Definition 20. (t-CX). Given an n1 × n2 × n3 tensor X, let C be an n1 × c × n3 tensor
whose lateral slices correspond to c lateral slices from tensor X. Then, the n1 × n2 × n3 tensor
is called a t-CX decomposition of X.
ΠC(X) = C ∗ C† ∗ X,
(cid:107)A ∗ A(cid:62) − C ∗ C(cid:62)(cid:107) ≤ max
k∈[n3]
(cid:62)
::k − C::k
(cid:62)
::k(cid:107)2
C
A
(cid:8)(cid:107) A::k
(cid:9)
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
11
Algorithm 3 t-project ΠC(X) computation in the Fourier domain.
1: Input: X ∈ Rn1×n2×n3; C ∈ Rn1×c×n3.
2: X ← fft(X, [], 3);
3: C ← fft(C, [], 3);
4: for k = 1, . . . , n3 do
†
X::k;
::k
Z::k = C::k
C
5:
6: end for
7: Z ← ifft( Z, [], 3);
8: return ΠC(X) = Z
The following remarks are of interest for the previous definition.
• The choice for the number of lateral slices c in the t-CX approximation depends on the
application under consideration; neverthelees, we are primarily interested in the case
c (cid:28) n2. For example, c could be constant, independent of the dimension of tensor, or
logarithmic in the size of n2, or a large constant factor less than n2.
• The t-CX decomposition expresses each X slice in terms of a linear combination of
basis slices (see, Definition 13), each of which corresponds to an actual lateral slice
of X. Hence, t-CX provides a low-rank approximation to the original tensor X, even
though its structural properties are different than those of the t-SVD.
• Given a set of lateral slices C, the approximation ΠC(X) = C ∗ C† ∗ X is the "best"
approximation to X in the following sense
(cid:107)X − C ∗ (C† ∗ X)(cid:107)F =
(18)
min
Y∈Rn1×c×n3
(cid:107)X − C ∗ Y(cid:107)F .
Next, we provide the t-CX decomposition algorithm and provide its relative error. Al-
gorithm 4 takes as input an n1 × n2 × n3 tensor A, a tubal rank parameter r, and an error
parameter . It returns as output an n1 × c × n3 tensor C comprising of a small number of
slices of A. Let L = U ∗ Σ ∗ V(cid:62)
, be a tubal rank-r approximation of tensor A. Central to
our proposed randomized algorithm is the concept of sampling slices of the tensor based on a
leverage score [15], defined by
pi ≥ β
rn3
∀ i ∈ [n2],
β ∈ (0, 1].
(cid:107) Vi::(cid:107)2
F ,
(19)
where V = fft(V, [], 3). The goal is to select a small number of lateral slices of A.
Note that we define rescaling and sampled tensors D and S in the Fourier domain. This
leads to a significant reduction in time complexity of the fast Fourier transform and its inverse.
For the case n3 = 1, the algorithm selects column-row pairs in Algorithm 4 and is identical to
the algorithm of [15].
Next, we present a lemma of general interest. The derived properties aid in obtaining tensor
based randomized (cid:96)2 regression and low rank estimation guarantees.
Lemma 21. Given a target tensor A ∈ Rn1×n2×n3 , let ∈ (0, 1] and L = U ∗ Σ ∗ V(cid:62)
be a
rankt-r approximation of A. Let Γ = (V(cid:62)∗ S∗ D)†−(V(cid:62)∗ S∗ D)(cid:62). If the sampling probabilities
{pi}n2
i=1 satisfy (19), and if c ≥ 48r log(4r/(βδ))/(β2), then with probability at least 1 − δ, the
following hold
(20a)
(20b)
(20c)
(20d)
rankt(V(cid:62) ∗ S) = rankt(V) = rankt(L),
(cid:107)Γ(cid:107) = (cid:107)Σ−1
V(cid:62)∗S∗D − ΣV(cid:62)∗S∗D(cid:107),
(L ∗ S ∗ D)† = (V(cid:62) ∗ S ∗ D)†Σ−1U(cid:62)
,
(cid:107)Σ−1
V(cid:62)∗S∗D − ΣV(cid:62)∗S∗D(cid:107)2 ≤
2
,
√
2
12
D. A. TARZANAGH AND G. MICHAILIDIS
Algorithm 4 (t-CX), a fast Monte-Carlo algorithm for tensor low rank approximation.
i∈[n2] pi = 1, a rank parameter r, positive
1: Input: A ∈ Rn1×n2×n3 , pi ≥ 0, i ∈ [n2] s.t. (cid:80)
integer c ≤ n2.
2: A ← fft(A, [], 3);
3: Initialize S and D to the all zeros tensors;
4: t = 1;
5: for i = 1, . . . , n2 do
6:
7:
8:
Pick i with probability min{1, cpi};
if i is picked then
Sit1 = 1,
Dtt1 = 1/ min{1,
t = t + 1;
Sit2:n3 = 0;
cpi},
Dtt2:n3 = 0;
√
9:
10:
end if
11:
12: end for
13: for k = 1, . . . , n3 do
S::k
14:
15: end for
16: Z ← ifft( Z, [], 3);
17: return Z
Z::k = A::k
D::k( A::k
S::k
D::k)† A::k;
where ΣV(cid:62)∗S∗D is an F-diagonal tensor and contains the r non-zero singular tubes of V(cid:62)∗S∗D.
Note that the equation (20d) shows that in terms of its singular tubes, the tensor V(cid:62)∗S∗D
-i.e., the slice sampled and rescaled version of V- is almost an orthogonal tensor. A useful
property of an orthogonal tensor V is that V†
(see, Definition 4). Equation (20b) shows
that although this property does not hold for V(cid:62) ∗ S ∗ D, the difference between (V(cid:62) ∗ S ∗ D)†
and (V(cid:62) ∗ S ∗ D)(cid:62) can be bounded.
= V(cid:62)
Using Theorem 17 and Lemma 21, we provide in proposition 22 a sampling complexity for
a tensor based randomized "(cid:96)2-regression" as in [15]. Indeed, Lemmas 2 and 3 of [15] follow
by using the Frobenius norm bound of Theorem 17 and applying Markov's inequality. The
claim of Lemma 1 of [15] follows by applying Lemma 21. If we consider the failing probability
δ = 0.05, the claims of all three lemmas hold simultaneously with probability at least 0.85 and
in the following proposition we condition on this event. The proof follows along similar lines to
the proof of Theorem 5 in [15].
Proposition 22. Given a target tensor A ∈ Rn1×n2×n3 , let L = U∗Σ∗V(cid:62)
be a rankt-r ap-
proximation of A. Choose c = O(r log(r/β)/(β2)) slices of A with their sampling probabilities
{pi}n2
i=1 satisfying (19). Then, with probability at least 0.85, the following holds
(cid:107)A − A ∗ S ∗ D ∗ (L ∗ S ∗ D)† ∗ L(cid:107)F ≤ (1 + )(cid:107)A − A ∗ L† ∗ L(cid:107)F.
(21)
Let L = U ∗ Σ ∗ V(cid:62)
be a t-SVD of tensor L ∈ Rn1×n2×n3. In the remainder of the paper,
we use the following two parameters related to the coherence of tensor L,
:=
rµ0(V)
n3
,
(22)
where Uc is the left singular tensor of C ∈ Rn1×c×n3 lateral slices of L.
c :=
n3
cµ0(Uc)
,
The following theorem shows that projection based on slice sampling leads to near optimal
estimation in tensor regression, when the covariate tensor has small coherence, µ0(V).
Theorem 23. Given A ∈ Rn1×n2×n3, let L = U∗ Σ∗ V(cid:62)
be a rankt-r approximation of A.
Choose c = O( log()/2), and let Ac ∈ Rn1×c×n2 be a tensor of c lateral slices of A sampled
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
13
uniformly without replacement. Further, let Lc ∈ Rn1×c×n3 consist of the corresponding slices
of L. Then, with probability at least 0.85, the following holds
(cid:107)A − Ac ∗ L†
c ∗ L(cid:107)F ≤ (1 + )(cid:107)A − A ∗ L† ∗ L(cid:107)F .
Next, by using Lemma 21, we provide a low rank estimation bound for the t-CX algorithm,
under the coherence assumption. Indeed, we will take advantage of a constant β in equation (19)
to provide relative error estimation guarantees for the t-CX algorithm under uniform slice
sampling, when tensor A under consideration exhibits sufficient incoherence.
Corollary 24. Given a target tensor A ∈ Rn1×n2×n3 , let L = U ∗ Σ ∗ V(cid:62)
be a rankt −r
approximation of A. Choose c = O( log() log(1/δ)/2), and let C be a tensor of c lateral slices
of A sampled uniformly without replacement. Then, the following holds
(23)
with probability at least 1 − δ.
(cid:107)A − C ∗ C† ∗ A(cid:107)F ≤ (1 + )(cid:107)A − L(cid:107)F ,
3.4. Slice-based tensor CUR decomposition. Similar to the matrix case, the t-CX
decomposition suffices when n1 (cid:28) n2 because C† ∗ A is small in size. However, when n1
and n2 are almost equal, computing and storing the dense matrix C† ∗ A in memory becomes
prohibitive. The CUR decomposition provides a very useful alternative. Next, we introduce a
tensor CUR (t-CUR) decomposition based on the rt-product.
Definition 25. (t-CUR). Let X be an n1×n2×n3 tensor. For any given C, an n1×c×n3
tensor whose slices comprise of c lateral slices of tensor X, and R, an l × n2 × n3 tensor whose
slices comprise of l horizontal slices of tensor X, the n1 × n2 × n3 tensor
C ∗ U ∗ R,
is a lateral-horizontal-based tensor approximation to X for any c × l × n3 tensor U.
The following remarks are of interest regarding the previous definition.
• The t-CUR decomposition is most appropriate as a data analysis tool, when the data
consist of one- and/or two modes that are qualitatively different than the remaining
ones. In this case, the t-CUR decomposition approximately expresses the original data
tensor in terms of a basis consisting of underlying subtensors that are actual data slices
and not artificial bases.
• The t-CUR approximation is a t-CX approximation, but one with a very special struc-
ture; i.e., every lateral slice of X can be expressed in terms of the basis provided by C
(see, Definition 13) using only the information contained in a small number of horizontal
slices of X and a low-dimensional encoding tensor.
• In terms of its t-SVD structure, U contains the "inverse-of-X" information. For the
proposed t-CUR decomposition, U will be a generalized inverse of the intersection
between selected tensors C and R.
Note that the structural simplicity of the t-CUR tensor decomposition becomes apparent
in the Fourier domain, as detailed in Algorithm 5. The latter takes as input an n1 × n2 × n3
tensor A, an n1 × c × n3 tensor C consisting of a small number of lateral slices of A, and an
error parameter . Letting C = U ∗ Σ ∗ V
, Algorithm 5 uses sampling probabilities
(cid:62)
(24)
pi ≥ β
cn3
(cid:62)
(cid:107) U
i::(cid:107),
∀i ∈ [n1],
β ∈ (0, 1],
to select a small number of horizontal slices of A. It returns an l×n2×n3 tensor R comprising of
a small number of horizontal slices of A and an c×l×n3 tensor U consisting of the corresponding
slices of C.
14
D. A. TARZANAGH AND G. MICHAILIDIS
Algorithm 5 (t-CUR), a fast Monte-Carlo algorithm for tensor low rank approximation.
1: Input: A ∈ Rn1×n2×n3, C ∈ Rn1×c×n3 consisting of c lateral slices of A, pi ≥ 0, i ∈ [n2]
i∈[n2] pi = 1, a rank parameter r, and positive integer l ≤ n1.
s.t. (cid:80)
2: A ← fft(A, [], 3);
3: C ← fft(C, [], 3);
4: Initialize S and D to the all zeros tensors;
5: t = 1;
6: for i = 1, . . . , n1 do
7:
8:
Pick i with probability min{1, lpi};
if i is picked then
Sit1 = 1,
Dtt1 = 1/ min{1,
t = t + 1;
Sit2:n3 = 0;
lpi},
Dtt2:n3 = 0;
√
9:
10:
11:
15:
16:
R::k = D::k
end if
12:
13: end for
14: for k = 1, . . . , n3 do
U::k =(cid:0) D::k
(cid:62)
A::k;
S
::k
(cid:62)
S
C::k
::k
U::k
R::k;
17:
18: end for
19: Z ← ifft( Z, [], 3);
20: return Z
Z::k = C::k
(cid:1)†
;
Corollary 26. Given A ∈ Rn1×n2×n3 , let L = U ∗ Σ ∗ V(cid:62)
be a rankt −r approxi-
In Algorithm 5, choose c = O( log() log(1/δ)/2), and let C ∈ Rn1×c×n3
mation of A.
be a tensor of c slices of A sampled uniformly without replacement. Further, choose l =
O(c log(c) log(1/δ)/2), and let R ∈ Rl×n2×n3 be a tensor of l horizontal slices of A sampled
uniformly without replacement. Then, the following holds
(cid:107)A − C ∗ U ∗ R(cid:107)F ≤ (1 + )(cid:107)A − L(cid:107)F ,
with probability at least 1 − δ.
Remark 27. In many applications such as tensor completion problems discussed in Section
5.3, it may not feasible to compute the t-SVD of the entire tensor A due to either computational
cost or large set of missing values. In these cases, algorithms requiring knowledge of the leverage
scores can not be applied. Hence, we may use an estimate where a subset of the lateral slices
are chosen uniformly and without replacement, and the horizontal leverage scores (24) are
calculated using the top-r right singular slices of this lateral tensor instead of the entire A
tensor.
Remark 28. Note that for any subspace, the smallest µ0 can be is 1.
In such case, we
are using the optimal subspace sampling with β = 1. Note also that we have provided Corol-
laries (24) and (26) based on uniform sampling. However, it can be easily seen that the rel-
ative error guarantees hold with nonuniform sampling probabilities (19) and (24), if we set
c = O(r log(r/β) log(1/δ)/(β2)), and l = O(c log(c/β) log(1/δ)/(β2)).
4. Tensor completion and robust factorization. Next, we provide a CUR tensor
nuclear norm minimization (CUR t-NN) procedure that solves a noisy tensor factorization
problem, using a small number of lateral and/or horizontal slices of the underlying tensor and
exhibits favorable computational complexity and in addition comes with performance guaran-
tees.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
15
4.1. Related work on matrix problems. Low rank plus sparse matrix decomposition.
In many image processing and computer vision applications, the given data matrix X can
be decomposed as a sum of a low-rank and a sparse component. To that end, Cand´es et
al. proposed Robust PCA [6] to model data matrices generated according to the following
mechanism:
(25)
rank(L) + (cid:107)E(cid:107)0
min
L
s.t.
X = L + E.
Note that the solution of (25) is an NP-hard problem. It is established in [7] that if L exhibits
a certain degree of low-rankness, while E is sparse enough, then the formulation of (25) can be
relaxed into a convex problem of the form:
(cid:107)L(cid:107)∗ + (cid:107)E(cid:107)1
X = L + E.
(26)
min
s.t.
L
This model implicitly assumes that the underlying data structure lies in a single low-rank
subspace. However, in many applications (e.g.
image classification) it is more likely that the
data are obtained from a union of multiple subspaces, and hence recovery of the structure
based on the above decomposition would be inaccurate. In order to segment the data into their
respective subspaces, one needs to compute an affinity matrix that encodes the pairwise affinities
between data vectors. Liu [32] proposed a more general rank minimization problem, where the
data matrix itself is used as the dictionary, resulting in the following convex optimization
problem:
(27)
(cid:107)L(cid:107)∗ + (cid:107)E(cid:107)2,1
min
L
s.t.
X = XL + E,
diag(L) = 0.
When the subspaces are globally independent, the data are noiseless and sampling is suf-
[32] show that the optimal solution, denoted by L∗, to the problem given
ficient, Liu et al.
by 27 corresponds to the widely used Shape Iteration Matrix (SIM) method [9]. The latter
is a "block-diagonal" affinity matrix that indicates the true segmentation of the data. To
handle data corrupted by noise, the popular Low Rank Representation (LRR) introduced in
[32] adopts a regularized formulation that introduces an extra penalty term to fit the noise
component. Further, after obtaining the self-representation matrix L, the affinity matrix C is
usually constructed as C = 1
the obtained affinity matrix C will be processed through a spectral clustering algorithm [40] to
produce the final clustering result and obtain the corresponding data generating subspaces.
2 (L + L(cid:62)), where · represents the absolute operator. Then,
It is established in [52, 47] that combining sparse and low-rank regularization can improve
the performance of image classification. The basic objective function of this combination [52]
is as follows:
λ1(cid:107)L(cid:107)∗ + λ2(cid:107)L(cid:107)1 + λ3(cid:107)E(cid:107)(cid:96)
L
s.t.
min
(28)
where λ1, λ2, λ3 are tuning parameters and (cid:107)E(cid:107)(cid:96) indicates different norms suitable for different
types for corrupting the data by noise; for example, the squared Frobenius norm for Gaussian
noise and the (cid:96)1 norm for random spiked noise. Equation (28) is similar to the objective
functions in [47, 16], where a detailed explanation of the formulation given in (28) is also
provided.
X = DL + E,
diag(Z) = 0,
4.1.1. Low rank tensor decomposition. Lu et al. [35] extended Robust PCA [6] to the
third order tensor based on t-SVD and proposed the following convex optimization problem:
(29)
(cid:107)L(cid:107)(cid:126) + λ(cid:107)E(cid:107)1
min
L,E
s.t.
X = L + E,
where λ is a tuning parameter.
This model implicitly assumes that the underlying data come from a single low-rank sub-
space. When the data is drawn from a union of multiple subspaces, which is common in image
16
D. A. TARZANAGH AND G. MICHAILIDIS
classification, the recovery based on the above formulation may lack in accuracy. To that end,
Xie et al. [49] extended LRR based subspace clustering to a multi-view one, by employing the
rank sum of different mode unfoldings to constrain the subspace coefficient tensor, resulting in
the following convex optimization problem:
(cid:107)L(cid:107)(cid:126) + λ(cid:107)E(cid:107)2,1
X = X ∗ L + E,
(30)
min
L
s.t.
Equation (30) is similar to the objective functions in [25, 42] and a detail explanation of
(30) is provided in that paper.
4.2. Proposed algorithm. Next, we propose an algorithm for large scale tensor decom-
position of noisy data. Our proposal, called CUR t-NN, extends Algorithm 5 to the noisy tensor
factorization. The main steps are outlined in Algorithm 6.
Note that CUR t-NN can be used in combination with an arbitrary optimization algorithm.
In this paper, we have chosen to solve the noisy tensor factorization formulations (10), (29) and
(30) using an Alternating Direction Method of Multiplier (ADMM) algorithm [3]. ADMM is
the most widely used approach for robust tensor PCA in both the fully and partially observed
settings. Indeed, ADMM achieves much higher accuracy than (accelerated) proximal gradient
algorithm using fewer iterations. It works well across a wide range of problem settings and does
not require careful tuning of the regularization parameters. Further, the following empirical
finding has been frequently observed: namely, the rank of the iterates often remains bounded by
the rank of the initializer, thus enabling efficient computations [6]. This feature is not shared by
the block coordinate decent algorithm. We provide a variant of ADMM for solving problem (29)
in the Appendix (see, Algorithm 7). With a small modification, Algorithm 7 can be also used
to solve the tensor completion problem 10, and tensor subspace clustering problem (30). In
the following algorithm, ADMM(X, λ) denotes the ADMM algorithm for solving regularized
tensor nuclear norm minimization problems (10), (29) and (30), where X is the (sampled) data
tensor and λ is a regularization parameter.
Algorithm 6 CUR t-NN, tensor nuclear norm minimization based on CUR factorization.
1: X ∈ Rn1×n2×n3, positive integer c and l, and a regularization parameter λ;
2: Let C be a tensor of c selected lateral slices of X using probabilities (19);
3: C ← ADMM(C, λ);
4: Let R be a tensor of l selected horizontal slices of X using probabilities (24);
5: R ← ADMM(R, λ);
†
6: Let U = W
7: L = C ∗ U ∗ R;
8: return L
, where W is the l × c × n3 tensor formed by sampling the corresponding l
horizontal slices of C;
4.3. Running Time of CUR t-NN. Algorithm 6 significantly reduces the per-iteration
complexity of nuclear norm minimization problems. Indeed, in each iteration, a base tensor
nuclear norm minimization algorithm requires O(n1n2n3 log(n3)) flops to transform the tensor
to the Fourier domain, O(n1n2n3 min(n1, n2)) flops for the t-SVD computation and factor-
ization, and O(n1n2n3 log(n3)) flops to transform it back to the original domain. On the
other hand, Algorithm 6 only requires O(max(cn1, ln2)n3 log(n3)), O(max(n1c, ln2)n3r) and
O(max(cn1, ln2)n3 log(n3)) flops, for the respective steps. Further, Algorithm 6 can be imple-
mented without storing the data tensor X and can be advantageous when r (cid:28) min(n1, n2),
which occurs frequently in real data sets.
4.4. Theoretical guarantees. Next, using Lemma 21, we establish that Algorithm 6
exhibits high probability recovery guarantees comparable to those of the exact algorithms that
use the full data tensor. Our first result bounds the µ0 and µ1-coherence (see, Definitions (14)
and (15)) of a randomized tensor in terms of the coherence of the full tensor.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
17
Lemma 29. Let Lc be a tensor formed by selecting c slices of a rankt −r tensor L =
U ∗ Σ ∗ V that satisfy the probabilistic conditions in (19). If c ≥ 48r log(4r/(βδ))/(β2), then
with probability at least 1 − δ:
• µ0(ULc) = µ0(U),
• µ0(VLc ) ≤
1 − /2
• µ1(Lc) ≤
r
1 − /2
where ∈ (0, 1].
1
µ0(V),
µ0(U)µ0(V),
Our next theorem provides a bound for the estimation error of the CUR t-NN Algorithm.
Theorem 30. Under the notion of Algorithm 6, choose c = O( log() log(1/δ)/2), and
be the corresponding lateral and horizontal sub-
is (µ, r)-coherent, then with probability at least 1 − δ,
and R∗
l = O(c log(c) log(1/δ)/2). Let C∗
tensors of the exact solution L∗
C∗
. If L∗
1−/2 , r)-coherent, and
and R∗
are ( rµ2
(cid:113)
(cid:107)L∗ − L(cid:107)F ≤ (2 + )
(cid:107)C∗ − C(cid:107)2
F + (cid:107)R∗ − R(cid:107)2
F ,
where L is a solution obtained by Algorithm 6.
5. Experimental Results. Next, we investigate the efficiency of the proposed random-
ized algorithms on both synthetic and real data sets. The results are organized in the following
three sub-sections: in Section 5.1, we compare the performance of the rt-product to that of the
t-product using synthetic data. In Section 5.2, we use the proposed t-CX and t-CUR algorithms
for finding important ions and positions in two Mass Spectrometry Imaging (MSI) data sets.
Finally, in Section 5.3, we apply the proposed tensor factorization CUR t-NN algorithm on data
sets related to image and video processing.
All algorithms have been implemented in the MATLAB R2018a environment and run on a
Mac machine equipped with a 1.8 GHz Intel Core i5 processor and 8 GB 1600 MHz DDR3 of
memory.
5.1. Experimental results for fast tensor multipication. A random data tensor X ∈
Rn1×n2×n3 of dimension n1 = 105, n2 = 103 and n3 = 5 is generated as follows: the first frontal
slice of X is set to X(:, :, 1) = sprandn(n1, n2, 1); i.e. comprising of n1 × n2 sparse normally
distributed random matrix. The remaining frontal slices are generated by sprandn(X(:, :, 1)),
which results in the same sparsity structure as X(:, :, 1), with normally distributed random
entries with mean 0 and variance 1.
We perform a t-SVD decomposition on X and select the left singular value tensor U for our
experiments. We consider both uniform and nonuniform sampling schemes. For nonuniform
sampling, we use a randomized approach (similar to Algorithm 4 in [14]) to obtain normal-
ized leverage scores of U. Further, the Frobenius and spectral bounds given in (15) and (16),
respectively, are used as performance metrics for Algorithm 2. Table (2) shows the Rela-
tive Frobenius Error (RFE) -(cid:107)U(cid:62)
F - and Relative Spectral Error (RSE) -
(cid:107)U(cid:62)
r ∗S∗D∗D∗S(cid:62)∗Ur
based on l ≈ r log(r) slices, and its deterministic counterpart. The depicted results are averaged
over 10 independent replications.
(cid:62)
2- for the rt-product algorithm to recover U
r
Ur(cid:107)F /(cid:107)Ur(cid:107)2
Ur(cid:107)2/(cid:107)Ur(cid:107)2
(cid:62)
Ur − U
r
(cid:62)
Ur− U
r
Ur = U(cid:62)
r
r
Very large improvements in computational speed can be seen when using the rt-product,
especially when coupled with a non-uniform sampling scheme. Further, the gains become more
pronounced for larger number of slices sampled and larger rank.
5.2. Finding Important Ions and Positions in MSI. MSI is used to visualize the
spatial distribution of chemical compounds, such as metabolites, peptides or proteins by their
molecular masses. The ability of MSI to localize panels of biomolecules in tissue samples has
led to a rapid and substantial impact in both clinical and pharmacological research, aided in
18
D. A. TARZANAGH AND G. MICHAILIDIS
t-product
rt-product
Rank, selected slices
r=100, l=460
r=200, l=1000
r=300, l=1700
r=400, l=2500
r=500, l=3100
time (s)
9.09
36.36
162.99
321.43
684.01
non-uniform sampling
time(s) RFE RSE
43e-4
25e-4
29e-4
13e-3
12e-4
70e-3
64e-3
56e-3
31e-3
71e-3
8.13
14.27
34.89
41.04
67.07
uniform sampling
time(s) RFE RSE
11e-2
82e-2
34e-2
38e-3
43e-3
39e-2
12e-2
41e-2
90e-2
23e-2
1.03
1.08
1.39
1.64
3.09
Table 1: Relative errors and running times of tensor multipication algorithms.
uncovering biomolecular changes associated with disease and finally provided low cost imaging
of drugs. Typical techniques used require finding important ions and positions from a 3D
image:
ions to be used in fragmentation studies to identify key compounds, and positions
for follow up validation measurements using microdissection or other orthogonal techniques.
Unfortunately, with modern imaging machines, these must be selected from an overwhelming
amount of raw data. Existing popular techniques used to reduce the volume of data, include
principal components analysis and non-negative matrix factorization, but return difficult-to-
interpret linear combinations of actual data elements. A recent paper [50] shows that CX and
CUR matrix decompositions can be used directly to address this selection need. One major
shortcoming of CX and/or CUR matrix decompositions is that they can only handle 2-way
(matrix) data. However, MSI data form a multi-dimensional array. Hence, in order to use
CX/CUR matrix decompositions, one has first to reformulate the multi-way array as a matrix.
Such preprocessing usually leads to information loss, which in turn could cause significant
performance degradation.
By using instead the t-CX and t-CUR decompositions (Algorithms 4 and 5, respectively)
one can obtain good low-rank approximations of the available data, expressed as combinations
of actual ions and positions, as opposed to difficult-to-interpret eigen-ions and eigen-positions
produced by matrix factorization techniques. We show that this leads to effective prioritization
of information for both actual ions and actual positions. In particular, important ions can be
discovered by using leverage scores as the importance sampling distribution. Further, selection
of important positions from the original tensor can be accomplished based on the random
sampling algorithm in [50], since the distribution of the leverage scores of positions is uniform.
To this end, we consider the following two ways of computing leverage scores of a given data
set:
• Deterministic: Compute the normalized tensor leverage scores exactly using probabil-
• Randomized: Compute an approximation to the normalized leverage scores of tensor
(mapped to a block diagonal matrix in the Fourier domain) by using Algorithm 4 of
[14].
ities (19) and (24).
5.2.1. Description and Analysis of MSI data sets. Next, we use the following two
data sets for illustration purposes, that are publicly available at the OpenMSI Web gateway2.
They represent two diverse acquisition modalities, including one mass spectrometry image of the
left coronal hemisphere of a mouse brain (see, Figure 3) acquired using a time-of-flight (TOF)
mass analyzer and one MSI data set of a lung acquired using an Orbitrap mass analyzer. These
data sets form a 122×120×80339 and a 122×120×500000 tensor, respectively. As described in
[50], the brain data set is processed using peak-finding to identify the most intense ions. Using
this technique, the original brain data is reduced from 122× 120× 80339 values to a data set of
size 122× 120× 2000. To compute the CX decomposition, we reshape the MSI data cube into a
2https://openmsi.nersc.gov/openmsi/client/
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
19
Fig. 3: Sample ions in the brain data set.
two-dimensional (14640× 2000) matrix, where each row corresponds to the spectrum of a pixel
in the image, and each column to the intensity of an ion across all pixels, thus describing the
distribution of the ion in physical space. No peak-finding was applied to the lung data set.
Fig. 4: Distribution of leverage scores of tensor X, relative to the best rank−5 space for the
brain data set. Left: Horizontal scores; Right: Lateral scores; Bottom: Frontal scores.
For the brain data set, we evaluate the quality of the approximation of the leverage scores
based on a rank r = 5 approximation. The distribution of the deterministic leverage scores for
the ions and pixels is shown in Figure 4. Table (2) shows the relative square error (RSE)
(cid:107)X − X(cid:107)F /(cid:107)X(cid:107)F using t-CX decomposition to recover rank −5 tensor X for selection of
c = 25, 35, 45 and, 55 ions, using both randomized and deterministic leverage scores. Table
(3) provides the reconstruction errors to the best rank−5 approximation, based on the t-CX
decomposition coupled with horizontal slice selection. The results obtained running both ran-
20
D. A. TARZANAGH AND G. MICHAILIDIS
Number of
selected slices
Randomized
RSE
25
35
45
55
19.13e-2
17.24e-2
16.35e-2
15.14e-2
CX
t-CX
Deterministic
time
RSE
8.46
8.95
14.99
15.86
18.84e-2
17.59e-2
16.93e-2
16.52e-2
time
4.56
5.12
7.01
8.16
Randomized
RSE
13.65e-2
17.43e-2
11.32e-2
16.26e-2
Deterministic
time
RSE
4.47
5.99
7.01
8.99
16.05e-2
13.13e-2
11.01e-2
10.16e-2
time
4.13
4.87
6.14
6.63
Table 2: RSE of matrix and tensor low rank decomposition relative to the best rank−5 space
for identifying important ions in the brain data set.
Number of
selected slices
Randomized
RSE
25
35
45
55
45.24e-2
35.87e-2
24.13e-2
23.16e-2
CX
t-CX
Deterministic
time
RSE
7.41
7.63
11.89
12.93
46.03e-2
35.68e-2
24.19e-2
24.39e-2
time
3.15
3.81
5.29
5.83
Randomized
RSE
37.65e-2
28.00e-2
23.98e-2
15.23e-2
Deterministic
time
RSE
4.06
4.55
5.59
5.71
16.05e-2
13.13e-2
21.01e-2
15.17e-2
time
3.01
3.19
4.15
4.23
Table 3: RSE of matrix and tensor low rank decomposition relative to the best rank−5 space
for finding important pixels in the brain data set.
domized and deterministic CX and t-CX algorithms are based on 10 independent replicates
and then averaging them. Note that for pixel selection, the deterministic CX decomposition
results in larger reconstruction errors than its randomized CX counterpart. The reason for this
behavior lies in the distribution of the leverage scores for the pixels, which are fairly uniform
(see, Figure 4).
Number of
selected slices
Decomposition using ions
Decomposition using ions and pixels
CX
t-CX
CUR
t-CUR
25
35
45
55
RSE
40.82e-2
41.98e-2
30.16e-2
30.74e-2
time
3.31
4.13
5.78
6.67
RSE
46.89e-2
29.63e-2
22.35e-2
20.81e-2
time
3.05
3.24
3.36
4.05
RSE
59.76e-2
42.57e-2
12.18e-2
19.00e-2
time
2.89
3.16
4.25
5.47
RSE
24.67e-2
22.71e-2
18.15e-2
18.15e-2
time
2.14
2.67
3.13
3.26
Table 4: RSE of matrix and tensor low rank decomposition relative to the best rank−15 space
for finding important ions and pixels in the lung data set.
For the lung data set, reconstruction errors to the best rank r = 15 approximation based
on randomized t-CX and t-CUR decompositions are given in Table (4) based on averages over
10 independent replicates. It can be seen that the t-CX and t-CUR match or outperform their
matrix variants in terms of accuracy, while improving on computing time.
These results introduce the concept of t-CX/ t-CUR tensor factorizations to MSI, describ-
ing their utility and illustrating principled algorithmic approaches to deal with the overwhelm-
ing amount of data generated by this technology and their ability to select important and
intepretable ions/pixels.
5.3. Robust PCA in the fully and partially observed settings. Many images exhibit
an inherent low rank structure and are suitably denoised by low-rank modeling methods, such
as robust PCA [6]. In this section, we assess the performance of the CUR t-NN on two popular
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
21
data sets, and for the typical use cases they represent. We compare the performance of the
proposed CUR t-NN algorithm to the following techniques:
• EXACT NN, the exact matrix completion [7];
• RPCA NN, the robust matrix completion [6];
• E-TUCKER NN, the TUCKER based tensor completion [33];
• R-TUCKER NN, the robust TUCKER based tensor completion [20];
• EXACT t-NN, the t-SVD based tensor completion [53];
• RPCA t-NN, the robust t-SVD based tensor completion [35];
All algorithms are terminated either when the relative square error (RSE),
RSE :=
(cid:107)L∗ − L(cid:107)F
(cid:107)L∗(cid:107)F
≤ 10−3,
or the number of iterations and CPU times exceed 1,000 and 20 minutes, respectively.
5.3.1. CUR t-NN on the Extended Yale B data set. We apply the CUR t-NN on the
Extended Yale B data set to evaluate the accuracy of the proposed low-rank representations,
as well as its computation time. The database consists of 2432 images of 38 individuals, each
under 64 different lighting conditions [17]. We used 30 images from each subject and kept them
at full resolution. Each image comprises of 192 × 168 pixels on a grayscale range. The data
are organized into a 192× 1140× 168 tensor that exhibits low tubal rank, which is an expected
feature due to the spatial correlation within lateral slices [18]. Laplacian (salt and pepper 3)
noise was introduced separately in all frontal slices of the observation tensor for 20% of the
entries.
Robust Completion Approach
RPCA NN [6]
R-TUCKER NN [20]
RPCA t-NN [35]
CUR t-NN
RSE
0.0056
0.0034
0.0021
0.0026
Time(s)
417
513
495
136
Table 5: RSE of tensor robust completion methods on Extended Yale B dataset.
We provide visualizations of the reconstructed first image from the first subject at the
20% noise level in Figure 5. We compare the performance of the CUR t-NN, RPCA t-NN,
R-TUCKER NN and RPCA NN for solving the robust tensor low rank approximation prob-
lem (29). Table 5 shows the accuracy of the CUR t-NN together with that of competing
algorithms for the Extended Yale B data set. As shown in Figure 5 and Table 5, CUR t-NN
estimates nearly the same face model as the RPCA t-NN requiring only a small fraction of time.
On the other hand, all algorithms exhibit small RSE, but the CUR t-NN proves essentially as
competitive as the best method RPCA t-NN, but achieves almost the same performance at
approximately 1/4 of the time.
Beyond just speed-ups and/or accuracy improvements of the CUR t- NN algorithm, its
output can be directly used in place of the singular slices and tubes that standard methods
provide. The latter represent linear combinations of the slices of the tensor, which for an
image data set capture an "average eigenface". On the other hand, CUR t-NN reconstructs the
tensor through selection of real faces in the data set, thus giving the opportunity to researchers
to inspect them and examine their representativeness. Hence, similar to the original CUR
decomposition of matrices, CUR t-NN enhances the interpretability of the tensor decomposition
[18, 48, 41].
3This noise can be caused by sharp and sudden disturbances in the image signal. It demonstrates itself as
sparsely occurring white and black pixels.
22
D. A. TARZANAGH AND G. MICHAILIDIS
(a)
(b)
Fig. 5: The 1st frame of the noisy tensor factorization result for the Extended Yale B data
set. (a) Left: The original frame. (a)Right: Noisy image (20% pixels corrupted). (b) Left:
RPCA t-NN [35]. (b) Right: CUR t-NN.
5.3.2. CUR t-NN on a video data set. Next, we compare the CUR t-NN to the
aforementioned listed competing methods for video data representation and compression from
randomly missing entries. The video data, henceforth referred to as the Basketball video (source:
YouTube) is mapped to a 144 × 256 × 80 tensor, obtained from with a nonstationary panning
camera moving from left to right horizontally following the running players. We randomly
sampled 50% entries from the Basketball video. We compare the performance of the CUR t-
NN, EXACT t-NN, E-TUCKER NN and EXACT NN for solving the tensor completion problem
(10).
The result is shown in Table 6. It can be seen that the CUR t-NN outperforms almost all
its competitors in terms of CPU running time and accuracy and essentially matches that of
EXACT t-NN.
Completion Approach
EXACT NN [7]
E-TUCKER NN [33]
EXACT t-NN [53]
CUR t-NN
RSE
0.1001
0.0900
0.0715
0.0850
Time(s)
687
718
695
205
Table 6: RSE of tensor completion results for the Basketball video.
6. Conclusion. This paper introduced two randomized algorithms for basic tensor oper-
ations -rt-product and rt-project- and then used in tensor CX and CUR decomposition algo-
rithms, whose aim is to select informative slices. The randomized tensor operations together
with the tensor decompositions algorithms comes with relative error recovery guarantees. These
algorithms can be effectively used in the analysis of large scale tensors with small tubal rank.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
23
(a)
(b)
Fig. 6: The 20th frame of the tensor completion result for the Basketball video. (a) Left: The
original video. (a) Right: Sampled video (50% sampling rate). (b) Left: EXACT t-NN based
reconstruction [53]. (b) Right: CUR t-NN based reconstruction.
In addition, we proposed the CUR t-NN algorithm that exploits the advantages of ran-
domization for dimensionality reduction and can be used effectively for large scale noisy tensor
decompositions. Indeed, CUR t-NN uses an adaptive technique to sample slices of the tensor
based on a leverage score for them and subsequently solves a convex optimization problem fol-
lowed by a projection step to recover a low rank approximation to the underlying tensor. The
proposed algorithm has linear running time, and provably maintains the recovery guarantees
of the exact algorithm with full data tensor.
Acknowledgments. The authors would like to thank the Associate Editor and three
anonymous referees for many constructive comments and suggestions that improved significantly
the structure and readability of the paper.
REFERENCES
[1] M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester, Image inpainting, in Proceedings of the
27th annual conference on Computer graphics and interactive techniques, ACM Press/Addison-Wesley
Publishing Co., 2000, pp. 417 -- 424.
[2] T. Bouwmans, A. Sobral, S. Javed, S. K. Jung, and E.-H. Zahzah, Decomposition into low-rank plus
additive matrices for background/foreground separation: A review for a comparative evaluation with
a large-scale dataset, arXiv preprint arXiv:1511.01245, (2015).
[3] S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, et al., Distributed optimization and statistical
learning via the alternating direction method of multipliers, Foundations and Trends R(cid:13) in Machine
learning, 3 (2011), pp. 1 -- 122.
[4] K. Braman, Third-order tensors as linear operators on a space of matrices, Linear Algebra and its Ap-
plications, 433 (2010), pp. 1241 -- 1253.
[5] C. F. Caiafa and A. Cichocki, Generalizing the column -- row matrix decomposition to multi-way arrays,
Linear Algebra and its Applications, 433 (2010), pp. 557 -- 573.
[6] E. J. Cand`es, X. Li, Y. Ma, and J. Wright, Robust principal component analysis?, Journal of the ACM
(JACM), 58 (2011), p. 11.
[7] E. J. Cand`es and B. Recht, Exact matrix completion via convex optimization, Foundations of Compu-
tational mathematics, 9 (2009), pp. 717 -- 772.
[8] J. D. Carroll and J.-J. Chang, Analysis of individual differences in multidimensional scaling via an
n-way generalization of eckart-young decomposition, Psychometrika, 35 (1970), pp. 283 -- 319.
24
D. A. TARZANAGH AND G. MICHAILIDIS
[9] J. P. Costeira and T. Kanade, A multibody factorization method for independently moving objects,
International Journal of Computer Vision, 29 (1998), pp. 159 -- 179.
[10] L. De Lathauwer, B. De Moor, and J. Vandewalle, A multilinear singular value decomposition, SIAM
journal on Matrix Analysis and Applications, 21 (2000), pp. 1253 -- 1278.
[11] L. De Lathauwer, B. De Moor, and J. Vandewalle, On the best rank-1 and rank-(r 1, r 2,..., rn)
approximation of higher-order tensors, SIAM journal on Matrix Analysis and Applications, 21 (2000),
pp. 1324 -- 1342.
[12] A. Deshpande, L. Rademacher, S. Vempala, and G. Wang, Matrix approximation and projective clus-
tering via volume sampling, in Proceedings of the seventeenth annual ACM-SIAM symposium on
Discrete algorithm, Society for Industrial and Applied Mathematics, 2006, pp. 1117 -- 1126.
[13] P. Drineas, R. Kannan, and M. W. Mahoney, Fast monte carlo algorithms for matrices i: Approxi-
mating matrix multiplication, SIAM Journal on Computing, 36 (2006), pp. 132 -- 157.
[14] P. Drineas, M. Magdon-Ismail, M. W. Mahoney, and D. P. Woodruff, Fast approximation of matrix
coherence and statistical leverage, Journal of Machine Learning Research, 13 (2012), pp. 3475 -- 3506.
[15] P. Drineas, M. W. Mahoney, and S. Muthukrishnan, Relative-error cur matrix decompositions, SIAM
Journal on Matrix Analysis and Applications, 30 (2008), pp. 844 -- 881.
[16] P. Foggia, G. Percannella, and M. Vento, Graph matching and learning in pattern recognition in
the last 10 years, International Journal of Pattern Recognition and Artificial Intelligence, 28 (2014),
p. 1450001.
[17] A. S. Georghiades, P. N. Belhumeur, and D. J. Kriegman, From few to many: Illumination cone
models for face recognition under variable lighting and pose, IEEE transactions on pattern analysis
and machine intelligence, 23 (2001), pp. 643 -- 660.
[18] N. Hao, M. E. Kilmer, K. Braman, and R. C. Hoover, Facial recognition using tensor-tensor decom-
positions, SIAM Journal on Imaging Sciences, 6 (2013), pp. 437 -- 463.
[19] D. Hsu, S. Kakade, T. Zhang, et al., Tail inequalities for sums of random matrices that depend on the
intrinsic dimension, Electronic Communications in Probability, 17 (2012).
[20] B. Huang, C. Mu, D. Goldfarb, and J. Wright, Provable low-rank tensor recovery, Optimization-
Online, 4252 (2014).
[21] B. Jiang, F. Yang, and S. Zhang, Tensor and its tucker core: the invariance relationships, Numerical
Linear Algebra with Applications, 24 (2017), p. e2086.
[22] I. Jolliffe, Principal component analysis, Wiley Online Library, 2002.
[23] D. Kalman, A singularly valuable decomposition: the svd of a matrix, The college mathematics journal,
27 (1996), pp. 2 -- 23.
[24] E. Karahan, P. A. Rojas-Lopez, M. L. Bringas-Vega, P. A. Valdes-Hernandez, and P. A. Valdes-
Sosa, Tensor analysis and fusion of multimodal brain images, Proceedings of the IEEE, 103 (2015),
pp. 1531 -- 1559.
[25] E. Kernfeld, S. Aeron, and M. Kilmer, Clustering multi-way data: a novel algebraic approach, arXiv
preprint arXiv:1412.7056, (2014).
[26] M. E. Kilmer, K. Braman, N. Hao, and R. C. Hoover, Third-order tensors as operators on matrices:
A theoretical and computational framework with applications in imaging, SIAM Journal on Matrix
Analysis and Applications, 34 (2013), pp. 148 -- 172.
[27] M. E. Kilmer and C. D. Martin, Factorization strategies for third-order tensors, Linear Algebra and its
Applications, 435 (2011), pp. 641 -- 658.
[28] T. G. Kolda and B. W. Bader, Tensor decompositions and applications, SIAM review, 51 (2009),
pp. 455 -- 500.
[29] T. G. Kolda, B. W. Bader, and J. P. Kenny, Higher-order web link analysis using multilinear algebra,
in Data Mining, Fifth IEEE International Conference on, IEEE, 2005, pp. 8 -- pp.
[30] P. M. Kroonenberg, Three-mode principal component analysis: Theory and applications, vol. 2, DSWO
press, 1983.
[31] F. G. Kuruvilla, P. J. Park, and S. L. Schreiber, Vector algebra in the analysis of genome-wide
expression data, Genome biology, 3 (2002), pp. research0011 -- 1.
[32] G. Liu, Z. Lin, S. Yan, J. Sun, Y. Yu, and Y. Ma, Robust recovery of subspace structures by low-rank
representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 (2013), pp. 171 --
184.
[33] J. Liu, P. Musialski, P. Wonka, and J. Ye, Tensor completion for estimating missing values in visual
data, IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 (2013), pp. 208 -- 220.
[34] R. Liu, Z. Lin, S. Wei, and Z. Su, Solving principal component pursuit in linear time via l 1 filtering,
arXiv preprint arXiv:1108.5359, (2011).
[35] C. Lu, J. Feng, Y. Chen, W. Liu, Z. Lin, and S. Yan, Tensor robust principal component analysis:
Exact recovery of corrupted low-rank tensors via convex optimization, (2016), pp. 5249 -- 5257.
[36] L. W. Mackey, M. I. Jordan, and A. Talwalkar, Divide-and-conquer matrix factorization, in Advances
in Neural Information Processing Systems, 2011, pp. 1134 -- 1142.
[37] M. W. Mahoney, M. Maggioni, and P. Drineas, Tensor-cur decompositions for tensor-based data,
SIAM Journal on Matrix Analysis and Applications, 30 (2008), pp. 957 -- 987.
[38] M. Mohri and A. Talwalkar, Can matrix coherence be efficiently and accurately estimated?, in Pro-
ceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 2011,
pp. 534 -- 542.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
25
[39] Y. Mu, J. Dong, X. Yuan, and S. Yan, Accelerated low-rank visual recovery by random projection, in
Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, IEEE, 2011, pp. 2609 --
2616.
[40] A. Y. Ng, M. I. Jordan, and Y. Weiss, On spectral clustering: Analysis and an algorithm, (2002),
pp. 849 -- 856.
[41] F. Nie, H. Huang, X. Cai, and C. H. Ding, Efficient and robust feature selection via joint 2, 1-norms
minimization, in Advances in neural information processing systems, 2010, pp. 1813 -- 1821.
[42] X. Piao, Y. Hu, J. Gao, Y. Sun, and Z. Lin, A submodule clustering method for multi-way data by
sparse and low-rank representation, arXiv preprint arXiv:1601.00149, (2016).
[43] M. Rahmani and G. K. Atia, Randomized robust subspace recovery and outlier detection for high dimen-
sional data matrices, IEEE Transactions on Signal Processing, 65, pp. 1580 -- 1594.
[44] O. Semerci, N. Hao, M. E. Kilmer, and E. L. Miller, Tensor-based formulation and nuclear norm
regularization for multienergy computed tomography, IEEE Transactions on Image Processing, 23
(2014), pp. 1678 -- 1693.
[45] A. Smilde, R. Bro, and P. Geladi, Multi-way analysis: applications in the chemical sciences, John
Wiley & Sons, 2005.
[46] D. A. Tarzanagh and G. Michailidis, Estimation of graphical models through structured norm mini-
mization, Journal of Machine Learning Research, 18 (2018), pp. 1 -- 48.
[47] Y.-X. Wang, H. Xu, and C. Leng, Provable subspace clustering: When lrr meets ssc, in Advances in
Neural Information Processing Systems, 2013, pp. 64 -- 72.
[48] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma, Robust face recognition via sparse
representation, IEEE transactions on pattern analysis and machine intelligence, 31 (2009), pp. 210 --
227.
[49] Y. Xie, D. Tao, W. Zhang, and L. Zhang, Multi-view subspace clustering via relaxed l 1-norm of tensor
multi-rank, arXiv preprint arXiv:1610.07126, (2016).
[50] J. Yang, O. Rubel, M. W. Mahoney, and B. P. Bowen, Identifying important ions and positions in
mass spectrometry imaging data using cur matrix decompositions, Analytical chemistry, 87 (2015),
pp. 4658 -- 4666.
[51] J. Zhang, A. K. Saibaba, M. Kilmer, and S. Aeron, Divide-and-conquer matrix factorization, in arXiv
preprint arXiv:1609.07086, 2016, pp. 1 -- 26.
[52] Y. Zhang, Z. Jiang, and L. S. Davis, Learning structured low-rank representations for image classifi-
cation, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2013,
pp. 676 -- 683.
[53] Z. Zhang and S. Aeron, Exact tensor completion using t-svd, IEEE Transactions on Signal Processing,
65, pp. 1511 -- 1526.
[54] Z. Zhang, G. Ely, S. Aeron, N. Hao, and M. Kilmer, Novel methods for multilinear data completion
and de-noising based on tensor-svd, in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2014, pp. 3842 -- 3849.
[55] T. Zhou and D. Tao, Godec: Randomized low-rank & sparse matrix decomposition in noisy case, in
International conference on machine learning, Omnipress, 2011.
6.1. Proof of Theorem 17. Let Fn3 denote the Discrete Fourier Matrix. Then, Defini-
tion 1 implies that
(Fn3 ⊗ I)circ(A)(F −1
n3
⊗ I)(Fn3 ⊗ I)unfold(A(cid:62)
) =
(cid:62)
::1
A
A::1
,
A::2
A
(cid:62)
::2
. . .
A::n3
A::n3
(31)
= bdiagk∈[n3]( A::k
A
(cid:62)
::k).
26
D. A. TARZANAGH AND G. MICHAILIDIS
Using (31) and the unitary invariance of the Frobenius norm, we have
(cid:107)A ∗ B − C ∗ R(cid:107)2
n3
n3
F = (cid:107)(Fn3 ⊗ I)circ(A)(F −1
− (Fn3 ⊗ I)circ(C)(F −1
(cid:107) bdiagk∈n3( A::k
n3(cid:88)
=
n3(cid:88)
B::k − C::k
1
n3
1
n3
(cid:107) A::k
k=1
=
(cid:107) A::k
B::k − A::k
=
1
n3
k=1
⊗ I)(Fn3 ⊗ I)unfold(B)
⊗ I)(Fn3 ⊗ I)unfold(R)(cid:107)2
F
B::k − C::k
R::k)(cid:107)2
F ,
R::k(cid:107)2
F
S::k
D::k
D::k
S
(cid:62)
::k
B::k(cid:107)2
F .
(32)
lateral and horizontal slices are scaled by score 1/(cid:112)min{1, cpj}. Note that if min{1, cpj} = 1,
In Algorithm 2, we define Ij, j ∈ [n2] as an indicator variable, which is set to 1 if both
thej-th lateral slice of A and the j-th horizontal slice of B are selected. In this case, the selected
then Ij = 1 with probability 1, and 1 − Ij/ min{1, cpj} = 0. Then, taking expectation on both
sides of (32) and considering the set Υ = {j ∈ [n2] : cpj < 1} ⊆ [n2], we get
E(cid:2)(cid:107)A ∗ B − C ∗ R(cid:107)2
(cid:3) =
F
2
2
i1i2
A:jk
Bj:k
Bji2k
Ai1jk
1 − Ij
cpj
A:jk
Bj:k(cid:107)2
F
j∈Υ
(cid:19)
(cid:88)
(cid:18)
(cid:88)
(cid:18)
(cid:88)
(cid:88)
(cid:88)
j1∈Υ
j∈Υ
j2∈Υ
(cid:19)
(cid:19)
1 − Ij
cpj
1 − Ij
cpj
pi1i2j1j2
E [pi1i2j1j2] ,
j1∈Υ
j2∈Υ
Ai1j2k
Bj2i2k.
E(1 − Ij/cpj)2 A
2
i1jk
2
B
ji2k
B
2
ji2k
A
i2=1
i2=1
j∈Υ
j∈Υ
2
i1jk
n2(cid:88)
1
cpj
2(cid:107) Bj:k(cid:107)2
(cid:107) A:jk(cid:107)2
1
c
pj
j∈Υ
2(cid:107)Bj:1(cid:107)2
(cid:107)A:j1(cid:107)2
pj
F(cid:107)Bj::(cid:107)2
(cid:107)A:j:(cid:107)2
pj
F
2
2
.
+ ··· +
(cid:88)
j∈Υ
1
c
2(cid:107)Bj:n3(cid:107)2
(cid:107)A:jn3(cid:107)2
pj
2
=
=
=
=
1
n3
1
n3
1
n3
1
n3
1
n3
E
E
E
E
k=1
k=1
k=1
k=1
i2=1
i1=1
i1=1
i1=1
i2=1
i2=1
j∈Υ
(cid:107)(cid:88)
(cid:18)
n3(cid:88)
n1(cid:88)
n2(cid:88)
n3(cid:88)
n1(cid:88)
n2(cid:88)
n3(cid:88)
n1(cid:88)
n2(cid:88)
n3(cid:88)
n1(cid:88)
n2(cid:88)
(cid:88)
n3(cid:88)
(cid:17) Ai1j1k
n2(cid:88)
n1(cid:88)
n3(cid:88)
(cid:88)
n3(cid:88)
(cid:88)
n3(cid:88)
(cid:88)
(cid:88)
(cid:88)
n1(cid:88)
Bj1i2k
i1=1
i1=1
i2=1
i1=1
k=1
k=1
k=1
k=1
j∈Υ
j∈Υ
1
n3
1
n3
1
c
1
c
≤ 1
n3
(33)
where pi1i2j1j2 =
(cid:16)
(cid:17)(cid:16)
(cid:3) =
E(cid:2)(cid:107)A ∗ B − C ∗ R(cid:107)2
1 − Ij1
cpj1
F
1 − Ij2
cpj2
=
=
=
(34)
Since for j ∈ [Υ], E[1 − Ij/cpj] = 0 and E[(1 − Ij/cpj)2] = (1/cpj) − 1 ≤ 1/cpj, we get
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
27
Equation (15) follows from (34) by using Jensen's inequality and the fact that the sampling
probabilities (13) and (14) are defined in the original domain.
6.2. Proof of Lemma 18. Using Definitions 1 and 9, we have that
(cid:107)A ∗ A(cid:62) − C ∗ C(cid:107) = (cid:107)(Fn3 ⊗ I)circ(A)(F −1
− (Fn3 ⊗ I)circ(C)(F −1
n3
(cid:62)
::k − C::k
= (cid:107) bdiagk∈n3( A::k
A
C
(cid:62)
(cid:62)
::k(cid:107)2
::k − C::k
A
= max
k∈[n3]
(cid:8)(cid:107) A::k
C
n3
(cid:62)
::k)(cid:107)2,
(cid:9),
⊗ I)(Fn3 ⊗ I)unfold(A(cid:62)
)
⊗ I)(Fn3 ⊗ I)unfold(C(cid:62)
)(cid:107)2,
(35)
where the first equality follows from the unitary invariance of the spectral norm, the second
equality from (31) and the third equality follows since the spectral norm of a block matrix is
equal to the maximum of block norms.
Let kπ = k ∈ [n3] be the index of the tensor's frontal slice with maximum spectral norm
A
(cid:62)
::k(cid:107)2. Using (35), and Example 4.3 in [19], we obtain
(cid:62)
::k − C::k
(cid:107) A::k
C
(36)
with probability at least 1 − δ.
(cid:107) A::kπ
A
(cid:62)
::kπ
− C::kπ
C
(cid:62)
::kπ
(cid:107)2 ≤
2
(cid:107) A::,kπ(cid:107)2
2,
6.3. Proof of Lemma 21. Let σi,k be the i-th largest singular value of slice k. For all
1 ≤ i ≤ r and 1 ≤ k ≤ n3, we have
1 − max
k∈[n3]
i,k(V(cid:62) ∗ S ∗ D) = max
σ2
k∈[n3]
σi,k(V(cid:62) ∗ V) − max
k∈[n3]
σi,k(V(cid:62) ∗ S ∗ D ∗ D ∗ S(cid:62) ∗ V)
≤ (cid:107)V ∗ V(cid:62) − V(cid:62) ∗ S ∗ D ∗ D ∗ S(cid:62) ∗ V(cid:107)2.
Using Lemma 18 and (37), we get
1 − max
k∈[n3]
i,k(V(cid:62) ∗ S ∗ D) ≤ /2(cid:107) V::kπ(cid:107)2
σ2
2
for all 1 ≤ i ≤ r and an index kπ ∈ [n3].
Since ∈ (0, 1], each tubal singular value of V(cid:62)∗S is positive, which implies that rankt(V(cid:62) ∗ S) =
To prove (20b), we use the t-SVD of V(cid:62) ∗ S ∗ D and note that
rankt(V) = rankt(L).
(cid:107)Ω(cid:107) = (cid:107)(cid:16)
= (cid:107)(cid:16)
= (cid:107)(cid:0)Σ−1
= (cid:107)VV(cid:62)∗S∗D
V(cid:62) ∗ S ∗ D
UV(cid:62)∗S∗DΣV(cid:62)∗S∗DV(cid:62)
V(cid:62) ∗ S ∗ D
(cid:17)† −(cid:16)
(cid:0)Σ−1
V(cid:62)∗S∗D − ΣV(cid:62)∗S∗D
(cid:17)(cid:62)(cid:107)
(cid:17)† −(cid:16)
(cid:1) U(cid:62)
UV(cid:62)∗S∗DΣV(cid:62)∗S∗DV(cid:62)
V(cid:62)∗S∗D(cid:107)
V(cid:62)∗S∗D
V(cid:62)∗S∗D − ΣV(cid:62)∗S∗D
(cid:1)(cid:107).
(cid:17)(cid:62)(cid:107)
V(cid:62)∗S∗D
(37)
(38)
(39)
The claim follows since VV(cid:62)∗S∗D and UV(cid:62)∗S∗D are orthogonal tensors.
To prove (20c), note that
†
(L ∗ S ∗ D)
(40)
=
(cid:17)†
(cid:16)
(cid:16)
U ∗ Σ ∗ V(cid:62) ∗ S ∗ D
U ∗ Σ ∗ UV(cid:62)∗S∗D ∗ ΣV(cid:62)∗S∗D ∗ V(cid:62)
=
= VV(cid:62)∗S∗D ∗ (Σ ∗ UV(cid:62)∗S∗D ∗ ΣV(cid:62)∗S∗D)
(cid:17)†
V(cid:62)∗S∗D
† ∗ U(cid:62)
.
28
D. A. TARZANAGH AND G. MICHAILIDIS
To remove the pseudoinverse in the above derivations, we use the first part of the Lemma. In
this case,
(41)
(Σ ∗ UV(cid:62)∗S∗D ∗ ΣV(cid:62)∗S∗D)
†
= (Σ ∗ UV(cid:62)∗S∗D ∗ ΣV(cid:62)∗S∗D)
= Σ−1
V(cid:62)∗S∗DΣ−1.
V(cid:62)∗S∗D ∗ U(cid:62)
−1
By combining (40) and (41), we get the result.
To prove (20d), we have that for all 1 ≤ i ≤ r and 1 ≤ k ≤ n3,
(cid:107)Ω(cid:107) = (cid:107)Σ−1
V(cid:62)∗S∗D − ΣV(cid:62)∗S∗D(cid:107)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
σi,k(V(cid:62) ∗ S ∗ D)
= max
i,k
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)σi,k(V(cid:62) ∗ S ∗ D) −
/2(cid:112)1 − /2
σ2
i,k(V(cid:62) ∗ S ∗ D) − 1
σi,k(V(cid:62) ∗ S ∗ D)
= max
i,k
≤
√
≤ /
2
from (39)
by definition
by simple manipulation
if < 1 =>(cid:112)1 − /2 > 1/
from (37)
√
2.
6.4. Proof of Theorem 23. From Definition 14, and using (19) with β = n3/µ0(V) ∈
(0, 1], we have that
for all i ∈ [n2].
(cid:107) Vi::(cid:107)2
F =
β
rn3
β
r
(cid:107)Vi::(cid:107)2
F ≤ β
r
r
n2n3
µ0(V) =
1
n2
= pi,
Using Proposition 22, our choice of c implies that
(42)
(cid:107)A − Ac ∗ L†
c ∗ L(cid:107)F = (cid:107)A − A ∗ C ∗ D ∗ (Lc ∗ D)† ∗ L(cid:107)F ,
≤ (1 + )(cid:107)A − A ∗ L† ∗ L(cid:107)F .
holds with probability at least 0.85.
X ∈ Rn1×c×n3, it follows that
6.5. Proof of Corollary 24. Since C+ ∗ A minimizes (cid:107)A − C ∗ X(cid:107)F over all tensors
(cid:107)A − C ∗ C† ∗ A(cid:107)F ≤ (cid:107)A − C ∗ L†
C ∗ L(cid:107)F .
Now, using (42), we get
(43)
(cid:107)A − C ∗ C† ∗ A(cid:107)F ≤ (1 + )(cid:107)A − L(cid:107)F ,
with probability at least 0.85.
We can trivially boost the success probability to 1−δ by repeating Algorithm 4 O(log(1/δ))
rounds. Specifically, let Ci denote the output of Algorithm 4 at round i; using (43) for each Ci
we have
(44)
(cid:107)A − Ci ∗ C†
i ∗ A(cid:107)F ≤ (1 + )(cid:107)A − L(cid:107)F ,
with probability at least 0.85.
Now, let C denote the set of all columns used in the approximation. Since each Ci = C∗ Si
for some tensor Si and C+∗ A minimizes (cid:107)A− C∗ X(cid:107)F over all tensors X ∈ Rn1×c×n3, it follows
that
for each i. Hence, if
(cid:107)A − C ∗ C† ∗ A(cid:107)F ≤ (cid:107)A − Ci ∗ C†
i ∗ A(cid:107)F ,
(cid:107)A − C ∗ C† ∗ A(cid:107)F ≤ (1 + )(cid:107)A − L(cid:107)F ,
fails to hold, then for each i, (44) also fails to hold. Since 0.15 < 1/e, the desired conclusion
must hold with probability at least 1 − (1/e)log(1/δ) = 1 − δ.
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
29
6.6. Proof of Corollary 26. Using (42), it follows that
(45)
(cid:107)A − C ∗ U ∗ R(cid:107)F ≤ (1 + )(cid:107)A − C ∗ C† ∗ A(cid:107)F ,
with probability at least 0.85.
Using (43) and (45), our choice of c and l implies the following holds with probabilities at
least 0.7,
(cid:107)A − C ∗ U ∗ R(cid:107)F ≤ (1 + )2A − L(cid:107)F ≤ (1 + (cid:48))A − L(cid:107)F ,
(46)
where (cid:48) = 3.
The inequality (46) holds with probability at least 1 − δ by following the boosting strategy
employed in the proof of Corollary 24.
Indeed, since in each trial inequality (46) fails with
probability less than 0.3 < 1/e, the claim of Corollary 26 will hold with probability greater
than 1 − (1/e)log(1/δ) = 1 − δ.
6.7. Proof of Lemma 29. Since from Lemma 21 we have rankt(Lc) = rankt(L), the
first claim follows similarly by using Lemma 1 of [38].
To prove the second claim, using Lemma 21, assume that S(cid:62) ∗ V consists of the first c
horizontal slices of V. Then, if Lc = U ∗ Σ ∗ V(cid:62) ∗ S has rankt(Lc) = rankt(L) = r, the tensor
V(cid:62) ∗ S must have full tubal rank. Thus, we can write
c ∗ Lc = (U ∗ Σ ∗ V(cid:62) ∗ S)+ ∗ U ∗ Σ ∗ V(cid:62) ∗ S
L+
= (Σ ∗ V(cid:62) ∗ S)+ ∗ U(cid:62) ∗ U ∗ Σ ∗ V(cid:62) ∗ S
= (Σ ∗ V(cid:62) ∗ S)+ ∗ Σ ∗ V(cid:62) ∗ S
= (V(cid:62) ∗ S)+ ∗ Σ(cid:62) ∗ Σ ∗ V(cid:62) ∗ S
= (V(cid:62) ∗ S)+ ∗ V(cid:62) ∗ S
= V(cid:62) ∗ S ∗ (V(cid:62) ∗ S ∗ S(cid:62) ∗ V)−1V(cid:62) ∗ S,
where the second and third equalities follow from U(cid:62) ∗ U = Ic. The fourth and fifth equalities
result from Σ having full tubal rank and V having full lateral slice rank, and the sixth follows
from S(cid:62) ∗ V having full horizontal slice rank. Next, denote the right singular vectors of Lc by
Vc ∈ Rc×r×n3. Define ei,c as the i-th lateral slice of Ic and ei,n2 as the i-th lateral slice of I.
Then we have,
µ0(Vc) =
=
=
=
=
cn3
r
cn3
r
cn3
r
cn3
r
cn3
r
max
1≤i≤c
max
1≤i≤c
max
1≤i≤c
max
1≤i≤c
max
1≤i≤c
F
c ∗ei,c(cid:107)2
c ∗ Lc ∗ei,c}
(cid:107)V(cid:62)
(cid:62)
trace{e
i,c ∗ L+
(cid:62)
i,c ∗ (V(cid:62) ∗ S)+ ∗ V(cid:62) ∗ S ∗ei,c}
trace{e
(cid:62)
trace{e
i,c ∗ S(cid:62) ∗ V ∗ W−1 ∗ V(cid:62) ∗ S ∗ei,c}
(cid:62)
trace{e
i,n2
∗ V ∗ W−1 ∗ V(cid:62) ∗ei,n2},
where W = V(cid:62) ∗ S ∗ S(cid:62) ∗ V and the final equality follows from V(cid:62) ∗ S ∗ei,c = V(cid:62) ∗ei,n2 for all
1 ≤ i ≤ c.
Next, we have
µ0(VLC) =
cn3
r
cn3
=
r
≤ cn3
r
max
1≤i≤c
(cid:62)
trace{e
i,n2
(cid:62)
trace{W−1 ∗ V(cid:62) ∗ei,n2 ∗e
i,n2
∗ V ∗ W−1 ∗ V(cid:62) ∗ei,n2}
∗ V}
∗ V(cid:107)2(cid:126),
(cid:62)
(cid:107)V(cid:62) ∗ei,n2 ∗e
i,n2
max
1≤i≤c
(cid:107)W−1(cid:107)2 max
1≤i≤c
30
D. A. TARZANAGH AND G. MICHAILIDIS
where the last inequality follows form Holder's inequality.
(cid:62)
Since V(cid:62) ∗ei,n2 ∗e
i,n2
∗ V has tubal rank one, using the definition of µ0-coherence, we have
µ0(VLC ) ≤ c
n2
Now, using (38), we have that (cid:107)W−1(cid:107)2 ≤ n2
(cid:107)W−1(cid:107)2µ0(V).
c(1−/2) . Thus, it follows that
µ0(VLC) ≤ µ0(V)/(1 − /2).
To prove the last claim under Lemma 21, we note that
µ1(Lc) =
n1cn2
3
r
max
1≤i≤n1
1≤j≤c
max
1≤i≤n1
(cid:62)
(cid:107)e
i,n1
∗ Uc ∗ V(cid:62)
c ∗ej,c(cid:107)2
F
(cid:107)U(cid:62)
c ∗ei,n1(cid:107)2
F max
1≤j≤c
(cid:107)V(cid:62)
c ∗ej,c(cid:107)2
F
3
≤ n1cn2
≤
r
r
(1 − /2)
µ0(U)µ0(V).
6.8. Proof of Theorem 30. Define A(X) as the event that a tensor X is ( rµ2
1−/2 , r)-
coherent. To prove Theorem 30, let L denote the solution obtained by CUR t-NN and let L∗
be the exact solution of problems (10), (29) and (30). In Algorithm 6, define ¯L as
(cid:21)
(cid:20)C1
(cid:3), and L∗
R2
C2 L∗
22
,
¯L =
(cid:3)(cid:62)
, and R =(cid:2) R1
C2
R2
. It can easily be seen that
where C =(cid:2)C1
subtensor of L∗
(47)
(cid:107)L∗ − ¯L(cid:107)2
F ≤ (cid:107)C∗ − C(cid:107)2
F + (cid:107)R∗ − R(cid:107)2
F ,
22 ∈ R(n1−l)×(n2−c)×n3 is the bottom right
Now, define W ( L, ¯L) as the event
(48)
(cid:107) L − ¯L(cid:107)F ≤ (1 + )(cid:107)L∗ − ¯L(cid:107)F .
If W ( L, ¯L) holds, we have
(cid:107)L∗ − L(cid:107)F ≤ (cid:107)L∗ − ¯L(cid:107)F + (cid:107) ¯L − L(cid:107)F
≤ (cid:107)L∗ − ¯L(cid:107)F + (1 + )(cid:107)L∗ − ¯L(cid:107)F
≤ (2 + )(cid:107)L∗ − ¯L(cid:107)F
(cid:107)C∗ − C(cid:107)2
≤ (2 + )
F + (cid:107)R∗ − R(cid:107)2
(cid:113)
F
by the triangle inequality
from (48)
from (47).
Next, we consider all three events W ( L, ¯L), A(R∗
) with log(3/δ). Using
Lemma 29, our choice of c and l implies that A(C∗
) holds with probability at least
1 − δ/3. Since L is a t-CUR approximation to ¯L, from Corollary 26, we get that W ( L, ¯L)
holds with probability at least 1 − δ/3.
), and A(C∗
) and A(R∗
Using the union bound, it follows that
(cid:92)
(cid:92)
Pr(W ( L, ¯L)
A(C∗
)
A(R∗
)) ≥ 1 − Pr(W ( L, ¯L)c) − Pr(A(C∗
≥ 1 − δ/3 − δ/3 − δ/3
= 1 − δ.
)c) − Pr(A(R∗
)c)
FAST RANDOMIZED ALGORITHMS FOR TENSOR OPERATIONS
31
Algorithm 7 , L ← ADMM(X, λ)
Initialize: L0 = E0 = Y0 = 0, ρ = 1.1, µ0 = 1e − 3, µmax = 1e + 10, = 1e − 8.
while not converged do
• Lk+1 ← argminL (cid:107)L(cid:107)∗ + µk
• Ek+1 ← argminE λ(cid:107)E(cid:107)1 + µk
• Yk+1 = Yk + µk(Lk+1 + Ek+1 − X);
• Update µk+1 by µk+1 = min(ρµk, µmax);
• Check (cid:107)Lk+1 − Lk(cid:107)∞ ≤ , (cid:107)Ek+1 − Ek(cid:107)∞ ≤ ,(cid:107)Lk+1 + Ek+1 − X(cid:107)∞ ≤ ;
(cid:107)2
2 (cid:107)L + Ek − X + Yk
F ;
(cid:107)2
2 (cid:107)Lk+1 + E − X + Yk
F ;
µk
µk
end while
6.9. Optimization by ADMM. We provide the optimization and parameter setting
details of ADMM used in Algorithm 7 for solving a robust tensor factorization. As discussed in
[54], the updates of Lk+1 and Ek+1 have closed form solutions. It is easy to see that the main
per-iteration cost of Algorithm 7 is in the update of Lk+1, which requires computing the fft
of L and the SVD of block matrix L = fft(L) in the Fourier domain.
The next result establishes the global convergence of the ADMM for solving problem (29)
(for details on the convergence analysis, see [46]). Note that similar results hold for solving
problems (10), and (30).
Theorem 31. The sequence (Lk, Ek, Yk) generated by Algorithm 7 from any starting point
converges to a stationary point of problem (29).
|
1507.01768 | 2 | 1507 | 2015-10-13T12:49:34 | The Restricted Isometry Property of Subsampled Fourier Matrices | [
"cs.DS",
"cs.IT",
"cs.IT",
"math.PR"
] | A matrix $A \in \mathbb{C}^{q \times N}$ satisfies the restricted isometry property of order $k$ with constant $\varepsilon$ if it preserves the $\ell_2$ norm of all $k$-sparse vectors up to a factor of $1\pm \varepsilon$. We prove that a matrix $A$ obtained by randomly sampling $q = O(k \cdot \log^2 k \cdot \log N)$ rows from an $N \times N$ Fourier matrix satisfies the restricted isometry property of order $k$ with a fixed $\varepsilon$ with high probability. This improves on Rudelson and Vershynin (Comm. Pure Appl. Math., 2008), its subsequent improvements, and Bourgain (GAFA Seminar Notes, 2014). | cs.DS | cs |
The Restricted Isometry Property of Subsampled Fourier
Matrices
Ishay Haviv∗
Oded Regev†
Abstract
A matrix A ∈ Cq×N satisfies the restricted isometry property of order k with constant ε if it
preserves the ℓ2 norm of all k-sparse vectors up to a factor of 1 ± ε. We prove that a matrix
A obtained by randomly sampling q = O(k · log2 k · log N) rows from an N × N Fourier ma-
trix satisfies the restricted isometry property of order k with a fixed ε with high probability.
This improves on Rudelson and Vershynin (Comm. Pure Appl. Math., 2008), its subsequent
improvements, and Bourgain (GAFA Seminar Notes, 2014).
1 Introduction
A matrix A ∈ Cq×N satisfies the restricted isometry property of order k with constant ε > 0 if for
every k-sparse vector x ∈ CN (i.e., a vector with at most k nonzero entries), it holds that
(1 − ε) · kxk2
2 ≤ kAxk2
2 ≤ (1 + ε) · kxk2
2 .
(1)
Intuitively, this means that every k columns of A are nearly orthogonal. This notion, due to Cand`es
and Tao [11], was intensively studied during the last decade and found various applications and
connections to several areas of theoretical computer science, including sparse recovery [8, 20, 27],
coding theory [14], norm embeddings [6, 23], and computational complexity [4, 31, 25].
The original motivation for the restricted isometry property comes from the area of com-
pressed sensing. There, one wishes to compress a high-dimensional sparse vector x ∈ CN to a
vector Ax, where A ∈ Cq×N is a measurement matrix that enables reconstruction of x from Ax.
Typical goals in this context include minimizing the number of measurements q and the running
time of the reconstruction algorithm. It is known that the restricted isometry property of A, for
ε < √2 − 1, is a sufficient condition for reconstruction. In fact, it was shown in [11, 10, 9, 8] that
under this condition, reconstruction is equivalent to finding the vector of least ℓ1 norm among
all vectors that agree with the given measurements, a task that can be formulated as a linear pro-
gram [13, 16], and thus can be solved efficiently.
∗School of Computer Science, The Academic College of Tel Aviv-Yaffo, Tel Aviv 61083, Israel.
†Courant Institute of Mathematical Sciences, New York University. Supported by the Simons Collaboration on
Algorithms and Geometry and by the National Science Foundation (NSF) under Grant No. CCF-1320188. Any opinions,
findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily
reflect the views of the NSF.
1
The above application leads to the challenge of finding matrices A ∈ Cq×N that satisfy the re-
stricted isometry property and have a small number of rows q as a function of N and k. (For sim-
plicity, we ignore for now the dependence on ε.) A general lower bound of q = Ω(k · log(N/k))
is known to follow from [18] (see also [17]). Fortunately, there are matrices that match this lower
bound, e.g., random matrices whose entries are chosen independently according to the normal
distribution [12]. However, in many applications the measurement matrix cannot be chosen ar-
bitrarily but is instead given by a random sample of rows from a unitary matrix, typically the
discrete Fourier transform. This includes, for instance, various tests and experiments in medicine
and biology (e.g., MRI [28] and ultrasound imaging [21]) and applications in astronomy (e.g.,
radio telescopes [32]). An advantage of subsampled Fourier matrices is that they support fast
matrix-vector multiplication, and as such, are useful for efficient compression as well as for effi-
cient reconstruction based on iterative methods (see, e.g., [26]).
In recent years, with motivation from both theory and practice, an intensive line of research
has aimed to study the restricted isometry property of random sub-matrices of unitary matrices.
Letting A ∈ Cq×N be a (normalized) matrix whose rows are chosen uniformly and independently
from the rows of a unitary matrix M ∈ CN×N, the goal is to prove an upper bound on q for which
A is guaranteed to satisfy the restricted isometry property with high probability. Note that the fact
that the entries of every row of A are not independent makes this question much more difficult
than in the case of random matrices with independent entries.
The first upper bound on the number of rows of a subsampled Fourier matrix that satisfies
the restricted isometry property was O(k · log6 N), which was proved by Cand`es and Tao [12].
This was then improved by Rudelson and Vershynin [30] to O(k · log2 k · log(k log N) · log N) (see
also [29, 15] for a simplified analysis with better success probability). A modification of their
analysis led to an improved bound of O(k · log3 k · log N) by Cheraghchi, Guruswami, and Vel-
ingker [14], who related the problem to a question on the list-decoding rate of random linear
codes over finite fields. Interestingly, replacing the log(k log N) term in the bound of [30] by log k
was crucial for their application.1 Recently, Bourgain [7] proved a bound of O(k · log k · log2 N),
which is incomparable to those of [30, 14] (and has a worse dependence on ε; see below). We
finally mention that the best known lower bound on the number of rows is Ω(k · log N) [5].
1.1 Our Contribution
In this work, we improve the previous bounds and prove the following.
Theorem 1.1 (Simplified). Let M ∈ CN×N be a unitary matrix with entries of absolute value O(1/√N),
and let ε > 0 be a fixed constant. For some q = O(k · log2 k · log N), let A ∈ Cq×N be a matrix whose q
rows are chosen uniformly and independently from the rows of M, multiplied bypN/q. Then, with high
probability, the matrix A satisfies the restricted isometry property of order k with constant ε.
The main idea in our proof is described in Section 1.3. We arrived at the proof from our recent
work on list-decoding [19], where a baby version of the idea was used to bound the sample com-
1Note that the list-decoding result of [14] was later improved by Wootters [33] using different techniques.
2
plexity of learning the class of Fourier-sparse Boolean functions.2 Like all previous work on this
question, our proof can be seen as a careful union bound applied to a sequence of progressively
finer nets, a technique sometimes known as chaining. However, unlike the work of Rudelson
and Vershynin [30] and its improvements [14, 15], we avoid the use of Gaussian processes, the
“symmetrization process,” and Dudley’s inequality. Instead, and more in line with Bourgain’s
proof [7], we apply the chaining argument directly to the problem at hand using only elementary
arguments. It would be interesting to see if our proof can be cast in the Gaussian framework of
Rudelson and Vershynin.
We remark that the bounds obtained in the previous works [30, 14] have a multiplicative
O(ε−2) term, where a much worse term of O(ε−6) was obtained in [7].
In our proof of Theo-
rem 1.1 we nearly obtain the best known dependence on ε. For simplicity of presentation we first
prove in Section 3 our bound with a weaker multiplicative term of O(ε−4), and then, in Section 4,
we modify the analysis and decrease the dependence on ε to eO(ε−2).
1.2 Related Literature
As mentioned before, one important advantage of using subsampled Fourier matrices in com-
pressed sensing is that they support fast, in fact nearly linear time, matrix-vector multiplication.
In certain scenarios, however, one is not restricted to using subsampled Fourier matrices as the
measurement matrix. The question then is whether one can decrease the number of rows us-
ing another measurement matrix, while still keeping the near-linear multiplication time. For
k < N1/2−γ where γ > 0 is an arbitrary constant, the answer is yes: a construction with the
optimal number O(k · log N) of rows follows from works by Ailon and Chazelle [1] and Ailon and
Liberty [2] (see [6]). For general k, Nelson, Price, and Wootters [27] suggested taking subsampled
Fourier matrices and “tweaking” them by bunching together rows with random signs. Using the
Gaussian-process-based analysis of [30, 14] and introducing further techniques from [22], they
showed that with this construction one can reduce the number of rows by a logarithmic factor
to O(k · log2(k log N) · log N) while still keeping the nearly linear multiplication time. Our result
shows that the same number of rows (in fact, a slightly smaller number) can be achieved already
with the original subsampled Fourier matrices without having to use the “tweak.” A natural open
question is whether the “tweak” from [27] and their techniques can be combined with ours to
further reduce the number of rows. An improvement in the regime of parameters of k = ω(√N)
would lead to more efficient low-dimensional embeddings based on Johnson–Lindenstrauss ma-
trices (see, e.g., [1, 2, 23, 3, 27]).
2The result in [19] is weaker in two main respects. First, it is restricted to the case that Ax is in {0, 1}q. This
significantly simplifies the analysis and leads to a better bound on the number of rows of A. Second, the order of
quantifiers is switched, namely it shows that for any sparse x, a random subsampled A works with high probability,
whereas for the restricted isometry property we need to show that a random A works for all sparse x.
3
1.3 Proof Overview
Recall from Theorem 1.1 and from (1) that our goal is to prove that a matrix A given by a random
sample Q of q rows of M satisfies with high probability that for all k-sparse x, kAxk2
2. Since
M is unitary, the latter is equivalent to saying that kAxk2
2. Yet another way of expressing
this condition is as
j∈Q(cid:2)(Mx2)j(cid:3) ≈ E
2 ≈ kMxk2
j∈[N](cid:2)(Mx2)j(cid:3) ,
2 ≈ kxk2
E
i.e., that a sample Q ⊆ [N] of q coordinates of the vector Mx2 gives a good approximation to the
average of all its coordinates. Here, Mx2 refers to the vector obtained by taking the squared ab-
solute value of Mx coordinate-wise. For reasons that will become clear soon, it will be convenient
to assume without loss of generality that kxk1 = 1. With this scaling, the sparsity assumption
implies that kMxk2
2 is not too small (namely at least 1/k), and this will determine the amount of
additive error we can afford in the approximation above. This is the only way we use the sparsity
assumption.
At a high level, the proof proceeds by defining a finite set of vectors H that forms a net, i.e.,
a set satisfying that any vector Mx2 is close to one of the vectors in H. We then argue using
the Chernoff-Hoeffding bound that for any fixed vector h ∈ H, a sample of q coordinates gives a
good approximation to the average of h. Finally, we complete the proof by a union bound over all
h ∈ H.
In order to define the set H we notice that since kxk1 = 1, Mx can be seen as a weighted
average of the columns of M (possibly with signs). In other words, we can think of Mx as the
expectation of a vector-valued random variable given by a certain probability distribution over the
columns of M. Using the Chernoff-Hoeffding bound again, this implies that we can approximate
Mx well by taking the average over a small number of samples from this distribution. We then
let H be the set of all possible such averages, and a bound on the cardinality of H follows easily
(basically N raised to the number of samples). This technique is sometimes referred to as Maurey’s
empirical method.
The argument above is actually oversimplified, and carrying it out leads to rather bad bounds
on q. As a result, our proof in Section 3 is slightly more delicate. Namely, instead of just one set H,
we have a sequence of sets, H1, H2, . . ., each being responsible for approximating a different scale
of Mx2. The first set H1 approximates Mx2 on coordinates on which its value is highest; since
the value is high, we need less samples in order to approximate it well, as a result of which the set
H1 is small. The next set H2 approximates Mx2 on coordinates on which its value is somewhat
smaller, and is therefore a bigger set, and so on and so forth. The end result is that any vector
Mx2 can be approximately decomposed into a sum ∑i h(i), with h(i) ∈ Hi. To complete the proof,
we argue that a random choice of q coordinates approximates all the vectors in all the Hi well.
The reason working with several Hi leads to the better bound stated in Theorem 1.1 is this: even
though as i increases the number of vectors in Hi grows, the quality of approximation that we
need the q coordinates to provide decreases, since the value of Mx2 there is small and so errors
are less significant. It turns out that these two requirements on q balance each other perfectly,
leading to the desired bound on q.
4
Acknowledgments. We thank Afonso S. Bandeira, Mahdi Cheraghchi, Michael Kapralov, Jelani
Nelson, and Eric Price for useful discussions, and anonymous reviewers for useful comments.
2 Preliminaries
Notation. The notation x ≈ε,α y means that x ∈ [(1 − ε)y − α, (1 + ε)y + α]. For a matrix M, we
denote by M(ℓ) the ℓth column of M and define kMk∞ = maxi,j Mi,j.
The Restricted Isometry Property. The restricted isometry property is defined as follows.
Definition 2.1. We say that a matrix A ∈ Cq×N satisfies the restricted isometry property of order k
with constant ε if for every k-sparse vector x ∈ CN it holds that
(1 − ε) · kxk2
2 ≤ kAxk2
2 ≤ (1 + ε) · kxk2
2.
Chernoff-Hoeffding Bounds. We now state the Chernoff-Hoeffding bound (see, e.g., [24]) and
derive several simple corollaries that will be used extensively later.
N · ∑N
N · ∑N
Theorem 2.2. Let X1, . . . , XN be N identically distributed independent random variables in [0, a] satisfy-
ing E[Xi] = µ for all i, and denote X = 1
i=1 Xi. Then there exists a universal constant C such that
for every 0 < ε ≤ 1/2, the probability that X ≈ε,0 µ is at least 1 − 2e−C·Nµε
Corollary 2.3. Let X1, . . . , XN be N identically distributed independent random variables in [0, a] satis-
fying E[Xi] = µ for all i, and denote X = 1
i=1 Xi. Then there exists a universal constant C such that
for every 0 < ε ≤ 1/2 and α > 0, the probability that X ≈ε,α µ is at least 1 − 2e−C·Nαε/a.
Proof: If µ ≥ α
is at least 1 − 2e−C·Nαε/a. Otherwise, Theorem 2.2 for ε = α
X ≈ε,0 µ, hence X ≈0,α µ, is at least 1 − 2e−C·Nµε
Corollary 2.4. Let X1, . . . , XN be N identically distributed independent random variables in [−a, +a]
satisfying E[Xi] = µ and E[Xi] = µ for all i, and denote X = 1
i=1 Xi. Then there exists a universal
constant C such that for every 0 < ε′ ≤ 1/2 and α > 0, the probability that X ≈0,ε′· µ+α µ is at least
1 − 4e−C·Nαε′/a.
Proof: The corollary follows by applying Corollary 2.3 to max(Xi, 0) and to − min(Xi, 0).
ε then by Theorem 2.2 the probability that X ≈ε,0 µ is at least 1 − 2e−C·Nµε
2/a, and the latter is at least 1 − 2e−C·Nαε/a.
2/a, which
> ε implies that the probability that
2/a.
µ
N · ∑N
We end with the additive form of the bound, followed by an easy extension to the complex
case.
Corollary 2.5. Let X1, . . . , XN be N identically distributed independent random variables in [−a, +a]
satisfying E[Xi] = µ for all i, and denote X = 1
i=1 Xi. Then there exists a universal constant C such
that for every b > 0, the probability that X ≈0,b µ is at least 1 − 4e−C·Nb2/a2
N · ∑N
.
5
Proof: We can assume that b ≤ 2a. The corollary follows by applying Corollary 2.4 to, say, α =
3b/4 and ε′ = b/(4a).
Corollary 2.6. Let X1, . . . , XN be N identically distributed independent complex-valued random variables
satisfying Xi ≤ a and E[Xi] = µ for all i, and denote X = 1
i=1 Xi. Then there exists a universal
constant C such that for every b > 0, the probability that X ≈0,b µ is at least 1 − 8e−C·Nb2/a2
Proof: By Corollary 2.5 applied to the real and imaginary parts of the random variables X1, . . . , XN
it follows that for a universal constant C, the probability that Re(X) ≈0,b/√2
Re(µ) and Im(X) ≈0,b/√2
Im(µ) is at least 1 − 8e−C·Nb2/a2
. By triangle inequality, it follows that with such probability we
have X ≈0,b µ, as required.
N · ∑N
.
3 The Simpler Analysis
In this section we prove our result with a multiplicative term of O(ε−4) in the bound. We start
with the following theorem.
Theorem 3.1. For a sufficiently large N, a matrix M ∈ CN×N, and sufficiently small ε, η > 0, the follow-
η−1 log N · log2(1/η)), let Q be a multiset of q uniform and independent
ing holds. For some q = O(ε−3
random elements of [N]. Then, with probability 1 − 2−Ω(ε−2·log N·log(1/η)), it holds that for every x ∈ CN,
j∈Q(cid:2)(Mx)j2(cid:3) ≈ε,η·kxk2
j∈[N](cid:2)(Mx)j2(cid:3).
Throughout the proof we assume without loss of generality that the matrix M ∈ CN×N satisfies
2), and γ = η/(2t). We start by
kMk∞ = 1. For ε, η > 0, we denote t = log2(1/η), r = log2(1/ε
defining several vector sets as follows.
E
E
1·kMk2
∞
The Vector Sets Gi. For every 1 ≤ i ≤ t + r, let Gi denote the set of all vectors g(i) ∈ CN that can
be represented as
g(i) =
√2
F · ∑
(ℓ,s)∈F
(−1)s/2 · M(ℓ)
(2)
for a multiset F of O(2i · log(1/γ)) pairs in [N] × {0, 1, 2, 3}. A trivial counting argument gives the
following.
Claim 3.2. For every 1 ≤ i ≤ t + r, Gi ≤ NO(2i·log(1/γ)).
The Vector Sets Hi. For a t-tuple of vectors (g(1+r), . . . , g(t+r)) ∈ G1+r × · · · × Gt+r and for 1 ≤
i ≤ t, let Bi be the set of all j ∈ [N] for which i is the smallest index satisfying g(i+r)
≥ 2 · 2−i/2.
For such i, define the vector h(i) by
j
2 · 1j∈Bi, 9 · 2−i).
Let Hi be the set of all vectors h(i) that can be obtained in this way.
h(i)
j = min(g(i+r)
j
6
(3)
Claim 3.3. For every 1 ≤ i ≤ t, Hi ≤ NO(ε−2·2i·log(1/γ)).
Proof: Observe that every h(i) ∈ Hi is fully defined by some (g(1+r), . . . , g(i+r)) ∈ G1+r × · · · ×
Gi+r. Hence
Hi ≤ G1+r · · · Gi+r ≤ NO(log(1/γ))·(21+r+22+r+···+2i+r) ≤ NO(log(1/γ))·2i+r+1
.
Using the definition of r, the claim follows.
Lemma 3.4. For every η > 0 and some q = O(ε−3 η−1 log N · log(1/γ)), let Q be a multiset of q uniform
and independent random elements of [N]. Then, with probability 1 − 2−Ω(ε−2·log N·log(1/γ)), it holds that
for all 1 ≤ i ≤ t and h(i) ∈ Hi ,
E
j∈Q(cid:2)h(i)
j (cid:3) ≈ε, η E
j (cid:3).
j∈[N](cid:2)h(i)
Proof: Fix an 1 ≤ i ≤ t and a vector h(i) ∈ Hi, and denote µ = Ej∈[N][h(i)
applied with α = η and a = 9 · 2−i (recall that h(i)
it holds that Ej∈Q[h(i)
that the probability that some h(i) ∈ Hi does not satisfy Ej∈Q[h(i)
]. By Corollary 2.3,
j ≤ a for every j), with probability 1 − 2−Ω(2i·qε η),
] ≈ε, η µ. Using Claim 3.3, the union bound over all the vectors in Hi implies
] ≈ε, η µ is at most
j
j
j
NO(ε−2·2i·log(1/γ)) · 2−Ω(2i·qε η) ≤ 2−Ω(ε−2·2i·log N·log(1/γ)) .
We complete the proof by a union bound over i.
Approximating the Vectors Mx.
Lemma 3.5. For every vector x ∈ CN with kxk1 = 1, every multiset Q ⊆ [N], and every 1 ≤ i ≤ t + r,
there exists a vector g ∈ Gi that satisfies (Mx)j ≈0,2−i/2 gj for all but at most γ fraction of j ∈ [N] and
for all but at most γ fraction of j ∈ Q.
Proof: Observe that for every ℓ ∈ [N] there exist pℓ,0, pℓ,1, pℓ,2, pℓ,3 ≥ 0 that satisfy
3
∑
s=0
pℓ,s = xℓ
and √2 ·
3
∑
s=0
pℓ,s · (−1)s/2 = xℓ.
Notice that the assumption kxk1 = 1 implies that the numbers pℓ,s form a probability distribution.
Thus, the vector Mx can be represented as
Mx =
N
∑
ℓ=1
xℓ · M(ℓ) = √2 ·
N
∑
ℓ=1
3
∑
s=0
pℓ,s · (−1)s/2 · M(ℓ) = E
(ℓ,s)∼D
[√2 · (−1)s/2 · M(ℓ)],
where D is the distribution that assigns probability pℓ,s to the pair (ℓ, s).
7
Let F be a multiset of O(2i · log(1/γ)) independent random samples from D, and let g ∈ Gi
be the vector corresponding to F as in (2). By Corollary 2.6, applied with a = √2 (recall that
kMk∞ = 1) and b = 2−i/2, for every j ∈ [N] the probability that
(Mx)j ≈0,2−i/2 gj
(4)
is at least 1 − γ/4. It follows that the expected number of j ∈ [N] that do not satisfy (4) is at most
γN/4, so by Markov’s inequality the probability that the number of j ∈ [N] that do not satisfy (4)
is at most γN is at least 3/4. Similarly, the expected number of j ∈ Q that do not satisfy (4) is at
most γQ/4, so by Markov’s inequality, with probability at least 3/4 it holds that the number of
j ∈ Q that do not satisfy (4) is at most γQ. It follows that there exists a vector g ∈ Gi for which (4)
holds for all but at most γ fraction of j ∈ [N] and for all but at most γ fraction of j ∈ Q, as required.
Lemma 3.6. For every multiset Q ⊆ [N] and every vector x ∈ CN with kxk1 = 1 there exists a t-tuple of
vectors (h(1), . . . , h(t)) ∈ H1 × · · · × Ht for which
j (cid:3) and
1. Ej∈Q(cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) Ej∈Q(cid:2)∑t
i=1 h(i)
j (cid:3).
2. Ej∈[N](cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) Ej∈[N](cid:2)∑t
Proof: By Lemma 3.5, for every 1 ≤ i ≤ t there exists a vector g(i+r) ∈ Gi+r that satisfies
i=1 h(i)
(Mx)j ≈0,2−(i+r)/2 g(i+r)
j
(5)
for all but at most γ fraction of j ∈ [N] and for all but at most γ fraction of j ∈ Q. We say that
j ∈ [N] is good if (5) holds for every 1 ≤ i ≤ t, and otherwise that it is bad. Notice that all but at
most tγ fraction of j ∈ [N] are good and that all but at most tγ fraction of j ∈ Q are good. Let
(h(1), . . . , h(t)) and (B1, . . . , Bt) be the vectors and sets associated with (g(1+r), . . . , g(t+r)) as defined
in (3). We claim that h(1), . . . , h(t) satisfy the requirements of the lemma.
We first show that for every good j it holds that (Mx)j2 ≈3ε,9η ∑t
. To obtain it, we
i=1 h(i)
j
observe that if j ∈ Bi for some i, then
2 · 2−i/2 ≤ g(i+r)
j
≤ 3 · 2−i/2.
(6)
The lower bound follows simply from the definition of Bi. For the upper bound, which trivially
holds for i = 1, assume that i ≥ 2, and notice that the definition of Bi implies that g(i+r−1)
<
2 · 2−(i−1)/2. Using (5), and assuming that ε is sufficiently small, we obtain that
j
g(i+r)
j
≤ (Mx)j + 2−(i+r)/2 ≤ g(i+r−1)
≤ 2−i/2(23/2 + 21/2 · ε + ε) ≤ 3 · 2−i/2.
j
+ 2−(i+r−1)/2 + 2−(i+r)/2
Hence, by the upper bound in (6), for a good j ∈ Bi we have h(i)
j = g(i+r)
j
Observe that by the lower bound in (6),
2 and h(i′)
j = 0 for i′ 6= i.
(Mx)j ∈ [g(i+r)
j
− 2−(i+r)/2, g(i+r)
j
+ 2−(i+r)/2] ⊆ [(1 − ε) · g(i+r)
j
, (1 + ε) · g(i+r)
j
],
8
and that this implies that (Mx)j2 ≈3ε,0 ∑t
does not belong to any Bi, recalling that t = log2(1/η), it follows that
j
i=1 h(i)
. On the other hand, in case that j is good but
(Mx)j ≤ g(t+r)
and thus (Mx)j2 ≈0,9η 0 = ∑t
+ 2−(t+r)/2 ≤ 2 · 2−t/2 + 2−(t+r)/2 ≤ 3 · 2−t/2 ≤ 3√η,
i=1 h(i)
j
Finally, for every bad j we have
.
j
(cid:12)(cid:12)(cid:12)(Mx)j2 −
t
∑
i=1
h(i)
j
(cid:12)(cid:12)(cid:12) ≤ max(cid:16)(Mx)j2,
t
∑
i=1
h(i)
j (cid:17) ≤ 2.
Since at most tγ fraction of the elements in [N] and in Q are bad, their effect on the difference
between the expectations in the lemma can be bounded by 2tγ. By our choice of γ, this is η,
completing the proof of the lemma.
Finally, we are ready to prove Theorem 3.1.
Proof of Theorem 3.1: By Lemma 3.4, applied with η = η/(2t), a random multiset Q of size
q = O(cid:16)ε−3
η−1 log N · log2(1/η)(cid:17)
η−1 · t · log N · log(1/γ)(cid:17) = O(cid:16)ε−3
satisfies with probability 1 − 2−Ω(ε−2·log N·log(1/η)) that for all 1 ≤ i ≤ t and h(i) ∈ Hi,
j (cid:3) ,
j∈[N](cid:2)h(i)
j (cid:3) ≈ε,η/t E
j i ≈ε,η E
j∈[N]h t
j i.
j∈Q(cid:2)h(i)
j∈Qh t
in which case we also have
∑
i=1
∑
i=1
h(i)
E
E
h(i)
We show that a Q with the above property satisfies the requirement of the theorem. Let x ∈ CN
be a vector, and assume without loss of generality that kxk1 = 1. By Lemma 3.6, there exists a t-
tuple of vectors (h(1), . . . , h(t)) ∈ H1 × · · · × Ht satisfying Items 1 and 2 there. As a result,
E
j∈Q(cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) E
j∈[N](cid:2)(Mx)j2(cid:3) ,
and we are done.
3.1 The Restricted Isometry Property
Equipped with Theorem 3.1, it is easy to derive our result on the restricted isometry property (see
Definition 2.1) of random sub-matrices of unitary matrices.
Theorem 3.7. For sufficiently large N and k, a unitary matrix M ∈ CN×N satisfying kMk∞ ≤ O(1/√N),
and a sufficiently small ε > 0, the following holds. For some q = O(ε−4 · k · log2(k/ε) · log N), let
A ∈ Cq×N be a matrix whose q rows are chosen uniformly and independently from the rows of M, mul-
tiplied bypN/q. Then, with probability 1 − 2−Ω(ε−2·log N·log(k/ε)), the matrix A satisfies the restricted
isometry property of order k with constant ε.
9
Proof: Let Q be a multiset of q uniform and independent random elements of [N], defining a
matrix A as above. Notice that by the Cauchy-Schwarz inequality, any k-sparse vector x ∈ CN
√k. Applying Theorem 3.1 with ε/2 and some η = Ω(ε/k), we get
with kxk2 = 1 satisfies kxk1 ≤
that with probability 1 − 2−Ω(ε−2·log N·log(k/ε)), it holds that for every x ∈ CN with kxk2 = 1,
j∈Q(cid:2)(Mx)j2(cid:3) ≈ε/2,ε/2 N · E
It follows that every vector x ∈ CN satisfies kAxk2
isometry property of order k with constant ε.
j∈[N](cid:2)(Mx)j2(cid:3) = kMxk2
2 ≈ε,0 kxk2
2, hence A satisfies the restricted
2 = N · E
kAxk2
2 = 1 .
4 The Improved Analysis
In this section we prove the following theorem, which improves the bound of Theorem 3.1 in
terms of the dependence on ε.
Theorem 4.1. For a sufficiently large N, a matrix M ∈ CN×N, and sufficiently small ε, η > 0, the
following holds. For some q = O(log2(1/ε) · ε−1
η−1 log N · log2(1/η)), let Q be a multiset of q uniform
and independent random elements of [N]. Then, with probability 1 − 2−Ω(log N·log(1/η)), it holds that for
every x ∈ CN,
E
j∈Q(cid:2)(Mx)j2(cid:3) ≈ε,η·kxk2
1·kMk2
∞
E
j∈[N](cid:2)(Mx)j2(cid:3).
(7)
We can assume that ε ≥ η, as otherwise, one can apply the theorem with parameters η/2, η/2
1 · kMk2
and derive (7) for ε, η as well (because the right-hand size is bounded from above by kxk2
∞).
As before, we assume without loss of generality that kMk∞ = 1. For ε ≥ η > 0, we define t =
2). For the analysis given in this section, we define γ = η/(60(t + r)).
log2(1/η) and r = log2(1/ε
Throughout the proof, we use the vector sets Gi from Section 3 and Lemma 3.5 for this value of γ.
The Vector Sets Di,m. For a (t + r)-tuple of vectors (g(1), . . . , g(t+r)) ∈ G1 × · · · × Gt+r and for
1 ≤ i ≤ t, let Ci be the set of all j ∈ [N] for which i is the smallest index satisfying g(i)
≥ 2 · 2−i/2.
For m = i, . . . , i + r define the vector h(i,m) by
j
and for other values of m define h(i,m) = 0. Now, for every m, let ∆(i,m) be the vector defined by
h(i,m)
j
= g(m)
j
2 · 1j∈Ci,
(8)
∆(i,m)
j
=( h(i,m)
j
0,
− h(i,m−1)
j
,
if h(i,m)
j
− h(i,m−1)
j
otherwise.
≤ 30 · 2−(i+m)/2;
(9)
Note that the support of ∆(i,m) is contained in Ci. Let Di,m be the set of all vectors ∆(i,m) that can be
obtained in this way.
Claim 4.2. For every 1 ≤ i ≤ t and i ≤ m ≤ i + r, Di,m ≤ NO(2m·log(1/γ)).
10
Proof: Observe that every vector in Di,m is fully defined by some (g(1), . . . , g(m)) ∈ G1 × · · · × Gm.
Hence
Di,m ≤ G1 · · · Gm ≤ NO(log(1/γ))·(21+22+···+2m) ≤ NO(log(1/γ))·2m+1
,
and the claim follows.
Lemma 4.3. For every ε, η > 0 and some q = O(ε−1 η−1 log N · log(1/γ)), let Q be a multiset of q
uniform and independent random elements of [N]. Then, with probability 1 − 2−Ω(log N·log(1/γ)), it holds
that for every 1 ≤ i ≤ t, m, and a vector ∆(i,m) ∈ Di,m associated with a set Ci,
E
j∈Q(cid:2)∆(i,m)
j
(cid:3) ≈0,b E
j∈[N](cid:2)∆(i,m)
j
(cid:3) for b = O(cid:16)ε · 2−i · Ci
N
+ η(cid:17) .
(10)
Proof: Fix i, m, and a vector ∆(i,m) ∈ Di,m associated with a set Ci as in (9). Notice that
[∆(i,m)
j
] ≤ 30 · 2−(i+m)/2 · Ci
N
.
E
j∈[N]
By Corollary 2.4, applied with
ε′ = ε · 2(m−i)/2, α = η, and a = 30 · 2−(i+m)/2,
we have that (10) holds with probability 1 − 2−Ω(2m·qε η). Using Claim 4.2, the union bound over
all the vectors in Di,m implies that the probability that some ∆(i,m) ∈ Di,m does not satisfy (10) is at
most
The result follows by a union bound over i and m.
NO(2m·log(1/γ)) · 2−Ω(2m·qε η) ≤ 2−Ω(2m·log N·log(1/γ)) .
Approximating the Vectors Mx.
Lemma 4.4. For every multiset Q ⊆ [N] and every vector x ∈ CN with kxk1 = 1 there exist vector
collections (∆(i,m) ∈ Di,m)m=i,...,i+r associated with sets Ci (1 ≤ i ≤ t), for which
1. Ej∈[N](cid:2)(Mx)j2(cid:3) ≥ ∑t
i=1 2−i · CiN − η,
2. Ej∈Q(cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) Ej∈Q(cid:2)∑t
3. Ej∈[N](cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) Ej∈[N](cid:2)∑t
Proof: By Lemma 3.5, for every 1 ≤ i ≤ t + r there exists a vector g(i) ∈ Gi that satisfies
i=1 ∑i+r
m=i
∆(i,m)
j
∆(i,m)
i=1 ∑i+r
m=i
(cid:3), and
(cid:3).
(Mx)j ≈0,2−i/2 g(i)
j
j
(11)
for all but at most γ fraction of j ∈ [N] and for all but at most γ fraction of j ∈ Q. We say that
j ∈ [N] is good if (11) holds for every i, and otherwise that it is bad. Notice that all but at most
11
(t + r)γ fraction of j ∈ [N] are good and that all but at most (t + r)γ fraction of j ∈ Q are good.
Consider the sets Ci and vectors h(i,m), ∆(i,m) associated with (g(1), . . . , g(t+r)) as defined in (8). We
claim that ∆(i,m) satisfy the requirements of the lemma.
Fix some 1 ≤ i ≤ t. For every good j ∈ Ci, the definition of Ci implies that g(i)
j
using (11) it follows that
(Mx)j ≥ g(i)
j
− 2−i/2 ≥ 2−i/2.
≥ 2 · 2−i/2, so
(12)
We also claim that (Mx)j ≤ 3 · 2−(i−1)/2. This trivially holds for i = 1, so assume that i ≥ 2, and
notice that the definition of Ci implies that g(i−1)
< 2 · 2−(i−1)/2, so using (11), it follows that
j
(Mx)j ≤ g(i−1)
j
+ 2−(i−1)/2 ≤ 3 · 2−(i−1)/2.
Since at most (t + r)γ fraction of j ∈ [N] are bad, (12) yields that
E
j∈[N](cid:2)(Mx)j2(cid:3) ≥
t
∑
i=1
2−i · Ci
N − (t + r)γ/2 ≥
t
∑
i=1
2−i · Ci
N − η,
as required for Item 1.
Next, we claim that every good j satisfies
(Mx)j2 ≈O(ε),O(η)
t
∑
i=1
h(i,i+r)
j
.
(13)
(14)
(15)
For a good j ∈ Ci and m ≥ i,
(cid:12)(cid:12)(Mx)j2 − h(i,m)
j
(cid:12)(cid:12) ≤ 2 · (Mx)j · 2−m/2 + 2−m ≤ 10 · 2−(i+m)/2,
where the first inequality follows from (11) and the second from (13). In particular, for m = i + r
(recall that r = log2(1/ε
2)), we have
j
j
(cid:12)(cid:12)(Mx)j2 − h(i,i+r)
and thus (Mx)j2 ≈O(ε),0 h(i,i+r)
good j ∈ S Ci we have (Mx)j2 ≈O(ε),0 ∑t
(Mx)j ≤ g(t)
i=1 h(i,i+r)
Next, we claim that for every good j,
and thus (Mx)j2 ≈0,9η 0 = ∑t
belong to any Ci, by our choice of t, it satisfies
j
j
(cid:12)(cid:12) ≤ 10 · ε · 2−i ≤ 10 · ε · (Mx)j2 ,
. Since every good j belongs to at most one of the sets Ci, for every
i=1 h(i,i+r)
j
. On the other hand, if j is good but does not
+ 2−t/2 ≤ 3 · 2−t/2 = 3√η ,
. This establishes that (14) holds for every good j.
(Mx)j2 ≈O(ε),O(η)
t
∑
i=1
i+r
∑
m=i
∆(i,m)
j
.
(16)
12
This follows since for every 1 ≤ i ≤ t, the vector h(i,i+r) can be written as the telescopic sum
where we used that h(i,i−1) = 0. We claim that for every good j, these differences satisfy
h(i,i+r) =
i+r
∑
m=i(cid:0)h(i,m) − h(i,m−1)(cid:1) ,
≤ 30 · 2−(i+m)/2,
h(i,m)
j
− h(i,m−1)
j
thus establishing that (16) holds for every good j. Indeed, for m ≥ i + 1, (15) implies that
h(i,m)
j
− h(i,m−1)
j
≤ 10 · (2−(i+m)/2 + 2−(i+m−1)/2) ≤ 30 · 2−(i+m)/2,
(17)
and for m = i it follows from (11) combined with (13).
Finally, for every bad j we have
(cid:12)(cid:12)(cid:12)(Mx)j2 −
t
∑
i=1
i+r
∑
m=i
∆(i,m)
j
(cid:12)(cid:12)(cid:12) ≤ 1 + 30 · max
1≤i≤t(cid:16) i+r
∑
m=i
2−(i+m)/2(cid:17) ≤ 60 .
Since at most (t + r)γ fraction of the elements in [N] and in Q are bad, their effect on the difference
between the expectations in Items 2 and 3 can be bounded by 60(t + r)γ. By our choice of γ this is
η, as required.
Finally, we are ready to prove Theorem 4.1.
Proof of Theorem 4.1: Recall that it can be assumed that ε ≥ η. By Lemma 4.3, applied with
ε = ε/r and η = η/(rt), a random multiset Q of size
η−1 · r2 · t · log N · log(1/γ)(cid:17)
q = O(cid:16)ε−1
= O(cid:16) log2(1/ε) · ε−1
η−1 log N · log2(1/η)(cid:17)
satisfies with probability 1 − 2−Ω(log N·log(1/η)), that for every 1 ≤ i ≤ t, m, and ∆(i,m) ∈ Di,m
associated with a set Ci,
in which case we also have
E
j∈Q(cid:2)∆(i,m)
j
j
E
(cid:3) ≈0,bi
j∈[N](cid:2)∆(i,m)
i ≈0,b E
j∈[N]h t
i+r
∑
m=i
∑
i=1
(cid:3) for bi = O(cid:16) ε
r · 2−i · Ci
N
∆(i,m)
j
i for b = O(cid:16)ε ·
t
∑
i=1
η
+
rt(cid:17),
2−i · Ci
N
+ η(cid:17) .
(18)
E
j∈Qh t
∑
i=1
i+r
∑
m=i
∆(i,m)
j
We show that a Q with the above property satisfies the requirement of the theorem. Let x ∈ CN
be a vector, and assume without loss of generality that kxk1 = 1. By Lemma 4.4, there exist vector
collections (∆(i,m) ∈ Di,m)m=i,...,i+r associated with sets Ci (1 ≤ i ≤ t), satisfying Items 1, 2, and 3
there. Combined with (18), this gives
and we are done.
E
j∈Q(cid:2)(Mx)j2(cid:3) ≈O(ε),O(η) E
j∈[N](cid:2)(Mx)j2(cid:3) ,
13
4.1 The Restricted Isometry Property
It is easy to derive now the following theorem. The proof is essentially identical to that of Theo-
rem 3.7, using Theorem 4.1 instead of Theorem 3.1.
Theorem 4.5. For sufficiently large N and k, a unitary matrix M ∈ CN×N satisfying kMk∞ ≤ O(1/√N),
and a sufficiently small ε > 0, the following holds. For some q = O(log2(1/ε)ε−2 · k · log2(k/ε) · log N),
let A ∈ Cq×N be a matrix whose q rows are chosen uniformly and independently from the rows of M,
multiplied bypN/q. Then, with probability 1 − 2−Ω(log N·log(k/ε)), the matrix A satisfies the restricted
isometry property of order k with constant ε.
References
[1] N. Ailon and B. Chazelle. The fast Johnson–Lindenstrauss transform and approximate near-
est neighbors. SIAM J. Comput., 39(1):302–322, 2009. Preliminary version in STOC’06.
[2] N. Ailon and E. Liberty. Fast dimension reduction using Rademacher series on dual BCH
codes. Discrete & Computational Geometry, 42(4):615–630, 2009. Preliminary version in
SODA’08.
[3] N. Ailon and E. Liberty. An almost optimal unrestricted fast Johnson–Lindenstrauss trans-
form. ACM Transactions on Algorithms, 9(3):21, 2013. Preliminary version in SODA’11.
[4] A. S. Bandeira, E. Dobriban, D. G. Mixon, and W. F. Sawin. Certifying the restricted isometry
property is hard. IEEE Transactions on Information Theory, 59(6):3448–3450, 2013.
[5] A. S. Bandeira, M. E. Lewis, and D. G. Mixon. Discrete uncertainty principles and sparse
signal processing. CoRR, abs/1504.01014, 2015.
[6] R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin. A simple proof of the restricted isom-
etry property for random matrices. Constructive Approximation, 28(3):253–263, 2008.
[7] J. Bourgain. An improved estimate in the restricted isometry problem. In Geometric Aspects of
Functional Analysis, volume 2116 of Lecture Notes in Mathematics, pages 65–70. Springer, 2014.
[8] E. J. Cand`es. The restricted isometry property and its implications for compressed sensing.
Comptes Rendus Mathematique, 346(9-10):589–592, 2008.
[9] E. J. Cand`es, J. K. Romberg, and T. Tao. Stable signal recovery from incomplete and inaccurate
measurements. Comm. Pure Appl. Math., 59(8):1207–1223, 2006.
[10] E. J. Cand`es, M. Rudelson, T. Tao, and R. Vershynin. Error correction via linear programming.
In 46th Annual IEEE Symposium on Foundations of Computer Science, FOCS, pages 295–308,
2005.
[11] E. J. Cand`es and T. Tao. Decoding by linear programming. IEEE Transactions on Information
Theory, 51(12):4203–4215, 2005.
14
[12] E. J. Cand`es and T. Tao. Near-optimal signal recovery from random projections: Universal
encoding strategies? IEEE Trans. on Information Theory, 52(12):5406–5425, 2006.
[13] S. S. Chen, D. L. Donoho, and M. A. Saunders. Atomic decomposition by basis pursuit. SIAM
J. Comput., 20(1):33–61, 1998.
[14] M. Cheraghchi, V. Guruswami, and A. Velingker. Restricted isometry of Fourier matrices and
list decodability of random linear codes. SIAM J. Comput., 42(5):1888–1914, 2013. Preliminary
version in SODA’13.
[15] S. Dirksen. Tail bounds via generic chaining. Electron. J. Probab., 20(53):1–29, 2015.
[16] D. L. Donoho, M. Elad, and V. N. Temlyakov. Stable recovery of sparse overcomplete repre-
sentations in the presence of noise. IEEE Transactions on Information Theory, 52(1):6–18, 2006.
[17] S. Foucart, A. Pajor, H. Rauhut, and T. Ullrich. The Gelfand widths of ℓp-balls for 0 < p ≤ 1.
J. Complexity, 26(6):629–640, 2010.
[18] A. Y. Garnaev and E. D. Gluskin. On the widths of Euclidean balls. Soviet Mathematics Doklady,
30:200–203, 1984.
[19] I. Haviv and O. Regev. The list-decoding size of Fourier-sparse boolean functions. In Proceed-
ings of the 30th Conference on Computational Complexity, CCC, pages 58–71, 2015.
[20] P. Indyk and I. Razenshteyn. On model-based RIP-1 matrices. In Automata, Languages, and
Programming - 40th International Colloquium, ICALP, pages 564–575, 2013.
[21] A. C. Kak and M. Slaney. Principles of Computerized Tomographic Imaging. Society of Industrial
and Applied Mathematics, 2001.
[22] F. Krahmer, S. Mendelson, and H. Rauhut. Suprema of chaos processes and the restricted
isometry property. CoRR, abs/1207.0235, 2012.
[23] F. Krahmer and R. Ward. New and improved Johnson-Lindenstrauss embeddings via the
restricted isometry property. SIAM J. Math. Analysis, 43(3):1269–1281, 2011.
[24] C. McDiarmid. Concentration.
In Probabilistic methods for algorithmic discrete mathematics,
volume 16 of Algorithms Combin., pages 195–248. Springer, Berlin, 1998.
[25] A. Natarajan and Y. Wu. Computational complexity of certifying restricted isometry property.
In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques,
APPROX, pages 371–380, 2014.
[26] D. Needell and J. A. Tropp. CoSaMP: iterative signal recovery from incomplete and inaccurate
samples. Commun. ACM, 53(12):93–100, 2010.
15
[27] J. Nelson, E. Price, and M. Wootters. New constructions of RIP matrices with fast multipli-
cation and fewer rows. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA, pages 1515–1528, 2014.
[28] D. G. Nishimura. Principles of Magnetic Resonance Imaging. Stanford University, 2010.
[29] H. Rauhut. Compressive sensing and structured random matrices.
In M. Fornasier, edi-
tor, Theoretical foundations and numerical methods for sparse recovery, volume 9, pages 1–92. De
Gruyter, 2010.
[30] M. Rudelson and R. Vershynin. On sparse reconstruction from Fourier and Gaussian mea-
surements. Comm. Pure Appl. Math., 61(8):1025–1045, 2008. Preliminary version in CISS’06.
[31] A. M. Tillmann and M. E. Pfetsch. The computational complexity of the restricted isometry
property, the nullspace property, and related concepts in compressed sensing. IEEE Transac-
tions on Information Theory, 60(2):1248–1259, 2014.
[32] S. Wenger, S. Darabi, P. Sen, K. Glassmeier, and M. A. Magnor. Compressed sensing for
aperture synthesis imaging. In Proceedings of the International Conference on Image Processing,
ICIP, pages 1381–1384, 2010.
[33] M. Wootters. On the list decodability of random linear codes with large error rates. In Pro-
ceedings of the 45th Annual ACM Symposium on Theory of Computing, STOC, pages 853–860,
2013.
16
|
1207.4127 | 1 | 1207 | 2012-07-11T14:47:24 | On finding minimal w-cutset | [
"cs.DS",
"cs.AI"
] | The complexity of a reasoning task over a graphical model is tied to the induced width of the underlying graph. It is well-known that the conditioning (assigning values) on a subset of variables yields a subproblem of the reduced complexity where instantiated variables are removed. If the assigned variables constitute a cycle-cutset, the rest of the network is singly-connected and therefore can be solved by linear propagation algorithms. A w-cutset is a generalization of a cycle-cutset defined as a subset of nodes such that the subgraph with cutset nodes removed has induced-width of w or less. In this paper we address the problem of finding a minimal w-cutset in a graph. We relate the problem to that of finding the minimal w-cutset of a treedecomposition. The latter can be mapped to the well-known set multi-cover problem. This relationship yields a proof of NP-completeness on one hand and a greedy algorithm for finding a w-cutset of a tree decomposition on the other. Empirical evaluation of the algorithms is presented. | cs.DS | cs | On finding minimal w-cutset
Bozhena Bidyuk and Rina Dechter
Information and Computer Science
University Of California Irvine
Irvine, CA 92697-3425
Abstract
The complexity of a reasoning task over a graph-
ical model is tied to the induced width of the un-
derlying graph.
It is well-known that the con-
ditioning (assigning values) on a subset of vari-
ables yields a subproblem of the reduced com-
plexity where instantiated variables are removed.
If the assigned variables constitute a cycle-cutset,
the rest of the network is singly-connected and
therefore can be solved by linear propagation al-
gorithms. A w-cutset is a generalization of a
cycle-cutset defined as a subset of nodes such
that the subgraph with cutset nodes removed has
In this paper we
induced-width of w or less.
address the problem of finding
a minimal w-
cutset in a graph. We relate the problem to
that of finding the minimal w-cutset of a tree-
decomposition. The latter can be mapped to the
well-known set multi-cover problem. This rela-
tionship yields a proof of NP-completeness on
one hand and a greedy algorithm for finding a w-
cutset of a tree decomposition on the other. Em-
pirical evaluation of the algorithms is presented.
1 Introduction
A cycle-cutset of an undirected graph is a subset of nodes
that, when removed, the graph is cycle-free. Thus, if
the assigned variables constitute a cycle-cutset, the rest of
the network is singly-connected and can be solved by lin-
ear propagation algorithms. This principle is at heart of
the well-known cycle-cutset conditioning algorithms for
Bayesian networks [16] and for constraint networks [7].
Recently, the idea of cutset-conditioning was extended to
accommodate search on any subset of variables using the
notion of w-cutset, yielding a hybrid algorithmic scheme
of conditioning and inference paramterized by w [18]. The
w-cutset is defined as a subset of nodes in the graph that,
once removed, the graph has tree-width of w or less.
The hybrid w-cutset-conditioning algorithm applies search
to the cutset variables and exact inference (e.g., bucket
elimination [8]) to the remaining network. Given a w-
cutset Cw , the algorithm is space exponential in w and
time exponential in w + jCw j [9]. The scheme was ap-
plied successfully in the context of satisfiability [18] and
constraint optimization [13]. More recently, the notion of
conditioning was explored for speeding up sampling al-
gorithms in Bayesian networks in a scheme called cutset-
sampling. The idea is to restrict sampling to w-cutset vari-
ables only (peform inference on the rest) and thus reduce
the sampling variance ([5, 6]).
Since the processing time of both search-based and
sampling-based schemes grows with the size of the w-
cutset, it calls for a secondary optimization task for finding
a minimal-size w-cutset. Also, of interest is the task of find-
ing the full sequence of minimal w-cutsets, where w ranges
from 1 to the problem’s induced-width (or tree-width), so
that the user can select the w that fits his/her resources. We
call the former the w-cutset problem and the latter the se-
quence w-cutset problem. The w-cutset problem extends
the task of finding minimum cycle-cutset (e.g. a 1-cutset),
a problem that received fair amount of attention [3, 2, 20].
The paper addresses the minimum size w-cutset and, more
generally, the minimum weight w-cutset problem. First, we
relate the size of a w-cutset of a graph to its tree-width and
the properties of its tree-decompositions. Then, we prove
that the problem of finding a minimal w-cutset of a given
tree decomposition is NP-complete by reduction from the
set multi-cover problem [20]. Consequently, we apply a
well-known greedy algorithm (GWC) for the set multi-
cover problem to solve the minimum w-cutset problem.
The algorithm finds w-cutset within O(1+ ln m) of optimal
where m is the maximum number of clusters of size greater
than w + 1 sharing the same variable in the input tree de-
composition. We investigate its performance empirically
and show that, with rare exceptions, GWC and its vari-
ants find a smaller w-cutset than the well-performing MGA
cycle-cutset algorithm [3] (adapted to the w-cutset prob-
lem) and a w-cutset algorithm (DGR) proposed in [11].
UAI 2004BIDYUK & DECHTER43 2 Preliminaries and Background
2.1 General graph concepts
We focus on automated reasoning problems R=<X ,F >
where X is a set of variables and F is a set of functions
over subsets of the variables. Such problems are associated
with a graphical description and thus also called graphical
models. The primary examples are constraint networks and
Bayesian or belief networks formally defined in [16]. The
structure of a reasoning problem can be depicted via several
graph representations.
DE FIN IT ION 2.1 (Primal-, dual-,hyper-graph of a probem)
The primal graph G=<X ,E>of a reasoning problem
<X ,F > has the variables X as its nodes and an arc
connects two nodes if they appear in the scope of the same
function f 2 F . A dual graph of a reasoning problem
has the scopes of the functions as the nodes and an arc
connects two nodes if the corresponding scopes share a
variable. The arcs are labelled by the shared variables.
The hypergraph of a reasoning problem has the variables
X as nodes and the scopes as edges. There is a one-to-one
correspondance between the hypergraph and the dual
graphs of a problem.
We next describe the notion of a tree-decomposition [10,
14]. It is applicable to any grapical model since it is defined
relative to its hypergraph or dual graph.
DE FIN IT ION 2.2 (tree-decomp., cluster-tree, tree-width)
Let R =< X; D; F > be a reasoning problem with its
hypergraph H = (X; F )). (We abuse notation when we
identify a function with its scope). A tree-decomposition
for R (resp., its hypergraph H) is a triple < T ; (cid:31); >,
where T =<V ; E> is a tree, and (cid:31) and are labeling
functions which associate with each vertex v 2 V two sets,
(cid:31)(v) (cid:18) X and (v) (cid:18) F such that
1. For each function fi 2 F , there is exactly one vertex
v 2 V such that fi 2 (v), and scope(fi ) (cid:18) (cid:31)(v).
2. For each variable Xi 2 X , the set fv 2 V jXi 2
(cid:31)(v)g induces a connected subtree of T . This is also
called the running intersection property.
We will often refer to a node and its functions as a cluster
and use the term tree-decomposition and cluster tree inter-
changeably. The maximum size of node (cid:31)i minus 1 is the
width of the tree decomposition. The tree width of a graph
G denoted tw(G) is the minimum width over all possible
tree decompositions of G [1].The tree-width of a graph is
also referred to as induced width of a graph. We sometime
denote the optimal tree-width of a graph by tw(cid:3).
There is a known relationship between tree-decompositions
and chordal graphs. A graph is chordal if any cycle of
length 4 or more has a chord. Every tree-decomposition
of a graph corresponds to a chordal graph that augments
the input graph where the tree clusters are the cliques in the
chordal graph. And vice-versa, any chordal graph is a tree-
decomposition where the clusters are the maximal cliques.
Therefore, much of the discussion that follows, relating to
tree-decompositions of graphs, can be rephrased using the
notion of chordal graph embeddings.
2.2 w-cutset of a graph
DE FIN IT ION 2.3 (w-cutset of a graph) Given a graph
G=<X; E>, Cw (cid:26) X is a w-cutset of G if the subgraph
over X nC has tree-width (cid:20) w. Clearly, a w-cutset is also
a w 0 -cutset when w 0 (cid:21) w. The cutset Cw is minimal if no
w-cutset of smaller size exists.
jD(Ci )j
For completeness, we also define
the weighted w-cutset
problem that generalizes minimum w-cutset problem
(where all node weights are assumed the same). For ex-
ample, in w-cutset conditioning, the space requirements
max ) where dmax is the max-
of exact inference are O(dw
imum node domain size in graph G. The total time re-
quired to condition on w-cutset C is O(N (cid:1) dw
max ) (cid:2) jD(C )j
where jD(C )j is the size of the cutset domain space and N
is the number of variables. The upper bound value d jC j
max
on jD(C )j produces a bound on the computation time of
max ) (cid:2) djC j
the cutset-conditioning algorithm: O(N (cid:1) dw
max =
O(N (cid:1) dw+jC j
max ). In this case, clearly, we want to minimize
the size of C. However, a more refined optimization task is
to minimize the actual value of jD(C )j:
jD(C )j = YCi2C
Since the minimum of jD(C )j corresponds to the minimum
of lg jD(C )j, we can solve this optimization task by assign-
ing each node Xi cost ci = lg jD(Xi )j and minimizing the
cost of cutset:
cost(C ) = lg jD(C )j = XCi2C
lg jD(Xi )j = Xi
Similar considerations apply in case of the w-cutset sam-
pling algorithm. Here, the space requirements for the exact
inference are the same. The time required to sample a node
Ci 2 C is O(dw
max ) (cid:2) jD(Ci )j. The total sampling time
max ) (cid:2) PCi2C jD(Ci )j. To minimize the total pro-
is O(dw
cessing time, we assign each node Xi cost ci = jD(Xi )j
and select the w-cutset of minimum cost:
cost(C ) = XCi2C
jD(Ci )j
DE FIN IT ION 2.4 (weighted w-cutset of a graph) Given a
reasoning problem <X; F > where each node Xi 2 X has
associated cost(Xi ) = ci , the cost of a w-cutset Cw is
given by: cost(Cw ) = PXi 2Cw ci . The minimum weight
w-cutset problem is to find a min-cost w-cutset.
ci
44BIDYUK & DECHTERUAI 2004 In practice, we can often assume that all nodes have the
same cost and solve the easier minimal w-cutset problem.
In section 3, we establish relations between the size of w-
cutset of a graph and the width of its tree-decomposition.
In section 4, we show that the problem is NP-hard even
when finding a minimum w-cutset of a chordal graph (cor-
responding to a tree-decomposition of a graph).
A
A
ACE DEF
A
B
D
C
E
B
D
C
E
F
(a)
F
(b)
ADE
ABD
B
C
E
F
(c)
AB
AC
CE
EF
3 w-Cutset and Tree-Decompositions
In this section, we explore relationship between w-cutset of
a graph and its tree-decomposition.
THEOREM 3.1 Given a graph G=<X; E>, if G has a w-
cutset Cw , then there is a tree-decomposition of G having
a tree-width tw (cid:20) jCw j + w.
Proof.
If there exists a w-cutset Cw , then we can remove
Cw from the graph yielding, by definition, a subgraph G 0
over X nCw that has a tree decomposition T with clusters
of size at most w+1. We can add the set Cw to each cluster
of T yielding a tree-decomposition with clusters of size at
most w + 1 + jCw j and tree-width w + jCw j. ut
We can conclude therefore that for any graph tw (cid:3) (cid:20) jCi j+i
for every i. Moreover,
THEOREM 3.2 Given a graph G, if c(cid:3)
i is the size of a small-
i , and tw(cid:3) is its tree-width, then:
est i-cutset C (cid:3)
c(cid:3)
1 + 1 (cid:21) c(cid:3)
2 + 2 (cid:21) ::: (cid:21) c(cid:3)
i + i (cid:21) ::: (cid:21) tw(cid:3)
(1)
Proof. Let us define (cid:1)i;i+1 = c(cid:3)
i+1 , then we claim
i (cid:0) c(cid:3)
that (cid:1)i;i+1 (cid:21) 1. Assume to the contrary that ci = ci+1 ,
that is (cid:1)i;i+1 =0. Since C (cid:3)
i is an i-cutset, we can build a tree
decomposition T with maximum cluster size (i+1). Note,
i j > 0 or we would have tw(cid:3) = i.
for all i < tw(cid:3) , jC (cid:3)
Pick some Xj 2 C (cid:3)
i and add Xj to every cluster yielding
tree decomposition T 0 with maximum cluster size (i+2).
i nXj is an (i+1)-cutset of size c(cid:3)
Clearly, C (cid:3)
i (cid:0) 1 = c(cid:3)
i+1 (cid:0) 1
i+1 . ut
which contradicts the minimality of C (cid:3)
Given a graph G = (V ; E ), the w-cutset sequence problem
seeks a sequence of minimal j -cutsets where j ranges from
j ; :::; Ctw(cid:3) = (cid:30). Let
1 to the graph’s tree-width: C (cid:3)
1 ; :::; C (cid:3)
w be a subset-minimal w-cutset, namely one that does not
C 0
contain another w-cutset. If we have a w-cutset sequence,
we can reason about which w to choose for applying the w-
cutset conditioning algorithm or w-cutset sampling. Given
a w-cutset sequence we define a function f (i) = jCi j + i
where i ranges from 1 to tw. This function characterizes
the complexity of the w-cutset conditioning algorithms: for
each i, the space complexity is exponential in i and the
time complexity is exponential in f (i). The space con-
sideration suggests selecting i as small as possible. No-
tice that for various intervals of i, f (i) maybe constant:
if jCi j = jCi+1 j + 1. Thus, given a w-cutset sequence,
whenever f (i) = f (i + 1), then w = i is preferred over
Figure 1: (a) Graph; (b) triangulated graph and correspond-
ing tree decomposition of width 2; (c) graph with 1-cutset
node fDg removed and corresponding tree-decomposition.
w = i + 1. Alternatively, given a bound on the space-
complexity expressed by r, we can select a most preferred
wp -cutset such that:
wp (r) = arg min
j
fr = j g
In the empirical section 6, we demonstrate the analysis of
function f (i) and its implications.
(2)
THEOREM 3.3 Given a tree-decomposition T =(V ; E )
where V =fV1 ; :::; Vt g is the set of clusters and given a con-
w of G satisfies:
stant w, a minimum w-cutset C (cid:3)
w j (cid:20) Xi;jVi j>w+1
jC (cid:3)
(jVi j (cid:0) (w + 1))
Proof. From each cluster Vi 2 V of size larger than
(w+1), select a subset of nodes Ci (cid:26) Vi of size
jCi j = jVi j (cid:0) (w + 1) so that jVi nCi j (cid:20) w + 1.
Let Cw = [i;jVi j>w+1Ci .
By construction, Cw is a w-cutset of G and:
w (cid:20) jCw j = j [i Ci j (cid:20) Pi jCi j = Pi;jVi j>w+1 jVi j (cid:0)
c(cid:3)
(w + 1). ut
Since a w-cutset yields a tree-decomposition having tw =
w, it looks reasonable when seeking w-cutset to start from
a good tree-decomposition and find its w-cutset (or a se-
quence). In particular, this avoids the need to test if a graph
has tw = w. This task is equivalent to finding a w-cutset
of a chordal (triangulated) graph.
DE FIN IT ION 3.1 (A w-cutset of a tree-decomposition)
Given a tree decomposition T =<V ; E> of a reasoning
problem <X; F > where V is a set of subsets of X then
w (cid:26) X is a w-cutset relative to T if for every i,
C T
w j (cid:20) w + 1.
jVi nC T
We should note upfront, however, that a minimum-size w-
cutset of T (even if T is optimal) is not necessarily a mini-
mum w-cutset of G.
Example 3.4 Consider a graph in Figure 1(a). An optimal
tree decomposition of width 2 is shown in Figure 1(b). This
tree-decomposition clearly does not have a 1-cutset of size
< 2. However, the graph has a 1-cutset of size 1, fDg, as
shown in Figure 1(c).
UAI 2004BIDYUK & DECHTER45 On the other hand, given a minimum w-cutset, removing
the w-cutset from the graph yields a graph having tw (cid:3) = w.
Because, otherwise, there exists a tree-decomposition over
X nCw having tw < w. Select such a tree and select a node
in Cw that can be added to the tree-decomposition without
increasing its tree-width beyond w. Such a node must exist,
contradicting the minimality of Cw .
It is still an open question if every minimal w-cutset is a w-
cutset of some minimum-width tree-decomposition of G.
4 Hardness of w-Cutset on Cluster-Tree
While it is obvious that the general w-cutset problem is
NP-complete (1-cutset is a cycle-cutset known to be NP-
complete), it is not clear that the same holds relative to a
given tree-decomposition. We now show that, given a tree-
decomposition T of a hyper-graph H, the w-cutset problem
for T is NP-complete.We use a reduction from set multi-
cover (SMC) problem.
DE FIN IT ION 4.1 (Set Cover (SC)) Given a pair < U; S >
where U is universal set and S is a set of subsets
S=fS1 ; :::; Smg of U , find a minimum set C (cid:26) S s.t. each
element of U is covered at least once: [Si2C Si = U .
DE FIN IT ION 4.2 (Set Multi-Cover(SMC)) Given a pair
< U; S > where U is universal set and S is a set of subsets
S = fS1 ; :::; Smg of U, find a minimum cost set C (cid:26) S s.t.
each Ui 2 U is covered at least ri > 0 times by C.
The SC is an instance of SMC problem when 8i; ri = 1.
THEOREM 4.1 (NP-completeness) The problem ” Given a
tree-decomposition T =<V ; E> and a constant k , does
there exist a w-cutset of T of size at most k ? ”
is NP-
complete.
Proof. Given a tree decomposition T =<V ; E> over X
and a subset of nodes C 2 X , we can verify in linear
time whether C is a w-cutset of T by checking if 8Vi 2 V ,
jVi nC j (cid:20) w + 1. Now, we show that the problem is NP-
hard by reduction from set multi-cover.
Assume we are given a set multi-cover problem <U; S>,
where U =fX1 ; :::; Xng and S =fS1 ; :::; Smg, a covering re-
quirement ri > 0 for each Ui 2 U .
a cluster tree T =<V ; E> over S where there
We define
is a node Vi 2 V corresponding to each variable Ui in U
that contains all subsets Sj 2 S that cover node Xi : Vi =
fSj 2 S jXi 2 Sj g. Additionally, there is a node VS 2 V
that contains all subsets in S: VS = S . Thus, V = fVi jUi 2
U g [ VS . The edges are added between each cluster Vi;i6=s
and cluster VS : E = fVi VS jUi 2 U g to satisfy running
intersection property in T .
Define w+1=jS j (cid:0) mini ri = m (cid:0) mini ri . For each Vi;i6=s ,
U={U1,U2,U3}
r1=2, r2=2, r3=1
S1={U1},
S2={U3},
S3={U1, U2},
S4={U2, U3},
S5={U1, U2, U3}
(a) SMC
'
1V
S1,S3,S5,
Q1
1,Q1
2,Q1
3
S1,S2,S3,S4,S5
'
SV
'
2V
S3,S4,S5,
2,Q2
Q2
1,Q2
3
'
3V
S2,S4,S5,
Q3
1,Q3
2
(b) 3-cutset problem
Figure 2:
(a) A set multi-cover problem <U; S>
where U =fU1 ; U2 ; U3g, S =fS1 ; :::; S5g,
the covering
requirements r1 =2,
(b) Correspond-
r3 =1.
r2 =2,
ing augmented tree decomposition T 0=<V 0 ; E> over
S 0=fS1 ; :::; S5 ; Q1
3g.
1 ; Q2
1 ; Q3
1 ; Q1
2 ; Q2
2 ; Q3
2 ; Q1
3 ; Q2
since jVi j (cid:20) m and ri > 0, then jVi j(cid:0)ri (cid:20) jVi j(cid:0)mini ri (cid:20)
m (cid:0) mini ri = w + 1. Consequently, jVi j (cid:20) ri + w + 1.
For each Vi s.t.
jVi j < ri + w + 1, de fine (cid:1)i =ri +
w + 1 (cid:0) jVi j and augment cluster Vi with a set of nodes
i :::Q(cid:1)i
i g yielding a cluster V 0
i = Vi [ Qi of size
Qi = fQ1
i j=ri + w + 1.
jV 0
We will show now that a set multi-cover <U; S> has a so-
lution of size k iff there exists a w-cutset of augmented
tree decomposition T 0=<V 0 ; E> of the same size.The aug-
mented tree for sample SMC problem in Figure 2(a) is
shown in Figure 2(b).
Let C be a set multi-cover solution of size k . Then, 8Ui 2
i j (cid:0) ri = w + 1.
i j (cid:21) ri which yields jV 0
U , jC \ V 0
i nC j (cid:20) jV 0
Since jC j (cid:21) mini ri , then jVS nC j (cid:20) jVS j (cid:0) mini ri =
m (cid:0) mini ri = w + 1. Therefore, C is a w-cutset of size k .
Let Cw be a w-cutset problem of size k . If Cw contains a
node Qj 2 Qi , we can replace it with some node Sp 2 V 0
i
without increasing the size of the cutset. Thus, without
loss of generality, we can assume Cw (cid:26) S . For each V 0
i
corresponding to some Ui 2 U , let Ci = Cw \ V 0
i . By
i nCw j (cid:20) w+1. Therefore, jCi j (cid:21)
de finition of w-cutset, jV 0
i j (cid:0) (w + 1) = ri . By de finition, Cw is a cover for the
jV 0
given SMC problem.
Minimum w-cutset problem is NP-hard by reduction from
set multi-cover and is verifiable in linear time. Therefore,
minimum w-cutset problem is NP-complete. ut
Example 4.2 Let us demonstrate those steps for the SMC
problem with U =fU1 ; U2 ; U3g and S=fS1 ; :::; S5g shown
in Figure 2(a). Define T =<V ; E>, V =fV1 ; V2 ; V3 ; Vs g,
over S:
V1=fS1 ; S2 ; S3g, f1=3, V2=fS3 ; S4 ; S5g, f2=3,
V3=fS2 ; S4 ; S5g, f3=3, VS =fS1 ; :::; S5g, fS =5.
Then, w = jS j (cid:0) 1 (cid:0) mini ri = 5 (cid:0) 1 (cid:0) 1 = 3. Augment:
1 ; Q3
1 ; Q2
V1 : (cid:1)1=w+1+r1 (cid:0) f1=4+2-3=3, Q1=fQ1
1g.
2 ; Q2
V2 : (cid:1)2=w+1+r2 (cid:0) f2=4+2-3=3, Q2=fQ1
2 ; Q3
2g.
3 ; Q2
V3 : (cid:1)3=w+1+r3 (cid:0) f3=4+1-3=2, Q3=fQ1
3g.
The augmented tree decomposition T 0 is shown in Fig-
46BIDYUK & DECHTERUAI 2004 ure 2(b). Any SMC solution such as C =fS3 ; S5g is a 3-
cutset of T and vice versa.
ACE
DEF
U={V1,V2,V3 ,V4}
r1=1, r2=1, r3=1,r4=1
In summary, we showed that when w is not a constant the
w-cutset problem is NP-complete. This implies that the w-
cutset sequence problem over tree-decompositions is hard.
5 Algorithm GWC for minimum cost
w-cutset
Next, we show that the problem of finding w-cutset can
be mapped to that of finding set multi-cover. The map-
ping suggests an application of greedy approximation al-
gorithm for set multi-cover problem to find w-cutset of a
tree decomposition. When applied to a tree decomposi-
tion T =<V ; E> over X , it is guaranteed to find a solution
within factor O(1 + ln m) of optimal where m is the max-
imum # of clusters of size > (w + 1) sharing the same
node. To avoid loss of generality, we consider the weighted
version of each problem.
The mapping is as follows. Given any w-cutset problem
of a tree-decomposition T =<V ; E> over X, each cluster
node Vi 2 V of the tree becomes a node of universal set
U. A covering set SXj =fVi 2 V jXj 2 Vi g is created for
each node Xj 2 X . The cost of SXj equals the cost of Xj .
The cover requirement is ri = jVi j (cid:0) (w + 1). Covering
a node in SMC with a set SXj corresponds to removing
node Xj from each cluster in T. Then, the solution to a set
multi-cover is a w-cutset of T. Let C be a solution to the
SMC problem. For each Ui 2 U , the set C contains at
least ri subsets SXj that contain Ui . Consequently, since
Ui = Vi , then jVi \ C j (cid:21) ri and jVi nC j (cid:20) jVi j (cid:0) ri =
jVi j (cid:0) jVi j + (w + 1) = w + 1. By definition, C is a w-
cutset. An example is shown in Figure 3. This duality is
important because the properties of SC and SMC problems
are well studied and any algorithms previously developed
for SMC can be applied to solve w-cutset problem.
A well-known polynomial time greedy algorithm exists for
weighted SMC [20] that chooses repeatedly set Si that cov-
ers the most ”li ve”
(covered less than ri times) nodes fi at
the cost ci : a set that minimizes the ratio ci =fi . In the con-
text of w-cutset, fi is the number of clusters whose size
still exceeds (w + 1) and ci is the cost of node Xi . As dis-
cussed earlier, ci maybe defined as the size of the domain
of node Xi or its log. When applied to solve the w-cutset
problem, we will refer to the algorithm as GWC (Greedy
W -Cutset). It is formally defined in Figure 4. We define
here the approximation algorithm metrics:
ADE
ABD
V1={ACE}
V2={ADE}
V3={ABD}
V4={DEF}
(a)
SA={V1,V2,V3},
SD={V2,V3,V4},
SB={V3},
SE={V1,V2,V4},
SC={V1},
SF={V4}
(b)
(a) A tree decomposition T =<V ,E> where
Figure 3:
V =fV1 ; :::; V4 g over X =fA; B ; C; D; E ; F g;
the
(b)
corresponding set multi-cover problem <U ,S> where
U =fV1 ; V2 ; V3 ; V4g and S =fSA , SB , SC , SD , SF g; here,
set SXi contains a cluster Vj iff Xi 2 Vj . The 1-cutset of
T is a solution to the set multicover with covering require-
ments r1 =r2=r3 =r4=1: when node Vi 2 V is ”co vered” by
set SXi , node Xi is removed from each cluster.
Greedy w-Cutset Algorithm (GWC)
Input: A set of clusters V = fV1 ; :::; Vm g of a tree-
decomposition over X = fX1 ; :::; Xng where 8Vi 2 V ,
Vi (cid:26) X ; the cost of each node Xi is ci .
Output : A set C (cid:26) X s.t. jVi nC j (cid:20) w.
Set C = ;,t=0.
While 9Vi s.t. jVi j > w + 1 do
1. 8Xi 2 X , compute fi = jfVj gj s.t.
Xi 2 Vj .
2. Find node Xi 2 X that minimizes the ratio ci =fi .
3. Remove Xi from all clusters: 8Vi 2 V ; Vi = Vi nXi .
4. Set X = X nXi , C = C [ fXig.
End While
Return C
jVj j > w and
Figure 4: Greedy w-cutset Algorithm.
GWC is a factor O(1+ ln m) approximation algorithm [17]
where m is the maximum number of clusters sharing the
same node (same as the maximum set size in SMC).
This bound is nearly the best possible for a polynomial al-
gorithm due to strong inapproximability results for the set
cover problem, the special case of set multi-cover problem.
Approximation guarantee better than O(ln m) is not possi-
ble for any polynomial time algorithm unless P=NP [4, 15].
Furthermore, 9 absolute constants C , (cid:1)0 s.t. 8 (cid:1) (cid:21) (cid:1)0 no
polynomial-time algorithm can approximate the optimum
within a factor of ln (cid:1) (cid:0) C ln ln (cid:1) unless P=NP [19].
6 Experiments
DE FIN IT ION 5.1 (factor (cid:14) approximation) An algorithm
A is a factor (cid:14) , (cid:14) > 0, approximation algorithm for
minimization problem P if A is polynomial and for ev-
ery instance I 2 DP it produces a solution s such that:
cost(s) (cid:20) (cid:14) (cid:3) costOP T (s); (cid:14) > 1.
We use Bayesian networks as input reasoning problems.
In all experiments, we started with a moral graph G of a
Bayesian network B which was used as the input to the
minimal w-cutset problem. The tree-decomposition of G
was obtained using min-fill algorithm [12].
UAI 2004BIDYUK & DECHTER47 Our benchmarks are two CPCS networks from UAI repos-
itory: cpcs360b with N=360 nodes and induced width
w(cid:3)=22 and cpcs422b with N=422 nodes and induced width
w(cid:3)=27, one instance each. Our other benchmarks are lay-
ered random networks where each node is assigned a set
of parents selected randomly from a previous layer. One
set of random networks consisted of 4 layers of L = 50
nodes each, total of N =50x4=200 nodes; each node as-
signed P = 3 parents. The second set of random net-
works consisted of 8 layers of L = 25 nodes each, total
of N =25x8=200 nodes; each node assigned P = 3 par-
ents. For random networks, the results are averaged over
100 instances. We compare the performance of two greedy
heuristic algorithms- MGA (Modified Greedy Algorithm
due to [3]) and DGR (Deterministic Greedy Algorithm due
to [11])- to our proposed algorithms: GWC (Greedy W -
Cutset) and its variants. All nodes domains of size 2.
The MGA algorithm is adapted from a minimum cost
cycle-cutset algorithm of [3] that iteratively removes all
nodes if degree 1 from the graph and adds to cutset the node
that minimizes the cost to degree ratio. The algorithm stops
when remaining subgraph is cycle-free. The MGA is a fac-
tor 4 approximation algorithm. In [20], a factor 2 approx-
imation algorithm is defined based on layering. However,
it can not be easily adapted to finding minimal w-cutset for
w > 1. For MGA, the only modification required to find
w-cutset is to stop when original graph with cutset nodes
removed can be decomposed into a cluster tree of width w
or less (using min- fill heuristics). In our implementation,
MGA algorithm uses the GWC heuristics to break ties: if
two nodes have the same degree, the node found in most of
the clusters of size > w is added to the cutset.
The DGR algorithm is the Deterministic Greedy Algorithm
for finding an elimination order of the variables that yields a
tree-decomposition of bounded width defined in [11]. DGR
obtains a w-cutset while computing the elimination order
of the variables. When eliminating some node X yields a
cluster that is too large (size > w + 1), the algorithm uses
greedy heuristics to pick a cutset node among all the nodes
that are not in the ordering yet. Specifically , the determin-
istic algorithm adds to the cutset a node X that maximizes
expression pjNX jCX , where NX is a set of neighbours of
X that are not eliminated yet and CX = QUi2NX jD(Ui )j.
The GWC algorithm was implemented as described earlier
picking at each iteration a node found in most clusters of
size > w + 1 with a secondary heuristics (tie breaking) that
selects the node contained in most of the clusters. Several
variants of GWC with different tie breaking heuristics were
tested that were allowed to rebuild a tree decomposition af-
ter removing a cutset node:
GWCA - breaks ties by selecting the node found in most
of the clusters of the tree decomposition;
GWCM - breaks ties by selecting the node found in most
of the clusters of maximum size;
GWCD - breaks ties by selecting the node of highest de-
gree (the degree of the node is computed on the subgraph
with all cutset nodes removed and all resulting singly-
connected nodes removed). Note that GWC and GWCA
only differ in that GWCA rebuilds a cluster-tree after re-
moving a cutset node. Also note that MGA and GWCD
have their primary and tie-breaking heuristics switched.
Table 1: w-cutset. Networks: I=cpcs360b, II=cpcs422b,
III=4-layer random networks, L=50, N=200, P=3; IV =8-
layer random networks, L=25, N=200, P=3.
9 10
8
7
1
2
3
4
5
6
w
MGA 30 22 20 18 16 15 14 13 12 10
I
36 22 19 18 16 14 13 12 11 10
w*=20 DGR
27 20 17 16 15 14 13 12 11 10
GWC
GWCA 27 21 18 16 15 14 13 12 11 10
GWCD 27 21 18 16 15 14 13 12 11 10
MGA 80 70 65 60 54 49 44 41 38 36
II
84 70 63 54 49 43 38 32 27 23
w*=22 DGR
78 66 58 52 46 41 36 31 26 22
GWC
GWCA 78 65 57 51 45 40 35 30 25 21
GWCD 78 65 57 51 45 40 35 30 25 21
MGA 87 59 54 52 50 48 47 45 44 43
III
w*=49 DGR
80 57 52 50 48 46 44 43 42 40
GWC
78 61 53 49 46 44 43 42 41 39
GWCA 74 56 50 47 44 42 41 39 38 37
GWCD 74 56 49 47 44 42 41 39 38 37
MGA 99 74 69 66 63 61 59 56 54 51
IV
w*=24 DGR
90 71 65 61 58 55 52 49 47 44
GWC
93 77 68 63 59 55 52 49 46 43
GWCA 87 70 62 57 54 51 48 45 42 39
GWCD 86 70 62 57 54 51 48 45 42 39
11 12 13 14 15 16 17 18 19 20
w
0
1
2
3
4
5
6
7
8
9
I
MGA
0
1
2
3
4
5
6
7
8
9
w*=20 DGR
0
1
2
3
4
5
6
7
8
9
GWC
GWCA 9
8
7
6
5
4
3
2
1
0
0
1
2
3
4
5
6
GWCD 9
7
8
2
8
9
4
5
6
7
MGA 33 30 28
II
2
3
4
5
8
9
7
21 19 16
w*=22 DGR
8
6
5
4
3
2
GWC
19 16 13 10
2
3
4
5
6
8
9
GWCA 18 15 12
GWCD 18 15 12
9
8
6
5
4
3
2
MGA 41 40 39 37 36 35 34 33 31 30
III
w*=49 DGR
39 38 36 36 34 33 32 31 30 29
GWC
38 37 36 35 34 33 32 31 30 29
GWCA 36 35 34 33 32 31 30 29 28 27
GWCD 36 34 34 33 32 31 30 29 28 27
MGA 49 47 44 41 39 36 34 31 28 26
IV
41 38 36 33 31 28 25 23 21 19
w*=24 DGR
GWC
40 37 35 32 29 27 25 23 20 18
GWCA 37 34 32 30 27 25 23 21 19 17
GWCD 37 35 32 30 28 25 24 21 19 17
48BIDYUK & DECHTERUAI 2004 The results are presented in Table 1. For each bench-
mark, the table provides the fi ve rows of results correspond-
ing to the fi ve algorithms (labelled in the second column).
Columns 3-12 are the w-cutset sizes for the w-value. The
upper half of the table entires provides results for w in
range [1; 10]; the lower half of the table provides results for
w in range [11; 20]. The results for cpcs360b and cpcs422b
correspond to a single instance of each network. The result
for random networks are averaged over 100 instances. The
best entries for each w are highlighted.
As Table 1 shows, it pays to rebuild a tree decomposi-
tion: with rare exceptions, GWCA finds a cutset as small
as GWC or smaller. On average, GWCA, GWCM, and
GWCD computed the same-size w-cutsets. The results for
GWCM are omitted since they do not vary sufficiently from
the others.
The performance of MGA algorithm appears to depend on
the network structure. In case of cpcs360b, it computes the
same size w-cutset as GWC variants for w (cid:21) 10. However,
in the instance of cpcs422b, MGA consistently finds
larger
cutsets except for w=20. On average, as reflected in the
results for random networks, MGA finds
larger cutset than
DGR or any of the GWC-family algorithms. In turn, DGR
occasionally finds a smaller cutset compared to GWC, but
always a larger cutset compared to GWCA and GWCD.
We measured the GWC algorithm approximation parame-
ter M in all of our benchmarks. In cpcs360b and cpcs422b
we have M = 86 and M = 87 yielding approximation
factor of 1 + ln M (cid:25) 5:4. In random networks, M varied
from 29 to 47 yielding approximation factor 2 [4:3; 4:9].
Thus, if C is the w-cutset obtained by GWC and Copt is
the minimum size w-cutset, then on average:
jC j
jCopt j
(cid:20) 5
Looking at the results as solutions to the sequence w-cutset
problems, we can inspect the sequence and suggest good
w’s by analysing the function f (i) = jCi j + i as de-
scribed in section 3. To illustrate this we focus on algo-
rithm GWCA for CPC364, CPCS424 and 4-layer random
networks (See Table 2).
For cpcs360b we observe a small range of values for f (i),
namely f (i) 2 f20; 21; 23; 28g. In this case the point of
choice is w = 4 because f (1) = 28, f (2) = 23, f (3) = 21
while at i = 4 we obtain reduction f (4) = 20 which stays
constant for i (cid:21) 4. Therefore, we can have the same time
complexity for w-cutset as for exact inference (w(cid:3) = 20)
while saving a lot in space, reducing space complexity from
exponential in 20 to exponential in 4 only. For w-cutset
sampling this implies sampling 20 variables (out of 360)
and for each variable doing inference exponential in 4.
The results are even more interesting for cpcs422b where
we see a fast decline in time complexity with relatively
Table 2: Function f (i) for i=1...16, GWCA. Networks:
I=cpcs360b,
random, L=50,
III=4-layer
II=cpcs422b,
N=200, P=3.
f(i)
9 10
8
7
6
5
4
3
2
1
i
28 23 21 20 20 20 20 20 20 20
I
II
79 67 60 55 50 46 42 38 34 31
III 75 57 53 51 49 48 48 47 47 47
f(i)
11 12 13 14 15 16 17 18 19 20
i
20 20 20 20 20 20 20 20 20 20
I
II
29 27 25 23 23 22 22 22 22 22
III 47 47 47 47 47 47 47 47 47 47
slow decline in space complexity for the range i =
1; :::; 11. The decline is more moderate for i (cid:21) 11 but
is still cost-effective: for i = 16 we get the same time per-
formance as i = 20 and therefore i = 16 represents a more
cost-effective point.
Finally, for the case of 4-layer random networks, on av-
erage the function f (i) decreases for i = 1:::8 and then
remains constant. This suggests that if space complexity
allows, the best point of operation is w = 8.
7 Related Work and Conclusions
In this paper, we formally defined the minimum w-cutset
problem applicable to any reasoning problems with graph-
ical models such as constraint networks and Bayesian net-
works. The minimum w-cutset problem extends the min-
imum cycle-cutset problem corresponding to w = 1. The
motivation for finding a minimal w-cutset is to bound the
space complexity of the problem (exponential in the width
of the graph) while minimizing the required additional pro-
cessing time (exponential in the width of the graph plus the
size of cutset).
The cycle-cutset problem corresponds to the well-known
weighted vertex-feedback set problem and can be approx-
imated within factor 2 of optimal by a polynomial algo-
rithm. We show that the minimal w-cutset problem is
harder. Since any set multi-cover problem (SMC) ([20])
can be reduced to a w-cutset problem of a clique tree, the
w-cutset problem of a tree-decomposition is at least as hard
as SMC and cannot have a constant-factor polynomial ap-
proximation algorithm unless P=NP. This complexity re-
sult applies to the problem of finding minimal w-cutset
of a graph. Consider chordal graphs discussed earlier in
Section 2.1. Each chordal graph has a corresponding tree
decomposition whose clusters are the maximal cliques of
the graph and vice-versa. If there was a constant-factor w-
cutset algorithm for a graph, then it could be used to find
the minimal w-cutset of a chordal graph that is the minimal
UAI 2004BIDYUK & DECHTER49 w-cutset of the corresponding tree-decomposition.
Empirically, we show that the minimal cycle-cutset heuris-
tics based on the degree of a node is not competitive with
the tree-decomposition of the graph. We also demonstrate
that the proposed GWC-family algorithms consistently out-
perform the previously defined algorithms based on the
node elimination order in [18, 13] and [11]. In [18, 13],
the next elimination node is added to the cutset if its bucket
size exceeds the limit. A similar approach was explored in
[11] in DGR algorithm (presented in the empirical section)
except that the cutset node was chosen heuristically among
all the nodes that were not eliminated yet. The immediate
drawback of either approach is that it does not permit to
change the order of the nodes already eliminated. As the
empirical results demonstrate, DGR usually finds
smaller
cutset than MGA but bigger than GWC/GWCA/GWCD.
good
The main objective of our future work is to find
heuristics for w-cutset problem that are independent from
the tree-decomposition of a graph. So far, we only looked
at the degree of a node as a possible heuristics and found
empirically that GWC heuristics are usually superior. Al-
ternatively, we want to incorporate into the heuristic mea-
sure of the ”goodness”
of the w-cutset other factors such as
correlation between variables that strongly affect the con-
vergence of the Markov-chain based w-cutset sampling. Fi-
nally, there are open questions remaining regarding the re-
lationship between w-cutset of a graph and a w-cutset of
its tree-decomposition. It is not clear, for example, whether
the minimal w-cutset of a graph is a w-cutset of one of its
minimum width tree-decompositions.
8 Acknowledgements
This work was supported in part by NSF grant IIS-0086529
and MURI ONR award N00014-00-1-061 7.
References
[1] S. A. Arnborg, ‘Efficient algorithms for combinatorial
problems on graphs with bounded decomposability -
a survey’, BIT, 25, 2–23,
(1985).
[2] A. Becker, R. Bar-Yehuda, and D. Geiger, ‘Random
algorithms for the loop cutset problem’, in Uncer-
tainty in AI, (1999).
[3] A. Becker and D. Geiger, ‘A sufficiently fast algo-
rithm for finding close to optimal junction trees’, in
Uncertainty in AI, pp. 81–89,
(1996).
[4] M. Bellare, C. Helvig, G. Robins, and A. Zelikovsky,
‘Provably good routing tree construction with multi-
port terminals’, in Twenty-Fifth Annual ACM Sympo-
sium on Theory of Computing, pp. 294–304,
(1993).
[5] B. Bidyuk and R. Dechter, ‘Cycle-cutset sampling for
bayesian networks’, Sixteenth Canadian Conf. on AI,
(2003).
[6] B. Bidyuk and R. Dechter, ‘Empirical study of w-
cutset sampling for bayesian networks’, UAI, (2003).
[7] R. Dechter, ‘Enhancement schemes for constraint
processing: Backjumping, learning and cutset decom-
Intelligence, 41, 273–312,
position’, Artificial
(1990).
[8] R. Dechter, ‘Bucket elimination: A unifying frame-
Intelligence, 113, 41–
work for reasoning’, Artificial
85, (1999).
[9] R. Dechter, Constraint Processing, Morgan Cauf-
mann, 2001.
[10] R. Dechter and J. Pearl, ‘Network-based heuristics
for constraint satisfaction problems’, Arti ficial
Intel-
ligence, 34, 1–38,
(1987).
[11] D. Geigher and M. Fishelson, ‘Optimizing exact ge-
netic linkage computations’, in 7th Annual Interna-
tional Conf. on Computational Molecular Biology,
(2003).
pp. 114–121,
[12] U. Kjaerulff. Triangulation of graphs - algorithms
giving small total space, 1990.
[13] J Larossa and R. Dechter, ‘Dynamic combination of
search and variable-elimination in csp and max-csp’,
Constraints, (2003).
[14] S.L. Lauritzen and D.J. Spiegelhalter, ‘Local compu-
tation with probabilities on graphical structures and
their application to expert systems’, Journal of the
Royal Statistical Society, Series B, 50(2), 157–224,
(1988).
[15] C. Lund and M. Yannakakis, ‘On the hardness of
approximating minimization problems’, J. of ACM,
41(5), 960–981,
(September 1994).
[16] J. Pearl, Probabilistic Reasoning in Intelligent Sys-
tems, Morgan Kaufmann, 1988.
[17] S. Rajagopalan and V.V. Vazirani, ‘Primal-dual rnc
approximation algorithms for (multi)set (multi)cover
and covering integer programs’, SIAM J. of Comput-
ing, 28(2), 525–540,
(1998).
‘Resolution vs. search;
[18] I. Rish and R. Dechter,
two strategies for sat’, J. of Automated Reasoning,
24(1/2) , 225–275,
(2000).
[19] Luca Trevisan, ‘Non-approximability resu.lts for op-
timization problems on bounded degree instances’, In
proceedings of 33rd ACM STOC, (2001).
[20] V. V. Vazirani, Approximation Algorithms, Springer,
2001.
50BIDYUK & DECHTERUAI 2004 |
1111.0321 | 3 | 1111 | 2016-03-13T09:49:08 | Anonymous Meeting in Networks | [
"cs.DS",
"cs.DC"
] | A team consisting of an unknown number of mobile agents, starting from different nodes of an unknown network, possibly at different times, have to meet at the same node. Agents are anonymous (identical), execute the same deterministic algorithm and move in synchronous rounds along links of the network. Which configurations are gatherable and how to gather all of them deterministically by the same algorithm?
We give a complete solution of this gathering problem in arbitrary networks. We characterize all gatherable configurations and give two universal deterministic gathering algorithms, i.e., algorithms that gather all gatherable configurations. The first algorithm works under the assumption that an upper bound n on the size of the network is known. In this case our algorithm guarantees gathering with detection, i.e., the existence of a round for any gatherable configuration, such that all agents are at the same node and all declare that gathering is accomplished. If no upper bound on the size of the network is known, we show that a universal algorithm for gathering with detection does not exist. Hence, for this harder scenario, we construct a second universal gathering algorithm, which guarantees that, for any gatherable configuration, all agents eventually get to one node and stop, although they cannot tell if gathering is over. The time of the first algorithm is polynomial in the upper bound n on the size of the network, and the time of the second algorithm is polynomial in the (unknown) size itself.
Our results have an important consequence for the leader election problem for anonymous agents in arbitrary graphs. For anonymous agents in graphs, leader election turns out to be equivalent to gathering with detection. Hence, as a by-product, we obtain a complete solution of the leader election problem for anonymous agents in arbitrary graphs. | cs.DS | cs | Anonymous Meeting in Networks∗
Yoann Dieudonn´e†
Andrzej Pelc‡
Abstract
A team consisting of an unknown number of mobile agents, starting from different nodes of an
unknown network, possibly at different times, have to meet at the same node. Agents are anonymous
(identical), execute the same deterministic algorithm and move in synchronous rounds along links of the
network. An initial configuration of agents is called gatherable if there exists a deterministic algorithm
(even dedicated to this particular configuration) that achieves meeting of all agents in one node. Which
configurations are gatherable and how to gather all of them deterministically by the same algorithm?
We give a complete solution of this gathering problem in arbitrary networks. We characterize all
gatherable configurations and give two universal deterministic gathering algorithms, i.e., algorithms that
gather all gatherable configurations. The first algorithm works under the assumption that a common
upper bound N on the size of the network is known to all agents. In this case our algorithm guarantees
gathering with detection, i.e., the existence of a round for any gatherable configuration, such that all
agents are at the same node and all declare that gathering is accomplished. If no upper bound on the size
of the network is known, we show that a universal algorithm for gathering with detection does not exist.
Hence, for this harder scenario, we construct a second universal gathering algorithm, which guarantees
that, for any gatherable configuration, all agents eventually get to one node and stop, although they
cannot tell if gathering is over. The time of the first algorithm is polynomial in the upper bound N on
the size of the network, and the time of the second algorithm is polynomial in the (unknown) size itself.
Our results have an important consequence for the leader election problem for anonymous agents in
arbitrary graphs. Leader election is a fundamental symmetry breaking problem in distributed computing.
Its goal is to assign, in some common round, value 1 (leader) to one of the entities and value 0 (non-
leader) to all others. For anonymous agents in graphs, leader election turns out to be equivalent to
gathering with detection. Hence, as a by-product, we obtain a complete solution of the leader election
problem for anonymous agents in arbitrary graphs.
Keywords: gathering, deterministic algorithm, anonymous mobile agent.
6
1
0
2
r
a
M
3
1
]
S
D
.
s
c
[
3
v
1
2
3
0
.
1
1
1
1
:
v
i
X
r
a
∗A preliminary version of this paper appeared in the Proc. 24th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA 2013). A part of this research was done during the Dieudonn´e's stay at the Research Chair in Distributed Computing of
the Universit´e du Qu´ebec en Outaouais as a postdoctoral fellow. Supported in part by NSERC discovery grant and by the Research
Chair in Distributed Computing of the Universit´e du Qu´ebec en Outaouais.
†MIS, Universit´e de Picardie Jules Verne Amiens, France and D´epartement d'informatique, Universit´e du Qu´ebec en Outaouais,
‡D´epartement d'informatique, Universit´e du Qu´ebec en Outaouais, Gatineau, Qu´ebec J8X 3X7, Canada. E-mail: [email protected].
Gatineau, Qu´ebec J8X 3X7, Canada. E-mail: [email protected].
Introduction
1
The background. A team of at least two mobile agents, starting from different nodes of a network, possibly
at different times, have to meet at the same node. This basic task, known as gathering or rendezvous, has
been thoroughly studied in the literature. This task has even applications in everyday life, e.g., when agents
are people that have to meet in a city whose streets form a network. In computer science, mobile agents
usually represent software agents in computer networks, or mobile robots, if the network is a labyrinth. The
reason to meet may be to exchange data previously collected by the agents, or to coordinate some future
task, such as network maintenance or finding a map of the network.
The model and the problem. The network is modeled as an undirected connected graph, referred to
hereafter as a graph. We seek gathering algorithms that do not rely on the knowledge of node labels, and
can work in anonymous graphs as well (cf. [4]). The importance of designing such algorithms is motivated
by the fact that, even when nodes are equipped with distinct labels, agents may be unable to perceive them
because of limited sensory capabilities, or nodes may refuse to reveal their labels, e.g., due to security or
privacy reasons. Note that if nodes had distinct labels, then agents might explore the graph and meet in
the smallest node, hence gathering would reduce to exploration. On the other hand, we assume that edges
incident to a node v have distinct labels in {0, . . . , d− 1}, where d is the degree of v. Thus every undirected
edge {u, v} has two labels, which are called its port numbers at u and at v. Port numbering is local, i.e.,
there is no relation between port numbers at u and at v. Note that in the absence of port numbers, edges
incident to a node would be undistinguishable for agents and thus gathering would be often impossible, as
the adversary could prevent an agent from taking some edge incident to the current node.
There are at least two agents that start from different nodes of the graph and traverse its edges in
synchronous rounds. They cannot mark visited nodes or traversed edges in any way. The adversary wakes
up some of the agents at possibly different times. A dormant agent, not woken up by the adversary, is woken
up by the first agent that visits its starting node, if such an agent exists. Agents are anonymous (identical)
and they execute the same deterministic algorithm. (Algorithms assuming anonymity of agents may be also
used if agents refuse to reveal their identities.) Every agent starts executing the algorithm in the round of its
wake-up. Agents do not know the topology of the graph or the size of the team. We consider two scenarios:
one when agents know an upper bound on the size of the graph and another when no bound is known. In
every round an agent may perform some local computations and move to an adjacent node by a chosen port,
or stay at the current node. When an agent enters a node, it learns its degree and the port of entry. When
several agents are at the same node in the same round, they can exchange all information they currently have.
However, agents that cross each other on an edge, traversing it simultaneously in different directions, do not
notice this fact. Also, agents cannot observe any part of the network except the currently visited node. We
assume that the memory of the agents is unlimited: they can be viewed as Turing machines.
An initial configuration of agents, i.e., their placement at some nodes of the graph, is called gatherable
if there exists a deterministic algorithm (even only dedicated to this particular configuration) that achieves
meeting of all agents in one node, regardless of the times at which some of the agents are woken up by the
adversary. In this paper we study the following gathering problem:
Which initial configurations are gatherable and how to gather all of them deterministically by
the same algorithm?
In other words, we want to decide which initial configurations are possible to gather, even by an algorithm
specifically designed for this particular configuration, and we want to find a universal gathering algorithm
that gathers all such configurations. We are interested only in terminating algorithms, in which every agent
eventually stops forever.
Our results. We give a complete solution of the gathering problem in arbitrary networks. We characterize
all gatherable configurations and give two universal deterministic gathering algorithms, i.e., algorithms that
1
gather all gatherable configurations. The first algorithm works under the assumption that a common upper
bound N on the size of the network is known to all agents. In this case our algorithm guarantees gathering
with detection, i.e., the existence of a round for any gatherable configuration, such that all agents are at the
same node and all declare that gathering is accomplished. If no upper bound on the size of the network is
known, we show that a universal algorithm for gathering with detection does not exist. Hence, for this harder
scenario, we construct a second universal gathering algorithm, which guarantees that, for any gatherable
configuration, all agents eventually get to one node and stop, although they cannot tell if gathering is over.
The time of the first algorithm is polynomial in the upper bound N on the size of the network, and the time
of the second algorithm is polynomial in the (unknown) size itself.
While gathering two anonymous agents is a relatively easy task (cf. [17]), our problem of gathering an
unknown team of anonymous agents presents the following major difficulty. The asymmetry of the initial
configuration because of which gathering is feasible, may be caused not only by non-similar locations of the
agents in the graph, but by their different situation with respect to other agents. Hence a new algorithmic
idea is needed in order to gather: agents that were initially identical, must make decisions based on the
memories of other agents met to date, in order to distinguish their future behavior. In the beginning the
memory of each agent is a blank slate and in the execution of the algorithm it records what the agent has
seen in previous steps of the navigation and what it heard from other agents during meetings. Even a slight
asymmetry occurring in a remote part of the graph must eventually influence the behavior of initially distant
agents. Notice that agents in different initial situations may be unaware of this difference in early meetings,
as the difference may be revealed only later on, after meeting other agents. Hence, for example, an agent
may mistakenly "think" that two different agents that it met in different stages of the algorithm execution,
are the same agent. Confusions due to this possibility are a significant challenge absent both in gathering
two (even anonymous) agents and in gathering many labeled agents.
Our results have an important consequence for the leader election problem for anonymous agents in ar-
bitrary graphs. Leader election [44] is a fundamental symmetry breaking problem in distributed computing.
Its goal is to assign, in some common round, value 1 (leader) to one of the entities and value 0 (non-leader)
to all others. For anonymous agents in graphs, leader election turns out to be equivalent to gathering with
detection (see Section 5). Hence, as a by-product, we obtain a complete solution of the leader election
problem for anonymous agents in arbitrary graphs.
Related work. Gathering was mostly studied for two mobile agents and in this case it is usually called
rendezvous. An extensive survey of randomized rendezvous in various scenarios can be found in [4], cf. also
[2, 3, 5, 36]. Deterministic rendezvous in networks was surveyed in [45]. Several authors considered the
geometric scenario (rendezvous in an interval of the real line, see, e.g., [11, 32], or in the plane, see, e.g., [6,
7]). Gathering more than two agents was studied, e.g., in [36, 43]. In [51] the authors considered rendezvous
of many agents with unique labels, and gathering many labeled agents in the presence of Byzantine agents
was studied in [23]. The problem was also studied in the context of multiple robot systems, cf. [15, 26], and
fault tolerant gathering of robots in the plane was studied, e.g., in [1, 16].
For the deterministic setting a lot of effort was dedicated to the study of the feasibility of rendezvous,
and to the time required to achieve this task, when feasible. For instance, deterministic rendezvous with
agents equipped with tokens used to mark nodes was considered, e.g., in [41]. Deterministic rendezvous
of two agents that cannot mark nodes but have unique labels was discussed in [21, 39, 48]. These papers
were concerned with the time of rendezvous in arbitrary graphs. In [21] the authors showed a rendezvous
algorithm polynomial in the size of the graph, in the length of the shorter label and in the delay between the
starting time of the agents. In [39, 48] rendezvous time is polynomial in the first two of these parameters
and independent of the delay.
Memory required by two anonymous agents to achieve deterministic rendezvous was studied in [28,
29] for trees and in [17] for general graphs. Memory needed for randomized rendezvous in the ring was
2
discussed, e.g., in [40].
Apart from the synchronous model used, e.g., in [17, 21, 48, 51] and in this paper, several authors
investigated asynchronous rendezvous in the plane [14, 26] and in network environments [10, 18, 20, 24, 33].
In the latter scenario the agent chooses the edge which it decides to traverse but the adversary controls the
speed of the agent. Under this assumption rendezvous in a node cannot be guaranteed even in very simple
graphs and hence the rendezvous requirement is relaxed to permit the agents to meet inside an edge. In
particular, in [33] the authors studied asynchronous rendezvous of two anonymous agents.
Gathering many anonymous agents in a ring [19, 37, 38] and in a tree [25] was studied in a different
model, where agents execute Look-Compute-Move cycles in an asynchronous way and can see positions
of other agents during the Look operation. While agents have the advantage of periodically seeing other
agents, they do not have any memory of past events, unlike in our model.
Tree exploration by many agents starting at the same node was studied in [27]. Some of the techniques
of this paper have been later used in our paper [22] in the context of anonymous graph exploration by many
agents.
The leader election problem was introduced in [42] and has been mostly studied in the scenario where
all nodes have distinct labels and the leader is found among them [30, 35, 46]. In [34], the leader elec-
tion problem was approached in a model based on mobile agents for networks with labeled nodes. Many
authors [8, 9, 12, 49, 50] studied leader election in anonymous networks. In particular, [50] characterized
message-passing networks in which leader election among nodes can be achieved when nodes are anony-
mous. Memory needed for leader election in unlabeled networks was studied in [31].
2 Preliminaries
Throughout the paper, the number of nodes of a graph is called its size.
In this section we recall four
procedures known from the literature, that will be used as building blocks in our algorithms. The aim of
the first two procedures is graph exploration, i.e., visiting all nodes of the graph by a single agent (cf., e.g.,
[13, 47]).
The first of these procedures assumes an upper bound N on the size of the graph. It is based on universal
exploration sequences (UXS) and is a corollary of the result of Reingold [47]. It allows the agent to traverse
all nodes of any graph of size at most N, starting from any node of this graph, using P (N ) edge traversals,
where P is some polynomial. After entering a node of degree d by some port p, the agent can compute the
port q by which it has to exit; more precisely q = (p + xi) mod d, where xi is the corresponding term of
the UXS of length P (N ).
The second procedure makes no assumptions on the size but it is performed by an agent using a fixed
token placed at the starting node of the agent. (It is well known that a terminating exploration even of all
anonymous rings of unknown size by a single agent without a token is impossible.) In our applications the
roles of the token and of the exploring agent will be played by agents or by groups of agents.
The first procedure works in time polynomial in the known upper bound N on the size of the graph and
the second in time polynomial in the size of the graph. At the end of each procedure all nodes of the graph
are visited. Moreover, at the end of the second procedure the agent is with the token and has a complete
map of the graph with all port numbers marked. We call the first procedure EXP LO(N ) and the second
procedure EST , for exploration with a stationary token. We denote by T (EXP LO(N )) the maximum
time of execution of the procedure EXP LO(N ) in a graph of size at most N.
Before describing the third procedure we define the following notion from [50]. Let G be a graph and v
a node of G. The view from v is the infinite rooted tree V(v) with labeled ports, defined as follows. Nodes of
this tree are all (not necessarily simple) finite paths in G starting at v and represented as the corresponding
sequences of port numbers. The root of the tree is the empty path, and children of path π of length x are all
paths of length x + 2, whose prefix is π. A child ρ of node π, such that ρ = πpq (with juxtaposition standing
for concatenation) is connected with node π by an edge that has port p at node π and port q at node ρ.
3
The truncated view V l(v) at depth l is the rooted subtree of V(v) induced by all nodes at distance at
most l from the root.
The third procedure, described in [17], permits a single anonymous agent starting at node v of an
n-node graph, to find a positive integer S(v) ∈ {1, . . . , n}, called the signature of the agent, such that
V(v) = V(w) if and only if S(v) = S(w). This procedure, called SIGN (N ) works for any graph G
of known upper bound N on its size and its running time is polynomial in N. After the completion of
SIGN (N ), graph G has been completely explored (i.e., all the nodes of G have been visited at least once)
and the agent is back at its starting node. We denote by T (SIGN (N )) the maximum time of execution of
the procedure SIGN (N ) in a graph of size at most N.
Finally, the fourth procedure is for gathering two agents in a graph of unknown size. It is due to Ta-
Shma and Zwick [48] and relies on the fact that agents have distinct labels. (Using it as a building block in
our scenario of anonymous agents is one of the difficulties that we need to overcome.) Each agent knows
its own label (which is a parameter of the algorithm) but not the label of the other agent. We will call this
procedure T Z((cid:96)), where (cid:96) is the label of the executing agent. In [48], the authors give a polynomial P in
two variables, increasing in each of the variables, such that, if there are agents with distinct labels (cid:96)1 and
(cid:96)2 operating in a graph of size n, appearing at their starting positions in possibly different times, then they
will meet after at most P (n,(cid:96)) rounds since the appearance of the later agent, where (cid:96) is the smaller label.
Also, if an agent with label (cid:96)i performs T Z((cid:96)i) for P (n,(cid:96)i) rounds and the other agent is inert during this
time, the meeting is guaranteed.
We will also use a notion similar to that of the view but reflecting the positions of agents in an initial
configuration. Consider a graph G and an initial configuration of agents in this graph. Let v be a node
occupied by an agent. The enhanced view from v is the couple (V(v), f ), where f is a binary valued
function defined on the set of nodes of V(v), such that f (π) = 1 if there is an agent at the unique node w
in G that is reached from v by path π, and f (π) = 0 otherwise. (Recall that nodes of V(v) are paths in G,
represented as sequences of port numbers.) Thus the enhanced view of an agent additionally marks in its
view the positions of other agents in the initial configuration.
An important notion used throughout the paper is the memory of an agent. Intuitively, the memory of
an agent in a given round is the total information the agent collected since its wake-up, both by navigating
in the graph and by exchanging information with other agents met until this round. This is formalized as
follows. The memory of an agent A at the end of some round which is the t-th round since its wake-up, is the
sequence (M0, M1, . . . , Mt), where M0 is the information of the agent at its start and Mi is the information
acquired in round i. The terms Mi are defined as follows. Suppose that in round i the agent A met agents
A1, . . . , Ak at node v of degree d. Suppose that agent A entered node v in round i leaving an adjacent node
w by port p and entering v by port q. Suppose that agent Aj entered node v in round i leaving an adjacent
node wj by port pj and entering v by port qj. If some agent did not move in round i then the respective
ports are replaced by −1. The term Mi for agent A, called its i-th memory box is defined as the sequence
(d, p, q,{(p1, q1, R1), . . . , (pk, qk, Rk)}), where Rj is the memory of the agent Aj at the end of round i− 1.
This definition is recursive with respect to global time (unknown to agents), starting at the wake-up of the
earliest agent by the adversary. Note that if an agent is woken up by the adversary at a node of degree d and
no other agents are at this node at this time then M0 = (d,−1,−1,∅) for this agent (at its wake-up the agent
sees only the degree of its starting position). If an agent is woken up by some other agents, it also learns their
memories to date. Note also that, since the memory of the agent is the total information it acquired to date,
any deterministic algorithm used by agents to navigate (including any deterministic gathering algorithm)
may be viewed as a function from the set of all memories into the set of integers greater or equal to −1
telling the agent to take a given port p (or stay idle, in which case the output is −1), if it has a given memory.
It should be noted that memories of agents may often be (at least) exponential in the size of the graph in
which they navigate.
4
If two agents meet, they must necessarily have different memories. Indeed, if they had identical memo-
ries, then they would be woken up in the same round and they would traverse identical paths to the meeting
node. This contradicts the assumption that agents start at different nodes.
We will end this section by introducing the order ≺ on the set M of all possible memories, and the
notion of P reft(M) that will be used throughout the paper. Let <α be any linear order on the set of all
memory boxes (one such simple order is to code all possible memory boxes as binary sequences in some
canonical way and use lexicographic order on binary sequences). Let <β be a lexicographic order on the set
M based on the order <α. Now the linear order ≺ on the set M is defined as follows: M1 ≺ M2 if (1)
the number of memory boxes of M1 is strictly smaller than that of M2 or (2) M1 and M2 have the same
number of memory boxes and M1 <β M2. When M1 ≺ M2 (resp. M2 ≺ M1), we say memory M2
is larger (resp. smaller) than M1. The order ≺ has the property that if the memory of agent A is smaller
than the memory of agent B in some round then it will remain smaller in all subsequent rounds. Let A be an
agent with a memory M = (M0, M1, . . . , Ms) in round t(cid:48) (with s ≤ t(cid:48)): the memory of agent A in round
t for t ≤ t(cid:48) is denoted P reft(M) and is equal to (M0, M1, . . . , Ms−(t(cid:48)−t)) if t(cid:48) − t < s, and to the empty
word otherwise.
3 Known upper bound on the size of the graph
In this section we assume that a common upper bound N on the size of the graph is known to all agents
at the beginning. The aim of this section is to characterize gatherable configurations and give a universal
gathering algorithm that gathers with detection all gatherable configurations. The time of such an algorithm
is the number of rounds between the wake-up of the first agent and the round when all agents declare that
gathering is accomplished. Consider the following condition on an initial configuration in an arbitrary graph.
G: There exist agents with different views, and each agent has a unique enhanced view.
We will prove the following result.
Theorem 3.1 An initial configuration is gatherable if and only if it satisfies the condition G. If a common
upper bound N on the size of the graph is known to all agents then there exists an algorithm for gathering
with detection all gatherable configurations. This algorithm works in time polynomial in N.
In order to appreciate the full strength of Theorem 3.1 notice that it can be rephrased as follows. If an
initial configuration does not satisfy condition G then there is no algorithm (even no algorithm dedicated
to this specific configuration, knowing it entirely) that permits to gather this configuration, even gather it
without detection: simply no algorithm can bring all agents simultaneously to one node. On the other
hand, assuming that all agents know a common upper bound on the size of the graph, there is a universal
algorithm that gathers with detection all initial configurations satisfying condition G. Our algorithm works
in time polynomial in any commonly known upper bound N on the size of the graph. Hence, if this upper
bound is polynomial in the size of the graph, the algorithm is polynomial in the size as well. In the next
section we will show how the positive part of the result changes when no upper bound on the size of the
graph is known.
It is also important to stress the global nature of the second part of condition G. While the first part
concerns views of individual agents, which depend only on the position of the agent in the graph, the second
part concerns enhanced views, which depend, for each agent, on the positions of all agents in the graph. In
our algorithms, the actions of each agent eventually depend on its enhanced view, and hence require learning
the initial positions of other agents with respect to the initial position of the given agent. (Since not only
agents but also nodes are anonymous, no "absolute" positions can ever be learned). This can be done only by
meeting other agents and learning these positions, sometimes indirectly, with agent A acting as intermediary
in conveying the initial position of agent B to agent C. In short, the execution of a gathering algorithm for
5
an agent depends on the actions of other agents. This is in sharp contrast to rendezvous of two anonymous
agents, cf., e.g., [17], where actions of each agent depend only on its view, whose sufficient part (a truncated
view) can be constructed individually by each agent without any help from the outside. This is also in sharp
contrast with gathering or other tasks performed by two or more labeled agents, when actions of each agent
depend on its label known by the agent in advance, and simple decisions based on comparison of labels are
made after meetings of subsets of agents when the task involves more than two agents, cf., e.g., [24].
The rest of the section is devoted to the proof of Theorem 3.1. We start with the following lemma.
Lemma 3.1 If an initial configuration does not satisfy condition G, then it is not gatherable.
Proof. Suppose that an initial configuration C does not satisfy condition G and that agents execute the same
deterministic algorithm. First suppose that the configuration C does not satisfy the first part of condition G,
i.e., that the views of all agents are identical. Suppose that the adversary wakes up all agents in the same
round. We show that no pair of agents can meet. Suppose, for contradiction, that agents A1 and A2 are the
first to meet and that this occurs in round t from the common start. Since the initial views of the agents are
identical and they execute the same deterministic algorithm, the sequences of port numbers encountered by
both agents are identical and hence they both enter the node at which they first meet by the same port. This
is a contradiction.
Now suppose that the configuration C does not satisfy the second part of condition G, i.e., that there
exists an agent whose enhanced view is not unique. Consider distinct agents A and A(cid:48) that have identical
enhanced views. Let B be any agent and suppose that q is a sequence of port numbers that leads from
agent A to agent B in the enhanced view of agent A. (Notice that there may be many such sequences,
corresponding to different paths leading from A to B.) Then there exists an agent B(cid:48) such that q is a
sequence of port numbers that leads from agent A(cid:48) to agent B(cid:48) in the enhanced view of agent A(cid:48). Agent
B(cid:48) has the same enhanced view as agent B. Since agents A and A(cid:48) were different, agents B and B(cid:48) are
different as well. Hence for every agent there exists another agent whose enhanced view is identical. Call
such agents homologs.
Suppose again that the adversary wakes up all agents in the same round. We will show that, although
now some meetings of agents are possible, homologs will never meet. Suppose that agents A and A(cid:48) are
homologs. Since A and A(cid:48) have the same enhanced view at the beginning, it follows by induction on the
round number that their memory will be identical in every round. Indeed, whenever A meets an agent B
in some round, its homolog A(cid:48) meets a homolog B(cid:48) of B in the same round and hence the memories of A
and A(cid:48) evolve identically. In particular, since all agents execute the same deterministic algorithm, agents A
and A(cid:48) follow identical sequences of port numbers on their paths. As before, if they met for the first time at
(cid:3)
some node, they would have to enter this node by the same port. This is a contradiction.
The other (much more difficult) direction of the equivalence from Theorem 3.1 will be shown by con-
structing an algorithm which, executed by agents starting from any initial configuration satisfying condition
G, accomplishes gathering with detection of all agents. (Our algorithm uses the knowledge of an upper
bound N of the size of the graph: this is enough to prove the other direction of the equivalence, as for any
specific configuration satisfying condition G even a gathering algorithm dedicated to this specific configura-
tion is enough to show that the configuration is gatherable, and such a dedicated algorithm knows the exact
size of the graph. Of course, our algorithm accomplishes much more: knowing just an upper bound on the
size of the graph it gathers all configurations satisfying condition G and does this with detection.) We first
give a high-level idea of the algorithm, then describe it in detail and prove its correctness. From now on we
assume that the initial configuration satisfies condition G.
Idea of the algorithm. At high level the algorithm works in two stages. The aim of the first stage for any
agent is to meet another agent, in order to perform later an exploration in which one of the agents will play
6
the role of the token and the other the role of the explorer (roles will be decided comparing memories of
the agents). To this end the agent starts an exploration of the graph using the known upper bound N on its
size. The aim of this exploration is to wake up all, possibly still dormant agents. Afterwards, in order to
meet, agents use the procedure T Z mentioned in Section 2. However, this procedure requires a label for
each agent and our agents are anonymous. One way to differentiate agents and give them labels is to find
their views: a view (truncated to N) from the initial position could serve as a label because the first part
of condition G guarantees that there are at least two distinct views. However, it is a priori not at all clear
how to find the view of an agent (truncated to N) in time polynomial in N. (Recall that the size of the view
truncated to N is exponential in N.) Hence we use procedure SIGN (N ) mentioned in Section 2, which
is polynomial in N and can still assign labels to agents after exploring the graph, producing at least two
different labels. Then, performing procedure T Z for a sufficiently long time guarantees a meeting for every
agent.
In the second stage each explorer explores the graph using procedure EXP LO(N ) and then backtracks
to its token left at the starting node of the exploration. After the backtrack, memories of the token and of
the explorer are updated to check for anomalies caused by other agents meeting in the meantime either the
token or the explorer. Note that, due to the fact that some agents may have identical memories at this stage
of the algorithm, an explorer may sometimes falsely consider another token as its own, due to their identical
memories. By contrast, an end of each backtrack is a time when an explorer can be sure that it is on its token
and the token can be sure that its explorer is with it.
Explorers repeat these explorations with backtrack again and again, with the aim of creating meetings
with other agents and detecting anomalies. As a consequence of these anomalies some agents merge with
others, mergers being decided on the basis of the memories of the agents. Each explorer eventually either
merges with some token or performs an exploration without anomalies. In the latter case it waits a prescribed
amount of time with its token: if no new agent comes during the waiting time, the end of the gathering is
declared, otherwise another exploration is launched. It will be proved that eventually, due to the second part
of condition G, all agents merge with the same token B and then, after the last exploration made by the
explorer A of B and after undisturbed waiting time, the end of the gathering is correctly declared.
We now give a detailed description of the algorithm.
Algorithm Gathering-with-Detection with parameter N (upper bound on the size of the graph)
During the execution of the algorithm an agent can be in one of the following six states: setup,
cruiser, shadow, explorer, token, searcher, depending on its memory. For every agent A in
state shadow there is exactly one agent B in some state different from shadow, called the guide of A. We
will also say that A is a shadow of B. Below we describe the actions of an agent A in each of the states and
the transitions between the states. At wake-up agent A enters the state setup.
State setup.
Agent A performs SIGN (N ) visiting all nodes (and waking up all still dormant agents) and finding
the signature of its initial position v, called the label of agent A. Agent A transits to state cruiser.
State cruiser.
Agent A performs T Z((cid:96)), where (cid:96) is its label, until meeting an agent in state cruiser or token at a
node v. When such a meeting occurs, we consider 2 cases.
Case 1. Agent A meets an agent B in state token.
We consider 2 subcases
Subcase 1.1. Agent B is not with its explorer at the time of the meeting. Then agent A transits to state
shadow of B.
7
Subcase 1.2. Agent B is with its explorer C at the time of the meeting. Then agent A transits to state
shadow of C.
Case 2. Agent A does not meet an agent in state token.
Then there is at least one other agent in state cruiser at node v (because, as mentionned above, the
considered meeting involves an agent in state cruiser or token). We consider 2 subcases.
Subcase 2.1. Agent A has the largest memory among all agents in state cruiser at node v.
Then agent A transits to state explorer.
Subcase 2.2. Agent A does not have the largest memory among all agents in state cruiser at node v.
If there is exactly one agent B in state cruiser with memory larger than A at node v, then A transits to
state token. Otherwise, it becomes shadow of the agent in state cruiser at node v with largest memory.
State shadow.
Agent A has exactly one guide and is at the same node as the guide in every round. In every round it
makes the same move as the guide. If the guide B transits itself to state shadow and gets agent C as its
guide, then agent A changes its guide to C as well. Agent A declares that gathering is over if the unique
agent in state explorer collocated with it makes this declaration.
Before describing the actions in the three remaining states, we define the notion of seniority of an agent
in state token (respectively explorer). The seniority in a given round is the number of rounds from the
time when the agent became token (respectively explorer).
State explorer
When agent A transits to state explorer, there is another agent B that transits to state token in the
same round at the same node v. Agent B is called the token of A. Agent A has a variable recent-token
that it initializes to the memory of B in this round. Denote by EXP LO∗(N ) the procedure EXP LO(N )
followed by a complete backtrack in which the agent traverses all edges traversed in EXP LO(N ) in the
reverse order and the reverse direction. The variable recent-token is updated in the beginning of each
execution of EXP LO∗(N ). An execution of EXP LO∗(N ) is called clean if the following condition is
satisfied: in each round during this execution, in which A met an agent C that is not in state shadow, the
memory of C is equal to that of B, and in each round during this execution, in which the token B was met
by an agent D, the memory of D was equal to that of A. Notice that after the execution of EXP LO∗(N ),
agent A is together with its token B and thus they can verify if the execution was clean, by inspecting their
memories. The execution time of EXP LO∗(N ) is at most 2T (EXP LO(N )).
After transiting to state explorer, agent A waits for T (SIGN (N )) + P (N, L) rounds, where L is
the largest possible label (it is polynomial in N). Then it executes the following protocol:
8
while A has not declared that gathering is over do
do
EXP LO∗(N )
/*now agent A is with its token.*/
if in round t(cid:48) agent A met an agent C in state token of higher
seniority than that of A or of equal seniority but such that
recent-token ≺ P reft(MC) where MC is the memory of agent C and
t is the last round before t(cid:48) when agent A updated its variable
recent-token then A transits to state searcher
if B was visited in round t(cid:48) by an agent C in state explorer of higher
seniority than that of B or of equal seniority but such that
P reft(MB) ≺ R where MB is the memory of agent B, t is the last
round before t(cid:48) when agent C updated its variable recent-token and R
is the variable recent-token of agent C in round t(cid:48) then A transits to
state searcher
until the execution of EXP LO∗(N ) is clean
agent A waits 2 · T (EXP LO(N )) rounds: this waiting period is interrupted
if A is visited by another agent;
if the waiting period of 2 · T (EXP LO(N )) rounds has expired
without any interruption then A declares that gathering is over.
endwhile
State token
When agent A transits to state token, there is another agent B that transits to state explorer in the
same round at the same node v. Agent B is called the explorer of A. Agent A remains idle at a node v and
does not change its state, except when its explorer B transits to state searcher. In this case it transits to
state shadow and B becomes its guide. Agent A declares that gathering is over if the unique agent in state
explorer collocated with it makes this declaration.
State searcher
Agent A performs an entire execution of EXP LO∗(N ) until its termination, regardless of any meet-
ings it could make during this execution. Then the agent starts another execution of EXP LO∗(N ) which
is stopped as soon as agent A meets an agent B in state token. If at the time of the meeting agent B is not
with its explorer C then agent A transits to state shadow of B (B becomes its guide). Otherwise, agent A
transits to state shadow of C (C becomes its guide).
The proof of the correctness of the algorithm is split into the following lemmas.
Lemma 3.2 In Algorithm Gathering-with-Detection every agent eventually stops after time polynomial in
N and declares that gathering is over.
Proof. At its wake-up an agent A enters state setup and remains in it for at most T (SIGN (N )) rounds
(the time to complete an exploration and find its label (cid:96)) and then transits to state cruiser. We will prove
that in state cruiser agent A can spend at most T (SIGN (N )) + 2P (N, (cid:96)) rounds. We will use the
9
following claim.
Claim 1. Let t be the first round, if any, in which an agent transits to state token. Then there exists an
agent B that remains in state token and is idle from round t on.
To prove the claim, let Z be the set of agents that transited to state token in round t. In every round
t(cid:48) > t, each agent from Z with the current largest memory remains in state token and stays idle. Indeed,
the reasons why such an agent, call it X, could leave the state token in round t(cid:48) all lead to a contradiction.
There are four such reasons.
• Case 1. The token X was visited by an agent in state explorer of higher seniority. We get a
contradiction with the fact that the agents belonging to Z have the highest seniority.
• Case 2. The explorer of the token X met an agent Y in state token of higher seniority. Since the
explorer of X has the same seniority as X, by transitivity the seniority of agent Y is higher than that
of agent X which is a contradiction with the definition of Z.
• Case 3. In round k < t(cid:48) the token X was visited by an agent Y in state explorer of equal seniority
but such that P refs(MX ) ≺ R where MX is the memory of agent X, s is the last round before
k when agent Y updated its variable recent-token and R is the variable recent-token of agent Y
in round k. This case is impossible. Indeed, by definition, agent X has one of the highest memories
among the agents from Z in round t(cid:48). Hence, according to the definition of order ≺ given in Section 2,
agent X had one of the highest memories among the agents from Z in all rounds between t and t(cid:48).
Moreover, since Y and X have the same seniority, this implies that the token of Y belongs to Z.
Hence, in round s the memory of agent X is greater than or equal to the memory of the token of Y ,
which is a contradiction with P refs(MX ) ≺ R.
• Case 4. In round k < t(cid:48) the explorer of the token X met an agent Y in state token of equal seniority
but such that R ≺ P refs(MY ) where MY is the memory of agent Y , s is the last round before
k when the explorer of X updated its variable recent-token, and R is the variable recent-token
of the explorer of X in round k. Similarly as before, we can get a contradiction with the fact that
R ≺ P refs(MY ).
the agents in Z in all previous rounds, the claim follows.
Since an agent with the largest memory in Z in a given round must have had the largest memory among
•
In order to prove our upper bound on the time spent by A in state cruiser, observe that after at most
T (SIGN (N )) rounds since A transits to state cruiser, all other agents have quit state setup. Consider
the additional 2P (N, (cid:96)) rounds during which agent A performs T Z((cid:96)). Let round τ be the end of the first
half of this segment S of 2P (N, (cid:96)) rounds. Some meeting must have occurred on or before round τ, due to
the properties of T Z. If agent A was involved in one of those meetings, it left state cruiser by round τ.
Otherwise, it must have met some other agent in state either cruiser or token during the second half of
the segment S. Indeed, if it does not meet another agent in state cruiser, it must meet another agent in
state token, which transited to this state by round τ. (Claim 1 guarantees the existence of such an agent
after round τ.) This proves our upper bound on the time spent by A in state cruiser.
From state cruiser agent A can transit to one of the three states: shadow, explorer or token.
To deal with the state shadow, we need the following claim.
Claim 2. If agent A becomes the shadow of an agent B in some round t, then agent B cannot itself switch
to state shadow in the same round.
To prove the claim, there are 3 cases to consider.
10
• Case 1. Agent A transits from state token to state shadow in round t. According to the algorithm,
agent B is an explorer transiting to state searcher in round t.
• Case 2. Agent A transits from state searcher to state shadow in round t. According to the
algorithm, B is
-- either an agent in state token that is not with its explorer in round t, in which case agent B
remains in state token in round t
-- or an agent in state explorer. However an agent in state explorer cannot switch directly
to state shadow. Hence B cannot transit to state shadow in round t.
• Case 3. Agent A transits from state cruiser to state shadow in round t. According to the algo-
rithm, B is either in one of the situations described in Case 2, in which case B does not switch to state
shadow in round t, or B is an agent in state cruiser that transits to state explorer.
•
In all cases, B does not switch to state shadow in round t, which proves the claim.
In view of Claim 2 and of the fact that the termination conditions for an agent in state shadow are the
same as of its guide, we may eliminate the case of state shadow from our analysis.
Consider an agent A in state explorer. After waiting time of T (SIGN (N )) +P (N, L) rounds,
where L is the largest possible label (it is polynomial in N), agent A knows that all other agents have
already transited from the state cruiser (they used at most T (SIGN (N )) rounds in state setup and
at most P (N, L) rounds in state cruiser, as their labels are at most L and at least one token is already
present in the graph).
In what follows, we show that, after at most a polynomial time ρ, agent A either leaves state explorer
or declares that gathering is over.
In order to prove this, we first compute an upper bound on the number of non-clean explorations
EXP LO∗(N ) that can be performed by agent A as an explorer. An exploration could be non-clean due to
several reasons, according to the description of the algorithm.
• In round t(cid:48) agent A met an agent C in state token of higher seniority than that of A, or of equal
seniority but such that recent-token ≺ P reft(MC), where t is the last round before round t(cid:48) when
the variable recent-token of A was updated. According to the algorithm, agent A transits to state
searcher as soon as it terminates its exploration EXP LO∗(N ) after round t(cid:48). Hence such a
meeting can make at most 1 non-clean exploration.
• In round t(cid:48) the token B of A was visited by an agent C in state explorer of higher seniority than
that of B, or of equal seniority but such that P reft(MB) ≺ R, where R is the variable recent-
token of agent C and t is the last round before round t(cid:48) when the variable recent-token of C was
updated. According to the algorithm, agent A transits to state searcher as soon as it terminates
its exploration EXP LO∗(N ) after round t(cid:48). Hence such a meeting can make at most 1 non-clean
exploration.
• Either agent A or its token B met an agent in state searcher. Since the lifespan of a searcher is
at most the time of two consecutive executions of EXP LO∗(N ), it can overlap at most three con-
secutive executions of this procedure. Hence one searcher can make non-clean at most 6 explorations
(3 by meeting A and 3 by meeting B). Since there are at most N searchers, this gives at most 6N
non-clean explorations.
• In round t(cid:48) agent A met an agent C in state token of lower seniority than that of A, or of equal
seniority but such that P reft(MC) ≺ recent-token, where t is the last round before t(cid:48) when the
11
variable recent-token of A was updated. After this meeting, the remaining time when agent C
remains in state token is at most the duration of one execution of EXP LO∗(N ) (after at most
this time the explorer of C becomes searcher and hence C transits to state shadow). This time can
overlap at most two consecutive executions of EXP LO∗(N ), hence such meetings can make at most
2N non-clean explorations.
• In round t(cid:48) the token B of A met an agent C in state explorer of lower seniority than that of B,
or of equal seniority but such that recent-token ≺ P reft(MB) (where t is the last round before
round t(cid:48) when the variable recent-token of C was updated). A similar analysis as in the previous
case shows that such meetings can make at most 2N non-clean explorations.
• Agent A met an agent C in state explorer. The memories of the two agents at this time are
different. After this meeting, the remaining time when agent C remains in state explorer is at
most the duration of two consecutive executions of EXP LO∗(N ) because after the return of C on
its token, the tokens of A and C have different memories and hence after another exploration, C must
become a searcher. Indeed, since by assumption A remains in state explorer till the end of the
algorithm, we must have R ≺ P reft(MB), where R is the variable recent-token of C at the time
t, where t is the first round after the meeting of A and C, in which agent C updated its variable
recent-token. This gives at most 3N non-clean explorations.
• A met an agent C in state token in round s, that looked like its token B at this time, but that turned
out not to be the token B after the backtrack of A on B. More precisely, recent-token = P reft(Mc)
in round s (where t is the last round before round s when the variable recent-token of A was updated)
but P refs(MB) (cid:54)= P refs(MC). After round s agent C remains in state token for at most the
duration of two executions of EXP LO∗(N ). This gives at most 3N non-clean explorations.
• In round t(cid:48) the token B was visited by an agent C of equal seniority in state explorer such that
recent-token = P reft(MB), where t is the last round before t(cid:48) when the variable recent-token
of C was updated and this agent turned out not to be A after the backtrack of A on B. Similarly as
before, this gives at most 3N non-clean explorations.
Hence there can be at most 19N + 2 non-clean executions of EXP LO∗(N ) for agent A (notice that,
e.g., an agent can make non-clean one exploration in the state explorer and then in the state searcher,
hence for simplicity we add all the above upper bounds). A similar analysis shows that during at most 19N +
2 waiting periods of a duration 2T (EXP LO(N )) agent A can be met by a new agent. Recall that before
performing the first execution of EXP LO∗(N ) agent A has been waiting for T (SIGN (N )) + P (N, L)
rounds. Hence if agent A has not left state explorer after at most ρ = T (SIGN (N )) + P (N, L) +
(38N +5)(4T (EXP LO(N )) rounds since it transited to state explorer, there has been a clean execution
of EXP LO∗(N ) followed by a waiting period without any new agent coming during the period of ρ rounds,
and thus agent A declares that gathering is over by the end of this period. Otherwise, agent A transits to state
searcher before spending ρ rounds in state explorer, in which case it uses at most 2·T (EXP LO(N ))
rounds for one execution of EXP LO∗(N ) and after additional at most 2 · T (EXP LO(N )) rounds it finds
an idle agent X in state token (claim 1 guarantees the existence of such an agent): it then becomes the
shadow of either X or of the explorer of X.
It remains to consider an agent A in state token. From this state, either at some point the agent transits
to state shadow or it remains in state token till the end of the algorithm. In this latter case, its explorer
declares that gathering is over after at most T (SIGN (N )) + P (N, L) + (38N + 5)(4T (EXP LO(N ))
rounds since it transited to state explorer. However, as soon as an explorer declares that gathering is
over, its token does the same. So, agent A declares that gathering is over after at most T (SIGN (N )) +
12
P (N, L) + (38N + 5)(4T (EXP LO(N )) rounds since it transited to state token (recall that, according
to the algorithm, agent A and its explorer have reached their current state at the same time).
Hence every agent eventually terminates. We conclude by observing that the execution time of the
entire algorithm is upper bounded by the sum of the following upper bounds:
• the time between the wake up of the first agent and the time of the wake up of an agent A that will be in
state explorer when declaring that gathering is over; this time is upper bounded by T (SIGN (N )).
• the time that such an agent A spends in state setup and cruiser
• the time that such an agent A spends in state explorer
We have shown above that each of these upper bounds is O(T (SIGN (N ))+P (N, L)+N·T (EXP LO(N ))),
where L is polynomial in N. Since the values of T (EXP LO(N )), T (SIGN (N )) and P (N, L) are all
polynomial in N, this proves that the running time of Algorithm Gathering-with-Detection is polynomial in
(cid:3)
N.
In the sequel we will use the following notion, which is a generalization of the enhanced view of a
node. Consider a configuration of agents in any round. Color nodes v and w with the same color if and only
if they are occupied by agents A1, . . . , Ar and B1, . . . , Br, respectively, where Ai and Bi have the same
memory in this round. A colored view from node v is the view from v in which nodes are colored according
to the above rule.
In view of Lemma 3.2, all agents eventually declare that gathering is over. Hence the final configuration
must consist of agents in states explorer , token and shadow, all situated in nodes v1, . . . , vk, such
that in each node vi there is exactly one agent Ei in state explorer, exactly one agent Ti in state token
and possibly some agents in state shadow. Call such a final configuration a clone configuration if there
are at least two distinct nodes vi, vj which have identical colored views. We will first show that the final
configuration cannot be a clone configuration and then that it must consist of all agents gathered in a unique
node and hence our algorithm is correct.
Lemma 3.3 The final configuration cannot be a clone configuration.
Proof. Suppose for contradiction that the final configuration in round f contains distinct nodes which have
identical colored views. Let A be one of the agents woken up earliest by the adversary. There exists an agent
A(cid:48) (also woken up earliest by the adversary) which has an identical memory as A and an identical colored
view. Notice that if two agents have the same memory at time t they must have had the same memory at
time t − 1. Since colors in a colored view are decided by memories of agents, this implies (by a backward
induction on the round number) that the colored views of A and A(cid:48) are the same in each round after their
wake-up, and in particular in the round of their wake-up. In this round no agent has moved yet and hence
each agent is in a different node. Hence colored views in this round correspond to enhanced views. Thus
we can conclude that the enhanced views from the initial positions of agents A and A(cid:48) were identical, which
(cid:3)
contradicts the assumption that in the initial configuration every agent has a unique enhanced view.
Lemma 3.4 In the final configuration all agents must be at the same node.
Proof. It follows from the formulation of the algorithm that at least one agent transits to state token. By
Claim 1 in the proof of Lemma 3.2, there exists an agent that remains in the state token till the end of the
algorithm. By Lemma 3.2, this agent declares that gathering is over. Let B be the first (or one of the first)
agents in state token that declares that gathering is over. Let A be its explorer. Let τ0 be the round in which
13
agent A starts its last exploration EXP LO∗(N ). Let τ1/2 be the round in which backtrack begins during
this execution. Let τ1 be the round in which this backtrack (and hence the execution of EXP LO∗(N )) is
finished, and let τ2 be the round in which A declares that gathering is over.
Claim 1. In round τ0 all agents in state token have the same memory.
In order to prove the claim we first show that all agents in state token in round τ0 have the same
seniority. Observe that there cannot be any agent in state token of higher seniority than B: at least one
of such agents would be seen by A during its last clean exploration EXP LO∗(N ) between rounds τ0 and
τ1 contradicting its cleanliness. Also there cannot be any agent C in state token of lower seniority than
B. Indeed, let D be the explorer of C. Either D becomes a searcher between τ0 and τ1 and thus it
meets the token B before time τ2 which contradicts the declaration of A and B at time τ2 or it remains an
explorer, in which case C remains a token between τ0 and τ1 and thus C is visited by A during its last
clean exploration, contradicting its cleanliness. This shows that all agents in state token in round τ0 have
the same seniority. Hence their explorers start and finish EXP LO∗(N ) at the same time. Consequently
no token existing in round τ0 can transit to state shadow before round τ1. Agent A must have seen all
these tokens during its last exploration. It follows that the memory of each such token in round τ0 must be
equal to the memory of B at this time: otherwise, agent A would detect such a discrepancy during its last
•
exploration, which would contradict the cleanliness of this exploration. This proves Claim 1.
Claim 1 implies that in time τ0 all agents in state explorer have the same memory. Indeed, since
at time τ0 agent A is together with B, each explorer must be with its token, since tokens have the same
memory.
Claim 2. In round τ0 there are no agents in state searcher.
Suppose for contradiction that there is a searcher S in round τ0. Recall that S performs two explo-
rations: one entire exploration EXP LO∗(N ) and another partial exploration EXP LO∗(N ) until meeting
a token or an explorer.
Case 1. S finished its first exploration EXP LO∗(N ) by round τ0.
Hence its second exploration ends by round τ1. It could not end by round τ0 because S would not be a
searcher in this round anymore. If it ended between τ0 and τ1, it must have met a token C. By Claim 1,
all explorers have the same seniority and hence at time τ1 the explorer D of C backtracked to C. This
exploration is not clean for D. Either D becomes a searcher at time τ1 and thus meets A and B before
time τ2, contradicting their declaration at time τ2, or D starts another EXP LO∗(N ) and it meets itself A
and B before time τ2, contradicting their declaration at time τ2. This shows that the second exploration of
S cannot end between τ0 and τ1, hence Case 1 is impossible.
Case 2. S finished its first exploration EXP LO∗(N ) between τ0 and τ1/2.
Hence it must visit some token C during its second exploration (and before starting the backtrack) by round
τ1. As before, this contradicts the declaration of A and B at time τ2.
Case 3. S finished its first exploration EXP LO∗(N ) between τ1/2 and τ1.
Hence the entire backtrack during this first exploration took place between rounds τ0 and τ1. During this
backtrack, S visited some token. As before, this contradicts the declaration of A and B at time τ2.
Case 4. S finished its first exploration EXP LO∗(N ) after round τ1.
This is impossible, as it would not be in state searcher in round τ0.
•
Claim 3. Let E be the set of agents in state explorer in round τ0. In round τ1/2 every agent from E can
reconstruct its colored view in round τ0.
To prove the claim first note that since agent A starts its last exploration in round τ0 and all agents from
E have the same memory in round τ0, they all start an exploration EXP LO∗(N ) in this round. In round
This concludes the proof of Claim 2.
14
τ1/2 every agent from E has visited all nodes of the graph and starts its backtrack. In round τ0 there are
no agents in state setup or cruiser, in view of the waiting time when A transited to state explorer,
and there are no agents in state searcher by Claim 2. Hence the visit of all nodes between rounds τ0 and
τ1/2 permits to see all agents that were tokens at time τ0. Since at this time every explorer were with its
token, this permits to reconstruct the memories and the positions of all agents in round τ0. This is enough to
•
reconstruct the colored views of all agents in round τ0, which proves the claim.
To conclude the proof of the lemma it is enough to show that in round τ0 only one node is occupied by
agents, since this will be the final configuration. Suppose that nodes v (cid:54)= v(cid:48) are occupied in this round. Let
A be the explorer at v and A(cid:48) the explorer at v(cid:48). Note that the colored views of A and A(cid:48) in round τ0 must
be different, for otherwise the configuration in round τ0 would be a clone configuration, and consequently
the final configuration would also be clone, contradicting Lemma 3.3. Since, by Claim 3, in round τ1/2
each of the agents A and A(cid:48) has reconstructed its colored view in round τ0, their memories in round τ1/2
are different. Between rounds τ1/2 and τ1, during its backtrack, agent A(cid:48) has visited again all tokens, in
particular the token of A. Hence A, after backtracking to its token in round τ1, realizes that another explorer
has visited its token, which contradicts the cleanliness of the last exploration of A. This contradiction shows
that in round τ0 only one node is occupied and hence the same is true in the final configuration. This
(cid:3)
concludes the proof of the lemma.
Now the proof of Theorem 3.1 follows directly from Lemmas 3.1, 3.2, and 3.4.
4 Unknown upper bound on the size of the graph
In this section we show that, if no upper bound on the size of the graph is known, then there is no universal
algorithm for gathering with detection all gatherable configurations. Nevertheless, we still show in this case
a universal algorithm that gathers all gatherable configurations: all agents from any gatherable configuration
eventually stop forever at the same node (although no agent is ever sure that gathering is over). The time
of such an algorithm is the number of rounds between the wake-up of the first agent and the last round in
which some agent moves. Our algorithm is polynomial in the (unknown) size of the graph.
We first prove the following negative result.
Theorem 4.1 There is no universal algorithm for gathering with detection all gatherable configurations in
all graphs.
Proof. Consider the following initial configurations. In configuration C the graph is a 4-cycle with clock-
wise oriented ports 0,1 at each node, and with additional nodes of degree 1 attached to two non-consecutive
nodes. There are two agents starting at a node of degree 2 and at its clockwise neighbor, cf. Fig. 1 (a). In
configuration Dn, for n = 4k, the graph is constructed as follows. Take a cycle of size n with clockwise
oriented ports 0,1 at each node. Call clockwise consecutive nodes of the cycle v0, . . . , vn−1 (names are used
only to explain the construction) and attach two nodes of degree 1 to v0 and one node of degree 1 to every
other node with even index. Initial positions of agents are at nodes vi, where i = 4j or i = 4j − 1, for some
j, cf. Fig. 1 (b).
Each of the configurations C and Dn, for n ≥ 8, is gatherable. Indeed, in each of these configurations
there exist agents with different views (agents starting at nodes of degree 2 and of degree 3) and each
agent has a unique enhanced view (this is obvious for configuration C and follows from the existence of a
unique node of degree 4 for configurations Dn). Hence each of these configurations satisfies condition G and
consequently, by Theorem 3.1, there is an algorithm for gathering with detection each specific configuration,
as such a dedicated algorithm knows the configuration and hence may use the knowledge of the size of the
graph.
It remains to show that there is no universal algorithm that gathers with detection all configurations
C and Dn. Suppose, for contradiction, that A is such an algorithm. Suppose that the adversary wakes up
15
Figure 1: Configurations C and D8 in the proof of Theorem 4.1. Black nodes are occupied by agents
all agents simultaneously and let t be the time after which agents in configuration C stop at the same node
and declare that gathering is over. Consider the configuration D8t and two consecutive agents antipodal
to the unique node of degree 4, i.e., starting from nodes v4t and v4t−1. Call X the agent starting at a
node of degree 2 in configuration C and call Y the agent starting at its clockwise neighbor (of degree 3)
in this configuration. Call X(cid:48) the agent starting at node v4t−1 and call Y (cid:48) the agent starting at node v4t in
configuration D8t. (Again names are used only to explain the construction.)
In the first t rounds of the executions of algorithm A starting from configurations C and D8t the
memories of the agents X and X(cid:48) and of the agents Y and Y (cid:48) are the same. This easily follows by induction
on the round number. Hence after t rounds agents X(cid:48) and Y (cid:48) starting from configuration D8t stop and
(falsely) declare that gathering is over. This contradicts universality of algorithm A.
(cid:3)
Our final result is a universal algorithm gathering all gatherable configurations, working without any
additional knowledge. It accomplishes correct gathering and always terminates but (as opposed to Algorithm
Gathering-with-Detection which used an upper bound on the size of the graph), this algorithm does not have
the feature of detecting that gathering is over. We first present a high-level idea of the algorithm, then
describe it in detail and prove its correctness. Recall that we assume that the initial configuration satisfies
condition G (otherwise gathering, even without detection, is impossible by Lemma 3.1).
Idea of the algorithm.
Since in our present scenario no upper bound on the size of the graph is known, already guaranteeing
any meeting between agents must be done differently than in Algorithm Gathering-with-Detection. After
wake-up each agent proceeds in phases i = 1, 2, . . . , where in phase i it "supposes" that the graph has size
at most 2i. In each phase an appropriate label based on procedure SIGN (2i) is computed and procedure
T Z is performed sufficiently long to guarantee a meeting at most at the end of phase (cid:100)log2 m(cid:101), where m is
the real size of the graph. If no meeting occurs in some phase for a sufficiently long time, the agent starts
the next phase.
16
0000111122000000000011111112220000023 (a) CONFIGURATION C (b) CONFIGURATION D8 Another important difference occurs after the meeting, when one of the agents becomes an explorer
and the other its token. Unlike in the case of known upper bound on the size of the graph, there is no way
for any explorer to be sure at any point of the execution that it has already visited the entire graph. Clearly
procedure EXP LO(m) cannot give this guarantee, as m is unknown, and procedure EST of exploration
with a stationary token, which does not require the knowledge of an upper bound, cannot give this guarantee
either, as an explorer cannot be always sure that it visits its own token, because memories of several agents
playing the role of the token can be identical at various stages of the execution, and hence these "tokens"
may be undistinguishable for the explorer.
Nevertheless, our algorithm succeeds in accomplishing the task by using a mechanism which is analo-
gous to the "butterfly effect". Even a slight asymmetry in a remote part of the graph is eventually commu-
nicated to all agents and guarantees that at some point some explorer will visit the entire graph (although
in some graphs no explorer can ever be sure of it at any point of an execution) and then all agents will
eventually gather at the token of one of these explorers. Making all agents decide on the same token uses
property G and is one of the main technical difficulties of the algorithm.
Algorithm Gathering-without-Detection
Similarly as in Algorithm Gathering-with-Detection, an agent can be in one of the following five states:
traveler, shadow, explorer, token, searcher. State traveler partly combines the roles of
previous states setup and cruiser. For every agent A in state shadow the notion of guide is defined
as before. Below we describe the actions of an agent A in each of the states and the transitions between the
states. At wake-up agent A enters the state traveler.
State traveler.
In this state agent A works in phases i = 1, 2, . . . . In phase i the agent supposes that the graph has
size at most n = 2i. Agent A performs SIGN (2i) in order to visit all nodes (and wake up all still dormant
agents), if the assumption is correct, and find the current signature of its initial position v, called the label
(cid:96)n = (cid:96)2i of agent A. Let L2i be the maximum possible label of an agent in phase i. (Note that L2i is
polynomial in n). Then agent A performs T Z((cid:96)n) for ∆n rounds, where ∆n = ∆2i = T (SIGN (2i)) +
2P (2i, L2i) + Σi−1
j=1Qj, for n ≥ 4 (i.e., i ≥ 2), and ∆2 = T (SIGN (2)) + 2P (2, L2). In the formula for
∆n, Qj is defined as T (SIGN (2j)) + ∆2j and is an upper bound on the duration of phase j. Note that
by induction on i we can prove that ∆2i = T (SIGN (2i)) + 2P (2i, L2i) + Σi−1
j=1[2j(T (SIGN (2(i−j))) +
P (2(i−j), L2(i−j)))]. Hence ∆n is upper-bounded by
2ii(T (SIGN (2i)) + 2P (2i, L2i)) = n log2(n)(T (SIGN (n)) + 2P (n, Ln)) which is a polynomial in n.
If no agent has been met during phase i, agent A starts phase i + 1. As soon as another agent is met in
some phase k, agent A interrupts this phase and transits either to state shadow or token or explorer.
Suppose that the first meeting of agent A occurs in round t at node v.
Case 1. There are some agents in round t at node v which are either in state searcher, or explorer or
token .
Let H be the set of these agents.
Subcase 1.1. There are some agents in H that are either in state explorer or token. Let I be the
set of all those agents in H. Agent A transits to state shadow and its guide is the agent having the largest
memory in set I.
Subcase 1.2. There is no agent in H that is either in state explorer or token. Agent A transits to
state shadow and its guide is the agent in state searcher having the largest memory in set H.
Case 2. There are only agents in state traveler in round t at node v.
Subcase 2.1. Agent A has the largest memory among all agents in round t at node v.
Then agent A transits to state explorer.
Subcase 2.2. Agent A does not have the largest memory among all agents in round t at node v.
17
If there is exactly one agent B with memory larger than A, then agent A transits to state token. Otherwise,
it transits to state shadow of the agent with largest memory.
(Note that cases 1 and 2 cover all possibilities because an agent in state shadow always accompanies
its guide and this guide cannot be an agent in state traveler.)
State shadow.
Agent A has exactly one guide and is at the same node as the guide in every round. In every round it
makes the same move as the guide. If the guide B transits itself to state shadow and gets agent C as its
guide, then agent A changes its guide to C as well.
In the description of the actions in the three remaining states, we will use the notion of seniority defined
for Algorithm Gathering-with-Detection.
State explorer.
When agent A transits to state explorer, there is another agent B that transits to state token in the
same round at the same node v. Agent B is called the token of A. Agent A has a variable recent-token that
it initializes to the memory of B in this round.
We first define the notion of a consistent meeting for agent A. Let t be the last round when agent A
updated its variable recent-token. A consistent meeting for A is a meeting in round t(cid:48) > t with an agent
C in state token of the same seniority as A, such that M is the current memory of C and P reft(M) =
recent-token. Intuitively, a consistent meeting is a meeting of an agent that A can plausibly consider to be
its token B. Note that, according to this definition, a meeting in the round when the variable recent-token
is updated, is not a consistent meeting.
We now briefly describe the procedure EST based on [13] that will be subsequently adapted to our
needs and which allows an agent to construct a BFS tree of the network provided that it cannot confuse
its token with another one. The agent constructs a BFS tree rooted at its starting node r marked by the
stationary token. In this tree it marks port numbers at all nodes. During the BFS traversal, some nodes are
added to the BFS tree. In the beginning, the agent adds the root r and then it makes the process of r. The
process of a node w consists in checking all the neighbors of w in order to determine whether some of them
have to be added to the tree or not. When an agent starts the process of a node w, it goes to the neighbor
reachable via port 0 and then checks the neighbor.
When a neighbor v of w gets checked, the agent verifies if v is equal to some node previously added to
the tree. To do this, for each node u belonging to the current BFS tree, the agent travels from v using the
reversal q of the shortest path q from r to u in the BFS tree (the path q is a sequence of port numbers). If at
the end of this backtrack it meets the token, then v = u: in this case v is not added to the tree as a neighbor
of w and is called w-rejected. If not, then v (cid:54)= u. Whether node v is rejected or not, the agent then comes
back to v using the path q. If v is different from all the nodes of the BFS tree, then it is added to the tree.
Once node v is added to the tree or rejected, the agent makes an edge traversal in order to be located at
w and then goes to a non-checked neighbor of w, if any. The order, in which the neighbors of w are checked,
follows the increasing order of the port numbers of w.
When all the neighbors of w are checked, the agent proceeds as follows. Let X be the set of the shortest
paths in the BFS tree leading from the root r to a node y having non-checked neighbors. If X is empty then
procedure EST is completed. Otherwise, the agent goes to the root r, using the shortest path from w to r in
the BFS tree, and then goes to a node y having non-checked neighbors, using the lexicographically smallest
path from X. From there, the agent starts the process of y.
Note that given a graph G of size at most n, every execution of EST in G lasts at most 8n5 rounds.
Indeed, processing every node v takes at most 4n3 rounds (because each node v has at most n− 1 neighbors,
checking a neighbor of v takes at most 2n2 rounds, and before (resp. after) each checking of a neighbor w
of v, the agent makes an edge traversal from v to w (resp. w to v)). Considering the fact that there are at
18
most n nodes to process in G and the fact that moving from a node that has been processed to the next node
to process costs at most 2n rounds, we get the upper bound of 8n5 rounds. Hereafter we define T (EST (n))
as being equal to 8n5.
The procedure EST (cid:48) is a simulation of EST with the following two changes. The first change concerns
the beginning of the execution of EST (cid:48) when the agent is with its token: it updates its variable recent-token
w.r.t to the current memory of its token. The second change concerns meetings with the token. Consider
a verification if a node w , which is getting checked, is equal to some previously constructed node u. This
verification consists in traveling from w using the reverse path q, where q is the path from the root r to u in
the BFS tree and checking the presence of the token. If at the end of the simulation of path q in EST (cid:48) agent
A makes a consistent meeting, then it acts as if it saw the token in EST ; otherwise it acts as if it did not see
the token in EST .
To introduce the next proposition, we first need to define the notion of a truncated spanning tree. We
say that a tree T is a truncated spanning tree of a graph G if T can be obtained from a spanning tree of G by
removing one or more of its subtrees.
Proposition 4.1 Given a graph G of size of at most n (unknown to the agents), the following two properties
hold: (1) every execution of EST (cid:48) in G lasts at most T (EST (n)) rounds and (2) every execution of EST (cid:48)
produces a spanning tree of G or a truncated spanning tree of G.
Proof. If the agent never confuses its token with another one, the proposition follows directly. So in this
proof, we focus only on the situations where there are possible confusions among tokens. When such
confusions may occur?
The execution of procedure EST (cid:48) consists of alternating periods of two different types. The first one
corresponds to periods when the agent processes a node and the second one corresponds to those when the
agent moves to the next node to process it. During the periods of the second type, an agent does not use any
token to move: it follows the same path regardless of whether it meets some token or not on its path. Hence,
an agent can confuse its token with another one only in periods of the first type. During such periods, an
agent may indeed be "mislead" by a token which is not its own token, when verifying whether a node has
to be rejected or not, by wrongly rejecting a node. This leads to the construction of a spanning tree which is
truncated, which proves the second property of the proposition.
Concerning the first property, note that, as for procedure EST , every execution of EST (cid:48) in G lasts at
most 8n5. Indeed, processing every node v also takes at most 4n3 rounds (as each node v has at most n − 1
neighbors, checking a neighbor of v takes at most 2n2 rounds, and before (resp. after) each checking of a
neighbor w of v, the agent makes an edge traversal from v to w (resp. w to v)). Besides, still for the same
reasons as for procedure EST , there are at most n nodes to process in G and moving from a node that has
been processed to the next node to process costs at most 2n rounds. Hence, we also obtain the upper bound
(cid:3)
of 8n5 rounds. Since T (EST (n)) = 8n5, the first property of the proposition follows.
The procedure EST ∗ is a simulation of EST (cid:48) with the following change. Suppose that the execution
of EST (cid:48) produced the route α of the agent. In procedure EST ∗, upon completing procedure EST (cid:48), the
agent traverses the reverse route α and then again α and α. Hence in procedure EST ∗ the agent traverses
the concatenation of routes α, α, α, α. These parts of the trajectory will be called, respectively, the first,
second, third and fourth segment of EST ∗. The variable recent-token is updated at the beginning of the
first and third segment of EST ∗. Note that in these rounds agent A is certain to be with its token. The
(possibly truncated) spanning tree resulting from the simulation EST ∗ is the (possibly truncated) spanning
tree resulting from the execution of the first segment.
In view of the description of procedure EST ∗ and Proposition 4.1, we have the following proposition.
19
Proposition 4.2 Given a graph G of size n (unknown to the agents), the following two properties hold: (1)
every execution of EST ∗ in G lasts at most 4T (EST (n)) rounds and (2) every execution of EST ∗ produces
a spanning tree of G or a truncated spanning tree of G.
Similarly as for EXP LO∗(n), an execution of EST ∗ is called clean if the following condition is
satisfied: in each round during this execution, in which A met an agent C that is not in state shadow, the
memory of C is equal to that of B, and in each round during this execution, in which the token B was met
by an agent D, the memory of D was equal to that of A. Notice that after the execution of EST ∗, agent A is
together with its token B and thus they can verify if the execution was clean, by inspecting their memories.
After transiting to state explorer, agent A executes the following protocol:
repeat forever
/*Before the first turn of the loop agent A has just entered state explorer and is with its token. After each
turn of the loop, agent A is with its token, waiting after a clean exploration.*/
if A has just transited to state explorer or A has just been visited by another agent then
do
EST ∗
if in round t(cid:48) agent A met an agent C in state token of higher
seniority than that of A or of equal seniority but such that
recent-token ≺ P reft(MC) where MC is the memory of agent C
and t is the last round before round t(cid:48) when agent A updated its
variable recent-token then A transits to state searcher
if in round t(cid:48) agent A met another agent C in state explorer, such
that either the seniority of C is higher than that of A, or these
seniorities are equal but P refmin(tA,tC )(RA) ≺ P refmin(tA,tC )(RC),
where RA (resp. RC) is the value of the variable recent-token of A
(resp. C) at the time of the meeting and tA (resp. tC) is the
last round before t(cid:48) when agent A (resp. C)
updated its variable recent-token then A transits to state searcher
if B was visited in round t(cid:48) by an agent C in state explorer of
higher seniority than that of B or of equal seniority but such that
P reft(MB) ≺ R, where MB is the memory of agent B, R is
the variable recent-token of agent C in round t(cid:48), and t is the last
round before t(cid:48) when the variable R was updated
then A transits to state searcher
until the execution of EST ∗ is clean
State token.
When agent A transits to state token, there is another agent B that transits to state explorer in the
same round at the same node v. Agent B is called the explorer of A. Agent A remains idle at a node v and
does not change its state, except when its explorer B transits to state searcher. In this case it transits to
state shadow and B becomes its guide.
State searcher
After transiting to state searcher agent A performs the sequence of explorations EXP LO(n) for
n = 1, 2, 3, . . . , until it meets an agent in state token or explorer in round t. Let S be the set of these
agents met by A in round t. Agent A transits to state shadow and its guide is the agent from S with largest
20
memory.
The analysis of the algorithm is split into the following lemmas.
Lemma 4.1 In Algorithm Gathering-without-Detection every agent eventually stops after time polynomial
in the size of the graph.
Proof. Let m be the size of the graph (unknown to the agents). Let i = (cid:100)log2 m(cid:101). Let A be any agent. We
may assume that at some point A is woken up (otherwise it would be idle all the time). We will first show
that A must meet some other agent at the end of phase i at the latest. To this end, we need to prove the
following claim.
Claim 1. Let t be the first round, if any, in which an agent transits to state token. Then there exists an
agent B that remains in state token and is idle from round t on.
To prove the claim, let Z be the set of agents that transited to state token in round t. In every round
t(cid:48) > t, the agent from Z with the current largest memory remains in state token and stays idle. Indeed,
the reasons why such an agent, call it X, could leave the state token in round t(cid:48) all lead to a contradiction.
There are six such reasons: four of them are identical to those given in Claim 1 of the proof of Lemma 3.2.
Hence to show the validity of the claim, we only need to deal with the two remaining reasons which are the
following ones.
• The explorer of token X, denoted E, met an agent Y in state explorer of higher seniority. Since agent
E has the same seniority as agent X, by transitivity the seniority of agent Y (and of its token) is higher
than that of agent X, which is a contradiction with the definition of Z.
• In round k < t(cid:48) the explorer of token X, denoted E, met an agent Y in state explorer of equal seniority
but such that P refmin(tE ,tY )(RE) ≺ P refmin(tE ,tY )(RY ), where RE (resp. RY ) is the value of the
variable recent-token of E (resp. Y ) at the time of the meeting and tE (resp. tY ) is the last round
before k when agent E (resp. Y ) updated its variable recent-token. This case is impossible. Indeed,
by definition, agent X is among the agents having the highest memory among the agents from Z in
round t(cid:48). Hence, according to the definition of order ≺ given in Section 2, agent X was among the
agents having the highest memory among the agents from Z in all rounds between t and t(cid:48). Moreover,
since E and Y have the same seniority, this implies that the token of Y belongs to Z. Hence, in round
min(tE, tY ) the memory of agent X is greater than or equal to the memory of the token of Y , which
is a contradiction with P refmin(tE ,tY )(RE) ≺ P refmin(tE ,tY )(RY ).
Since an agent with the largest memory in Z in a given round must have had the largest memory among
•
the agents in Z in all previous rounds, the claim follows.
Now we are ready to prove the following claim.
Claim 2. Agent A must meet some other agent at the end of phase i at the latest.
Assume by contradiction that agent A does not meet any agent by the end of phase i. So, there exists
at least one agent executing the first i phases in state traveler. Let F be the first agent to finish the
execution of phase i. According to the algorithm, phase i is made up of two parts. The first one consists
in performing SIGN (2i) and finding the current signature (cid:96)2i of the initial position of the executing agent.
The signature (cid:96)2i plays the role of the agent's label in the second part of phase i which consists in performing
T Z((cid:96)2i) for ∆2i = T (SIGN (2i)) + 2P (2i, L2i) + Σi−1
j=1Qj is an upper bound on
the sum of durations of phases 1 to i − 1, and L2i is the maximum possible label of an agent in phase i.
Observe that at the end of the execution by agent F , at some round t, of the first part of phase i, all the
agents in the graph are necessarily woken up due to the properties of procedure SIGN (2i) (as m ≤ 2i).
Hence, we consider two cases.
j=1Qj rounds, where Σi−1
21
• No agent meets another agent by round t + T (SIGN (2i)) + Σi−1
j=1Qj. In that case, we know that from
round t + 1 + T (SIGN (2i)) + Σi−1
j=1Qj on, all the agents execute the second part of phase i and for
each of them there remain at least P (2i, L2i) rounds before the end of the second part of phase i. Since
the agents cannot all determine the same signature in phase i, there are at least two agents having two
distinct labels and thus two agents meet by round t(cid:48) = t + T (SIGN (2i)) + P (2i, L2i) + Σi−1
j=1Qj
due to the properties of procedure T Z. So, at least one agent transits to state token by round t(cid:48).
However, the last period of P (2i, L2i) rounds when agent A executes the second part of phase i starts
after round t(cid:48). So, in view of the properties of procedure T Z and of Claim 1, we know that agent A
meets an agent in state token by the end of phase i if it does not meet an agent in another state before.
We get a contradiction with the assumption made at the beginning of this proof.
• At least two agents meet by round t + T (SIGN (2i)) + Σi−1
j=1Qj. In that case, we know that an agent
transits to state token by this round. Using similar arguments as before, we can prove that agent A
meets an agent in state token by the end of phase i if it does not meet an agent in another state before.
Again we get a contradiction with the assumption made at the beginning of this proof.
So, we get a contradiction in all cases. Hence, agent A must meet some other agent at the end of phase
•
i at the latest, which proves the claim.
According to Claim 2, we know that after time at most Σi
j=1Qj =
j=1T (SIGN (2j) + ∆2j agent A transits from state traveler either to state shadow or token or
Σi
explorer. To deal with the state shadow, we need the following claim.
Claim 3. If agent A becomes the shadow of an agent B at some round t, then agent B cannot itself switch
to state shadow in the same round.
To prove the claim, there are 3 cases to consider.
• Case 1. Agent A transits from state token to state shadow in round t. According to the algorithm,
agent B is an explorer transiting to state searcher in round t.
• Case 2. Agent A transits from state searcher to state shadow in round t. According to the
algorithm, B is
-- either an agent in state explorer. However, an agent in state explorer cannot switch
directly to state shadow. Hence B cannot transit to state shadow in round t.
-- or an agent in state token. Let K be the set of agents in state explorer or token that are
at the same node as A in round t. According to the algorithm, agent B is the agent having the
highest memory in K. This implies that agent B is not with its explorer in round t because
an agent in state explorer has a higher memory than its token (refer to the way they are
created from state traveler). However, an agent in state token may transit to state shadow
only if it is with its explorer. Hence agent B remains in state token in round t.
• Case 3. Agent A transits from state traveler to state shadow in round t. Let J be the set of the
other agents that are at the same node v as agent A in round t. We have 3 subcases to consider.
-- There are only agents in state traveler in J .
According to the algorithm, agent B is a traveler transiting to state explorer.
-- There is no agent in state explorer or token in J but at least one agent in state searcher.
According to the algorithm, agent B is an agent in state searcher that does not transit to state
shadow in round t (because an agent in state searcher can transit to state shadow only if it
is with an agent in state explorer or token).
22
-- There is at least one agent in state explorer or token in J .
According to the algorithm, agent B is an agent in state explorer or token that cannot
switch to state shadow in round t for similar reasons as in Case 2.
•
In all cases, B does not switch to state shadow in round t, which proves the claim.
In view of Claim 3 and of the fact that the termination conditions for an agent in state shadow are the
same as of its guide, we may exclude the state shadow from our analysis.
Consider an agent in state explorer. Either at some point it transits to state searcher, in which
case, after executing this transition, it uses at most Σm
i=1T (EXP LO(i)) rounds to perform procedures
EXP LO(i) for i = 1, 2, . . . , m, by which time it must have met some token or explorer (because at least
one token is idle all the time starting from the first round when an agent transits to state token, according to
Claim 1) and hence must have transited to state shadow, or it remains in state explorer till the end of
the algorithm.
We will first show that the total number of rounds in which the agent moves as an explorer is polynomial
in m. This is not enough to show that, after polynomial time, A transits to state searcher or remains idle
forever (as an explorer), since we still need to bound the duration of each period of idleness between any
consecutive periods of moving. This will be addressed later.
Two events can trigger further moves of agent A while it is in state explorer: a meeting causing
a non-clean exploration EST ∗ or a visit of A by some agent, when A stays with its token after a clean
exploration.
We first treat the first of these two types of events and bound the total time of explorations caused
by them. An exploration made by agent A could be non-clean due to several reasons, according to the
description of the algorithm.
• In round s agent A met an agent C in state token of higher seniority than that of A, or of equal
seniority but such that recent-token ≺ P reft(MC), where t is the last round before round s when
the variable recent-token of A was updated. According to the algorithm, agent A transits to state
searcher as soon as it terminates its exploration EST ∗ after round s. Hence such a meeting can
cause at most 1 exploration EST ∗ of A to be non-clean.
• In round s the token B of A was visited by an agent C in state explorer of higher seniority than
that of B, or of equal seniority but such that P reft(MB) ≺ R, where R is the variable recent-
token of agent C, and t is the last round before round s when the variable recent-token of C was
updated. According to the algorithm, agent A transits to state searcher as soon as it terminates its
exploration EST ∗ after round s. Hence such a meeting can cause at most 1 exploration EST ∗ of A
to be non-clean.
• Either agent A or its token B met an agent C in state traveler. Since C transits immediately to
state shadow, all agents in state traveler can cause at most m explorations EST ∗ of agent A to
be non-clean.
• Either agent A or its token B met an agent C in state searcher. Since C transits immediately to
state shadow, all agents in state searcher can cause at most m explorations EST ∗ of agent A to
be non-clean.
• A met an agent C in state token of lower seniority than that of A, or of equal seniority but such that
P refk(MC) ≺ recent-token, where k is the last round before this meeting when agent A updated its
variable recent-token. After this meeting, the remaining time when agent C remains in state token
is less than the longest duration of one execution of EST ∗ (after less than this time the explorer of C
23
becomes searcher and hence C transits to state shadow). Thus, as an agent in state token, agent C
can cause at most 4T (EST (m)) explorations EST ∗ of A to be non-clean. Hence all such meetings
can cause at most m · 4T (EST (m)) explorations EST ∗ of A to be non-clean.
• In round s the token B of A met an agent C in state explorer of lower seniority than that of B,
or of equal seniority but such that recent-token ≺ P reft(MB), (where t is the last round before s
when the variable recent-token of C was updated). A similar analysis as in the previous case shows
that such meetings can cause at most m · 4T (EST (m)) explorations EST ∗ of A to be non-clean.
• In round s agent A met an agent C in state explorer of lower seniority than that of A, or of equal
seniority but such that P refmin(tA,tC )(RC) ≺ P refmin(tA,tC )(RA), where RA (resp. RC) is the
value of the variable recent-token of A (resp. C) at the time of the meeting and tA (resp. tC) is the
last round before s when agent A (resp. C) updated its variable. After the meeting agent C "loses",
i.e., it will transit to state searcher after backtracking to its token. Hence agent C remains in state
explorer for less than 4T (EST (m)) rounds after the meeting. Similarly as before, such meetings
can cause at most m · 4T (EST (m)) explorations EST ∗ of A to be non-clean.
• In round s agent A met an agent C in state explorer of equal seniority and such that P refmin(tA,tC )(RA) =
A (resp. t(cid:48)
P refmin(tA,tC )(RC), where RA (resp. RC) is the value of the variable recent-token of A (resp. C)
at the time of the meeting and tA (resp. tC) is the last round before s when agent A (resp. C) updated
its variable recent-token. Let t(cid:48)
C) be the first round since s when agent A (resp. C) updates
its variable recent-token. From round h = max(t(cid:48)
C) on, this kind of meeting with agent C can-
not occur anymore. Indeed, in view of the fact that the memories of A and C are necessarily different
in round s and the fact that once agent C finishes the execution of EST ∗ involving this meeting, its
variable recent-token will be updated, we know that P refmin(t(cid:48)
C )(RC)
from round h on. Since the difference between max(t(cid:48)
C) and s is less than the longest duration
of one execution of EST ∗, similarly as before, such meetings can cause at most m · 4T (EST (m))
explorations EST ∗ of A to be non-clean.
C )(RA) (cid:54)= P refmin(t(cid:48)
A,t(cid:48)
A,t(cid:48)
A, t(cid:48)
A, t(cid:48)
• In round s agent A met an agent C in state explorer of higher seniority than that of A or of equal
seniority but such that P refmin(tA,tC )(RA) ≺ P refmin(tA,tC )(RC), where RA (resp. RC) is the
value of the variable recent-token of A (resp. C) at the time of the meeting and tA (resp. tC) is the
last round before s when agent A (resp. C) updated its variable. According to the algorithm, agent A
transits to state searcher as soon as it terminates its exploration EST ∗ after round s. Hence such
a meeting can cause at most 1 exploration EST ∗ of A to be non-clean.
• A met an agent C in state token in round s, that looked like its token B at this time, but that turned
out not to be the token B after the backtrack of A on B. More precisely, recent-token = P reft(Mc)
in round s (where t is the last round before s when the variable recent-token of A was updated) but
P refs(MB) (cid:54)= P refs(MC). After round s agent C may look like token B of A for less than
4T (EST (m)) rounds because after less than this time A backtracks to its token B and, from this time
on, it can see the difference between B and C. Similarly as before such meetings can cause at most
m · 4T (EST (m)) explorations EST ∗ of A to be non-clean.
• In round t(cid:48) the token B was visited by an agent C of equal seniority in state explorer such that
recent-token = P reft(MB), where t is the last round before t(cid:48) when the variable recent-token
of C was updated, and this agent turned out not to be A after the backtrack of A on B. Similarly as
before, such meetings can cause at most m · 4T (EST (m)) explorations EST ∗ of A to be non-clean.
24
Hence the first of the two types of events (meeting causing a non-clean exploration) can cause at most
(24T (EST (m)) + 2)m + 3 explorations EST ∗ of A to be non-clean. (As before we add up all upper
bounds for simplicity). Considering the fact that each non-clean exploration EST ∗ is directly followed by
at most one clean exploration EST ∗, this kind of meeting can cause at most 2((24T (EST (m)) + 2)m +
3)4T (EST (m)) rounds of motion of A. The second type of events (a visit of A by some agent, when A
stays with its token after a clean exploration) can cause at most (2m(1 + 4T (EST (m))) + 1)4T (EST (m))
rounds of motion of A. Indeed, according to the algorithm, a visit of A by some agent C when A stays idle
with its token B can be of the following kinds.
• Agents A and B are visited by an agent C in state traveler. Since, at this visit, agent C transits
immediately to state shadow, such visits can trigger at most m exploration EST ∗ of agent A.
• Agents A and B are visited by an agent C in state searcher. Since, at this visit, agent C transits
immediately to state shadow, such visits can trigger at most m exploration EST ∗ of agent A.
• In round s, agents A and B are visited by an agent C in state explorer of lower seniority than
that of B, or of equal seniority but such that RC ≺ P reftC (MB), where MB is the memory of
agent B in round s, RC is the variable recent-token of agent C in round s, and tC is the last round
before s when the variable RC was updated. After the visit agent C "loses", i.e., it will transit to
state searcher after backtracking to its token. Hence agent C remains in state explorer for less
than 4T (EST (m)) rounds after the visit and thus, such visits can trigger at most m · 4T (EST (m))
explorations EST ∗ of A.
• In round s agents A and B are visited by an agent C in state explorer of equal seniority to that of
B and such that RC = P reftC (MB), where MB is the memory of agent B in round s, RC is the
variable recent-token of agent C in round s, and tC is the last round before s when the variable RC
was updated. Let t(cid:48)
C be the first round after s when agent C updates its variable recent-token. From
round t(cid:48)
C on, this kind of visit by agent C cannot occur anymore. Indeed, in view of the fact that the
memories of A and C are necessarily different in round s and the fact that once agent C finishes the
execution of EST ∗, its variable recent-token will be updated, we know that RC (cid:54)= P reft(cid:48)
(MB)
from round t(cid:48)
C and s is less than the longest duration of one
execution of EST ∗, similarly as before, such visits can trigger at most m· 4T (EST (m)) explorations
EST ∗ of A.
C on. Since the difference between t(cid:48)
C
• In round s, agents A and B are visited by an agent C in state explorer of higher seniority than
that of A or of equal seniority to that of B and such that P reftC (MB) ≺ RC, where MB is the
memory of agent B in round s, RC is the variable recent-token of agent C in round s, and tC is the
last round before s when the variable R was updated. According to the algorithm, agent A transits to
state searcher as soon as it terminates its exploration EST ∗ after round s. Hence such visits can
trigger at most 1 exploration EST ∗ of A.
Hence adding the first exploration that must be made by A (which is not trigerred by any meeting), we
get an upper bound of (2((28T (EST (m)) + 3)m + 3) + 1)4T (EST (m)) rounds during which agent A
moves in state explorer.
It remains to consider an agent in state token. It may either transit to state shadow or remain in state
token forever. In the latter case it is idle all the time.
Since Σi
j=1T (SIGN (2j) + ∆2j , T (EST (m)) and T (EXP LO(m)) are all polynomial in m, the
above analysis shows that there exists a polynomial Y , such that, for each agent A executing Algorithm
Gathering-without-Detection in any graph of size m, the number of rounds during which this agent moves
25
is at most Y (m). In order to finish the proof, we need to bound the number of rounds during which an agent
A can be idle before moving again. To do this we will use the following claim.
Claim 4. If in round t of the execution of Algorithm Gathering-without-Detection no agent moves, then no
agent moves in any later round of this execution.
To prove the claim notice that if no agent moves in round t, then in this round no agent is in state
traveler or searcher. Moreover each agent in state explorer must be idle and stay with its token
in this round (all other nodes must be in state shadow). In order for some agent to move in round t + 1,
some explorer would have to visit some other token in round t, contradicting the definition of t. Hence all
•
agents are idle in round t + 1. By induction, all agents are idle from round t on. This proves the claim.
Since for each agent executing Algorithm Gathering-without-Detection in a graph of size m, the num-
ber of rounds in which it moves is at most Y (m) and there are at most m agents, Claim 4 implies that after
time at most m · Y (m) since the wake up of the first agent, all agents must stop forever.
(cid:3)
By Lemma 4.1 there exists a round after which, according to Algorithm Gathering-without-Detection,
no agent moves. Call the resulting configuration final. The following lemma implies that Algorithm
Gathering-without-Detection is correct.
Lemma 4.2 In every final configuration exactly one node is occupied by agents.
Proof. A final configuration must consist of agents in states explorer , token and shadow, all situated
in nodes v1, . . . , vk, such that in each node vi there is exactly one agent Ei in state explorer, exactly
one agent Ti in state token and possibly some agents in state shadow. As before we call such a final
configuration a clone configuration if there are at least two distinct nodes vi, vj which have identical colored
views. The same argument as in the proof of Lemma 3.3 shows that a final configuration cannot be a clone
configuration.
It is enough to prove that k = 1. Suppose for contradiction that k > 1. We will consider two cases.
In the first case the memories of all explorers Ei are identical and in the second case they are not. In both
cases we will derive a contradiction.
Case 1. All explorers Ei in the final configuration have identical memory.
In this case all these explorers performed the last exploration EST ∗ simultaneously, in view of the fact
that the algorithm that they execute is deterministic.
We start with the following claim.
Claim 1. If a node has been rejected by the explorer Ej in the construction of its (truncated) spanning tree
during its last exploration EST ∗, then this node, let us call it w, must have been either added previously
by Ej to its (truncated) spanning tree, or added by another explorer Es in the construction of its (truncated)
spanning tree during its last exploration EST ∗.
The node w was rejected by Ej for the following reason. Ej traveled from w using the reversal q
of the path q, where q is a path (coded as a sequence of ports) from vj to some node u already in the
(truncated) spanning tree of Ej, and at the end of this path q, Ej met a token with memory M, such that
P reft(M) = recent-token, where t is the last round when Ej updated its variable recent-token.
There are two possible cases. If the token met by Ej is its own token (residing at vj), then w is equal
to some node u already added previously to the (truncated) spanning tree of Ej. If, on the other hand, the
token met by Ej is the token of some other explorer Es, then we will show that w is added by Es to its
(truncated) spanning tree. Indeed, since Ej has added a node u to its (truncated) spanning tree, such that the
path from vj to u is q, the explorer Es must have added a node u(cid:48) to its (truncated) spanning tree, such that
the path from vs to u(cid:48) is q as well, because both Es and Ej have identical memories. However, this node u(cid:48)
•
must be equal to w, since the path from w to vs is q. This proves the claim.
26
The contradiction in Case 1 will be obtained in the following way. Using (truncated) spanning trees
produced by explorers Ei during their last exploration EST ∗ (recall that these trees are isomorphic, since
memories of the explorers are identical), we will construct the colored view for each explorer. Using the
fact that memories of the explorers are identical, these colored views will be identical. This will imply that
the final configuration is a clone configuration, which is impossible.
The construction proceeds as follows (we will show it for explorer E1). Let Ti be the (truncated)
spanning tree produced by Ei. Each tree Ti has its root vi colored black and all other nodes colored white.
We will gradually attach various trees to T1 in order to obtain the colored view from v1. First attach to every
node of T1 its neighbors that have been rejected by E1 during the construction of T1. Explorer E1 has visited
these nodes, hence the respective port numbers can be faithfully added. Consider any such rejected node w.
By Claim 1, there are two possibilities. If node w was previously added by E1 to T1 as some node u, then
we proceed as follows. Let T (cid:48)
1 at w, identifying
its root u with w. If node w was added by another explorer Es in the construction of its (truncated) spanning
tree Ts, we proceed as follows. As mentioned in the proof of Claim 1, the explorer Es must have added a
node u(cid:48) to its (truncated) spanning tree, such that the path from vs to u(cid:48) is q. Let T (cid:48)
s be the tree Ts but rooted
at u(cid:48) instead of vs. We attach tree T (cid:48)
1 be the tree T1 but rooted at u instead of v1. We attach tree T (cid:48)
s at w, identifying its root u(cid:48) with w.
After processing all nodes rejected by E1 and adding the appropriate trees, we attach all rejected neigh-
bors of nodes in the newly obtained increased tree. These nodes could have been rejected either by E1 itself
or by another explorer Ej whose (re-rooted) tree T (cid:48)
j has been attached. For each newly attached node
rejected by Ej, the construction continues as before, replacing the role of T1 by Tj.
The above construction proceeds infinitely, producing an infinite rooted tree (rooted at a node corre-
sponding to v1). We make one final addition in order to obtain the view V(v1) from v1 (cf. Claim 2) : each
node u of the tree is assigned a label corresponding to the shortest path from the root to u. The resulting
infinite tree is denoted by T .
Claim 2. T corresponds to view V(v1)
We prove the claim by induction. First of all, note that T truncated at depth 1 from its root corresponds
to the truncated view V 1(v1). Assume as induction hypothesis that the tree T truncated at depth l from
its root, call it T l, corresponds to the truncated view V l(v1). We will show that T (l+1) corresponds to the
truncated view V (l+1)(v1).
According to the process described above, whenever a node u, corresponding to a node x in the network
N , is added to T under construction, exactly one child v is eventually added to u for each neighbor y of x
in N . More precisely, node x is connected to node y by an edge having port p at node x and port q at node y
iff node u and its child v are connected between them by an edge having port p at node u and port q at node
v. In addition, if the label of u is path ρ then the label of v is path ρpq.
Hence, if in tree T we identify a node with its label, we have the following two properties: (1) for
every path π of length 2l (according to the induction hypothesis, π is located at distance l from the root of
T and corresponds to a path from v1 in N ), its children correspond to all paths from v1 of length 2l + 2
whose prefix is π in N and (2) for every path π of length 2l, every child ρ of path π, such that ρ = πpq, is
connected to π in T by an edge that has port p at node π and port q at node ρ.
Moreover, since T l is V l(v1), the above two properties also hold if l is replaced by any value ranging
from 0 to (l − 1). Hence T (l+1) is V (l+1)(v1), which proves the claim.
•
To produce the colored view, notice that there are only two colors in this colored view: white corre-
sponding to empty nodes in the final configuration and black corresponding to nodes v1,..., vk (all these
nodes get identical colors: since memories of explorers are the same, memories of their tokens are also the
same and memories of corresponding nodes in state shadow are also identical). It remains to indicate how
the colors are distributed in the constructed view. This is done as follows. When a tree T (cid:48)
j is attached, exactly
one of its nodes (namely the node corresponding to vj) is black. Exactly these nodes become black in the
27
obtained colored view.
This construction of colored views is done for all explorers Ei. Consider two explorers Ei and Ej.
Since these explorers have the same memory, the trees T (cid:48)
s attached at a given stage of the construction of
the views of Ei and Ej are isomorphic. They are also attached in the same places of the view. Hence by
induction of the level of the view it follows that both colored views are identical. This implies that the final
configuration is a clone configuration which gives a contradiction in Case 1.
Case 2. There are at least two explorers Ei and Ej with different memories in the final configuration.
Consider the equivalence relation on the set of explorers E1, . . . , Ek, such that two explorers are equiv-
alent if their memories in the final configuration are identical. Let C1, . . . ,Ch, where h > 1, be the equiva-
lence classes of this relation. Suppose w.l.o.g. that C1 is a class of explorers with smallest seniority. We will
use the following claim.
Claim 3. During the last exploration EST ∗ of explorers in C1, at least one of the following statements
holds:
• an explorer from C1 has visited a token of an explorer not belonging to C1;
• a token of an explorer from C1 has been visited by an explorer not belonging to C1.
In order to prove the claim consider two cases. If every node of the graph has been visited by some
explorer from C1, we will show that the first statement holds. Indeed, since explorers from C1 have the
smallest seniority, during their last execution of EST ∗ all tokens, which are in the final configuration, are
already at their respective nodes (because otherwise there would be at least one token and its explorer in the
final configuration that were created after the creation of the tokens from C1, which would be a contradiction
with the fact that explorers from C1, and hence also their tokens, have the smallest seniority in the final
configuration). Hence some explorers from C1 must visit the tokens of explorers outside of C1. Hence we
can restrict attention to the second case, when some nodes of the graph have not been visited by any explorer
from C1. Notice that if there were no other classes than C1, this could not occur. Indeed, we would be then
in Case 1 (in which all explorers have identical memory). Thus Claim 1 would hold, which implies that all
nodes must be visited by some explorer, in view of the graph connectivity.
Hence the fact that some node is not visited by explorers from C1 must be due to a meeting of some other
agent (which is neither an explorer from C1 nor a token of such an explorer) during their last exploration
EST ∗. What kind of a meeting can it be? It cannot be a meeting with an agent in state traveler or
searcher because this would contradict that the last exploration was clean. For the same reason it cannot
be a meeting of an explorer from C1 with another explorer. This leaves only the two types of meetings
•
specified in the claim, which finishes the proof of the claim.
Let (Ex, T ok) be a couple of an explorer outside of C1 and of its token, such that either an explorer
from C1 visited T ok or a token of an explorer from C1 has been visited by Ex during the last exploration of
explorers in the class C1. Such a couple exists by Claim 3. The seniority of Ex and T ok must be the same
as that of explorers from C1, for otherwise their last exploration would not be clean. For the same reason,
when explorers from C1 started their last exploration, the explorer Ex must have started an exploration
as well (possibly not its final exploration): otherwise the exploration of explorers from C1 would not be
clean. Moreover we show that when explorers from C1 finished their last exploration, explorer Ex must
have finished an exploration as well. To prove this, consider two cases, corresponding to two possibilities
in Claim 3. Suppose that an explorer Ej from C1 has visited T ok and that its exploration did not finish
simultaneously with the exploration of Ex. Consider the consecutive segments S1, S2, S3, S4 of the last
exploration EST ∗ of Ej. (Recall that these segments were specified in the definition of EST ∗.) Since Ej
has visited T ok during EST ∗, it must have visited it during each segment Si. At the end of S1, explorer Ej
knows how long EST ∗ will take. At the end of S2 its token learns it as well. When Ej visits T ok again
28
in segment S3, there are two possibilities. Either T ok does not know when the exploration of Ex finishes,
or it does know that it finishes at a different time than the exploration of Ej. In both cases the explorer Ej
that updated its variable recent-token at the end of S2 can see that recent-token (cid:54)= P reft(M), where M
is the memory of T ok and t is the end of S2. This makes the last exploration of Ej non clean, which is a
contradiction. This proves that Ex and Ej finish their exploration simultaneously, if Ej has visited T ok.
The other case, when Ex has visited the token of Ej is similar. Hence we conclude that explorations of Ej
and of Ex started and finished simultaneously.
Let τ be the round in which the last exploration of Ej (and hence of all explorers in C1) finished. The
exploration of Ex that finished in round τ cannot be its final exploration because then it would have the
same memory as Ej in the final configuration and thus it would be in the class C1 contrary to the choice of
Ex. Hence Ex must move after round τ. It follows that there exists a class Ci (w.l.o.g. let it be C2) such that
explorers from this class started their last exploration after round τ. Note that during this last exploration,
explorers from C2 could not visit all nodes of the graph, for otherwise they would meet explorers from C1
after round τ, inducing them to move after this round, contradicting the fact that explorers from C1 do not
move after round τ.
The fact that some node is not visited by explorers from C2 must be due to a meeting of some other
agent (which is neither an explorer from C2 nor a token of such an explorer) during their last exploration
EST ∗. Otherwise, for explorers in C2 the situation would be identical as if their equivalence class were
the only one, and hence, as in Case 1, they would visit all nodes. Moreover, the fact that some node is not
visited by explorers from C2 must be due to a meeting of some explorer outside of C1 or of its token (if not,
explorers from C1 would move after round τ, which is a contradiction). An argument similar to that used
in the proof of Claim 3 shows that there exists a couple (Ex(cid:48), T ok(cid:48)), such that Ex(cid:48) is an explorer outside
of C1 ∪ C2, T ok(cid:48) is its token, and either an explorer from C2 visited T ok(cid:48) or a token of an explorer from C2
has been visited by Ex(cid:48) during the last exploration of explorers in the class C2. Let τ(cid:48) be the round in which
the last exploration of explorers from C2 is finished. Similarly as before, the explorer Ex(cid:48) terminates some
exploration in round τ(cid:48) but continues to move afterwards.
Repeating the same argument h − 1 times we conclude that there exists a round τ∗ after which all
explorers from C1 ∪ ··· ∪ Ch−1 never move again, but the last exploration of explorers from Ch starts on or
after τ∗. During this last exploration there must be a node not visited by any explorer from Ch, otherwise
some explorers from C1 ∪ ··· ∪ Ch−1 would move after τ∗. This is due to a meeting. It cannot be a meeting
with an agent in state traveler or searcher because this would contradict that the last exploration was
clean. For the same reason it cannot be a meeting of an explorer from Ch with another explorer. Hence two
possibilities remain. Either an explorer from Ch visits a token of an explorer from C1 ∪···∪Ch−1 or a token
of an explorer from Ch is visited by an explorer from C1∪···∪Ch−1. The first situation is impossible because
it would contradict the cleanliness of the last exploration of explorers from Ch and the second situation is
impossible because explorers from C1 ∪ ··· ∪ Ch−1 do not move after τ∗. Hence in Case 2 we obtain a
(cid:3)
contradiction as well, which completes the proof.
Lemmas 4.1 and 4.2 imply the following result.
Theorem 4.2 Algorithm Gathering-without-Detection performs a correct gathering of all gatherable con-
figurations and terminates in time polynomial in the size of the graph.
5 Consequences for leader election
Leader election [44] is a fundamental symmetry breaking problem in distributed computing. Its goal is to
assign, in some common round, value 1 (leader) to one of the entities and value 0 (non-leader) to all others.
The assignment should happen once for each identity, in a unique common round, and cannot be changed
afterwards. In the context of anonymous agents in graphs, leader election can be formulated as follows:
29
• There exists a common unique round in which one of the agents assigns itself value 1 (i.e., it declares
itself a leader) and each other agent assigns itself value 0 (i.e., it declares itself non-leader).
The following proposition says that the problems of leader election and of gathering with detection
are equivalent in the following strong sense. Consider any initial configuration of agents in a graph. If
gathering with detection can be accomplished for this configuration in some round t, then leader election
can be accomplished for this configuration in some round t(cid:48) > t, and conversely, if leader election can be
accomplished for this configuration in some round t, then gathering with detection can be accomplished for
this configuration in some round t∗ > t.
Proposition 5.1 Leader election is equivalent to gathering with detection.
Proof. Suppose that gathering with detection is accomplished and let t be the round when all agents are
together and declare that gathering is over. As mentioned in the Preliminaries, all agents must have different
memories, since they are at the same node, and, being together, they can compare these memories. Since,
in view of detection, the round t is known to all agents, in round t + 1 the agent with the largest memory
assigns itself value 1 and all other agents assign themselves value 0.
Conversely, suppose that leader election is accomplished and let t be the round in which one of the
agents assigns itself value 1 and all other agents assign themselves value 0. Starting from round t the agent
with value 1 stops forever and plays the role of the token, all other agents playing the role of explorers. First,
every explorer finds the token by executing procedure EXP LO(n) for T (EXP LO(n)) rounds in phases
n = 1, 2, ..., until it finds the token in round t(cid:48) (this round may be different for every explorer). Then every
explorer executes procedure EST (using the token) and finds the map of the graph and hence its size m.
i=1 T (EXP LO(i)) + T (EST (m)). By this round all
explorers must have found the token and executed procedure EST , i.e., they are all together with the token.
In round t∗ all agents declare that gathering is over.
(cid:3)
Then it waits with the token until round t∗ = t +(cid:80)m
Proposition 5.1 implies that the class of initial configurations for which leader election is at all possible
(even only using an algorithm dedicated to this specific configuration) is equal to the class of gatherable
configurations, i.e., to the class of configurations satisfying property G. Similarly as for gathering, we will
say that a leader election algorithm is universal if it performs leader election for all such configurations.
It follows that a small modification of Algorithm Gathering-with-Detection is a universal leader election
algorithm, provided that an upper bound on the size of the graph is known to the agents. The modification is
the following: use Algorithm Gathering-with-Detection to gather all agents in some round t, and, in round
t + 1, elect as leader the agent that has the largest memory in the round of the gathering declaration (this
agent assigns itself value 1 and all other agents assign themselves value 0). Let LE be the name of this
modified algorithm.
The following corollary summarizes the above discussion and gives a complete solution of the leader
election problem for anonymous agents in arbitrary graphs.
Corollary 5.1 For a given initial configuration, leader election is possible if and only if this configuration
satisfies condition G. If an upper bound on the size of the graph is known, then Algorithm LE accomplishes
leader election for all these configurations. There is no universal algorithm accomplishing leader election
for all configurations satisfying condition G in all graphs.
References
[1] Noa Agmon and David Peleg. Fault-tolerant gathering algorithms for autonomous mobile robots.
SIAM Journal on Computing, 36(1):56 -- 82, 2006.
30
[2] Steve Alpern. The rendezvous search problem. SIAM Journal on Control and Optimization, 33(3):673 --
683, 1995.
[3] Steve Alpern. Rendezvous search on labelled networks. Naval Research Logistics, 49:256 -- 274, 2002.
[4] Steve Alpern and Shmuel Gal. Theory of Search Games and Rendezvous. Kluwer Academic Publisher,
2003.
[5] Eddie Anderson and Richard Weber. The rendezvous problem on discrete locations. Journal of Applied
Probability, 28(4):839 -- 851, 1990.
[6] Edward Anderson and S´andor Fekete. Asymmetric rendezvous on the plane. In 14th Annual Sym-
posium on Computational Geometry, June 7-10, 1998, Minneapolis, Minnesota, USA. Proceedings,
pages 365 -- 373, 1998.
[7] Edward Anderson and S´andor Fekete. Two dimensional rendezvous search. Operations Research,
49(1):107 -- 118, 2001.
[8] Dana Angluin. Local and global properties in networks of processors (extended abstract).
In 12th
Annual ACM Symposium on Theory of Computing, April 28-30, 1980, Los Angeles, California, USA.
Proceedings, pages 82 -- 93, 1980.
[9] Hagit Attiya and Marc Snir. Better computing on the anonymous ring.
12(2):204 -- 238, 1991.
Journal of Algorithms,
[10] Evangelos Bampas, Jurek Czyzowicz, Leszek Gasieniec, David Ilcinkas, and Arnaud Labourel. Al-
most optimal asynchronous rendezvous in infinite multidimensional grids. In Distributed Computing,
24th International Symposium, DISC 2010, Cambridge, MA, USA, September 13-15, 2010. Proceed-
ings, pages 297 -- 311, 2010.
[11] Vic Baston and Shmuel Gal. Rendezvous search when marks are left at the starting points. Naval
Research Logistics, 48(8):722 -- 731, 2001.
[12] Paolo Boldi and Sebastiano Vigna. Computing anonymously with arbitrary knowledge. In Proceedings
of the Eighteenth Annual ACM Symposium on Principles of Distributed Computing, PODC, '99At-
lanta, Georgia, USA, May 3-6, 1999, pages 181 -- 188, 1999.
[13] J´er´emie Chalopin, Shantanu Das, and Adrian Kosowski. Constructing a map of an anonymous graph:
Applications of universal sequences. In Principles of Distributed Systems - 14th International Confer-
ence, OPODIS 2010, Tozeur, Tunisia, December 14-17, 2010. Proceedings, pages 119 -- 134, 2010.
[14] Mark Cieliebak, Paola Flocchini, Giuseppe Prencipe, and Nicola Santoro. Distributed computing by
mobile robots: Gathering. SIAM Journal on Computing, 41(4):829 -- 879, 2012.
[15] Reuven Cohen and David Peleg. Convergence properties of the gravitational algorithm in asyn-
chronous robot systems. SIAM Journal on Computing, 34(6):1516 -- 1528, 2005.
[16] Reuven Cohen and David Peleg. Convergence of autonomous mobile robots with inaccurate sensors
and movements. SIAM Journal on Computing, 38(1):276 -- 302, 2008.
[17] Jurek Czyzowicz, Adrian Kosowski, and Andrzej Pelc. How to meet when you forget:
rendezvous in arbitrary graphs. Distributed Computing, 25(2):165 -- 178, 2012.
log-space
31
[18] Jurek Czyzowicz, Andrzej Pelc, and Arnaud Labourel. How to meet asynchronously (almost) every-
where. ACM Transactions on Algorithms, 8(4):37, 2012.
[19] Gianlorenzo D'Angelo, Gabriele Di Stefano, and Alfredo Navarra. Gathering on rings under the look-
compute-move model. Distributed Computing, 27(4):255 -- 285, 2014.
[20] Gianluca De Marco, Luisa Gargano, Evangelos Kranakis, Danny Krizanc, Andrzej Pelc, and Ugo Vac-
caro. Asynchronous deterministic rendezvous in graphs. Theoretical Computer Science, 355(3):315 --
326, 2006.
[21] Anders Dessmark, Pierre Fraigniaud, Dariusz R. Kowalski, and Andrzej Pelc. Deterministic ren-
dezvous in graphs. Algorithmica, 46(1):69 -- 96, 2006.
[22] Yoann Dieudonn´e and Andrzej Pelc. Deterministic network exploration by anonymous silent agents
with local traffic reports. In Automata, Languages, and Programming - 39th International Colloquium,
ICALP 2012, Warwick, UK, July 9-13, 2012, Proceedings, Part II, pages 500 -- 512, 2012.
[23] Yoann Dieudonn´e, Andrzej Pelc, and David Peleg. Gathering despite mischief. ACM Transactions on
Algorithms, 11(1):1, 2014.
[24] Yoann Dieudonn´e, Andrzej Pelc, and Vincent Villain. How to meet asynchronously at polynomial cost.
In ACM Symposium on Principles of Distributed Computing, PODC 2013, Montreal, QC, Canada, July
22-24, 2013, pages 92 -- 99, 2013.
[25] Paola Flocchini, David Ilcinkas, Andrzej Pelc, and Nicola Santoro. Remembering without memory:
Tree exploration by asynchronous oblivious robots. Theoretical Computer Sci.ence, 411(14-15):1583 --
1598, 2010.
[26] Paola Flocchini, Giuseppe Prencipe, Nicola Santoro, and Peter Widmayer. Gathering of asynchronous
robots with limited visibility. Theoretical Computer Science, 337(1-3):147 -- 168, 2005.
[27] Pierre Fraigniaud, Leszek Gasieniec, Dariusz R. Kowalski, and Andrzej Pelc. Collective tree explo-
ration. Networks, 48(3):166 -- 177, 2006.
[28] Pierre Fraigniaud and Andrzej Pelc. Deterministic rendezvous in trees with little memory. In Dis-
tributed Computing, 22nd International Symposium, DISC 2008, Arcachon, France, September 22-24,
2008. Proceedings, pages 242 -- 256, 2008.
[29] Pierre Fraigniaud and Andrzej Pelc. Delays induce an exponential memory gap for rendezvous in trees.
ACM Transactions on Algorithms, 9(2):17, 2013.
[30] Greg N. Frederickson and Nancy A. Lynch. Electing a leader in a synchronous ring. Journal of the
ACM, 34(1):98 -- 115, 1987.
[31] Emanuele G. Fusco and Andrzej Pelc. How much memory is needed for leader election. Distributed
Computing, 24(2):65 -- 78, 2011.
[32] Shmuel Gal. Rendezvous search on the line. Operations Research, 47(6):974 -- 976, 1999.
[33] Samuel Guilbault and Andrzej Pelc. Asynchronous rendezvous of anonymous agents in arbitrary
graphs. In Principles of Distributed Systems - 15th International Conference, OPODIS 2011, Toulouse,
France, December 13-16, 2011. Proceedings, pages 421 -- 434, 2011.
32
[34] Med Amine Haddar, Ahmed Hadj Kacem, Yves M´etivier, Mohamed Mosbah, and Mohamed Jmaiel.
Electing a leader in the local computation model using mobile agents. In 6th ACS/IEEE International
Conference on Computer Systems and Applications, AICCSA 2008, Doha, Qatar, March 31 - April 4,
2008, Proceedings, pages 473 -- 480, 2008.
[35] Daniel S. Hirschberg and J. B. Sinclair. Decentralized extrema-finding in circular configurations of
processors. Communications of the ACM, 23(11):627 -- 628, 1980.
[36] Amos Israeli and Marc Jalfon. Token management schemes and random walks yield self-stabilizing
mutual exclusion. In 9th Annual ACM Symposium on Principles of Distributed Computing, Quebec
City, Quebec, Canada, August 22-24, 1990, Proceedings, pages 119 -- 131, 1990.
[37] Ralf Klasing, Adrian Kosowski, and Alfredo Navarra. Taking advantage of symmetries: Gathering of
many asynchronous oblivious robots on a ring. Theoretical Computer Science, 411(34-36):3235 -- 3246,
2010.
[38] Ralf Klasing, Euripides Markou, and Andrzej Pelc. Gathering asynchronous oblivious mobile robots
in a ring. Theoretical Computer Science, 390(1):27 -- 39, 2008.
[39] Dariusz R. Kowalski and Adam Malinowski. How to meet in anonymous network. Theoretical Com-
puter Science, 399(1-2):141 -- 156, 2008.
[40] Evangelos Kranakis, Danny Krizanc, and Pat Morin. Randomized rendezvous with limited memory.
ACM Transactions on Algorithms, 7(3):34, 2011.
[41] Evangelos Kranakis, Nicola Santoro, Cindy Sawchuk, and Danny Krizanc. Mobile agent rendezvous
in a ring. In 23rd International Conference on Distributed Computing Systems (ICDCS 2003), 19-22
May 2003, Providence, RI, USA. Proceedings, pages 592 -- 599, 2003.
[42] G´erard Le Lann. Distributed systems - towards a formal approach. In IFIP Congress 77, Toronto,
Canada, August 8-12, 1977. North-Holland, Proceedings, pages 155 -- 160, 1977.
[43] Wei Shi Lim and Steve Alpern. Minimax rendezvous on the line. SIAM Journal on Control and
Optimization, 34(5):1650 -- 1665, 1996.
[44] Nancy Ann Lynch. Distributed Algorithms. Morgan Kaufmann, 1996.
[45] Andrzej Pelc. Deterministic rendezvous in networks: A comprehensive survey. Networks, 59(3):331 --
347, 2012.
[46] Gary L. Peterson. An O(n log n) unidirectional algorithm for the circular extrema problem. ACM
Transactions on Programming Languages and Systems, 4(4):758 -- 762, 1982.
[47] Omer Reingold. Undirected connectivity in log-space. Journal of the ACM, 55(4), 2008.
[48] Amnon Ta-Shma and Uri Zwick. Deterministic rendezvous, treasure hunts, and strongly universal
exploration sequences. ACM Transactions on Algorithms, 10(3):12, 2014.
[49] Masafumi Yamashita and Tiko Kameda. Electing a leader when processor identity numbers are not
In Distributed Algorithms, 3rd International Workshop, Nice, France,
distinct (extended abstract).
September 26-28, 1989, Proceedings, pages 303 -- 314, 1989.
33
[50] Masafumi Yamashita and Tsunehiko Kameda.
i-
characterizing the solvable cases. IEEE Transactions on Parallel and Distributed Systems, 7(1):69 -- 89,
1996.
Computing on anonymous networks: Part
[51] Xiangdong Yu and Moti Yung. Agent rendezvous: A dynamic symmetry-breaking problem. In Au-
tomata, Languages, and Programming - 23rd International Colloquium, ICALP 1996, Paderborn,
Germany, July 8-12, 1996, Proceedings, pages 610 -- 621, 1996.
34
|
1805.09924 | 2 | 1805 | 2018-07-01T07:39:26 | Longest Unbordered Factor in Quasilinear Time | [
"cs.DS"
] | A border u of a word w is a proper factor of w occurring both as a prefix and as a suffix. The maximal unbordered factor of w is the longest factor of w which does not have a border. Here an O(n log n)-time with high probability (or O(n log n log^2 log n)-time deterministic) algorithm to compute the Longest Unbordered Factor Array of w for general alphabets is presented, where n is the length of w. This array specifies the length of the maximal unbordered factor starting at each position of w. This is a major improvement on the running time of the currently best worst-case algorithm working in O(n^{1.5} ) time for integer alphabets [Gawrychowski et al., 2015]. | cs.DS | cs |
Longest Unbordered Factor in Quasilinear Time
Tomasz Kociumaka∗1, Ritu Kundu†2, Manal Mohamed‡2, and
Solon P. Pissis§ 2
1Institute of Informatics, University of Warsaw, Warsaw, Poland
2Department of Informatics, King's College London, London, UK
July 3, 2018
Abstract
A border u of a word w is a proper factor of w occurring both as a
prefix and as a suffix. The maximal unbordered factor of w is the longest
factor of w which does not have a border. Here an O(n log n)-time with
high probability (or O(n log n log2 log n)-time deterministic) algorithm to
compute the Longest Unbordered Factor Array of w for general alphabets
is presented, where n is the length of w. This array specifies the length of
the maximal unbordered factor starting at each position of w. This is a
major improvement on the running time of the currently best worst-case
1.5) time for integer alphabets [Gawrychowski
algorithm working in O(n
et al., 2015].
1
Introduction
There are two central properties characterising repetitions in a word –period and
border – which play direct or indirect roles in several diverse applications ranging
over pattern matching, text compression, assembly of genomic sequences and so
on (see [3, 6]). A period of a non-empty word w of length n is an integer p such
that 1 ≤ p ≤ n, if w[i] = w[i + p], for all 1 ≤ i ≤ n − p. For instance, 3, 6, 7,
and 8 are periods of the word aabaabaa. On the other hand, a border u of w is
a (possibly empty) proper factor of w occurring both as a prefix and as a suffix
of w. For example, ε, a, aa, and aabaa are the borders of w = aabaabaa.
In fact, the notions of border and period are dual: the length of each border
of w is equal to the length of w minus the length of some period of w. For
example, aa is a border of the word aabaabaa; it corresponds to period 6 =
aabaabaa − aa. Consequently, the basic data structure of periodicity on
words is the border array which stores the length of the longest border for each
prefix of w. The computation of the border array of w was the fundamental
concept behind the first linear-time pattern matching algorithm – given a word
∗[email protected]
†[email protected]
‡[email protected]
§[email protected]
1
w (pattern), find all its occurrences in a longer word y (text). The border array
of w is better known as the "failure function" introduced in [15] (see also [1]). It
is well-known that the border array of w can be computed in O(n) time, where
n is the length of w, by a variant of the Knuth-Morris-Pratt algorithm [15].
Another notable aspect of the inter-dependency of these dual notions is the
relationship between the length of the maximal unbordered factor of w and
the periodicity of w. A maximal unbordered factor is the longest factor of w
which does not have a border; its length is usually represented by µ(w), e.g. the
maximal unbordered factor is aabab and µ(w) = 5 for the word w = baabab.
This dependency has been a subject of interest in the literature for a long
time, starting from the 1979 paper of Ehrenfeucht and Silberger [9] in which
they raised the question – at what length of w, µ(w) is maximal (i.e., equal
to the minimal period of the word as it is well-known that it cannot be longer
than that). This line of questioning, after being explored for more than three
decades, culminated in 2012 with the work by Holub and Nowotka [11] where an
asymptotically optimal upper bound (µ(w) ≤ 3
7 n) was presented; the historic
overview of the related research can be found in [11].
Somewhat surprisingly, the symmetric computational problem-given a word
w, compute the longest factor of w that does not have a border-had not been
studied until very recently.
In 2015, Kucherov et al. [14] considered this ar-
guably natural problem and presented the first sub-quadratic-time solution. A
naıve way to solve this problem is to compute the border array starting at each
position of w and locating the rightmost zero, which results in an algorithm
with O(n2) worst-case running time. On the other hand, the computation of
the maximal unbordered factor can be done in linear time for the cases when
µ(w) or its minimal period is small (i.e., at most half the length of w) us-
ing the linear-time computation of unbordered conjugates [8]. However, as has
been illustrated in [14] and [2], most of the words do not fall in this category
owing to the fact that they have large µ(w) and consequently large minimal
period. In [14], an adaptation of the basic algorithm has been provided with
average-case running time O(n2/σ4), where σ is the alphabet's size; it has also
been shown to work better, both in practice and asymptotically, than another
straightforward approach that employs data structures from [13, 12] to query
all relevant factors.
The currently fastest worst-case algorithm to compute the maximal unbor-
dered factor of a given word takes O(n1.5) time; it was presented by Gawrychowski
et al. [10] and it works for integer alphabets (alphabets of polynomial size in n).
This algorithm works by categorising bordered factors into short borders and
long borders depending on a threshold, and exploiting the fact that, for each
position, the short borders are bounded by the threshold and the long borders
are small in number. The resulting algorithm runs in O(n log n) time on aver-
age. More recently, an O(n)-time average-case algorithm was presented using a
refined bound on the expected length of the maximal unbordered factor [2].
Our Contribution. In this paper, we show how to efficiently answer the
Longest Unbordered Factor question using combinatorial insight. Specifically,
we present an algorithm that computes the Longest Unbordered Factor Array in
O(n log n) time with high probability. The algorithm can also be implemented
deterministically in O(n log n log2 log n) time. This array specifies the length of
the maximal unbordered factor at each position in w. We thus improve on the
running time of the currently fastest algorithm, which reports only the maximal
2
unbordered factor of w and works only for integer alphabets, taking O(n1.5)
time.
Structure of the Paper. In Section 2, we present the preliminaries, some
useful properties of unbordered words, the algorithmic toolbox, and a formal
definition of the problem. We lay down the combinatorial foundation of the
algorithm in Section 3 and expound the algorithm in Section 4; its analysis is
explicated in Section 5. We conclude this paper with a final remark in Section
6.
2 Background
Definitions and Notation. We consider a finite alphabet Σ of letters. Let
Σ∗ be the set of all finite words over Σ. The empty word is denoted by ε. The
length of a word w is denoted by w. For a word w = w[1]w[2] . . w[n], w[i . . j]
denotes the factor w[i]w[i + 1] . . w[j], where 1 ≤ i ≤ j ≤ n. The concatenation
of two words u and v is the word composed of the letters of u followed by the
letters of v. It is denoted by uv or also by u · v to show the decomposition of
the resulting word. Suppose w = uv, then u is a prefix and v is a suffix of w;
if u 6= w then u is a proper prefix of w; similarly, if v 6= w then v is a proper
suffix of w. Throughout the paper we consider a non-empty word w of length
n over a general alphabet Σ; in this case, we replace each letter by its rank such
that the resulting word consists of integers in the range {1, . . . , n}. This can be
done in O(n log n) time after sorting the letters of Σ.
An integer 1 ≤ p ≤ n is a period of w if and only if w[i] = w[i + p] for
all 1 ≤ i ≤ n − p. The smallest period of w is called the minimum period (or
the period ) of w, denoted by λ(w). A word u (u 6= w) is a border of w, if
w = uv = v′u for some non-empty words v and v′; note that u is both a proper
prefix and a suffix of w.
It should be clear that if w has a border of length
w − p then it has a period p. Thus, the minimum period of w corresponds to
the length of the longest border (or the border ) of w. Observe that the empty
word ε is a border of any word w.
If u is the shortest border then u is the
shortest non-empty border of w.
The word w is called bordered if it has a non-empty border, otherwise it is
unbordered. Equivalently, the minimum period p = w for an unbordered word
w. Note that every bordered word w has a shortest border u such that w = uvu,
where u is unbordered. By µ(w) we denote the maximum length among all the
unbordered factors of w.
Useful Properties of Unbordered Words. Recall that a word u is a border
of a word w if and only if u is both a proper prefix and a suffix of w. A border
of a border of w is also a border of w. A word w is unbordered if and only if it
has no non-empty border; equivalently ε is the only border of w. The following
properties related to unbordered words form the basis of our algorithm and were
presented and proved in [7].
Proposition 1 ([7]). Let w be a bordered word and u be the shortest non-empty
border of w. The following propositions hold:
1. u is an unbordered word;
2. u is the unique unbordered prefix and suffix of w;
3
3. w has the form w = uvu.
Proposition 2 ([7]). For any word w, there exists a unique sequence (u1, · · · , uk)
of unbordered prefixes of w such that w = uk · · · u1. Furthermore, the following
properties hold:
1. u1 is the shortest border of w;
2. uk is the longest unbordered prefix of w;
3. for all i, 1 ≤ i ≤ k, ui is an unbordered prefix of uk.
The computation of the unique sequence described in Proposition 2 pro-
vides a unique unbordered-decomposition of a word. For instance, for w =
baababbabab the unique unbordered-decomposition of w is baa · ba · b · ba · ba · b.
Longest Successor Factor (Length and Reference) Arrays. Here, we
present the arrays that will act as a toolbox for our algorithm. The longest
successor factor of w (denoted by lsf ) starting at position i, is the longest
factor of w that occurs at i and has at least one other occurrence in the suffix
w[i + 1 . . n]. The longest successor factor array gives for each position i in
w, the length of the longest factor starting both at position i and at another
position j > i. Formally, the longest successor factor array (LSFℓ) is defined as
follows.
LSFℓ[i] =(cid:26) 0
max{k w[i . . i + k − 1] = w[j . . j + k − 1},
if i = n,
for i < j ≤ n.
Additionally, we define the LSF-Reference Array, denoted by LSFr. This
array specifies, for each position i of w, the reference of the longest successor
factor at i. The reference of i is defined as the position j of the last occurrence
of w[i . . i + LSFℓ[i] − 1] in w; we say i refers to j. Formally, LSF-Reference Array
(LSFr) is defined as follows.
LSFr[i] =(cid:26) nil
max{j w[j . . j + LSFℓ[i] − 1] = w[i . . i + LSFℓ[i] − 1]}
Computation: Note that the longest successor factor array is a mirror image of
the well-studied longest previous factor array which can be computed in O(n)
time for integer alphabets [4, 5]. Moreover, in [4], an additional array that
keeps a position of some previous occurrence of the longest previous factor was
presented; such position may not be the leftmost. Arrays LSFℓ and LSFr can
be computed using simple modifications (pertaining to the symmetry between
the longest previous and successor factors) of this algorithm1 within O(n) time
for integer alphabets; see Appendix A.4 for an example.
if LSFℓ[i] = 0,
for i < j ≤ n.
Remark 1. For brevity, we will use lsf and luf to represent the longest successor
factor and the longest unbordered factor, respectively.
1The modified algorithm also computes some starting position j > i for each factor w[i . . i+
LSFℓ[i] − 1], 1 ≤ i ≤ n. Each such factor corresponds to the lowest common ancestor of the
two terminal nodes in the suffix tree of w representing the suffixes w[i . . n] and j[j . . n]; this
ancestor can be located in constant time after linear-time preprocessing [?]. A linear-time
preprocessing of the suffix tree also allows for constant-time computation of the rightmost
starting position of each such factor.
4
Problem Definition. The Longest Unbordered Factor Array problem
can be defined formally as follows.
Longest Unbordered Factor Array
Input: A word w of length n.
Output: An array LUF[1 . . n] such that LUF[i] is the length of the maximal
unbordered factor starting at position i in w, for all 1 ≤ i ≤ n.
Example 1. Consider w = aabbabaabbaababbabab, then the longest unbor-
dered factor array is as follows. (Observe that w is unbordered thus µ(w) =
w = 20.)
i
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
b
a a b b a b a a b
a
a
b
b
a
b
w[i]
LUF[i] 20 3 12 9 12 3 14 3 11 3 10 5
b
a
b
a
2
3
5
2
2
2
2
1
3 Combinatorial Tools
The core of our algorithm exploits the unique unbordered-decomposition of all
suffixes of w in order to compute the length of the maximal (longest) unbordered
prefix of each such suffix. Let the unbordered-decomposition of w[i . . n] be
uk · · · u1 as in Proposition 2. Then LUF[i] = uk.
In order to compute the
unbordered-decomposition for all the suffixes efficiently, the algorithm uses the
repetitive structure of w characterised by the longest successor factor arrays.
Basis of the algorithm. Abstractly, it is easy to observe that for a given
position, if the length of the longest successor factor is zero (no factor starting
at this position repeats afterwards) then the suffix starting at that position is
necessarily unbordered. On the other hand, if the length of the longest successor
factor is smaller than the length of the unbordered factor at the reference (the
position of the the last occurrence of the longest successor factor) then the
ending positions of the longest unbordered factors at this position and that at
its reference will coincide; these two cases are formalised in Lemmas 3 and 4
below. The remaining case is not straightforward and its handling accounts for
the bulk of the algorithm.
Lemma 3. If LSFℓ[i] = 0 then LUF[i] = n − i + 1, for 1 ≤ i ≤ n.
Lemma 4. If LSFr[i] = j and LSFℓ[i] < LUF[j] then LUF[i] = j + LUF[j] − i,
for 1 ≤ i ≤ n.
Proof. Let k = j + LUF[j] − 1.
We first show that w[i . . k] is unbordered. Assume that w[i . . k] is bordered
and let β be the length of one of its borders (β < LSFℓ[i] as LSFr[i] = j). This
implies that w[i . . i + β − 1] = w[k − β + 1 . . k]. Since w[i . . i + LSFℓ[i] − 1] =
w[j . . j + LSFℓ[i] − 1], we get w[j . . j + β − 1] = w[k − β + 1 . . k] (i.e., w[j . . k]
is bordered) which is a contradiction. Moreover, w[k + 1 . . n] can be factorised
into prefixes of w[j . . k] (by definition of LUF); every such prefix is also a proper
prefix of w[i . . i + LSFℓ[i] − 1] which will make every factor w[i . . k′], k < k′ ≤ n,
to be bordered. This completes the proof.
5
We introduce the notion of a hook to handle finding the unbordered-decomposition
of suffixes w[i . . n] for the remaining case (i.e., when LSFℓ[i] ≥ LUF[LSFr[i]]).
Definition 3.1 (Hook). Consider a position j in a length-n word w. Its hook
Hj is the smallest position q such that w[q . . j − 1] can be decomposed into
unbordered prefixes of w[j . . n].
The following observation provides a greedy construction of this decomposi-
tion.
Observation 5. The decomposition of a word v into unbordered prefixes of
another word u is unique. This decomposition can be constructed by iteratively
trimming the shortest prefix of u which occurs as a suffix of the decomposed
word.
Moreover, the decomposability into unbordered prefixes of u is hereditary in
a certain sense:
Observation 6. If a word v can be decomposed into unbordered prefixes of u,
then every prefix of v also admits such a decomposition.
Example 2. Consider w = aabbabaabbaababbabab as in Example 1. Observe
that H18 = 13: the factor w[13 . . 17] = ba · b · ba can be decomposed into unbor-
dered prefixes of w[18 . . 20] = bab. Moreover, no prefix of w[18 . . 20] matches a
suffix of w[1 . . 12] = · · · aa.
The hook Hj has its utility when j is a reference as shown in the following
lemma.
Lemma 7. Consider a position i such that LSFℓ[i] ≥ LUF[j], where j = LSFr[i].
Then
LUF[i] =(Hj − i
LUF[j]
if i < Hj,
otherwise.
Proof. Let u = w[j . . j + LUF[j] − 1], v = w[i . . i + LSFℓ[i] − 1], and q = Hj.
Observe that u occurs at position i and that w[q . . n] can be decomposed into
unbordered prefixes of u.
Case a: i < q. We shall prove that w[i . . q − 1] is the longest unbordered prefix
of w[i . . n]; see Fig. 1. By Observation 6, any longer factor w[i . . k], q ≤ k ≤ n
has a suffix w[q . . k] composed of unbordered prefixes of u. Thus, w[i . . k] must
be bordered, because u is its prefix. To conclude, for a proof by contradiction
suppose that w[i . . q − 1] has a border v′. Note that v′ ≤ LSFℓ[i], so v′ is a
prefix of v. Hence, it occurs both as a suffix of w[1 . . q − 1] and a prefix of
w[j . . n], which contradicts the greedy construction of q = Hj (Observation 5).
Case b: i ≥ q. The decomposition of w[q . . n] into unbordered prefixes of u
yields a decomposition of w[i . . n] into unbordered prefixes of u, starting with
u. This is the unbordered-decomposition of w[i . . n] (see Proposition 2), which
yields LUF[i] = u = LUF[j].
6
i
i
u
v
LSFℓ[i] = v
LSFr[i] = j
q
q
ur
ur−1
· · ·
u2
u1
Hj = q
j
j
u
v
LUF[j] = u
Figure 1: Case a (i < q): The unbordered-decomposition of w[i . . n] consists of
w[i . . q − 1] as the longest unbordered prefix, followed by a sequence of unbor-
dered prefixes of u, including u itself at position j. Therefore, LUF[i] = q − i.
4 Algorithm
The algorithm operates in two phases: a preprocessing phase followed by the
main computation phase. The preprocessing phase accomplishes the following:
Firstly, compute the longest successor factor array LSFℓ together with LSFr
array. If LSFr[i] = j then we say i refers to j and mark j in a boolean array
(IsReference) as a reference.
In the main phase, the algorithm computes the lengths of the longest un-
bordered factors for all positions in w. Moreover, it determines HOOK[j] = Hj
for each potential reference, i.e., each position j such that j = LSFr[i] and
LSFℓ[i] ≥ LUF[j] for some i < j; see Lemma 7.
Positions are processed from right to left (in decreasing order) so that if i
refers to j, then LUF[j] (and HOOK[j], if necessary) has already been computed
before i is considered. For each position i, the value of LUF[i] is determined as
follows:
1. If LSFℓ[i] = 0, then LUF[i] = n − i + 1.
2. Otherwise
(a) If LSFℓ[i] < LUF[j], then LUF[i] = j + LUF[j] − i.
(b) If LSFℓ[i] ≥ LUF[j] and i ≥ HOOK[j], then LUF[i] = LUF[j].
(c) If LSFℓ[i] ≥ LUF[j] and i < HOOK[j], then LUF[i] = HOOK[j] − i.
If i is a potential reference, then HOOK[i] is also computed, as described
in the Section 4.1 (see Algorithm 2 in Appendix A.1 for a pseudo-code of the
main algorithm). It is evident that the computational phase of the algorithm
fundamentally reduces to finding the hooks for potential references; for brevity,
the term reference will mean a potential reference hereafter.
4.1 Finding Hook (FindHook Function)
Main idea When FindHook is called on a reference j, it must return
Hj. A simple greedy approach follows directly from Observation 5; see also
Figure 2. Initially, the factor w[1 . . j − 1] is considered and the shortest suffix
of w[1 . . j − 1] which is a prefix of w[j . . n] is computed. Then this suffix,
denoted u1 = w[i1 . . j − 1], is truncated (chopped) from the considered factor
w[1 . . j − 1]; the next factor considered will be w[1 . . i1 − 1].
In general, we
iteratively compute and truncate the shortest prefixes of w[j . . n] from the right
7
q
ip
· · ·
ur
ip−1
· · ·
up
ik
i2
i1
j
· · ·
uk
u2
u1
u
Figure 2: A chain of consecutive shortest prefixes of w[j . . n] starting at positions
i1 > i2 > · · · > ir = q. No prefix of w[j . . n] is a suffix of w[1 . . q − 1], so the
hook value of position j is Hj = q. Meanwhile, HOOK[ik] is set to ip−1 in order
to avoid iterating through ik+1, . . . , ip−1 again.
end of the considered factor; shortening the length of the considered factor in
each iteration and terminating as soon as no prefix of w[j . . n] can be found. If
the considered factor at termination is w[1 . . q − 1], position q is returned by
the function as Hj.
The factors w[q . . j −1] considered by successive calls of FindHook function
may overlap. Moreover, the same chains of consecutive unbordered prefixes
may be computed several times throughout the algorithm. To expedite the
chain computation in the subsequent calls of FindHook on another reference
j′ (j′ < j), we can recycle some of the computations done for j by shifting the
value HOOK[·] of each such index (at which a prefix was cut for j) leftwards
(towards its final value). Consider the starting position ik at which uk was
cut (i.e., uk = w[ik . . ik−1 − 1] is the shortest unbordered prefix of w[j . . n]
computed at ik−1). Let ip be the first position considered after ik such that
up > uk.
In this case, every factor uk+1, . . . , up−1 is a prefix of uk; see
Figure 2. Therefore, w[ip−1 . . ik − 1] can be decomposed into prefixes of uk (and
of w[ik . . n]). Consequently, we set HOOK[ik] = ip−1 so that the next time a
prefix of length greater than or equal to uk is cut at ik, we do not have to
repeat truncating the prefixes uk+1, . . . , up−1 and we may start directly from
position ip−1.
In order to express the intermediate values in the HOOK table, we generalize
the notion of Hj: for a position j and a length ℓ, we define Hℓ
j as the largest
position q such that w[q . . j − 1] can be decomposed into unbordered prefixes of
w[j . . n] whose lengths do not exceed ℓ. Observe that H0
j = Hj if
ℓ ≥ LUF[j].
j = j and Hℓ
Implementation For each position ik, we set HOOK[ik] = Huk
, equal
to ip−1 in the case considered above. Computing these values for all indices ik
can be efficiently realised using a stack. Every starting position ip, at which
up is cut, is pushed onto the stack as a (length, position) pair (up, ip). Before
pushing, every element (uk, ik) such that uk < up is popped and the hook
value of index ik is updated (HOOK[ik] = Huk
= ip−1 = ip + up). A pseudo-
code implementation of this function is given below as Algorithm 1.
ik
ik
Analysis Throughout the algorithm, each unbordered prefix up at position
ip is computed just once by the FindHook function. Nevertheless, a longer2
2It will be easy to deduce after Lemma 9 that the length of the prefix cut (the next time)
at the same position will be at least twice the length of the current prefix cut at it.
8
Algorithm 1 Return Hj and set HOOK[i] ← Hβ
the stack
i
for each (β, i) pushed onto
1: function FindHook(j)
2:
3:
st ← NewStack()
q ← HOOK[j]
β ← FindBeta(q, j) ⊲ the length of the shortest prefix of w[j . . n] ending at q − 1
while (β 6= 0) do
4:
5:
6:
7:
8:
9:
10:
11:
HandlePopping(st, j, q, β)
Push(st, (β, q − β))
q ← HOOK[q − β]
β ← FindBeta(q, j)
HandlePopping(st, j, q, ∞)
return q
12: function HandlePopping(st, j, q, β)
13:
while IsNotEmpty(st) and Length(Top(st)) < β do
14:
15:
(length,pos) ← Pop(st)
HOOK[pos] ← q
⊲ returns Hj
⊲ q = Hlength
pos
unbordered prefix u′
on reference j′ (where q < j′ < j).
p may be computed at ip again when FindHook is called
In what follows, we introduce certain characteristics of the computed unbor-
dered prefixes which aids in establishing the relationship between the stacks of
various references. Let Sj be the set of positions pushed onto the stack during
a call of FindHook on reference j.
Definition 4.1 (Twin Set). A twin set of reference j for length ℓ, denoted
by T ℓ
is the set of all the positions i ∈ Sj which were pushed onto the
j ,
stack paired with length ℓ in the call of FindHook on reference j (i.e., T ℓ
j =
{i (ℓ, i) was pushed onto the stack of j}).
Note that a unique shortest unbordered prefix of w[j . . LUF[j] − 1] occurs at
each i belonging to the same twin set. However, as and when a longer prefix at
i is cut (say ℓ′) for another reference j′ < j, i will be added to T ℓ′
j ′ .
LUF[j]
Remark 2. Sj =
T ℓ
j .
Sℓ=1
Hereafter, a twin set will essentially imply a non-empty twin set.
Lemma 8. If j′ and j are references such that j′ ∈ Sj, then Hj ≤ Hj ′ .
Proof. Since j′ ∈ Sj , the suffix w[j′ . . n] (and, by Observation 6, its every pre-
fix w[j′ . . k]) can be decomposed into unbordered prefixes of w[j . . n]. Conse-
quently, any decomposition into unbordered prefixes of w[j′ . . n] yields a decom-
position into unbordered prefixes of w[j . . n]. In particular, w[Hj ′ . . n] admits
such a decomposition, which implies Hj ≤ Hj ′ .
If the stack Sj is the most recent stack containing a reference j′, we say that
j′ is the parent of j. More formally, the parent of j′ is defined as min{j : j′ ∈ Sj }.
If j′ does not belong to any stack (and thus has no parent), we will call it a
base reference.
9
Hj
i
v
· · ·
j′
v
· · ·
k
p
v′
z
z
j
· · ·
Figure 3: The pair (z, i) is the first to be pushed onto the stack of j′. The
factor z is unbordered, has v as a proper prefix and some v′ as a proper suffix,
where both v and v′ are unbordered prefixes of w[j . . n] whose lengths ℓ and ℓ′,
respectively, satisfy ℓ < ℓ′.
Lemma 9. If j and j′ are two references such that j is the parent of j′ and
j′ ∈ T ℓ
j , then each position i ∈ Sj ′ satisfies the following properties:
1. i ∈ T ℓ
j ;
2. there exists k ∈ T ℓ′
j , with ℓ′ > ℓ, such that (k + ℓ′ − i, i) is pushed onto the
stack of j′.
j ′ .
Proof. Let p be the value of HOOK[j′] prior to the execution of FindHook(j′).
Since j′ ∈ T ℓ
j ′ . As j is
the parent of j′, no further call has updated HOOK[j′]. Thus, we conclude that
p = Hℓ
j , the earlier call FindHook(j) has set HOOK[j′] = Hℓ
Consequently, the first pair pushed onto the stack of j′ (cf. Algorithm 1) is
(z, i), where z = w[i . . p − 1] is the shortest suffix of w[1 . . p − 1] which also
occurs as a prefix of w[j′ . . n] (see Figure 3). Moreover, observe that z > ℓ by
the greedy construction of Hℓ
j ′ .
Recall that j′ ∈ T ℓ
j implies that w[j′ . . n] can be decomposed into unbordered
prefixes of w[j . . n], with the first prefix of length ℓ, denoted v = w[j′ . . j′ +ℓ−1].
With an occurrence at position j′, the factor z also admits such a decomposi-
tion, still with the first prefix v (due to z > v). Additionally, note that
w[p . . j′ − 1] can be decomposed into unbordered prefixes of v. Concatenating
the decompositions of z = w[i . . p − 1], w[p . . j′ − 1], and w[j′ . . n], we conclude
that w[i . . n] can be decomposed into unbordered prefixes of w[j . . n] with the
first prefix (in this unique decomposition) equal to v. Hence, i ∈ Sj ′ belongs to
the same twin set as j′; i.e., it satisfies the first claim of the lemma.
Additionally, in the aforementioned decomposition of w[i . . n] consider the
factor v′ = w[k . . p − 1] which ends at position p − 1. By the greedy construction
for ℓ′ = v′ > ℓ.
of Hℓ
Moreover, recall that (z, i) = (k + ℓ′ − i, i) is pushed onto the stack of j′.
Consequently, i also satisfies the second claim of the lemma.
j ′ , its length v′ is strictly larger than ℓ, so k ∈ T ℓ′
j
A similar reasoning is valid for each i that will appear in Sj ′ .
Lemma 10. If j is the parent of two references j′′ < j′, both of which belong
to T ℓ
j , then Sj ′ ∩ Sj ′′ = ∅.
Proof. Let u = w[j . . j + LUF[j] − 1] and v be the shortest unbordered prefix
of u cut at j′ and j′′ (i.e., v = ℓ). Let u′ = w[j′ . . j′ + LUF[j′] − 1] and
10
u′′ = w[j′′ . . j′′ + LUF[j′′] − 1]. Here, the current call to the FindHook function
has been made on the reference j′′. Consider a position i such that i ∈ Sj ′ and i
would also appear in Sj ′′ ; let the corresponding prefixes of u′ and u′′ cut at i be
z′ and z′′ (examine Figure 4). Observe that i was in T ℓ
j (Lemma 9), therefore,
each of z′ and z′′ has v as a proper prefix. Let v′ and v′′ be the corresponding
proper suffixes of z′ and z′′ where v′ and v′′ are unbordered prefixes of u; both
of length greater than v.
q
i
v
· · ·
p
· · ·
v′
z′′
z′
v
v′′
x
k
· · ·
j′′
v
· · ·
j
u
u′
j′
v
u′′
Figure 4: The pair (z′, i) and (z′′, i) are pushed onto the stack of j′ and j′′,
respectively, where i is a position common to both Sj ′ and Sj ′′ .
Since i ∈ Sj ′ , w[i . . j′ − 1] can be decomposed from right to left into unbor-
dered prefixes of u′ such that each prefix (say v) of u having length greater than
v that had been computed when j was considered, is covered ; i.e., v appears as
either a proper suffix or a factor of some shortest prefix in such decomposition.
In other words, the shortest prefix of u′ that ends with v starts from the nearest
v preceding v whose corresponding position was pushed in Sj ′ . Note that the
same condition is also valid for any prefix of v that is longer than v.
1. z′ < z′′: The factor w[p . . k − 1] can be decomposed into unbordered
prefixes of u′, where p and k are as in Figure 4. Let x be the rightmost
prefix of such decomposition (x has v as proper prefix, v′′ as a proper
suffix, and the corresponding position of this v; i.e., i + z′′ − x is in Sj ′ ).
Moreover, z′ < x otherwise z′′ cannot be unbordered. Observe the two
equal-length factors w[i . . i + z′′] and w[j′′ . . j′′ + z′′]. It follows from
i+z′′−x ∈ Sj ′ that j′′+z′′−x ∈ Sj ′ . Consequently, w[i . . i+z′′−x]
and w[j′′ . . j′′ + z′′ − x] have the same right to left decomposition in
prefixes of u′ implying that if if i ∈ Sj ′ then j′′ should have been in Sj ′
which is a contradiction.
2. z′ ≥ z′′: This implies that z′′ is an unbordered prefix of u′. Therefore,
if i was pushed onto the stack of j′ then j′′ should also had been pushed
onto its stack which is a contradiction.
4.2 Finding Shortest Border (FindBeta Function)
Given a reference j and a position q, function FindBeta returns the length β
of the shortest prefix of w[j . . n] that is a suffix of w[1 . . q − 1], or β = 0 if there
is no such prefix; note that the sought shortest prefix is necessarily unbordered.
To find this length, we use 'prefix-suffix queries' of [13, 12]. Such a query,
given a positive integer d and two factors x and y of w, reports all prefixes of
11
x of length between d and 2d that occur as suffixes of y. The lengths of sought
prefixes are represented as an arithmetic progression, which makes it trivial to
extract the smallest one. A single prefix-suffix query can be implemented in
O(1) time after randomized preprocessing of w which takes O(n) time in expec-
tation [13], or O(n log n) time with high probability [12]. Additionally, replacing
hash tables with deterministic dictionaries [16], yields an O(n log n log2 log n)-
time deterministic preprocessing.
To implement FindBeta, we set x = [j . . n], y = [1 . . q − 1] and we ask
prefix-suffix queries for subsequent values d = 1, 3, . . . , 2k − 1, . . . until d exceeds
min(x, y). Note that we can terminate the search as soon as a query reports
a non-empty answer. Hence, the running time is O(1 + log β) if the query is
successful (i.e., β 6= 0) and O(log n) otherwise.
Furthermore, we can expedite the successful calls to FindBeta if we already
know that β /∈ {1, . . . , ℓ}. In this case, we can start the search with d = ℓ + 1.
j ′ for some j′, we can
Specifically, if j is not a base reference and belongs to T ℓ
start from d = 2ℓ + 1 because Lemma 9.2 guarantees that β > ℓ + ℓ′ ≥ 2ℓ.
5 Analysis
Algorithm 2 computes the longest unbordered factor at each position i; position
i is a start-reference or it refers to some other position. The correctness of the
computed LUF[i] follows directly from Lemmas 3 through 7.
The analysis of the algorithm running time necessitates probing of the total
time consumed by FindHook and the time spent by FindBeta function which,
in turn, can be measured in terms of the total size of the stacks of various
references.
Lemma 11. The total size of all the stacks used throughout the algorithm is
O(n log n). Moreover, the total running time of the FindBeta function is
O(n log n).
Proof. First, we shall prove that any position p belongs to O(log n) stacks.
By Lemma 9.1, the stack of any reference is a subset of the stack of its
parent. Moreover, by Lemmas 9.1 and 10, the stacks of references sharing the
same parent are disjoint. A similar argument (see Lemma 13 in Appendix A.3)
shows that the stacks of base references are disjoint.
Consequently, the references j1 > . . . > js whose stacks Sji contain p form
a chain with respect to the parent relation: j1 is a base reference, and the
parent of any subsequent ji is ji−1. Let us define ℓ1, . . . , ℓs so that p ∈ T ℓi
ji . By
i > ℓi such that ki ∈ T ℓ′
Lemma 9.2, for each 1 ≤ i < s, there exist ki and ℓ′
i
ji
and ℓi+1 = ki − p + ℓ′
i ≥ ℓi + ℓ′
i > 2ℓi. Due to 1 ≤ ℓi ≤ n, this yields
s ≤ 1 + log n = O(log n), as claimed.
Next, let us analyse the successful calls β = FindBeta(q, j) with p = q − β.
Observe that after each such call, p is inserted to the stack Sj and to the twin set
j , i.e, j = ji and β = ℓi for some 1 ≤ i ≤ s. Moreover, if i > 0, then ji ∈ T ℓi−1
T β
ji−1 ,
which we are aware of while calling FindBeta. Hence, we can make use of the
fact that ℓi /∈ {1, . . . , 2ℓi−1} to find β = ℓi in time O(log ℓi
). For i = 1,
ℓi−1
the running time is O(1 + log ℓ1). Hence, the overall running time of successful
12
i=2 log ℓi
ℓi−1
O(1 + log ℓs) = O(log n), which sums up to O(n log n) across all positions p.
queries β = FindBeta(q, j) with p = q − β is O(1 + log ℓ1 +Ps
) =
As far as the unsuccessful calls 0 = FindBeta(q, j) are concerned, we ob-
serve that each such call terminates the enclosing execution of FindHook.
Hence, the number of such calls is bounded by n and their overall running time
is clearly O(n log n).
Theorem 12. Given a word w of length n, Algorithm 2 solves the Longest
Unbordered Factor Array problem in O(n log n) time with high probability.
It can also be implemented deterministically in O(n log n log2 log n) time.
Proof. Assuming an integer alphabet, the computation of LSFℓ and LSFr arrays
along with the constant time per position initialisation of the other arrays sum
up the preprocessing stage (Lines 2–7) to O(n) time. The running time required
for the assignment of the luf for all positions (Lines 9–18) is O(n). The time
spent in construction of the data structure to answer prefix-suffix queries used in
FindBeta function is O(n log n) with high probability or O(n log n log2 log n)
deterministic.
Additionally, the total running time of the FindHook function for all the
references, being proportional to the aggregate size of all the stacks, can be
deduced from Lemma 11. This has been shown to be O(n log n) in the worst
case, same as the total running time of FindBeta. The claimed bound on the
overall running time follows.
To show that the upper bound shown in Lemma 11 in the worst case is tight,
we design an infinite family of words that exhibit the worst-case behaviour (see
Appendix A.2 for more details).
6 Final Remark
Computing the longest unbordered factor in o(n log n) time for integer alphabets
remains an open question.
References
[1] A. V. Aho, J. E. Hopcroft, and J. D. Ullman. The design and analysis
of computer algorithms, 1974. Reading: Addison-Wesley, pages 207–209,
1987.
[2] P. Cording and M. Knudsen. Maximal unbordered factors of random
strings. In S. Inenaga, K. Sadakane, and T. Sakai, editors, Proceedings of
the International Symposium String Processing and Information Retrieval
SPIRE, pages 93–96, 2016.
[3] M. Crochemore, C. Hancart, and T. Lecroq. Algorithms on Strings. Cam-
bridge University Press, 2007.
[4] M. Crochemore and L. Ilie. Computing longest previous factor in linear
time and applications. Inf. Process. Lett., 106(2):75–80, 2008.
13
[5] M. Crochemore, L. Ilie, C. S. Iliopoulos, M. Kubica, W. Rytter, and
T. Wale´n. Computing the longest previous factor. Eur. J. Comb., 34(1):15–
26, 2013.
[6] M. Crochemore and W. Rytter. Jewels of Stringology. World Scientific,
2002.
[7] J. Duval. Relationship between the period of a finite word and the length
of its unbordered segments. Discrete Mathematics, 40(1):31 – 44, 1982.
[8] J.-P. Duval, T. Lecroq, and A. Lefebvre. Linear computation of unbordered
conjugate on unordered alphabet. Theoretical Computer Science, 522:77 –
84, 2014.
[9] A. Ehrenfeucht and D. Silberger. Periodicity and unbordered segments of
words. Discrete Math, 26(2):101–109, 1979.
[10] P. Gawrychowski, G. Kucherov, B. Sach, and T. Starikovskaya. Computing
the longest unbordered substring. In C. Iliopoulos, S. Puglisi, and E. Yil-
maz, editors, Proceedings of the International Symposium String Processing
and Information Retrieval SPIRE, pages 246–257, 2015.
[11] S. Holub and D. Nowotka. The Ehrenfeucht–Silberger problem. Journal of
Combinatorial Theory, Series A, 119(3):668–682, 2012.
[12] T. Kociumaka, J. Radoszewski, W. Rytter, and T. Wale´n. Efficient data
In String Processing and
structures for the factor periodicity problem.
Information Retrieval SPIRE, pages 284–294. Springer, 2012.
[13] T. Kociumaka, J. Radoszewski, W. Rytter, and T. Walen. Internal pat-
tern matching queries in a text and applications.
In Proceedings of the
Annual ACM-SIAM Symposium on Discrete Algorithms SODA, pages 532–
551, 2015.
[14] G. Kucherov, A. Loptev, and T. Starikovskaya. On maximal unbordered
factors. In Proceedings of the Annual Symposium on Combinatorial Pat-
tern Matching CPM, Lecture Notes in Computer Science, pages 343–354.
Springer, 2015.
[15] J. Morris Jr and V. Pratt. A linear pattern-matching algorithm. Technical
Report 40, University of California, Berkeley, 1970.
[16] M. Rui. Constructing efficient dictionaries in close to sorting time.
In
L. Aceto, I. Damgard, L. A. Goldberg, M. M. Halld´orsson, A. Ing´olfsd´ottir,
and I. Walukiewicz, editors, Automata, Languages and Programming,
ICALP 2008, Part I, volume 5125 of LNCS, pages 84–95. Springer, 2008.
14
A Appendix
A.1 Pseudo-code
Algorithm 2 Compute the Longest Unbordered Factor Array
1: procedure LongestUnborderedFactor(word w, integer n = w)
⊲ Preprocessing:
LSFℓ, LSFr ← LongestSuccessorFactor(w)
for i ← 1 to n do
if LSFℓ[i] 6= 0 then
IsReference[LSFr[i]] ← true
HOOK[1 . . n] ← 1, · · · , n
⊲ Main Algorithm:
for i ← n to 1 do
if LSFℓ[i] = 0 then
LUF[i] ← n − i + 1
else
j ← LSFr[i]
if LSFℓ[i] < LUF[j] then
LUF[i] ← j + LUF[j] − i
else if i ≥ HOOK[j] then
LUF[i] ← LUF[j]
else
LUF[i] ← HOOK[j] − i
if IsPotentialReference(i) then 3
HOOK[i] ← FindHook(i)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
⊲ starting reference
A C++ implementation of our algorithm can also be made available upon request.
A.2 Words Exhibiting Worst-Case Behaviour
A word can be made to exhibit the worst-case behaviour if we force the maximum
number of positions to be pushed onto Θ(log n) stacks. This can be achieved as
follows.
1. Maximise the number of references: Every position in each twin set T l
j is
a reference.
2. Maximise the size of each stack: The largest position (reference) in any
If j′ is largest
twin set pushes the rest of the positions onto its stack.
reference in T ℓ
j then Sj ′ = T ℓ
j − {j′}.
3. Maximise the number of twin sets obtained for a stack: This increases
the number of unbordered prefixes that can be cut at some position i,
therefore, increasing the number of re-pushes of i. This can be achieved
by keeping T ℓ
+ 1.
j = 2T ℓ+1
j
Using the above, Algorithm 3 creates a word w over Σ = {a,b}, such that the
total size of the stack of the base reference j (w[j] = a) and the references that
3 IsPotentialReference(i) returns true if there exists i′ such that LSFr[i′] = i and
LSFℓ[i′] ≥ LUF[i].
15
appear in Sj is tSj − t, where t is the number of non-empty twin sets obtained
for the stack of j.
Assume a binary alphabet Σ = {a, b}, the following words exhibit the
maximum total size of the stacks used: w3 = (aabaabb)2, tmax = 3; w4 =
(aabaabbaabaabbb)2, tmax = 4; w5 = (aabaabbaabaabbbaabaabbaabaabbbb)2, tmax =
5; etc., where tmax is the maximum number of stacks onto which some propor-
tional number of elements has been pushed by Algorithm 2. Position 1 in w4, for
example, is pushed onto four stacks paired with length 1, 3, 7 then 15. The total
size of the stacks used by each word from this family of words is thus Θ(n log n).
Figure 5 (in Appendix A.2) shows the relation between the lengths of the words
and the total size of the stacks used by Algorithm 2 for the specified family of
the words.
Algorithm 3 Create Word w Over Σ = {a,b}
1: w ← ""
2: block ← "a"
3: i ← 1
4: while i ≤ t − 1 do
5:
6:
7:
8: w ← w + block
9: w ← w + w
w ← w + block + w
block ← block + "b"
i ← i + 1
15
10
x
a
m
t
5
1
0.5
tmax
total size of stacks
0
0
0.5
1
1.5
2
2.5
n/105
6
0
1
/
s
k
c
a
t
s
f
o
e
z
i
s
l
a
t
o
t
Figure 5: Plot showing the maximum number of stacks onto which some element
has been pushed (tmax) and the total size of stacks used by Algorithm 2 for
specially designed words.
A.3 Stacks of the Base References are Disjoint
Lemma 13. If j1 6= j2 are base references, then Sj1 ∩ Sj2 = ∅.
Proof. Suppose that the FindHook function is called for each position in w.
We define a base position analogously as a position that does not appear in any
stack. For a proof by contradiction, let i be the largest element of Sj1 ∩ Sj2 ,
with (ℓ1, i) and (ℓ2, i) pushed onto the stacks of j1 and j2, respectively. Note
16
that i + ℓ1 ∈ {j1} ∪ Sj1 and i + ℓ2 ∈ {j2} ∪ Sj2 . Thus, our choice of j1 6= j2 as
base positions and i as the largest element of Sj1 ∩ Sj2 guarantees ℓ1 6= ℓ2. We
assume that ℓ1 < ℓ2 without loss of generality.
Let u be the longest unbordered factor at j1. Note that due to i ∈ Sj1 , the
suffix w[i . . n] can be decomposed into unbordered prefixes of u. In particular,
w[i . . i + ℓ2 − 1] admits such a decomposition w[i . . i + ℓ2 − 1] = v1 · · · vr with
v1 = ℓ1. Moreover, observe that vr > ℓ1; otherwise, vr would be a border of
w[i . . i + ℓ2 − 1].
Let vs be the first of these factors satisfying vs > ℓ1 and let k = j1 +
v1 · · · vs−1. Note that w[j1 . . k − 1] admits a decomposition w[j1 . . k − 1] =
v1 · · · vs−1 into unbordered prefixes of vs. Consequently, j1 ∈ Sk if k is a base
position, and j1 ∈ Sk′ if k is not a base position and k ∈ Sk′ for some base
position k′.
In either case, this contradicts the assumption that j1 is a base
position.
In fact, Algorithm 2 calls the FindHook function on a subset of positions;
i.e., potential references. However, as we show below, all base references are
actually base positions. For a proof by contradiction, suppose that j′ is a base
reference, but it would have been pushed onto the stack of a base position j > j′.
Below, we show that the longest unbordered factor at j, denoted u, does
not have any other occurrence in w. First, suppose that it occurs at a position
k > j. Observe that w[j . . k − 1] can be decomposed into unbordered prefixes
of u. Consequently, j ∈ Sk if k is a base position, and j ∈ Sk′ if k is not a base
position and k ∈ Sk′ for some base position k′. In either case, this contradicts
the assumption that j is a base position. Next, suppose that u occurs at a
position k < j and let us choose the largest such k. Observe that LSFℓ[k] ≥ u
and LSFr[k] = j since j is the only position of u larger than k. However, this
means that j is a potential reference, contrary to our assumption.
In particular, we conclude that u′ = w[j′ . . j + u − 1] does not have any
border of length u or more. On the other hand, shorter borders are excluded
since u is unbordered and u′ can be decomposed into unbordered prefixes of
u. Consequently, u′ is unbordered. However, j′ is a potential reference, so
u′ occurs to the left of j′. This yields an occurrence of u to the left of j, a
contradiction.
A.4 Example of Longest Successor Factor Arrays
Example 3. Let w = aabbabaabbaababbabab. The associated arrays are as
follows.
i
w[i]
9 10 11 12 13 14 15 16 17 18 19 20
1 2
b
b
a a
LSFℓ[i] 5 6
3
0
LSFr[i] 7 14 15 16 17 10 11 14 15 18 19 17 18 19 20 18 19 20 nil nil
7
a
3
8
a
4
a
4
b
3
a
2
b
1
a
2
b
1
b
2
a
1
3
b
5
4
b
4
5
a
3
6
b
4
b
3
a
0
17
|
1603.07768 | 1 | 1603 | 2016-03-24T22:16:13 | Online Budgeted Allocation with General Budgets | [
"cs.DS"
] | We study the online budgeted allocation (also called ADWORDS) problem, where a set of impressions arriving online are allocated to a set of budget-constrained advertisers to maximize revenue. Motivated by connections to Internet advertising, several variants of this problem have been studied since the seminal work of Mehta, Saberi, Vazirani, and Vazirani (FOCS 2005). However, this entire body of work focuses on a single budget for every advertising campaign, whereas in order to fully represent the actual agenda of an advertiser, an advertising budget should be expressible over multiple tiers of user-attribute granularity. A simple example is an advertising campaign that is constrained by an overall budget but is also accompanied by a set of sub-budgets for each target demographic. In such a contract scheme, an advertiser can specify their true user-targeting goals, allowing the publisher to fulfill them through relevant allocations.
In this paper, we give a complete characterization of the ADWORDS problem for general advertising budgets. In the most general setting, we show that, unlike in the single-budget ADWORDS problem, obtaining a constant competitive ratio is impossible and give asymptotically tight upper and lower bounds. However for our main result, we observe that in many real-world scenarios (as in the above example), multi-tier budgets have a laminar structure, since most relevant consumer or product classifications are hierarchical. For laminar budgets, we obtain a competitive ratio of e/(e-1) in the small bids case, which matches the best known ADWORDS result for single budgets. Our algorithm has a primal-dual structure and generalizes the primal-dual analysis for single- budget ADWORDS first given by Buchbinder, Jain, and Naor (ESA 2007). | cs.DS | cs |
Online Budgeted Allocation with General Budgets
Nathaniel Kell
Debmalya Panigrahi
Department of Computer Science, Duke University, Durham, NC, USA.
Email: {kell,debmalya}@cs.duke.edu
Abstract
We study the online budgeted allocation (also called ADWORDS) problem, where a set of impressions arriving
online are allocated to a set of budget-constrained advertisers to maximize revenue. Motivated by connections to
Internet advertising, several variants of this problem have been studied since the seminal work of Mehta, Saberi,
Vazirani, and Vazirani (FOCS 2005). However, this entire body of work focuses on a single budget for every
advertising campaign, whereas in order to fully represent the actual agenda of an advertiser, an advertising budget
should be expressible over multiple tiers of user-attribute granularity. A simple example is an advertising campaign
that is constrained by an overall budget but is also accompanied by a set of sub-budgets for each target demographic.
In such a contract scheme, an advertiser can specify their true user-targeting goals, allowing the publisher to fulfill
them through relevant allocations.
In this paper, we give a complete characterization of the ADWORDS problem for general advertising budgets.
In the most general setting, we show that, unlike in the single-budget ADWORDS problem, obtaining a constant
competitive ratio is impossible and give asymptotically tight upper and lower bounds. However for our main result,
we observe that in many real-world scenarios (as in the above example), multi-tier budgets have a laminar structure,
since most relevant consumer or product classifications are hierarchical. For laminar budgets, we obtain a competitive
ratio of e/(e − 1) in the small bids case, which matches the best known ADWORDS result for single budgets. Our
algorithm has a primal-dual structure and generalizes the primal-dual analysis for single-budget ADWORDS first
given by Buchbinder, Jain, and Naor (ESA 2007). However many new ideas are required to overcome the barriers
introduced by laminar budgets-our algorithm uses a novel formulation that overcomes non-monotonicity in the
syntactically defined dual variables, as well as a dynamically maintained labeling scheme that properly tracks the
"most-limiting" budgets in the hierarchy.
Internet advertising, adwords, user targeting, primal-dual algorithms
Index Terms
I. INTRODUCTION
The online budgeted allocation problem, also called the ADWORDS problem, has had a significant impact on
the theory and practice of online (Internet) advertising. In this problem, an advertisement publisher is tasked with
matching user-generated advertisement slots on a web page (typically called impressions) to advertisers (typically
called bidders). More formally, the publisher is given a set of offline bidders u ∈ U at the outset of the problem,
each of which is specified by a budget Bu indicating the maximum revenue the publisher can receive from bidder
u. A set of impressions v ∈ V then arrive in an online sequence and must each be irrevocably assigned to a unique
bidder. Upon assigning impression v to bidder u, the publisher receives revenue ruv (typically called a bid value)
from bidder u. The objective is to maximize the total revenue generated over all impressions.
The ADWORDS problem was introduced in the seminal work of Mehta, Saberi, Vazirani, and Vazirani [22]. While
this problem generalizes the classic online matching problem introduced by Karp, Vazirani, and Vazirani [19], the
research focus has been on the so called small bids case, i.e., on algorithmic performance as the ratio maxv∈V buv
tends to 0 for all advertisers. This assumption models most real-world scenarios, where the revenue generated
from a single impression is infinitesimal compared to the total budget of an advertiser. The small bids assumption
distinguishes ADWORDS from online matching and makes them incomparable from a technical perspective; given
its natural applicability and popularity in Internet advertising, we will also focus primarily on the small bids case.
Since the introduction of these problems, several variants of ADWORDS and online matching have been studied,
motivated primarily by the evolving challenges advertisement publishers face in practice. For example, Agrawal and
Devanur [4] recently considered arbitrary linear and non-linear convex budget constraints for the stochastic input
model, and Devanur and Jain [12] studied concave returns on revenue, both motivated by problem features such as
Bu
Illustration of how the "Californian female" example translates into a bidder's budget in an instance of ADGENERAL
Fig. 1.
(and in this case also ADLAMINAR). For simplicity, we only have a Los Angeles residency subdivision for the 30-39 age
range, giving us four dimensions in total. Observe that Ks = {2, 3} for the 30-39 age-range budget, and and Ks = {1, 2, 3, 4}
for the overall budget; furthermore, we have that B{2,3}
= 2500. The algorithm has currently earned
$500, $100, and $400 on dimensions 1, 3, and 4, respectively (and $0 on dimensions 2); therefore, we have used $1025 out
of the overall budget's capacity of $2500. If an impression v is newly assigned to this bidder u such that r(1)
uv = $10 and
r(k)
uv = 0 for k = 2, 3, 4, then we will have then earned $510 on dimension 1 and $1035 overall.
= 1000 and B{1,2,3,4}
u
u
under-delivery penalties and pay-per-click advertisements. Another motivation for these variants that is receiving
increasing attention is that of impression diversification or representativeness. More specifically, a campaign contract
is usually of the form: "deliver ten million advertisements to Californian females in the month of July." Although
such an agreement clearly indicates a target group to which the publisher should restrict its assignments, often the
advertiser still wants the impressions to be equally spread among the sub-populations of the targeted group (e.g.,
in the above contract, the advertiser will likely be unhappy if their ads are only shown to white females in their
twenties living in Los Angeles). Although it can be at odds with short-term revenue gains, ensuring diversity is
crucial with respect to long-term revenue for the publisher, as advertisers that see a high return on investment (in
this case via reaching desired audiences) are more likely to continually purchase future contracts.
There are several recent works that address impression diversification (e.g., see [15], [7], and [17]). In many of
these results, the objective function of the problem incorporates a diversity penalty, usually in the form of a distance
function that incurs a cost if the algorithm's assignment differs too much from the advertiser's ideal allocation.
However another natural approach, which to the best of our knowledge has yet to be considered, is a scheme where
contracts are specified over multiple tiers of user-attribute granularity. Recalling our previous contract example, in
a multi-tier scheme the advertiser could further specify constraints in terms of age groups, e.g., "no more than four
million of the total ten million advertisements should be shown to age groups 20-29, 30-39, and 40+, respectively."
Further constraints could also be placed on each age group in terms of residency, e.g., "of the maximum four
million advertisements assigned in each age group, no more than two million should be shown to residents in Los
Angeles." This scheme has several appealing features, the foremost being that it allows advertisers to explicitly
indicate their true user-targeting goals with a high degree of expressibility (whereas penalty functions often assume
that the ideal allocation must follow the same distribution as the overall targeted population, i.e., is forced to be a
representative sample).
With this motivation, we introduce a generalization of the ADWORDS problem that we call ADWORDS with
general budgets (denoted ADGENERAL). As earlier, there is an offline set of advertisers U, and a set of impressions
V that arrive online. We also have a set of dimensions Ku for each bidder u that represent the smallest level of
{+ $2500 20-29 + 30-39 + 40+ + $1000 $1000 $1000 $∞ (dimension 1) LA 30-39 + + $500 (dimension 2) (dimension 3) (dimension 4) LA 30-39 + + {{$500 $400 $125 $500 + $400 + $125 = $1025 (k)
user-attribute granularity over which the bidder defines her budgets. In our previous example, one dimension would
correspond to "Californian females living in Los Angeles in the age range 20-29"; another dimension would be
"Californian females living outside of Los Angeles in age range 40+". For simplicity of notation, we will consider
a universal set of dimensions K = ∪uKu and assume that each bidder has this same set of dimensions; dimensions
in K\Ku will simply earn no revenue and have no budget constraints for bidder u. If an impression v is assigned to
uv (called the bid value in dimension k) on each dimension k ∈ K. Note
an advertiser u, then the algorithm earns r
(k)
that we allow multiple non-zero r
uv revenue entries for a single impression-bidder pair (u, v). This is to reflect the
fact that certain categories are more definite than others. For instance, the campaign that an advertisement belongs
to is known but attributes of an online user may be less certain. If an anonymous online user has a 60% chance of
being below 30 and 40% of being above 30 (e.g., based on browsing behavior), then the revenue earned for this
user should be split in the same proportion between these two user category dimensions. In this case, the bid vector
has a non-zero entry in multiple dimensions. (The reader is referred to Chatwin [11] for an overview on uncertainty
in determining online user attributes. Adhikari and Dutta discuss how attribute uncertainty is weighted in real-time
bidding strategies [1], while Ghosh et al. [15] highlights the different levels of information about the users available
to different entities in Internet advertising. Techniques for determining user attributes based on historical and prior
behavior have also been extensively studied in the marketing research community (see Barajas Jamora [6] and
references contained therein), where the mapping of user behavior to attributes is inherently probabilistic.)
The revenue generated from an advertiser u is subject to an arbitrary set of budget constraints Su, where each
(see Figure
constraint s ∈ Su caps the total revenue generated from a subset of dimensions Ks to a budget B
1 to see how subsets of dimensions are used to define the budgets in our example). As usual, the objective in the
ADGENERAL problem is to maximize the total revenue generated by the algorithm. We will measure our algorithms
using competitive analysis, which is the maximum ratio over all instances of the objectives of an optimal solution
and the algorithmic solution (see, e.g., [9]).
(s)
u
A natural question is whether this generalization changes the structure of the ADWORDS problem. To understand
this, let us consider an instance with a single advertiser. In this case, an algorithm that assigns all impressions to
the lone advertiser is clearly optimal for the ADWORDS problem. However, let us now consider the ADGENERAL
problem with 2 dimensions, and budget constraints of $1 each on dimensions {1, 2} and {2, 3}. Now, suppose the
first impression has a revenue of $1 on dimension 2 alone. Should the algorithm assign the impression to the lone
advertiser? If it does, then the instance will generate two impressions yielding a revenue of $1 on dimensions 1 and
3 each, while if it does not, then the instance will generate no other impression. Clearly, this example shows that no
algorithm can do better than a competitive ratio of 2, even with a single advertiser. One may object that the small
bids assumption is being violated, but replacing an impression of bid value $1 with 1/ impressions of bid value
→ 0 still produces a constant lower bound of 3/2. In fact, this lower bound is a manifestation of a more general
observation: it may be a better option for the algorithm to not allocate an impression, or to not earn revenue
from some of the dimensions, even when possible to do so. This is in sharp contrast to the classical ADWORDS
problem, where an impression should always be allocated if possible. Thus, the ADGENERAL problem introduces
an aspect of "admission control" to the ADWORDS framework. Our results for ADGENERAL are characterized in
the following theorem.
Theorem 1. The competitive ratio of the ADGENERAL problem is Θ(lg p) under the small bids assumption, where
p = maxu∈U,k∈K {s : k ∈ Ks} denotes the the maximum number of budget constraints for an advertiser to which
any dimension belongs.
Although there is a super-constant lower bound for ADGENERAL, we observe that many multi-tier budgets based
on real-world instances will have additional structure. As we saw earlier in our motivating example, each budget
constraint was a subdivision of a more general constraint, i.e., the total budget for Californian females was divided
into age groups to obtain the second level of budgets, and then each of these age groups was divided based on
residency to obtain the next level. In essence, many consumer and product classifications are naturally hierarchical.
If the budgets of an ADGENERAL instance are defined over such a taxonomy, then we have the additional structure
that the budget sets {Ks : s ∈ Su} for each bidder u form a laminar family, i.e., for every pair of intersecting sets
in {Ks : s ∈ Su}, one is contained in the other.
Thus, we also consider the ADWORDS with laminar budgets problem (ADLAMINAR). It turns out that laminar
budgets make admission control redundant - the algorithm can now earn revenue whenever possible. However,
there are other conceptual difficulties. Consider an instance with 2 dimensions, where an advertiser has a budget of
$1 for dimension 1 and an overall budget of $2 for dimensions {1, 2}. At any point in the algorithm, what is the
total budget of dimension 2, i.e., cap on total revenue earned from dimension 2? This value clearly depends on the
revenue earned on dimension 1, and therefore changes during the course of the algorithm. This is in sharp contrast
to the classical ADWORDS setting where the total budget of a bidder remains unchanged during the course of the
algorithm (note the distinction between total budget and remaining budget). The first technical hurdle, therefore, is
to define a dynamic notion of total budget on individual dimensions. In addition, we also need to define a notion
of current budget utilization for individual dimensions to determine which dimensions we should prefer in making
the allocation. Again, this notion is canonical in the classical ADWORDS problem – it is simply the fraction of the
budget of a bidder that has already been earned as revenue. In our more general setting, a single dimension might
be in multiple budget constraints, and therefore, we must first identify the most constraining budget. Once we do so,
we need a mechanism for importing the budget utilization of this constraining budget to the dimension itself. For
instance, in the example above, if the revenue earned on dimension 1 is $1 and that on dimension 2 is $0 at some
point in the algorithm, then should be budget utilization for dimension 2 be 0.5 (from its most constraining budget
constraint) or 0 (from the fact that the algorithm has not earned any revenue at all from dimension 2 yet)? It turns
out that these concepts (revenue cap, most constraining budget, and budget utilization of a dimension) are closely
tied to each other and have to be defined through a common inductive process. Our main technical contribution
for the ADLAMINAR problem is to carefully define these entities in a way that ensures semantic consistency and
eventually gives our main result for this problem: an algorithm with a competitive ratio of e/(e− 1), matching that
for the classic ADWORDS problem [22].
Theorem 2. The competitive ratio of ADLAMINAR is e/(e − 1) under the small bids assumption.
Finally, we study the ADGENERAL problem without the small bids assumption. In the absence of this assumption,
there are two possible variants – either (a) the algorithm can choose the amount of revenue it earns on any given
dimension (which can be less than the corresponding bid value) from an impression, or (b) the algorithm is
constrained to earn the entire bid value as revenue on any dimension, which means that an assignment of an
impression to an advertiser is only allowed if adding the bid value to the previously earned revenue on each
dimension does not violate any constraint. The former is more natural in the context of Internet advertising – we
call it the ADGEN-P problem1 and match the bounds in the small bids case.
Theorem 3. The competitive ratio of the ADGEN-P problem is Θ(lg p), where p = maxu∈U,k∈K {s : k ∈ Ks}
denotes the the maximum number of budget constraints for an advertiser to which any dimension belongs.
(cid:80)
We also study the latter problem (where the entire bid value is always added to the revenue), primarily because of
interesting connections to the classical online admission control problem [5]. In Theorem 4, we give the competitive
ratio of the ADGEN-AON problem (AON for "all or nothing"). As a byproduct of our result, we also obtain tight
bounds for the online admission control problem, slightly improving the classical bounds of [5].
Theorem 4. Let = maxu,v,s
denote the maximum bid-to-budget ratio and p = maxu∈U,k∈K {s : k ∈
Ks} denote the the maximum number of budget constraints for an advertiser that a dimension belongs to. If
lg(2p) < < 1, then the competitive ratio of the ADGEN-AON problem is Θ
Note that we do not consider ≤ 1
For reasons of brevity, we do not discuss the admission control problem here – the implication of the above
theorem to this problem is straightforward – and relegate the details for the large bids case, both ADGEN-P and
ADGEN-AON, to the appendix.
lg(2p) in Theorem 4, as in this case ADGEN-AON is essentially identical to
ADGENERAL with small bids.
(cid:16) p
k∈Ks
B(s)
u
(cid:17)
.
1−
r(k)
uv
1
1P for "partial"
Related Work. Given the large volume of work in this area, we will only sample a small fraction of the online
matching and ADWORDS literature, focusing on results in the (adversarial) online model. For a comprehensive
survey, including results in stochastic input models, the reader is referred to the survey by Mehta [20].
Karp, Vazirani, and Vazirani [19] introduced the online matching problem, and gave a tight e/(e− 1)-competitive
algorithm (see also [16], [8], and [13]). The first generalization was to the b-matching problem by Kalyanasundaram
and Pruhs [18]. Later generalizations include a vertex-weighted version by Aggarwal et al. [3], a pay-per-click
model using stochastic rewards by Mehta and Panigrahi [21] (see also [23]), and a bi-objective model suggested
by Aggarwal et al. [2]. Devanur and Jain [12] explored non-linear concave objectives to encode, e.g., penalties for
under-delivery. In terms of techniques, most of the initial results used combinatorial methods, but recent work has
focused on a (randomized) primal dual technique introduced by Devanur et al. [13].
The ADWORDS problem, which generalizes b-matching, was introduced by Mehta et al. [22], who gave an
e/(e − 1) approximation for small bids. They also showed that this competitive ratio is the best possible. Without
the small bids assumption, the greedy algorithm for the ADWORDS problem has a competitive ratio of 2, and
while this is tight for deterministic algorithms, obtaining a better ratio using a randomized algorithm is open.
Buchbinder et al. [10] gave an alternative primal-dual analysis for the algorithm of Mehta et al. [22] with the same
competitive ratio. More recently, other variants of the ADWORDS problem have been considered. For instance,
Feldman et al. [14] and Aggarwal et al. [3] introduced variants to model display ads with vertex weights and/or
capacities.
Paper Organization. In Section II, we prove an e/(e − 1) upper bound for ADLAMINAR under the small bids
assumption (Theorem 2). In Section III, we prove an O(lg p) upper bound for ADGENERAL under the small bids
assumptions (Theorem 1). Our results for ADGENERAL without the small bids assumptions (Theorems 3 and 4 for
ADGEN-P and ADGEN-AON, respectively), as well as our Ω(lg p) lower bound for ADGENERAL, can be found
in the appendix.
II. ADWORDS WITH LAMINAR BUDGET CONSTRAINTS (ADLAMINAR)
(k)
Recall the ADLAMINAR problem: we given a set of offline bidders U and a set of impressions V that arrive
online, where each bidder-impression pair (u, v) is specified by a bid value r
uv for each dimension k ∈ K. The
revenue generated from a bidder u is subject to an arbitrary set of budget constraints Su, where each constraint
(s)
s ∈ Su caps the total revenue generated from a subset of dimensions Ks to a given budget B
u . We assume that
the sets {Ks : s ∈ Su} form a laminar family, i.e., for every pair of intersecting sets in {Ks : s ∈ Su}, one is
contained in the other.
In this section, we give an algorithm for the ADLAMINAR problem with a competitive ratio of e/(e − 1) under
the small bids assumption (Theorem 2). This bound is tight because of a matching lower bound for the ADWORDS
problem [22]. Throughout this section, we assume that for every dimension k ∈ K, a constraint s with Ks = {k}
appears in Su for each bidder u. This is wlog since a budget can be made arbitrarily large. We will call these
constraints singleton budgets of bidder u.
A. Primal and Dual Formulations
Our algorithm uses a primal-dual formulation of the ADLAMINAR problem. In other words, we give a primal
LP formulation of the ADLAMINAR problem and its corresponding dual, and update the solutions to both LPs on
the arrival of a new impression. The primal updates, which are guided by the dual solution, define the algorithm.
The main challenge is to show that the dual updates maintain feasibility while ensuring that the ratio of the primal
and dual objectives remains bounded by the desired competitive ratio ρ = e/(e − 1).
(s)
Let us define the current budget utilization for constraint s of bidder u (denoted κ
u ) to be the fraction of budget
(s)
for impressions v assigned to bidder
B
u
u thus far. Let us call a dimension k active for bidder u if for all budgets s ∈ Su such that k ∈ Ks, the algorithm
(s)
u < 1. In other words, u's active dimensions are the ones on which the algorithm can still receive
currently has κ
revenue from u.
currently used by the algorithm, or formally, κ
k∈Ks,v r
u =(cid:80)
(s)
(k)
uv /B
(s)
u
An algorithm for ADLAMINAR might gain revenue from only a subset of dimensions when assigning an
impression to a bidder. To implement this flexibility in the LP, we introduce the notion of assignment types.
Note that Eq. (1) ensures that the algorithm receives no revenue from inactive dimensions for a bidder, and Eq. (2)
ensures that every impression v is assigned using a single type to a single bidder.
The dual D of this LP is defined as:
u
∀ u ∈ U, v ∈ V, t ⊆ K :
min
α(s)
u B(s)
u +
(cid:88)
(cid:88)
(cid:88)
s∈Su
s∈Su
(cid:32)
∀ u ∈ U, s ∈ Su :
∀ v ∈ V :
(cid:88)
(cid:88)
v
σv
(cid:33)
α(s)
u
(k)
r
uvt
k∈Ks
α(s)
u ≥ 0
σv ≥ 0.
(cid:88)
k
(k)
r
uvt
+ σv ≥
(3)
Let t ⊆ K. For impression v and bidder u, we define a type-t assignment as one where the dimensions in t are
active and the dimensions in K \ t are inactive. Thus, our decision variables for the LP will be of the form xuvt,
where the algorithm sets xuvt to be 1 if impression v is assigned to bidder u using a type-t assignment (and 0
otherwise). We then define r
(k)
uvt = 0. Our primal LP P is now defined as:
(k)
uvt = r
(k)
(cid:88)
(cid:88)
uv if k ∈ t; otherwise, r
(cid:88)
xuvt
(cid:88)
∀ u ∈ U, s ∈ Su :
max
xuvt
u,v,t
v,t
k
∀ v ∈ V :
u,t
(k)
r
uvt
(cid:88)
k∈Ks
xuvt ≤ 1
(k)
uvt ≤ B(s)
r
u
∀ u ∈ U, v ∈ V, t ⊆ K :
xuvt ≥ 0.
(1)
(2)
(s)
u = 1 once a budget B
Now, if we naıvely enforce α
(s)
u = 1, then the dual variable α
Unfortunately, the dual stated above cannot be used directly in a primal dual algorithm. If a constraint s has budget
(s)
utilization κ
u also needs to be equal to 1 in order to balance the contributions to
the two sides of the dual constraint by dimensions k ∈ Ks. (Note that the primal objective does not increase for
these dimensions and hence the value of σv in the dual objective cannot depend on these dimensions either, if the
ratio of the primal to dual objective is to be maintained.)
(s)
u
is tight, then the primal-dual ratio could be proportional
to the number of nested levels, if a nested set of budgets are all tight. To obtain a constant competitive ratio,
(s)
what our scheme will (roughly speaking) need to do is only set α
to 1 at the highest level of nesting for each
u
nested set of tight constraint. If we think of the primal objective as being "attributed" to dual variables in order to
maintain the primal dual ratio, then what we roughly want is that at any point of time, the primal objective from
a given dimension for some bidder u is attributed to a unique dual variable representing a budget constraint for u
(s)
u variables need to
containing that dimension. However, in order to implement this property in an online setting, α
be non-monotone since the budgets in lower nesting levels might become tight first followed by the higher levels.
(s)
Therefore, we need a means of raising and lowering each α
u so that at the end of the instance, the revenue earned
from a dimension for bidder u is attributed to exactly one of these variables. In general, non-monotonicity of dual
variables is undesirable in online algorithms because a satisfied dual constraint might become unsatisfied later. To
overcome this problem, we give a new dual D(cid:48) where we decompose α
into decision variables that are indeed
monotone in our eventual primal-dual analysis.
Formally, our transformed dual D(cid:48) is defined as follows. Since each Su is laminar, we can represent its set
system as a forest Fu, where each node in the forest corresponds to a constraint s ∈ Su, and the singleton budgets
are the leaves. Let As be the set of ancestors of s in Fu, including s itself. Define a new decision variable
(s)
(where for a maximal
γ
= 0). Using the new variables, we can rewrite Eqn. (3) in our original
set s with no parent in Fu, we set γ
(s(cid:48))
u , and let p(s) be the parent budget of s. Observe that α
u =(cid:80)
u − γ
(s)
u = γ
s(cid:48)∈As α
(p(s))
u
(p(s))
u
(s)
u
(s)
dual formulation as:
∀ u ∈ U, v ∈ V, t ⊆ K :
(cid:32)(cid:16)
(cid:88)
s∈Su
γ(s)
u − γ(p(s))
u
(cid:33)
(k)
r
uvt
(cid:17) (cid:88)
k∈Ks
(cid:88)
k
(k)
r
uvt.
+ σv ≥
(4)
Next, we observe that the outermost summation on the LHS of Eqn. (4) telescopes, and the only remaining γ
are those for singleton budgets. This gives us our final dual formulation D(cid:48):
(s)
u
(cid:88)
min
σv +
v
u
(cid:16)
(cid:88)
(cid:88)
(cid:88)
s∈Su
(cid:17)
(cid:88)
k
B(s)
u
γ(s)
u − γp(s)
u
(k)
γ({k})
u
uvt + σv ≥
r
≥ 0.
u − γ(p(s))
γ(s)
u
∀ u ∈ U, v ∈ V, t ⊆ K :
∀ u ∈ U, s ∈ Su :
k
(k)
r
uvt
(5)
(6)
B. Labeling Scheme
(s)
u
corresponding to the "most-limiting" budget B
(s)
u . However, simply using utilization to define γ
Recalling our above discussion, our goal will be to attribute the revenue earned on a dimension k for bidder
(s)
(s)
u to exactly one dual variable γ
u , ideally to the γ
such
u
(s)
that k ∈ Ks. This suggests that we should make γ
a monotonically increase function of the budget's current
u
(s)
utilization κ
u does not capture the interactions between budgets
(s)
in the laminar setting. The overarching issue with just using κ
is the fact that B
u might be the most utilized
(s)
constraint for only some of the dimensions in Ks, since other budgets that sit below B
in the hierarchy may have
u
higher utilization. This raises the following question: should the revenue currently constrained by these descendant
budgets, say revenue earned on some particular dimension k(cid:48), affect how the algorithm determines the extent to
(s)
which B
u limits other unbounded dimensions like k? The answer is not immediate. It is tempting to say "no" since
the dimension-k(cid:48) revenue is already bounded by a tighter budget; on the other hand, B
(s)
u might in fact become
the tightest budget for dimension k(cid:48) later in the instance and ignoring the dimension-k(cid:48) revenue till that time will
prevent a smooth transition of the tightest budget for k(cid:48).
To overcome this challenge, we introduce a labeling scheme (cid:96)
(s)
u will
represent the modified notion of the budget's utilization that we need to properly measure the remaining capacity
for future revenue. Our primal-dual analysis will then follow by making each dual variable a monotone function
of these labels.
u : ∪uSu → [0, 1]. Label (cid:96)
(s)
u for budget B
(s)
u
(s)
More concretely, we address the above challenge by having our labels maintain the following two high-level
features:
• For label (cid:96)
(s)
u
and bidder u, revenue from a dimension k ∈ s will only contribute to the label if (cid:96)
is at
least as large as the labels of all budgets containing k that are subsets of s. This corresponds to identifying
the "most constrained" budget for any dimension by interpreting these abstract labels as surrogates of actual
budget utilizations.
• In defining label (cid:96)
(s)
u , we need to identify the capacity of constraint s for future revenue earnings from the
(s)
dimensions that are deriving their label from s. We define this capacity as the total budget B
u minus the
budgets of constraints below s that have a higher label. This automatically discounts the revenue earning
capacities of dimensions that are deriving labels from descendant constraints of s.
(s)
u
One challenge with maintaining these properties is that they are somewhat circular. To determine the value of
a label, we need to first determine which dimensions count toward the label, but determining dimension inclusion
requires comparisons between label values. Another challenge is maintaining smoothness. As impressions are
assigned to bidders, budgets that were previously slack will become tight, which requires us to reassign dimensions
to labels and change their capacities. In order to make our primal-dual analysis smooth, we will need to ensure
that labels remain consistent after we reassign dimensions to labels and change label capacities.
(cid:80){k}∈L(s) R
(cid:80)
(k)
u
s(cid:48)∈T (s) B
(s)
u −
B
(cid:96)(s)
u =
.
(s(cid:48))
u
(7)
To overcome the issue of circularity, we will not give an explicit label definition but rather give a set of label
properties that we maintain throughout the algorithm. These properties are based on two sets of budgets, L(s)
and T (s), that the algorithm will dynamically update for all bidders u and budgets B
(we drop the subscript u
for simplicity). The two sets partition the descendant dimensions of s (i.e., each descendant dimension belongs to
exactly one set in L(s)∪ T (s)). Intuitively, L(s) contains singleton budgets {k} representing dimensions that count
(s)
toward label (cid:96)
u . On the other hand, T (s) contains the closest descendants of s that have a bigger label than s,
i.e., every dimension in s that is not in L(s) derives its label from a budget in T (s) or from one their respective
descendants.
(k)
u be the total revenue currently earned on dimension k for bidder u. Formalizing the above discussion,
Let R
we say the labels for bidder u are valid if the following three properties hold for all s ∈ Su.
1) Property 1: For all {k} ∈ L(s), all constraints s(cid:48) on the path from {k} to s in Fu have (cid:96)
(s(cid:48))
2) Property 2: For all s(cid:48)
∈ T (s), we have (cid:96)
u > (cid:96)
3) Property 3: The following identity holds:
(s(cid:48))
u ≤ (cid:96)
u , and for all s(cid:48)(cid:48) on the path from s(cid:48) to s, we have (cid:96)
(s)
u .
(s(cid:48)(cid:48))
u ≤ (cid:96)
(s)
u .
(s)
u
(s)
(s)
u
Note that once we have fixed sets T (s) and L(s) for all s, we can verify all three properties and use Eqn. (7) to
directly compute each label for each s ∈ Su. Also observe we can initialize all labels to be 0, and set T (s) = ∅
and L(s) = {{k} : k ∈ K} to start with a valid labeling. Finally, we note that we will soon show that all labels
remain non-negative (in the proof of Lemma 5).
Next, we define a procedure for updating labels after the revenue earned in a single dimension increases
infinitesimally. More specifically, suppose an impression v is assigned to bidder u, and assume we have a valid
(k)
labeling for the sets in Su before the assignment. The assignment of the impression changes the values of R
for
u
the active dimensions k. This necessitates label updates, which we define for an infinitesimally small increment in
the value of R
for a particular dimension k. Note that the overall assignment of the impression is a sequence of
such incremental changes.
(k)
u
The labels that we increase on such an increment are (cid:96)
such that {k} ∈ L(s), i.e., we will increase R
in
(s)
and leave T (s) and L(s) fixed. If the labeling is valid before a given
the numerator of Eqn. (7) for all such (cid:96)
u
(s)
increment, and if after the increment the relative ordering of all (cid:96)
remains the same, then by the definition of
u
Properties 1 and 2 the labeling remains valid. Thus, in terms of updating L(s) and T (s), we only need to consider
(k)
when the relative order of labels changes as a result of adding to R
u . In particular, there are two types of reordering
that need considered. We will call these two reordering possibilities Events 1 and 2 and describe how the algorithm
updates L(s) and T (s) in each case. Later, we will show that these set redefinitions maintain the current value of
the label.
• Event 1: For some s such that {k} ∈ L(s), there now exists a descendant s(cid:48) of s such that (cid:96)
(s)
u ,
where s(cid:48) is on the path from {k} to s and (cid:96)
u previously. In this event, s(cid:48) is now added to T (s). All
singleton-budgets {k} that are descendants of s(cid:48) and belong to L(s) are now removed from L(s). Additionally,
all descendants of s(cid:48) that belong to T (s) are also removed from T (s).
• Event 2: For some constraint s such that {k} ∈ L(s), there now exists a descendant s(cid:48) of s such that
∈ T (s) previously. In this event, s(cid:48) is removed from T (s). Conversely to Event 1, all
(s)
u ≥ (cid:96)
(cid:96)
constraints in T (s(cid:48)) are added to T (s), and all {k} ∈ L(s(cid:48)) are added to L(s).
In order to make the process smooth, we will think of the updates in Events 1 and 2 as being done at the transition
(s(cid:48))
point where (cid:96)
u . This completes the description of our labeling scheme and the process by which the
algorithm determines them. We encourage the reader to refer to Figure 2 for a small example of an Event 2
update.2 We now prove the following lemma, which will be useful for our primal-dual analysis.
(s(cid:48))
u , where s(cid:48)
(s(cid:48))
u ≤ (cid:96)
(s(cid:48))
u > (cid:96)
(s)
u = (cid:96)
(s)
(k)
u
2In terms of how L(s) and T (s) are updated, Event 1 is the reverse of Event 2. So, reversing the example in the Figure 2 will provide
the reader with an Event 1 example.
Illustration of the modifications made to L(s) and T (s) in Event 2, occurring between budget sets s = {1, 2, 3, 4} and
Fig. 2.
s(cid:48) = {1, 2}. In this example, the algorithm is incrementing the revenue on dimension 3. The state of the labels before dimension
3 revenue is added is shown on top. Notice that {3} ∈ L({1, 2, 3, 4}) since (cid:96){3}u = 1/10 is smaller than (cid:96){1,2,3,4}
= 1/5 (note
that the value of (cid:96){1,2,3,4}
corresponds to the left most rectangle, where the middle and right rectangles show the revenue levels
of the dimensions that are not included in the label). After increasing dimension 3 to $10/3, we have that (cid:96){1,2,3,4}
= 2/3, which
means it is about to surpass label (cid:96){1,2}
, therefore triggering Event 2. The state of the labels after the Event 2 modifications
are shown on bottom (at the transition point). Notice that {1, 2} has been removed from T ({1, 2, 3, 4}), and we have added
the sets in L({1, 2}) to L({1, 2, 3, 4}) (namely, {1} and {2}).
u
u
u
u
T({1,2,3,4})={{1,2},{4}}L({1,2,3,4})={{3}}{{B({1,2})u=$15{B({1})u=$10{$5{B({2})u=$10{$5{B({4})u=$10{$9{B({3})u=$10$1{B({1,2,3,4})u=$30B({1,2,3,4})u B({1,2})u B({4})u=30 15 10=$5{$1$5+$5=$10{$5{$5{$9$10/3{{{$40/3B({1,2,3,4})u B({4})u=30 10=$20L({1,2,3,4})={{1},{2},{3}}T({1,2,3,4})={{4}}$5+$5=$10{(dim.1)(dim.2)(dim.3)(dim.4)(dimension3increasesto$10/3) Lemma 5. For every constraint s, the label (cid:96)
Proof: Clearly when neither Event 1 or 2 occurs, (cid:96)
(s)
u is monotonically non-decreasing over the course of the algorithm.
(s)
u can only increase (this follows directly from the definition
(s)
of the update procedure). Thus, it suffices to show that (cid:96)
u does not decrease when it participates in Event 1 or 2.
(s)
In particular, we will show that (cid:96)
u has an identical value after L(s) and T (s) have been modified in either event.
We will show that this holds for Event 1, noting that the argument for Event 2 is identical.
Suppose the updates for Event 1 occur for a constraint s and a descendant s(cid:48), triggered by a dimension k ∈ s.
The changes are: s(cid:48) is added to T (s), all descendant singleton-budgets of s(cid:48) that were in L(s) are removed from
L(s), and all descendants of s(cid:48) that were in T (s) are removed from T (s). Recall that before the event, we have
that
(k)
u
(cid:80){k}∈L(s) R
(cid:80)
(cid:80){k}∈L(s(cid:48)) R
(cid:80)
(w)
w∈T (s) B
u
(k)
u
w(cid:48)∈T (s(cid:48)) B
(s)
u −
(s(cid:48))
u −
.
(w(cid:48))
u
(cid:96)(s)
u =
(cid:96)(s(cid:48))
u =
B
B
Therefore, using the definitions of L(s) and T (s) before they are modified by the event, the new (cid:96)
(s)
new) can be written as:
(cid:96)
(cid:80){k}∈L(s) R
(cid:80)
(k)
(cid:80){k}∈L(s(cid:48)) R
u +(cid:80)
(s(cid:48))
(k)
u
w(cid:48)∈T (s(cid:48)) B
.
(w(cid:48))
u
(s)
(cid:96)
new =
u −
u − B
at the moment Event 1 occurs, we have (cid:96)
u −
B
w∈T (s) B
(w)
(s)
(s)
u
(denoted
(8)
For the competitive analysis, it suffices to show that a) the ratio between dual and primal objectives is at most
ρ, and b) the dual solution is feasible.
(s(cid:48))
u
(s(cid:48))
u
(s)
u = (cid:96)
(s)
u = (cid:96)
(s)
new = (cid:96)
(which follows from the fact that
To complete the proof, note that the above argument does not exclude the possibility of (cid:80){k}∈L(s) R
Since (cid:96)
a/b = c/d = α implies (a − c)/(b − d) = α).
(k)
u = 0
(k)
and the denominator in Eqn. (8) being negative (if this were to happen, the increments to R
u would decrease
the label by making it more negative). However, in both events this cannot be the case. First observe that Event 2
can only occur between two non-zero labels (since in Event 2, s(cid:48)
∈ T (s) before the event, which implies a strict
(s(cid:48))
(s(cid:48))
(s)
inequality (cid:96)
u = 0, but in Event 1, the denominator of (cid:96)
u = (cid:96)
u
(s)
u . This is because Event 1 can only occur between two labels
must always be smaller than the denominator of (cid:96)
such that {k} ∈ L(s) where k the dimension is currently being incremented. Since (cid:96)
at the
(s(cid:48))
(s(cid:48))
(s)
transition point in Event 1, it must be the case that (cid:96)
u . This implies (cid:96)
u
u
(s)
u before the modifications to L(s) and T (s).
must have a smaller denominator than (cid:96)
(s(cid:48))
u ). Event 1 can (and will) occur when (cid:96)
is increasing at a higher rate than (cid:96)
(s)
is surpassing (cid:96)
u
(s)
u < (cid:96)
(s(cid:48))
u
C. Algorithm Definition and Analysis
Duv =(cid:80)
Using our dual formulation and labeling scheme, we are now ready to define and analyze our algorithm. Consider
(s(cid:48))
the arrival of impression v. Define gu(s) = maxs(cid:48)∈As (cid:96)
u , i.e., the maximum label of an ancestor of s in the
forest Fu (including s itself). Our algorithm assigns impression v to bidder u = arg maxu(cid:48)∈U{Du(cid:48)v}, where
For the rest of the section, let ρ = e/(e − 1). For a primal assignment of impression v to bidder u, we change
(k)
uv and tu is the current active dimensions for bidder u.
k∈tu(1 − eg({k})
u −1)r
the dual solution by setting σv = ρ · Duv and update γ
eg(s)
u − 1
e − 1
u −1 − e
is computed after the assignment of the current impression v.
= ρ(eg(s)
where g
γ(s)
u =
to be
(s)
u
(s)
u
−1),
:
(p(s))
u
u as the subset of constraints in Su where the value of γ
Primal-Dual Ratio. Our goal is to show that when an impression v is assigned to a bidder u, the change in dual
objective is at most ρ = e/(e− 1) times that of the primal objective. First, note that the dual objective is a function
of the labels, and we have argued above that the labels do not change when either of Event 1 or 2 happens.
Therefore, we only need to account for the change in the dual objective when the labels change but neither of the
two events happen. Let us define S∗
is different from
γ
We can rewrite the dual objective as(cid:80)
(cid:16)
(cid:80)
v∈V σv +(cid:80)
∗
u (cid:54)= γ(p(s))
u = {s ∈ Su : γ(s)
S
since for all the other terms, the
s∈S∗
u − γ
γ
= 0. For any constraint s ∈ S∗
u(s) be its closest ancestor in Fu that is also in S∗
u, let p∗
value of γ
u.
u , T (s) = {s(cid:48)
u(s(cid:48)) = s} for any s ∈ S∗
B(s)
Now, observe that by Property 2 of labels and the definitionof g
u.
(cid:88)
(cid:88)
Then, the dual objective can be further rewritten as
.
(cid:88)
(cid:88)
u − γ
∈ S∗
u : p∗
B(s(cid:48))
(p(s))
u
(p(s))
u
(s)
B
u
(cid:17)
γ(s)
u
σv +
(s)
u
}.
(s)
(s)
(s)
u
u
u
u
v∈V
u
s∈S∗
u
u −
s(cid:48)∈T (s)
As earlier, we will analyze the change in the dual and primal objectives when the revenue on a dimension k
uv . Note that for any singleton constraint {k}, there is a unique
is incremented by an infinitesimal amount ∆r
({k})
u . Therefore, the only dual variable in S∗ (i.e., in the
s ∈ S∗
u satisfying k ∈ L(s); furthermore, g
= g
u
(s)
(s)
dual objective given above) that changes is γ
u . Let us denote the change in g
u . Using the small bids
assumption, we can write:
(s)
u by ∆g
(s)
u = (cid:96)
(k)
(s)
B(s)
(cid:88)
∆γ(s)
u ·
u −
s(cid:48)∈T (s)
B(s(cid:48))
u
=
(s)
∂γ
u
(s)
∂g
u
· ∆g(s)
u ·
B(s)
(cid:80)
u −
(cid:88)
s(cid:48)∈T (s)
B(s(cid:48))
u
∆r
(k)
uv
s(cid:48)∈T (s) B
·
(s(cid:48))
u
B(s)
u −
B(s(cid:48))
u
(cid:88)
s(cid:48)∈T (s)
Summing over all the infinitesimal changes in revenue, the total change in the dual objective for the assignment
(k)
uv , where tu is the set of active dimensions. Since the
of impression v is given by σv + ρ ·
algorithm sets
(cid:80)
(s)
B
u −1) ·
= ρ · (eg(s)
u −
u −1) · ∆r(k)
= ρ · (eg({k})
uv .
u −1) · r
k∈tu(eg({k})
(cid:88)
σv = ρ ·
k∈tu
the total change in the dual objective can be written as:
u −1)r(k)
uv ,
(1 − eg({k})
(cid:88)
k∈tu
(cid:88)
k∈tu
r(k)
uv ,
(1 − eg({k})
u −1)r(k)
uv + ρ ·
u −1) · r(k)
(eg({k})
uv = ρ ·
(cid:88)
k∈tu
ρ ·
which is exactly ρ times the increase in the primal objective.
Dual Feasibility. Finally, we argue that the dual is feasible when the algorithm terminates.
Lemma 6. At the end of the algorithm, the dual is feasible.
Proof: The feasibility of Eqn. (5) follows directly from definition of g
is a non-deceasing
(s)
u . We now show Eqn. (6). Let tu be the set of active dimensions for bidder u when impression v
= 1. This follows from the fact that if k is inactive,
u = 1 as well). Since s is an ancestor of {k}
function of g
arrived. First, observe that for all k (cid:54)∈ tu, we have that g
a constraint s ∈ Su containing k has reached κ
({k})
in Fu, we also have g
u
(s)
u = 1 (and thus (cid:96)
(s)
u , and the fact γ
({k})
u
({k})
u
= 1.
= γ
(s)
u
(s)
Let u(cid:48) be the bidder to which the algorithm assigned impression v. Let (cid:100)g
assigned (and define (cid:100)g
(s)
u(cid:48) similarly). We have the following:
(s)
u be the value of g
(s)
u when v was
γ({k})
u
(k)
r
uvt + σv =
=
(cid:88)
(cid:88)
k
k(cid:54)∈tu
(cid:88)
γ({k})
u
(k)
r
uvt + ρ
(cid:92)
u(cid:48) −1)r
g({k})
(k)
u(cid:48)v
(cid:88)
k∈tu
(1 − e
k∈tu
u −1 − 1/e)r
(eg({k})
({k})
u
and the fact that γ
(k)
uvt + ρ
(k)
r
uvt + ρ
(cid:88)
k∈tu
({k})
u
(by substituting σv)
(cid:92)
u(cid:48) −1)r
g({k})
(k)
u(cid:48)v,
(1 − e
(9)
= 1 for all k (cid:54)∈ tu. We can now
γ({k})
u
(k)
uvt + σv ≥
r
≥
≥
(cid:88)
(cid:88)
(cid:88)
k(cid:54)∈tu
k(cid:54)∈tu
k(cid:54)∈tu
(cid:88)
(cid:88)
(cid:88)
k∈tu
k∈tu
k∈tu
(k)
uvt + ρ
r
(k)
r
uvt + ρ
(k)
r
uvt + ρ
(cid:88)
(cid:88)
k∈tu
k∈tu
(k)
uvt + ρ
(k)
uvt + ρ
u −1 − 1/e)r
(eg({k})
u −1 − 1/e)r
(eg({k})
(cid:88)
(1 − 1/e)r
(k)
uvt =
k
(k)
r
uvt
(cid:92)
u −1)r
g({k})
(k)
u(cid:48)v
u −1)r(k)
(1 − e
(1 − eg({k})
(since ρ = e/(e − 1)).
uv
(cid:88)
k
(cid:88)
k
where the second equality follows by substituting γ
establish Eqn. (6) as follows:
The first inequality follows from Eqn. (9) and the fact that the algorithm assigns v to u(cid:48) = arg maxu∈U{Duv}. The
second inequality is because 1 − eg({k})
since r
u −1 is a non-increasing function of g
, and the third inequality follows
({k})
u
(k)
uv ≥ r
(k)
uvt.
III. ADWORDS WITH GENERAL BUDGET CONSTRAINTS (ADGENERAL)
Recall the ADGENERAL problem: we given a set of offline bidders U and a set of impressions V that arrive
online, where each bidder-impression pair (u, v) is specified by a bid value r
uv for each dimension k ∈ K. The
revenue generated from a bidder u is subject to an arbitrary set of budget constraints Su, where each constraint
s ∈ Su caps the total revenue generated from a subset of dimensions Ks to a given budget B
In this section, we will prove an O(lg p) upper bound for ADGENERAL (Theorem 1).
(s)
u .
(k)
A. Algorithm Definition
As in Section II, let κ
exponential potential function defined by:
(s)
u denote the current utilization of budget B
(s)
u . The algorithm (we call it ALGO) uses an
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:16)
(s)
B
u
p
φ =
φ(s)
u =
u
s
u
s
(2p + 2)κ(s)
u − 1
,
(cid:17)
(s)
u
(s)
where κ
u
φ = 0 initially.
is defined as the fraction of B
At any stage of ALGO, a dimension k is said to be active for bidder u if and only if(cid:80)
u ≤ 1; otherwise,
dimension k is said to be inactive for bidder u. (Note that this is a different definition of active dimensions than
what is used in Section II). ALGO only attempts to earn revenue on active dimensions, and hence, the total revenue
if impression v is allocated to bidder u is given by:
that has already been used by the algorithm at any stage. Note that
φ(s)
B(s)
s:k∈s
u
ruv =
r(k)
uv , where Au is the set of current active dimensions for bidder u.
(cid:88)
k∈Au
The algorithm makes a greedy assignment with respect to ruv, i.e., it assigns impression v to arg maxu ruv. Note
that it is possible that Au = ∅ for all bidders u, and therefore the algorithm does not assign impression v to any
bidder, even though there are dimensions and bidders where it could have earned revenue. This completes the
description of our algorithm.
For our analysis, it will be sufficient to quantify the small bids assumption as the following property for any
B. Algorithm Analysis
impression v, bidder u, dimension k, and constraint s such that k ∈ s:
.
1
lg(2p + 2)
(cid:88)
k∈s
(k)
r
uv
(s)
u ≤
B
(10)
We first establish the feasibility of the solution, which follows almost directly from how we define active dimensions.
Lemma 7. If ALGO assigns an impression v to a bidder u, then it can earn revenue on all the active dimensions
Au of u without violating any constraint.
Proof: We need to show that for all constraints s of bidder u,
(cid:88)
(k)
r
uv
(s)
u ≤ 1.
k∈Au∩s
B
κ(s)
u +
> 1
Suppose not. Then, for some constraint s,
(k)
r
uv
(s)
u
κ(s)
u +
B
(cid:88)
i.e., (2p + 2)κ(s)
lg(2p+2) > 2p + 2
k∈Au∩s
1
lg(2p + 2)
u + 1
(s)
u
(s)
u
i.e., (2p + 2)κ(s)
i.e., pφ
B
i.e., φ
B
(s)
u
(s)
u
u > p + 1
+ 1 > p + 1
> 1,
i.e., κ(s)
u +
> 1
(by the small bids assumption Eqn. (10))
which contradicts the fact that dimension k is active for bidder u.
This lemma implies that ALGO is indeed able to earn revenue on all active dimensions of a bidder u when it
assigns an impression to u.
Next, we will bound the total revenue of an optimal solution that we denote by OPT against the total revenue
of ALGO. Let uOPT(v) (resp., uALGO(v)) be the bidder that OPT (resp., ALGO) allocates impression v to. For every
dimension k that OPT earns revenue from, one of the following holds:
1) Case 1: dimension k is active for bidder uOPT(v) in ALGO when impression v arrives, and ALGO assigns v
to the same bidder, i.e., uALGO(v) = uOPT(v).
2) Case 2: dimension k is active for bidder uOPT(v) in ALGO when impression v arrives, but ALGO assigns v
to a different bidder, i.e., uALGO(v) (cid:54)= uOPT(v).
3) Case 3: dimension k is inactive for bidder uOPT(v) in ALGO when impression v arrives.
We partition the dimensions that OPT earns revenue from into active and inactive dimensions (according to their
status in ALGO for bidder uOPT(v)). For active dimensions, the next lemma gives a straightforward charging argument
using the greediness of the choice made by ALGO.
Lemma 8. For any impression v, the total revenue earned by OPT on the active dimensions is at most the total
revenue earned by ALGO overall.
Proof: For case 1 above (OPT and ALGO choose the same bidder), the two algorithms earn the same revenue
on the active dimensions. For case 2 above, the fact ALGO makes a greedy choice implies that it earns at least as
much revenue by assigning to different bidder as it would have made by assigning to u, which includes the revenue
on all the active dimensions.
The more involved case is that of inactive dimensions. In this case, we use a different global charging argument
over all dimensions, based on the potential function. In particular, we show that for a bidder u, the total revenue
of OPT from inactive dimensions (recall that this only includes revenue from impressions that arrived after the
dimension became inactive in ALGO) can be charged, up to a logarithmic loss, to the revenue that ALGO earned
overall from bidder u.
Lemma 9. Fix a bidder u. The total revenue that OPT earns in inactive dimensions for bidder u is at most the
final potential of bidder u in ALGO.
Proof: For any dimension k, let v ∈ Vu,k denote the subset of impressions assigned to u by OPT that arrived
after k became an inactive dimension for bidder u in ALGO. We need to bound the total revenue earned by OPT
on dimension k from impressions in Vu,k, summed over all k. For any impression v ∈ Vu,k, we have:
(cid:88)
φ
(s)
u
(s)
u
s:k∈s
B
> 1,
is the final potential for constraint s of bidder u. Thus,the revenue that OPT earns from impressions
(s)
u
where φ
v ∈ Vu,k on dimension k, summed over all dimensions, can be bounded as follows:
(k)
r
uv
(s)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
r(k)
uv <
φ(s)
u
r(k)
uv
=
φ
v∈Vu,k
k
s:k∈s
B
s
k∈s
v∈Vu,k
B
v∈Vu,k
k
(s)
u
(s)
u
u ≤
(cid:88)
s
φ(s)
u ,
where the last inequality follows from the feasibility of OPT.
Finally, we need to lower bound the final potential of a bidder in terms in terms of the revenue it generates for
ALGO. This is done in the following lemma.
Lemma 10. The increase in potential of a bidder u during the course of ALGO is at most 4 lg(2p + 2) times the
revenue that ALGO earns from u.
Proof: Suppose ALGO assigns impression v to bidder u. Let Ka denote the set of active dimensions for bidder
u when this assignment is made. Let η
. Then, the increase in potential is given by:
k:k∈Ka∩s
r(k)
uv
B(s)
u
(cid:16)
(2p + 2)κ(s)
u +η(s)
(cid:17)
(cid:16)
(cid:16)
u
(2p + 2)η(s)
u − (2p + 2)κ(s)
(cid:17)
(cid:17)
u − 1
u ·lg(2p+2) − 1
2η(s)
(cid:88)
· 2 lg(2p + 2) ·
u
(cid:33)
(cid:33)
·
(cid:88)
s
∆φ(s)
u =
=
=
≤
s
(s)
(s)
B
u
p
u =(cid:80)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(s)
B
u
p
s
s
(2p + 2)κ(s)
(s)
B
u
p
(s)
B
u
p
φ(s)
u +
(cid:32)
(cid:32)
u = (cid:80)
(cid:32)(cid:88)
φ(s)
u +
(s)
r(k)
uv ·
s
(cid:88)
k∈Ka
(cid:33)
(cid:88)
s
(k)
r
uv
(s)
u
B
,
(11)
k:k∈Ka∩s
where the last inequality follows since η
rearranging the RHS of inequality (11), we have:
k:k∈Ka∩s
r(k)
uv
B(s)
u
and ax ≤ 1 + ax for 0 ≤ x ≤ 1, a ≥ 1. By
∆φ(s)
u ≤ 2 lg(2p + 2) ·
φ
(s)
u
(s)
u
s:k∈s
B
+ {s : k ∈ s}
p
≤ 4 lg(2p + 2) ·
r(k)
uv ,
since k is active and p ≥ {s : k ∈ s}.
completing proof of the upper bound in Theorem 1.
A competitive ratio of O(lg p) for ADGENERAL in the small bids case now follows from Lemmas 7, 9, and 10,
(cid:88)
k∈Ka
REFERENCES
[1] Shalinda Adikari and Kaushik Dutta. Real time bidding in online digital advertisement. In DESRIST, pages 19–38, 2015.
[2] Gagan Aggarwal, Yang Cai, Aranyak Mehta, and George Pierrakos. Biobjective online bipartite matching. In WINE, pages 218–231,
2014.
[3] Gagan Aggarwal, Gagan Goel, Chinmay Karande, and Aranyak Mehta. Online vertex-weighted bipartite matching and single-bid
budgeted allocations. In SODA, pages 1253–1264, 2011.
[4] Shipra Agrawal and Nikhil R. Devanur. Fast algorithms for online stochastic convex programming. In SODA, pages 1405–1424, 2015.
[5] Baruch Awerbuch, Yossi Azar, and Serge A. Plotkin. Throughput-competitive on-line routing. In STOC, pages 32–40, 1993.
[6] Joel Barajas Zamora. Online display advertising causal attribution and evaluation. 2015.
[7] Vijay Bharadwaj, Peiji Chen, Wenjing Ma, Chandrashekhar Nagarajan, John Tomlin, Sergei Vassilvitskii, Erik Vee, and Jian Yang.
SHALE: an efficient algorithm for allocation of guaranteed display advertising. In SIGKDD, pages 1195–1203, 2012.
[8] Benjamin E. Birnbaum and Claire Mathieu. On-line bipartite matching made simple. SIGACT News, 39(1):80–87, 2008.
[9] Allan Borodin and Ran El-Yaniv. Online computation and competitive analysis. Cambridge University Press, 1998.
[10] Niv Buchbinder, Kamal Jain, and Joseph Naor. Online primal-dual algorithms for maximizing ad-auctions revenue.
In ESA, pages
253–264, 2007.
5990–6007, 2013.
[11] Richard E. Chatwin. An overview of computational challenges in online advertising. In American Control Conference, ACC, pages
[12] Nikhil R. Devanur and Kamal Jain. Online matching with concave returns. In STOC, pages 137–144, 2012.
[13] Nikhil R. Devanur, Kamal Jain, and Robert D. Kleinberg. Randomized primal-dual analysis of RANKING for online bipartite matching.
[14] Jon Feldman, Nitish Korula, Vahab S. Mirrokni, S. Muthukrishnan, and Martin P´al. Online ad assignment with free disposal. In WINE,
[15] Arpita Ghosh, Randolph Preston McAfee, Kishore Papineni, and Sergei Vassilvitskii. Bidding for representative allocations for display
[16] Gagan Goel and Aranyak Mehta. Online budgeted matching in random input models with applications to adwords. In SODA, pages
In SODA, pages 101–107, 2013.
pages 374–385, 2009.
advertising. In WINE, pages 208–219, 2009.
[17] S. Ali Hojjat, John Turner, Suleyman Cetintas, and Jian Yang. Delivering guaranteed display ads under reach and frequency requirements.
In AAAI, pages 2278–2284, 2014.
[18] Bala Kalyanasundaram and Kirk Pruhs. An optimal deterministic algorithm for online b-matching. Theor. Comput. Sci., 233(1-2):319–
[19] Richard M. Karp, Umesh V. Vazirani, and Vijay V. Vazirani. An optimal algorithm for on-line bipartite matching. In STOC, pages
[20] Aranyak Mehta. Online matching and ad allocation. Foundations and Trends in Theoretical Computer Science, 8(4):265–368, 2013.
[21] Aranyak Mehta and Debmalya Panigrahi. Online matching with stochastic rewards. In FOCS, pages 728–737, 2012.
[22] Aranyak Mehta, Amin Saberi, Umesh V. Vazirani, and Vijay V. Vazirani. Adwords and generalized online matching. J. ACM, 54(5),
[23] Aranyak Mehta, Bo Waggoner, and Morteza Zadimoghaddam. Online stochastic matching with unequal probabilities. In SODA, pages
982–991, 2008.
325, 2000.
352–358, 1990.
2007.
1388–1404, 2015.
APPENDIX A
ADGEN-AON LOWER BOUND
In this section, we prove the lower bound for Theorem 4. As a byproduct of this lower bound, we also obtain
tight bounds for the online admission control problem, slightly improving the classical bounds of [Awerbuch et al.
1993].
lg(2p) implies that p
Recall that we assume > 1
lg(2p). In our instance, we have a single advertiser with a set of p budget constraints
(we index budget constraints by s), where each constraint has a budget of 1. We will assume that p
1− is integral.
1− > 2.) Each impression has bid value of on a single, unique dimension.
(Note that > 1
(This allows us to use impressions and dimensions interchangeably in the rest of the construction.) We are now
left to specify the mapping of impressions to constraints. Let us denote (cid:96) = 1−
. First, we construct a hierarchical
segmentation of the constraints 1, 2, . . . , p, where segment j of level i comprises constraints (j− 1)· p1−i/(cid:96) + 1, (j−
1) · p1−i/(cid:96) + 2, . . . , j · p1−i/(cid:96). Overall, there are (cid:96) + 1 levels i = 0, 1, . . . , (cid:96), and level i partitions the overall set
of p constraints into pi/(cid:96) segments j = 1, 2, . . . , pi/(cid:96). Each such segment comprises p1−i/(cid:96) constraints. Note that
the segments in level i are a refinement of the segments in level i − 1. For the purpose of visualization, the reader
can imagine a complete p1/(cid:96)-ary tree on p leaves, where each leaf represents a constraint and each internal node
corresponds to a segment.
The online arrival of impressions is divided into (cid:96) + 2 rounds. The first round is special and is called the initial
round (described below). Every subsequent round, indexed 0, 1, . . . , (cid:96), comprises a set of impression blocks. Each
impression block of round i corresponds to a unique segment in level i. Such an impression block comprises 1/
identical impressions, all of which appear in all constraints of the segment of level i that the impression block
corresponds to. Clearly, the total number of impression blocks in round i is at most the total number of segments
in level i, i.e., at most pi/(cid:96). However, not all segments receive an impression block corresponding to it; only the
active segments (we define this notion below) in level i receive an impression block. Therefore, the arrival rule for
round i is simple: every active segment in level i receives a unique impression block. The relative order of arrival
of impressions in a round is arbitrary. As we mentioned above, the initial round is special - in this round, there
is a single impression with bid value δ > 0 that appears in all constraints.
Now, we are only left to describe the rule for defining active segments. After the initial round, all segments
in all levels are inactive, except the single segment in level 0 comprising all dimension. Therefore, in round 0,
there is a single impression block corresponding to all the constraints. Next, we define the inductive process for
activating segments. Recall that the impression blocks arriving in round i ≥ 0 correspond to the active segments in
level i. The algorithm assigns some subset of these impressions to the sole advertiser. For any segment j in level i
that received an impression block, if the algorithm assigned t > 0 impressions from the block, then all impression
blocks at level i + t that are refinements of the current block are made active. After round i, all segments in level
i are made inactive. This completes the description of the instance.
First, we need to show that the construction is valid. In particular, we need to show that i + t ≤ (cid:96) (i.e., there is
a level where the refined segments can be activated) if the algorithm assigns t impressions of a segment in round
i. We prove a more general lemma.
Lemma 11. At any point of time, if a segment in level i is active, then every constraint in that segment has a
current utilization of i + δ.
Proof: The algorithm must assign the impression in the initial round to stay competitive; therefore, the lemma
holds for i = 0. Inductively, the segments that are made active at level i + t had utilization i before round i (by
the inductive hypothesis) and have an additional utilization of t from round i.
is valid.
As a corollary of this lemma, we can infer that i+t ≤ (cid:96) since utilization cannot exceed 1; hence, the construction
Let (cid:96)s be the last round where constraint s was in an active segment, and let the corresponding segment be τs.
For every constraint s, the optimal solution assigns the entire impression block corresponding to τs in level (cid:96)s.
Note that the segments τs partition the set of constraints, and hence, the optimal assignment is feasible. Clearly,
the optimal solution earns a revenue of 1 on the impression block corresponding to segment τs.
To compare the revenue of the algorithm, we redistribute the revenue earned by the algorithm on an impression
by dividing it equally among all the constraints that the impression appears in. Next, we sum the revenues on
constraints in the same segment τs. We will now compare this revenue on τs with the unit revenue that the optimal
solution earns. To upper bound the (redistributed) revenue of the algorithm, we first note that the algorithm does
not earn any revenue on the impression block τs itself in round (cid:96)s. If (cid:96)s ≤ (cid:96) − 1, this follows from the fact that
refinements of τs were not marked active. If (cid:96)s = (cid:96), then by Lemma 11, the utilization at the beginning of the
round is 1 − + δ, which prevents the algorithm from earning any further revenue (note that revenue comes in
units of ).
Lemma 12. The total revenue earned by segment τs in the algorithm is at most
2
p1/(cid:96) .
Proof: Let r0, r1, . . . , rj denote the rounds in which segment τs earns revenue, and let t0, t1, . . . , tj be the
number of impressions assigned in the respective rounds. In other words, r0 = 0, r1 = t0 +r0, r2 = t1 +r1, . . . , (cid:96)s =
rj + tj. Then, the revenue earned by constraint s is
j(cid:88)
q=0
(cid:96)s−1(cid:88)
i=0
tq ·
1
p1−rq/(cid:96) ≤ ·
1
p1−i/(cid:96) ≤
2
p1−((cid:96)s−1)/(cid:96) ,
where the last inequality follows since p−1/(cid:96) ≤ 1/2 for (cid:96) ≤ lg p which follows from > 1
p1−(cid:96)s/(cid:96) constraints, the revenue earned by the algorithm on segment τs is at most
The lower bound now follows by comparing Lemma 12 to the optimal revenue, and setting δ → 0.
p1/(cid:96) .
2
lg(2p). Since τs comprises
APPENDIX B
ADGENERAL AND ADGEN-P LOWER BOUND (THEOREM 1 AND THEOREM 3)
In this section, we prove a Ω(lg p) lower bound for ADGENERAL that will be constructed via a reduction from
the online problem considered by Awerbuch et al. in [5], which we call ADMISSION-CONTROL. At the end of
the section, we will describe how the problem definition of ADMISSION-CONTROL can be modified so that the
reduction implies the same lower bound for ADGEN-P. We formally define this problem as follows.
ADMISISON-CONTROL: At the outset, the online algorithm is given an edge-capacitated graph G = (V, E), where
ce denotes the capacity of edge e. Demand requests then arrive in an online sequence R = (cid:104)r1, . . . , rh(cid:105), where each
request ri is specified by a path Pi between two vertices (si, ti) and a capacity demand di. Upon the arrival of
ri, the algorithm must decide to either reject the request or route it along Pi. The objective of the algorithm is to
maximize the total demand of admitted requests subject to the constraint that for any edge e, the sum of capacity
demands from admitted requests using e does not exceed ce.
We note that this is a less general problem than the one considered in [5] (e.g. in the original problem, the
algorithm can choose routing paths, requests have arrival times, etc.); however to extend their lower bound, the
above problem definition will suffice.
For completeness, we will give the proof of the "small-demands" ADMISSION-CONTROL lower bound from [5]
(i.e. in the given instance, the maximum demand-to-capacity ratio is arbitrarily small). This will be useful as our
reduction will not be completely "black-box", i.e., we need to have some knowledge of the proof's online sequence
in order to perform the reduction.
Let L(n) be the line graph defined on n + 1 vertices {v1, . . . , vn+1} (with n edges). Then the following lemma
holds.
Lemma 13. ( [5]) Let A an online algorithm for ADMISSION-CONTROL under the small demands assumption.
Then there exists an instance I for L(n) such that the capacity earned by the optimal solution is Ω(lg n) times the
demand earned by A.
Proof:
2i
Without loss of generality, assume n is a power of two. The instance will consist of lg n + 1 phases indexed by
i = 0, . . . , lg n. In phase i, we will issue 2i groups of requests. The jth group in phase i consists of 1/δ requests
each with capacity demand δ and identical routing paths Pj (and so the total capacity for each group in every phase
is 1).
Fix a phase i and group j ∈ {0, 1, . . . , 2i − 1}. Then path Pj is defined to be the segment of L(n) starting at
. In other words, in a given phase we are splitting L(n) into 2i edge -disjoint
vertex v jn
subsegments each of the length n/2i, where each subsegment defines a path for a group.
and ending at vertex v (j+1)n
This implies that
Let xi be the total demand admitted by the algorithm from requests in phase i. Observe that in order to admit a
unit of demand from the requests in phase i, the algorithm must use up a total capacity of n/2i (since each path
i=0 2−inxi ≤ n.
Pj in phase i has n/2i edges). Since the total capacity of edges in the graph is n, we have that(cid:80)lg n
Let Sk = 2−k(cid:80)k
i=0 xi be the normalized total capacity obtained by the algorithm from phases 0 through k. By
−ixi ≤ 1.
lg n(cid:88)
(12)
i=0
2
2i
equation (12), we have that
lg n(cid:88)
k=0
Sk =
lg n(cid:88)
k(cid:88)
k=0
i=0
lg n(cid:88)
k=0
2
−kxi ≤ 2
2
−kxk ≤ 2.
Thus, there exists a k(cid:48) such that Sk(cid:48) ≤ 2/ lg n, implying the algorithm only earned at most 2k(cid:48)+1/ lg n total capacity
after the completion of phase k(cid:48). The optimal solution at this point is to reject all requests before phase k(cid:48) and
admit all requests in phase k(cid:48) to obtain capacity 2k(cid:48). Therefore, the adversary can stop the sequence after phase k(cid:48)
to obtain the desired instance.
Given Lemma 13 and its proof, let I be the instance implied by the statement of the lemma, and let I(cid:48) be the
entire of sequence of requests given by the construction (all lg n + 1 phases, regardless of the algorithm's behavior).
We construct our lower bound instance for ADGEN-AON (in the same bids case) as follows.
• There will be one bidder u for the instance.
• Each group in I(cid:48) will correspond to both a dimension and an impression. Specifically, for all impressions that
correspond to the vth group in I(cid:48), we set r
(v)
uv = δ and r
• Each edge e will correspond to a budget constraint B
(se)
u
uv = 0 for all k (cid:54)= v.
, where the capacity of the budget is ce = 1 and set
se is defined to be the set of dimensions whose corresponding request group use edge e along their paths in
I(cid:48).
Observe that the single request in phase 1 of I(cid:48) traverses all of L(n); therefore based on the construction, the
first dimension belongs to n different budget constraints, which implies p = n. It now follows from Lemma 13
that any algorithm for ADGEN-AON is Ω(lg p) competitive.
(k)
Extension to ADGEN-P: To show a Ω(lg p) lower bound for ADGENERAL (without the small bids assumption),
we modify the definition of ADMISSION-CONTROL so that algorithm chooses to accept each request with some
fraction fi ∈ [0, 1] (i.e., the algorithm routes demand fi·di for request ri). We then change the lower bound instance
in Lemma 13 so that each request has unit demand (instead of issuing 1/δ requests with δ demand). The remainder
of the reduction is equivalent. Using the same arguments as before, we obtain the desired Ω(lg p) lower bound for
the ADGEN-P setting.
APPENDIX C
ADGEN-P AND ADGEN-AON UPPER BOUND (THEOREM 3 AND THEOREM 4)
In this section, we give our upper bounds for ADGEN-P and ADGEN-AON. We will first present the algorithm
under the context ADGEN-AON. At the end of the section, we will outline how the same analysis extends to
ADGEN-P. In both settings, the algorithm and analysis will be almost identical to our algorithm for ADGENERAL
in Section III.
A. Algorithm Definition
Again let κ
(s)
u denote the fraction of B
(s)
u
an exponential potential function defined by:
currently used by the algorithm. The algorithm (we call it ALGO) uses
(cid:88)
(cid:88)
(cid:88)
(cid:88)
φ =
φ(s)
u =
(cid:18)
(s)
B
u
p
(p + 1)
(s)
u
κ
1− − 1
.
(cid:19)
s
u
Note that φ = 0 initially.
At any stage of ALGO, a dimension k is said to be active for bidder u if and only if (cid:80)
u ≤ 1, else
dimension k is said to be inactive for bidder u. (Note that this is a different definition of active dimensions than
what is used in Section II). ALGO only attempts to earn revenue on active dimensions, and hence, the total revenue
if impression v is allocated to bidder u is given by:
φ(s)
B(s)
s:k∈s
u
s
u
ruv =
r(k)
uv , where Au is the set of current active dimensions for bidder u.
(cid:88)
k∈Au
The algorithm makes a greedy assignment with respect to ruv, i.e., it assigns impression v to arg maxu ruv. Note
that it is possible that Au = ∅ for all bidders u, and therefore the algorithm does not assign impression v to any
bidder, even though there are dimensions and bidders where it could have earned revenue. This completes the
description of our algorithm.
B. Algorithm Analysis
We first establish feasibility of the solution.
Lemma 14. If ALGO assigns an impression v to a bidder u, then it can earn revenue on all the active dimensions
Au of u without violating any constraint.
Proof: We need to show that for all constraints s of bidder u, κ
k∈Au∩s
r(k)
uv
B(s)
u ≤ 1. Suppose not. Then,
u +(cid:80)
(s)
for some constraint s,
κ(s)
u + ≥ κ(s)
u +
(cid:88)
k∈Au∩s
(k)
r
uv
(s)
u
B
> 1
(by the definition of )
(cid:32)
1
1−
since κ
(s)
u
1 −
(cid:33)
> 1 from the first line
i.e., (p + 1)
κ
(s)
u +
1− > (p + 1)
i.e., (p + 1)
(s)
u
κ
1− > p + 1
(s)
u
(s)
u
i.e., pφ
B
i.e., φ
B
(s)
u
(s)
u
+ 1 > p + 1
> 1,
which contradicts the fact that dimension k is active for bidder u.
This lemma implies that ALGO is indeed able to earn revenue on all active dimensions of a bidder u when it
assigns an impression to u.
Next, we will bound the total revenue of an optimal solution that we denote by OPT against the total revenue
of ALGO. Let uOPT(v) (resp., uALGO(v)) be the bidder that OPT (resp., ALGO) allocates impression v to. For every
dimension k that OPT earns revenue from, one of the following holds:
1) Case 1: dimension k is active for bidder uOPT(v) in ALGO when impression v arrives, and ALGO assigns v
2) Case 2: dimension k is active for bidder uOPT(v) in ALGO when impression v arrives, but ALGO assigns v
to the same bidder, i.e., uALGO(v) = uOPT(v).
to a different bidder, i.e., uALGO(v) (cid:54)= uOPT(v).
3) Case 3: dimension k is inactive for bidder uOPT(v) in ALGO when impression v arrives.
We partition the dimensions that OPT earns revenue from into active and inactive dimensions (according to
their status in ALGO for bidder uOPT(v)). For active dimensions (cases 1 and 2 above), the next lemma gives
a straightforward charging argument using the greediness of the choice made by ALGO.
Lemma 15. For any impression v, the total revenue earned by OPT on the active dimensions is at most the total
revenue earned by ALGO overall.
Proof: For case 1 above (OPT and ALGO choose the same bidder), the two algorithms earn the same revenue
on the active dimensions. For case 2 above, the fact that ALGO makes a greedy choice implies that it earns at least
as much revenue by assigning to a different bidder as it would have made by assigning to u, which includes the
revenue on all the active dimensions.
The more involved case is that of inactive dimensions (case 3 above). In this case, we use a different global
charging argument over all dimensions, based on the potential function. In particular, we show that for a bidder
u, the total revenue of OPT from inactive dimensions (recall that this only includes revenue from impressions that
arrived after the dimension became inactive in ALGO) can be charged, up to a loss equal to the desired competitive
ratio, to the revenue that ALGO earned overall from bidder u.
Lemma 16. Fix a bidder u. The total revenue that OPT earns in inactive dimensions for bidder u is at most the
final potential of bidder u in ALGO.
Proof: For any dimension k, let Vu,k denote the subset of impressions assigned to u by OPT that arrived after
k became an inactive dimension for bidder u in ALGO. We need to bound the total revenue earned by OPT on
dimension k from impressions in Vu,k, summed over all k. For any impression v ∈ Vu,k, we have:
(cid:88)
φ
(s)
u
(s)
u
s:k∈s
B
> 1,
is the final potential for constraint s of bidder u. Thus, the revenue that OPT earns from impressions
(s)
u
where φ
v ∈ Vu,k on dimension k, summed over all dimensions, can be bounded as follows:
(k)
r
uv
(s)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
r(k)
uv <
φ(s)
u
r(k)
uv
=
v∈Vu,k
k
s:k∈s
B
s
k∈s
v∈Vu,k
B
φ
(s)
u
(s)
u
v∈Vu,k
k
(cid:88)
s
u ≤
φ(s)
u ,
where the last inequality follows by the feasibility of OPT.
Next, we need to lower bound the final potential of a bidder in terms of the revenue that ALGO earns from her.
This is given in the next lemma.
Lemma 17. The increase in potential of a bidder u during the course of ALGO is at most 4· p
that ALGO earns from u.
1−
times the revenue
Proof: Suppose ALGO assigns impression v to bidder u. Let Ka denote the set of active dimensions for bidder
. The increase in potential for bidder u is given as
(s)
k:k∈Ka∩s
r(k)
uv
B(s)
u
u when this assignment is made. Let η
follows:
u = (cid:80)
where the inequality follows since η
rearranging the RHS of inequality (13) and using the fact that {s : k ∈ s} ≤ p, we obtain:
u ≤ and ax ≤ 1 + ax for 0 ≤ x ≤ 1, a ≥ 1. By
k:k∈Ka∩s
r(k)
uv
B(s)
,
(13)
(cid:88)
s
∆φ(s)
u =
(cid:18)
s
s
s
=
=
φ(s)
u +
(s)
B
u
p
(s)
B
u
p
(cid:88)
(cid:88)
(cid:32)
(cid:88)
(cid:32)
(cid:88)
u = (cid:80)
(cid:32)(cid:88)
(cid:88)
(cid:88)
r(k)
uv ·
φ(s)
u +
≤
(s)
s
r(k)
uv
k∈Ka
1−
(cid:19)
(s)
κ
u
1−
(cid:33)
(cid:19)
(cid:17) η
− 1
(cid:88)
(k)
r
uv
(s)
B
u
(s)
u
k:k∈Ka∩s
(cid:33)
(p + 1)
(s)
u +η
(s)
u
κ
1− − (p + 1)
(p + 1)
(p + 1)
(s)
u
η
1− − 1
(cid:18)
(cid:32)(cid:16)
·
(s)
κ
u
1−
(cid:33)
(cid:33)
(s)
B
u
p
(s)
B
u
p
(p + 1)
1−
(p + 1)
1−
·
·
φ
(s)
u
(s)
u
s:k∈s
B
+ {s : k ∈ s}
p
(since k is active, and p ≥ {s : k ∈ s})
(cid:16)
since (p + 1)
1− ≤ 2 · p
1− for large enough p
(cid:17)
,
(cid:88)
s
∆φ(s)
u ≤
(p + 1)
1−
(p + 1)
≤ 2 ·
k∈Ka
r(k)
uv
(cid:88)
(cid:17)
k∈Ka
1−
1−
(cid:16) p
p
≤ 4 ·
as desired.
A competitive ratio of O
C. Extension to ADGEN-P
for ADGEN-AON now follows from Lemmas 15, 16, and 17.
(cid:80)
We begin by noting that for ADGEN-P, we will assume that the maximum bid-to-budget ratio maxu,v,s
<
1. Obviously, this assumption is wlog for ADGEN-AON since any u, v pair that results in > 1 cannot be assigned
(and it is easy to show an arbitrarily large lower bound when = 1). For ADGEN-P, however, allowing instances
where > 1 still admits a nontrivial problem definition since the algorithm can choose to earn partial revenues.
However, since such a scenario would clearly never arise in practice (i.e, an impression generating more revenue
than a budget) and only complicates the analysis, we proceed with this added assumption.
u
r(k)
uv
k∈Ks
B(s)
To adapt our ADGEN-AON algorithm and analysis for ADGEN-P, the algorithm will now choose to earn
1
lg(2p+2)
1
fraction of all revenues and set the parameter =
lg(2p+2) in algorithm and proof (so essentially the algorithm
treats the instance as if its a ADGENERAL small-bids instance). Otherwise, the algorithm behaves identically as
before. Since we are assuming the maximum bid-to-budget ratio is at most 1, this scaling procedure ensures that the
lg(2p+2) factor.
revenue generated by an impression never increases the utilization of a constraint by more than a
To show a O(lg p) competitive ratio, it suffices to show that Lemmas 14 through 17 still hold in this setting. It
is not too hard to verify that Lemmas 14, 16, and 17 follow by the same arguments, noting that in these proofs,
(k)
lg(2p+2) factor (except in Lemma
uv still denotes the revenue earned by the algorithm after its been reduced by a
r
(k)
uv denotes the amount of revenue the optimal solution chooses to earn). Lemma 15 uses the same argument,
16, r
lg(2p+2) factor of that earned by the
except now the greedy property implies that the algorithm earns at least a
optimal solution on a active dimension (instead of strictly more); however, losing this factor in this case is fine
since the we are ultimately aiming for a O(lg p) competitive ratio. Hence, our algorithm extends to the ADGEN-P
setting.
1
1
1
|
1609.00810 | 1 | 1609 | 2016-09-03T10:14:59 | Greedy MAXCUT Algorithms and their Information Content | [
"cs.DS",
"cs.DM",
"cs.IT",
"cs.IT"
] | MAXCUT defines a classical NP-hard problem for graph partitioning and it serves as a typical case of the symmetric non-monotone Unconstrained Submodular Maximization (USM) problem. Applications of MAXCUT are abundant in machine learning, computer vision and statistical physics. Greedy algorithms to approximately solve MAXCUT rely on greedy vertex labelling or on an edge contraction strategy. These algorithms have been studied by measuring their approximation ratios in the worst case setting but very little is known to characterize their robustness to noise contaminations of the input data in the average case. Adapting the framework of Approximation Set Coding, we present a method to exactly measure the cardinality of the algorithmic approximation sets of five greedy MAXCUT algorithms. Their information contents are explored for graph instances generated by two different noise models: the edge reversal model and Gaussian edge weights model. The results provide insights into the robustness of different greedy heuristics and techniques for MAXCUT, which can be used for algorithm design of general USM problems. | cs.DS | cs | Greedy MAXCUT Algorithms
and their Information Content
Yatao Bian, Alexey Gronskiy and Joachim M. Buhmann
Department of Computer Science, ETH Zurich
{ybian, alexeygr, jbuhmann}@inf.ethz.ch
6
1
0
2
p
e
S
3
]
S
D
.
s
c
[
1
v
0
1
8
0
0
.
9
0
6
1
:
v
i
X
r
a
Abstract-MAXCUT defines a classical NP-hard problem for
graph partitioning and it serves as a typical case of the symmetric
non-monotone Unconstrained Submodular Maximization (USM)
problem. Applications of MAXCUT are abundant in machine
learning, computer vision and statistical physics. Greedy algo-
rithms to approximately solve MAXCUT rely on greedy vertex
labelling or on an edge contraction strategy. These algorithms
have been studied by measuring their approximation ratios in
the worst case setting but very little is known to characterize
their robustness to noise contaminations of the input data in
the average case. Adapting the framework of Approximation Set
Coding, we present a method to exactly measure the cardinality
of the algorithmic approximation sets of five greedy MAXCUT
algorithms. Their information contents are explored for graph
instances generated by two different noise models: the edge
reversal model and Gaussian edge weights model. The results
provide insights into the robustness of different greedy heuristics
and techniques for MAXCUT, which can be used for algorithm
design of general USM problems.
I. INTRODUCTION
Algorithms are mostly analyzed by measuring their run-
time and memory consumption for the worst possible input
instance. In many application scenarios, algorithms are also
selected according to their "robustness" to noise perturbations
of the input instance and their insensitivity to randomization
during algorithm execution. How should this "robustness"
property be measured? Machine learning requires that algo-
rithms with random variables as input generalize over these
fluctuations. The algorithmic answer has to be stable w.r.t.
this uncertainty in the input
instance. Approximation Set
Coding (ASC) quantifies the impact of input randomness
on the solution space of an algorithm by measuring the
attainable resolution for the algorithm's output. We employ
this framework in an exemplary way by estimating the ro-
bustness of MAXCUT algorithms to specific input instances.
Thereby, we effectively perform an average case analysis of
the generalization properties of MAXCUT algorithms.
such that the cut value cut(S1, S2) := (cid:80)
A. MAXCUT and Unconstrained Submodular Maximization
Given an undirected graph G = (V, E) with vertex set
V = {v1, v2,··· , vn} and edge set E with nonnegative
weights wij,∀(i, j) ∈ E, the MAXCUT problem aims to find
a partition of vertices into two disjoint subsets S1 and S2,
wij is
maximized. MAXCUT is emlpoyed in various applications,
such as in semisupervised learning ([1]), in social network
([2]), in statistical physics and in circuit layout design ([3]).
(cid:80)
i∈S1
j∈S2
MAXCUT is considered to be a typical case of the USM
problem because its objective can be formulated as a set
function: f (S) := cut(S, V \S), S ⊆ V , which is submodular,
nonmonotone, and symmetric (f (S) = f (V \S)). Beside
MAXCUT, USM captures many practical problems such as
MAXDICUT ([4]), variants of MAXSAT and the maximum
facility location problem ([5], [6]).
B. Greedy Heuristics and Techniques
The five algorithms investigated here (as summarized in
Table I) belong to two greedy heuristics: double greedy and
backward greedy. The double greedy algorithms exploit the
symmetric property of USM, and conducts classical forward
greedy and backward greedy simultaneously: it works on two
solutions initialized as ∅ and the ground set V , respectively,
then processes the elements (vertices for MAXCUT problem)
one at a time, for which it determines whether it should
be added to the first solution or removed from the second
solution. The backward greedy algorithm removes the smallest
weighted edge in each step. The difference of the four double
greedy algorithms lies in the greedy techniques they use:
sorting, randomization and the way to initialize the first two
vertices.
C. Approximation Set Coding for Algorithm Analysis
In analogy to Shannon's theory of communication, the ASC
framework ([7], [8], [9]) determines distinguishable sets of
solutions and, thereby, provides a general principle to conduct
model validation ([10], [11]). As an algorithmic variant of
the ASC framework, [12], [13] defines the algorithmic t-
approximation set of an algorithm A at step t as the set
of feasible solutions after t steps, CA
t (G) := At(G), where
At(G) is the solution set which are still considered as viable
by A after t computational steps.
ASC utilizes the two instance-scenario to investigate the
information content of greedy MAXCUT algorithms. Since we
investigate the average case behavior of algorithms, we have to
specify the probability distribution of the input instances. We
generate graph instances in a two step process. First, generate
a "master graph" G, e.g., a complete graph with Gaussian
distributed weights. In a second step, we generate two input
graphs G(cid:48), G(cid:48)(cid:48) by independently applying a noise process to
edge weights of the master graph G.
TABLE I: Summary of Greedy MAXCUT Algorithms
Name
D2Greedy
RDGreedy
SG
SG3
EC
Greedy
Heuristic
Double
Backward
Techniques
Sort. Rand.
Init. Vertices
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
:
t
t
(cid:17)(cid:105)
E(cid:104)
I A := max
I A
t = max
The algorithmic analogy of information content ([7]), i.e.
algorithmic information content I A , is computed as the max-
imum stepwise information I A
t
(cid:16) C ∆A
t (G(cid:48), G(cid:48)(cid:48))
(1)
t (G(cid:48)(cid:48))
t (G(cid:48))CA
t (G(cid:48), G(cid:48)(cid:48)) :=
The expectation is taken w.r.t. (G(cid:48), G(cid:48)(cid:48)); ∆A
t (G(cid:48)(cid:48)) denotes the intersection of approximation
CA
sets, and C is the solution space,
i.e., all possible cuts.
The information content I A
t measures how much information
is extracted by algorithm A at iteration t from the input
distribution that is relevant to the output distribution.
t (G(cid:48))∩CA
log
CA
II. GREEDY MAXCUT ALGORITHMS
We investigate five greedy algorithms (Table I) for MAX-
CUT. According to the type of greedy heuristic, they can
be divided into two categories: I) Double Greedy: SG, SG3,
D2Greedy, RDGreedy; II) Backward Greedy: Edge Contrac-
tion. Besides the type of greedy heuristic,
the difference
between the algorithms are mainly in three techniques: sorting
the candidate elements, randomization and the way initializing
the first two vertices. In the following, we briefly introduce one
typical algorithm in each category and we present the others
by showing the difference (details are in the Supplement VI-A
because of space limit).
A. Double Greedy Algorithms
D2Greedy (Alg. 1) is the Deterministic double greedy,
RDGreedy is the Randomized double greedy,
they were
proposed by [14] to solve the general USM problem with
1/3 and 1/2 worst-case approximation guarantee, respectively.
They use the same double greedy heuristic as SG ([15]) and
SG3 (variant of SG), which are classical greedy MAXCUT
algorithms. We prove in Supplement VI-B that, for MAXCUT,
SG and D2Greedy use equivalent labelling criteria except for
initializing the first two vertices.
As shown in Alg. 1, D2Greedy maintains two solution sets:
S initialized as ∅, T initialized as the ground set V . It labels
all the vertices one by one: for vertex vi, it computes the
objective gain of adding vi to S and the gain of removing vi
from T , then labels vi to have higher objective gain.
SG and D2Greedy differ in the initialization of the first
two vertices: SG picks first of all the maximum weighted
edge and distributes its two vertices to the two active subsets.
Compared to D2Greedy, the RDGreedy uses randomization
Algorithm 1: D2Greedy ([14])
Input: Complete graph G = (V, E) with nonnegative edges
Output: A disjoint cut and the cut value
1 S0 := ∅, T 0 := V ;
2 for i = 1 to n do
3
4
5
6
ai := f (Si−1 ∪ {vi}) − f (Si−1);
bi := f (T i−1\{vi}) − f (T i−1);
if ai ≥ bi then
Si := Si−1 ∪ {vi}, T i := T i−1 ;
else
7
Si := Si−1, T i := T i−1\{vi} ;
8
9 return Sn, V \Sn, and cut(Sn, V \Sn)
//expand S
//shrink T
technique when labelling each vertex: it labels each vertex
with probability proportional to the objective gain. Compared
to SG, SG3 sorts the unlabelled vertices according to a certain
score function (which is proportional to the possible objective
gains), and selects the vertex with the maximum score to be
the next one to be labelled.
B. Edge Contraction (EC)
EC ([16], Alg. 2) contracts the smallest edge in each step.
The two vertices of this contracted edge become one "super"
vertex, and the weight of an edge connecting this super vertex
to any other vertex is assigned as the sum of weights of the
original two edges. EC belongs to the backward greedy in the
sense that it tries to remove the least expensive edge from the
cut set in each step. We can easily derive a heuristic for the
MAX-K-CUT problem by using n − k steps instead of n − 2
steps.
Algorithm 2: Edge Contraction (EC) ([16])
Input: Complete graph G = (V, E) with nonnegative edge
Output: A disjoint cut S1, S2 and cut value cut(S1, S2)
ContractionList(i) := {i};
1 for i = 1 : n do
2
3 for i = 1 : n − 2 do
4
5
6
7
Find a minimum weight edge (x, y) in G;
v := contract(x, y), V := V ∪ {v}\{x, y} ; //contract
for j ∈ V \{v} do
8
wvj := wxj + wyj;
ContractionList(v) :=
ContractionList(x) ∪ ContractionList(y);
9 Denote by x and y the only 2 vertices in V ;
10 return S1 := ContractionList(x),
S2 := ContractionList(y), cut(S1, S2) := wxy
III. COUNTING SOLUTIONS IN APPROXIMATION SETS
To compute the information content according to Eq. 1,
we need to exactly compute the cardinalities of four different
solution sets. For MAXCUT problem, the solution space has
the cardinality C = 2n−1−1. In the following we will present
guaranteed methods for exact counting CA
t (G(cid:48)(cid:48))
t (G(cid:48), G(cid:48)(cid:48)) (sub-/superscripts omitted for notational
and ∆A
clarity).
t (G(cid:48)),CA
A. Counting Methods for Double Greedy Algorithms
The counting methods for the double greedy algorithms are
similar, so we only discuss the method for SG3 here; details
about other methods and the corresponding proofs are in the
Supplement VI-C and VI-D, respectively.
For the SG3 (Alg. 6, see Supplement), after step t (t =
1,··· , n − 1) there are k = n − t − 1 unlabelled vertices, and
it is clear that C(G(cid:48)) = C(G(cid:48)(cid:48)) = 2k.
To count the intersection set ∆(G(cid:48), G(cid:48)(cid:48)), assume the solution
set pair of G(cid:48)
2), the solution set pair of G(cid:48)(cid:48)
is (S(cid:48)
is
2},
1∪ S(cid:48)
1 , S(cid:48)(cid:48)
(S(cid:48)(cid:48)
2}, respectively. Denote L := T (cid:48) ∩ T (cid:48)(cid:48) be
T (cid:48)(cid:48) = V \{S(cid:48)(cid:48)
the common vertices of the two unlabelled vertex sets, so
l = L (0 ≤ l ≤ k) is the number of common vertices in
the unlabelled k vertices. Denote M(cid:48) := T (cid:48)\L, M(cid:48)(cid:48) := T (cid:48)(cid:48)\L
be the sets of different vertex sets between the two unlabelled
vertex sets. Then,
2 ), so the unlabelled vertex sets are T (cid:48) = V \{S(cid:48)
1 ∪ S(cid:48)(cid:48)
1, S(cid:48)
∆(G(cid:48), G(cid:48)(cid:48)) =
1\M(cid:48), S(cid:48)(cid:48)
1\M(cid:48)(cid:48), S(cid:48)
if (S(cid:48)(cid:48)
(S(cid:48)
otherwise
2\M(cid:48)) is matched by
2\M(cid:48)(cid:48)) or (S(cid:48)
2\M(cid:48)(cid:48), S(cid:48)
1\M(cid:48)(cid:48))
2l
0
B. Counting Method for Edge Contraction Algorithm
For EC (Alg. 2), after step t (t = 1,··· , n − 2) there
are k = n − t "super" vertices (i.e. contracted ones). It is
straightforward to see that C(G(cid:48)) = C(G(cid:48)(cid:48)) = 2k−1 − 1.
To count the intersection ∆(G(cid:48), G(cid:48)(cid:48)), suppose there are l
(0 ≤ l ≤ k) common super vertices in the unlabelled k ver-
tices. Remove the l common super vertices from each set, then
there are h = k − l distinct super vertices in each set, denote
them by P := {p1, p2,··· , ph}, Q := {q1, q2,··· , qh},
respectively. Notice that p1∪p2∪···∪ph = q1∪q2∪···∪qh,
so after some contractions in both P and Q, there must be
some common super vertices between P and Q. Assume the
maximum number of common super vertices after all possible
contractions is c∗, then it holds
∆(G(cid:48), G(cid:48)(cid:48)) = 2c∗+l−1 − 1 .
(2)
To compute c∗, we propose a polynomial time algorithm (Alg.
3) with a theoretical guarantee in Theorem 1 (for the proof see
Supplement VI-E). The algorithm finds the maximal number
of common super vertices after all possible contractions, that
is used to count ∆(G(cid:48), G(cid:48)(cid:48)) for EC.
Theorem 1. Given two distinct super vertex sets P :=
{p1, p2,··· , ph}, Q := {q1, q2,··· , qh} (any 2 super ver-
tices inside P or Q do not intersect, and there is no common
super vertex between P and Q), such that p1∪p2∪···∪ph =
q1 ∪ q2 ∪ ··· ∪ qh, Alg. 3 returns the maximum number of
common super vertices between P and Q after all possible
contractions.
IV. EXPERIMENTS
We conducted experiments on two exemplary models: the
edge reversal model and the Gaussian edge weights model.
Each model involves the master graph G and a noise type used
to generate the two noisy instances G(cid:48) and G(cid:48)(cid:48). The width of
Algorithm 3: Common Super Vertex Counting
Input: Two distinct super vertex sets P , Q
Output: Maximum number of common super vertices after all
possible contractions
1 c := 0;
2 while P (cid:54)= ∅ do
3
4
5
6
7
Randomly pick pi ∈ P ;
Find qj ∈ Q s.t. pi ∩ qj (cid:54)= ∅;
if qj\pi (cid:54)= ∅ then
if pi\qj (cid:54)= ∅ then
For pi, find pi(cid:48) ∈ P\{pi} s.t. pi(cid:48) ∩ (qj\pi) (cid:54)= ∅;
pii(cid:48) := pi ∪ pi(cid:48), P := P ∪ {pii(cid:48)}\{pi, pi(cid:48)} ;
For qj, find qj(cid:48) ∈ Q\{qj} s.t. qj(cid:48) ∩ (pi\qj) (cid:54)= ∅;
qjj(cid:48) := qj ∪ qj(cid:48), Q := Q ∪ {qjj(cid:48)}\{qj, qj(cid:48)} ;
if pii(cid:48) == qjj(cid:48) then
Remove pii(cid:48), qjj(cid:48) from P , Q, respectively;
c := c + 1;
8
9
10
11
12
13
14 return c
the instance distribution is controlled by the strength of the
noise model. These models provide the setting to investigate
the algorithmic behavior.
A. Experimental Setting
Edge Reversal Model: To obtain the master graph, we
generate a balanced bipartite graph Gb with disjoint vertex
sets S1, S2. Then we assign uniformly distributed weights in
n2 ] to all edges inside S1 or S2 and we assign uniformly
[0, 8
distributed weights in [1− 8
n2 , 1] to all edges between S1 and
S2, thus generating graph G(cid:48)
b. Then randomly flip edges in
G(cid:48)
b to generate the master graph G. Here, flip edge eij means
changing its weight wij to 1 − wij with probability pm, and
(f lip eij) ∼ Ber(pm); pm = 0.2 is used to generate the
master graph G. Noisy graphs G(cid:48), G(cid:48)(cid:48) are generated by flipping
the edges in G with probability p, ((f lip eij) ∼ Ber(p)).
Gaussian Edge Weights Model: The master graph G
is generated with Gaussian distributed edge weights wij ∼
m), µ = 600, σm = 50, negative edges are set to be
N (µ, σ2
µ. Noisy graphs G(cid:48), G(cid:48)(cid:48) are obtained by adding Gaussian
distributed noise nij ∼ N (0, σ2), negative noisy edges are
set to be 0.
For both noise models, we conducted 1000 experiments on
i.i.d. generated noisy graphs G(cid:48) and G(cid:48)(cid:48), and then we aggregate
the results to estimate the expectation in Eq. 1.
B. Results
We plot the information content and stepwise information
per node in Fig. 1 and 2, respectively. For the edge reversal
model, we also investigate the number of equal edge pairs be-
tween G(cid:48) and G(cid:48)(cid:48): d = 0,··· , m (m is the total edge number),
d measures the consistency of the two noisy instances. The
expected fraction of equal edge pairs is Ed = p2 + (1 − p)2,
and it is plotted as the dashed magenta line in Fig. 1(a).
C. Analysis
Before discussing these results, let us revisit the stepwise in-
formation and information content. From the counting methods
(a) Edge Reversal Model, n: 100
(b) Gaussian Edge Weights Model, n: 100
Fig. 1: Information Content per Node
t (G(cid:48))
t (G(cid:48)(cid:48)) (e.g., A = SG3), an we insert these values
in Section III, we derive the analytical form of C, CA
and CA
into the definition of stepwise information,
(cid:16)C ∆A
(cid:17)
CA
t (G(cid:48)(cid:48))
t (G(cid:48))CA
t (G(cid:48), G(cid:48)(cid:48))) − log(CA
t (G(cid:48)(cid:48))))
t (G(cid:48), G(cid:48)(cid:48)) + 2t + log(2n−1 − 1) − 2(n − 1)
t = E log
I A
= E(log(C∆A
= E log ∆A
t (G(cid:48), G(cid:48)(cid:48))
t (G(cid:48))CA
(3)
The information content
is computed as
the maxi-
mum stepwise information I A := maxt I A
. Notice that
t
t (G(cid:48), G(cid:48)(cid:48)) measures the ability of A to find common
log ∆A
solutions for the two noisy instances G(cid:48), G(cid:48)(cid:48), given the under-
lying input graph G.
t (G(cid:48), G(cid:48)(cid:48)) = log CA
in the noise free limit (G(cid:48) = G(cid:48)(cid:48)),
Our results support the following observations and analysis:
• All
investigated algorithms reach the maximum infor-
mation content
i.e.,
for p = 0, 1 in the edge reversal model and for σ = 0
in the Gaussian edge weights model. In this circumstance,
t (G(cid:48)) = n − t − 1, so I A
E log ∆A
t =
t+log(2n−1−1)−(n−1), and the information content reaches
its maximum log(2n−1 − 1) at the final step t = n − 1.
• Fig. 1(a) demonstrates that the information content qual-
itatively agrees with the consistency between two noisy
instances (the dashed magenta line), which reflects that
log ∆A
• Stepwise information (Fig. 2) of the algorithms increase
initially, but after reaching the optimal step t∗ (the step with
highest information), it decreases and finally vanishes.
• For the greedy heuristics, backward greedy is more infor-
mative than double greedy under both models. EC (backward
greedy) achieves the highest information content. We explain
this behavior by delayed decision making of the backward
greedy edge contraction. With high probability it preserves
consistent solutions by contracting low weight edges that
t (G(cid:48), G(cid:48)(cid:48)) is affected by the noisy instances.
have a low probability to be included in the cut. The same
phenomena arises for the reverse-delete algorithm to calculate
the minimum spanning tree of a graph (see [13]).
• The information content of the four double greedy al-
gorithms achieve different rank orders for the two models.
SG3 is inferior to other double greedy algorithms under
Gaussian edge weights model, but this only happens when
p ∈ [0.2, 0.87] for the edge reversal model. This results from
that information content of one specific algorithm is affected
by both the input master graph G and the noisy instances
G(cid:48), G(cid:48)(cid:48), which are completely different under the two models.
• Different greedy techniques cast different influences on
the information content. The four double greedy algorithms
differ by the techniques they use (Table
I). (1) The ran-
domization technique makes RDGreedy very fragile w.r.t.
information content, though it improves the worst-case ap-
proximation guarantee for the general USM problem ([14]).
RDGreedy labels each vertex with a probability proportional
to the objective gain, this randomization makes the consistency
between CA
t (G(cid:48)(cid:48)) very weak, resulting in
t (G(cid:48), G(cid:48)(cid:48)). (2) The
small approximation set intersection log ∆A
initializing strategy for the first 2 vertices as used in SG
decreases the information content (SG is outperformed by
D2Greedy under both models) due to early decision making.
(3) The situation is similar for the sorting techniquey used in
SG3 under Gaussian edge weights model, it is outperformed
by both SG and D2Greedy. But for the edge reversal model,
this observation only holds when p ∈ [0.2, 0.87].
t (G(cid:48)) and CA
• SG and D2Greedy behave very similar under both models,
which is caused by an equivalent processing sequence apart
from initializing of the first two vertices (proved in Supplement
VI-B).
00.10.20.30.40.50.60.70.80.9110−310−210−1100Noise [p]Information Content per Node [bits] SG3ECSGD2GreedyR2Greedy SG3 EC SG D2Greedy RDGreedy 02040608010012014016018020010−310−210−1100Noise [std. dev.]Information Content per Node [bits] SG3ECSGD2GreedyR2Greedy SG3 EC SG D2Greedy RDGreedy (a) Edge Reversal Model, n: 100, p = 0.65
(b) Gaussian Edge Weights Model, n: 100, σ = 125
Fig. 2: Stepwise Information per Node
V. DISCUSSION AND CONCLUSION
REFERENCES
This work advocates an information theoretically guided
average case analysis of the generalization ability of greedy
MAXCUT algorithms. We have contributed to the foundation
of approximation set coding by presenting provably correct
methods to exactly compute the cardinality of approximation
sets. The counting algorithms for approximate solutions enable
us to explore the information content of greedy MAXCUT al-
gorithms. Based on the observations and analysis, we propose
the following conjecture:
• Different greedy heuristics (backward, double) and differ-
ent processing techniques (sorting, randomization, intilization)
sensitively influence the information content. The backward
greedy with its delayed decision making consistently outper-
forms the double greedy strategies for different noise models
and noise levels.
Since EC demonstrated to achieve the highest robustness,
it is valuable to develop the corresponding algorithm for the
general USM.
In this work ASC has been employed as a descriptive
tool to compare algorithms. We could also use the method
for algorithm design. A meta-algorithm modifies the algo-
rithmic steps of a MAXCUT procedure and measures the
resulting change in information content. Beneficial changes
are accepted and detrimental changes are rejected. It is also
imaginable that design principles like delayed decision making
are systematically identified and then combined to improve the
informativeness of novel algorithms.
ACKNOWLEDGMENT
This work was partially supported by SNF Grant # 200021 138117.
The authors would like to thank Andreas Krause, Mat´us Mihal´ak and
Peter Widmayer for valuable discussions.
[1] J. Wang, T. Jebara, and S.-F. Chang, "Semi-supervised learning using
greedy max-cut," JMLR, vol. 14, no. 1, pp. 771–800, Mar. 2013.
[2] R. Agrawal, S. Rajagopalan, R. Srikant, and Y. Xu, "Mining newsgroups
using networks arising from social behavior," in WWW. ACM, 2003,
pp. 529–535.
[3] F. Barahona, M. Grotschel, M. Junger, and G. Reinelt, "An application
of combinatorial optimization to statistical physics and circuit layout
design," Operations Research, vol. 36, no. 3, pp. 493–513, 1988.
[4] E. Halperin and U. Zwick, "Combinatorial approximation algorithms
for the maximum directed cut problem," in ACM-SIAM symposium on
Discrete algorithms, 2001, pp. 1–7.
[5] A. A. Ageev and M. Sviridenko, "An 0.828-approximation algorithm
for the uncapacitated facility location problem," Discrete Applied Math-
ematics, vol. 93, no. 2, pp. 149–156, 1999.
[6] G. Cornuejols, M. Fisher, and G. L. Nemhauser, "On the uncapacitated
location problem," Annals Discr. Math., vol. 1, pp. 163–177, 1977.
[7] J. M. Buhmann, "Information theoretic model validation for clustering."
in ISIT, 2010, pp. 1398–1402.
[8] J. Buhmann, "Context sensitive information: Model validation by in-
formation theory," in Pattern Recognition, ser. LNCS 6718. Springer
Berlin / Heidelberg, 2011, pp. 12–21.
[9] J. M. Buhmann, "Simbad: Emergence of pattern similarity," in
Springer, 2013,
Similarity-Based Pattern Analysis and Recognition.
pp. 45–64.
[10] M. H. Chehreghani, A. G. Busetto, and J. M. Buhmann, "Information
theoretic model validation for spectral clustering." in AISTATS, 2012,
pp. 495–503.
[11] G. Zhou, S. Geman, and J. M. Buhmann, "Sparse feature selection by
information theory," in ISIT.
IEEE, 2014, pp. 926–930.
[12] L. M. Busse, M. H. Chehreghani, and J. M. Buhmann, "The information
IEEE, 2012, pp. 2746–2750.
[13] A. Gronskiy and J. M. Buhmann, "How informative are minimum
content in sorting algorithms," in ISIT.
spanning tree algorithms?" in ISIT.
IEEE, 2014, pp. 2277 – 2281.
[14] N. Buchbinder, M. Feldman, J. Naor, and R. Schwartz, "A tight linear
time (1/2)-approximation for unconstrained submodular maximization,"
in FOCS.
IEEE, 2012, pp. 649–658.
[15] S. Sahni and T. Gonzalez, "P-complete approximation problems," Jour-
nal of the ACM (JACM), vol. 23, no. 3, pp. 555–565, 1976.
[16] S. Kahruman, E. Kolotoglu, S. Butenko, and I. V. Hicks, "On greedy
construction heuristics for the max-cut problem," Int. J. Comput. Science
and Engineering, vol. 3, no. 3, pp. 211–218, 2007.
10010110210−410−310−210−1100StepStepwise Information per Node (p = 0.65) SG3ECSGD2GreedyR2Greedy SG3 EC SG D2Greedy RDGreedy 10010110210−510−410−310−210−1StepStepwise Information per Node (sigma = 125) SG3ECSGD2GreedyR2Greedy SG3 EC SG D2Greedy RDGreedy VI. SUPPLEMENTARY MATERIAL
A. Details of Double Greedy Algorithms
Algorithm 4: SG ([15])
Input: A complete graph G = (V, E) with nonnegative edge
weights wij,∀i, j ∈ V, i (cid:54)= j
Output: A disjoint cut and the cut value
1 Pick the maximum weighted edge (x, y);
2 S1 := {x},S2 := {y}, cut(S1, S2) := wxy;
3 for i = 1 : n − 2 do
4
//w(i, Sk) :=(cid:80)
If w(i, S1) > w(i, S2), then add i to S2, else add it to S1;
cut(S1, S2) := cut(S1, S2) + max{w(i, S1), w(i, S2)};
wij , k = 1, 2
j∈Sk
5
6 return S1, S2, and cut(S1, S2)
Algorithm 5: RDGreedy ([14])
Input: A complete graph G = (V, E) with nonnegative edge
weights wij,∀i, j ∈ V, i (cid:54)= j
Output: A disjoint cut and the cut value
1 S0 := ∅, T 0 := V ;
2 for i = 1 to n do
3
4
5
ai := f (Si−1 ∪ {vi}) − f (Si−1);
bi := f (T i−1\{vi}) − f (T i−1);
i := max{bi, 0};
i := max{ai, 0}, b(cid:48)
a(cid:48)
a(cid:48)
With probability
i+b(cid:48)
a(cid:48)
T i := T i−1//If a(cid:48)
i = b(cid:48)
i
i
do: Si := Si−1 ∪ {vi},
a(cid:48)
i = 0, assume
i+b(cid:48)
a(cid:48)
b(cid:48)
) do:
a(cid:48)
i+b(cid:48)
i
i
i
i
= 1
Else (with the compliment probability
Si := Si−1, T i := T i−1\{vi};
8 return 2 subsets: Sn, V \Sn, and cut(Sn, V \Sn)
6
7
6
7
Algorithm 6: SG3
Input: A complete graph G = (V, E) with nonnegative edge
weights wij,∀i, j ∈ V, i (cid:54)= j
Output: A disjoint cut S1, S2 and the cut value cut(S1, S2)
1 Pick the maximum weighted edge (x, y);
2 S1 := {x},S2 := {y},V := V \{x, y}, cut(S1, S2) := wxy;
3 for i = 1 : n − 2 do
4
5
score(j) := w(j, S1) − w(j, S2) ;
for j ∈ V do
//
w(j, Sk) :=(cid:80)
j(cid:48)∈Sk
wjj(cid:48) , k = 1, 2
Choose the vertex j∗ with the maximum score;
If w(j∗, S1) > w(j∗, S2), then add j∗ to S2, else add it to
S1;
V := V \{j∗};
cut(S1, S2) := cut(S1, S2) + max{w(j∗, S1), w(j∗, S2)};
8
9
10 return S1, S2, and cut(S1, S2)
B. Equivalence Between Labelling Criterions of SG and
D2Greedy
Claim:
for processing the first 2 vertices,
D2Greedy and SG conduct the same labelling strategy for each
vertices.
Except
Proof. To verify this, assume in the beginning of a certain
step i, the solution set pair of SG is (S1, S2), of D2Greedy is
(S, T ) (for simplicity omit the step index here).
D2Greedy is: S1 ↔ S and S2 ↔ (V \T ).
Note that the relationship between solution sets of SG and
For SG, the labelling criterion for vertex i is:
w(i, S2) − w(i, S1) =
wij
(4)
(cid:88)
wij − (cid:88)
i,j∈S2
i,j∈S1
For D2Greedy, the labelling criterion for vertex i is:
ai − bi = [f (S ∪ {vi}) − f (S)] − [f (T\{vi}) − f (T )]
−
(5)
wij
wij
i∈S,j∈V \S
i∈T,j∈V \T
=
=
=
wij
i∈S,j=i
i,j∈V \S\{vi}
i∈S∪{vi},j∈V \S\{vi}
i∈T\{vi},j∈V \T∪{vi}
(cid:88)
(cid:88)
(cid:88)
wij − (cid:88)
(cid:88)
wij − (cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
wij − (cid:88)
wij − (cid:88)
−
wij − (cid:88)
wij − (cid:88)
wij − (cid:88)
wij − (cid:88)
i,j∈(V \T )∪(T\S\{vi})
i,j∈(S)∪(T\S\{vi})
i∈T\{vi},j=i
i,j∈V \T
i,j∈V \T
i,j∈V \T
i,j∈S
i,j∈S
wij
wij
wij
wij
= 2
= 2
i,j∈S2
i,j∈S1
= 2[w(i, S2) − w(i, S1)]
wij
−
(6)
(7)
(8)
where Eq. 8 comes from the relationship between solution sets
of SG and D2Greedy.
So the labelling criterion for SG and D2Greedy is equivalent
with each other.
C. Counting Methods for Double Greedy Algorithms
D2Greedy: summarized in Alg. 1, we have proved that
it has the same labelling criterion with SG, the relationship
between solution sets of SG and D2Greedy is: S1 ↔ S and
S2 ↔ (V \T ), we will use S1 and S2 in the description of its
counting methods.
In step t (t = 1,··· , n) there are k = n − t unlabelled
vertices, it is not difficult to know that the number of possible
solutions for each instance is
C(G(cid:48)) = C(G(cid:48)(cid:48)) =
2k − 1
if S1 (cid:54)= ∅ and S2 (cid:54)= ∅
otherwise
(cid:26) 2k
(cid:54)= pii(cid:48)). That means, there is only one
such that pii(cid:48)(cid:48)
possible way to construct the removed common super
vertex.
We will use inductive assumption to prove the claim. First
of all, in the beginning (step 0), the conditions hold. Assume
the conditions hold in step t. In step t + 1, there are 2 possible
situations:
• There are no common super vertex removed.
Condition 1 holds because the contracted super vertices
pair do not equal. Condition 2, 3 hold as well because
there are no contracted super vertices removed.
• There are common super vertex removed.
Condition 1 holds because the only common super
vertices pair have been removed from P, Q, respectively.
To prove condition 2, notice that the smaller vertices for
pii(cid:48) are pii(cid:48)\pi = pi(cid:48) and pii(cid:48)\p(cid:48)
i = pi, respectively, for
qjj(cid:48) are qjj(cid:48)\qj = qj(cid:48) and qjj(cid:48)\q(cid:48)
j = qj, according to
Condition 1, they can not be common super vertices, so
there are no smaller common super vertices.
To prove condition 3, assume there exists pii(cid:48)(cid:48) = qjj(cid:48)(cid:48),
such that pii(cid:48)(cid:48) (cid:54)= pii(cid:48) (respectively, qjj(cid:48)(cid:48) (cid:54)= qjj(cid:48)), so pi(cid:48)(cid:48) (cid:54)=
pi(cid:48) (pj(cid:48)(cid:48) (cid:54)= pj(cid:48)). From Alg. 3 we know that pi ∪ pi(cid:48)(cid:48) =
pii(cid:48)(cid:48) ⊇ qj\pi and pi ∪ pi(cid:48) = pii(cid:48) ⊇ qj\pi (respectively,
qj∪qj(cid:48)(cid:48) = qjj(cid:48)(cid:48) ⊇ pi\qj and qj∪qj(cid:48) = qjj(cid:48) ⊇ pi\qj), so
that pi(cid:48)(cid:48) ⊇ qj\pi and pi(cid:48) ⊇ qj\pi (respectively, qj(cid:48)(cid:48) ⊇
pi\qj and qj(cid:48) ⊇ pi\qj), that contradicts the known truth
that pi(cid:48) and pi(cid:48)(cid:48) (respectively, qj(cid:48) and qj(cid:48)(cid:48)) must be totally
different with each other (from Condition 1).
Then we use the Claim to prove that the c returned by Alg.
3 is exactly the maximum number of common super vertices
after all possible contractions. Because the 3 conditions hold
for each step, we know that finally all the common super
vertices are removed out from P and Q. From Condition 2
we know that all the removed common super vertices are the
smallest ones, from Condition 3 we get that there is not a
second way to construct the common super vertices, so the
resulted c is the maximum number of common super vertices
after all possible contractions.
1, S(cid:48)
1\S(cid:48)(cid:48)
2 ), so the unlabelled vertex sets are T (cid:48) = V \S(cid:48)
To count the intersection set (i.e. C(G(cid:48))∩C(G(cid:48)(cid:48))), assume
2), the solution sets of G(cid:48)(cid:48) is
the solution sets of G(cid:48) is (S(cid:48)
1\S(cid:48)
1 , S(cid:48)(cid:48)
(S(cid:48)(cid:48)
2,
T (cid:48)(cid:48) = V \S(cid:48)(cid:48)
2 , respectively. Denote L := T (cid:48) ∩ T (cid:48)(cid:48) be the
common vertices of the two unlabelled vertex sets, so l =
L (0 ≤ l ≤ k) is the number of common vertices in the
unlabelled k vertices. Denote M(cid:48) := T (cid:48)\L, M(cid:48)(cid:48) := T (cid:48)(cid:48)\L be
the sets of different vertex sets between the two unlabelled
vertex sets. Then,
1\M(cid:48)(cid:48)) matches
(S(cid:48)
1\M(cid:48), S(cid:48)(cid:48)
2\M(cid:48)(cid:48)) matches
1\M(cid:48), S(cid:48)(cid:48)
2\M(cid:48)).
(S(cid:48)(cid:48)
Assume w.l.o.g.
2\M(cid:48)):
(S(cid:48)(cid:48)
2\M(cid:48)(cid:48), S(cid:48)
(S(cid:48)
1\M(cid:48)(cid:48), S(cid:48)
(S(cid:48)
2\M(cid:48)(cid:48)) or
1\M(cid:48)(cid:48), S(cid:48)
1) if
that
C(G
(cid:48)
(cid:26) 2l
) =
) ∩ C(G
(cid:48)(cid:48)
1 ∪ S(cid:48)(cid:48)
if S(cid:48)
otherwise
) ∩ C(G
(cid:48)(cid:48)
(cid:48)
1 (cid:54)= ∅ and S(cid:48)
2 ∪ S(cid:48)(cid:48)
2 (cid:54)= ∅
2l − 1
2) otherwise, C(G
SG3: presented in Section III-A.
SG: summarized in Alg. 4, the methods to count its approx-
) = 0
imation sets is the same as that of SG3.
RDGreedy: summarized in Alg. 5, the methods to count its
2
1 ∪ S(cid:48)
1 ∪ S(cid:48)(cid:48)
1 ∪ S(cid:48)(cid:48)
1\M(cid:48)(cid:48), S(cid:48)
approximation sets is the same as that of D2Greedy.
D. Proof of the Correctness of Method to Count C(G(cid:48)) ∩
C(G(cid:48)(cid:48)) of SG3
1 ∪S(cid:48)(cid:48)
Proof. First of all, notice that M(cid:48) must be included in S(cid:48)(cid:48)
and M(cid:48)(cid:48) must be included in S(cid:48)
2, because M(cid:48) has no
2 ∪ M(cid:48)(cid:48) =
intersection with M(cid:48)(cid:48), and we know that S(cid:48)(cid:48)
2 ∪ M(cid:48). After removing M(cid:48) from S(cid:48)(cid:48)
1 ∪ S(cid:48)
S(cid:48)
2 , and M(cid:48)(cid:48)
1 ∪ S(cid:48)
2\M(cid:48)(cid:48))
from S(cid:48)
2, the vertices in the pairs, (S(cid:48)
1\M(cid:48), S(cid:48)(cid:48)
2\M(cid:48)), can not be changed by distributing any
and (S(cid:48)(cid:48)
other unlabelled vertices , so if they can not match with each
other, there will be no common solutions.
If they can match, in the following, there is only one way
to distribute M(cid:48) and M(cid:48)(cid:48) to have common solutions. And the
vertices in the common set L = T (cid:48) ∩ T (cid:48)(cid:48) can be distributed
consistently in the two instances, so in this situation C(G(cid:48))∩
C(G(cid:48)(cid:48)) = 2l.
E. Proof of Theorem 1
Proof. First of all, We will prove the following claim, then
use the claim to prove Theorem 1.
In each step t (t = 0,··· , n − 2), the following
Claim:
conditions hold:
1) The remained super vertices in P, Q are distinct with each
other, that means any 2 super vertices inside P or Q do
not have intersection, and there are no common super
vertex between P and Q.
2) The common super vertex removed from P, Q, i.e., pii(cid:48) =
qjj(cid:48), is the smallest common super vertex containing pi
or pi(cid:48) (respectively, qj or qj(cid:48))
3) The common super vertex removed from P, Q, i.e., pii(cid:48) =
qjj(cid:48), are "unique" (i.e., there does not exist pii(cid:48)(cid:48) = qjj(cid:48)(cid:48),
|
1307.2863 | 1 | 1307 | 2013-07-10T17:47:20 | Dynamic Data Structure for Tree-Depth Decomposition | [
"cs.DS"
] | We present a dynamic data structure for representing a graph $G$ with tree-depth at most $D$. Tree-depth is an important graph parameter which arose in the study of sparse graph classes.
The structure allows addition and removal of edges and vertices such that the resulting graph still has tree-depth at most $D$, in time bounds depending only on $D$. A tree-depth decomposition of the graph is maintained explicitly.
This makes the data structure useful for dynamization of static algorithms for graphs with bounded tree-depth. As an example application, we give a dynamic data structure for MSO-property testing, with time bounds for removal depending only on $D$ and constant-time testing of the property, while the time for the initialization and insertion also depends on the size of the formula expressing the property. | cs.DS | cs |
Dynamic Data Structure for Tree-Depth
Decomposition
Zdenek Dvor´ak
Martin Kupec
Vojtech Tuma1
Computer Science Institute, Charles University
Prague, Czech Republic
{rakdver,kupec,voyta}@iuuk.mff.cuni.cz.
Abstract
We present a dynamic data structure for representing a graph G with
tree-depth at most D. Tree-depth is an important graph parameter which
arose in the study of sparse graph classes.
The structure allows addition and removal of edges and vertices such
that the resulting graph still has tree-depth at most D, in time bounds
depending only on D. A tree-depth decomposition of the graph is main-
tained explicitly.
This makes the data structure useful for dynamization of static algo-
rithms for graphs with bounded tree-depth. As an example application,
we give a dynamic data structure for MSO-property testing, with time
bounds for removal depending only on D and constant-time testing of the
property, while the time for the initialization and insertion also depends
on the size of the formula expressing the property.
The concept of tree-depth, introduced in [12], appears prominently in the
sparse graph theory and in particular the theory of graph classes with bounded
expansion, developed mainly by Nesetril and Ossona de Mendez [11, 13, 14, 15,
16, 17]. One of its many equivalent definitions is as follows. The tree-depth
td(G) of an undirected simple graph G is the smallest integer t for that there
exists a rooted forest T of height t with vertex set V (G) such that for every
edge xy of G, either x is ancestor of y in T or vice versa -- in other words, G is
a subgraph of the closure of F .
Alternatively, tree-depth can be defined using (and is related to) rank func-
tion, vertex ranking number, minimum elimination tree or weak-coloring num-
bers. Futhermore, a class of graphs closed on subgraphs has bounded tree-depth
if and only if it does not contain arbitrarily long paths. Tree-depth is also re-
lated to other structural graph parameters -- it is greater or equal to path-width
(and thus also tree-width), and smaller or equal to the smallest vertex cover.
1The work leading to this invention has received funding from KONTAKT II LH12095 and
SVV 267313.
1
Determining tree-depth of a graph is NP-complete in general. Since tree-
depth of a graph G is at most log(G) times its tree-width, tree-depth can
be approximated up to log2(G)-factor, using the approximation algorithm for
tree-width [1]. Furthermore, for a fixed integer t, the problem of deciding
whether G has tree-depth at most t can be solved in time O(G). Minimal
minor/subgraph/induced subgraph obstructions for the class of all graphs of
tree-depth at most t are well characterized, see [4]. Clearly, tree-depth is mono-
tone with respect to all these relations. For more information about tree-depth,
see the book [18].
A motivation for investigating structural graph parameters such as tree-
depth is that restricted structure often implies efficient algorithms for prob-
lems that are generally intractable. Structural parameters have a flourishing
relationship with algorithmic meta-theorems, combining graph-theoretical and
structural approach with tools from logic and model theory -- see for instance [8].
A canonical example of a meta-theorem using a structural parameter is the re-
sult of Courcelle [3] which gives linear-time algorithms for properties expressible
in MSO logic on classes of graphs with bounded tree-width.
Tree-depth is similar to tree-width, in the sense that it measures "tree-
likeness" of a graph and also allows decomposition with algorithmically ex-
ploitable properties. However, tree-depth is more restrictive, since bounded
tree-depth implies bounded tree-width, but forbids the presence of long paths.
Long paths turn out to be related to the hardness of model checking for MSO
logic [10, 6]. This motivated a search for meta-theorems similar to [3] on more
restricted classes of graphs, such as the result of Lampis [9] that provides al-
gorithms with better dependence on the size of the formula for classes such
as those with bounded vertex cover or bounded max-leaf number. This result
was subsequently generalized by Gajarsk´y and Hlinen´y to graphs with bounded
tree-depth [7].
In the usual static setting, the problem is to decide whether a graph given
on input has some fixed property P . Our work is of dynamic kind, that is,
the considered graph gradually changes over time and we have to be able to
answer any time whether it has the property P . One application comes im-
mediately in mind. Graphs modelling many natural phenomena, such as the
web graph, graphs of social networks or graphs of some physical structure all
change rapidly. However, there is another area where this dynamic approach is
useful. For example, one reduces a graph by removing edges, and each time an
edge is removed, some procedure has to be performed. Instead of running the
procedure from scratch every time, it makes sense to keep some dynamic infor-
mation. Classical examples are the usage of a disjoint-find-union data structure
in minimal spanning tree algorithms [2] or Link-cut trees for network flow algo-
rithms [19]. A more recent example is a data structure for subgraph counting [5]
with applications in graph coloring and social networking.
The main theorem of our paper follows.
Theorem 1. Let φ be a MSO2 formula and D ∈ N. There exists a data structure
for representing a graph G with td(G) ≤ D supporting the following operations:
2
• insert edge e, provided that td(G + {e}) ≤ D,
• delete edge e,
• query -- determine whether G satisfies the formula φ.
The time complexity of deletion depends on D only, in particular, it does not
depend on φ or G. The time complexity of insertion depends on φ and D, but
does not depend on G. The time complexity of the initialization of the data
structure depends on φ, D and G. The query is done in constant time, as is
addition or removal of an isolated vertex.
The dependence of the initialization and edge insertion is roughly a tower
of height D where the highest element of the tower is the number of nested
quantifiers of φ squared.
The basic idea of the data structure is to explicitly maintain a forest of
smallest depth whose closure contains G, together with its compact constant-
size summary obtained by identifying "equivalent" subtrees. This summary is
sufficient to decide the property expressed by φ, as outlined in the following
paragraph.
Two graphs are said to be n-equivalent, if they satisfy the same first or-
der formulas with at most n quantifier alterations -- that is, for instance, of the
form ∀x1..i1 ∃xi1+1..i2 ∀xi2+1..i3 . . . ∃xin−1+1..in φ(x1..n), where φ is quantifier-free.
This concept of n-equivalency is of practical use for model checking. It serves to
reduce the investigated graph to a small one, so that time-expensive approaches
as brute force become possible (a technique known as kernelization). An exam-
ple of such application is the following theorem (from section 6.7 of [18]): for
every D, n exists N such that every graph G with td(G) ≤ D is n-equivalent to
one of its induced subgraphs of order at most N . This can be extended even
to labeled graphs. In our work we use a similar theorem, taken from [7]. Infor-
mally, the result says that when one is interested in checking whether a specific
formula is true on a class of trees of bounded depth, then one can also assume
bounded degree. This allows us to only maintain the summary of the tree-depth
decomposition as described above.
In the rest of the paper, the first section reviews necessary definitions and
tools we use, and the second section describes in detail the data structure and
its operations. We conclude the paper with the application to dynamic model
checking.
1 Preliminaries
In this paper, all trees we work with are rooted. For simplicity, we assume
in this section that all graphs we work with are connected. If we encounter a
disconnected graph, we consider each of its connected components individually.
Let T be a tree, the depth of T is the maximum length of a path from the root
of T to a leaf of T . Two trees are isomorphic if there exists a graph-isomorphism
between them such that the root is preserved under it. Mostly we will work with
3
trees with vertices labelled from some set of l labels -- two l-labelled trees are l-
isomorphic if they are isomorphic as trees and the isomorphism preserves labels.
The closure clos(T ) is the graph obtained from T by adding all edges (x, y)
such that x is an ancestor of y, and x 6= y. For instance, the closure of a path is a
complete graph. The tree-depth td(G) is the minimum number t such that there
exists a forest T of depth t such that G ⊆ clos(T ). For instance, the tree-depth
of a path on n vertices is ⌈log2(n + 1)⌉. A limb of a vertex v ∈ T is the subgraph
induced by some of the children of v. A second-order logic formula φ is in MSO1
logic, if all second-order quantifiers are over sets of elements (vertices) and the
language contains just the relation edge(u, v).
The following result is a simplification of Lemma 3.1 from [7].
Lemma 2. Let φ be an MSO1 sentence, l, D ∈ N. Then there exists a number
S with the following property. Let T be an l-labelled tree of depth at most D with
vertices labelled with l labels, and v a vertex of T . If v has more than S pairwise
l-isomorphic limbs, then for the tree T ′ obtained by deleting one of those limbs
we have that
T ′ satisfies φ ⇐⇒ T satisfies φ.
The Lemma implies in particular that with respect to φ-checking there are
only finitely many l-labelled trees of depth at most D -- that is, every l-labelled
tree of depth at most D is φ-equivalent to some l-labelled tree of depth at most
D and maximum degree at most S. We call such trees φ-minimal.
Let G be a graph of tree-depth D, the tree decomposition T of G is a 2D−1-
labelled tree such that G ⊆ clos(T ), where a vertex v is labelled by a 0-1 vector
of length D − 1 that encodes the edges between v and the vertices on the path
from v to the root (1 whenever the edge is present, 0 otherwise). Let lD be a
set of labels we describe later, compressed tree decomposition of the graph G is
an lD-labelled tree C obtained from a tree decomposition T of G as follows. For
every vertex, all its limbs that are pairwise-isomorphic are deleted except for
one representative, in which we additionally store the number of these limbs.
Vertices of C are called cabinets, and the underlying tree decomposition T is
called a decompression of C. A set of all vertices corresponding to the same
cabinet (that is, inducing lD-isomorphic limbs) and having the same vertex as a
father in the decompression is called a drawer. Thus every cabinet is disjointly
partitioned into drawers. For an example how a graph, its tree decomposition
and compressed tree decomposition look like, see the figures on page 5 (in the
compressed tree decomposition, the number next to the drawers denotes how
many vertices are there in each drawer).
Now we describe the labelling. We start inductively, with l0 being just a
set of vectors of length D − 1. Assume that φ is some fixed formula we specify
later (in Section 2.5) and let S be the number obtained from applying Lemma
2 to it. Let B be a cabinet that induces a subtree of depth t′ ≤ t in C. The
label of B consists of the label of a corresponding vertex b of T and of a vector
vec with entry for every lt−1-labelled φ-minimal tree M of depth smaller than
4
Figure 1: Graph
Figure 2: Tree-depth decomposition
1
2
1
1
2
1
2
Figure 3: Compressed graph
t′ with value
vecM = min{S, number of limbs of b which are φ-equivalent to M }.
During the update operations, we will be occasionally forced to have more
than one cabinet for a given isomorphism type (that is, a cabinet will have two
pairwise-isomorphic limbs). Both the decomposition and the individual cabinets
that have isomorphic children will be called dirty.
2 Data structure
Our data structure basically consists of storing some extra information for every
vertex v of the represented graph G, and of a compressed tree-depth decompo-
sition T of G with depth at most D. We will store the following for every
v ∈ G.
• Label of the cabinet corresponding to v, that is, the vector of its neigh-
bors on the path to root and the vector with the numbers of limbs of v
isomorphic to individual φ-minimal trees.
5
• Pointer to the father of v in T (more precisely, pointer to a vertex u of
G that is the father of v in the decompression of T ). However, in some
operations we need to change the father of all vertices in a drawer at once
-- thus instead of storing father individually for every vertex, for every
drawer we will maintain a pointer to the common father of the vertices
in this drawer, and every vertex in the drawer will have a pointer to this
pointer.
• Linked list of sons of v in T . This is again implemented by having a linked
list of drawers at v, and for every drawer in it, a linked list of vertices in
this drawer.
Additionally, we keep the vertex v which is the root of T and we call it r --
we again assume connectedness of G in this section, otherwise we keep a list of
roots corresponding to individual components.
2.1 Extraction of a path
In this subsection we describe an auxiliary operation of extracting a path. It
can be seen as a temporary decompression of a part of T in order to make some
vertex accessible. The result of extracting v from T is a dirty compressed tree
decomposition T ′ of G, such that on the cabinets in T ′ corresponding to r − v
path there are no cabinets to which corresponds more than one vertex of G.
First, we find the vertices of the r − v path, and the corresponding cabinets
in T . This is done by simply following the father-pointers from v, and then
by going backwards from r, always picking the cabinet that corresponds to the
label of the vertex on the r − v path. Then, for every cabinet B on this path
with more than one vertex, let b be the vertex of the r − v path lying in B,
and c its father -- which we assume to be the only vertex in its cabinet, C. We
remove b from the lists of sons of c of the label of b, and move b into a new list
for c, and do the corresponding change in T ′, that is, creating a new cabinet of
the same label as a son of C, thus making C a dirty cabinet.
The complexity of this operation is clearly linear in D.
2.2 Edge deletion
Edge deletion is simple -- let vu be the edge to be deleted, with v the lower
vertex (in the tree-order imposed by T ). We extract the vertex v from T . Now
u lies on the r − v path, and as there are no other vertices in the cabinets on
the corresponding path in T , we remove the edge vu from the graph and change
the labels for the cabinets and vertices accordingly. The only affected labels are
on the r − v path, and we will precompute during initialization what the label
should change into. It can also happen that removal of such edge disconnects
the graph -- this also depends only on labels and thus will be precomputed in
advance. When such situation occurs, we split T into two components -- the
new root depends only on the labels, and the vertices for which labels change
are only on the r − v path.
6
Now, we need to clean the dirty cabinets. As the only dirty cabinets are
on the r − v path, we traverse this path, starting from v and going upwards,
and for every vertex w in a dirty cabinet, we compare the label of w with the
labels of other present drawers at the father of w, and move w to the correct
drawer/cabinet.
The complexity of this operation is clearly linear in D.
2.3 Rerooting
Rerooting is also an auxiliary operation, which will allow us to easily handle the
edge insertion. This operation takes a compressed tree decomposition T and a
vertex rN of G, for which we have a guarantee that there is a tree decomposition
with depth at most D such that rN is its root, outputs one such compressed
tree decomposition T ′ and updates data for vertices in G accordingly. In this
subsection we denote by rO the root of T , that is, the old root.
We proceed as follows:
1. extract rN from T ,
2. remove rN from T entirely,
3. consider the connected components thereof -- those that do not contain
rO have depth < D and thus can be directly attached under rN . Recurse
into the component with rO.
Only the third point deserves further explanation. The components are de-
termined by the labels only, so we will precompute which labels are in which
components and what vertices are the roots of the components. Every connected
component of T − rN that does not contain rO must have as its highest vertex
(under the tree-order) a son of rN , thus these components are already in their
proper place. For the component C with rO, either it has depth < D and thus
can be attached under rN , with rO being a son of rN . We have to deal with
two details -- firstly, there might be some edges to rN from vertices that were
above rN in T -- but none of these vertices was in a cabinet with more than one
vertex, thus we only change the labels accordingly.
Secondly, the limbs of rN in T that are in C have no father after removal
of rN . But as they are in C, for every such limb there is an edge from it to
some vertex on the rN − rO path T . Choose lowest such vertex, and make it
new father for that limb. This refathering is done by using the pointers for the
drawers -- note that every cabinet that is a root of such limb consist only of single
drawer, thanks to the extraction of rN . Thus the total number of operations we
have to do is linear in D and the maximum number of children of rN , which is
ld. As in the case of edge deletion, we have to clean dirty cabinets (which are
in C) in the end. This can again be done by simply comparing labels on that
former rN − rO path.
However, it might happen that C has depth exactly D. But we are guar-
anteed that there exist a tree decomposition with rN as a root, which implies
7
that there exists a tree decomposition of C with depth D − 1. If we know which
vertex can serve as a root of such decomposition, we can apply the operation
recursively. We describe the procedure to find a root in Section 2.5. An addi-
tional thing we have to care about is that some vertices of C have an edge to
rN -- this information has to be preserved in the recursive call. But the number
of such vertices is bounded by a function of D and thus it is not a problem -- we
only modify their labels accordingly. After this recursive call, we again clean
dirty cabinets.
The complexity of this operation for one call is linear in D + ld+ time to
find new root, and there are at most D recursive calls.
2.4 Edge insertion
Let u, v be two vertices not connected with an edge, such that G + uv has
treedepth at most D, we now describe how to add such edge. If the edge uv
respects the tree-order (that is, either u lies on v − r path or vice versa), we just
extract the lower of the two vertices, add the edge, and get rid of dirty cabinets.
Otherwise, there exists a vertex r1 which is a root of some tree decomposition
of G + uv. We describe the procedure for finding it in Section 2.5. Reroot into
this vertex to obtain decomposition T1. Now, u and v must be in the same
connected component C1 of T1 − r1. Again, unless the edge uv respects the
tree-order now, we can find a vertex r2 in C1 which is a root of some tree
decomposition of C1 + uv of depth at most D − 1, and reroot into it to obtain
decomposition T2 of C1, with u and v lying in the same component C2 of C1 −r2.
Carrying on in the obvious manner, this process stops after at most D iterations.
The complexity of this operation is O(D· complexity of rerooting).
Finally, we just remark that addition and removal of a vertex (without in-
cident edges) is implemented trivially by just adding/removing new component
with the corresponding label.
2.5 Finding a root
Let us recall what we have to face in this section. We want to find a vertex
v such that there is a tree T of depth at most t′ ≤ t such that its closure
contains the connected component C of the graph G + (a, b) − {v1, v2, . . . , vk}
as a subgraph and v is a root of T . The vertices v1, v2, . . . , vk correspond to the
roots found in previous applications of this procedure, (a, b) denotes the edge
we are trying to add.
At this point we define the formula φ according to which we constructed
the labelling of our trees. Let γ(C) be a formula which is true whenever C is
connected -- this is easily seen to be expressible in MSO1 logic -- and τd(G)
the following formula:
τd(G) = (∃v ∈ G)(∀C ⊆ G)(γ(C − {v}) ⇒ τd−1(C − {v})),
with τ1(v) being always true. Then τd(G) says that there exists a tree T with
depth at most d such that G ⊆ clos(T ). Furthermore, as we need to express the
8
addition of an edge, we work with the logic with two extra constants a, b, and
modify the formula for γ accordingly to obtain γ ′ and τ ′. The resulting formula
τ ′
t is the formula φ.
Using Lemma 2, we construct all φ-minimal trees -- note that we have to
consider every possible evaluation of the constants a, b, that is, we construct all
trees of depth at most t such that no vertex has more than S pairwise-isomorphic
limbs, and then for every two of labels, we choose two arbitrary vertices having
that label, and choose them to be a and b. For every such minimal tree, we
evaluate the formula, that is, we find which vertex is to be the root of the
tree decomposition. It might happen that the formula is false, that is, no such
vertex exists, which means we evaluated a and b so that the graph has tree-
depth greater than d. But such evaluation will not occur during the run of the
structure -- recall that we restricted the edge additions -- and thus we can
safely discard these minimal trees. Thus for every minimal tree we store the
label of the vertices that can be made root, and when applying the rerooting
subroutine, we find an arbitrary vertex of this label. This has complexity at
most D, because when looking for the given vertex, we follow first pointer from
the corresponding linked list of children for a vertex.
This means that the total complexity of the edge insertion is O(D(D + lD)).
Finally, let us remark on the complexity of initialization. From [7] we conclude
that lD, that is, the number of φ-minimal trees, is roughly a tower of 2's of
height linear in D, to the power φ2. The complexity of operations we do for
every φ-minimal tree is bounded by a polynomial in D and lD.
2.6 Dynamic model checking
We now describe how to modify the structure so that it also allows queries of the
form "does G satisfy the formula ϕ", where ϕ is some fixed MSO formula. The
modifications affect only the Section 2.5. Instead of using just the formula φ to
obtain the φ-minimal trees, we apply the Lemma 2 to the formula ϕ also and in
the construction of the minimal trees and the labelling, we use the higher of the
two numbers obtained from the Lemma. Then for every such obtained minimal
tree we evaluate whether it satisfies ϕ or not, this time without evaluating the
constants a, b.
References
[1] H. L. Bodlaender, J. R. Gilbert, H. Hafsteinsson, and T. Kloks,
frontsize, and shortest elimination
Approximating treewidth, pathwidth,
tree, Journal of Algorithms, 18 (1995), pp. 238 -- 255.
[2] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Intro-
duction to algorithms, The MIT Press, 2009.
[3] B. Courcelle, The monadic second-order logic of graphs. i. recognizable
sets of finite graphs, Information and computation, 85 (1990), pp. 12 -- 75.
9
[4] Z. Dvor´ak, A. C. Giannopoulou, and D. M. Thilikos, Forbid-
den graphs for tree-depth, European Journal of Combinatorics, 33 (2012),
pp. 969 -- 979.
[5] Z. Dvor´ak and V. Tuma, A dynamic data structure for counting sub-
graphs in sparse graphs, in Algorithms and Data Structures Symposium,
WADS 2013, London (Ont.), Proceedings, Springer, 2013. To appear.
[6] M. Frick and M. Grohe, The complexity of first-order and monadic
second-order logic revisited, in Logic in Computer Science, 2002. Proceed-
ings. 17th Annual IEEE Symposium on, IEEE, 2002, pp. 215 -- 224.
[7] J. Gajarsky and P. Hlineny, Faster Deciding MSO Properties of Trees
of Fixed Height, and Some Consequences, in IARCS Annual Conference
on Foundations of Software Technology and Theoretical Computer Sci-
ence (FSTTCS 2012), D. D'Souza, T. Kavitha, and J. Radhakrishnan,
eds., vol. 18 of Leibniz International Proceedings in Informatics (LIPIcs),
Dagstuhl, Germany, 2012, Schloss Dagstuhl -- Leibniz-Zentrum fuer Infor-
matik, pp. 112 -- 123.
[8] M. Grohe and S. Kreutzer, Methods for algorithmic meta theorems,
Model Theoretic Methods in Finite Combinatorics, 558 (2011), pp. 181 --
206.
[9] M. Lampis, Algorithmic meta-theorems for restrictions of treewidth, Algo-
rithmica, 64 (2012), pp. 19 -- 37.
[10] M. Lampis, Model checking lower bounds for simple graphs, CoRR,
abs/1302.4266 (2013).
[11] J. Nesetril and P. Ossona de Mendez, Linear time low tree-width
partitions and algorithmic consequences, in Proceedings of the thirty-eighth
annual ACM symposium on Theory of computing, STOC '06, ACM, 2006,
pp. 391 -- 400.
[12]
, Tree-depth, subgraph coloring and homomorphism bounds, European
Journal of Combinatorics, 27 (2006), pp. 1022 -- 1041.
[13]
, Grad and classes with bounded expansion i. decompositions, Euro-
pean Journal of Combinatorics, 29 (2008), pp. 760 -- 776.
[14]
, Grad and classes with bounded expansion ii. algorithmic aspects,
European Journal of Combinatorics, 29 (2008), pp. 777 -- 791.
[15]
, Grad and classes with bounded expansion iii. restricted graph ho-
momorphism dualities, European Journal of Combinatorics, 29 (2008),
pp. 1012 -- 1024.
[16]
, First order properties on nowhere dense structures, J. Symbolic
Logic, 75 (2010), pp. 868 -- 887.
10
[17]
, On nowhere dense graphs, European J. Combin., 32 (2011), pp. 600 --
617.
[18]
, Sparsity: Graphs, Structures, and Algorithms, vol. 28, Springer,
2012.
[19] D. D. Sleator and R. Endre Tarjan, A data structure for dynamic
trees, Journal of computer and system sciences, 26 (1983), pp. 362 -- 391.
11
|
1901.07118 | 1 | 1901 | 2019-01-21T23:22:47 | A Space-efficient Parameterized Algorithm for the Hamiltonian Cycle Problem by Dynamic Algebraziation | [
"cs.DS"
] | An NP-hard graph problem may be intractable for general graphs but it could be efficiently solvable using dynamic programming for graphs with bounded width (or depth or some other structural parameter). Dynamic programming is a well-known approach used for finding exact solutions for NP-hard graph problems based on tree decompositions. It has been shown that there exist algorithms using linear time in the number of vertices and single exponential time in the width (depth or other parameters) of a given tree decomposition for many connectivity problems. Employing dynamic programming on a tree decomposition usually uses exponential space. In 2010, Lokshtanov and Nederlof introduced an elegant framework to avoid exponential space by algebraization. Later, F\"urer and Yu modified the framework in a way that even works when the underlying set is dynamic, thus applying it to tree decompositions.
In this work, we design space-efficient algorithms to solve the Hamiltonian Cycle and the Traveling Salesman problems, using polynomial space while the time complexity is only slightly increased. This might be inevitable since we are reducing the space usage from an exponential amount (in dynamic programming solution) to polynomial. We give an algorithm to solve Hamiltonian cycle in time $\mathcal{O}((4w)^d\, nM(n\log{n}))$ using $\mathcal{O}(dn\log{n})$ space, where $M(r)$ is the time complexity to multiply two integers, each of which being represented by at most $r$ bits. Then, we solve the more general Traveling Salesman problem in time $\mathcal{O}((4w)^d poly(n))$ using space $\mathcal{O}(\mathcal{W}dn\log{n})$, where $w$ and $d$ are the width and the depth of the given tree decomposition and $\mathcal{W}$ is the sum of weights. Furthermore, this algorithm counts the number of Hamiltonian Cycles. | cs.DS | cs | A Space-efficient Parametrized Algorithm for the
Hamiltonian Cycle Problem by Dynamic Algebraziation
Mahdi Belbasi, Martin Furer
Department of Computer Science and Engineering
University Park, PA 16802, USA
Pennsylvania State University
{mahdi,furer}@cse.psu.edu
Abstract. An NP-hard graph problem may be intractable for general graphs but it could
be efficiently solvable using dynamic programming for graphs with bounded width (or
depth or some other structural parameter). Dynamic programming is a well-known ap-
proach used for finding exact solutions for NP-hard graph problems based on tree decom-
positions. It has been shown that there exist algorithms using linear time in the number
of vertices and single exponential time in the width (depth or other parameters) of a given
tree decomposition for many connectivity problems. Employing dynamic programming on
a tree decomposition usually uses exponential space. In 2010, Lokshtanov and Nederlof in-
troduced an elegant framework to avoid exponential space by algebraization. Later, Furer
and Yu modified the framework in a way that even works when the underlying set is dy-
namic, thus applying it to tree decompositions.
In this work, we design space-efficient algorithms to solve the Hamiltonian Cycle and the
Traveling Salesman problems, using polynomial space while the time complexity is only
slightly increased. This might be inevitable since we are reducing the space usage from an
exponential amount (in dynamic programming solution) to polynomial. We give an algo-
rithm to solve Hamiltonian cycle in time O((4w)d nM (n log n)) using O(dn log n) space,
where M (r) is the time complexity to multiply two integers, each of which being repre-
sented by at most r bits. Then, we solve the more general Traveling Salesman problem in
time O((4w)dpoly(n)) using space O(Wdn log n), where w and d are the width and the
depth of the given tree decomposition and W is the sum of weights. Furthermore, this
algorithm counts the number of Hamiltonian Cycles.
1 Introduction
Dynamic programming (DP) is largely used to avoid recomputing subproblems. It may
decrease the time complexity, but it uses auxiliary space to store the intermediate values.
This auxiliary space may go up to exponential in the size of the input. This means both
the running time and the space complexity are exponential for some algorithms solving
those NP-complete problems. Space complexity is a crucial aspect of algorithm design,
because we typically run out of space before running out of time. To fix this issue, Lok-
shtanov and Nederlof [16] introduced a framework which works on a static underlying
set. The problems they considered were Subset Sum, Knapsack, Traveling Salesman (in
time O(2nw) using polynomial space),Weighted Steiner Tree, and Weighted Set Cover.
They use DFTs, zeta transforms and Mobius transforms [19, 20], taking advantage of
the fact that working on zeta (or discrete Fourier) transformed values is significantly
easier since the subset convolution operation converts to pointwise multiplication oper-
ation. In all their settings, the input is a set or a graph which means the underlying
set for the subproblems is static. Furer and Yu [12] changed this approach modifying a
dynamic programming algorithm applied to a tree decomposition (instead of the graph
itself). The resulting algorithm uses only polynomial space and the running time does
9
1
0
2
n
a
J
1
2
]
S
D
.
s
c
[
1
v
8
1
1
7
0
.
1
0
9
1
:
v
i
X
r
a
2
M. Belbasi, M. Furer
not increase drastically. By working with tree decompositions, they obtain parametrized
algorithms which are exponential in the tree-depth and linear in the number of vertices.
If the tree decomposition has a bounded width, then the algorithm is both fast and
space-efficient. In this setting, the underlying set is not static anymore, because they are
working with different bags of nodes. They show that using algebraization helps to save
space even if the underlying set is dynamic. They consider perfect matchings in their
paper. In recent years, there have been several results in this field where algebraic tools
are used to save space when DP algorithms are applied to NP-hard problems. In 2018,
Pilipczuk and Wrochna [18] applied a similar approach to solve the Minimum Dominat-
ing Set problem. Although they have not directly used these algebraic tools but it is a
similar approach (in time O(3dpoly(n)) using poly(n) space).
We have to mention that there is no general method to automatically transform dynamic
programming solutions to polynomial space solutions while increasing the tree-width pa-
rameter to the tree-depth in the running time.
One of the interesting NP-hard problems in graph theory is Hamiltonian Cycle. It seems
harder than many other graph problems. We are given a graph and we want to find a
cycle visiting each vertex exactly once. The naive deterministic algorithm for the Hamil-
tonian Cycle problem and the more general the Traveling Salesman problem runs in
time O(n!) using polynomial space. Later, deterministic DP and inclusion-exclusion al-
gorithms for these two problems running in time O∗(2n)1 using exponential space were
given in [2,13,15]. The existence of a deterministic algorithm for Hamiltonian cycle run-
ning in time O((2 − )n), for a fixed > 0 is still an open problem. There are some
randomized algorithms which run in time O((2− )n), for a fixed > 0 like the one given
in [4]. Although, there is no improvement in deterministic running time, there are some
results on parametrized algorithms. In 2011, Cygan et al. [11] designed a parametrized
algorithm for the Hamiltonian Cycle problem, which runs in time 4twV O(1). They also
√
presented a randomized algorithm for planar graphs running in time O(26.366
n). In
2015, Bodlaender et al. [6] introduced two deterministic single exponential time algo-
rithm for Hamiltonian Cycle: One based on pathwidth running in time O(6pwpwO(1)n2)2
and the other is based on treewidth running in time O(15twtwO(1)n2), where pw and tw
are the pathwidth and the treewidth respectively. The authors also solve the Traveling
Salesman problem in time O(n(2 + 2ω)pwpwO(1)) if a path decomposition of width pw
of G is given, and in time O(n(7 + 2ω+1)twtwO(1)), where ω denotes the matrix multi-
plication exponent. One of the best known upper bound for ω [21] is 2.3727. They do
not consider the space complexity of their algorithm and as far as we checked it uses
exponential space.
Recently, Curticapean et al. [10] showed that there is no positive such that the prob-
lem of counting the number of Hamiltonian cycles can be solved in O∗((6 − )pw) time
assuming SETH. Here pw is the width of the given path decomposition of the graph.
They show this tight lower bound via matrix rank.
2 Preliminaries
In this section we review notations that we use later.
1 O∗ notation hides the polynomial factors of the expression.
2 O notation hides the logarithmic factors of the expression.
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
3
2.1 Tree Decomposition
A tree decomposition of a graph G = (V, E), of G is a tree T = (VT , ET ) such that each
node x in VT is associated with a set Bx (called the bag of x) of vertices in G, and T
has the following properties:
-- The union of all bags is equal to V . In other words, for each v ∈ V , there exists at
-- For every edge {u, v} ∈ E, there exists a node x such that u, v ∈ Bx.
-- For any nodes x, y ∈ VT , and any node z ∈ VT belonging to the path connecting x
least one node x ∈ VT with Bx containing v.
and y in T , Bx ∩ By ⊆ Bz.
The width of a tree decomposition is the size of its largest bag minus one. The
treewidth of a graph G is the minimum width over all tree decompositions of G called
tw(G). In the following, we use the letter k for the treewidth. Arnborg et al. [1] showed
that constructing a tree decomposition with the minimum treewidth is an NP-hard
problem but there are some approximation algorithms for finding near-optimal tree de-
compositions [5, 7, 9]. In 1996, Bodlaender [5] introduced a linear time algorithm to find
the minimum treewidth if the treewidth is bounded by a constant.
To simplify many application algorithms, the notion of a nice tree decomposition has
been introduced which has the following properties. The tree is rooted and every node
in a nice tree decomposition has at most two children. Any node x in a nice tree decom-
position τ is of one of the following types (let c be the only child of x or let c1 and c2 be
the two children of x):
-- Leaf node, a leaf of τ without any children.
-- Forget node (forgetting vertex v), where v ∈ Bc and Bx = Bc \ {v},
-- Introduce vertex node (introducing vertex v), where v /∈ Bc and Bx = Bc ∪ {v} ,
-- Join node, where x has two children with the same bag as x, i.e. Bx = Bc1 = Bc2.
We should mention that in some papers like [12], for the sake of simplicity an introduce
edge node is defined which is not a part of the standard definition of the nice tree
decomposition. Here introduce edge nodes are not needed and we can handle the problem
easier without such nodes. In fact, we add edges to the bags as soon as the endpoints
are introduced.
It has been shown that any given tree decomposition can be converted to a nice tree
decomposition with the same treewidth in polynomial time [14].
Definition 1. The depth of a tree decomposition is the maximum number of distinct
vertices in the union of all bags on a path from the root to the leaf. We use d to denote
the depth of a given tree decomposition.
We have to mention that this is different from the depth of a tree. This is the depth of
a tree decompisition as defined above.
Definition 2. The tree-depth of a graph G, is the minimum over the depths of all tree
decompositions of G. We use td to denote the tree-depth of a graph.
After defining the treewidth and the tree-depth, now we have to talk about the
relationship between these parameters in a given graph G:
Lemma 1. (see [17, Corollary 2.5] and [8])
For any connected graph G, td(G) ≥ tw(G) + 1 ≥ td(G)
log2 V (G) .
4
M. Belbasi, M. Furer
Example: The path Pn with n vertices, has treewidth 1 and tree-depth of log2(n+1).
One should note that finding the treewidth and the tree-depth of a given graph G are
known to be an NP-hard problem. A question which arises here, is whether there exists
a tree decomposition such that its width is the treewidth and its depth is the tree-depth
of the original graph? In other words, is it possible to construct a tree decomposition of
a given graph which minimizes both the depth and the width? If the answer is yes, how
to gain such a tree decomposition. Although we are not focusing on this question here,
but it seems an interesting problem to think about.
2.2 Algebraic tools to save space
When we use dynamic programming to solve a graph problem on a tree decomposition,
it usually uses exponential space. Lokshtanov and Nederlof converted some algorithms
using subset convolution or union product into transformed version in order to reduce
the space complexity. Later, Furer and Yu [12] also used this approach in a dynamic
setting, based on tree decompositions for the Perfect Matching problem. In this work, we
introduce algorithms to solve the Hamiltonian cycle and the Traveling Salesman prob-
lems. First, let us recall some definitions.
Let R[2V ] be the set of all functions from the power set of the universe V to the ring R.
The operator ⊕ is the pointwise addition and the operator (cid:12) is the pointwise multipli-
cation.
Definition 3. A relaxation of a function f ∈ R[2V ] is a sequence of functions {f i : f i ∈
R[2V ], 0 ≤ i ≤ V}, where ∀ 0 ≤ i ≤ V and ∀X ⊆ V, f i[X] is defined as:
0
f i[X] =
f [X]
arbitrary value
if i < X,
if i = X,
if i > X.
Definition 4. The zeta transform of a function f ∈ R[2V ] is defined as:
(1)
(2)
(3)
ζf [X] =
f [Y ].
(cid:88)
Y ⊆X
(cid:88)
Y ⊆X
Definition 5. The Mobius transform of a function f ∈ R[2V ] is defined as:
µf [X] =
(−1)X\Y f [Y ].
Lemma 2. The Mobius transform is the inversion of the zeta transform and vice versa,
i.e.
µ(ζf [X]) = ζ(µf [X]) = f [X].
(4)
See [19, 20] for the proof.
Instead of storing exponentially many intermediate results, we store the zeta transformed
values. We can assume that instead of working on the original nice tree decomposition, we
are working on a mirrored nice tree decomposition to which the zeta transform has been
applied. We work on zeta transformed values and finally to recover the original values,
we use Equation 4. The zeta transform converts the "hard" union product operation (∗u)
to the "easier" pointwise multiplication operation ((cid:12)) which results in saving space.
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
5
Definition 6. Given f, g ∈ R[2V ] and X ∈ 2V , the Subset Convolution of f and g
denoted (f ∗R g) is defined as:
(f ∗R g)[X] =
f (X1)g(X \ X1).
(5)
(cid:88)
X1⊆X
(cid:88)
X1∪X2=X
Definition 7. Given f, g ∈ R[2V ] and X ∈ 2V , the Union Product of f and g denoted
(f ∗u g) is defined as:
(f ∗u g)[X] =
f (X1)g(X2).
(6)
Theorem 1. ( [3]) Applying the zeta transform to a union product operation, results
in the pointwise multiplication of zeta transforms of the outputs, i.e., given f, g ∈ R[2V ]
and X ∈ 2V ,
ζ(f ∗u g)[X] = (ζf ) (cid:12) (ζg)[X].
(7)
All of the previous works which used either DFT or the zeta transform on a given
tree decomposition (such as [12], and [16]), have one common central property that
the recursion in the join nodes, can be presented by a formula using a union product
operation. The union product and the subset convolution are complicated operations
in comparison with pointwise multiplication or pointwise addition. That is why taking
the zeta transform of a formula having union products makes the computation easier.
As noted earlier in Theorem 1, taking the zeta transform of such a term, will result in
a term having pointwise multiplication which is much easier to handle than the union
product. After doing a computation over the zeta transformed values, we can apply the
Mobius transform on the outcome to get the main result (based on Theorem 2). In
other words, the direct computation has a mirror image using the zeta transforms of the
intermediate values instead of the original ones. While the direct computation keeps track
of exponentially many intermediate values, the computation over the zeta transformed
values partitions into exponentially many branches, and they can be executed one after
another. Later, we show that this approach improves the space complexity using only
polynomial space instead of exponential space.
3 Counting the number of Hamiltonian cycles
We are given a connected graph G = (V, E) and a nice tree decomposition τ of G of
width w. If H is a Hamiltonian cycle, then the induced subgraph of H on τx (called
H[Vx], where Vx is the union of all bags in τx) is a set of disjoint paths with endpoints
in Bx (see Figure 1).
Definition 8. A pseudo-edge is a pair of endpoints of a path of length ≥ 2 in H[Vx]. We
use the (cid:104),(cid:105) notation for the pseudo-edges. E.g., in Figure 1, p = (cid:104)u, v(cid:105) is a pseudo-edge
(it does not imply that there is an edge between u and v, it just says that there is a path
of length at least two in H[Vx] where u and v are its endpoints). The (cid:104),(cid:105) notation is a
symmetrical notation since our paths are undirected, i.e., (cid:104)u, v(cid:105) = (cid:104)v, u(cid:105). Each path is
associated with a pseudo-edge.
Lemma 3. The degree of all vertices in H[Vx] is at most 2.
6
M. Belbasi, M. Furer
Fig. 1: H[Vx] is a set of paths
with endpoints in Bx
Proof. H[Vx] is a subgraph of the cycle H.
Let Tx be the vertices contained in the bags associated with nodes in the subtree τx
rooted at x which are not in Bx. Remember that vertices in pseudo-edges are vertices
in Bx. Let X be the union of pseudo-edges (in Bx). Let [Bx]2 be the set of two-element
subsets of Bx, and let X ⊆ [Bx]2. Let SX be the union of vertices involved in X. Then,
SX is a subset of Bx. For any X, define YX to be the union of SX and Tx. Let Fx be
the vertices of G which are introduced through Tx and are not present in the bag of
the parent of x. Define fx[X] to be the number of sets of disjoint paths (except in their
endpoints where they can share a vertex) whose pseudo-edge set is X (remember SX is
the union of vertices involved in X) visiting vertices of Fx exactly once (it can also visit
vertices which are not in Fx but we require them to visit at least vertices in Fx since
they are not present in the proper ancestors of x). Computing fr[∅] gives us the number
of possible Hamiltonian cycles in G. Now, we show how to compute the values of fx for
all types of nodes.
3.1 Computing fx[X]
In two rounds we will compute fx[X] efficiently. In the first round we will introduce the
recursive formulas for any kind of nodes (when space usage is still exponential) and in
the second round we will explain how to compute the zeta transformed values (in the
next section, when the space usage drops to polynomial):
-- Leaf node: Assume x is a leaf node.
fx[X] =
(cid:40)
1
0
if X = ∅,
otherwise.
(8)
Since x is a leaf node, there is no path through Tx, so for all non-empty sets X of
pseudo-edges, fx[X] is zero and for X = ∅, there is only one set of paths, which is
empty.
-- Forget node: Assume x is a forget node (forgetting vertex v) with a child c, where
Bx = Bc \ {v}. Any pseudo-edge (cid:104)u, w(cid:105) ∈ X ⊆ [Bx]2 can define a path starting from
u, going to v possibly through Tc and then going to w possibly through Tc. Here,
either or both pieces of the path (from u to v, and/or from v to w) can consist of
single edges.
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
7
(cid:88)
(cid:88)
(cid:104)u,w(cid:105)∈X
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
fx[X] =
Fig. 2: Forget node x forgetting vertex
v with the child c.
dQ fc[X \ {(cid:104)u, w(cid:105)} ∪ Q],
(9)
(cid:40)
1
0
where dQ =
if (cid:104)u, v(cid:105) ∈ Q ∪ E and (cid:104)v, w(cid:105) ∈ Q ∪ E
otherwise.
-- Introduce vertex node: Assume x is an introduce vertex node (introducing vertex
v) with a child c, where Bx = Bc ∪ {v}. The vertex v cannot be an endpoint of a
pseudo-edge because paths have length at least two.
(cid:40)
fx[X] =
fc[X]
0
if v /∈ SX ,
otherwise.
(10)
-- Join node: Assume x is a join node with two children c1 and c2, where Bx = Bc1 =
Bc2. For any given X ⊆ Bx × Bx, X can be partitioned in two sets X(cid:48) and X \ X(cid:48)
and each of them can be the set of pseudo-edges for one of the children.
Fig. 3: Join node x
with two chil-
dren c1 and
c2.
The number of such paths associated with X through Tx is equal to the sum of
the products of the number of corresponding paths associated with X(cid:48) and X \ X(cid:48)
8
M. Belbasi, M. Furer
through Tc1 and Tc2 respectively.
(cid:88)
X(cid:48)⊆X
fx[X] =
fc1[X(cid:48)]fc2[X \ X(cid:48)] = (fc1 ∗R fc2)[X].
(11)
Here we get subset convolution and we have to convert it to union product to be able
to use zeta transfrom. We do this conversion in the next subsection.
3.2 Computing ζfx[X]
In this subsection (the second round of computation), first we compute the relaxations
of fx[X] for all kinds of nodes and then we apply the zeta transform to the relaxations.
In the following section let {f i
-- Leaf node: Assume x is a leaf node. Since fx[∅] = 1 and for any X (cid:54)= ∅ : fx[X] = 0,
x}0≤i≤k be a relaxation of fx.
we can choose f i
x[X] = fx[X] for all i and X. Then
(ζf i
x)[X] = 1, for all i and X.
(12)
-- Forget node: Assume x is a forget node (forgetting vertex v) with a child c, where
Bx = Bc \ {v}. Thus,
(cid:88)
(cid:88)
f i
x[X] =
(cid:104)u,w(cid:105)∈X
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
dQ f i(cid:48)
c [X \ {(cid:104)u, w(cid:105)} ∪ Q],
(13)
where i(cid:48)(Q) = i−1+Q, i.e., i(cid:48)(Q) is the number of pseudo-edges in X \{(cid:104)u, w(cid:105)}∪Q.
Now we apply zeta transform:
(ζf i
x)[X] =
f i
x[Y ]
(cid:88)
(cid:88)
Y ⊆X
(cid:88)
(cid:104)u,w(cid:105)∈Y
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
Y ⊆X
(cid:88)
(cid:88)
(cid:88)
=
=
=
(cid:88)
(cid:88)
(cid:104)u,w(cid:105)∈X
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
{(cid:104)u,w(cid:105)}⊆Y ⊆X
dQ
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
Y ⊆(X\{(cid:104)u,w(cid:105)})
dQ f i(cid:48)(Q)
c
[Y \ {(cid:104)u, w(cid:105)} ∪ Q]
dQ f i(cid:48)(Q)
c
[Y \ {(cid:104)u, w(cid:105)} ∪ Q]
(cid:88)
(cid:88)
f i(cid:48)(Q)
c
[Y ∪ Q]
(14)
Y ⊆(X\{(cid:104)u,w(cid:105)}) f i(cid:48)(Q)
c
[Y ∪ Q] by ζ-transforms, depending on
(cid:104)u,w(cid:105)∈X
We now express EQ =(cid:80)
the size of Q. We use the abbreviation X(cid:48) = X \ {(cid:104)u, w(cid:105)}
If Q = ∅, then
(cid:88)
EQ =
f i−1
[Y ] = ζf i−1[X(cid:48)].
c
Y ⊆X(cid:48)
If Q = {(cid:104)u, v(cid:105)} or Q = {(cid:104)v, w(cid:105)}, then
(cid:88)
Y ⊆X(cid:48)
EQ =
c[Y ∪ Q] = ζf i
f i
c[X(cid:48) ∪ Q] − ζf i
c[X(cid:48)].
(15)
(16)
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
9
If Q = {(cid:104)u, v(cid:105),(cid:104)v, w(cid:105)}, then
(cid:88)
EQ =
Y ⊆X(cid:48)
= ζf i+1
c
[Y ∪ Q]
f i+1
c
[X(cid:48) ∪ Q] − ζf i+1
c
− ζf i+1
[X(cid:48) ∪ {(cid:104)v, w(cid:105)}] + ζf i+1
c
[X(cid:48) ∪ {(cid:104)u, v(cid:105)}]
[X(cid:48)].
c
(17)
With these sums computed, we can now express ζf i
(cid:88)
(ζf i
x)[X] =
x)[X] more concisely.
((d∅ − d{(cid:104)u,v(cid:105)} − d{(cid:104)v,w(cid:105)} + d{(cid:104)u,v(cid:105)},{(cid:104)v,w(cid:105)}) ζf i−1
c[X(cid:48) ∪ {(cid:104)v, w(cid:105)}]
c[X(cid:48) ∪ {(cid:104)u, v(cid:105)}] + d{(cid:104)v,w(cid:105)} ζf i
[X(cid:48) ∪ {(cid:104)u, v(cid:105),(cid:104)v, w(cid:105)}]).
(cid:104)u,w(cid:105)∈X
+d{(cid:104)u,v(cid:105)} ζf i
+d{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)} ζf i+1
c
[X(cid:48)]
c
(18)
Note that d∅ = 1 if {(cid:104)u, v(cid:105),(cid:104)v, w(cid:105)} ⊆ E, d{(cid:104)u,v(cid:105)} = 1 if (cid:104)v, w(cid:105) ∈ E, d{(cid:104)v,w(cid:105)} = 1 if
(cid:104)u, v(cid:105) ∈ E, and d{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)} is always 1, while otherwise dQ = 0. This implies that for
every (cid:104)u, w(cid:105) in the previous equation at least one of the 4 coefficients is 0. Therefore,
in each forget node, we have at most a 3X fold branching.
-- Introduce vertex node: Assume x is an introduce vertex node (introducing vertex
v) with a child c, where Bx = Bc ∪ {v}. Let Xv be the set of pseudo-edges having v
as one their endpoints. Therefore,
(cid:40)
f i
x[X] =
(cid:88)
Y ⊆X
(ζf i
x)[X] =
f i
x[Y ] =
if v /∈ SX ,
otherwise.
f i
c[X]
0
(cid:88)
Y ⊆(X\Xv)
f i
c[Y ] = (ζf i
c)[X \ Xv].
(19)
(20)
i(cid:88)
-- Join node: Assume x is a join node with two children c1 and c2, where Bx = Bc1 =
Bc2. To compute fx on a join node, we can use Equation 11. In order to convert
the subset convolution operation to pointwise multiplication, first we need to convert
the subset convolution to union product and then we are able to use Theorem 1.
To convert subset convolution to union product we introduce a relaxation of fx. Let
f i
x[X] =(cid:80)i
c2 )[X] be a relaxation of fx.
c1 ∗u f i−j
j=0(f j
(ζf i
x)[X] =
(ζf j
c1)[X] · (ζf i−j
c2 )[X], for 0 ≤ i ≤ w,
(21)
where w is the treewidth of τ .
j=0
To summerize, we present the following algorithm for the Hamiltonian Cycle problem
where a tree decomposition τ is given.
Theorem 2. Given a graph G = (V, E) and a tree decomposition τ of G, we can compute
the total number of the Hamiltonian Cycles of G in time O((4w)dnM (n log n)) and in
space O(wdn log n) by using Algorithm 1, where w and d are the width and the depth of
τ respectively, and M (n) is the time complexity to multiply two numbers which can be
encoded by at most n bits.
Proof can be found in appendix.
10
M. Belbasi, M. Furer
Algorithm 1: Counting the total number of the possible Hamiltonian cycles in a
graph given by a nice tree decomposition
Input
return
procedure (ζf )(x, X, i): //(ζf )(x, X, i) is the representation of (ζf i
: A nice tree decompoisition τ with root r
: (ζf )(r,∅, 0)
r [∅].
//f 0
x)[X]
if x is a leaf node:
return 1.
if x is a forget node: //forgetting v
return the value according to Equation 18.
if x is an introduce vertex node: //introducing v
return (ζf )(c, X \ Xv, i).
j=0(ζf )(c1, X, j) · (ζf )(c2, X, i − j).
if x is a join node:
return(cid:80)i
end procedure
4 The traveling Salesman problem
In the previous section, we showed how to count the total number of possible Hamil-
tonian cycles of a given graph. In this section, we discuss a harder problem. We know
that the Hamiltonian Cycle problem is reducible to the traveling Salesman problem by
setting all cost of edges to be 1. We could have just explained how to solve Traveling
Salesman problem but we chose to first explain the easier one to help understanding the
process. Now, we have most of the notations. First, we recap the formal definition of the
traveling Salesman problem.
Definition 9. Traveling Salesman. Given an undirected graph G = (V, E) with weighted
edges, where for all {u, v} ∈ E, cu,v is the weight (= cost) of {u, v} (nonnegative inte-
ger). In the traveling Salesman problem we are asked to find a cycle (if there is any) that
visits all of the vertices exactly once (it is a Hamiltonian cycle) with minimum cost.
As mentioned, the output should be a minimum cost Hamiltonian cycle. We have
the same notations as before. Thus, we are ready to explain our algorithmic framework:
The difference between counting the number of Hamiltonian cycles and finding the min-
imum cost of a Hamiltonian cycle (answer to the Traveling salesman problem), is that
we should work with lengths instead of the numbers of solutions. In order to solve this
problem, we work with the ring of polynomials Z[x], where x is a variable. Our algo-
i [X] is the number
of solutions in τx associated with X with cost i, and W is the sum of the weights of all
edges. The edge lengths have to be nonnegative integers as mentioned above.
rithm computes the polynomial Px(y)[X] =(cid:80)W
i [X]yi, where ax
i=0 ax
4.1 Computing the Hamiltonian Cycles of All Costs
To find the answer to the traveling salesman problem, we have find the first nonzero
coefficient of Pr(x), where r is the root of the given nice tree decomposition. As we did
for the Hamiltonian cycle problem, we show how compute this polynomial recursively
for all kinds of node.
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
11
-- Leaf node: Assume x is a leaf node. Since we require leaf nodes to have empty bags,
then
Px(y)[X] =
if X = ∅,
otherwise.
(22)
(cid:40)
1
0
-- Forget node: Assume x is a forget node (forgetting vertex v) with a child c, where
Bx = Bc \ {v}.
(cid:88)
(cid:88)
(cid:40)
Px(y)[X] =
(cid:104)u,w(cid:105)∈X
Q⊆{(cid:104)u,v(cid:105),(cid:104)v,w(cid:105)}
dQ(Pc(y))[X \ {(cid:104)u, w(cid:105)} ∪ Q],
(23)
(cid:40)
1
0
where dQ =
if (cid:104)u, v(cid:105) ∈ Q ∪ E and (cid:104)v, w(cid:105) ∈ Q ∪ E
otherwise.
-- Introduce vertex node: Assume x is an introduce vertex node (introducing vertex
v) with a child c, where Bx = Bc ∪ {v}.
Px(y)[X] =
Pc(y)[X]
0
if v /∈ SX ,
otherwise.
(24)
-- Join node: Assume x is a join node with two children c1 and c2, where Bx = Bc1 =
Bc2.
fx[X] =
Pc1(y)[X(cid:48)]Pc2(y)[X \ X(cid:48)].
(25)
(cid:88)
X(cid:48)⊆X
We skip the zeta transform part because it is similar to the Hamiltonian Cycle case.
Theorem 3. Given a graph G = (V, E) and a tree decomposition τ of G, we can
solve the Traveling Salesman problem for G in time O(4w)d · poly(n) and in space
O(Wdwn log n) where W is the sum of the weights, w and d are the width and the
depth of the tree decomposition τ respectively.
You can find the proof in appendix.
4.2 Conclusion
In this work, we solved Hamiltonian Cycle and Traveling Salesman problems with poly-
nomial space complexity where the running time is polynomial in size of the given graph
and exponential in tree-depth. Our algorithms for both problems rely on modifying a DP
approach such that instead of storing all possible intermediate values, we keep track of
zeta transformed values which was first introduced in [16], and then in [12] for dynamic
underlying sets.
References
1. Stefan Arnborg, Jens Lagergren, and Detlef Seese, Problems easy for tree-decomposable graphs ex-
tended abstract, International Colloquium on Automata, Languages, and Programming, Springer,
1988, pp. 38 -- 51.
2. Richard Bellman, Dynamic programming treatment of the travelling salesman problem, Journal of
the ACM (JACM) 9 (1962), no. 1, 61 -- 63.
3. Andreas Bjorklund, Thore Husfeldt, Petteri Kaski, and Mikko Koivisto, Fourier meets Mobius: fast
subset convolution, Proceedings of the thirty-ninth annual ACM symposium on Theory of computing,
ACM, 2007, pp. 67 -- 74.
12
M. Belbasi, M. Furer
4. Andreas Bjorklund, Petteri Kaski, and Ioannis Koutis, Directed Hamiltonicity and out-branchings via
generalized Laplacians, 44th International Colloquium on Automata, Languages, and Programming,
ICALP 2017, July 10-14, 2017, Warsaw, Poland, 2017, pp. 91:1 -- 91:14.
5. Hans L Bodlaender, A linear-time algorithm for finding tree-decompositions of small treewidth, SIAM
Journal on computing 25 (1996), no. 6, 1305 -- 1317.
6. Hans L Bodlaender, Marek Cygan, Stefan Kratsch, and Jesper Nederlof, Deterministic single ex-
ponential time algorithms for connectivity problems parameterized by treewidth, Information and
Computation 243 (2015), 86 -- 111.
7. Hans L Bodlaender, John R Gilbert, Hj´almt`yr Hafsteinsson, and Ton Kloks, Approximating
treewidth, pathwidth, and minimum elimination tree height, International Workshop on Graph-
Theoretic Concepts in Computer Science, Springer, 1991, pp. 1 -- 12.
8. Hans L Bodlaender, John R Gilbert, Hj´almtyr Hafsteinsson, and Ton Kloks, Approximating
treewidth, pathwidth, frontsize, and shortest elimination tree, Journal of Algorithms 18 (1995), no. 2,
238 -- 255.
9. Vincent Bouchitt´e, Dieter Kratsch, Haiko Muller, and Ioan Todinca, On treewidth approximations,
Discrete Applied Mathematics 136 (2004), no. 2-3, 183 -- 196.
10. Radu Curticapean, Nathan Lindzey, and Jesper Nederlof, A tight lower bound for counting hamilto-
nian cycles via matrix rank, Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on
Discrete Algorithms, Society for Industrial and Applied Mathematics, 2018, pp. 1080 -- 1099.
11. Marek Cygan, Jesper Nederlof, Marcin Pilipczuk, Michal Pilipczuk, Joham MM van Rooij, and
Jakub Onufry Wojtaszczyk, Solving connectivity problems parameterized by treewidth in single ex-
ponential time, Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium on,
IEEE, 2011, pp. 150 -- 159.
12. Martin Furer and Huiwen Yu, Space saving by dynamic algebraization based on tree-depth, Theory
of Computing Systems 61 (2017), no. 2, 283 -- 304.
13. Richard M Karp, Dynamic programming meets the principle of inclusion and exclusion, Operations
Research Letters 1 (1982), no. 2, 49 -- 51.
14. Joachim Kneis, Daniel Molle, Stefan Richter, and Peter Rossmanith, A bound on the pathwidth
of sparse graphs with applications to exact algorithms, SIAM Journal on Discrete Mathematics 23
(2009), no. 1, 407 -- 427.
15. Samuel Kohn, Allan Gottlieb, and Meryle Kohn, A generating function approach to the traveling
salesman problem, Proceedings of the 1977 annual conference, ACM, 1977, pp. 294 -- 300.
16. Daniel Lokshtanov and Jesper Nederlof, Saving space by algebraization, Proceedings of the forty-
second ACM symposium on Theory of computing, ACM, 2010, pp. 321 -- 330.
17. Jaroslav Nesetril and Patrice Ossona De Mendez, Tree-depth, subgraph coloring and homomorphism
bounds, European Journal of Combinatorics 27 (2006), no. 6, 1022 -- 1041.
18. Micha(cid:32)l Pilipczuk and Marcin Wrochna, On space efficiency of algorithms working on structural
decompositions of graphs, ACM Transactions on Computation Theory (TOCT) 9 (2018), no. 4,
18:1 -- 18:36.
19. Gian-Carlo Rota, On the foundations of combinatorial theory, I. Theory of Mobius functions,
Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete 2 (1964), no. 4, 340 -- 368.
20. Richard P Stanley, Enumerative combinatorics. vol. 1, with a foreword by Gian-Carlo Rota. corrected
reprint of the 1986 original, Cambridge Studies in Advanced Mathematics 49 (1997).
21. Virginia Vassilevska Williams, Multiplying matrices faster than Coppersmith-Winograd, Proceedings
of the Forty-fourth Annual ACM Symposium on Theory of Computing (New York, NY, USA),
STOC '12, ACM, 2012, pp. 887 -- 898.
A Space-efficient Parametrized Algorithm for Hamiltonian Cycle by Dynamic Algebraization
13
A Appendix
Here is the proof of Theorem 2.
Proof. The correctness of the algorithm is shown in Section 3.1 and Section 3.2. Now,
we show the running time and the space complexity of our algorithm.
Running time: The only case that branching happens in where we are handling a forget
node. We have at most w branches in forget nodes since the number of pseudo-edges in
each bag is bounded by w + 1 because of Lemma 3. Furthermore, based on the formula
for forget node, the two unordered pairs of (u, v) and (v, w) can contribute as an edge or
a pseudo-edge (four possible cases). On the other hand the number of forget nodes in a
path from the root to one leaf is bounded by the d (the depth of tree decomposition τ ).
Also, the number of vertices is n and we work with numbers of size at most n! (number
of paths) which can be represented by at most n log n bits. We handle multiplication of
these numbers which happens in O(M (n log n)). All being said, the total running time
is O((4w)dnM (n log n)).
Space complexity: We keep the results for each strand at a time. The number of nodes
in a path from the root to a leaf is bounded by the depth of the tree decomposition (d).
Along the path, we keep track of bags (of size w) and number of disjoint paths (at most
n! paths exist which can be shown by at most n log n bits). Therefore, the total space
complexity is O(wdn log n).
And here is the proof of Theorem 3.
Proof. The correctness of the algorithm is shown in Section 3.1 and Section 3.2. Now,
we show the running time and the space complexity of our algorithm.
Running time: The running time analysis is very similar to the analysis done for Hamil-
tonian Cycle.
Space complexity: The space complexity analysis is also similar to the analysis done for
Hamiltonian Cycle except here we have to keep track of the sum of the wights which at
most is W. So there is a factor of W in the space complexity here.
|
1012.2202 | 2 | 1012 | 2010-12-13T08:22:20 | Transforming a random graph drawing into a Lombardi drawing | [
"cs.DS"
] | The visualization of any graph plays important role in various aspects, such as graph drawing software. Complex systems (like large databases or networks) that have a graph structure should be properly visualized in order to avoid obfuscation. One way to provide an aesthetic improvement to a graph visualization is to apply a force-directed drawing algorithm to it. This method, that emerged in the 60's views graphs as spring systems that exert forces (repulsive or attractive) to the nodes.
A Lombardi drawing of a graph is a drawing where the edges are drawn as circular arcs (straight edges are considered circular arcs of infinite radius) with perfect angular resolution. This means, that consecutive edges around a vertex are equally spaced around it. In other words, each angle between the tangents of two consecutive edges is equal to $2\pi/d$ where d is the degree of that specific vertex. The requirement of using circular edges in graphs when we want to provide perfect angular resolution is necessary, since even cycle graphs cannot be drawn with straight edges when perfect angular resolution is needed.
In this survey, we provide an algorithm that takes as input a random drawing of a graph and provides its Lombardi drawing, giving a proper visualization of the graph. | cs.DS | cs |
1
Transforming a random graph drawing into a Lombardi drawing
Nicolaos Matsakis1
1 Department of Information and Computer Science, UC Irvine, CA, USA
∗ E-mail: [email protected]
Abstract
The visualization of any graph plays important role in various aspects, such as graph drawing software.
Complex systems (like large databases or networks) that have a graph structure should be properly
visualized in order to avoid obfuscation. One way to provide an aesthetic improvement to a graph
visualization is to apply a force-directed drawing algorithm to it. This method, that emerged in the
60's views graphs as spring systems that exert forces (repulsive or attractive) to the nodes. The first
force-directed drawing algorithm of Tutte [1] was aiming at providing a barycentric representation of
the graph and it is suitable for getting a straight-line drawing of 3-connected planar graphs without
crossings. Each vertex is gradually placed at the berycenter of its neighbour vertices. This method used
no spring systems, though it can be easily understood that springs can be applied to the system getting
the same result as that of the barycentric method. Unfortunately, Tutte's method usually gives bad
angular resolutions.
Many other force-directed methods have appeared since then. Eades' method [2] creates 2D layouts
using spring systems and it is suitable for graphs with less than 30 nodes. Attractive forces are applied
to adjacent vertices only and repulsive to all pairs of vertices (so we can view the edges as springs with
zero rest length and the nodes as electrically charged identical particles). Of course, the exact type of the
forces exerted is a very important aesthetic criterion. Eades sets logarithmic attractive forces and inverse
square root repulsive forces. Fruchterman and Reingold [3] apply the notion of simulated annealing, in
order to provide an even better 2D layout of less cumulative spring energy, compared to that of Eades.
The main disadvantage is, again, that the algorithm is limited to graphs having less than 40 vertices.
Hadany and Harel [4] provide a force-directed layout for much larger graphs of over 1000 vertices. In
order to obtain that, they follow a multi-scale approach where they initially consider an abstraction of
the graph, which is iteratively augmented by the graph details. Gajer, Goodrich and Kobourov [5] use
a coarsening strategy which avoids quadratic space and time complexities. They use a vertex filtration
which enables the algorithm to restrict the number of vertices relocated. They, also, introduce the idea
of realizing the graph in high-dimensional Euclidean space. An excellent survey on various force-directed
methods can be found in [6].
A Lombardi drawing of a graph is a drawing where the edges are drawn as circular arcs (straight edges
are considered degenerate circular ones) with perfect angular resolution. This means, that consecutive
edges around a vertex are equally spaced around it. In other words, each angle1 between the tangents of
two consecutive edges is equal to 2π/d where d is the degree of that specific vertex. This kind of drawing
took its name from the American artist Mark Lombardi, who mainly used this technique in his paintings.
The requirement of using circular edges in graphs when we want to provide perfect angular resolution is
necessary, since even cycle graphs cannot be drawn with straight edges when perfect angular resolution
is needed.
In this survey, we provide an algorithm that takes as input a random drawing of a graph and provides
its Lombardi drawing.
1 every angle in this survey will be expressed in rads
2
Figure 1. The spring forces on the central vertex due to the connection with the 4 circular
edges-springs, form a cumulative force equal to the sum of the spring forces of 4 straight spring-edges
(dotted lines), each one having the same length as that of its respective circular arc. In the specific
example, however, we can easily see that the cumulative force on the central vertex is not zero,
something which means that our algorithm will move the central vertex to a position where P4
i=1
→
Fi= 0
1 Lombardi Graphs
We will use the spring-force layout of a graph. A very important property of a single vertex which is
connected to a number of d vertices by circular edges is that, if the edges are regarded as circular springs,
the total spring force exterted on the central vertex will be equal to the one exerted on it if the circular
springs are replaced by straight ones, each one having the same length as the respective circular arc and
a direction onto the tangent of each respective circular edge (as in figure 1). The reason behind this, is
that each force is exerted instantly on the direction of each tangent. In figure 1 there is an example of
a node having degree equal to 4. Whether we take the 4 circular edges or the 4 straight ones (of equal
corresponding length) under account, the instant result on the move of the central vertex will be the
same.
We will use this relation extensively while forming the angular resolution around a vertex. Our
procedure actually uses, initially, an already known force-directed method and, afterwards, starts fixing
the angular resolution around each vertex of the current drawing. The target is to obtain a drawing with
the least cumulative spring energy, which will, also, have perfect angular resolution.
The type of force-directed method initially applied, depends on the number of nodes of G. Let's
assume that this is not great enough (at most 40 vertices), so we can apply the barycentric method of
Tutte. However, there are methods, as predescribed, for much greater numbers of nodes. After applying
the barycentric method, we order the nodes of the graph according to a descending order of degree. We
start by picking the first vertex in the order and choosing its neighbour vertex that is farthest away in
straight distance from it2 (for example we pick vertex b in figure 2). We erase the current edge (circular
or straight) between them and connect them with a straight edge (so in a case of a previous straight
edge connection, we do not change anything). We continue in a clockwise order to the next adjacent
edge (that is c in the example), which we erase and connect the 2 corresponding vertices (the central
and its neighbour vertex) with a circular edge which has angle 2π/d with the previously formed edge
2the reason behind this is that the minimum total energy that we want to achieve forces us to connect initially the 2
farthest points in plane with a straight edge; any other edge gives greater length and this means greater energy
3
b
a
e
c
d
Figure 2. The central vertex a has degree 4, so starting with neighbour vertex b and continuing in
clockwise order, we arrange each new circular edge having angle equal to π/2 with the exactly previous
one, until we reach again b where we stop. Bold (straight) edges refer to the previous edge visualization.
(in anticlockwise order and we obviously refer to angles between tangents in the position of the central
vertex). This is totally valid since any two points in plane can be connected with one circular arc, if we
specify a tangent passing through one of them. We continue doing this procedure, creating equal angles
between tangents, until we reach the very first edge, where we stop.
It is obvious, now, that the Lombardi property has been applied on the central examined vertex.
However, the zero cumulative spring force has been lost because the spring forces have changed, in both
the direction and the length of the springs. So, we must fix that, by finding the expression of a spring
force on a vertex when the spring is circular. Here is where the first observation is used. We must bring
the central vertex to the position of zero cumulative spring force, leaving stable its neighbour vertices.
So, our target is to minimize the energy of the spring system of this examined central vertex which is
connected to d vertices under the invariance of the Lombardi property, which means that we are not
allowed to alter the angles between the tangents of the edges and the x-axis, formed in the plane. In
i=1 ti, where ti is
the arc length of the i-th adjacent edge (starting from the one we beginned with in the previous step and
continuing in clockwise order3). When we accomplice that and move the position of the central vertex
to a position of mimimum total spring energy, we continue with the rest of the vertices in the descending
order and then we repeat the whole algorithm from the beginning several times to gain convergence. The
number of times may vary for each drawing, so we can just calculate the difference in the co-ordinates
of every vertex between each 2 repetitions of the algorithm and stop where we have reached the desired
accuracy, for each one of the vertices.
other words we seek to minimize the sum of the circular springs' lengths which equals Pd
3so in our example t1 is the circular edge (straight edges are regarded as circular arcs of infinite radius) connecting a
with vertex b
4
t
x
Figure 3. For any given arc length t, the corresponding chord length obeys the relation
x = t ∗ sin(θ)/θ where θ > 0 is the angle between x and the tangent of the arc in the intersection point
between the chord and the arc.
2 The Algorithm
A basic trigonometric relation is that the length of a chord x is equal to
x = t ∗ sin(θ)/θ
(1)
where t is the length of the arc that is defined between the 2 points of the chord and π/2 ≥ θ > 0 is the
angle between the chord x and the tangent at any of the 2 intersection points of the defined cycle and
the chord, as we can see in Figure 3.
The Euclidean distance between any points (p1, p2) in plane (where p1 = (x1, y1), p2 = (x2, y2)) is
given by the relation
Since we seek minimization of the sum of arc lengths, while keeping the Lombardi property, we want
D(p1, p2) = p(x1 − x2)2 + (y1 − y2)2
(2)
to minimize the quantity
d
Xi=1
xi ∗ θi/sin(θi)
(3)
where xi refers to the distance between the central vertex and the i-th neighbour vertex (assuming a
clockwise enumeration as indicated before from i = 1 to d) and θi is the angle between the tangent of a
circular edge connecting the central vertex with the i-th vertex neighbour of it and the straight distance
between the central vertex and the currently examined neighbour of it (0 < θ ≤ π/2). In order to avoid
obfuscation with the co-ordinates we will refer to the distance xi as D from now on, associated with each
specific central examined vertex.
From (2),(3) we get:
minx0,y0
d
Xi=1
θi ∗p(xi − xo)2 + (yi − yo)2/sin(θi)
(4)
The Lombardi property of equally spaced tangents on arc edges must be preserved while moving the
central vertex on plane. In other words, we move the central vertex trying to obtain Pd
F i
refers to the spring force exerted between the central vertex and the i-th neighbour vertex of it, assuming
zero rest length on each spring.
i=1
→
→
Fi= 0 where
However, sin(θi) is equal to the distance between the i-th neighbour node's (stable) position and
the tangent of the specific arc (see figure 4) divided by the new Euclidean distance between the 2 ver-
tices (new position of central vertex and the stable position of its neighbour vertex) which is equal to
5
Figure 4. The central vertex is moved around trying to minimize the total cimulative spring force
exterted on it and preserving the Lombardi property. Here we can see the central vertex moving to a
new position along with one of its neighbours which stays stable. Of course the previous edge may have
also been circular and not straight as here, which means that we should have used its tangent.
Figure 5. The distance between point (xa, ya) and to the straight line defined on the line segment
connecting points (x0, y0) and (xb, yb).
′
0, y
′
p(xi − xo)2 + (yi − yo)2, where the co-ordinates (xo, yo) refer to the new position of the central vertex.
0). In order to avoid complicated equations we
Assume that the central node's initial position is (x
think of the tangent of the first formed edge as tantamount to the semiaxis Ox and since we preserve the
Lombardi property while moving around the central vertex, all the tangents keep the same ordering in
plane (e.g. each one moves in parallel to its previous arrangement). This fact helps us constructing the
arrangement of the d tangents if we set the first one onto semiaxis Ox where we think of the new position
of the central vertex on (0,0). Therefore, we obtain y = 0 for the first tangent{footnotewhich we call initial
from now on, refering to each specific central examined vertex, y = tan(−(i−1)∗2π/d)x = tan(−2∗π/d)x
for the second one (by setting i = 2) and so on, until we get to the d-th tangent.
The only thing remaining is to identify the distance between each tangent and the corresponding
neighbour vertex. In general, the distance between a straight line (identified by 2 points) and a point in
plane, is given by the following equation, where the co-ordinates refer to figure 5:
D =
(xb − x0)(y0 − ya) − (x0 − xa)(yb − y0)
In our case and since we define each tangent on a semiaxis starting at point (0,0) we surely know point
p(xb − x0)2 + (yb − y0)2
(5)
(x0, y0). Point (xa, ya) is the actual neighbour vertex i correlated with the specific tangent and point
(xb, yb) can be obtained by setting xb = 1 and using the tangent form (yb/xb) = tan(−(i − 1) ∗ 2π/d)
for the specific i, to obtain yb. The overall distance between the central node's new position and the i-th
neighbour node's (stable) position is, then:
6
Using (6), we get for the sin(θi):
D = − yi + xi ∗ tan(−(i − 1) ∗ 2π/d)
p1 + tan2((i − 1)2π/d)
sin(θi) =
D
p(xi − x0)2 + (yi − y0)2
Replacing (7) into (4), we take a relation solely depending on the central node's and i-th neighbour's
coordinates. The sum of it, according to all neighbour vertices, can be minimized, moving the central
vertex to a position of minimum total energy:
(6)
(7)
(8)
)
(9)
minx0,y0
d
Xi=1
D ∗ arcsin(
√(xi−x0)2+(yi−y0)2 )
D
√(xi−x0)2+(yi−y0)2
D
which is equal to
minx0,y0
d
Xi=1
[(xi − x0)2 + (yi − y0)2] ∗ arcsin(
D
p(xi − x0)2 + (yi − y0)2
We note that D is defined in (6) and that there is no problem with the range of values of the arcsine
function since 0 < θ ≤ π/2.
After altering the position of the central vertex and moving it to a balancing position, we apply the
exact same procedure on the next vertex of the descending order and so on until we reach the V -th
vertex in the order. There, we have to repeat the whole procedure from the beginning since the positions
of the vertices should have changed, as many times as it needs for convergence under a desired accuracy.
Finally, we must note that each repetition of the algorithm on a different central vertex should be
coming along a rotation of the Cartesian coordinates in order to place the straight distance between the
old position of the central vertex and the farthest edge onto semiaxis 0x, as we have described before.
So, each time we switch to another central vertex in the descending order we should place the first
tangent onto the new 0x semiaxis by multiplying the coefficients of all tangents (the first point needed
to describe them is always (0,0) regarding all of them which is not rotated obviously, so we only apply
the transformation on each second point) by
(cid:18) cos(φ) −sin(φ)
cos(φ) (cid:19)
sin(φ)
which corresponds to the clockwise rotation matrix of the 2-dimensional Cartesian coordinates, where
angle φ is the angle between the 2 semiaxes (intersected at (0,0) of the two initial tangents of any 2
consecutive examined central vertices).
The pseudocode follows:
repeat
Apply a barycentric method to G=(V,E)
Arrange the vertices of the graph according to descending order of degree;
Let v = V and vi be the ordering;
7
//start fixing tangents
//there are no tangents in the beginning; these are formed in the first iteration of the FOR procedure
//and their direction stays unaltered unless we unmark them.
for i = 1 to v do
Find the neighbour vertex of vi which is farthest away in straight distance from it;
Delete current edge between them and connect them with a straight edge, marking it with a
tangent intersecting the central vertex;
Change (Rotate) Cartesian co-ordinates in order to place this (initial) tangent onto semiaxis Ox;
Get the next edge of the central vertex in clockwise order;
start applying the Lombardi property around the central vertex
for j = 1 to d-1 do
Delete current edge between vi and j and connect them with a circular edge having angle 2jπ/d
with the initial tangent;
Mark the tangent on this circular edge intersecting the central vertex;
j ← j + 1
end for
Move the central vertex to a position of zero cumulative force on it according to equation (8) by
not altering the directions of marked tangents in plane;
When the position of the central vertex gives zero cumulative force according to (8), set i ← i + 1
and unmark all tangents;
end for
until the difference on the co-ordinates of each vertex converges
3 Conclusions
We have sketched an algorithm for creating a Lombardi drawing of a graph, given a random drawing
of the graph. However, the algorithm's convergence remains to be tested for various numbers of graph
nodes, since force-directed methods, in general, experience different convergence attitudes, which are
highly correlated with the number of nodes of the graphs they are applied on. Especially in our case,
graphs with very high degrees on vertices should be checked for appropriate convergence.
References
1. Tutte WT (1963) How to draw a graph. Proceedings London Mathematical Society 13: 743 -- 768.
2. Eades P (1984) A heuristic for graph drawing. Congressus Numerantium : 149-160.
3. Fruchterman TMJ, Reingold EM (1991) Graph drawing by force-directed placement. Softw, Pract
Exper 21: 1129-1164.
4. Hadany R, Harel D (1999) A multi-scale algorithm for drawing graphs nicely. WG : 262-277.
5. Gajer P, Goodrich MT, Kobourov SG (2004) A multi-dimensional approach to force-directed layouts
of large graphs. Comput Geom 29: 3-18.
6. Tamassia R (2010) Handbook of Graph Drawing and Visualization, Chapman Hall/CRC.
|
1807.00121 | 1 | 1807 | 2018-06-30T03:40:50 | An optimal algorithm for 2-bounded delay buffer management with lookahead | [
"cs.DS"
] | The bounded delay buffer management problem, which was proposed by Kesselman et~al.\ (STOC 2001 and SIAM Journal on Computing 33(3), 2004), is an online problem focusing on buffer management of a switch supporting Quality of Service (QoS). The problem definition is as follows: Packets arrive to a buffer over time and each packet is specified by the {\em release time}, {\em deadline} and {\em value}. An algorithm can transmit at most one packet from the buffer at each integer time and can gain its value as the {\em profit} if transmitting a packet by its deadline after its release time. The objective of this problem is to maximize the gained profit. We say that an instance of the problem is $s$-bounded if for any packet, an algorithm has at most $s$ chances to transmit it. For any $s \geq 2$, Hajek (CISS 2001) showed that the competitive ratio of any deterministic algorithm is at least $(1 + \sqrt{5})/2 \approx 1.619$. It is conjectured that there exists an algorithm whose competitive ratio matching this lower bound for any $s$. However, it has not been shown yet. Then, when $s = 2$, B{\"o}hm et al.~(ISAAC 2016) introduced the {\em lookahead} ability to an online algorithm, that is the algorithm can gain information about future arriving packets, and showed that the algorithm achieves the competitive ratio of $(-1 + \sqrt{13})/2 \approx 1.303$. Also, they showed that the competitive ratio of any deterministic algorithm is at least $(1 + \sqrt{17})/4 \approx 1.281$. In this paper, for the 2-bounded model with lookahead, we design an algorithm with a matching competitive ratio of $(1 + \sqrt{17})/4$. | cs.DS | cs |
An optimal algorithm for 2-bounded delay buffer management with
lookahead
Koji M. Kobayashi
Abstract
The bounded delay buffer management problem, which was proposed by Kesselman et al.
(STOC 2001 and SIAM Journal on Computing 33(3), 2004), is an online problem focusing on
buffer management of a switch supporting Quality of Service (QoS). The problem definition is
as follows: Packets arrive to a buffer over time and each packet is specified by the release time,
deadline and value. An algorithm can transmit at most one packet from the buffer at each integer
time and can gain its value as the profit if transmitting a packet by its deadline after its release
time. The objective of this problem is to maximize the gained profit. We say that an instance
of the problem is s-bounded if for any packet, an algorithm has at most s chances to transmit
it. For any s ≥ 2, Hajek (CISS 2001) showed that the competitive ratio of any deterministic
algorithm is at least (1 + √5)/2 ≈ 1.619. It is conjectured that there exists an algorithm whose
competitive ratio matching this lower bound for any s. However, it has not been shown yet.
Then, when s = 2, Bohm et al. (ISAAC 2016) introduced the lookahead ability to an online
algorithm, that is the algorithm can gain information about future arriving packets, and showed
that the algorithm achieves the competitive ratio of (−1 + √13)/2 ≈ 1.303. Also, they showed
that the competitive ratio of any deterministic algorithm is at least (1 + √17)/4 ≈ 1.281.
matching competitive ratio of (1 + √17)/4.
In this paper, for the 2-bounded model with lookahead, we design an algorithm with a
1
Introduction
The online buffer management problem proposed by Aiello et al. [1] formulates the management
of buffers to store arriving packets in a network switch with Quality of Service (QoS) support as
an online problem. This problem has received much attention among online problems and has
been studied for the last fifteen years, which leads to developing various variants of this problem
(see comprehensive surveys [16, 28]). Kesselman et al. [23] proposed the bounded delay buffer
management problem as one of the variants, whose definition is as follows: Packets arrive to a
buffer over time. A packet p is specified by the release time r(p), value v(p) and deadline d(p). An
algorithm is allowed to transfer at most one packet at each integer time. If the algorithm transmits
a packet between its release time and deadline, it can gain its value as the profit. The objective
of this problem is to maximize the gained profit. The performance of an online algorithm for this
problem is evaluated using competitive analysis [11, 29]. If for any problem instance, the profit of
an optimal offline algorithm OP T is at most c times that of an online algorithm A, then we say
that the competitive ratio of A is at most c. We call a problem instance the s-bounded instance
(or s-bounded delay buffer management problem) in which for any packet p, d(p) − r(p) + 1 ≤ s.
For any s ≥ 2, Hajek [18] showed that the competitive ratio of any deterministic algorithm is at
least (1 + √5)/2 ≈ 1.619. Also, it is conjectured that for any s ≥ 2, there exists a deterministic
algorithm with a competitive ratio of (1 + √5)/2 (see, e.g. [16]), which has not been proved yet.
1
There is much research among online problems to reduce the competitive ratio of an online
algorithm for the original problems by adding extra abilities to the algorithm. One of the major
methods is called the lookahead ability, with which an online algorithm can obtain information
about arriving packets in the near future. This ability is introduced to various online problems:
The bin packing problem [17], the paging problem [2, 12], the list update problem [3], the scheduling
problem [27] and so on. Then, Bohm et al. [10] introduced the lookahead ability to the bounded
delay buffer management problem, that is, they gave an online algorithm for this problem an ability
to obtain the information about future arriving packets and analyzed its performance.
Previous Results and Our Results.
Bohm et al. [10] studied the 2-bounded bounded
delay buffer management problem with lookahead. They designed a deterministic algorithm whose
As mentioned above, for the s-bounded delay model without lookahead,
In this paper, we showed an optimal online algorithm for this problem, that is, its competitive
competitive ratio is at most (−1 + √13)/2 ≈ 1.303. Also, they proved that the competitive ratio
of any deterministic algorithm is at least (1 + √17)/4 ≈ 1.281.
ratio is exactly (1 + √17)/4.
Related Results.
Hajek [18] showed that the competitive ratio of any deterministic algorithm is at least (1+√5)/2 ≈
1.619 in the case of s ≥ 2. Independently, this bound was also shown in [13, 4, 30]. For s = ∞,
Englert and Westermann [15] developed a deterministic online algorithm whose competitive ratio
is at most 2√2 − 1 ≈ 1.829, which is the current best upper bound. For each s = 2 [23], 3 [5, 9],
and 4 [10], an algorithm with a competitive ratio of (1 + √5)/2 was designed. For any s ≥ 5,
an algorithm with a competitive ratio of larger than (1 + √5)/2 but less than 2 was shown [5, 9].
Moreover, in the case where an algorithm must decide which packet to transmit on the basis of
the current buffer situation, called the memoryless case, some results were shown [5, 9, 15]. The
agreeable deadline variant has also been studied. In this variant, the larger the release times of
packets are, the larger their deadlines are. Specifically, for any packets p and p′, d(p) ≤ d(p′)
if r(p) < r(p′). The lower bound of (1 + √5)/2 by Hajek [18] is applicable to this variant. Li
et al. [25, 21] displayed an optimal algorithm, whose competitive ratio matches the lower bound.
The case in which for any packet p, d(p) − r(p) + 1 = s has also been studied, called the s-uniform
delay variant, which is a specialized variant of the agreeable deadline variant. The current best
upper and lower bounds for this variant are (1 + √5)/2 [25, 21] and 1.377 [14], respectively.
The research on randomized algorithms for the bounded delay buffer management problem has
also been conducted extensively [13, 5, 9, 6, 19, 20, 21, 22].
In the case in which s is general,
the current best upper and lower bounds are e/(e − 1) ≈ 1.582 [5, 9, 22] and 5/4 = 1.25 [13],
respectively, against an oblivious adversary were shown. Upper and lower bounds of e/(e−1) [6, 22]
and 4/3 ≈ 1.333 [6], respectively, against an adaptive adversary were shown. For any fixed s, lower
bounds are the same with the bounds in the case in which s is general while upper bounds are
1/(1 − (1 − 1
s )s) [22] against the both adversaries.
A generalization of the bounded delay buffer management problem has been studied, called the
weighted item collection problem [7, 8, 22]. In this problem, an online algorithm does not know
the deadline of each packet but knows the relative order of the deadlines of packets. Many other
variants of the buffer management problem have been studied extensively (see e.g. [16, 28]).
2 Model Description
We formally give the definition of the 2-bounded delay buffer management problem with lookahead,
which is addressed in this paper. An input of this problem is a sequence of phases. Time begins
2
with zero and a phase occurs at an integer time. Each phase consists of three subphases. The
first occurring subphase is the arrival subphase. At an arrival subphase, arbitrary many packets
can arrive to a buffer. The buffer has no capacity limit and hence, all arriving packets can always
be accepted to the buffer. A packet p is characterized by the release time, deadline and value,
denoted by r(p), d(p) and v(p) respectively. Arrival times and deadlines are non-negative integers
and values are positive reals. d(p) − r(p) ≤ 1 holds because we focus on 2-bounded instances. The
second subphase is the transmission phase. At a transmission subphase, an algorithm can transmit
at most one packet from its buffer if any packet. At the transmission subphase at a time t, the
algorithm can obtain the information about packets arriving at time t + 1 using the lookahead
ability. The third subphase is the expiration subphase. At an expiration subphase, a packet which
has reached its deadline is discarded from its buffer. That is, at the expiration subphase at a time
t, all the packets p in the buffer such that d(p) = t are discarded.
The profit of an algorithm is the sum of the values of packets transmitted by the algorithm.
The objective of this problem is to maximize the gained profit. Let VA(σ) denote the profit of an
algorithm A for an input σ. Let OP T be an optimal offline algorithm. We say that the competitive
ratio of an online algorithm ON is at most c if for any input σ, VOP T (σ) ≤ VON (σ)c.
For ease of analysis, we assume that when OP T does not store any packet in its buffer, the
input is over. It is easy to see that this assumption does not affect the performance analysis of an
algorithm.
3 Matching Upper Bound
3.1 Notation and Definitions for Algorithm
We give definitions before defining our algorithm CompareWithPartialOPT (CP ). For any
integer time t, B(t) denotes the set of packets in CP 's buffer immediately before the arrival subphase
at time t. That is, each packet p in the set is not transmitted before t, t > r(p) and t ≤ d(p).
For integer times t, t′(≥ t) and t′′(≥ t′) and an input σ, let OP T ∗(t, t′, t′′) be an offline algorithm
such that if the packets in OP T ∗(t, t′, t′′)'s buffer immediately before the arrival subphase at time
t is equal to those in B(t), and the subinput of σ during time [t, t′] is given to OP T ∗(t, t′, t′′),
that is, packets p such that r(p) ∈ [t, t′] arrive to OP T ∗(t, t′, t′′)'s buffer during time [t, t′], then
OP T ∗(t, t′, t′′) is allowed to transmit packets only from time t to t′′ inclusive, and chooses the packets
whose total profit is maximized. If there exist packets with the same value in OP T ∗(t, t′, t′′)'s buffer,
OP T ∗(t, t′, t′′) follows a fixed tie breaking rule. Also, P (t, t′, t′′) denotes the set of t′′ − t + 1 packets
transmitted by OP T ∗(t, t′, t′′) during time [t, t′′]. Note that for any t and t′(≥ t), the following
relations hold because of the optimality of packets transmitted by OP T ∗(t, t′, t′′) during time [t, t′′]:
and
We define for any t and i,
P (t, t′, t′) ⊆ P (t, t′ + 1, t′ + 1)
P (t, t′, t′) ⊆ P (t, t′, t′ + 1)
P (t + 1, t′, t′) ⊆ P (t, t′, t′).
and
mi(t) = P (t, t + i, t + i)\P (t, t + i − 1, t + i − 1)
qi(t) = P (t, t + i, t + i + 1)\P (t, t + i, t + i).
3
(1)
(2)
(3)
Also, we define
P (t, t − 1, t − 1) = ∅.
Furthermore, we describe each value in the algorithm definition for ease of presentation as follows:
mi = mi(t)
m′′
m′
i = mi(t − 1)
i = mi(t − 2)
qi = qi(t)
q′
i = qi(t − 1)
and
In addition,
and
q′′
i = qi(t − 2).
1 + √17
R =
4
α = −3 + √17
2
.
CP uses the internal variable st for holding the name of a packet which CP transmits at a time
t. st′ =null holds at first for any integer t′. CP uses the two constants tmp1 and tmp2 if at time
t, CP cannot decide which packet to transmit at t + 1 in Cases 1.2.3.4 and 2.2.2.3. On the other
hand, once the name of a packet is set to st+1 at time t, CP certainly transmits the packet at t + 1.
3.2 Algorithm
CompareWithPartialOPT (CP )
Initialize: For any integer time t′, st′ :=null.
Consider the transmission subphase at a time t.
If the buffer stores no packets, do nothing.
Otherwise, do one of the following three cases and then transmit the packet whose name is set to
st.
Case 1 (st =null):
Case 1.1 (d(m0) = t): st := m0.
Case 1.2 (d(m0) 6= t):
Case 1.2.1 (d(m1) = t): st := m1 and st+1 := m0.
Case 1.2.2 (d(m1) = t + 1): st := m0 and st+1 := m1.
Case 1.2.3 (d(m1) 6= t + 1):
Case 1.2.3.1 (v(m0) ≥ v(m1) and v(q1) ≥ αv(m1)): st := q1 and st+1 := m0.
Case 1.2.3.2 (v(m0) ≥ v(m1) and v(q1) < αv(m1)): st := m0 and st+1 := m1.
Case 1.2.3.3 (v(m0) < v(m1) and v(q1)+v(m0)+v(m1)
Case 1.2.3.4 (v(m0) < v(m1) and v(q1)+v(m0)+v(m1)
v(m0)+v(m1) ≤ R): st := m0 and st+1 := m1.
v(m0)+v(m1) > R): st := q1 and st+1 := tmp1.
Case 2 (st =tmp1):
Case 2.1 ( v(m′
v(q′
Case 2.2 ( v(m′
v(q′
0)+v(m′
1)+v(m′
0)+v(m′
1)+v(m′
1)+v(m′
0)+v(m′
1)+v(m′
0)+v(m′
2)
1) ≤ R): st := m′
2)
1) > R):
0 and st+1 := m′
1.
4
Case 2.2.1 (d(m′
Case 2.2.2 (d(m′
1): st := m′
Case 2.2.2.1 (q′
1.
2)+v(m′
1 and v(q′
Case 2.2.2.2 (q′
v(q′
Case 2.2.2.3 (Otherwise): st := m′
2) = t + 1): st := m′
2) 6= t + 1):
2 6= q′
2 = q′
1 and st+1 := m′
2.
0)+v(m′
1)+v(m′
2)
1)+v(m′
1)+v(m′
0 and st+1 := tmp2.
2)
≤ R): st := m′
1 and st+1 := m′
2.
0 )+v(m′′
1 )+v(m′′
0 )+v(m′′
1 )+v(m′′
Case 3 (st =tmp2):
Case 3.1 ( v(m′′
v(q′′
Case 3.2 ( v(m′′
v(q′′
Case 3.2.1 (d(m′′
Case 3.2.2 (d(m′′
Case 3.2.3 (d(m′′
1 )+v(m′′
0 )+v(m′′
1 )+v(m′′
0 )+v(m′′
2 )+v(m′′
3 )
2 ) ≤ R): st := m′′
1 )+v(m′′
2 )+v(m′′
3 )
2 ) > R):
1 )+v(m′′
1 and st+1 := m′′
2.
3 ) = t + 1): st := m′′
3 ) 6= t + 1 and q′′
3 ) 6= t + 1 and q′′
2 and st+1 := m′′
3.
1 ): st := m′′
2.
2 and st+1 := m′′
1 ): st := m′′
3.
3 6= q′′
3 = q′′
3.3 Overview of the Analysis
Consider a given input σ. Let τ be the time at which CP transmits the last packet. We partition
the time sequence [0, τ ] into k sequences Ti(i = 1, . . . , k) to evaluate the competitive ratio of CP ,
in which k depends on σ and if Ti = (ti, t′
i), then ti ≤ t′
k = τ and for any j = 2, . . . , k,
tj = t′
j−1 + 1. The size of each Ti depends on which case CP executes at each time. Specifically, it
is defined as follows: Suppose that Ti = (t, t′) and then
i, t1 = 0, t′
• If Case 1.1 is executed at t, then t′ = t.
• If Case 1.2.1, 1.2.2, 1.2.3.1, 1.2.3.2 or 1.2.3.3 is executed at t, then t′ = t + 1.
• If Case 1.2.3.4 is executed at t and Case 2.1, 2.2.1 or 2.2.2.2 at t + 1, then t′ = t + 2.
• If Cases 1.2.3.4 and 2.2.2.1 are executed at t and t + 1, respectively, then t′ = t + 1.
• If Cases 1.2.3.4 and 2.2.2.3 are executed at t and t + 1, respectively, and Case 3.1, 3.2.1 or
3.2.3 is executed at t + 2, then t′ = t + 3.
• If Cases 1.2.3.4 and 2.2.2.3 are executed at t and t + 1, respectively, and Case 3.2.2 is
executed at t + 2, then t′ = t + 2.
For a time t, a packet whose release time is t and deadline is t + 1 is called a 2t-packet. On the
other hand, we also partition the time sequence [0, τ ′] into k time sequences T ′
i (i = 1, . . . , k), in
which τ ′ is the time at which OP T transmits the last packet. We will compare the total value
of packets transmitted by CP during the time Ti with that by OP T during the time T ′
is
i = (t, t′), in which if CP transmits
defined as follows: Suppose that Ti = (t, t′). Then, we define T ′
a 2t−1-packet p at time t − 1 and OP T transmits p at time t, then t = t + 1 and otherwise, t = t.
Moreover, if CP transmits a 2t′-packet p′ and OP T transmits p′ at time t′ + 1, then t′ = t′ + 1.
Otherwise, t′ = t′.
i . T ′
i
We give the lemma about T ′
i .
Lemma 3.1 A time t ∈ [0, τ ′] is contained in some T ′
j.
Proof. We prove this lemma by contradiction and assume that a time t ∈ [0, τ ′] is not contained
in any T ′
j. By the assumption of a given input, OP T transmits a packet at t. If CP transmits a
packet at t, then t is contained in some Ti and t is contained in either T ′
i−1 by the definition
of T ′. Thus, CP does not store any packet at t and does not transmit a packet. That is, no packets
i or T ′
5
arrive at t. A packet p transmitted by OP T at t is a 2t−1-packet. If not, CP can transmit p at t.
Hence, CP transmits p at time t − 1. By the definition of CP , CP executes Case 1.2.3.2, 1.2.3.3,
2.2.2.2 or 3.2.3 and transmits p at t − 1, which is the last time of Ti. Therefore, t is contained in
T ′
i by the definition of T ′, which contradicts the above assumption.
For any i(∈ [1, k]), let Vi (V ′
Ti (T ′
i ). By definition,
i ) denote the total value of packets transmitted by CP (OP T ) during
By Lemma 3.1,
Since
we have using Lemma 3.7,
VCP (σ) =
VOP T (σ) ≤
k
Xi=1
Vi.
k
Xi=1
V ′
i .
i∈[1,k](cid:26) V ′
Vi(cid:27) ,
i
VOP T (σ)
VCP (σ) ≤ Pk
i=1 V ′
i
i=1 Vi ≤ max
Pk
V ′
i
Vi ≤ R.
Therefore, we have the following theorem:
Theorem 3.2 The competitive ratio of CP is at most (1 + √17)/4.
3.4 Analysis
To show Lemma 3.7, we first prove the following lemmas. Let V (t, t′, t′′) denote the total value of
packets in P (t, t′, t′′). That is, V (t, t′, t′′) = Pp∈P (t,t′,t′′) v(p).
Lemma 3.3 For integer times t and t′(≥ t), suppose that Ti = (t, t′) and a packet that CP trans-
mits at t′ is not a 2t′ -packet. Then, V ′
i ≤ V (t, t′, t′).
Proof. Suppose that Ti = (t, t′). First, we consider CP does not transmit a 2t−1-packet at time
t − 1. Then, by definition,
(4)
Furthermore, all the 2t−1-packet, which arrive at t − 1, are stored in CP 's buffer at time t. Hence,
since we discussed the 2-bounded instance in this paper, all the packets in OP T 's buffer at t are
stored in CP 's buffer. Thus, by the optimality of OP T ∗(t, t′, t′) from Eq. (4),
T ′
i = (t, t′).
i ≤ V (t, t′, t′).
V ′
Next, we consider the case in which CP transmits a 2t−1-packet p at t− 1. First, let us consider
the case in which OP T does not transmit p at t. By definition,
i = (t, t′).
T ′
(5)
6
At time t, OP T 's buffer may store p but CP 's buffer does not. However, it does not affect a packet
which OP T transmits at t whether p is stored in OP T 's buffer because OP T does not transmit p
at t. Hence, by the optimality of OP T ∗(t, t′, t′) from Eq. (5),
Second, we discuss the case in which OP T transmits p at t. By definition,
V ′
i ≤ V (t, t′, t′).
T ′
i = (t + 1, t′).
Therefore,
in which the second inequality follows from Eq. (3).
V ′
i ≤ V (t + 1, t′, t′) ≤ V (t, t′, t′),
Lemma 3.4 For integer times t and t′(≥ t), suppose that Ti = (t, t′) and a packet that CP trans-
mits at t′ is a 2t′ -packet. Then, V ′
i ≤ V (t, t′, t′ + 1).
Proof. Suppose that Ti = (t, t′). We have in a similar way to the proof of Lemma 3.3, either
T ′
i = (t, t′) or T ′
i = (t + 1, t′). Thus,
Hence, by Eq. (2),
V ′
i ≤ V (t + 1, t′, t′) ≤ V (t, t′, t′).
V ′
i ≤ V (t, t′, t′ + 1).
Second, we consider the case in which OP T transmits a 2t′ -packet p′ at t′+1 which CP transmits
i = (t + 1, t′ + 1).
i = (t, t′ + 1). In the either case, we have in a similar way to the proof of Lemma 3.3,
at t′. If CP transmits a 2t−1-packet p at t − 1 and OP T transmits p at t, then T ′
Otherwise, T ′
V ′
i ≤ V (t, t′ + 1, t′ + 1).
Since OP T transmits a packet at time t′ + 1 which arrives at time t′, packets arriving at t′ + 1 does
not matter to OP T . Thus,
P (t, t′ + 1, t′ + 1) = P (t, t′, t′ + 1),
which leads to
V ′
i ≤ V (t, t′, t′ + 1).
Lemma 3.5 Let t be an integer time.
transmits m0(t) and m1(t) at times t and t + 1, respectively.
If CP executes Case 2.2.2.1 at time t + 1, then OP T
Proof. Suppose that Case 2.2.2.1 is executed at a time t + 1. Then, note that q1(t) 6= q2(t) by the
condition of Case 2.2.2.1. First, let us consider the case in which r(q2(t)) ≤ t+1. By the definition of
P , P (t, t+1, t+2) = {q1(t), m0(t), m1(t)}. However, q2(t) is neither m0(t) nor m1(t) by the definition
of q2(t). Also, OP T ∗(t, t+2, t+3) obtains a higher profit by q2(t) in P (t, t+2, t+3) instead of q1(t) in
P (t, t+1, t+2). That is, v(q1(t)) < v(q2(t)). Thus, V (t, t+1, t+2) < v(q2(t))+v(m0(t))+v(m1(t)),
which contradicts the optimality of OP T ∗(t, t + 1, t + 2).
7
In the following, we consider the case in which r(q2(t)) ≥ t + 2. P (t, t + 2, t + 2) contains one
packet which OP T ∗(t, t + 2, t + 2) transmits at time t + 2. Also, P (t, t + 2, t + 3) contains q2(t),
which is not in P (t, t + 2, t + 2). OP T ∗(t, t + 2, t + 3) transmits q2(t) at or after t + 2. Furthermore,
r(m2(t)) ≥ t + 2 because d(m2(t)) = t + 3 by the condition of Case 2.2.2. In a similar way to the
proofs of Lemmas 3.3 and 3.4, we can show that packets in OP T 's buffer at t is included in ones in
CP 's buffer at t. Therefore, OP T transmits q2(t) and m2(t) at or after t + 2 and transmits m0(t)
and m1(t) at t and t + 1.
Lemma 3.6 Let t be an integer time. If CP executes Case 3.2.2 at time t+2, then OP T transmits
m0(t), m1(t) and m2(t) at times t, t + 1 and t + 2, respectively.
Proof. We can show the proof of this lemma in a similar way to the proof of Lemma 3.5. Suppose
that Case 2.2.2.3 is executed at a time t + 1 and Case 3.2.2 is executed at t + 2. Note that
q1(t) = q2(t) by the condition of Case 2.2.2.3 and q1(t) 6= q3(t) by the condition of Case 3.2.2. If
r(q2(t)) ≤ t + 2, then OP T ∗(t, t + 2, t + 3) can transmit q2(t) during time [t, t + 2] instead of q1(t)
concerning P (t, t + 2, t + 3), which contradicts the optimality of OP T ∗(t, t + 2, t + 3).
Thus, r(q2(t)) ≥ t+3. P (t, t+3, t+3) contains one packet transmitted by OP T ∗(t, t+3, t+3) at
time t+3. Also, P (t, t+3, t+4) contains q3(t), which is not in P (t, t+3, t+3), and OP T ∗(t, t+3, t+4)
transmits q3(t) at or after t + 3. Moreover, since d(m3(t)) = t + 4 by the condition of Case 2.2.2,
r(m3(t)) ≥ t + 3. Therefore, OP T transmits q3(t) and m3(t) at or after t + 3 and transmits m0(t),
m1(t) and m2 during time [t, t + 2].
We are ready to prove Lemma 3.7.
Lemma 3.7 V ′
i /Vi ≤ R.
Proof. Suppose that CP executes Case 1 at a time t and t is contained in Tj. Note that if CP
executes Case 1.2.3.4 at t, then CP executes Case 2 at time t + 1. Moreover, if CP executes
Case 2.2.2.3 at t + 1, then CP executes Case 3 at time t + 2. For ease of presentation, mi and qi
denote mi(t) and qi(t), respectively.
Before proceeding to the proof, we give some inequalities used often later. When CP executes
Case 1.2.3.4 at t, by the condition of Case 1.2.3.4,
v(m0) < v(m1)
and
When CP executes Case 2.2 at t + 1,
v(q1) > (R − 1)(v(m0) + v(m1)).
When CP executes Case 3.2 at t + 2,
v(m2) > (R − 1)(v(m0) + v(m1)) + Rv(q1).
v(m3) > (R − 1)(v(m0) + v(m1) + v(m2)) + Rv(q1).
(6)
(7)
(8)
(9)
8
When CP executes Case 2.2.2.3 at t + 1,
v(m0) > (R − 1)(v(q2) + v(m1) + v(m2))
(by the condition of Case 2.2.2.3)
= (R − 1)(v(q1) + v(m1) + v(m2))
> (R − 1)(v(q1) + v(m1)) + (R − 1)((R − 1)(v(m0) + v(m1)) + Rv(q1))
= (R2 − 1)v(q1) + (R2 − R)v(m1) + (R2 − 2R + 1)v(m0)
> (R2 − 1)(R − 1)(v(m0 + v(m1)) + (R2 − R)v(m1) + (R2 − 2R + 1)v(m0)
= (R3 − 3R + 2)v(m0) + (R3 − 2R + 1)v(m1).
By rearranging this inequality, we have
v(m0) >
R3 − 2R + 1
−R3 + 3R − 1
v(m1).
(by Eq. (8))
(by Eq. (7))
(10)
Now we discuss the profit ratio for the execution of each case. When CP executes Case 1.1, CP
j ≤ V (t, t, t) by Lemma 3.3.
transmits m0 at t and Vj = v(m0) = V (t, t, t). On the other hand, V ′
Thus, V ′
j /Vj ≤ 1.
When Case 1.2.1 or 1.2.2 is executed, CP transmits both m0 and m1 and Vj = v(m0)+ v(m1) =
V (t, t + 1, t + 1). By Lemma 3.3, V ′
j ≤ V (t, t + 1, t + 1). Thus, V ′
j /Vj ≤ 1.
We consider Case 1.2.3.1. CP transmits q1 and m0 at times t and t + 1, respectively, by
definition. Thus,
Vj = v(q1) + v(m0).
Since P (t, t + 1, t + 1) = {m0, m1} by the definition of P , by Lemma 3.3,
V ′
j ≤ V (t, t + 1, t + 1) = v(m0) + v(m1).
Hence,
V ′
j
Vj ≤
v(m0) + v(m1)
v(q1) + v(m0) ≤
v(m1) + v(m1)
αv(m1) + v(m1)
=
2
α + 1
= R,
in which the second inequality follows from v(q1) ≥ αv(m1) and v(m0) ≥ v(m1), which is the
condition of Case 1.2.3.1, and the last equality follows from the definitions of α and R.
Let us consider Case 1.2.3.2. Since CP transmits m0 and m1 at t and t + 1, respectively, by
definition,
Vj = v(m0) + v(m1).
P (t, t + 1, t + 1) = {m0, m1} and P (t, t + 1, t + 2) = {q1, m0, m1} by definition. If OP T does not
transmit m1 at t + 2, we have using Lemma 3.3,
V ′
j ≤ V (t, t + 1, t + 1) = v(m0) + v(m1).
Thus, V ′
j /Vj ≤ 1. If OP T transmits m1 at t + 2, we have using Lemma 3.4,
V ′
j ≤ V (t, t + 1, t + 2) = v(q1) + v(m0) + v(m1).
By these inequalities,
V ′
j
Vj ≤
v(q1) + v(m0) + v(m1)
v(m0) + v(m1)
<
αv(m0) + v(m0) + v(m0)
v(m0) + v(m0)
=
α + 2
2
= R,
9
in which the second inequality follows from v(q1) < αv(m1) and v(m0) ≥ v(m1), which is the
execution condition of Case 1.2.3.2, and the last equality is immediately from the definitions of α
and R.
We consider Case 1.2.3.3. Since CP transmits m0 and m1 at t and t + 1, respectively,
Vj = v(m0) + v(m1).
In a similar way to the proof of Case 1.2.3.2, if OP T does not transmit m1 at t + 2, it follows from
Lemma 3.3 that
Thus, V ′
j /Vj ≤ 1. If OP T transmits m1 at time t + 2, we have by Lemma 3.4,
V ′
j ≤ V (t, t + 1, t + 1) = v(m0) + v(m1).
V ′
j ≤ V (t, t + 1, t + 2) = v(q1) + v(m0) + v(m1).
By the condition of Case 1.2.3.3,
V ′
j
Vj ≤
v(q1) + v(m0) + v(m1)
v(m0) + v(m1)
≤ R.
For the rest of the proof, suppose that CP executes Case 1.2.3.4 at t and next executes Case 2
at t + 1. Hence, CP transmits q1 at t. First, we consider the case in which Case 2.1 is executed at
t + 1. By the definition of Case 2.1, CP transmits m0 and m1 at t + 1 and t + 2, respectively, and
thus,
Vj = v(q1) + v(m0) + v(m1).
On the other hand, since P (t, t + 2, t + 2) = {m0, m1, m2} by definition, it follows from Lemma 3.3
that
Thus,
V ′
j ≤ V (t, t + 2, t + 2) = v(m0) + v(m1) + v(m2).
V ′
j
Vj ≤
v(m0) + v(m1) + v(m2)
v(q1) + v(m0) + v(m1) ≤ R,
which follows from the condition of Case 2.1.
We discuss Case 2.2.1. Since CP transmits m1 and m2 at t + 1 and t + 2, respectively, by
definition,
Since P (t, t + 2, t + 2) = {m0, m1, m2}, we have using Lemma 3.3,
Vj = v(q1) + v(m1) + v(m2).
V ′
j ≤ V (t, t + 1, t + 1) = v(m0) + v(m1) + v(m2).
10
Hence,
V ′
j
Vj ≤
<
=
<
=
=
v(m0) + v(m1) + v(m2)
v(q1) + v(m1) + v(m2)
v(m0) + v(m1) + (R − 1)(v(m0) + v(m1)) + Rv(q1)
v(q1) + v(m1) + (R − 1)(v(m0) + v(m1)) + Rv(q1)
R(v(m0) + v(m1)) + Rv(q1)
(by Eq. (8))
R(v(q1) + v(m1)) − v(m0) + (R + 1)v(q1)
R(v(m0) + v(m1)) + R(R − 1)(v(m0) + v(m1))
R(v(q1) + v(m1)) − v(m0) + (R + 1)(R − 1)(v(m0) + v(m1))
2R2v(m0)
R2(v(m0) + v(m1))
(by Eq. (7))
(R2 + R − 1)(v(q1) + v(m1)) − v(m0)
2(R2 + R − 1)v(m0) − v(m0)
<
(by Eq. (6))
2R2
2R2 + 2R − 3
<
= R.
R2
R
In Case 2.2.2.1, CP transmits m1 at t + 1 and hence,
By Lemma 3.5,
Thus,
Vj = v(q1) + v(m1).
V ′
j = v(m0) + v(m1).
V ′
j
Vj ≤
v(m0) + v(m1)
v(q1) + v(m1)
<
<
=
v(m0) + v(m1)
(R − 1)(v(m0) + v(m1)) + v(m1)
v(m0) + v(m0)
(R − 1)(v(m0) + v(m0)) + v(m0)
= √5 − 1 < R.
2
R
(by Eq. (7))
(by Eq. (6))
In Case 2.2.2.2, CP transmits m1 and m2 at t + 1 and t + 2, respectively,
Vj = v(q1) + v(m1) + v(m2).
By the definition of P , P (t, t + 2, t + 2) = {m0, m1, m2}. Since q2 = q1 by the condition of
Case 2.2.2.2, P (t, t + 2, t + 3) = {q2, m0, m1, m2} = {q1, m0, m1, m2}. If OP T does not transmit
m2 at t + 3, by Lemma 3.3,
V ′
j ≤ V (t, t + 2, t + 2) = v(m0) + v(m1) + v(m2).
Thus,
V ′
j
Vj ≤
v(m0) + v(m1) + v(m2)
v(q1) + v(m1) + v(m2)
<
v(q1) + v(m0) + v(m1) + v(m2)
v(q1) + v(m1) + v(m2)
.
If OP T transmits m2 at t + 3, by Lemma 3.4,
V ′
j ≤ V (t, t + 2, t + 3) = v(q1) + v(m0) + v(m1) + v(m2).
11
Therefore,
V ′
j
Vj ≤
v(q1) + v(m0) + v(m1) + v(m2)
v(q1) + v(m1) + v(m2)
≤ R,
which is immediately from the condition of Case 2.2.2.2.
In the following, suppose that CP executes Cases 2.2.2.3 and 3 at t + 1 and t + 2, respectively,
which indicates that CP transmits m0 at t + 1. Let us consider the case in which Case 3.1 is
executed at t + 2. Since CP transmits m1 and m2 at t + 2 and t + 3, respectively,
Vj = v(q1) + v(m0) + v(m1) + v(m2).
P (t, t + 4, t + 4) = {m0, m1, m2, m3} and thus we have using Lemma 3.3,
V ′
j ≤ V (t, t + 4, t + 4) = v(m0) + v(m1) + v(m2) + v(m3).
Thus,
V ′
j
Vj ≤
v(m0) + v(m1) + v(m2) + v(m3)
v(q1) + v(m0) + v(m1) + v(m2) ≤ R,
which is immediately from the condition of Case 3.1.
We discuss Case 3.2.1 at the end of this proof and next consider Case 3.2.2. Since CP transmits
m2 at t + 2,
Moreover, by Lemma 3.6,
Vj = v(q1) + v(m0) + v(m2).
V ′
j = v(m0) + v(m1) + v(m2).
Hence,
V ′
j
Vj ≤
<
=
<
=
<
v(m0) + v(m1) + v(m2)
v(q1) + v(m1) + v(m2)
v(m0) + v(m1) + (R − 1)(v(m0) + v(m1)) + Rv(q1)
v(q1) + v(m1) + (R − 1)(v(m0) + v(m1)) + Rv(q1)
Rv(m0) + Rv(m1) + Rv(q1)
(by Eq. (8))
Rv(m0) + (R − 1)v(m1) + (R + 1)v(q1)
Rv(m0) + Rv(m1)) + R(R − 1)(v(m0) + v(m1))
Rv(m0) + (R − 1)v(m1) + (R + 1)(R − 1)(v(m0) + v(m1))
R2v(m0) + R2v(m1)
(by Eq. (7))
(R2 + R − 1)v(m0)) + (R2 + R − 2)v(m1)
R3
2R3 + R2 − 4R + 1
< 1.23 < R.
(by Eq. (10))
In Case 3.2.3, CP transmits m2 and m3 at t + 2 and t + 3, respectively, and thus,
Vj = v(q1) + v(m0) + v(m2) + v(m3).
On the other hand, P (t, t+3, t+3) = {m0, m1, m2, m3}. Since q3 = q1 by the condition of Case 3.2.3,
P (t, t + 3, t + 4) = {q3, m0, m1, m2, m3} = {q1, m0, m1, m2, m3}. If OP T does not transmit m3 at
t + 4, by Lemma 3.3,
V ′
j ≤ V (t, t + 3, t + 3) = v(m0) + v(m1) + v(m2) + v(m3).
12
Hence,
V ′
j
Vj ≤
v(m0) + v(m1) + v(m2) + v(m3)
v(q1) + v(m0) + v(m2) + v(m3)
<
v(p) + v(m0) + v(m1) + v(m2) + v(m3)
v(q1) + v(m0) + v(m2) + v(m3)
.
If OP T transmits m3 at t + 4, by Lemma 3.4,
V ′
j ≤ V (t, t + 3, t + 4) = v(q1) + v(m0) + v(m1) + v(m2) + v(m3).
Then,
v(q1) + v(m0) + v(m1) + v(m2) + v(m3)
> R(v(m0) + v(m1) + v(m2)) + (R + 1)v(q1)
> R2(v(m0) + v(m1)) + (R2 + R + 1)v(q1)
(by Eq. (8))
> R2(v(m0) + v(m1)) + (R2 + R + 1)(R − 1)(v(m0) + v(m1))
= (R3 + R2 − 1)(v(m0) + v(m1))
(by Eq. (9))
(by Eq. (7))
(11)
Hence, we have
V ′
j
Vj ≤
<
<
v(q1) + v(m0) + v(m1) + v(m2) + v(m3)
(by Eq. (11))
v(q1) + v(m0) + v(m2) + v(m3)
(R3 + R2 − 1)(v(m0) + v(m1))
(R3 + R2 − 1)(v(m0) + v(m1)) − v(m1)
(by Eq. (10))
R4 + R3 − R
R4 + 2R3 − 4R + 1
< 1.23 < R.
Finally we discuss Case 3.2.1. Since CP transmits m2 and m3 at t + 2 and t + 3, respectively,
Vj = v(q1) + v(m0) + v(m2) + v(m3).
Since P (t, t + 4, t + 4) = {m0, m1, m2, m3},
V ′
j ≤ V (t, t + 4, t + 4) = v(m0) + v(m1) + v(m2) + v(m3)
by Lemma 3.3. In the same way as the proof of Case 3.2.2,
V ′
j
Vj ≤
<
v(m0) + v(m1) + v(m2) + v(m3)
v(q1) + v(m0) + v(m2) + v(m3)
v(q1) + v(m0) + v(m1) + v(m2) + v(m3)
v(q1) + v(m0) + v(m2) + v(m3)
< R.
References
[1] W. Aiello, Y. Mansour, S. Rajagopolan and A. Ros´en, "Competitive queue policies for differ-
entiated services," Journal of Algorithms, Vol. 55, No. 2, pp. 113–141, 2005.
13
[2] S. Albers, "On the influence of lookahead in competitive paging algorithms," Algorithmica,
Vol. 18, No. 3, pp, 283–305, 1997.
[3] S. Albers, "A competitive analysis of the list update problem with lookahead," Theoretical
Computer Science, Vol. 197, No. 1–2, pp, 95–109, 1998.
[4] N. Andelman, Y. Mansour and A. Zhu, "Competitive queueing policies for QoS switches," In
Proc. of the 14th ACM-SIAM Symposium on Discrete Algorithms, pp. 761–770, 2003.
[5] Y. Bartal, F. Chin, M. Chrobak, S. Fung, W. Jawor, R. Lavi, J. Sgall and T. Tich´y, "Online
competitive algorithms for maximizing weighted throughput of unit jobs," In Proc. of the 21st
International Symposium on Theoretical Aspects of Computer Science, pp. 187–198, 2004.
[6] M. Bienkowski, M. Chrobak and L. Jez, "Randomized competitive algorithms for online buffer
management in the adaptive adversary model," Theoretical Computer Science, Vol. 412, No.
39, pp. 5121–5131, 2011.
[7] M. Bienkowski, M. Chrobak, C. Durr, M. Hurand, A. Jez, L. Jez and G. Stachowiak, "Col-
lecting weighted items from a dynamic queue," Algorithmica, Vol. 65, No. 1, pp. 60–94, 2013,
[8] M. Bienkowski, M. Chrobak, C. Durr, M. Hurand, A. Jez, L. Jez and G. Stachowiak, "A
Φ-competitive algorithm for collecting items with increasing weights from a dynamic queue,"
Theoretical Computer Science, Vol. 475, pp. 92–102, 2013,
[9] F. Y. L. Chin, M. Chrobak, S. P. Y. Fung, W. Jawor, J. Sgall and T. Tich´y, "Online competitive
algorithms for maximizing weighted throughput of unit jobs," Journal of Discrete Algorithms,
Vol. 4, No. 2, pp. 255–276, 2006,
[10] M. Bohm, M. Chrobak, L. Jez, F. Li, J. Sgall and P. Vesel´y, "Online packet scheduling with
bounded delay and lookahead," In Proc. of the 27th International Symposium on Algorithms
and Computation, pp. 21:1–21:13, 2016.
[11] A. Borodin and R. El-Yaniv, "Online computation and competitive analysis," Cambridge Uni-
versity Press, 1998.
[12] D. Breslauer, "On competitive on-line paging with lookahead," Theoretical Computer Science,
Vol.209, No. 1–2, pp, 365–375, 1998,
[13] F. Chin and S. Fung, "Online scheduling for partial job values: Does timesharing or random-
ization help?," Algorithmica, Vol.37, pp, 149–164, 2003,
[14] M. Chrobak, W. Jawor, J. Sgall and T. Tich´y, "Improved online algorithms for buffer man-
agement in QoS switches," ACM Transactions on Algorithms, Vol.3, No.4, 2007.
[15] M. Englert and M. Westermann, "Considering suppressed packets improves buffer management
in quality of service switches," SIAM Journal on Computing, Vol.41, No.5, pp, 1166–1192,
2012.
[16] M. Goldwasser, "A survey of buffer management policies for packet switches," ACM SIGACT
News, Vol.41, No. 1, pp.100–128, 2010.
[17] E. F. Grove, "Online bin packing with lookahead," In Proc. of the 6th ACM-SIAM Symposium
on Discrete Algorithms, pp. 430–436, 1995.
14
[18] B. Hajek, "On the competitiveness of online scheduling of unit-length packets with hard dead-
lines in slotted time," In Proc. of the 35th Conference on Information Sciences and Systems,
pp. 434–438, 2001.
[19]
L. Jez, "Randomised buffer management with bounded delay against adaptive adversary,"
CoRR, abs/0907.2050, 2009.
[20]
L. Jez, "Randomized algorithm for agreeable deadlines packet scheduling," In Proc. of the
27th Symposium on Theoretical Aspects of Computer Science, pp. 489–500, 2010.
[21]
L. Jez, F. Li, J. Sethuraman and C. Stein, "Online scheduling of packets with agreeable
deadlines," ACM Transactions on Algorithms, Vol. 9, No. 1, 2012.
[22]
L. Jez, "A Universal randomized packet scheduling algorithm," Algorithmica, Vol. 67, No. 4,
pp. 498–515, 2013.
[23] A. Kesselman, Z. Lotker, Y. Mansour, B. Patt-Shamir, B. Schieber and M. Sviridenko, "Buffer
overflow management in QoS switches," SIAM Journal on Computing,Vol. 33, No. 3, pp. 563–
583, 2004.
[24] A. Kesselman, Y. Mansour and R. van Stee, "Improved competitive guarantees for QoS buffer-
ing," Algorithmica, Vol.43, No.1-2, pp. 63–80, 2005.
[25] F. Li, J. Sethuraman and C. Stein, "An optimal online algorithm for packet scheduling with
agreeable deadlines," In Proc. of the 16th ACM-SIAM Symposium on Discrete Algorithms, pp.
801–802, 2005.
[26] F. Li, J. Sethuraman and C. Stein, "Better online buffer management," In Proc. of the 18th
ACM-SIAM Symposium on Discrete Algorithms, pp. 199–208, 2007.
[27] R. Motwani, V. Saraswat and E. Torng, "Online scheduling with lookahead: Multipass assem-
bly lines," INFORMS Journal on Computing,Vol. 10, No. 3, pp. 331–340, 1998.
[28] S. I. Nikolenko and K. Kogan, "Single and multiple buffer processing," In Encyclopedia of
Algorithms, pp. 1–9, Springer, 2015.
[29] D. Sleator and R. Tarjan, "Amortized efficiency of list update and paging rules," Communi-
cations of the ACM,Vol. 28, No. 2, pp. 202–208, 1985.
[30] A. Zhu, "Analysis of queueing policies in QoS switches," Journal of Algorithms, Vol. 53, pp.
137–168, 2004.
15
|
1508.03566 | 1 | 1508 | 2015-08-14T16:43:56 | Analyzing the Performance of Lock-Free Data Structures: A Conflict-based Model | [
"cs.DS",
"cs.DC"
] | This paper considers the modeling and the analysis of the performance of lock-free concurrent data structures. Lock-free designs employ an optimistic conflict control mechanism, allowing several processes to access the shared data object at the same time. They guarantee that at least one concurrent operation finishes in a finite number of its own steps regardless of the state of the operations. Our analysis considers such lock-free data structures that can be represented as linear combinations of fixed size retry loops. Our main contribution is a new way of modeling and analyzing a general class of lock-free algorithms, achieving predictions of throughput that are close to what we observe in practice. We emphasize two kinds of conflicts that shape the performance: (i) hardware conflicts, due to concurrent calls to atomic primitives; (ii) logical conflicts, caused by simultaneous operations on the shared data structure. We show how to deal with these hardware and logical conflicts separately, and how to combine them, so as to calculate the throughput of lock-free algorithms. We propose also a common framework that enables a fair comparison between lock-free implementations by covering the whole contention domain, together with a better understanding of the performance impacting factors. This part of our analysis comes with a method for calculating a good back-off strategy to finely tune the performance of a lock-free algorithm. Our experimental results, based on a set of widely used concurrent data structures and on abstract lock-free designs, show that our analysis follows closely the actual code behavior. | cs.DS | cs | Analyzing the Performance of Lock-Free
Data Structures: A Conflict-based Model
Aras Atalar, Paul Renaud-Goud and Philippas Tsigas
Chalmers University of Technology
{aarasgoudtsigas}@chalmers.se
Abstract
This paper considers the modeling and the analysis of the performance of lock-free concurrent
data structures. Lock-free designs employ an optimistic conflict control mechanism, allowing several
processes to access the shared data object at the same time. They guarantee that at least one
concurrent operation finishes in a finite number of its own steps regardless of the state of the
operations. Our analysis considers such lock-free data structures that can be represented as linear
combinations of fixed size retry loops.
Our main contribution is a new way of modeling and analyzing a general class of lock-free
algorithms, achieving predictions of throughput that are close to what we observe in practice.
We emphasize two kinds of conflicts that shape the performance: (i) hardware conflicts, due to
concurrent calls to atomic primitives; (ii) logical conflicts, caused by simultaneous operations on
the shared data structure.
We show how to deal with these hardware and logical conflicts separately, and how to combine
them, so as to calculate the throughput of lock-free algorithms. We propose also a common
framework that enables a fair comparison between lock-free implementations by covering the whole
contention domain, together with a better understanding of the performance impacting factors.
This part of our analysis comes with a method for calculating a good back-off strategy to finely
tune the performance of a lock-free algorithm. Our experimental results, based on a set of widely
used concurrent data structures and on abstract lock-free designs, show that our analysis follows
closely the actual code behavior.
5
1
0
2
g
u
A
4
1
]
S
D
.
s
c
[
1
v
6
6
5
3
0
.
8
0
5
1
:
v
i
X
r
a
Contents
I
II
Introduction
Related Work
III Problem Statement
III-A
III-B
Running Program and Targeted Platform . . . . . . . . . . . . . . . . . . . . .
Examples and Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Immediate Upper Bounds . . . . . . . . . . . . . . . . . . . . . . . .
III-B1
Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III-B2
III-B3
Process
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV Execution without hardware conflict
IV-A
IV-B
IV-C
Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Initial Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
IV-A1
IV-A2
Notations and Definitions
. . . . . . . . . . . . . . . . . . . . . . . .
Cyclic Executions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Throughput Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V
Expansion and Complete Throughput Estimation
V-A
V-B
V-C
Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Throughput Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Several Retry Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . .
V-C1
Wasted Retries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V-C2
V-C3
Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI Experimental Evaluation
VI-A
VI-B
Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Synthetic Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-B1
Single retry loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-B2
Several retry loops . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Treiber's Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-C
Shared Counter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-D
DeleteMin in Priority List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-E
Enqueue-Dequeue on a Queue . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-F
VI-G Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VI-H
Back-Off Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
VII Conclusion
References
1
2
3
3
3
5
5
5
6
6
6
6
7
9
16
19
19
20
21
21
22
23
23
23
23
23
25
25
27
27
28
30
31
33
33
I. Introduction
2
Lock-free programming provides highly concurrent access to data and has been increasing its
footprint in industrial settings. Providing a modeling and an analysis framework capable of describing
the practical performance of lock-free algorithms is an essential, missing resource necessary to the
parallel programming and algorithmic research communities in their effort to build on previous
intellectual efforts. The definition of lock-freedom mainly guarantees that at least one concurrent
operation on the data structure finishes in a finite number of its own steps, regardless of the state of
the operations. On the individual operation level, lock-freedom cannot guarantee that an operation
will not starve.
The goal of this paper is to provide a way to model and analyze the practically observed
performance of lock-free data structures. In the literature, the common performance measure of
a lock-free data structure is the throughput, i.e. the number of successful operations per unit of
time. It is obtained while threads are accessing the data structure according to an access pattern
that interleaves local work between calls to consecutive operations on the data structure. Although
this access pattern to the data structure is significant, there is no consensus in the literature on
what access to be used when comparing two data structures. So, the amount of local work (that we
will refer as parallel work for the rest of the paper) could be constant ([MS96], [SL00]), uniformly
distributed ([HSY10], [DLM13]), exponentially distributed ([Val94], [DB08]), null ([KH14], [LJ13]),
etc., and more questionably, the average amount is rarely scanned, which leads to a partial covering
of the contention domain.
We propose here a common framework enabling a fair comparison between lock-free data
structures, while exhibiting the main phenomena that drive performance, and particularly the
contention, which leads to different kinds of conflicts. As this is the first step in this direction,
we want to deeply analyze the core of the problem, without impacting factors being diluted within
a probabilistic smoothing. Therefore, we choose a constant local work, hence constant access rate
to the data structures. In addition to the prediction of the data structure performance, our model
provides a good back-off strategy, that achieves the peak performance of a lock-free algorithm.
Two kinds of conflict appear during the execution of a lock-free algorithm, both of them leading
to additional work. Hardware conflicts occur when concurrent operations call atomic primitives on
the same data: these calls collide and conduct to stall time, that we name here expansion. Logical
conflicts take place if concurrent operations overlap: because of the lock-free nature of the algorithm,
several concurrent operations can run simultaneously, but only one retry can logically succeed. We
show that the additional work produced by the failures is not necessarily harmful for the system-wise
performance.
We then show how throughput can be computed by connecting these two key factors in an iterative
way. We start by estimating the expansion probabilistically, and emulate the effect of stall time
introduced by the hardware conflicts as extra work added to each thread. Then we estimate the
number of failed operations, that in turn lead to additional extra work, by computing again the
expansion on a system setting where those two new amounts of work have been incorporated, and
reiterate the process; the convergence is ensured by a fixed-point search.
We consider the class of lock-free algorithms that can be modeled as a linear composition of fixed
size retry loops. This class covers numerous extensively used lock-free designs such as stacks [Tre86]
(Pop, Push), queues [MS96] (Enqueue, Dequeue), counters [DLM13] (Increment, Decrement) and
priority queues [LJ13] (DeleteMin).
To evaluate the accuracy of our model and analysis framework, we performed experiments both
on synthetic tests, that capture a wide range of possible abstract algorithmic designs, and on several
reference implementations of extensively studied lock-free data structures. Our evaluation results
reveal that our model is able to capture the behavior of all the synthetic and real designs for
3
all different numbers of threads and sizes of parallel work (consequently also contention). We also
evaluate the use of our analysis as a tool for tuning the performance of lock-free code by selecting
the appropriate back-off strategy that will maximize throughput by comparing our method with
against widely known back-off policies, namely linear and exponential.
The rest of the paper is organized as follows. We discuss related work in Section II, then the
problem is formally described in Section III. We consider the logical conflicts in the absence of
hardware conflicts in Section IV, while in Section V, we firstly show how to compute the expansion,
then combine hardware and logical conflicts to obtain the final throughput estimate. We describe
the experimental results in Section VI.
II. Related Work
Anderson et al. [ARJ97] evaluated the performance of lock-free objects in a single processor real-
time system by emphasizing the impact of retry loop interference. Tasks can be preempted during
the retry loop execution, which can lead to interference, and consequently to an inflation in retry
loop execution due to retries. They obtained upper bounds for the number of interferences under
various scheduling schemes for periodic real-time tasks.
Intel [Int13] conducted an empirical study to illustrate performance and scalability of locks. They
showed that the critical section size, the time interval between releasing and re-acquiring the lock
(that is similar to our parallel section size) and number of threads contending the lock are vital
parameters.
Failed retries do not only lead to useless effort but also degrade the performance of successful
ones by contending the shared resources. Alemany et al. [AF92] have pointed out this fact, that is in
accordance with our two key factors, and, without trying to model it, have mitigated those effects
by designing non-blocking algorithms with operating system support.
Alistarh et al. [ACS14] have studied the same class of lock-free structures that we consider in this
paper. The analysis is done in terms of scheduler steps, in a system where only one thread can be
scheduled (and can then run) at each step. If compared with execution time, this is particularly
appropriate to a system with a single processor and several threads, or to a system where the
instructions of the threads cannot be done in parallel (e.g. multi-threaded program on a multi-core
processor with only read and write on the same cache line of the shared memory). In our paper,
the execution is evaluated in terms of processor cycles, strongly related to the execution time. In
addition, the "parallel work" and the "critical work" can be done in parallel, and we only consider
retry-loops with one Read and one CAS, which are serialized. In addition, they bound the asymptotic
expected system latency (with a big O, when the number of threads tends to infinity), while in our
paper we estimate the throughput (close to the inverse of system latency) for any number of threads.
A. Running Program and Targeted Platform
III. Problem Statement
In this paper, we aim at evaluating the throughput of a multi-threaded algorithm that is based
on the utilization of a shared lock-free data structure. Such a program can be abstracted by the
Procedure AbstractAlgorithm (see Figure 1) that represents the skeleton of the function which is
called by each spawned thread. It is decomposed in two main phases: the parallel section, represented
on line 3, and the retry loop, from line 4 to line 7. A retry starts at line 5 and ends at line 7.
As for line 1, the function Initialization shall be seen as an abstraction of the delay between the
spawns of the threads, that is expected not to be null, even when a barrier is used. We then consider
that the threads begin at the exact same time, but have different initialization times.
The parallel section is the part of the code where the thread does not access the shared data
structure; the work that is performed inside this parallel section can possibly depend on the value
4
Procedure AbstractAlgorithm
1 Initialization();
2 while ! done do
3
4
5
6
7
Parallel_Work();
while ! success do
current ← Read(AP);
new ← Critical_Work(current);
success ← CAS(AP, current, new);
Figure 1: Thread procedure
T0
T1
T2
T3
Figure 2: Execution with one wasted retry, and one inevitable failure
Cycle
that has been read from the data structure, e.g. in the case of processing an element that has been
dequeued from a FIFO (First-In-First-Out) queue.
In each retry, a thread tries to modify the data structure, and does not exit the retry loop until
it has successfully modified the data structure. It does that by firstly reading the access point AP of
the data structure, then according to the value that has been read, and possibly to other previous
computations that occurred in the past, the thread prepares the new desired value as an access
point of the data structure. Finally, it atomically tries to perform the change through a call to the
Compare-And-Swap (CAS) primitive. If it succeeds, i.e. if the access point has not been changed
by another thread between the first Read and the CAS, then it goes to the next parallel section,
otherwise it repeats the process. The retry loop is composed of at least one retry, and we number
the retries starting from 0, since the first iteration of the retry loop is actually not a retry, but a
try.
We analyze the behavior of AbstractAlgorithm from a throughput perspective, which is defined
T0
T1
T2
T3
Cycle
Figure 3: Execution with minimum number of failures
as the number of successful data structure operations per unit of time. In the context of Proce-
dure AbstractAlgorithm, it is equivalent to the number of successful CASs.
5
The throughput of the lock-free algorithm, that we denote by T, is impacted by several parameters.
• Algorithm parameters: the amount of work inside a call to Parallel_Work (resp. Critical_Work)
denoted by pw (resp. cw).
• Platform parameters: Read and CAS latencies (rc and cc respectively), and the number P of
processing units (cores). We assume homogeneity for the latencies, i.e. every thread experiences
the same latency when accessing an uncontended shared data, which is achieved in practice by
pinning threads to the same socket.
B. Examples and Issues
We first present two straightforward upper bounds on the throughput, and describe the two kinds
of conflict that keep the actual throughput away from those upper bounds.
1) Immediate Upper Bounds: Trivially, the minimum amount of work rlw(-) in a given retry is
rlw(-) = rc + cw + cc, as we should pay at least the memory accesses and the critical work cw in
between.
Thread-wise: A given thread can at most perform one successful retry every pw + rlw(-) units
of time. In the best case, P threads can then lead to a throughput of P/(pw + rlw(-)).
System-wise: By definition, two successful retries cannot overlap, hence we have at most 1
successful retry every rlw(-) units of time.
Altogether, the throughput T is bounded by
1
rc + cw + cc ,
P
pw + rc + cw + cc
(cid:19)
, i.e.
(cid:18)
T ≤ min
(
1
T ≤
rc+cw+cc
pw+rc+cw+cc
P
if pw ≤ (P − 1)(rc + cw + cc)
otherwise.
(1)
2) Conflicts:
Logical conflicts: Equation 1 expresses the fact that when pw is small enough, i.e. when pw ≤
(P −1)rlw(-), we cannot expect that every thread performs a successful retry every pw +rlw(-) units
of time, since it is more than what the retry loop can afford. As a result, some logical conflicts,
hence unsuccessful retries, will be inevitable, while the others, if any, are called wasted.
However, different executions can lead to different numbers of failures, which end up with different
throughput values. Figures 2 and 3 depict two executions, where the black parts are the calls to
Initialization, the blue parts are the parallel sections, and the retries can be either unsuccessful -- in
red -- or successful -- in green. We experiment different initialization times, and observe different
synchronizations, hence different numbers of wasted retries. After the initial transient state, the
execution depicted in Figure 3 comprises only the inevitable unsuccessful retries, while the execution
of Figure 2 contains one wasted retry.
We can see on those two examples that a cyclic execution is reached after the transient behavior;
actually, we show in Section IV that, in the absence of hardware conflicts, every execution will become
periodic, if the initialization times are spaced enough. In addition, we prove that the shortest period
is such that, during this period, every thread succeeds exactly once. This finally leads us to define
the additional failures as wasted, since we can directly link the throughput with this number of
wasted retries: a higher number of wasted retries implying a lower throughput.
CAS
Read & cw
Expansion
Previously
expanded CAS
Figure 4: Expansion
6
Hardware conflicts: The requirement of atomicity compels the ownership of the data in an
exclusive manner by the executing core. This fact prohibits concurrent execution of atomic instruc-
tions if they are operating on the same data. Therefore, overlapping parts of atomic instructions
are serialized by the hardware, leading to stalls in subsequently issued ones. For our target lock-free
algorithm, these stalls that we refer to as expansion become an important slowdown factor in case
threads interfere in the retry loop. As illustrated in Figure 4, the latency for CAS can expand and
cause remarkable decreases in throughput since the CAS of a successful thread is then expanded by
others; for this reason, the amount of work inside a retry is not constant, but is, generally speaking,
a function depending on the number of threads that are inside the retry loop.
3) Process: We deal with the two kinds of conflicts separately and connect them together through
the fixed-point iterative convergence.
In Section V-A, we compute the expansion in execution time of a retry, noted e, by following
a probabilistic approach. The estimation takes as input the expected number of threads inside the
retry loop at any time, and returns the expected increase in the execution time of a retry due to the
serialization of atomic primitives.
In Section IV, we are given program without hardware conflict described by the size of the parallel
section pw(+) and the size of a retry rlw(+). We compute upper and lower bounds on the throughput
T, the number of wasted retries w, and the average number of threads inside the retry loop Prl.
Without loss of generality, we can normalize those execution times by the execution time of a retry,
and define the parallel section size as pw(+) = q + r, where q is a non-negative integer and r is such
that 0 ≤ r < 1. This pair (together with the number of threads P) constitutes the actual input of
the estimation.
Finally, we combine those two outcomes in Section V-B by emulating expansion through work not
prone to hardware conflicts and obtain the full estimation of the throughput.
IV. Execution without hardware conflict
We show in this section that, in the absence of hardware conflicts, the execution becomes periodic,
which eases the calculation of the throughput. We start by defining some useful concepts: (f, P)-
cyclic executions are special kind of periodic executions such that within the shortest period, each
thread performs exactly f unsuccessful retries and 1 successful retry. The well-formed seed is a set
of events that allows us to detect an (f, P)-cyclic execution early, and the gaps are a measure of the
quality of the synchronization between threads. The idea is to iteratively add threads into the game
and show that the periodicity is maintained. Theorem 1 establishes a fundamental relation between
gaps and well-formed seeds, while Theorem 2 proves the periodicity, relying on the disjoint cases
of Lemma 1, 3, and 4. Finally, we exhibit upper and lower bounds on throughput and number of
failures, along with the average number of threads inside the retry loop.
A. Setting
1) Initial Restrictions:
7
Remark 1. Concerning correctness, we assume that the reference point of the Read and the CAS
occurs when the thread enters and exits any retry, respectively.
Remark 2. We do not consider simultaneous events, so all inequalities that refer to time comparison
are strict, and can be viewed as follows: time instants are real numbers, and can be equal, but every
event is associated with a thread; also, in order to obtain a strict order relation, we break ties
according to the thread numbers (for instance with the relation <).
2) Notations and Definitions: We recall that P threads are executing the pseudo-code described in
Procedure AbstractAlgorithm, one retry is of unit-size, and the parallel section is of size pw(+) = q+r,
where q is a non-negative integer and r is such that 0 ≤ r < 1. Considering a thread Tn which succeeds
at time Sn; this thread completes a whole retry in 1 unit of time, then executes the parallel section of
size pw(+), and attempts to perform again the operation every unit of time, until one of the attempt
is successful.
Definition 1. An execution with P threads is called (C, P)-cyclic execution if and only if (i) the
execution is periodic, i.e. at every time, every thread is in the same state as one period before, (ii)
the shortest period contains exactly one successful attempt per thread, (iii) the shortest period is
1 + q + r + C.
Definition 2. Let S = (Ti, Si)i∈(cid:74)0,P−1(cid:75), where Ti are threads and Si ordered times, i.e. such that
S0 < S1 < ··· < SP−1. S is a seed if and only if for all i ∈(cid:74)0, P − 1(cid:75), Ti does not succeed between
S0 and Si, and starts a retry at Si.
We define f (S) as the smallest non-negative integer such that S0 + 1 + q + r + f (S) > SP−1 + 1,
i.e. f (S) = max (0,dSP−1 − S0 − q − re). When S is clear from the context, we denote f (S) by f.
Definition 3. S is a well-formed seed if and only if for each i ∈(cid:74)0, P − 1(cid:75), the execution of thread Ti
contains the following sequence: firstly a success beginning at Si, the parallel section, f unsuccessful
retries, and finally a successful retry.
Those definitions are coupled through the two natural following properties:
Property 1. Given a (C, P)-cyclic execution, any seed S including P consecutive successes is a
well-formed seed, with f (S) = C.
Proof: Choosing any set of P consecutive successes, we are ensured, by the definition of a (f, P)-
cyclic execution, that for each thread, after the first success, the next success will be obtained after
f failures. The order will be preserved, and this shows that a seed including our set of successes is
actually a well-formed seed.
Property 2. If there exists a well-formed seed in an execution, then after each thread succeeded once,
the execution coincides with an (f, P)-cyclic execution.
Proof: By the definition of a well-formed seed, we know that the threads will first succeed in
order, fails f times, and succeed again in the same order. Considering the second set of successes
in a new well-formed seed, we observe that the threads will succeed a third time in the same order,
after failing f times. By induction, the execution coincides with an (f, P)-cyclic execution.
Together with the seed concept, we define the notion of gap that we will use extensively in the
next subsection. The general idea of those gaps is that within an (f, P)-cyclic execution, the period
is higher than P × 1, which is the total execution time of all the successful retries within the period.
The difference between the period (that lasts 1 + q + r + f) and P, reduced by r (so that we obtain
an integer), is referred as lagging time in the following. If the threads are numbered according to
their order of success (modulo P), as the time elapsed between the successes of two given consecutive
8
TP−1
(2)
G
0
T0
P−1X
n=0
(1)
G
1
T1
G(1)
n
T2
Figure 5: Gaps
(1)
G
2
threads is constant (during the next period, this time will remain the same), this lagging time can be
seen in a circular manner (see Figure 5): the threads are represented on a circle whose length is the
lagging time increased by r, and the length between two consecutive threads is the time between the
end of the successful retry of the first thread and the begin of the successful retry of the second one.
n between Tn and its kth predecessor
(k)
based on the gap with the first predecessor:
More formally, for all (n, k) ∈(cid:74)0, P − 1(cid:75)2, we define the gap G
0 = S0 + q + r + f − SP−1
(1)
G
; G
n = Sn − Sn−1 − 1
(1)
,
which leads to the definition of higher order gaps:
( ∀n ∈(cid:74)1, P − 1(cid:75)
∀n ∈(cid:74)0, P − 1(cid:75)
;
∀k > 0 ; G(k)
n =
(1)
j mod P .
G
nX
j=n−k+1
For consistency, for all n ∈(cid:74)0, P − 1(cid:75), G
Equally, the gaps can be obtained directly from the successes: for all k ∈(cid:74)1, P − 1(cid:75),
(0)
n = 0.
(
G(k)
n =
Sn − Sn−k − k
if n > k
Sn − SP +n−k + 1 + q + r + f − k otherwise
(2)
Note that, in an (f, P)-cyclic execution, the lagging time is the sum of all first order gaps, reduced
by r.
Now we extend the concept of well-formed seed to weakly-formed seed.
Definition 4. Let S = (Ti, Si)i∈(cid:74)0,P−1(cid:75) be a seed.
S is a weakly-formed seed for P threads if and only if: (Ti, Si)i∈(cid:74)0,P−2(cid:75) is a well-formed seed for
P − 1 threads, and the first thread succeeding after TP−2 is TP−1.
(cid:17), for each n ∈(cid:74)0, P − 1(cid:75), G
Property 3. Let S = (Ti, Si)i∈(cid:74)0,P−1(cid:75) be a weakly-formed seed.
Denoting f = f
(Ti, Si)i∈(cid:74)0,P−2(cid:75), the previous well-formed seed with P−1 threads, we know that for all n ∈(cid:74)1, P − 2(cid:75),
n ≤ eG
eG
n , for all n ∈(cid:74)0, P − 1(cid:75) and k; hence
(cid:16)(Ti, Si)i∈(cid:74)0,P−2(cid:75)
0 = eG
0, and if we note indeed eG
Proof: We have SP−2 + 1 < SP−1 < Rf
(1)
0 , which leads to G
(1)
n = G
(1)
P−1 + G
the weaker property.
(1)
n , and G
the gaps within
(f)
n < 1.
(k)
n
(k)
(k)
(1)
T0
T1
T2
Figure 6: Lemma 1 configuration
9
Lemma 1. Let S be a weakly-formed seed, and f = f
(f+1)
G
n
f (S0) = f + 1.
< 1, then there exists later in the execution a well-formed seed S0 for P threads such that
(cid:16)(Ti, Si)i∈(cid:74)0,P−2(cid:75)
(cid:17). If, for all n ∈(cid:74)0, P − 1(cid:75),
Proof: The proof is straightforward; S is actually a well-formed seed such that f (S) = f + 1.
0 − SP−1 < G
0 < 1, the first success of T0 after the success of TP−1 is its f + 1th retry.
(1)
Since Rf
(f)
n < 1.
B. Cyclic Executions
Theorem 1. Given a seed S = (Ti, Si)i∈(cid:74)0,P−1(cid:75), S is a well-formed seed if and only if for all
n ∈(cid:74)0, P − 1(cid:75), 0 ≤ G
Proof:
Let S = (Ti, Si)i∈(cid:74)0,P−1(cid:75) be a seed.
(⇐) We assume that for all n ∈(cid:74)0, P − 1(cid:75), 0 < G
(f)
n < 1, and we first show that the first successes
occur in the following order: T0 at S0, T1 at S1, . . . , TP−1 at SP−1, T0 again at Rf
0. The first threads
that are successful executes their parallel section after their success, then enters their second retry
loop: from this moment, they can make the first attempt of the threads, that has not been successful
yet, fail. Therefore, we will look at which retry of which already successful threads could have an
impact on which other threads.
We can notice that for all n ∈ (cid:74)0, P − 1(cid:75), if the first success of Tn occurs at Sn, then its next
n = Sn + 1 + q + r + k, where k ≥ 0. More specifically, thanks
attempts will potentially occur at Rk
to Equation 2, for all n ≤ f, Rk
n = SP +n−f + G
n + k. Also, for all k ≤ f − n,
(f)
n − SP +n−f+k = − (SP +n−f+k − SP +n−f − k) + G(f)
Rk
n
n − SP +n−f+k = G(f−k)
Rk
n
= G(f)
n − G
,
(k)
P +n−f+k
(3)
and this implies that if k > 0,
.
n
(f−k)
n
SP +n−f+k − Rk−1
We know, by hypothesis, that 0 < G
n = 1 − G(f−k)
< 1, equivalently 0 < 1 − G
(4)
< 1. Therefore
Equation 3 states that if a thread Tn0 starts a successful attempt at SP +n−f+k, then this thread will
make the kth retry of Tn fail, since Tn enters a retry while Tn0 is in a successful retry. And Equation 4
shows that, given a thread Tn0 starting a new retry at SP +n−f+k, the only retry of Tn that can make
Tn0 fail on its attempt is the (k − 1)th one. There is indeed only one retry of Tn that can enter a
retry before the entrance of Tn0, and exit the retry after it.
T0 is the first thread to succeed at S0, because no other thread is in the retry loop at this time.
Its next attempt will occur at R0
0, and all thread attempts that start before SP−f (included) cannot
fail because of T0, since it runs then the parallel section. Also, since all gaps are positive, the threads
T1 to TP−f will succeed in this order, respectively starting at times S1 to SP−f.
(f−k)
n
10
n
Then, using induction, we can show that TP−f+1, . . . , TP−1 succeed in this order, respectively
we show that it implies that TP−f+j+1 will succeed at SP−f+j+1. The successful attempt of TP−f+j
But for each Tj0, this attempt was precisely the one that could have made TP−f+j+1 fail on its attempt
at SP−f+j+1 (explanation of Equation 3). Given that all threads Tn, where n > P − f + j + 1, do
not start any retry loop before SP−f+j+1, TP−f+j+1 will succeed at SP−f+j+1. By induction, (Pj) is
starting at times SP−f+1, . . . , SP−1. For j ∈(cid:74)0, f − 1(cid:75), let (Pj) be the following property: for all
n ∈(cid:74)0, P − f + j(cid:75), Tn starts a successful retry at Sn. We assume that for a given j, (Pj) is true, and
at SP−f+j leads, for all j0 ∈(cid:74)0, j(cid:75), to the failure of the j0th retry of Tj−j0 (explanation of Equation 3).
true for all j ∈(cid:74)0, f − 1(cid:75).
Finally, when TP−1 succeeds, it makes the (f − 1− n)th retry of Tn fail, for all n ∈(cid:74)0, f − 1(cid:75); also
. (Naturally, for all n ∈(cid:74)f, P − 1(cid:75), the next
We can observe that for all n < P, j ∈(cid:74)0, P − 1 − n(cid:75), and all k ≥ j,
Rk−j
n+j − Rk
n = Sn+j + k − j − (Sn + k)
Rk−j
n+j − Rk
n = G
hence for all n ∈(cid:74)1, f(cid:75), Rf−n
n − Rf
(n)
0 = G
n > 0.
n − Rf
Rf−n
As we have as well, for all n ∈(cid:74)f + 1, P − 1(cid:75), R0
the next potentially successful attempt for Tn is at Rf−n
potentially successful attempt for Tn is at R0
n.)
f, we obtain that among all the threads,
0. Following TP−1, T0 is consequently the next successful
the earliest possibly successful attempt is Rf
thread in its f th retry.
To conclude this part, we can renumber the threads (Tn+1 becoming now Tn if n > 0, and T0
becoming TP−1), and follow the same line of reasoning. The only difference is the fact that TP−1
(according to the new numbering) enters the retry loop f units of time before SP−1, but it does not
interfere with the other threads, since we know that those attempts will fail.
(f)
n = 0. This implies that
There remains the case where there exists n ∈(cid:74)0, P − 1(cid:75) such that G
(⇒) We prove now the implication by contraposition; we assume that there exists n ∈(cid:74)0, P − 1(cid:75)
f = 0, thus we have a well-formed seed.
0 = G(n)
n > 0.
n > R0
(j)
n+j,
(5)
such that G
(f)
n > 1 or G
n < 0, and show that S is not a well-formed seed.
(f)
is negative; let n00 be the highest one.
We assume first that an f th order gap is negative. As it is a sum of 1st order gaps, then there
exists n0 such that G
(1)
n0
If n00 > 0, then either the threads T0, . . . ,Tn00−1 succeeded in order at their 0th retry, and then
Tn00−1 makes Tn00 fail at its 0th retry (we have a seed, hence by definition, Sn00−1 < Sn00, and G
(1)
n00 < 0,
thus Sn00−1 < Sn00 < Sn00−1 + 1 ), or they did not succeed in order at their first try. In both cases, S
is not a well-formed seed.
i)i∈(cid:74)0,P−1(cid:75),
0 = SP−1 − (q + 1 + f + r). Like S, S0 is a well-formed
is negative, and we fall back into the previous case, which shows that S0 is not
If n00 = 0, let us assume that S is a well-formed seed. Let also a new seed be S0 = (Ti, S0
where for all n ∈(cid:74)0, P − 2(cid:75), S0
(1)
seed; however, G
1
a well-formed seed. This is absurd, hence S is not a well-formed seed.
n+1 = Sn, and S0
n+k > 1}, and f0 = min{k ; G
(k)
those that concern the earliest thread, and among them the one with the lowest order.
We assume now that every gap is positive and choose n0 defined by: n0 = min{n ; ∃k ∈
(cid:74)0, P − 1(cid:75) /G
n0+k > 1}: among the gaps that exceed 1, we pick
Let us assume that threads T0, . . . , TP−1 succeed at their 0th retry in this order, then T0, . . . , Tn0
complete their second successful retry loop at their f th retry, in this order. If this is not the case, then
S is not a well-formed seed, and the proof is completed. According to Equation 5, we have, on the one
(k)
11
n0+f0 −(cid:0)Rf
n0+1−(Rf
n0 + 1(cid:1) = G
n0 = G
n0+1−Rf0
(1)
n0+1, thus Rf
n0+1−1−Rf0
n0+f0 implying Rf−f0
(f0)
hand, Rf0−1
(1)
(1)
n0 = G
n0+1, which implies Rf0
n0 +1) = G
n0+1;
n0+f0 − 1. As
n0+f0 − Rf0
(f0)
and on the other hand, R0
n0 = G
(f0−1)
n0+f0 − G
(1)
(f0)
n0+f0 < 1 by definition of f0 (and n0), we can derive that
we know that G
n0+1 = G
n0 + 1) > Rf−f0
n0 + 1). We have assumed that Tn0 succeeds at its f th retry, which
n0+1 − (Rf
Rf
n0 + 1. The previous inequality states then that Tn0+1 cannot be successful at its f th
will end at Rf
retry, since either a thread succeeds before Tn0+f0 and makes both Tn0+f0 and Tn0+1 fail, or Tn0+f0
succeeds and makes Tn0+1 fail. We have shown that S is not a well-formed seed.
Lemma 2. Assuming r 6= 0, if a new thread is added to an (f, P)-cyclic execution, it will eventually
succeed.
n0+f0 − (Rf
Proof:
Let R0
P be the time of the 0th retry of the new thread, that we number TP . If this retry is successful,
we are done; let us assume now that this retry is a failure, and let us shift the thread numbers (for
the threads T0, . . . , TP−1) so that T0 makes TP fail on its first attempt. We distinguish two cases,
(P )
0 > R0
depending on whether G
(P )
0 > R0
is increasing on (cid:74)0, P − 1(cid:75) and
P − S0}. For all k ∈(cid:74)0, n0(cid:75), we have
We assume that G
that G
P − Sk = k + R0
P −(G
P − Sk > 0 and Rk
P − S0 < 1.
Rk
This shows that T0, . . . , Tn0, because of their successes at S0, . . . , Sn0, successively make 0th, . . . ,
0 retries (respectively) of TP fail. The next attempt for TP is at Rn0+1
nth
, which fulfills the following
inequality: Rn0+1
0 = 0, hence let n0 = min{n ∈(cid:74)0, P − 1(cid:75) ; G
P − S0 or not.
P − S0. We know that n 7→ G
(n)
n < R0
− (Sn0 + 1) < Sn0+1 − (Sn0 + 1) since
(k)
k + S0 + k) = R0
P − Sk < R0
(k)
k hence Rk
P − S0− G
(n)
n
(0)
P
P
P
Rn0+1
Rn0+1
P
− Sn0+1 = (n0 + 1 + R0
− Sn0+1 > 0.
P ) − (G
(n0+1)
n0+1 + S0 + n0 + 1)
We consider now the reverse case by assuming that G
Tn0+1 should have been the successful thread, but TP starts a retry before Sn0+1, and is therefore
succeeding.
P − S0. With the previous line of
reasoning, we can show that T0, . . . , TP−1, because of their successes at S0, . . . , SP−1, successively
make 0th, . . . , (P − 1)th retries (respectively) of TP fail. Then we are back in the same situation
when T0 made TP fail for the first time (T0 makes TP fail), except that the success of T0 starts at
0 ≥ r.
0 = q + r + f − P > 0 and q, f and P are integers, we have that G
S0
(P )
(P )
(P )
0 = S0 + G
0
P − S0, which is absurd. S0 makes
(P )
By the way, if we had G
0 > r, we would have G
(P )
(P )
indeed R0
0 = r.
P fail, therefore G
0
We define
should be less than 1. Consequently, we are ensured that G
0 ≥ 1 + r > R0
(P )
(P )
0 < R0
. As G
$
k0 =
P − S0
R0
r
%
;
also, for every k ∈(cid:74)1, k0(cid:75), r < R0
P − (S0 + (k0 + 1) × r): the cycle of
successes of T0, . . . , TP−1 is executed k0 times. Then the situation is similar to the first case, and
TP will succeed.
P − (S0 + k × r) and r > R0
Lemma 3. Let S be a weakly-formed seed, and f = f
> 1, and if the
second success of TP−1 does not occur before the second success of Tf−1, then we can find in the
execution a well-formed seed S0 for P threads such that f (S0) = f.
(f+1)
f
(cid:16)(Ti, Si)i∈(cid:74)0,P−2(cid:75)
(cid:17). If G
Proof:
T0
T1
T2
T3
Figure 7: Lemma 3 configuration
12
Let us first remark that, by the definition of a weakly-formed seed, all threads will succeed once,
in order. Then two ordered groups of threads will compete for each of the next successes, until Tf−1
succeeds for the second time.
second success of Tf−1. Let then S1 and S2 be the two groups of threads that are in competition,
defined by
Let e be the smallest integer of (cid:74)f, P − 1(cid:75) such that the second success of Te occurs after the
S1 = {Tn ; n ∈(cid:74)0, f − 1(cid:75)}
S2 = {Tn ; n ∈(cid:74)f, e − 1(cid:75)}
(
For all n ∈(cid:74)0, e − 1(cid:75), we note
We define σ, a permutation of (cid:74)0, e − 1(cid:75) that describes the reordering of the threads during the
round of the second successes, such that, for all (i, j) ∈ (cid:74)0, e − 1(cid:75)2, σ (i) < σ (j) if and only if
if Tn ∈ S1
if Tn ∈ S2
rank (n) =
(n+1)
n
(n+1)
n
− 1
G
G
rank (i) < rank (j).
.
We also define a function that will help in expressing the σ−1 (k)'s:
(cid:74)f, e − 1(cid:75)
(cid:74)0, e − 1(cid:75) −→
m2 :
We note that rank(cid:12)(cid:12)(cid:74)0,f−1(cid:75) is increasing, as well as rank(cid:12)(cid:12)(cid:74)f,e−1(cid:75). This shows that #{T' ∈ S2 ; σ (') ≤
7−→ max {' ∈(cid:74)f, e − 1(cid:75) ; T' ∈ S2 ; σ (') ≤ k} .
k} = m2 (k) − (f − 1). Consequently, if Tσ−1(k) ∈ S2, then
k
m2 (k) = #{T' ∈ S2 ; σ (') ≤ k} + f − 1
= #{T' ∈ S2 ; ' ≤ σ−1 (k)} + f − 1
= σ−1 (k) − f + 1 + f − 1
m2 (k) = σ−1 (k) .
in S2, hence
σ−1 (k) = k + 1 − (m2 (k) − f + 1) − 1 = f + k − m2 (k) − 1.
Conversely, if Tσ−1(k) ∈ S1, among {Tσ(n) ; n ∈(cid:74)0, k(cid:75)}, there are exactly m2 (k) − f + 1 threads
In both cases, among {Tσ(n) ; n ∈ (cid:74)0, k(cid:75)}, there are exactly m2 (k) − f + 1 threads in S2, and
m1 (k) = k − (m2 (k) − f) threads in S1.
We prove by induction that after this first round, the next successes will be, respectively, achieved
by Tσ−1(0), Tσ−1(1), . . . , Tσ−1(e−1). In the following, by "kth success", we mean kth success after the
first success of TP−1, starting from 0, and the Rj
i 's denote the attempts of the second round.
Let (PK) be the following property: for all k ≤ K, the kth success is achieved by Tσ−1(k) at
Rf+k−σ−1(k)
. We assume (PK) true, and we show that the (K +1)th success is achieved by Tσ−1(K+1)
σ−1(k)
at Rf+K+1−σ−1(K+1)
σ−1(K+1)
We first show that if Tσ−1(K) ∈ S1, then
Rm1(K)−1
m2(K)+1 > Rf+K−σ−1(K)
> Rm1(K)
m2(K).
σ−1(K)
(6)
.
13
On the one hand,
Rf+K−σ−1(K)
σ−1(K)
Rf+K−σ−1(K)
σ−1(K)
On the other hand,
= K − σ−1 (K) + Rf
= K − σ−1 (K) + Rf
= K + SP−1 + 1 + G
= K + SP−1 + 1 + G
σ−1(K)
0 + σ−1 (K) + G
(σ−1(K))
(1)
0 + G
σ−1(K)
(σ−1(K)+1)
σ−1(K)
.
(σ−1(K))
σ−1(K)
f
+ G
= (m2 (K) − f) + RK−(m2(K)−f)
(m2(K)−f)
m2(K)
(m2(K)−f)
= (m2 (K) − f) + K − (m2 (K) − f) + R0
m2(K)
= (m2 (K) − f) + K − (m2 (K) − f) + SP−1 + 1 + (G
= K + SP−1 + 1 + G
f + G
− 1.
(m2(K)+1)
m2(K)
(f+1)
f
− 1) + G
(m2(K)−f)
m2(K)
Rf+K−m2(K)
m2(K)
Rf+K−m2(K)
m2(K)
Therefore,
Rf+K−σ−1(K)
σ−1(K)
− Rm1(K)
Rf+K−σ−1(K)
σ−1(K)
− Rm1(K)
−(cid:16)
m2(K) = Rf+K−σ−1(K)
σ−1(K)
(σ−1(K)+1)
σ−1(K)
m2(K) = rank(cid:16)
− Rf+K−m2(K)
G
− 1(cid:17)
σ−1 (K)(cid:17) − rank (m2 (K)) .
m2(K)
(m2(K)+1)
m2(K)
= G
In a similar way, we can obtain that if Tσ−1(K) ∈ S2, then
m1(K) > Rf+K−σ−1(K)
Rm2(K)
(7)
In addition, we recall that if Tσ−1(K) ∈ S2, σ−1 (K) = m2 (K), thus the second inequality of
Equation 6 becomes an equality, and if Tσ−1(K) ∈ S1, σ−1 (K) = f + K − m2 (K) − 1, hence the
second inequality of Equation 7 becomes an equality.
Now let us look at which attempt of other threads Tσ−1(K) made fail. From now on, and until
explicitly said otherwise, we assume that Tσ−1(K) ∈ S1. According to Equation 6, we have
> Rm2(K)+1
m1(K)−1.
σ−1(K)
Rm1(K)−1
m2(K)+1 >
m2(K)+j − Rm1(K)−1
Rm1(K)−j
m2(K)+1 <
(j−1)
m2(K)+j <
G
Rf+K−σ−1(K)
σ−1(K)
m2(K)+j − Rf+K−σ−1(K)
Rm1(K)−j
Rm1(K)−j
m2(K)+j − Rf+K−σ−1(K)
σ−1(K)
>
<
<
Rm1(K)
m2(K)
Rm1(K)−j
m2(K)+j − Rm1(K)
m2(K)
(j)
m2(K)+j
G
This holds for every j ∈(cid:74)1, m1 (K)(cid:75), implying j ≤ f, since there could not be more than f threads
in S1. Therefore, as by assumptions gaps of at most f th order are between 0 and 1,
σ−1(K)
0 < Rm1(K)−j
m2(K)+j − Rf+K−σ−1(K)
σ−1(K)
< 1;
14
showing that the success of Tσ−1(K) makes thread Tm2(K)+j fail on its attempt at Rm1(K)−j
j ∈(cid:74)1, m1 (K)(cid:75).
Since Tσ−1(K) ∈ S1, σ−1 (K) = m1 (K) − 1. Also, for all j ∈(cid:74)0, f − 1 − m1 (K)(cid:75),
m2(K)+j, for all
Rm2(K)−j
m1(K)+j − Rf+K−σ−1(K)
σ−1(K)
Rm2(K)−j
m1(K)+j − Rf+K−σ−1(K)
σ−1(K)
=(cid:16)
= Rm2(K)−j
m1(K)+j − Rm2(K)+1
m1(K)−1
Rm2(K)−j
m1(K)−1 + (j + 1) + G
(j+1)
m1(K)+j
= G
(cid:17) −(cid:16)
m1(K)−1 + (j + 1)(cid:17)
Rm2(K)−j
(j+1)
m1(K)+j
.
m1(K)+j
m2(K)+j
m1(K)+j, for all j ∈(cid:74)0, f − 1 − m1 (K)(cid:75),
As a result, Tσ−1(K) makes Tm1(K)+j fail on its attempt at Rm2(K)−j
and the next attempt will occur at Rm2(K)−j+1
Altogether, the next attempt after the end of the success of Tσ−1(K) for Tm1(K)+j is Rm2(K)−j+1
, for all j ∈(cid:74)1, m1 (K)(cid:75).
for j ∈(cid:74)0, f − 1 − m1 (K)(cid:75), and for Tm2(K)+j is Rm1(K)−j+1
Additionally, a thread will begin a new retry loop, the 0th retry being at R0
m2(K)+m1(K)+1 =
f+K+1. We note that f + K + 1 could be higher than P − 1, referring to a thread whose number
R0
n refers to the jth retry of Trank(n−P +1), after its first
is more than P − 1. Actually, if n > P − 1, Rj
indices, of S1 ∩ σ−1 ((cid:74)K + 1, e − 1(cid:75)) and S2 ∩
two successes.
σ−1 ((cid:74)K + 1, e − 1(cid:75)) will
(cid:74)0, f − 1 − m1 (K)(cid:75),
thus if someone succeeds in S1, it will be Tm1(K). In the same way, for all j ∈(cid:74)1, m1 (K) + 1(cid:75),
(j−1)
m2(K)+j > 0,
meaning that if someone succeeds in S2, it will be Tm2(K)+1.
the two smallest
then compete for being successful.
Rm1(K)−j+1
m2(K)+j − Rm1(K)
Rm2(K)−j+1
m1(K)+j − Rm2(K)+1
m1(K) = G
Indeed, within S1,
The two heads,
(j)
m1(K)+j > 0,
m2(K)+1 = G
for j ∈
m1(K)+j
i.e.
,
Let us compare now those two candidates:
m1(K) − Rm1(K)
Rm2(K)+1
m2(K)+1 = m2 (K) + 1 − f + SP−1 + m1 (K) + G
−(cid:16)
−(cid:16)
(cid:17)
(m1(K)+1)
m1(K)
f + m2 (K) + 1 − f + G
(cid:17)
(m2(K)+1−f)
m2(K)+1
(m2(K)+1−f)
m2(K)+1
m1 (K) + R0
= SP−1 − 1 + G
SP−1 + G
(m1(K)+1)
m1(K)
(f+1)
f
−(cid:16)
G
− 1 + G
m2(K)+1 − 1(cid:17)
(m2(K)+2)
= G
(m1(K)+1)
m1(K)
m1(K) − Rm1(K)
Rm2(K)+1
m2(K)+1 = rank (m1 (K)) − rank (m2 (K) + 1) .
By definition, σ−1 (K + 1) is either m1 (K) or m2 (K) + 1 and corresponds to the next successful
thread. We can follow the same line of reasoning in the case where Tσ−1(K) ∈ S2 and prove in this
way that (PK+1) is true.
(P0) is true, and the property spreads until (Pe−1), where all threads of S1 and S2 have been
successful, in the order ruled by σ−1, i.e. Tσ−1(0), . . . , Tσ−1(e−1). And before those successes the
threads Te−1 =Tσ−1(e−1), . . . , TP−1 have been successful as well. The seed composed of those successes
is a well-formed seed. Given a thread, the gap between this thread and the next one in the new order
could indeed not be higher than the gap in the previous order with its next thread. Also the f th
order gaps remain smaller than 1. And as Te−1 succeeds the second time after f failures, it means
that the new seed S00 is such that f (S00) = f.
15
T0
T1
T2
T3
Figure 8: Lemma 4 configuration
Lemma 4. Let S be a weakly-formed seed, and f = f
> 1 and if the
second success of TP−1 occurs before the second success of Tf−1, then we can find in the execution a
well-formed seed S0 for P threads such that f (S0) = f.
(f+1)
f
(cid:16)(Ti, Si)i∈(cid:74)0,P−2(cid:75)
(cid:17). If G
f−1 − SP−1 = G
Proof: Until the second success of TP−1, the execution follows the same pattern as in Lemma 3.
Actually, the case invoked in the current lemma could have been handled in the previous lemma,
but it would have implied tricky notations, when we referred to Trank(n−P +1). Let us deal with this
case independently then, and come back to the instant where TP−1 succeeds for the second time.
f−1 < 1. For the thread Tσ(j) to succeed at its kth retry after
(f)
We had 0 < R0
the first success of TP−1 and before Tf−1, it should necessary fill the following condition: j + 1 <
f−1. This holds also for the second success of TP−1, which implies that
σ(j) − SP−1 < j + 1 + G
(f)
Rk
f−1, where h is the number of failures of TP−1 before its
P 0 < SP−1 + 1 + q + r + h − SP−1 < P 0 + G
(f)
second success and P 0 is the number of successes between the two successes of TP−1. As G
(f)
f−1 < 1,
and q, P 0 and h are non-negative integers, we have r < G
To conclude, as any gap at any order is less than the gap between the two successes of TP−1,
f−1 and h = P 0 − 1 − q.
(f)
(1)
(n)
n = r. With the new thread, the first order gaps are changed by decomposing G
0
formed seed succeeded previously as T0, . . . , TP 0−1. As explained before, for all (k, n) ∈(cid:74)0, P 0 − 1(cid:75)2,
which is r < 1, we found a well-formed seed for P 0 threads.
Finally any other thread will eventually succeed (see Lemma 2). We can renumber the threads such
that TP 0 is the first thread that is not in the well-formed seed to succeed, and the threads of the well-
(k)
into
G
n < G
(1)
(1)
0 . All gaps can only be decreased, hence we have a new well-formed seed for
P 0 and the new G
G
P 0 + 1 threads. We repeat the process until all threads have been encountered, and obtain in the
end S0, a well-formed seed with P threads such that f (S0) = P − 1 − q, which is an optimal cyclic
execution.
Still, as Tf succeeds between two successes of TP−1 that are separated by r, we had, in the initial
(1)
(f)
f < 1, we conclude that
f−1 < 1 and G
configuration: G
(f+1)
the lagging time was initially less than 2+ r. By hypothesis, we know that G
> 1, which implies
f
that, before the entry of the new thread, the lagging time was 1 + r. In the final execution with one
more thread, the lagging time is r and we have one more success in the cycle, thus f (S0) = f.
Theorem 2. Assuming r 6= 0, if a new thread is added to an (f, P − 1)-cyclic execution, then all the
threads will eventually form either an (f, P)-cyclic execution, or an (f + 1, P)-cyclic execution.
< r. As, in addition, we have both G
(P−1−f)
P−1
Proof: According to Lemma 2, the new thread will eventually succeed. In addition, we recall
that Properties 1 and 2 ensure that before the first success of the new thread, any set of P − 1
16
consecutive successes is a well-formed seed with P − 1 threads. We then consider a seed (we number
the threads accordingly, and number the new thread as TP−1) such that the success of the new
thread occurs between the success of TP−2 and T0; we obtain in this way a weakly-formed seed
S = (Tn, Sn)n∈(cid:74)0,P−1(cid:75)&. We differentiate between two cases.
Firstly, if for all n ∈ (cid:74)0, P − 1(cid:75), G
< 1, according to Lemma 1, we can find later in the
execution a well-formed seed S0 for P threads such that f (S0) = f + 1, hence we reach eventually
Let us assume now that this condition is not fulfilled. There exists n0 ∈ (cid:74)0, P − 1(cid:75) such that
an (f + 1, P)-cyclic execution.
(f+1)
> 1.
G
n0
Then two cases are feasible. If the second success of TP−1 occurs before the second success of Tf−1,
then Lemma 3 shows that we will reach an (f, P)-cyclic execution. Otherwise, from Lemma 3, we
conclude that an (f, P)-cyclic execution will still occur.
> 1. We shift the thread numbers, such that n0 is now f, and we have then G
(f+1)
f
(f+1)
n
C. Throughput Bounds
Firstly we calculate the expression of throughput and the expected number of threads inside the
retry loop (that is needed when we gather expansion and wasted retries). Then we exhibit upper
and lower bounds on both throughput and the number of failures, and show that those bounds are
reached. Finally, we give the worst case on the number of wasted retries.
Lemma 5. In an (f, P)-cyclic execution, the throughput is
T =
P
q + r + 1 + f
.
(8)
Proof: By definition, the execution is periodic, and the period lasts q + r + 1 + f units of time.
As P successes occur during this period, we end up with the claimed expression.
Lemma 6. In an (f, P)-cyclic execution, the average number of threads Prl in the retry loop is given
by
Prl = P ×
f + 1
q + r + f + 1 .
Proof: Within a period, each thread spends f + 1 units of time in the retry loop, among the
q + r + f + 1 units of time of the period, hence the Lemma.
Lemma 7. The number of failures is not less than f (-), where
(
f (-) =
P − q − 1 if q ≤ P − 1
0
otherwise
, and accordingly,
T ≤
( P
P +r
P
q+r+1
if q ≤ P − 1
otherwise.
(9)
Proof: According to Equation 8, the throughput is maximized when the number of failures
is minimized. In addition, we have two lower bounds on the number of failures: (i) f ≥ 0, and
(ii) P successes should fit within a period, hence q + 1 + f ≥ P. Therefore, if P − 1 − q < 0,
T ≤ P/(q + r + 1 + 0), otherwise,
T ≤
= P
P
q + r + 1 + P − 1 − q
.
P + r
Remark 3. We notice that if q > P − 1, the upper bound in Equation 9 is actually the same as
the immediate upper bound described in Section III-B1. However, if q ≤ P − 1, Equation 9 refines
the immediate upper bound.
Lemma 8. The number of failures is bounded by
f ≤ f (+) =
(P − 1 − q − r) +
(cid:18)
(cid:22)1
2
q
(P − 1 − q − r)2 + 4P
17
(cid:19)(cid:23)
, and accordingly,
the throughput is bounded by
T ≥
P
q + r + 1 + f (+) .
also have G
(P )
n = ' + r.
Proof: We show that a necessary condition so that an (f, P)-cyclic execution, whose lagging
time is ', exists, is f × (' + r) < P. According to Property 1, any set of P consecutive successes
is a well-formed seed with P threads. Let S be any of them. As we have f failures before success,
Theorem 1 ensures that for all n ∈(cid:74)0, P − 1(cid:75), G
n < 1. We recall that for all n ∈(cid:74)0, P − 1(cid:75), we
nX
P−1X
= f × P−1X
On the one hand, we have
P−1X
(1)
j mod P
G(f)
n =
j=n−f+1
G(1)
n=0
n=0
(f)
G
On the other hand, PP−1
n=0 G
n
n=0
n = f × (' + r).
G(f)
P−1X
n <PP−1
n=0
(f)
n=0 1 = P.
Altogether, the necessary condition states that f × (' + r) < P, which can be rewritten as f ×
(q + 1 + f − P + r) < P. The proof is complete since minimizing the throughput is equivalent to
maximizing the number of failures.
Lemma 9. For each of the bounds defined in Lemmas 7 and 8, there exists an (f, P)-cyclic execution
that reaches the bound.
Proof: According to Lemmas 7 and 8, if an (f, P)-cyclic execution exists, then the number of
failures is such that f (-) ≤ f ≤ f (+).
We show now that this double necessary condition is also sufficient. We consider f such that f (-) ≤
f ≤ f (+), and build a well-formed seed S = (Ti, Si)i∈(cid:74)0,P−1(cid:75).
For all n ∈(cid:74)0, P − 1(cid:75), we define Si as
(cid:18) q + 1 + f − P + r
(cid:18) q + 1 + f − P + r
We first show that f (S) = f. By definition, f (S) = max (0,dSP−1 − S0 − q − re); we have then
Sn = n ×
(cid:25)(cid:19)
(cid:19)
(cid:19)
+ 1
P
.
(cid:24)
(cid:24)
(cid:24)
(cid:18)
(cid:18)
(cid:18)
0,
0,
0,
f (S) = max
= max
f (S) = max
P
+ 1
(P − 1) ×
(P − 1 − q − r) + (q + 1 + f − P + r) − q + 1 + f − P + r
f − q + 1 + f − P + r
− q − r
(cid:25)(cid:19)
P
.
(cid:25)(cid:19)
P
Firstly, we know that q + 1 + f − P ≥ 0, thus if f = 0, then the second term of the maximum is not
positive, and f (S) = 0 = f. Secondly, if f > 0, then according to Lemma 7, (q +1+ f − P + r)/P <
1/f ≤ 1. As we also have (q +1+ f − P + r)/P ≥ 0, we conclude that f (S) =l
f − q+1+f−P +r
m = f.
P
18
G(f)
n =
Additionally, for all n ∈(cid:74)0, P − 1(cid:75),
(
n ×(cid:16) q+1+f−P +r
n ×(cid:16) q+1+f−P +r
(
=
P
P
Sn − Sn−f − f
if n > f
Sn − SP +n−f + 1 + q + r otherwise
+ 1(cid:17) − (n − f) ×(cid:16) q+1+f−P +r
+ 1(cid:17) − (P + n − f) ×(cid:16) q+1+f−P +r
+ 1(cid:17) − f
+ 1(cid:17) + 1 + q + r
P
=
f × q+1+f−P +r
−(P − f) − (q + 1 + f − P + r) + f × q+1+f−P +r
P
+ 1 + q + r
P
P
n = f × w + r
G(f)
P
As w ≤ 0 and f ≤ 0, G
seed that leads to an (f, P)-cyclic execution.
n < 1. Theorem 1 implies that S is a well-formed
(f)
We have shown that for all f such that f (-) ≤ f ≤ f (+) there exists an (f, P)-cyclic execution; in
n > 0. Since f ≤ f (+), G
(f)
particular there exist an (f (+), P)-cyclic execution and an (f (-), P)-cyclic execution.
Corollary 1. The highest possible number of wasted repetitions is l√
P = q + 1.
Proof:
P − 1m and is achieved when
− f (-)(cid:23)
.
The highest possible number of wasted repetitions w(P) with P threads is given by
Let a and h be the functions respectively defined as a(P) = q+1−P +r, which implies a0(P) = −1,
Let us first assume that a(P) > 0. In this case, q ≤ P − 1, hence f (-) = 0. We have
q
(cid:22)1
w(P) = f (+) − f (-) =
(cid:18)
−a(P) +
(cid:19)
and h(P) = (−a(P) +pa(P)2 + 4P)/2 − f (-), so that w(P) = bh(P)c.
2pa(P)2 + 4P
2h0(P) = 1 + −2a(P) + 4
2h0(P) = 2 × 2 − a(P) +pa(P)2 + 4P
2pa(P)2 + 4P
a(P)2 + 4P
2
Therefore, h0(P) is negative if and only if pa(P)2 + 4P < a(P) − 2. It cannot be true if a(P) < 2.
h0(P) = (cid:16)
If a(P) ≥ 2, then the previous inequality is equivalent to a(P)2 + 4P < a(P)2 − 4a(P) + 4, which
can be rewritten in q + 1 + r < 1, which is absurd. We have shown that h is increasing in ]0, q + 1].
Let us now assume that a(P) ≤ 0. In this case, q > P − 1, hence f (-) = P − q − 1, and
/2 − r. Assuming h0(P) to be positive leads to the same absurd
inequality q + 1 + r < 1, which proves that h is decreasing on [q + 2, +∞[.
Also, the maximum number of wasted repetitions is achieved as P = q + 1 or P = q + 2. Since
the maximum number of wasted repetitions is w(q + 1). In addition,
√
√
(cid:16)−(r + 1) +
(cid:16)−r +
r2 + 4P
r2 +
(cid:17) = h(q + 2),
(cid:17)
√
4P
(cid:17)
>
1
2
r2 + 4P
√
(cid:17)
4P
< h(q + 1) <
√
P − r
2 < h(q + 1) <
√
P − 1 ≤ h(q + 1) <
1
2
√
P
√
P
(cid:17)
a(P) +pa(P)2 + 4P
(cid:16)−r +
h(q + 1) = 1
2
(cid:16)−r +
√
1
2
We conclude that the maximum number of wasted repetitions is l√
P − 1m.
19
V. Expansion and Complete Throughput Estimation
A. Expansion
Interference of threads does not only lead to logical conflicts but also to hardware conflicts which
impact the performance significantly. We model the behavior of the cache coherency protocols which
determine the interaction of overlapping Reads and CASs. By taking MESIF [GH09] as basis, we
come up with the following assumptions. When executing an atomic CAS, the core gets the cache line
in exclusive state and does not forward it to any other requesting core until the instruction is retired.
Therefore, requests stall for the release of the cache line which implies serialization. On the other
hand, ongoing Reads can overlap with other operations. As a result, a CAS introduces expansion
only to overlapping Read and CAS operations that start after it, as illustrated in Figure 4. As a
remark, we ignore memory bandwidth issues which are negligible for our study.
Furthermore, we assume that Reads that are executed just after a CAS do not experience expansion
(as the thread already owns of the data), which takes effect at the beginning of a retry following a
failing attempt. Thus, read expansions need only to be considered before the 0th retry. In this sense,
read expansion can be moved to parallel section and calculated in the same way as CAS expansion
is calculated.
To estimate expansion, we consider the delay that a thread can introduce, provided that there is
already a given number of threads in the retry loop. The starting point of each CAS is a random
variable which is distributed uniformly within an expanded retry. The cost function d provides the
amount of delay that the additional thread introduces, depending on the point where the starting
point of its CAS hits. By using this cost function we can formulate the expansion increase that each
new thread introduces and derive the differential equation below to calculate the expansion of a
CAS.
Lemma 10. The expansion of a CAS operation is the solution of the following system of equations:
e0 (Prl) = cc ×
(cid:17) = 0
(cid:16)
P
e
(0)
rl
cc
2 + e (Prl)
rc + cw + cc + e (Prl)
,
(0)
rl
where P
expansion begins.
is the point where
Proof:
We compute e (Prl + h), where h ≤ 1, by assuming that there are already Prl threads in the retry
loop, and that a new thread attempts to CAS during the retry, within a probability h.
20
e (Prl + h) = e (Prl) + h×
0
Z rlw(+)
d (t)
= e (Prl) + h×(cid:16)Z rc+cw−cc
rlw(+) dt
d (t)
Z rc+cw
rlw(+) dt
d (t)
Z rc+cw+e(Prl)
rlw(+) dt
d (t)
Z rlw(+)
rlw(+) dt
d (t)
= e (Prl) + h×(cid:16)Z rc+cw
rc+cw+e(Prl)
rc+cw−cc
0
+
rc+cw
+
+
rlw(+) dt(cid:17)
rlw(+) dt(cid:17)
cc
t
Z rc+cw+e(Prl)
rlw(+) dt
rc+cw−cc
+
2 + e (Prl) × cc
e (Prl + h) = e (Prl) + h× cc2
rc+cw
This leads to
obtain
e (Prl + h) − e (Prl)
h
=
e0 (Prl) = cc ×
rlw(+)
2 + e (Prl) × cc
cc2
rlw(+)
cc
2 + e (Prl)
rc + cw + cc + e (Prl) .
. When making h tend to 0, we finally
B. Throughput Estimate
There remains to combine hardware and logical conflicts in order to obtain the final upper and
lower bounds on throughput. We are given as an input an expected number of threads Prl inside
the retry loop. We firstly compute the expansion accordingly, by solving numerically the differential
equation of Lemma 10. As explained in the previous subsection, we have pw(+) = pw + e, and
rlw(+) = rc + cw + e + cc. We can then compute q and r, that are the inputs (together with the
total number of threads P) of the method described in Section IV. Assuming that the initialization
times of the threads are spaced enough, the execution will superimpose an (f, P)-cyclic execution.
Thanks to Lemma 6, we can compute the average number of threads inside the retry loop, that we
note by hf(Prl). A posteriori, the solution is consistent if this average number of threads inside the
retry loop hf(Prl) is equal to the expected number of threads Prl that has been given as an input.
Several (f, P)-cyclic executions belong to the domain of the possible outcomes, but we are
interested in upper and lower bounds on the number of failures f. We can compute them through
Lemmas 7 and 8, along with their corresponding throughput and average number of threads inside
the retry loop. We note by h(+)(Prl) and h(-)(Prl) the average number of threads for the lowest
number of failures and highest one, respectively. Our aim is finally to find P
, such that
h(+)(P
. If several solutions exist, then we want to keep the smallest,
since the retry loop stops to expand when a stable state is reached.
Note that we also need to provide the point where the expansion begins. It begins when we start to
have failures, while reducing the parallel section. Thus this point is (2P −1)rlw(-) (resp. (P −1)rlw(-))
for the lower (resp. upper) bound on the throughput.
(+)
rl and h(-)(P
(-)
rl and P
(+)
rl ) = P
(-)
rl ) = P
(-)
rl
(+)
rl
21
Theorem 3. Let (xn) be the sequence defined recursively by x0 = 0 and xn+1 = h(+)(xn). If pw ≥
rc + cw + cc, then
(+)
rl = lim
P
n→+∞ xn.
Proof: First of all, the average number of threads belongs to ]0, P[, thus for all x ∈ [0, P],
0 < h(+)(x) < P. In particular, we have h(+)(0) > 0, and h(+)(P) < P, which proves that there exist
one fixed point for h(+).
In addition, we show that h(+) is a non-decreasing function. According to Lemma 6,
1 + f (-)
q + r + f (-) + 1 ,
h(+)(Prl) = P ×
(cid:23)
(cid:22) pw + e
where all variables except P depend actually on Prl. We have
and r = pw + e
rlw(-) + e
rlw(-) + e
q =
− q,
hence, if pw ≥ rlw(-), q and r are non-increasing as e is non-decreasing, which is non-decreasing
with Prl. Since f (-) is non-decreasing as a function of q, we have shown that if pw ≥ rlw(-), h(+) is
a non-decreasing function.
Finally, the proof is completed by the theorem of Knaster-Tarski.
The same line of reasoning holds for h(-) as well. As a remark, w point out that when pw < rlw(-),
we scan the interval of solution, and have no guarantees about the fact that the solution is the
smallest one; still it corresponds to very extreme cases.
C. Several Retry Loops
We consider here a lock-free algorithm that, instead of being a loop over one parallel section and
one retry loop, is composed of a loop over a sequence of alternating parallel sections and retry loops.
We show that this algorithm is equivalent to an algorithm with only one parallel section and one
retry loop, by proving the intuition that the longest retry loop is the only one that fails and hence
expands.
1) Problem Formulation: In this subsection, we consider an execution such that each spawned
thread runs Procedure Combined in Figure 9. Each thread executes a linear combination of S
independent retry loops, i.e. operating on separate variables, interleaved with parallel sections.
We note now as rlw(+)
parallel section, respectively, for each i ∈ (cid:74)1, S(cid:75). As previously, qi and ri are defined such that
the size of a retry of the ith retry loop and the size of the ith
i = (qi + ri) × rlw(+)
, where qi is a non-negative integer and ri is smaller than 1.
and pw(+)
pw(+)
i
i
i
Procedure Combined
1 Initialization();
2 while ! done do
3
4
5
6
7
8
for i ← 1 to S do
Parallel_Work(i);
while ! success do
current ← Read(AP[i]);
new ← Critical_Work(i,current);
success ← CAS(AP, current, new);
Figure 9: Thread procedure with several retry loops
The Procedure Combined executes the retry loops and parallel sections in a cyclic fashion, so we
can normalize the writing of this procedure by assuming that a retry of the 1st retry loop is the
longest one. More precisely, we consider the initial algorithm, and we define i0 as
22
i0 = min argmaxi∈(cid:74)1,S(cid:75) rlw(+)
i
.
We then renumber the retry loops such that the new ordering is i0, . . . , S, 1, . . . , i0 − 1, and we add
in Initialization the first parallel sections and retry loops on access points from 1 to i0 -- according
to the initial ordering.
One success at the system level is defined as one success of the last CAS, and the throughput is
defined accordingly. We note that in steady-state, all retry loops have the same throughput, so the
throughput can be computed from the throughput of the 1st retry loop instead.
Lemma 11. Unsuccessful retry loops can only occur in the 1st retry loop.
2) Wasted Retries:
Proof:
iX
i0=1
iX
i0=1
We note (tn)n∈[1,+∞[ the sequence of the thread numbers that succeeds in the 1st retry loop, and
(sn)n∈[1,+∞[ the sequence of the corresponding time where they exit the retry loop. We notice that
by construction, for all n ∈ [1, +∞[, sn < sn+1. Let, for i ∈(cid:74)2, S(cid:75) and n ∈ [1, +∞[, (Pi,n) be the
following property: for all i0 ∈(cid:74)2, i(cid:75), and for all n0 ∈(cid:74)1, n(cid:75), the thread Ttn0 succeeds in the ith retry
loop at its first attempt.
We assume that for a given (i, n), (Pi+1,n) and (Pi,n+1) is true, and show that (Pi+1,n+1) is true.
As the threads Ttn and Ttn+1 do not have any failure in the first i retry loops, their entrance time in
the i + 1th retry loop is given by
sn +
(rlw(+)
i0 + pw(+)
i0 ) + pw(+)
i+1 = X1 and sn+1 +
(rlw(+)
i0 + pw(+)
i0 ) + pw(+)
i+1 = X2,
respectively. Thread Ttn does not fail in the i + 1th retry loop, hence exits at
1 = sn + X2 − sn+1 < X2.
i+1 < X1 + rlw(+)
X1 + rlw(+)
As the previous threads Tn−1, . . . ,T1 exits the ith retry loop before Tn, and next threads Tn0, where
n0 > n + 1, enters this retry loop after Tn+1, this implies that the thread Ttn+1 succeeds in the i + 1th
retry loop at its first attempt, and (Pi+1,n+1) is true.
retry loop since there is no other thread to compete with. Therefore, for all i ∈ (cid:74)2, S(cid:75), (Pi,1) is
Regarding the first thread that succeeds in the first retry loop, we know that he successes in any
true. Then we show by induction that all (P2,n) is true, then all (P3,n), etc., until all (PS,n), which
concludes the proof.
Theorem 4. The multi-retry loop Procedure Combined is equivalent to the Procedure Abstract-
Algorithm, where
pw(+) = pw(+)
1 +
i + rlw(+)
i
and rlw(+) = rlw(+)
1 .
(cid:16)pw(+)
SX
i=2
(cid:17)
Proof: According to Lemma 11 there is no failure in other retry loop than the first one; therefore,
all retry loops have a constant duration, and can thus be considered as parallel sections.
23
3) Expansion: The expansion in the retry loop starts as threads fail inside this retry loop. When
threads are launched, there is no expansion, and Lemma 11 implies that if threads fail, it should be
inside the first retry loop, because it is the longest one. As a result, there will be some stall time
in the memory accesses of this first retry loop, i.e. expansion, and it will get even longer. Failures
will thus still occur in the first retry loop: there is a positive feedback on the expansion of the first
retry loop that keeps this first retry loop as the longest one among all retry loops. Therefore, in
accordance to Theorem 4, we can compute the expansion by considering the equivalent single-retry
loop procedure described in the theorem.
VI. Experimental Evaluation
We validate our model and analysis framework through successive steps, from synthetic tests,
capturing a wide range of possible abstract algorithmic designs, to several reference implementations
of extensively studied lock-free data structure designs that include cases with non-constant parallel
section and retry loop.
A. Setting
We have conducted experiments on an Intel ccNUMA workstation system. The system is composed
of two sockets, that is equipped with Intel Xeon E5-2687W v2 CPUs with frequency band 1.2-3.4. GHz
The physical cores have private L1, L2 caches and they share an L3 cache, which is 25 MB. In a
socket, the ring interconnect provides L3 cache accesses and core-to-core communication. Due to
the bi-directionality of the ring interconnect, uncontended latencies for intra-socket communication
between cores do not show significant variability.
Our model assumes uniformity in the CAS and Read latencies on the shared cache line. Thus,
threads are pinned to a single socket to minimize non-uniformity in Read and CAS latencies. In the
experiments, we vary the number of threads between 4 and 8 since the maximum number of threads
that can be used in the experiments are bounded by the number of physical cores that reside in one
socket.
In all figures, y-axis provides the throughput, which is the number of successful operations
completed per millisecond. Parallel work is represented in x-axis in cycles. The graphs contain the
high and low estimates (see Section IV), corresponding to the lower and upper bound on the wasted
retries, respectively, and an additional curve that shows the average of them.
As mentioned before, the latencies of CAS and Read are parameters of our model. We used the
methodology described in [DGT13] to measure latencies of these operations in a benchmark program
by using two threads that are pinned to the same socket. The aim is to bring the cache line into the
state used in our model. Our assumption is that the Read is conducted on an invalid line. For CAS,
the state of the cache line could be exclusive, forward, shared or invalid. Regardless of the state
of the cache line, CAS requests it for ownership, that compels invalidation in other cores, which in
turn incurs a two-way communication and a memory fence afterwards to assure atomicity. Thus, the
latency of CAS does not show negligible variability with respect to the state of the cache line, as
also revealed in our latency benchmarks.
As for the computation cost, the work inside the parallel section is implemented by a dummy
for-loop of Pause instructions.
B. Synthetic Tests
1) Single retry loop: For the evaluation of our model, we first create synthetic tests that emulate
different design patterns of lock-free data structures (value of cw) and different application contexts
(value of pw). As described in the previous subsection, in the Procedure AbstractAlgorithm, the
24
Figure 10: Synthetic program
cw = 50, threads = 4cw = 50, threads = 6cw = 50, threads = 8cw = 100, threads = 4cw = 100, threads = 6cw = 100, threads = 8cw = 200, threads = 4cw = 200, threads = 6cw = 200, threads = 8cw = 600, threads = 4cw = 600, threads = 6cw = 600, threads = 8cw = 1600, threads = 4cw = 1600, threads = 6cw = 1600, threads = 840006000800010000120004000600080001000012000400060008000100001200050007000900011000500070009000110005000700090001100040006000800040006000800040006000800020003000400020003000400020003000400010001500100015001000150010002000300001000200030004000020004000600001000200030000200040000200040006000010002000300040005000020004000600002500500075001000002500500075001000005000100001500005000100001500020000050001000015000200000100002000030000010000200003000040000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageReal 25
amount of work in both the parallel section and the retry loop are implemented as dummy loops,
whose costs are adjusted through the number of iterations in the loop.
Generally speaking, in Figure 10, we observe two main behaviors: when pw is high, the data
structure is not contended, and threads can operate without failure. When pw is low, the data
structure is contended, and depending on the size of cw (that drives the expansion) a steep decrease
in throughput or just a roughly constant bound on the performance is observed.
The position of the experimental curve between the high and low estimates, depends on cw. It can
be observed that the experimental curve mostly tends upwards as cw gets smaller, possibly because
the serialization of the CASs helps the synchronization of the threads.
Another interesting fact is the waves appearing on the experimental curve, especially when the
number of threads is low or the critical work big. This behavior is originating because of the variation
of r with the change of parallel work, a fact that is captured by our analysis.
Figure 11: Multiple retry loops with 8 threads
2) Several retry loops: We have created experiments by combining several retry loops, each
operating on an independent variable which is aligned to a cache line. In Figure 11, results are
compared with the model for single retry loop case where the single retry loop is equal to the
longest retry loop, while the other retry loops are part of the parallel section. The distribution of
fails in the retry loops are illustrated and all throughput curves are normalized with a factor of 175
(to be easily seen in the same graph). Fails per success values are not normalized and a success is
obtained after completing all retry loops.
C. Treiber's Stack
The lock-free stack by Treiber [Tre86] is one of the most studied efficient data structures. Pop and
Push both contain a retry loop, such that each retry starts with a Read and ends with CAS on the
shared top pointer. In order to validate our model, we start by using Pops. From a stack which is
cw1=50cw1=200cw1=400cw1=10000102030400510150.02.55.07.5cw2=50cw2=400cw2=100002000400060000250050007500100000500010000150000100002000030000Parallel Work + Small Retry Loop (cycles)Norm. SuccessFails RL1/SuccessFails RL2/SuccessTotal Fails / SuccessLowHighAverage 26
Figure 12: Pop on Treiber's stack
initiated with 50 million elements, threads continuously pop elements for a given amount of time.
We count the total number of pop operations per millisecond. Each Pop first reads the top pointer
and gets the next pointer of the element to obtain the address of the second element in the stack,
before attempting to CAS with the address of the second element. The access to the next pointer of
the first element occurs in between the Read and the CAS. Thus, it represents the work in cw. This
memory access can possibly introduce a costly cache miss depending on the locality of the popped
element.
To validate our model with different cw values, we make use of this costly cache miss possibility.
We allocate a contiguous chunk of memory and align each element to a cache line. Then, we initialize
the stack by pushing elements from contiguous memory either with a single or large stride to disable
the prefetcher. When we measure the latency of cw in Pop for single and large stride cases, we obtain
the values that are approximately 50 and 300 cycles, respectively. As a remark, 300 cycles is the
cost of an L3 miss in our system when it is serviced from the local main memory module. To create
more test cases with larger cw, we extended the stack implementation to pop multiple elements with
a single operation. Thus, each access to the next element could introduce an additional L3 cache
miss while popping multiple elements. By doing so, we created cases in which each thread pops 2,
3, etc. elements, and cw goes to 600, 900, etc. cycles, respectively. In Figure 12, comparison of the
experimental results from Treiber's stack and our model is provided.
As a remark, we did not implemented memory reclamation for our experiments but one can
implement a stack that allows pop and push of multiple elements with small modifications using
hazard pointers [Mic04]. Pushing can be implemented in the same way as single element case. A
cw = 50, threads = 6cw = 300, threads = 6cw = 600, threads = 6cw = 900, threads = 6cw = 1200, threads = 6cw = 1500, threads = 640006000800010000120002000300040005000600070002000300040001000150020002500300010001500200025001000150020000100020003000400002500500075000500010000150000500010000150002000005000100001500020000250000100002000030000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageReal 27
if t = NULL then
return EMPTY;
Algorithm 1: Multiple Pop
1 Pop (multiple)
2 while true do
3
4
5
6
7
8
9
10
11
12
13
14 RetireNodes (t, multiple);
hp* = t;
if top != t then
break;
hp++;
next = t.next;
t = Read(top);
for multiple do
if CAS(&top, t, next) then
break;
Pop requires some modifications for memory reclamation. It can be implemented by making use of
hazard pointers just by adding the address of the next element to the hazard list before jumping
to it. Also, the validity of top pointer should be checked after adding the pointer to the hazard list
to make sure that other threads are aware of the newly added hazard pointer. By repeating this
process, a thread can jump through multiple elements and pop all of them with a CAS at the end.
D. Shared Counter
In [DLM13], the authors have implemented a "scalable statistics counters" relying on the following
idea: when contention is low, the implementation is a regular concurrent counter with a CAS; when
the counter starts to be contended, it switches to a statistical implementation, where the counter is
actually incremented less frequently, but by a higher value. One key point of this algorithm is the
switch point, which is decided thanks to the number of failed increments; our model can be used by
providing the peak point of performance of the regular counter implementation as the switch point.
We then have implemented a shared counter which is basically a Fetch-and-Increment using a CAS,
and compared it with our analysis. The result is illustrated in Figure 13, and shows that the parallel
section size corresponding to the peak point is correctly estimated using our analysis.
E. DeleteMin in Priority List
We have applied our model to DeleteMin of the skiplist based priority queue designed in [LJ13].
DeleteMin traverses the list from the beginning of the lowest level, finds the first node that is not
logically deleted, and tries to delete it by marking. If the operation does not succeed, it continues
with the next node. Physical removal is done in batches when reaching a threshold on the number of
deleted prefixes, and is followed by a restructuring of the list by updating the higher level pointers,
which is conducted by the thread that is successful in redirecting the head to the node deleted by
itself.
We consider the last link traversal before the logical deletion as critical work, as it continues
with the next node in case of failure. The rest of the traversal is attributed to the parallel section
as the threads can proceed concurrently without interference. We measured the average cost of a
traversal under low contention for each number of threads, since traversal becomes expensive with
28
(a) 4 threads
(b) 6 threads
(c) 8 threads
Figure 13: Increment on a shared counter
more threads. In addition, average cost of restructuring is also included in the parallel section since
it is executed infrequently by a single thread.
We initialize the priority queue with a large set of elements. As illustrated in Figure 14, the smallest
pw value is not zero as the average cost of traversal and restructuring is intrinsically included. The
peak point is in the estimated place but the curve does not go down sharply under high contention.
This presumably occurs as the traversal might require more than one steps (link access) after a failed
attempt, which creates a back-off effect.
F. Enqueue-Dequeue on a Queue
In order to demonstrate the validity of the model with several retry loops, and that the results
covers a wider spectrum of application and designs from the ones we focused in our model, we studied
the following setting: the threads share a queue, and each thread enqueues an element, executes the
parallel section, dequeues an element, and reiterates. We consider the queue implementation by
Michael and Scott [MS96], that is usually viewed as the reference queue while looking at lock-free
queue implementations.
Dequeue operations fit immediately into our model but Enqueue operations need an adjustment due
to the helping mechanism. Note that without this helping mechanism, a simple queue implementation
would fit directly, but we also want to show that the model is malleable, i.e. the fundamental behavior
remains unchanged even if we divert slightly from the initial assumptions. We consider an equivalent
execution that catches up with the model, and use it to approximate the performance of the actual
execution of Enqueue.
cw = 0, threads = 450001000015000010002000300040005000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageRealcw = 0, threads = 6400080001200016000020004000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageRealcw = 0, threads = 850007500100001250015000020004000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageReal 29
(a) 4 threads
(b) 6 threads
(c) 8 threads
Figure 14: DeleteMin on a priority list
Enqueue is composed of two steps. Firstly, the new node is attached to the last node of the queue
via a CAS, that we denote by CASA, leading to a transient state. Secondly, the tail is redirected to
point to the new node via another CAS, that we denote by CASB, which brings back the queue into
a steady state.
A new Enqueue can not proceed before the two steps of previous success are completed. The first
step is the linearization point of operation and the second step could be conducted by a different
thread through the helping mechanism. In order to start a new Enqueue, concurrent Enqueues help
the completion of the second step of the last success if they find the queue in the transient state.
Alternatively, they try to attach their node to the queue if the queue is in the steady state at the
instant of check. This process continues until they manage to attach their node to the queue via a
retry loop in which state is checked and corresponding CAS is executed.
The flow of an Enqueue is determined by this state checks. Thus, an Enqueue could execute multiple
CASB (successful or failing) and multiple CASA (failing) in an interleaved manner, before succeeding
in CASA at the end of the last retry. If we assume that both states are equally probable for a check
instant which will then end up with a retry, the number of CAS s that ends up with a retry are
expected to be distributed equally among CASA and CASB for each thread. In addition, each thread
has a successful CASA (which linearizes the Enqueue) and a CASB at the end of the operation which
could either be successful or failed by a concurrent helper thread.
We imitate such an execution with an equivalent execution in which threads keep the same relative
ordering of the invocation, return from Enqueue together with same result. In equivalent execution,
threads alternate between CASA and CASB in their retries, and both steps of successful operation is
cw = 50, threads = 425005000750010000125001000200030004000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageRealcw = 50, threads = 650007500100001250010002000300040005000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageRealcw = 50, threads = 85000750010000125002000400060008000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageReal 30
(a) 4 threads
(b) 6 threads
(c) 8 threads
Figure 15: Enqueue-Dequeue on Michael and Scott queues
conducted by the same thread. The equivalent execution can be obtained by thread-wise reordering
of CAS s that leads to a retry and exchanging successful CASB s with the failed counterparts at
the end of an Enqueue, as the latter ones indeed fail because of this success of helper threads.
The model can be applied to this equivalent execution by attributing each CASA-CASB couple to a
single iteration and represent it as a larger retry loop since the successful couple can not overlap
with another successful one and all overlapping ones fail. With a straightforward extension of the
expansion formula, we accomodate the CASA in the critical work which can also expand, and use
CASB as the CAS of our model.
In addition, we take one step further outside the analysis by including a new case, where the
parallel section follows a Poisson distribution, instead of being constant. pw is chosen as the mean
to generate Poisson distribution instead of taking it constant. The results are illustrated in Figure 15.
Our model provides good estimates for the constant pw and also reasonable results for the Poisson
distribution case, although this case deviates from (/extends) our model assumptions. The advantage
of regularity, which brings synchronization to threads, can be observed when the constant and Poisson
distributions are compared. In the Poisson distribution, the threads start to fail with larger pw, which
smoothes the curve around the peak of the throughput curve.
G. Discussion
In this subsection we discuss the adequacy of our model, specifically the cyclic argument, to
capture the behavior that we observe in practice. Figure 16 illustrates the frequency of occurrence of
cw = 225, threads = 42000400060000200040006000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageConstantPoissoncw = 225, threads = 63000400050006000700002000400060008000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageConstantPoissoncw = 225, threads = 82000300040005000600070008000030006000900012000Parallel Work (cycles)Throughput (ops/msec)CaseLowHighAverageConstantPoisson 31
Figure 16: Consecutive Fails Frequency
a given number of consecutive fails, together with average fails per success values and the throughput
values, normalized by a constant factor so that they can be seen on the graph. In the background,
the frequency of occurrence of a given number of consecutive fails before success is presented. As
a remark, the frequency of 6+ fails is gathered with 6. We expect to see a frequency distribution
concentrated around the average fails per success value, within the bounds computed by our model.
While comparing the distribution of failures with the throughput, we could conjecture that the
bumps come from the fact that the failures spread out. However, our model captures correctly the
throughput variations and thus strips down the right impacting factor. The spread of the distribution
of failures indicates the violation of a stable cyclic execution (that takes place in our model), but
in these regions, r actually gets close to 0, as well as the minimum of all gaps. The scattering in
failures shows that, during the execution, a thread is overtaken by another one. Still, as gaps are
close to 0, the imaginary execution, in which we switch the two thread IDs, would create almost the
same performance effect. This reasoning is strengthened by the fact that the actual average number
of failures follows the step behavior, predicted by our model. This shows that even when the real
execution is not cyclic and the distribution of failures is not concentrated, our model that results in
a cyclic execution remains a close approximation of the actual execution.
H. Back-Off Tuning
Together with the analysis comes a natural back-off strategy: we estimate the pw corresponding to
the peak point of the average curve, and when the parallel section is smaller than the corresponding
cw = 4000, threads = 602468010000200003000040000Parallel Work (cycles)0.250.500.75Consecutive Fail FrequencyCaseAv. Fails per SuccessModel AverageNormalized Throughput 32
(a) 8 threads
(b) 4 threads
Figure 17: Comparison of back-off schemes for Poisson Distribution
pw, we add a back-off in the parallel section, so that the new parallel section is at the peak point.
We have applied exponential, linear and our back-off strategy to the Enqueue/Dequeue experiment
specified above. Our back-off estimate provides good results for both types of distribution. In
Figure 17 (where the values of back-off are steps of 115 cycles), the comparison is plotted for the
Poisson distribution, which is likely to be the worst for our back-off. Our back-off strategy is better
than the other, except for very small parallel sections, but other back-off strategies should be tuned
for each value of pw.
We obtained the same shapes while removing the distribution law and considering constant values.
The results are illustrated in Figure 18.
(a) 8 threads
(b) 4 threads
Figure 18: Comparison of back-off schemes for constant pw
cw = 225, threads = 8300040005000600070000250050007500Parallel Work (cycles)Throughput (ops/msec)TypeExponentialLinearNewNoneValue012481632cw = 225, threads = 420004000600080001000010002000300040005000Parallel Work (cycles)Throughput (ops/msec)TypeExponentialLinearNewNoneValue012481632cw = 225, threads = 8300040005000600070000250050007500Parallel Work (cycles)Throughput (ops/msec)TypeExponentialLinearNewNoneValue012481632cw = 225, threads = 4200040006000800010002000300040005000Parallel Work (cycles)Throughput (ops/msec)TypeExponentialLinearNewNoneValue012481632 33
VII. Conclusion
In this paper, we have modeled and analyzed the performance of a general class of lock-free
algorithms, and have so been able to predict the throughput of such algorithms, on actual executions.
The analysis rely on the estimation of two impacting factors that lower the throughput: on the
one hand, the expansion, due to the serialization of the atomic primitives that take place in the
retry loops; on the other hand, the wasted retries, due to a non-optimal synchronization between
the running threads. We have derived methods to calculate those parameters, along with the final
throughput estimate, that is calculated from a combination of these two previous parameters. As
a side result of our work, this accurate prediction enables the design of a back-off technique that
performs better than other well-known techniques, namely linear and exponential back-offs.
As a future work, we envision to enlarge the domain of validity of the model, in order to cope with
data structures whose operations do not have constant retry loop, as well as the framework, so that
it includes more various access patterns. The fact that our results extend outside the model allows us
to be optimistic on the identification of the right impacting factors. Finally, we also foresee studying
back-off techniques that would combine a back-off in the parallel section (for lower contention) and
in the retry loops (for higher robustness).
References
[AF92]
[ACS14] Dan Alistarh, Keren Censor-Hillel, and Nir Shavit. Are lock-free concurrent algorithms practically wait-
free? In David B. Shmoys, editor, Symposium on Theory of Computing (STOC), pages 714 -- 723. ACM,
June 2014.
Juan Alemany and Edward W. Felten. Performance issues in non-blocking synchronization on shared-
memory multiprocessors.
In Norman C. Hutchinson, editor, Proceedings of the ACM Symposium on
Principles of Distributed Computing (PoDC), pages 125 -- 134. ACM, 1992.
[ARJ97] James H. Anderson, Srikanth Ramamurthy, and Kevin Jeffay. Real-time computing with lock-free shared
objects. ACM Transactions on Computer Systems (TOCS), 15(2):134 -- 165, 1997.
[DB08] Kristijan Dragicevic and Daniel Bauer. A survey of concurrent priority queue algorithms. In Proceedings
of the International Parallel and Distributed Processing Symposium (IPDPS), pages 1 -- 6, April 2008.
[GH09]
[DGT13] Tudor David, Rachid Guerraoui, and Vasileios Trigonakis. Everything you always wanted to know about
synchronization but were afraid to ask. In Michael Kaminsky and Mike Dahlin, editors, Proceedings of the
ACM Symposium on Operating Systems Principles (SOSP), pages 33 -- 48. ACM, November 2013.
[DLM13] Dave Dice, Yossi Lev, and Mark Moir. Scalable statistics counters. In Guy E. Blelloch and Berthold Vöcking,
editors, Proceedings of the ACM Symposium on Parallelism in Algorithms and Architectures (SPAA), pages
43 -- 52. ACM, July 2013.
James R. Goodman and Herbert Hing Jing Hum. Mesif: A two-hop cache coherency protocol for point-to-
point interconnects. Technical report, University of Auckland, November 2009.
[HSY10] Danny Hendler, Nir Shavit, and Lena Yerushalmi. A scalable lock-free stack algorithm. Journal of Parallel
and Distributed Computing (JPDC), 70(1):1 -- 12, 2010.
Intel. Lock scaling analysis on Intel R(cid:13) Xeon R(cid:13) processors. Technical Report 328878-001, Intel, April 2013.
[Int13]
[KH14] Alex Kogan and Maurice Herlihy. The future(s) of shared data structures. In Magnús M. Halldórsson and
Shlomi Dolev, editors, Proceedings of the ACM Symposium on Principles of Distributed Computing (PoDC),
pages 30 -- 39. ACM, July 2014.
Jonatan Lindén and Bengt Jonsson. A skiplist-based concurrent priority queue with minimal memory
contention.
In Roberto Baldoni, Nicolas Nisse, and Maarten van Steen, editors, Proceedings of the
International Conference on Principle of Distributed Systems (OPODIS), volume 8304 of Lecture Notes
in Computer Science, pages 206 -- 220. Springer, December 2013.
[Mic04] Maged M. Michael. Hazard pointers: Safe memory reclamation for lock-free objects. IEEE Transactions
[LJ13]
on Parallel and Distributed Systems (TPDS), 15(6):491 -- 504, 2004.
[MS96] Maged M. Michael and Michael L. Scott. Simple, fast, and practical non-blocking and blocking concurrent
queue algorithms. In James E. Burns and Yoram Moses, editors, Proceedings of the ACM Symposium on
Principles of Distributed Computing (PoDC), pages 267 -- 275. ACM, May 1996.
Nir Shavit and Itay Lotan. Skiplist-based concurrent priority queues. In Proceedings of the International
Parallel and Distributed Processing Symposium (IPDPS), pages 263 -- 268, May 2000.
[SL00]
[Tre86] R. Kent Treiber.
Systems programming: Coping with parallelism.
International Business Machines
Incorporated, Thomas J. Watson Research Center, 1986.
[Val94]
J. D. Valois. Implementing Lock-Free Queues. In Proceedings of International Conference on Parallel and
Distributed Systems (ICPADS), pages 64 -- 69, December 1994.
34
|
1704.03664 | 3 | 1704 | 2018-11-26T13:15:28 | Approximating Optimization Problems using EAs on Scale-Free Networks | [
"cs.DS",
"cs.NE",
"cs.SI"
] | It has been observed that many complex real-world networks have certain properties, such as a high clustering coefficient, a low diameter, and a power-law degree distribution. A network with a power-law degree distribution is known as scale-free network. In order to study these networks, various random graph models have been proposed, e.g. Preferential Attachment, Chung-Lu, or Hyperbolic.
We look at the interplay between the power-law degree distribution and the run time of optimization techniques for well known combinatorial problems. We observe that on scale-free networks, simple evolutionary algorithms (EAs) quickly reach a constant-factor approximation ratio on common covering problems
We prove that the single-objective (1+1)EA reaches a constant-factor approximation ratio on the Minimum Dominating Set problem, the Minimum Vertex Cover problem, the Minimum Connected Dominating Set problem, and the Maximum Independent Set problem in expected polynomial number of calls to the fitness function.
Furthermore, we prove that the multi-objective GSEMO algorithm reaches a better approximation ratio than the (1+1)EA on those problems, within polynomial fitness evaluations. | cs.DS | cs |
Approximating Optimization Problems using EAs on
Scale-Free Networks
Ankit Chauhan, Tobias Friedrich, and Francesco Quinzan
Hasso Plattner Institute, Potsdam, Germany
Abstract. It has been observed that many complex real-world networks have certain prop-
erties, such as a high clustering coefficient, a low diameter, and a power-law degree dis-
tribution. A network with a power-law degree distribution is known as scale-free network.
In order to study these networks, various random graph models have been proposed, e.g.
Preferential Attachment, Chung-Lu, or Hyperbolic.
We look at the interplay between the power-law degree distribution and the run time of op-
timization techniques for well known combinatorial problems. We observe that on scale-free
networks, simple evolutionary algorithms (EAs) quickly reach a constant-factor approxima-
tion ratio on common covering problems
We prove that the single-objective (1 + 1) EA reaches a constant-factor approximation
ratio on the Minimum Dominating Set problem, the Minimum Vertex Cover problem, the
Minimum Connected Dominating Set problem, and the Maximum Independent Set problem
in expected polynomial number of calls to the fitness function.
Furthermore, we prove that the multi-objective Gsemo algorithm reaches a better approxi-
mation ratio than the (1 + 1) EA on those problems, within polynomial fitness evaluations.
Keywords: Evolutionary algorithms, covering problems, power-law bounded networks.
1
Introduction
Bio-inspired randomized search heuristics, such as evolutionary algorithms, are well-suited to ap-
proach combinatorial optimization problems. These algorithms have been extensively analyzed on
artificial pseudo-Boolean functions (see e.g. Droste et al. [1], and Jansen and Wegener [2]) as well
as on some combinatorial optimization problems (see e.g. Giel and Wegener [3], Neumann [4],
and Neumann and Wegener [5]). The standard approach to perform theoretical analyses is the
average-case black-box complexity: The run time is estimated by counting the expected number of
calls to the valuation oracle or fitness function. In recent years, a significant effort has been made
to study the run time of EAs on combinatorial optimization problems.
The performance of single-objective population-based EAs on the Minimum Vertex Cover
problem is studied in Oliveto et al. [6,7]. They show that these algorithms achieve arbitrarily bad
approximation guarantee on this problem. In contrast, Friedrich et al. [8] prove that multi-objective
EAs can obtain an optimal solution in expected polynomial time on this instance.
Back et al. [9] study the performance of a simple a single-objective evolutionary algorithm for
the Maximum Independent set problem and claim its superiority by experimental observations.
Similarly, Peng [10] analyzes a single objective EA on this problem and proves that it reaches a
((∆(G) + 1)/2)-approximation within expected O(cid:0)n4(cid:1) fitness evaluations, where n is the number
of nodes in a given graph G, and ∆(G) its maximum node degree. These analyses, as well as those
for the Minimum Vertex Cover problem, do not require any assumptions on the topology of G.
Power-law bounded networks. A wide range of real-world networks, such as the internet graph, the
web, power grids, protein-protein interaction graphs, and social networks exhibit properties such
as high clustering coefficient (see Kumar et al. [11]), small diameter (see Leskovech et al. [12]), and
approximate power-law degree distribution (see Faolutsos et al. [13]). Various models have been
proposed to capture, and possibly explain, these properties (see e.g. Newman [14] and Watts and
Strogatz [15]).
2
One of the most significant contributions in this sense is that of Brach et al. [16]. Many real-
world networks approximately exhibit a power-law degree distribution. That is, the number of
nodes of degree k is approximately proportional to k−β, for k sufficiently large, where β > 1 is a
constant inherent to the network. Networks with this property are commonly referred to as scale-
free. Brach et al. [16] define a deterministic condition that captures the behavior of the degree
distribution of such networks. This condition requires an upper-bound on the number of nodes
with degree δ(v) ∈(cid:2)2i, 2i+1(cid:1), for all i = 0, · · · , ⌊log(n)−1⌋+1. This property is commonly referred
to as PLB. Carrying out run time analysis of EAs on networks that exhibit this property may give
a better understanding of their performance in practice.
Our contribution. We study various well-known NP-hard graph covering problems, and prove
that simple EAs perform well, if the underlying graph exhibits the PLB property. Specifically, we
study the Minimum Dominating Set problem, the Minimum Vertex Cover problem, the Maximum
Independent Set problem, and the minimum Connected Dominating Set problem.
We prove that for these problems a single-objective EA reaches a constant-factor approximation
ratio within expected polynomial fitness evaluations. We prove that a multi-objective EA reaches
a better approximation guarantee, given more time budget.
The paper is organized as follows. In Section 2 we define the algorithms and give some basic
definitions and technical tools. In Section 4 we study the Minimum Dominating Set problem, both
in the single- and multi-objective case. The results for the Minimum Vertex Cover problem follows
directly from the analysis for the Minimum Dominating Set problem and are presented in Section
4. The Minimum Connected Dominating Set problem is discussed in Section 6. We conclude with
the analysis for the Maximum Independent Set problem, in Section 7.
2 Preliminaries
In this paper, we only study undirected graphs G = (V, E) without loops, where V is the set of
nodes and E is the set of edges. We always use n to denote V and m to denote E. We use
δ(v) to denote the degree of each node v ∈ V , and we let ∆(G) be the maximum degree of G.
Moreover, given a pseudo-Boolean array x we denote with x1 the number of ones in the input
string. Otherwise, we use the well-known mathematical and graph theoretic notations, e.g., we use
A to denote the size of any set A. We use the following definition of ǫ-approximation.
Definition 1. Consider a problem P on a graph G = (V, E) let U be a possible solution, and
denote with opt an optimal solution of P. If P is a minimization problem, we say that U is an
ǫ-approximation if it holds that U / opt ≤ ǫ. If P is a maximization problem, we say that U is
an ǫ-approximation if it holds that opt / U ≤ ǫ.
We consider both single- and multi-objective optimization. Whereas in the single-objective case
the algorithm searches for the global optimum, in the latter the algorithm searches for a set of
optimal solutions. Such solutions are part of the Pareto front.
Consider any two points x′ = (x′
1, . . . , x′
Rm. We say that x′ dominates x′′, in symbols x′ ≻ x′′, if it holds that x′
With this notion of dominance, we define the Pareto front as follows.
m) and x′′ = (x′′
1 , . . . , x′′
m) in the m-dimensional space
i for all i = 1, . . . , m.
i ≥ x′′
Definition 2. Consider a function f : X ⊆ Rn → Rm, with X being a compact set in the metric
space Rn. Consider the set Y := {y ∈ Rm y = f (x), x ∈ X}, and denote with ≻ the standard
partial order on Y . The Pareto front is defined as P(Y ) = {y′ ∈ Y {y′′ ≻ y′, y′′ 6= y′} = ∅}.
2.1 Algorithms
The (1+1) EA is a randomized local search heuristic inspired by the process of biological evolution
(cf. Algorithm 1). Initially, an individual x is sampled uniformly at random (u.a.r.). An offspring
y is then generated, by flipping all bits of x independently with probability 1/n. The fitness values
3
Algorithm 1: The (1 + 1) EA.
input: a fitness function f : 2V → R≥0;
output: an (approximate) global maximum of the function f ;
// sample initial solution
choose x ∈ {0, 1}n uniformly at random;
while convergence criterion not met do
// perform mutation
create offspring y by flipping each bit of x independently w.p. 1/n;
// perform selection
if f (y) ≤ f (x) then
x ← y;
return x;
Algorithm 2: The Gsemo.
input: a fitness function f : 2V → R≥0;
output: an (approximate) global minimum of the function f ;
// sample initial solution
choose x ∈ {0, 1}n uniformly at random;
P ← P ∪ {x};
while convergence criterion not met do
// perform mutation
choose x ∈ P uniformly at random;
create y by flipping each bit of x independently w.p. 1/n;
// perform selection
if f (y) is not dominated by f (z), for all points z ∈ P then
P ← P ∪ {y};
delete all solutions z ∈ P s.t. f (z) is dominated by f (y);
return P ;
of x and y are then compared. If the value f (y) is less than or equal to f (x), then x is discarded
and y is preserved in memory; otherwise y is discarded and x is preserved in memory. Note that
this algorithms are defined for minimization problems. However, symmetric definitions hold for
maximization problems.
The Gsemo is a multi-objective evolutionary algorithm (cf. Algorithm 2). As in the case of the
(1 + 1) EA, an initial solution is chosen u.a.r. from the objective space and stored it in the Pareto
front P . An element x is then chosen u.a.r. from P and a new solution y is generated by flipping
each bit of x independently w.p. 1/n. If y is not strongly dominated by any other solution in
P , then y is stored in the Pareto front, and all elements which are strongly dominated by y are
discarded. Otherwise, the population remains unchanged.
2.2 The Multiplicative Drift theorem.
In the case of the (1 + 1) EA, for any fitness function f : {0, 1}n → R≥0, we describe its run
time as a Markov chain {Xt}t≥0, where Xt is the f -value reached at time step t. Following this
convention, we perform parts of the analyses with the Multiplicative Drift Theorem (see Doerr et
4
al. [17]), a strong tool to analyze the run time of evolutionary algorithms such as the (1 + 1) EA.
Intuitively, this theorem gives an estimate for the expected value of the run time of the (1 + 1) EA,
provided that the change of the average value of the process {Xt}t≥0 is within a multiplicative
factor of the previous solution. More formally, the following theorem holds.
Theorem 3 (Multiplicative Drift). Let S ⊆ R be a finite set of positive numbers with minimum
smin. Let {X (t)}t∈N be a random variables over S ∪ {0}. Let T be the random variable that denotes
the first point in time t ∈ N for which X (t) = 0.
Suppose that there exists a real number δ > 0 such that
holds for all s ∈ S with P r[X (t) = s] > 0. Then for all s0 ∈ S with P r[X (0) = s0] > 0, we have
E[X (t) − X (t+1) X (t) = s] ≥ δs
E[T X (0) = s0] ≤
1 + ln(s0/smin)
δ
.
A proof of this result is given in Doerr et al. [17, Theorem 3].
3 Power-Law Bounded Networks
In many real-world networks the degree distribution approximately follows a power-law. We frame
this concept with the following definitions (see Brach et al. [16]).
Definition 4 (PLB Network). An undirected graph G = (V, E) is a power-law bounded network,
if there exists parameters 1 < β = O(1) and t ≥ 0 s.t. the number of nodes v with δ(v) ∈ [2d, 2d+1)
is at most
c1n(t + 1)β−1
2d+1−1
(i + t)−β,
Xi=2d
for all d ≥ 0, and with c1 > 0 a universal constant.
Definition 4 intuitively captures the general idea that vertices with bounded degree can be grouped
into sets with cardinality upper-bounded by a power-law. Again, we remark that this property is
deterministic. We say a graph is a PLB networks if it is power-law bounded as in Definition 4.
We prove that PLB networks have the property that feasible solutions of various covering prob-
lems yield a constant-factor approximation ratio. To this end, we consider the following definition.
Definition 5. Consider an undirected graph G = (V, E). We say that a set D ⊆ V is dominating
if every node of V \ D is adjacent to at least one node of D.
Dominating sets play an important role in our analysis, because the sum of the degree of their
nodes always yields constant-factor approximation with respect (w.r.t.) to their cardinality, on
PLB networks with bounded parameters β, t. More formally, the following lemma holds.
Lemma 6. Let G = (V, E) be a PLB network with parameters 2 < β = O (1), t ≥ 0, and a
universal constant c1. Denote with D a dominating set of G. Define the constants
a :=
Then it holds
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
Pv∈D(δ(v) + 1)
D
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
≤ 2ab + 1 = Θ (1) .
5
Proof Define the following constants
d≥k := {v ∈ V : δ(v) ≥ k}
and γ := min{k ∈ N : d≥k ≤ D}.
Then d≥k is the number of nodes of G with degree at least k, and γ is the smallest index k by
which the nodes of G with degree at least 2k is at most D. We first prove an upper-bound on
the numerator, in terms of γ. From Definition 5 it holds
⌈log(n−1)⌉
2j+1−1
δ(v) ≤ c1n(t + 1)β−1
Xv∈D
≤ 2c1n(t + 1)β−1
2j+1
(i + t)−β
Xi=2j
2⌈log(n−1)⌉+1−1
(i + t)1−β
Xj=γ
Xi=2γ
≤ 2c1n(t + 1)β−1Z 2⌈log(n−1)⌉+1−1
≤ 2c1n(t + 1)β−1 (2γ + t)2−β
2γ
.
β − 2
(x + t)1−βdx
(1)
We continue by computing a lower-bound on D in terms of γ. Again, from Definition 5 we have
that it holds
2⌈log(n−1)⌉+1−1
D ≥ nγ = c1n(t + 1)β−1
(i + t)−β
Xi=2γ
≥ c1n(t + 1)β−1Z 2γ+1
1 −(cid:16) t+2
≥ c1n(t + 1)β−1
2γ
t+1(cid:17)1−β
β − 1
(x + t)−βdx
(2γ + t)1−β.
(2)
We combine the upper-bound on the numerator Pv∈D δ(v) + 1 given in (1), together with the
lower bound on D as in (2), to obtain that it holds
Pv∈D(δ(v) + 1)
D
≤ 2
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
(2γ + t) + 1.
(3)
We conclude by giving an upper-bound for 2γ + t. From the definition of dominating set (see
Definition 5) it holds Pv∈D δ(v) ≥ n/2. Combining this observation with (1), it holds
n
2
≤ 2c1n(t + 1)β−1 (2γ + t)2−β
,
β − 2
from which it follows that
2γ + t ≤(cid:20)4c1
(t + 1)β−1
β − 2
1
β−2
.
(cid:21)
The claim follows by substituting (4) in (3).
(4)
(cid:3)
Note that Lemma 6 uses a parameter β > 2. From a practical point of view this assumption is not
restrictive, since most real-world networks fulfill this requirement. However, for β < 2 one may
define a degenerate PLB network for which our analysis fails.
Lemma 6 can be readily used used to prove approximation guarantees for the Minimum Domi-
nating Set problem - i.e. the problem of searching for a minimum dominating set. In the following,
6
however, we show that this lemma can be used to derive approximation guarantees for other
common covering problems.
Given a graph G = (V, E), a connected dominating set is any dominating set C ⊆ V s.t. the
nodes of C induce a connected sub-graph of G. From Lemma 6 the following result readily follows.
Corollary 7. Let G = (V, E) be a PLB network with parameters 2 < β = O (1), t ≥ 0, and a
universal constant c1. Denote with C a connected dominating set of G. Define the constants
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
a :=
Then it holds
Pv∈C (δ(v) + 1)
C
≤ 2ab + 1 = Θ (1) .
Finally, we apply Lemma 6 to independent sets. Given a graph G = (V, E), an independent
set M ⊆ V consists of vertices of G, no two of which are adjacent. A maximum independent set
is any independent set of maximum size. It is not true, in general, that an independent set is
a dominating set. However, a maximum independent set is always a dominating set. Hence, the
following corollary holds.
Corollary 8. Let G = (V, E) be a PLB network with parameters 2 < β = O (1), t ≥ 0, and a
universal constant c1. Denote with M∗ a maximum independent set of G. Define the constants
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
a :=
Then it holds
Pv∈M∗
(δ(v) + 1)
M∗
≤ 2ab + 1 = Θ (1) .
4 The Minimum Dominating Set Problem
We study the minimum dominating set problem (MDS): For a given graph G = (V, E), find a
dominating set of V (see Definition 5) of minimum cardinality. In this section, we denote with
opt any optimal solution to the MDS. Given a graph G = (V, E) with V = n, we consider
an indexing i : V → {1, . . . , n} of the vertices. We represent any set of nodes S ⊆ V with a
pseudo-Boolean array (x1, . . . , xn) of length n and s.t. xi = 1 if and only if the i-th node of G is
in S.
4.1 Single-objective optimization.
We search for a minimum dominating set by minimizing the following single-objective function
F (x) = nu(x) + x1,
(5)
where u(x) is the number of non-dominated nodes, and x1 is the number of 1s in the input string.
Note that u(x) = 0 if and only if the solution x is a dominating set. We prove the main result of
this section, by giving an upper-bound on the run time of the (1 + 1) EA until a dominating set
is found. We then use Lemma 6 to give an upper-bound on the approximation ratio achieved by
this solution.
Theorem 9. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0 and with a universal
constant c1. Define the constants
7
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the (1 + 1) EA finds a (2ab + 1)-approximation of the MDS after expected O (n log n) fitness
evaluations. In particular, the (1 + 1) EA is a Θ(1)-approximation algorithm for the MDS on PLB
networks.
Proof We first observe that because of the weights on the fitness, the algorithm initially searches
for a dominating set and then tries to minimize it. Furthermore, no step that reduces the number of
dominated nodes is accepted. We estimate the expected run time until a feasible solution is found,
and conclude by proving that any feasible solution reaches the desired approximation guarantee.
To estimate the run time we use the Multiplicative Drift theorem (see Theorem 3), by defining
{X (t)}t≥0 as the process X (t) = u(xt) for all t ≥ 0. Suppose that the current solution xt is not a
dominating set - i.e. u(xt) > 0. Then there exists a node v ∈ V s.t. by adding it to the current
solution, the fitness u(xt) decreases by 1. Since the probability of performing a single chosen bit-flip
is at least 1/en, then it holds
E[Xt − Xt−1 Xt] ≥
Xt
en
.
Following the notation of Theorem 3, we can trivially set S = {1, . . . , n} and obtain that the run
time until a feasible solution is found is upper-bounded as O (n log n). We conclude by proving that
the (1 + 1) EA reaches the desired approximation guarantee. Denote with x∗ the first dominating
set reached by the (1 + 1) EA. We observe that any optimal solution opt dominates all nodes of
G. It follows that it holds
x∗
opt
≤ Pv∈opt(δ(v) + 1)
opt
≤ 2ab + 1,
where the last inequality in 6 follows from Lemma 6. The claim follows.
(6)
(cid:3)
4.2 Multi-objective optimization.
We approach MDS with the multi-objective Gsemo (cf. Algorithm 2). In this case, we use the
following bi-objective fitness function
f = (u(x), x1),
(7)
where u(x) is the number of non-dominated nodes, and x1 the number of 1s in the input string.
We prove that the Gsemo fins a good approximation of the MDS in polynomial time, when
optimizing a fitness as in (7). Useful in the analysis is the following lemma.
Lemma 10. Let G = (V, E) be a graph, and let Sk := {v1, . . . , vk} be a sequence of nodes s.t.
vj has the maximum degree in the complete sub-graph induced by the nodes of G that are not
dominated by {v1, . . . , vj−1}. Let nk be the number of nodes in V that are not dominated by Sk.
Then it holds
for all 0 < k ≤ n.
nk ≤ n(cid:18)1 −
1
opt(cid:19)k
,
8
Proof We prove this by induction on k. For the base case, we set k = 1. Since the number of
non-dominated nodes is n, there exists a node v ∈ V s.t. v dominates at least n/opt nodes in
the graph. Otherwise, opt is not an optimal solution. Thus, the number of non-dominated nodes
by S1 is at most
n1 ≤ n −
n
opt
= n(cid:18)1 −
1
opt(cid:19) .
For the inductive step, suppose that the statement holds for k = i. Again, since opt is the size of
the minimum dominating set, by the pigeon-hole principle there exists a node V \ Si s.t. Si ∪ {v}
dominates at least ni/ opt many nodes. Therefore, it holds that
ni+1 ≤ ni −
ni
opt
= ni(cid:18)1 −
1
opt(cid:19) = n(cid:18)1 −
1
opt(cid:19)i+1
.
The claim follows.
(cid:3)
We use the lemma above to analyze the run time and approximation guarantee of the Gsemo on
MDS. The following theorem holds.
Theorem 11. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0 and with a
universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the Gsemo finds an ln(2ab + 1)-approximation for the MDS after expected O(cid:0)n3(cid:1) fitness
evaluations. In particular, Gsemo is a Θ(1)-approximation algorithm for the MDS on PLB net-
works.
Proof The proof consists of two parts. We first estimate the expected run time until the solution
0n - i.e. the empty set - is added to the Pareto front (Phase 1), and then we estimate the time
until the desired approximation is reached from that solution (Phase 2).
(Phase 1) To analyze the first part of the process, we denote with x∗ the individual in the Pareto
front with the minimum number of 1s, at a given time step. Since the total number of elements in
the Pareto front is upper-bounded by n + 1, the probability of choosing x∗ for mutation is at least
1/(n + 1) at each iteration. Once x∗ is selected for mutation, then the probability of performing a
single bit-flip and remove one element to x∗ is at least x∗1 /(en). Therefore, the expected waiting
time to reach the solution 0n is upper-bounded as
(n + 1)
en
j
n
Xj=1
= O(cid:0)n2 log n(cid:1) .
(Phase 2) We now give an upper bound on the run time until the desired approximation is
reached, assuming that the solution 0n has been found. From Lemma 10, it follows that there
exists a node v1 ∈ V s.t. after adding this node to 0n the number of non-dominated nodes is at
most
n1 := u ({v1}) ≤ n(cid:18)1 −
1
opt(cid:19) .
With an argument similar to the one presented to analyze the first part of the process, in expected
O(cid:0)n2(cid:1) time an individual with fitness value (n1, 1) can be added to the solution. If {v1} is not an
optimal solution, then by Lemma 10 there exists a node v2 ∈ V \ {v1} s.t.
n2 := u ({v1, v2}) ≤ n(cid:18)1 −
1
opt(cid:19)2
.
9
solution with fitness (n1, 1) was added. Similarly, an individual with fitness (ni, i) can be added to
Hence, an individual (n2, 2) can be added to the current population within expected O(cid:0)n2(cid:1), after a
the current population within expected O(cid:0)n2(cid:1), after a solution with fitness (ni−1, i−1) was added.
We can iterate this process as indicated in Lemma 10, until a dominating set Sk = {v1, . . . , vk} ⊆ V
is found.
nk := u (Sk) ≤ n(cid:18)1 −
1
opt(cid:19)k
.
(8)
Note that there are at most n nodes that can be added to the solution 0n as described above.
10 that is a dominating set.
Hence after a total of O(cid:0)n3(cid:1) fitness evaluations the Gsemo reaches a solution Sk as in Lemma
Again, from the definition of minimum dominating set it holds n ≤ Pv∈opt δ(v) + 1. Combining
We conclude by showing that the solution Sk yields the desired approximation guarantee.
this observation with (8), we can estimate k by solving the inequality
Sk
opt
≤
n
opt(cid:18)1 −
1
opt(cid:19)k
≤ Pv∈opt(δ(v) + 1)
opt
(cid:18)1 −
1
opt(cid:19)k
.
(9)
We combine (9) with Lemma 6 to obtain that
1 ≤ (2ab + 1) exp(cid:18)−
k
opt(cid:19) .
By taking the logarithm on both sides, we get k ≤ opt ln(2ab + 1) and we can conclude that the
cardinality of Sk is at least Sk ≤ opt ln(2ab + 1). The claim follows.
(cid:3)
5 The Minimum Vertex Cover Problem
Given a graph G = (V, E), a vertex cover is a set S ⊆ V of vertices s.t. each edge in E is adjacent
to at least one node in S. The minimum vertex cover problem (MVC) consists of finding a vertex
cover of minimum size. In this section, we denote with opt any optimal solution to the MVC, and
we use intuitive bit-string representation based on vertices, as discussed in Section 4.
5.1 Single-objective optimization.
We study the run time and approximation guarantee for the (1 + 1) EA. Since any vertex cover is
also a dominating set, then one can approach the MVC by minimizing a fitness function as in (5).
Moreover, the following run time analysis follows directly from Theorem 9.
Corollary 12. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0 and with a
universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the (1 + 1) EA finds a (2ab)-approximation of the MVC in expected O (n log n) fitness eval-
uations. In particular, the (1 + 1) EA is a Θ(1)-approximation algorithm for the MVC on PLB
networks.
10
5.2 Multi-objective optimization.
We prove that the multi-objective Gsemo yields an improved approximation guarantee then the
(1 + 1) EA, at the cost of a higher number of expected fitness evaluations. Again, since a vertex
cover is also a dominating set, then this problem can be approached by minimizing a two-objectives
function as in (7). Hence, the following run time analysis for the Gsemo readily follows from
Theorem 11.
Theorem 13. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0 and with a
universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the Gsemo finds an (ln(2ab) + 1)-approximation solution of MVC in expected O(cid:0)n3(cid:1) fit-
ness evaluations. In particular, Gsemo is a Θ(1)-approximation algorithm for the MVC on PLB
networks.
6 The Minimum Connected Dominating Set Problem
Given a graph G = (V, E), recall that a connected dominating set is a set D ⊆ V of vertices
that dominates every node v ∈ V , and the nodes u ∈ D induce a connected sub-graph on G. The
minimum connected dominating set problem (CDS) consists of finding a connected dominating set
C ⊆ V of minimum size. In this section, we denote with opt any optimal solution to the CDS.
Throughout this section, we use intuitive bit-string representation based on vertices, as discussed
in Section 4.
6.1 Single-objective optimization.
We study the optimization process for the CDS with the (1 + 1) EA. We approach this problem
by minimizing following fitness function.
F (x) = n2(u(x) + w(x) − 1)+x1,
(10)
where u(x) is the number of non-dominated nodes, and w(x) is the number of connected compo-
nents in the sub-graph induced by the chosen solution. Note that any connected dominating set
C yields u(C) + w(C) − 1 = 0. We search for a solution to the CDS by minimzing F (x). It can be
observed that a sub-graph induced by any subset D ⊆ V that is represented by the input string
x is a connected dominating set if and only if u(x) = 0 and w(x) = 1. Because of the weights on
the fitness function, the (1 + 1) EA tends to reach a feasible solution first, and then it removes
unnecessary nodes.
The argument for the theorem below is similar to that given in Theorem 9: We first show that
the (1 + 1) EA reaches a locally optimal solution in O (n log n) fitness evaluations, and then we
use the PLB property to show that any such solution gives the desired approximation ratio (see
Corollary 7). The following theorem holds.
Theorem 14. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0 and with a
universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the (1 + 1) EA finds a (2ab)-approximation of the CDS in expected O (n log n) fitness eval-
uations. In particular, the (1 + 1) EA is a Θ(1)-approximation algorithm for the CDS on PLB
networks.
Proof To estimate the run time we use the multiplicative Drift theorem (see Theorem 3). Define
the function f (x) = u(x) + w(x) − 1, with u(x) and w(x) as in (10), and let xt be a solution found
at time step t. Suppose that the current solution xt is not feasible. Then with probability atl east
1/en a node is added to xt s.t. the value f (xt) decreases by 1. Hence, if we define {Xt}t≥0 as the
process Xt = u(xt) for all t ≥ 0, then it holds
11
E[Xt − Xt−1 Xt] ≥
Xt
en
.
Following the notation of Theorem 3, we set S = {1, . . . , n} to obtain that the run time until a
feasible solution is found is upper-bounded as O (n log n).
We conclude by proving that the (1 + 1) EA reaches the desired approximation guarantee.
Denote with x∗ the first feasible solution reached by the (1 + 1) EA. We observe that any optimal
solution opt dominates all nodes of G. It follows that it holds
x∗
opt
≤ Pv∈opt(δ(v) + 1)
opt
≤ 2ab + 1,
with the last inequality in (11) following from Corollary 7. The claim follows.
(11)
(cid:3)
6.2 Multi-objective optimization.
As in the case of the MDS and the MVC, we again analyze the run time of the Gsemo. We search
for a solution of the CDS by minimizing the following multi-objective fitness
F (x) = (u(x) + w(x), x1),
(12)
with u(x) and w(x) as described earlier in this section. The following result is useful for the run
time analysis.
Lemma 15. Let G = (V, E) be a connected graph, and consider the function f (x) := u(x) + w(x),
with u(x), w(x) as in (12). Let Sk := {v1, . . . , vk} be a set of nodes of G s.t.
vj :=
arg max
{f (Sk \ {vk}) − f (Sk)} ,
u∈V \{v1,...,vj−1}
for all j = 1, . . . , k. Then there exists a node v ∈ V \ Sk s.t. it holds that
f (Sk ∪ {v}) ≤ f (Sk) −
f (Sk)
opt
+ 1.
A proof of the lemma above is given by Guha and Khuller [18, Theorem 3.3], where it is used to
analyze the performance of a greedy algorithm. The following theorem holds.
Theorem 16. Let G = (V, E) be a connected PLB network with parameters β > 2, t ≥ 0 and
with a universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the Gsemo finds a ln(2eab + e)-approximation for the CDS after expected O(cid:0)n3(cid:1) fitness
evaluations. In particular, the Gsemo is a Θ(1)-approximation algorithm for the CDS on PLB
networks.
12
Proof The hereby presented argument is similar to the one given in Theorem 11. Since the function
x1 is an objective, then there are at most n + 1 individuals in the Pareto front, at each point
during the iteration. Using this knowledge, and following an argument as in (Phase 1) in Theorem
11, one can prove that the solution 0n is added to the Pareto front after expected
n
Xj=1(cid:18)
j
e(n + 1)n(cid:19)−1
= O(cid:0)n2 log n(cid:1)
fitness evaluations.
Once the empty set is reached, the Gsemo iteratively generates solutions with increasing
cardinality, until a good approximation of a minimum dominating set is reached. Define f (x) =
u(x) + w(x), and denote with Sj = {v1, . . . , vj} a sequence of nodes s.t.
vj :=
arg max
{f (Sk \ {vk}) − f (Sk)} ,
u∈V \{v1,...,vj−1}
until a connected dominated set Sk is reached (see Lemma 15). Again, selecting an individual for
mutation and adding a chosen node to it occurs after expected O(cid:0)n2(cid:1) fitness evaluations. Since
there are at most n nodes that can be added to the empty set 0n, then after expected O(cid:0)n3(cid:1)
fitness evaluations the Gsemo reaches a solution Sk that is a dominating set.
We conclude by showing that Sk yields the desired approximation guarantee. From Lemma 15
it holds
1
1
opt(cid:19)k
f (Sk) ≤ f (S0)(cid:18)1 −
opt(cid:19)k
≤ f (S0)(cid:18)1 −
≤ (f (S0) − opt)(cid:18)1 −
= (n + 1 − opt)(cid:18)1 −
1
+
opt(cid:19)j
Xj=0(cid:18)1 −
+ opt 1 −(cid:18)1 −
opt(cid:19)k
opt(cid:19)k
+ opt.
+ opt
1
opt(cid:19)k!
k−1
1
1
Note also that from Corollary 7 it holds
n + 1 − opt
opt
≤ Px∈opt(δ(x) + 1)
opt
≤ 2ab + 1,
and the following chain of inequalities also holds
1 ≤
f (Sk)
opt
≤ (2ab + 1)(cid:18)1 −
1
opt(cid:19)k
+ opt .
(13)
If we solve (13) w.r.t. k, then it follows that k ≤ opt ln(2eab + e). Since k = Sk, then we
conclude that Sk / opt ≤ ln(2eab + e), and the claim follows.
(cid:3)
7 The Maximum Independent Set Problem
For a given graph G = (V, E), recall that an independent set is a subset S ⊆ V s.t. no two
nodes u, v ∈ S are adjacent. The maximum independent set problem (MIS) consists of finding an
independent set S ⊆ V of maximum cardinality. In this section, we denote with opt any solution
for the MIS. In this section, we use a standard bit-string representation based on vertices, as
discussed in Section 4.
7.1 Single-objective optimization.
13
To implement the (1 + 1) EA, we use the fitness function proposed by Back et al. [9], which is
defined as follows:
n
n
F (x) = x1 − n
xj eij,
xi
Xi=1
Xj=1
(14)
where eij is 1 if there is an edge between vi and vj, and it is 0 otherwise. In this case, the objective
j=1 xjeij = 0 if and only
if the solution is an independent set. Because of the weight n, the (1 + 1) EA reaches a feasible
solution first and then it adds nodes to it, to maximize the fitness.
of the (1 + 1) EA is to maximize F (x). Note that it holds thatPn
i=1 xiPn
In order to perform the analysis for the (1 + 1) EA, we use the following definition of local
optimality.
Definition 17. Let G = (V, E) be a graph, and let F be as in (14). We say that a set of nodes
S ⊆ V is a 3-local optimum of it holds
for all U ⊆ S and T ⊆ V \ S s.t. U ∪ T ≤ 3.
F ((S \ U ) ∪ T ) ≤ F (S)
Definition 17 captures the notion of optimality up to three bit-flips. Using this definition, we prove
the following result.
Theorem 18. Let G = (V, E) be a connected PLB network with parameters β > 2, t ≥ 0 and
with a universal constant c1. Define the constants
a :=
β − 1
β − 2"1 −(cid:18) t + 2
t + 1(cid:19)1−β#−1
and
b :=(cid:20)4c1
(t + 1)β−1
β − 1
1
β−2
.
(cid:21)
Then the (1 + 1) EA finds an (ab + 1/2)-approximation for the MIS after expected O(cid:0)n4(cid:1) fitness
evaluations. In particular, the (1 + 1) EA is a Θ(1)-approximation algorithm for the MIS on PLB
networks.
Proof We first perform a run time analysis for the (1 + 1) EA, until a 3-local optimum is found
(see Definition 17), and then we prove that any such solution yields the desired approximation
guarantee.
To perform the run time analysis, we divide the optimization process in two phases. In (Phase
1) the (1 + 1) EA searches for a feasible solution, whereas in (Phase 2) it searches for a 3-local
optimum, given that a feasible solution has been found.
(Phase 1) Let x0 be the initial solution of the (1 + 1) EA and assume that x0 is not feasible. Since
the empty set is always an independent set, then we can use an argument as in Theorem 9 (Phase
1) to conclude that the (1 + 1) EA reaches a feasible solution after O (n log n) fitness evaluations.
(Phase 2) After a feasible solution is reached, then (1 + 1) EA searches for a 3-local optimum,
within the portion of the search space that consists of feasible solutions. In this phase the (1 +
1) EA maximizes the function x1. We call favorable any move up to a 3-bit flip that yields an
improvement on x1. The probability that the (1 + 1) EA performs a favorable move is 1/(en)3.
Hence, it is always possible to obtain an improvement on the fitness of at least 1, after expected
a linear number of favorable moves the (1 + 1) EA reaches a 3-local optimum. It follows that the
O(cid:0)n3(cid:1) fitness evaluations, unless the current solution is a 3-local optimum. Note that after at most
(1 + 1) EA finds a 3-local optimum as in Definition 17 after expected O(cid:0)n4(cid:1) fitness evaluations.
To prove that 3-local optima reach the desired approximation guarantee, we present an argu-
ment similar to that of Khanna et al. [19]. Let S be a 3-local optimum, and define I := S ∩ opt.
Since we cannot add a node outside of S to get the bigger independent set (see Definition 17),
then every node in S has at least one incoming edge from V \ S. Moreover, because of the local
optimality of S, there are at most S \ I nodes in opt \ I that have exactly one edge coming
14
from S. Thus, opt − S nodes in opt \ I must have at least two outgoing edges to S. This
implies that the minimum number of edges between S and opt \ I is S − I + 2(opt − S). If
we denote with E(S, opt \ I) the set of all edges between S and opt \ I, then it holds
2 opt
S
≤
S − I + 2(opt − S)
S
≤
E(S, opt \ I)
S
≤ Pv∈S(δ(v) + 1)
S
.
(15)
From the definition of 3-local optimum it follows that S is a maximum independent set. Hence,
we apply Corollary 8 to (15) and conclude that opt / S ≤ ab + 1/2, as claimed.
(cid:3)
7.2 Multi-objective optimization.
We also consider the multi-objective case for MIS. As in the previous sections, we consider the
bi-objective function f = (x1 , u(x)), with u(x) defined as
u(x) = −
xi
n
Xi=1
n
Xj=1
xjeij,
where eij is 1 if there is an edge between vi and vj, and it is 0 otherwise. The Gsemo attempts
to maximize both objectives simultaneously. To perform the run time analysis of Gsemo, we use
an argument similar to that of Halld´orsson and Radhakrishnan [20] for the greedy algorithm.
Theorem 19. Let G = (V, E) be a graph. For any node v ∈ V , denote with N (v) the set of nodes
adjacent to v. Let Sk := {v1, . . . , vk} be a sequence of k nodes s.t. vj is the node with smallest degree
i=1 (N (vi) ∪ vi) = ∅.
i=1 (N (vi) ∪ vi)(cid:17), and with Sk−1
in the complete sub-graph induced by V \(cid:16)Sj−1
Then Sk is a maximum independent set. Furthermore, it holds
Sk ≥ V (cid:18)Pv∈V δ(v)
V
+ 1(cid:19)−1
.
A proof of Theorem 19 is given in Halld´orsson and Radhakrishnan [20, Theorem 1].
We also consider a lemma that allows to obtain an upper-bound on the average degree of a
PLB network G. The following lemma holds.
Lemma 20. Let G = (V, E) be a PLB network with parameters β > 2, t ≥ 0, and with universal
constant c1. Then it holds
δ(v) ≤ 2c1n(t + 1)β−1
Xv∈V
i(i + t)−β.
∆(G)
Xi=1
A proof of Lemma 20 is given in more general terms in Brach et al. [16, Lemma 3.2]. We use
Theorem 19 and Lemma 20 to prove the following theorem.
Theorem 21. Let G = (V, E) be a connected PLB network with parameters β > 2, t ≥ 0 and with
a universal constant c1. Then the Gsemo finds a (cid:16) 2c1(β+t−1)
(β−1)(β−2) + 1(cid:17)-approximation of the MIS after
expected O(cid:0)n3(cid:1) fitness evaluations. In particular, the Gsemo is a Θ(1)-approximation algorithm
for the MIS on PLB networks.
Proof This argument is similar to the one given in Theorem 11 and Theorem 16. Following an
argument as in (Phase 1) in Theorem 11, and as in Theorem 16, one can prove that the solution
0n is added to the Pareto front after expected
n
Xj=1(cid:18)
j
e(n + 1)n(cid:19)−1
= O(cid:0)n2 log n(cid:1)
fitness evaluations. Note that the empty set is trivially an independent set. Once 0n is reached,
the Gsemo iteratively adds nodes vj as in Theorem 19 to the current solution, until a maximum
independent set is found. Again, since the population size is at most n+1, then selecting a particular
15
fitness at least that of a set Sk = {v1, . . . , vk}, that consists of a sequence of k nodes s.t. vj has
solution and adding a chosen node to it occurs after expected O(cid:0)n2(cid:1) fitness evaluations. Hence,
after expected O(cid:0)n3(cid:1) fitness evaluations the (1 + 1) EA reaches a maximum independent set with
minimum degree in the complete sub-graph induced by the nodes V \(cid:16)Sj−1
i=1 (N (vi) ∪ vi)(cid:17), as in
We conclude by proving that Sk = {v1, . . . , vk} yields the desired approximation guarantee,
using the PLB property. Let di be the number of nodes adjacent the node vi. From Lemma 20 it
holds
Lemma 20.
∆(G)
Xv∈V
δ(v) ≤ 2c1n(t + 1)β−1
Xi=1
≤ 2c1n(t + 1)β−1Z ∆(G)
1
i(i + t)−β
x(x + t)−βdx ≤
2c1n(β + t − 1)
(β − 1)(β − 2)
.
We combine this chain of inequalities with Theorem 19 to conclude that
Sk ≥ V (cid:20)Pv∈V δ(v)
V
+ 1(cid:21)−1
≥ n(cid:20) 2c1(β + t − 1)
(β − 1)(β − 2)
+ 1(cid:21)−1
.
Hence, it follows that
and the claim follows.
8 Conclusion
opt
Sk
≤
n
Sk
≤
2c1(β + t − 1)
(β − 1)(β − 2)
+ 1,
(cid:3)
In this paper, we look at the approximation ratio and run time analysis of commonly studied
evolutionary algorithms, for well known NP-complete covering problems, on power-law bounded
networks with exponent β > 2 (see Definition 4). We consider the single-objective (1 + 1) EA (see
Algorithm 1) and the multi-objective Gsemo (see Algorithm 2).
We prove in Section 4, that the (1+1) EA reaches a constant-factor approximation ratio for the
Minimum Dominating Set problem within O (n log n) fitness evaluations on power-law bounded
networks. Furthermore, we obtain and improved approximation guarantee for the Gsemo. In
Section 4, we discuss similar bounds for the Minimum Vertex Cover problem. We show that the
(1 + 1) EA and the Gsemo reach constant-factor approximation ratios in expected polynomial
fitness evaluations for the Minimum Connected Dominating Set and Maximum Independent Set
problems, in Section 6 and Section 7.
In all cases, we observe that the (1 + 1) EA and the Gsemo reach a constant-factor approxima-
tion ratio in polynomial time. This suggests that EAs implicitly exploit the topology of real-world
networks to reach a better solution quality than theoretically predicted in the general case.
References
1. S. Droste, T. Jansen, I. Wegener, On the analysis of the (1+1) evolutionary algorithm, Theoretical
Computer Science 276 (2002) 51 -- 81.
2. T. Jansen, I. Wegener, Evolutionary algorithms - how to cope with plateaus of constant fitness and
when to reject strings of the same fitness, IEEE Transactions on Evolutionary Computation 5 (2001)
589 -- 599.
16
3. O. Giel, I. Wegener, Evolutionary algorithms and the maximum matching problem, in: Proc. of STACS,
2003, pp. 415 -- 426.
4. F. Neumann, Expected runtimes of a simple evolutionary algorithm for the multi-objective minimum
spanning tree problem, European Journal of Operational Research 181 (2007) 1620 -- 1629.
5. F. Neumann, I. Wegener, Randomized local search, evolutionary algorithms, and the minimum span-
ning tree problem, Theoretical Computer Science 378 (2007) 32 -- 40.
6. P. S. Oliveto, J. He, X. Yao, Analysis of population-based evolutionary algorithms for the vertex cover
problem, in: Proc. of CEC, 2008, pp. 1563 -- 1570.
7. P. S. Oliveto, J. He, X. Yao, Evolutionary algorithms and the vertex cover problem, in: Proc. of CEC,
2007, pp. 1870 -- 1877.
8. T. Friedrich, N. Hebbinghaus, F. Neumann, J. He, C. Witt, Approximating covering problems by
randomized search heuristics using multi-objective models, in: Proc. of GECCO, 2007, pp. 797 -- 804.
9. T. Back, S. Khuri, An evolutionary heuristic for the maximum independent set problem, in: Proc. of
WCCI, 1994, pp. 531 -- 535.
10. X. Peng, Performance analysis of (1+1) EA on the maximum independent set problem, in: Proc. of
ICCCS, 2015, pp. 448 -- 456.
11. R. Kumar, P. Raghavan, S. Rajagopalan, A. Tomkins, Trawling the web for emerging cyber-
communities, Computer Networks 31 (1999) 1481 -- 1493.
12. J. Leskovec, J. Kleinberg, C. Faloutsos, Graphs over time: Densification laws, shrinking diameters and
possible explanations, in: Proc. of KDD, 2005, pp. 177 -- 187.
13. M. Faloutsos, P. Faloutsos, C. Faloutsos, On power-law relationships of the internet topology, in: Proc.
of SIGCOMM, 1999, pp. 251 -- 262.
14. M. E. J. Newman, The structure and function of complex networks, SIAM Review 45 (2003) 167 -- 256.
15. D. J. Watts, S. H. Strogatz, Collective dynamics of 'small-world' networks, Nature 393 (1998) 440 -- 442.
16. P. Brach, M. Cygan, J. Lacki, P. Sankowski, Algorithmic complexity of power law networks, in: Proc.
of SODA, 2016, pp. 1306 -- 1325.
17. B. Doerr, D. Johannsen, C. Winzen, Multiplicative drift analysis, Algorithmica 64 (4) (2012) 673 -- 697.
18. S. Guha, S. Khuller, Approximation algorithms for connected dominating sets, Algorithmica 20 (1998)
374 -- 387.
19. S. Khanna, R. Motwani, M. Sudan, U. Vazirani, On syntactic versus computational views of approx-
imability, SIAM Journal on Computing 28 (1998) 164 -- 191.
20. M. M. Halld´orsson, J. Radhakrishnan, Greed is good: Approximating independent sets in sparse and
bounded-degree graphs, Algorithmica 18 (1) (1997) 145 -- 163.
|
1101.0021 | 1 | 1101 | 2010-12-29T23:32:39 | Popular b-matchings | [
"cs.DS"
] | Suppose that each member of a set of agents has a preference list of a subset of houses, possibly involving ties and each agent and house has their capacity denoting the maximum number of correspondingly agents/houses that can be matched to him/her/it. We want to find a matching $M$, for which there is no other matching $M'$ such that more agents prefer $M'$ to $M$ than $M$ to $M'$. (What it means that an agent prefers one matching to the other is explained in the paper.) Popular matchings have been studied quite extensively, especially in the one-to-one setting. We provide a characterization of popular b-matchings for two defintions of popularity, show some $NP$-hardness results and for certain versions describe polynomial algorithms. | cs.DS | cs |
Popular b-Matchings
Katarzyna Paluch
Institute of Computer Science, Wrocław University
1 Introduction
In the paper we study popular b-matchings, which in other words are popular many-to-many matchings.
The problem can be best described in graph terms: We are given a bipartite graph G = (A ∪ H, E), a
capacity function on vertices b : A ∪ H → N and a rank function on edges r : E → N. A stands for
the set of agents and H for the set of houses. Each member of a set of agents a ∈ A has a preference list
Pa of a subset Ha of houses H. For a ∈ A and h ∈ H edge e = (a, h) belongs to E iff h is on Pa and
r((a, h)) = i reads that h belongs to (one of) a's ith choices. We say that a prefers h1 to h2 (or ranks
h1 higher than h2) if r((a, h1)) < r((a, h2)). If r(e1) < r(e2) we say that e1 has a higher rank
than e2. If there exist a ∈ A and h1, h2 ∈ Ha, h1 6= h2 such that r(e1 = (a, h1)) = r(e2 = (a, h2)),
then we say that e1, e2 belong to a tie and graph G contains ties. Otherwise we say that G does
not contain ties. A b-matching M of G is such a subset of edges that degM (v) ≤ b(v) for every
v ∈ A ∪ H, meaning that every vertex v has at most b(v) edges of M incident with it. Let r denote the
greatest rank (i.e. the largest number) given to any edge of E. With each agent a and each b-matching
M we associate a signature denoted as sigM (a), which is an r-tuple (x1, x2, . . . , xr) such that xi
(1 ≤ i ≤ r) is equal to the number of edges of rank i matched to a in a b-matching M. We introduce a
lexicographic order ≻ on signatures as follows. We will say that (x1, x2, . . . , xr) ≻ (y1, y2, . . . , yr)
if there exists j such that 1 ≤ j ≤ r and for each 1 ≤ i ≤ j − 1 there is xi = yi and xj > yj. We say
that an agent a prefers b-matching M ′ to M if sigM ′(a) ≻ sigM (a). M ′ is more popular than M,
denoted by M ′ ≻ M, if the number of agents that prefer M ′ to M exceeds the number of agents that
prefer M to M ′.
Definition 1 A b-matching M is popular if there exists no b-matching M ′ that is more popular than
M. The popular b-matching problem is to determine if a given triple (G, b, r) admits a popular
b-matching and find one if it exists.
Previous work The notion of popularity was first introduced by Gardenfors [3] in the one-to-one and
two-sided context, where two-sided means that both agents and houses express their preferences over
the other side and a matching M is popular if there is no other matching M ′ such that more participants
(i.e. agents plus houses) prefer M ′ to M than M to M ′. (He used the term of a majority 1assignment.)
He proved that every stable matching is a popular matching if there are no ties.
One-sided popular matchings were first studied in the one-to-one setting by Abraham et al. in [1]. They
proved that a popular matching needn't exist and decribed fast polynomial algorithms to compute a
popular matching, if it exists. The dynamic scenario, in which agents and houses can come and leave
the graph and agents can change their preference lists, is examined in [2]. Mestre [16] considered a
version in which every agent has an associated weight, reflecting an agent's priority. Manlove and Sng
in [13] extended an algorithm from [1] to the one-to-many setting (notice that this not equivalent to the
many-to-one setting.) In [17], in turn, the version, in which agents have weights, houses have arbitrary
capacities, every agent has capacity one and there are no ties, is examined.
Mahdian [12] showed for the one-to-one case without ties that a popular matching exists with high
probability when preference lists are random, and the number of houses is a small multiplicative factor
larger than the number of agents. Instances when a popular matching does not exist were dealt with
by McCutchen ([14]), who defined two notions of a matching that is, in some sense, as popular as
possible, namely a leastunpopularity- factor matching and a least-unpopularity-margin matching and
proved that computing either type of matching is N P -hard, even if there are no ties, by Huang et al.
[6] and by Kavitha et al. [10]. Kavitha and Nasre [8] gave an algorithm to compute an optimal popular
matching for various interpretations of optimality. McDermid and Irving [15] gave a characterisation
of the set of popular matchings for the one-to-one version without ties, which can be exploited to yield
algorithms for related problems.
Our contribution We provide a characterization of popular b-matchings and prove that the popular
b-matching problem is N P -hard even when agents use only two ranks and have capacity at most 2 and
houses have capacity one. This in particular answers the question about many-to-one popular match-
ings asked in [13]. Next we modify the notion of popularity and consider so-called weakly popular
b-matchings. We give their characterization and show that finding a weakly popular b-matching or
reporting that it does not exist is N P -hard even if all agents use at most three ranks, there are no ties
and houses have capacity one. We construct polynomial algorithms for the versions in which agents
use two ranks.
2 Characterization
First we introduce some terminology and recall a few facts from the matching theory.
By a path P we will mean a sequence of edges. Usually a path P will be denoted as (v1, v2, . . . , vk),
where v1, . . . , vk are vertices from the graph, not necessarrily all different, and for each i (1 ≤ i ≤
k−1) (vi, vi+1) ∈ E and no edge of G occurs twice in P . We will sometimes treat a path as a sequence
of edges and sometimes as a set of edges.
If M is a b-matching and degM (v) < b(v) we will say that v is unsaturated, if degM (v) = v(v) we
say that v is saturated. If b(v) = 1, then we will also use the terms matched and unmatched instead
of saturated and unsaturated. If e ∈ M we will call it an M-edge and otherwise -- a non-M-edge. By
M (v) we mean the set {v′ : (v, v′) ∈ M }. A path is said to be alternating (with respect to M) or
M-alternating if its edges are alternately M-edges and non-M-edges. An alternating path is said to
be (M-)augmenting if its end vertices are unsaturated and its first and last edges are non-M-edges.
For two sets Z1, Z2 Z1 ⊕ Z2 is defined as (Z1 \ Z2) ∪ (Z2 \ Z1). If M is a b-matching and P is
an alternating with respect to M path such that its beginning edge (v1, v2) is an M-edge or v1 is
unsaturated and its ending edge (vk−1, vk) is an M-edge or vk is unsaturated, then M ⊕ P is a also a
b-matching and if P is additionally augmenting, then M ⊕ P has more edges than M and is said to
have larger size or greater cardinality than M. A b-matching of maximum size is called a maximum
b-matching.
We will also need a notion of an even path: a path (a1, h1, a2, h2, . . . , v), where v denotes either hk
or ak+1, is defined to be even and denoted Pe(a1, v) if it is alternating, (a1, h1) is a non-M-edge and
for each i, 2 ≤ i ≤ k edges (hi−1, ai), (ai, hi) have the same rank. If this path is written in the reverse
e (v, a1). By writing (Pe(a, h), a′) we mean a path that consists of an even path
order we denote it by P r
Pe(a, h) and edge (h, a′).
2
Theorem 1 A b-matching M is popular iff graph G does not contain a path of one of the following
four types:
1. (Pe(a1, hk−1), ak, hk, ak+1), where (1) the path is alternating and (2) a1 is unsaturated or
there exists an edge e in M (a) such that r(e) > r(a1, h1), (3) r(hk−1, ak) > r(ak, hk) and (4)
agents a1, ak, ak+1 are pairwise different or a1 6= ak, a1 = ak+1 and r(a1, h1) < r(a1, hk),
2. (Pe(a1, a), P r
e (a, a′
that r(e) > r(a1, h1), (2) a′
r(e′) > r(a′
1) and (3) agents a1, a, a′
1, h′
1 are pairwise different,
1)), where (1) a1 is unsaturated or there exists an edge e in M (a1) such
1) such that
1 is unsaturated or there exists an edge e′ in M (a′
3. Pe(a1, h), (1) a1 is unsaturated or there exists an edge e in M (a) such that r(e) > r(a1, h1)
and (2) h is unsaturated,
4. (Pe(a1, hk), a1), where the path is alternating and r(hk, a1) > r(a1, h1).
Proof. Suppose the graph contains a path P1 of the first type.
(notice that e /∈ P1), otherwise let P ′
sigM (a1), sigM ′ (ak) ≻ sigM (ak), sigM (ak+1) ≻ sig′
we have sig′
M (a) = sigM (a). Therefore M ′ is more popular than M.
1 = P1. M ′ = M ⊕ P ′
If a1 is saturated let P ′
1 = P1 ∪ e
1 is a b-matching such that sigM ′(a1) ≻
M (ak+1) and for each a different from a1, ak, ak+1
If the graph contains a path of the second type, we proceed analogously.
Suppose the graph contains a path P3 of the third type. If a1 is saturated let P ′
3 = P3. M ′ = M ⊕ P ′
3 = P3 ∪ e, otherwise
3 is a b-matching such that sigM ′(a1) ≻ sigM (a1) and for each a different
let P ′
from a1 we have sig′
M (a) = sigM (a). Therefore M ′ is more popular than M.
Suppose now that there exists a b-matching M ′ which is more popular than M. This means that
the set A1 of agents who prefer M ′ to M outnumbers the set A2 of agents who prefer M to M ′.
For each a ∈ A1 we build a path Pa in the following way. We will use only edges of M ⊕ M ′. We
start with an edge (a, h1) ∈ M ′ \ M having the highest possible rank (i.e. lowest possible number).
Now assume that our so far built path Pa ends with hi. If hi is unsaturated in M, we end. Otherwise we
a (a′ ∈ A1). (The set
consider edges (hi, ai) belonging to M \M ′ and not already used by other paths P ′
of such unused edges is nonempty as hi is saturated in M and thus there are at least as many M-edges
as M ′-edges incident with hi and each time we arrive at hi while building some path Pa (a ∈ A1) we
use one M-edge and one M ′-edge.) If among these edges, there is such one that ai ∈ A1 we add it
to Pa and stop. Otherwise if there is such one that ai ∈ A2 we add it to Pa and stop. Otherwise we
add any unused edge (hi, ai) to Pa. ai has the same signature in M and in M ′. Therefore there exists
an edge (ai, hi+1) ∈ M ′ \ M having the same rank as (hi, ai) and there exists an unused edge of this
kind because the number of edges in M incident with ai is the same as the number of edges in M ′
incident with ai (as ai has the same signature in both b-matchings), we add this edge to Pa.
Clearly we stop building Pa at some point because we either arrive at an unsaturated vertex h ∈ H
or at a vertex of A1 ∪ A2, which may be a itself. Suppose there exists a ∈ A1 such that Pa ends on a.
Since we have started from an edge e of M ′(a) \ M (a) having the highest rank, the ending edge of Pa
must have a lower rank than e, hence Pa is a path of the fourth type. If there exists a path Pa ending on
an unsaturated vertex, it is of type three. If there exists a ∈ A1 such that Pa ends on a1 ∈ A1, a1 6= a,
then let e = (h′, a1) denote the ending edge of Pa. Since a1 ∈ A1 the edge of M (a1) \ M ′(a1) having
the highest rank, let us call it e′ has a higher rank than e. Suppose e′ = (a1, h). If there exists an edge
e3 = (h, a) ∈ M \ M ′, path Pa ∪ e′ ∪ e3 forms a path of type (1). Otherwise there exists an edge
e3 = (h, a2) ∈ M \ M ′, where a2 6= a and of course a2 6= a1 and path Pa ∪ e′ ∪ e3 also forms a
path of type (1). If none of the above paths exists, each Pa ends on some agent a2 ∈ A2. Because A2
3
outnumbers A1 there exist a1, a′
These paths are edge-disjoint and together form a path of type (2).
1 ∈ A1, a1 6= a′
1 and a2 ∈ A2 such that Pa1 and Pa′
1
both end on a2.
✷
Theorem 2 The problem of deciding whether a given triple (G, b, r) has a popular b-matching is N P -
hard, even if all edges are of one of two ranks, each applicant a ∈ A has capacity at most 2 and each
house h ∈ H has capacity 1.
Proof. We prove the theorem by showing a polynomial reduction of the exact 3-cover problem to
the popular b-matching problem. In the exact 3-cover problem we have a finite set K and a family
: i = 1, 2, . . . , m} of subsets of K such that Ti = 3 for 1 ≤ i ≤ m. We want to
Σ = {Ti
establish if there exists J ⊆ {1, 2, . . . , m} such that {Sj}j∈J forms a partition of K. The exact 3-
cover problem is NP-complete even when each element of K belongs to either two o three sets of
Σ and the underlining graph is planar [5], [4]. Given an instance of the exact 3-cover problem, we
construct the following graph G = (A ∪ H, E). We have a vertex vk ∈ H for each element k ∈ K,
with b(vk) = 1. For each set Ti we have 5 vertices ai1, ai2, ai3, ai4 , ai5 ∈ A with b(aij ) = 1 for j 6= 4
and b(ai4) = 2 and 2 vertices hi, h′
i) = 1. Additionally we have m − K/3
vertices g1, g2, . . . gm−K/3 ∈ H having capacity 1. The subgraph of G corresponding to a set Ti is
shown in Figure1. Each vertex vk is connected by a rank one edge with one of vertices ai1, ai2, ai3 for
each i such that k ∈ Ti. Each vertex ai5 is connected with each vertex of {g1, g2, . . . gm−K/3} by a
rank one edge.
i ∈ H with b(hi) = b(h′
Suppose there exists J which is an exact 3-cover of K. We build a popular b-matching M as
follows. For each Tj = {j1, j2, j3} such that j ∈ J we add edges (vj1, aj1), (vj2, aj2), (vj3 , aj3) to a
b-matching M. For each i = 1, 2, . . . , m we add edges (ai4, hi), (ai4 , h′
i) to M and for each i /∈ J we
add an edge (ai5, g), where g is some vertex of {g1, g2, . . . gm−K/3}. We claim that M is popular.
Since all vertices of H are saturated, there does not exist a path of type 3 from Theorem 1. Each
vertex ai4 is saturated and matched via one rank one edge and one rank two edge. If some vertex
aij is unsaturated, then it has no M-edges incident with it. Therefore it cannot belong to a path of
type 4. If it is saturated, then it is matched via a rank one edge and cannot belong to a path of type
4. If ai1 is unsaturated and its rank one edge is (ai1, v1), then v1 is matched via a rank one edge to
has only one rank one edge incident with it. Thus ai1 cannot
some ai′
be a beginning of a path of type 1. If ai1 has no rank one M-edge incident with it, then i /∈ J and
ai5 is saturated. Therefore path (ai1, hi, ai4, h′
i, ai5) is not a path of type 2. If ai5 is unmatched, then
ai1, ai2, ai3 are saturated and thus path (ai5 , h′
i, ai4, hi, ai1) is not a path of type 2. Also unmatched
ai5 does not belong to any path of type 1.
where j ∈ {1, 2, 3} and ai′
j
j
i, ai5, g1, ai′
Next we show, that if there is no exact 3-cover of K, then G has no popular b-matching. First let us
notice that if a b-matching M of G is popular, then all vertices of H are saturated in M. Next we can
see that each edge (ai4, h′
i) belongs to every popular b-matching of G, for if it does not belong to some
b-matching M, then path (ai4, h′
is a vertex matched to g1, is of type 1 from
Theorem 1. Also in every popular b-matching M ai4 must be saturated. For if edge (ai4, hi) does not
belong to M, then hi is matched to a vertex of {ai1, ai2, ai3 }, say ai1. Then path (ai4, hi, ai1 , vs, ai′
) ,
where vs is a neighbour of ai1 and ai′
is matched to vs, is a path of type 1 from Theorem 1. If there is
no exact 3-cover of K, then for some i only two or one vertices of {ai1, ai2, ai3 } will be matched via
a rank one edge. The number of such i clearly surpasses m − K/3. Suppose that ai1 is not matched
via a rank one edge and ai2, ai3 are. Then if path (ai1 , hi, ai4 , h′
i, ai5) is not to be a path of type 2, ai5
must be saturated. But only m − K/3 vertices of {ai5 : i = 1, 2, . . . m} can be saturated.
), where ai′
5
5
j
j
4
v_3
v_7 v_11
1
a_i1
1
1
a_i2
a_i3
2
2
2
h_i
2
a_i4
1
h'i
1
g_s
h_i
h_i
a_i4
a_i4
h'i
a_i5
a_i5
a_i5
g_2
1
g_1
h'_i
2
Figure 1: Solid lines denote edges of rank 1 and dotted lines edges of rank 2. In the second and third part of the figure thick lines indicate
M-edges. The second part describes the situation when Ti = {3, 7, 11} belongs to an exact 3-cover J and the third part the situation when
i /∈ J.
✷
3 Weakly popular b-matchings
Instead of just checking whether an agent prefers one b-matching to the other, we might also want to
assess how much he/she prefers one to the other.
Before defining a loss/gain factor of an agent we introduce the following. For a set E′ = {e1, e2, . . . , ek}
of edges incident with a, let s(E′) denote a d-tuple (x1, x2, . . . , xd) such that xi = r(eji) if 1 ≤ i ≤ k
and xi = r+1 if i > k, where ej1, ej2, . . . , ejk are edges of E′ ordered so that r(eij ) ≤ r(eij+1) for 1 ≤
j ≤ k − 1 and d = max{deg(a) : a ∈ A}. For two d-tuples d1 = (x1, . . . , xd), d2 = (y1, y2, . . . , yd)
we define their difference as d1 − d2 = Σd
i=1signum(xi − yi), where signum(x) = 1, 0, or −1 if
correspondingly x > 0, x = 0 or x < 0. Now we define a loss/gain factor of agent a with respect to
b-matchings M, M ′ as fa(M, M ′) = s(M \ M ∩ M ′) − s(M ′ \ M ∩ M ′). Also we define a loss/gain
factor for two b-matchings: f (M, M ′) = Σa∈Afa(M, M ′). We will say that a b-matching M is more
weakly popular than a b-matching M ′ if f (M, M ′) > 0 and a b-matching M ′ will be called weakly
popular if there does not exist a b-matching M ′ that is more weakly popular than M.
Theorem 3 A b-matching M is weakly popular iff graph G does not contain a path of type (1), (3) or
(4) from Theorem 1.
The proof is similar to that of Theorem 1.
Theorem 4 The problem of deciding whether a given triple (G, b, r) admits a weakly popular b-
matching is N P -hard, even if all edges are of one of three ranks, each applicant a ∈ A has capacity
at most 3, each house h ∈ H has capacity 1 and there are no ties.
Proof. We will reduce the 3-SAT problem to the Weakly Popular b-matching problem. assume we
have a formula F that uses k variables p1, p2, . . . , pk and has the form: (q1,1 ∨ q1,2 ∨ q1,3) ∧ . . . ∧
5
p'_{i,1}
1
b_{i,1}
p_{i,1}
1
g_{i,3}
1
p'_{i,3}
a_{i,1}
a_{i,,6}
b_{i,3}
1
p_{i,3}
g_{i,1}
h_j
a_{i,2}
p_{i,2}
1
a_{i,3}
c_{j,1}
c_{j,2}
c_{j,3}
a_{i,4}
a_{i,5}
b_{i,2}
1
g_{i,2}
p'_{i,2}
p_{1,k}
p'_{4,s}
p_{7,t}
Figure 2: A subgraph corresponding to variable pi such that r(i) = 3 and a subgraph corresponding to the exemplary 1st clause.
(qn,1 ∨ qn,2 ∨ qn,3), where each qj1,j2 (1 ≤ j1 ≤ n, 1 ≤ j2 ≤ 3) represents either pi or pi (meaning
not pi) for some i. For each pi let r(pi) denote the number of times pi occurs in F and r′(pi) -- the
number of times pi occurs in F . Let r(i) = max{r(pi), r′(pi)}.
We construct the following graph G = (A∪H, E). For each variable pi we have the following vertices
that will belong to H: pi,1, pi,2, . . . , pi,r(i) and p′
i,r(i) and bi,1, . . . , bi,r(i) and gi,1, . . . , gi,r(i).
For each of the n clauses we have one additional house hi. For each variable pi we have 2r(i) agents
ai,1, . . . , ai,2r(i), each of capacity 3. For each odd j agent ai,j is connected via a rank one edge
with pi,(j+1)/2 and for each even j agent ai,j is connected via a rank one edge with p′
i,j/2. Agents
ai,j, ai,j+1 for each odd j are connected via rank two edges with bi,(j+1)/2 for each odd j and agents
ai,j, ai,(j+1) mod 2r(i) for each even j are connected via rank three edges to gi,j/2 for each even j. For
each clause we have 3 agents: ci,1, ci,2, ci,3, each of capacity 2. All of them are connected via a rank
two edge with hi. If the 1st clause is of the form, say (p1 ∨ not p4 ∨ p7), then c1,1 is connected via a
rank one edge with p1,j for some j, c1,2 with p′
4,j for some j and c1,3 with p1,j for some j. No vertex
pi,j or p′
i,j is connected to two clause vertices. The construction is illustrated in Figure2.
i,1, . . . , p′
If F is true when pi = f (pi) (1 ≤ i ≤ k) for some function f : {p1, . . . , pk} → {0, 1}, we build
the following b-matching M. If f (pi) = 1 (1 ≤ i ≤ k), we add to M all rank one, two and three
edges incident with agents ai,j such that j is even, otherwise M will contain all rank one, two and
three edges incident with agents ai,j such that j is odd. Since F is true, in the first clause we have
f (p1) = 1 ∨ f (p4) = 0 ∨ f (p7) = 1. Suppose that f (p4) = 0, then we add to M a rank one edge
incident with c1,2, we are able to do so, because at the moment all vertices p′
4,j are unmatched. We
also add to M a rank two edge incident with c1,2. We proceed in this way for every clause. At the end
for every rank one edge (a, h) such that a and h are unsaturated we add it to M. One can check that
thus built M is weakly popular.
Suppose there exists a weakly popular b-matching M. We will show that there exists a function
f : {p1, . . . , pk} → {0, 1} such that F is true when pi = f (pi) (1 ≤ i ≤ k). The key observation
6
a_{i,3}
a_{i,,6}
a_{i,4}
a_{i,5}
b_{i,2}
b_{i,3}
g_{i,2}
Figure 3: Thick lines indicate M-edges. If we match gi,2 to ai,4, we get a path (ai,5, gi,2, ai,4, bi,2, ai,3),
which is of type (1).
is that for any variable pi either all houses pi,j will be matched to agents ai,j or all houses p′
i,j will
be matched to agents ai,j. We show this as follows. We notice that in the subgraph corresponding to
variable pi either all agents ai,j for even j are matched to houses bi,j/2 or all agents ai,j for odd j are
matched to houses bi,(j+1)/2. Figure 3 shows what happens when it is not the case. Then there exist
two agents ai,j, ai,j+1 such that both are not matched via a rank two edge and the graph contains a path
of type 2, which means that M is not weakly popular. Suppose that all agents ai,j such that j is even
are matched via a rank two edge. Then they must also be matched via a rank one edge. For suppose
that ai,2 is not matched to p′
i,1) is either a beginning of a path of type (1)
or is a path of type (3). Thus all houses p′
i,j are matched to agents ai,j. We set f (pi) = 1. Next, let
us consider clauses. Suppose the 1st clause is of the form (p1 ∨ not p4 ∨ p7) and ci,1 is matched to
hi. Then ci,1 must also matched to some p1,j, because otherwise (ci,2, hi, ci,j, p1,j) is a beginning of a
path of type 1 or 3, which shows that the 1st clause is satisfied under function f .
✷
i,1. Then (ai,1, bi,1, ai,2, p′
Theorem 5 The problem of deciding whether a given triple (G, b, r) admits a weakly popular b-
matching is N P -hard, even if each applicant a ∈ A has at most 3 edges of rank 1 incident with
him/her and at most 1 edge of rank 2 and capacity at most 4 and each house h ∈ H has capacity 1.
The proof is by reducing the 3-exact cover problem to the problem from the theorem.
4 Polynomial algorithms
We start with the case when there are no ties and each agent uses at most two ranks, houses have
arbitrary capacities.
We are going to need an algoritm for the following problem. We have a bipartite graph G =
(A ∪ H, E),b : A ∪ A → N such that b(a) = 1 for every a ∈ A, a partition of A into A1, A2, . . . , Ap
and a nonnegative integer ki for each 1 ≤ i ≤ p. We want to find a maximum b-matching M of G
such that for each 1 ≤ i ≤ p the number of vertices of Ai matched in M is at least ki or ascertain that
such a matching does not exist. Here we will use the term of a matching instead of a b-matching.
Let Mmax denote any maximum matching of G. Clearly if a matching satisfying our require-
ments exists (further on we will call such a matching a partition matching), Σp
i=1ki must not surpass
Mmax. Suppose that matching M is of maximum cardinality but it matches less than ki vertices
of Ai. Then if the solution to our problem exists, there is in the graph a sequence P1, P2, . . . , Ps of
edge-disjoint alternating paths such that each Pj = (aj, . . . , a′
j) begins with a non-M-edge and ends
7
with an M-edge, a1 is of Ai and is unmatched in M, as is of Ai′ such that M matches more than ki′
vertices of Ai′ and for each j ≤ s − 1 vertices a′
j and aj+1 are of the same set At for some t and they
either denote the same vertex or aj+1 is unmatched in M. Let us call such a sequence an improving
sequence. (To see that an improving sequence exists, let M ′ denote any partition matching and con-
sider M ⊕ M ′. Since in M ′ at least ki vertices are matched there exists in M ⊕ M ′ an alternating path
1 /∈ Ai. If a′
1 beginning with an unmatched in Ai vertex and ending on some vertex a′
P ′
1 ∈ At and At
is such that M matches exactly kt vertices of At, then there exists an alternating path P ′
2 (edge-disjoint
from P ′
1 or at a vertex a2 ∈ At unmatched in M and ending on a vertex a3 inAs
such that t 6= s and i 6= s. Proceeding in this way we obtain an improving sequence.) The algorithm
for this problem can be then described as follows.
1) beginning either at a′
Algorithm Partition Matching
Input: graph G = (A∪H, E), b : A∪H → N, a partition of A into A1, A2, . . . , Ap, a sequence of nonnegative integers (k1, k2, . . . , kp)
Output: b-matching M of maximum cardinality that for each 1 ≤ i ≤ p matches at least ki vertices of Ai, or a report that such a matching
does not exist.
Find any maximum b-matching M of G.
while M does not satisfy requirements:
Find an improving sequence P .
If P does not exist write "does not exist" and halt.
Otherwise M := M ⊕ P .
We need the above algorithm for the following. Suppose agents Ah = {a1, a2, . . . , ak} have all
a rank two edge incident with house h that has capacity c < k (and has no rank one edges incident
with it). Then we are able to match only p agents of Ah to h. If we match a2 to h, then a2 should also
be matched to a rank one house, because otherwise for any a1 not matched to h, path (a1, h, a2, h′)
(where h′ is a rank one house for a2) forms a beginning of a path of type 1 or a path of type 3. Thus
we should find such a M1 matching among rank one edges that p agents of Ah are matched in M1.
Algorithm A
Input: graph G = (A ∪ H, E), function b : A ∪ H → N, a partition of E = E1 ∪ E2
Output: a weakly popular b-matching M or a report that it does not exist
Let G1 = (A ∪ H, E1) and b1 be defined as b1(a) = 1 for a ∈ A and b1(h) = b(h).
Let G2 = (A ∪ H, E2) and b2 be defined as b2(a) = 1 for a ∈ A and b2(h) = b(h) − degE1 (h) for h ∈ H.
If in G2 every h ∈ H satisfies: degE2 (h) ≤ b2(h), then
find a maximum b2-matching M2 of G2, a max. b1-matching M1 of G1 and output M = M1 ∪ M2.
Otherwise let H ′ = (h1, h2, . . . , hp) denote all houses h ∈ H such that degE2 (h) > b2(h). For each hi ∈ H ′ build Ai = NG2 (hi),
where NG2 (hi) denotes the set of neighbours of hi in G2. Set ki = b2(hi). Set Ap+1 = A \ S1≤i≤p Ai and kp+1 = 0.
Using algorithm Partition Matching for the input: G1, b1, the partition of A into A1, . . . , Ap+1 and (k1, . . . , kp+1) compute a partition
matching M1.
If it does not exist write "does not exist" and halt.
Otherwise for each i, 1 ≤ i ≤ p let A′
each a ∈ Ap+1 add (a, h), such that (a, h) ∈ E2, to M2.
Output M = M1 ∪ M2.
i ⊆ Ai denote the set of agents that are matched in M1. For each a ∈ A′
i add (a, hi) to M2. For
8
2
1
2
1
2
1
2
1
1
2
2
2
1
Figure 4: The graph on the left has a weakly popular b-matching and the one on the right has not.(Houses have capacity 1.)
Theorem 6 Algorithm A solves the Weakly Popular b-Matching problem for the cases when b(a) = 2
for each a ∈ A, each agent uses 2 ranks and there are no ties.
Proof. Suppose that algorithm A computes a b-matching M. To prove that it is weakly popular, it
suffices by Theorem 3 to show that the graph does not contain a path of type (1), (3) or (4).
Since M is of maximum cardinality, the graph does not contain a path of type 3 and since there
are no ties, the graph does not contain a path of type 4. The graph does not contain a path of type
1 either, because, if such a path existed, then it would have the form (a1, h1, a2, h2, a3) such that a1
is unsaturated in M, edges (a1, h1), (h1, a2) are of rank 2 and edges (a2, h2), (h2, a3) are of rank 1.
However M has the property that if agent a is matched with a rank 2 edge, then he/she is also matched
with a rank 1 edge.
On the other hand if algorithm A fails to compute a weakly popular b-matching, then it is because
an appropriate partition matching does not exist. First let us notice that if a weakly popular b-matching
exists, then it is of maximum cardinality. Next we can show, that if in the graph G a weakly popular
b-matching M exists, then there exists a weakly popular b matching M ′ such that M ′ contains some
maximum b1-matching of G1, where b1 and G1 are as in the description of Algorithm A. For assume,
that a weakly popular b-matching M restricted to rank one edges (called M1) is not a maximum b1-
matching of G1. Then in G1 there exists an M1-augmenting path, which must be of the form (a, h),
where a is not matched in M1 and h is not saturated in M1. Since M is weakly popular h is saturated
in M. Therefore there exists a′ such that (a′, h) ∈ M and (a′, h) ∈ E2. It is not difficult to see that
M \ {(a′, h)} ∪ {(a, h)} must also be a weakly popular b-matching of G. We can proceed in this way
until we have a weakly popular b-matching of G that contains a maximum b1-matching of G1.
Thus if there exists a weakly popular b-matching M of G, then there exists such a weakly popular
b-matching M ′ that M ′ restricted to rank two edges is a maximum b2-matching of G2. If for some h
we have degE2(h) > b2(h), then in any maximum b2-matching of G2 exactly b2(h) vertices of NG2(h)
will be matched and the remaining degE2 (h) − b2(h) vertices of G2(h) will be unmatched. Therefore
for such h there will always be in G a path (a, h, a′) such that a, a′ ∈ NG2(h) and a is not matched in
G2. So if such a path is not to become a beginning od a path of type 1, each a′ that is matched in G2
should also be matched in G1. Therefore if there exists a weakly popular b-matching of G, there exists
an appropriate partition matching of G1.
✷
Next we are going to deal with the case when ties are allowed among rank two edges and each
agent has capacity 2 . Suppose that a function b : A ∪ H → N is such that b(a) = 1 for each a ∈ A
and M is a maximum b-matching. Then a vertex v ∈ A ∪ H will be called an O-vertex if there exists
an odd-length alternating path from un unmatched (in M) vertex v0 to v and an E-vertex if there
exists an even-length alternating path from un unmatched (in M) vertex v0 to v. All other vertices
will be called U-vertices (unreachable via an alternating path from un unmatched vertex). By the
Gallai-Edmonds decomposition theorem O-, E- and U-vertices forms a partition of A ∪ H, which is
9
A2
A3
A4
A9
A8
A1
A6
A5
A10
A7
Figure 5: Thick lines indicate M2-edges. Houses have capacity 1. All agents except for A8 are O-vertices. Agent A8 is a U-vertex.
Here two sets will belong to a partition of A: {A1, A2, A3, A4, A5, A6} and {A9, A10}. d({A9, A10} = 1.
independent of a particular maximum b-matching M. (See [11], [7] for example.)
We will need an algorithm for the following problem. The input is as for Algorithm Partition
Matching and additionally for each i, 1 ≤ i ≤ p we have a family Zi of subsets of Ai, each Z ∈ Zi
having cardinality ki. We want to find a maximum matching M of G such that for each 1 ≤ i ≤ p
vertices of Ai matched in M contain some Z ∈ Zi (such a matching is going to be called a z-partition
matching) or ascertain that such a matching does not exist.
Let us first explain what we need this for. Suppose a maximum matching M2 on some subgraph
of G2 looks as shown in Figure 5. Then if we want to avoid creating paths of type (1), we should have
that agents A2, A3, A4, A7, A10 have a rank one M1-edge incident with them. Notice that we do not
have to worry about A8. If agents Z0 = {A1, A3, A6, A4, A9} have a rank one M1-edge incident with
them, then we can change M2, so that it will still be maximum and will saturate vertices of Z0 and the
graph will not contain a path of type 1.
With each subset Ai of the partition of A there will be associated a graph G′
i) and
a function b2 (such that b(a) = 1 for each a ∈ A.) We define Zi so that it contains each Z ⊆ Ai such
that there exists a maximum b2-matching M of G′
i having the property that vertices of Ai matched
in M form the set Z. Notice that all sets in Zi have the same cardinality. For each A′
i ⊆ Ai define
d(A′
i, ∃Z∈ZiB ⊆ Z}. Suppose that in Ai, the set of matched in the current
b1-matching M1 vertices equals A′
i) = max{B : B ⊆ A′
i = (Ai ∪ Hi, E′
i. Then let si = A′
i − d(A′
i).
Lemma 1 For each A′
i ⊆ Ai, we can compute d(A′
i) in polynomial time.
i. Let B′ ⊆ A′
i) = B. Otherwise check if there exists in G′
Proof. It suffices to compute the largest subset B ⊆ A such that there exists a maximum b2-matching
M of G′
i, that matches all vertices of B. To this end, first compute any maximum b2-matching M
of G′
i, set B = B′ and
d(A′
i an alternating path P = (a1, h1, a2, h2, . . . , an)
such that a1 ∈ A′
i. If P exists, set M = M ⊕ P . This way new M is clearly
of maximum cardinality (as P is of even length) and B′ has increased by one. Repeat computing an
alternating path of this type as long as possible. At the end B′ will be our desired set B.
✷
i denote vertices matched in M. If B′ = M or B′ = A′
i \ B′ and an /∈ A′
If Ai of the partition of A is such that si > 0 we will say that it is excessive and if it such that
i) < Z, where Z ∈ Zi, we will say that it is deficient. Notice that a set can be both excessive
d(A′
and deficient. If a set is neither deficient nor excessive, then we call it equal.
Suppose we have a maximum b1-matching M1 of G1, but it is not a z-partition matching. Then in
the partition of A there is at least one deficient set. We define a z-improving sequence as a sequence
10
P1, P2, . . . , Ps of edge-disjoint alternating paths such that each Pj = (aj, . . . , a′
j) begins with a non-
M1-edge and ends with an M1-edge, a1 is from a deficient set, as is from an excessive set, for each
j ≤ s − 1 vertices a′
j and aj+1 are of the same equal set At of the partition of A (they can denote the
same vertex) and for each j ≤ s − 1 if a′
j} ∪ {aj+1} ∈ Zt.
Lemma 2 A z-improving sequence, if exists, can be found in polynomial time.
j ∈ At, then A′
t \ {a′
Lemma 3 If z-improving sequence (with respect to any maximum b1-matching of G1) does not exist,
then G1 does not contain a z-partition matching.
Algorithm for computing a z-partition matching looks exactly as algorithm for a partition match-
ing, only it finds a z-improving sequence instead of an improving sequence.
Algorithm A' is similar to Algorithm A, but it computes a different partition of A′. The partition
in Algorithm A' is established as follows. First we compute a maximum b2-matching M ′
2 of G2. Next
we find the (E, O, U )-partition of A into O−, E− and U-vertices. Two O-vertices a1, a2 of A will
belong to one set of partition of A iff there exists h ∈ H such that there exists an M ′
2-alternating path
(a). Additional set,
from a1 to h and from a2 to h. For such a set Ai we will have ki = Pa∈Ai
say (p + 1)th of the partition will be formed by the remaining agents of A and we will have kp+1 = 0.
degM ′
2
References
[1] D.J. Abraham, R.W. Irving, T. Kavitha, and K. Mehlhorn. Popular matchings. SIAM Journal on
Computing, 37:1030-1045, 2007.
[2] D.J. Abraham, T. Kavitha: Voting Paths. SIAM J. Discrete Math. 24(2): 520-537 (2010).
[3] P. Gardenfors: Match making: Assignments based on bilateral preferences. Behavioural Sci-
ences, 20 (1975), pp. 166-173.
[4] G. Cornuejols. General factors of graphs. J. Comb. Theory, Ser.B 45(2):185-198 (1988).
[5] M.E. Dyer and A.M.Friese, Planar 3DM is NP-complete. J.Algorithms 7 (1986), 174-184.
[6] C.-C. Huang, T. Kavitha, D. Michail, and M. Nasre. Bounded unpopularity matchings. In Pro-
ceedings of SWAT 2008: the 12th Scandinavian Workshop on Algorithm Theory, volume 5124
of Lecture Notes in Computer Science, pages 127-137. Springer, 2008.
[7] R.W. Irving, T. Kavitha, K. Mehlhorn, D. Michail, and K. Paluch. Rank-maximal matchings.
ACM Transactions on Algorithms, 2(4):602-610, 2006.
[8] T. Kavitha and M. Nasre. Optimal popular matchings. In Proceedings of Match-UP 2008: Work-
shop on Matching Under Preferences - Algorithms and Complexity, held at ICALP 2008, pages
46-54, 2008.
[9] T. Kavitha and C.D. Shah. Efficient algorithms for weighted rank-maximal matchings and related
problems. In Proceedings of ISAAC 2006: the Seventeenth International Symposium on Algo-
rithms and Computation, volume 4288 of Lecture Notes in Computer Science, pages 153-162.
Springer, 2006.
11
[10] Telikepalli Kavitha, Julián Mestre, Meghana Nasre: Popular Mixed Matchings. ICALP (1) 2009:
574-584.
[11] L.Lovasz, M.D.Plummer: Matching Theory. Ann. Discrete Math. 29, North-Holland, Amster-
dam, 1986.
[12] M. Mahdian. Random popular matchings. In Proceedings of EC 06: the 7th ACM Conference on
Electronic Commerce, pages 238-242. ACM, 2006.
[13] D.F. Manlove and C.T.S. Sng. Popular matchings in the Capacitated House Allocation problem.
In Proceedings of ESA 06: the 14th Annual European Symposium on Algorithms, volume 4168
of Lecture Notes in Computer Science, pages 492-503. Springer, 2006.
[14] R.M. McCutchen. The least-unpopularity-factor and least-unpopularity-margin criteria for
matching problems with one-sided preferences. In Proceedings of LATIN 2008: the 8th Latin-
American Theoretical INformatics symposium, volume 4957 of Lecture Notes in Computer Sci-
ence, pages 593-604. Springer, 2008.
[15] E. McDermid and R.W. Irving. Popular Matchings: Structure and Algorithms. COCOON 2009:
506-515.
[16] J. Mestre. Weighted popular matchings. In Proceedings of ICALP 06: the 33rd International Col-
loquium on Automata, Languages and Programming, volume 4051 of Lecture Notes in Computer
Science, pages 715-726. Springer, 2006.
[17] C.T.S. Sng and D.F. Manlove. Popular matchings in the weighted capacitated house allocation
problem. J. Discrete Algorithms 8(2): 102-116 (2010)
12
|
1007.5450 | 1 | 1007 | 2010-07-30T13:49:18 | Known Algorithms on Graphs of Bounded Treewidth are Probably Optimal | [
"cs.DS",
"cs.CC",
"cs.DM"
] | We obtain a number of lower bounds on the running time of algorithms solving problems on graphs of bounded treewidth. We prove the results under the Strong Exponential Time Hypothesis of Impagliazzo and Paturi. In particular, assuming that SAT cannot be solved in (2-\epsilon)^{n}m^{O(1)} time, we show that for any e > 0; {\sc Independent Set} cannot be solved in (2-e)^{tw(G)}|V(G)|^{O(1)} time, {\sc Dominating Set} cannot be solved in (3-e)^{tw(G)}|V(G)|^{O(1)} time, {\sc Max Cut} cannot be solved in (2-e)^{tw(G)}|V(G)|^{O(1)} time, {\sc Odd Cycle Transversal} cannot be solved in (3-e)^{tw(G)}|V(G)|^{O(1)} time, For any $q \geq 3$, $q$-{\sc Coloring} cannot be solved in (q-e)^{tw(G)}|V(G)|^{O(1)} time, {\sc Partition Into Triangles} cannot be solved in (2-e)^{tw(G)}|V(G)|^{O(1)} time. Our lower bounds match the running times for the best known algorithms for the problems, up to the e in the base. | cs.DS | cs |
Known Algorithms on Graphs of Bounded Treewidth are
Probably Optimal
Daniel Lokshtanov∗
D´aniel Marx†
Saket Saurabh‡
May 28, 2018
Abstract
We obtain a number of lower bounds on the running time of algorithms solving problems on
graphs of bounded treewidth. We prove the results under the Strong Exponential Time Hypothesis
of Impagliazzo and Paturi. In particular, assuming that SAT cannot be solved in (2 − ǫ)nmO(1) time,
we show that for any ǫ > 0;
• INDEPENDENT SET cannot be solved in (2 − ǫ)tw(G)V (G)O(1) time,
• DOMINATING SET cannot be solved in (3 − ǫ)tw(G)V (G)O(1) time,
• MAX CUT cannot be solved in (2 − ǫ)tw(G)V (G)O(1) time,
• ODD CYCLE TRANSVERSAL cannot be solved in (3 − ǫ)tw(G)V (G)O(1) time,
• For any q ≥ 3, q-COLORING cannot be solved in (q − ǫ)tw(G)V (G)O(1) time,
• PARTITION INTO TRIANGLES cannot be solved in (2 − ǫ)tw(G)V (G)O(1) time.
Our lower bounds match the running times for the best known algorithms for the problems, up to the
ǫ in the base.
1 Introduction
It is well-known that many NP-hard graph problems can be solved efficiently if the treewidth (tw(G)) of
the input graph G is bounded. For an example, an expository algorithm to solve VERTEX COVER and
INDEPENDENT SET running in time O∗(4tw(G)) is described in the algorithms textbook by Kleinberg
and Tardos [15] (the O∗ notation suppresses factors polynomial in the input size), while the book of
Niedermeier [20] on fixed-parameter algorithms presents an algorithm with running time O∗(2tw(G)).
Similar algorithms, with running times on the form O∗(ctw(G)) for a constant c, are known for many other
graph problems such as DOMINATING SET, q-COLORING and ODD CYCLE TRANSVERSAL [1, 9, 10,
27]. Algorithms for graph problems on bounded treewidth graphs have found many uses as subroutines
in approximation algorithms [7, 8], parameterized algorithms [6, 19, 26], and exact algorithms [12, 23,
28].
In this paper, we show that any improvement over the currently best known algorithms for a number
of well-studied problems on graphs of bounded treewidth would yield a faster algorithm for SAT. In
particular, we show if there exists an ǫ > 0 such that
• INDEPENDENT SET can be solved in O∗((2 − ǫ)tw(G)) time, or
• DOMINATING SET can be solved in O∗((3 − ǫ)tw(G)) time, or
• MAX CUT can be solved in O∗((2 − ǫ)tw(G)) time, or
∗Department of Informatics, University of Bergen, Norway. [email protected]
†School of Computer Science, Tel Aviv University, Tel Aviv, Israel. [email protected]
‡The Institute of Mathematical Sciences, India. [email protected]
1
• ODD CYCLE TRANSVERSAL can be solved in O∗((3 − ǫ)tw(G)) time, or
• there is a q ≥ 3 such that q-COLORING can be solved in O∗((q − ǫ)tw(G)) time, or
• PARTITION INTO TRIANGLES can be solved in O∗((2 − ǫ)tw(G)) time,
then SAT can be solved in O∗((2 − δ)n) time for some δ > 0. Here n is the number of variables in
the input formula to SAT. Such an algorithm would violate the Strong Exponential Time Hypothesis
(SETH) of Impagliazzo and Paturi [13]. Thus, assuming SETH, the known algorithms for the mentioned
problems on graphs of bounded treewidth are essentially the best possible.
To show our results we give polynomial time many-one reductions that transform n-variable boolean
formulas φ to instances of the problems in question. Such reductions are well-known, but for our results
we need to carefully control the treewidth of the graphs that our reductions output. A typical reduction
creates n gadgets corresponding to the n variables; each gadget has a small constant number of vertices.
In most cases, this implies that the treewidth can be bounded by O(n). However, to prove the a lower
bound of the form O∗((2 − ǫ)tw(G)), we need that the treewidth of the constructed graph is (1 + o(1))n.
Thus we can afford to increase the treewidth by at most one per variable. For lower bounds above
O∗((2 − ǫ)tw(G)), we need even more economical constructions. To understand the difficulty, consider
the DOMINATING SET problem, here we want to say that if DOMINATING SET admits an algorithm
with running time O∗((3 − ǫ)tw(G)) = O∗(2log(3−ǫ)tw(G)) for some ǫ > 0, then we can solve SAT on
input formulas with n-variables in time O∗((2 − δ)n) for some δ > 0. Therefore by naıvely equating
the exponent in the previous sentence we get that we need to construct an instance for DOMINATING
SET whose treewidth is essentially n
log 3. In other words, each variable should increase treewidth by less
than one. The main challenge in our reductions is to squeeze out as many combinatorial possibilities
per increase of treewidth as possible. In order to control the treewidth of the graphs we construct, we
upper bound the pathwidth (pw(G)) of the constructed instances and use the fact that for any graph G,
tw(G) ≤ pw(G). Thus all of our lower bounds also hold for problems on graphs of bounded pathwidth.
Complexity Assumption: The Exponential Time Hypothesis (ETH) and its strong variant (SETH) are
conjectures about the exponential time complexity of k-SAT. The k-SAT problem is a restriction of
SAT, where every clause in input boolean formula φ has at most k literals. Let sk = inf{δ : k-SAT
can be solved in 2δn time}. The Exponential Time Hypothesis conjectured by Impagliazzo, Paturi and
Zane [14] is that s3 > 0. In [14] it is shown that ETH is robust, that is s3 > 0 if and only if there is a
k ≥ 3 such that sk > 0. In the same year it was shown that assuming ETH the sequence {sk} increases
infinitely often [13]. Since SAT has a O∗(2n) time algorithm, {sk} is bounded by above by one, and
Impagliazzo and Paturi [13] conjecture that 1 is indeed the limit of this sequence.
In a subsequent
paper [3], this conjecture is coined as SETH.
While ETH is now a widely believed assumption, and has been used as a starting point to prove
running time lower bounds for numerous problems [5, 4, 11, 18, 17], SETH remains largely untouched
(with one exception [21]). The reason for this is two-fold. First, the assumption that limk→∞ sk = ∞
is a very strong one. Second, when proving lower bounds under ETH we can utilize the Sparsification
Lemma [14] which allows us to reduce from instances of 3-SAT where the number of clauses is linear in
the number of variables. Such a tool does not exist for SETH, and this seems to be a major obstruction for
showing running time lower bounds for interesting problems under SETH. We overcome this obstruction
by circumventing it – in order to show running time lower bounds for algorithms on bounded treewidth
graphs sparsification is simply not required. We would like to stress that our results make sense, even if
one does not believe in SETH. In particular, our results show that one should probably wait with trying
to improve the known algorithms for graphs of bounded treewidth until a faster algorithm for SAT is
around.
Related Work. Despite of the importance of fast algorithms on graphs of bounded treewidth or path-
width, there is no known natural graph problem for which we know an algorithm outperforming the
2
naıve approach on bounded pathwidth graphs. For treewidth, the situation is slightly better: Alber et
al. [1] gave a O∗(4tw(G)) time algorithm for DOMINATING SET, improving over the natural O∗(9tw(G))
algorithm of Telle and Proskurowski [25]. Recently, van Rooij et al. [27] observed that one could use
fast subset convolution [2] to improve the running time of algorithms on graphs of bounded treewidth.
Their results include a O∗(3tw(G)) algorithm for DOMINATING SET and a O∗(2tw(G)) time algorithm
for PARTITION INTO TRIANGLES. Interestingly, the effect of applying subset convolution was that the
running time for several graph problems on bounded treewidth graphs became the same as the running
time for the problems on graphs of bounded pathwidth.
In [27], van Rooij et al. believe that their algorithms are probably optimal, since the running times
of their algorithms match the size of the dynamic programming table, and that improving the size of
the table without losing time should be very difficult. Our results prove them right: improving their
algorithm is at least as hard as giving an improved algorithm for SAT.
2 Preliminaries
In this section we give various definitions which we make use of in the paper. Let G be a graph with
vertex set V (G) and edge set E(G). A graph G′ is a subgraph of G if V (G′) ⊆ V (G) and E(G′) ⊆
E(G). For subset V ′ ⊆ V (G), the subgraph G′ = G[V ′] of G is called a subgraph induced by V ′
if E(G′) = {uv ∈ E(G) u, v ∈ V ′}. By N (u) we denote (open) neighborhood of u in graph G that
is the set of all vertices adjacent to u and by N [u] = N (u) ∪ {u}. Similarly, for a subset D ⊆ V , we
define N [D] = ∪v∈DN [v].
A tree decomposition of a graph G is a pair (X , T ) where T is a tree and X = {Xi i ∈ V (T )} is a
collection of subsets of V such that: 1. Si∈V (T ) Xi = V (G), 2. for each edge xy ∈ E(G), {x, y} ⊆ Xi
for some i ∈ V (T ); 3. for each x ∈ V (G) the set {i x ∈ Xi} induces a connected subtree of T . The
width of the tree decomposition is maxi∈V (T ){Xi − 1}. The treewidth of a graph G is the minimum
width over all tree decompositions of G. We denote by tw(G) the treewidth of graph G.
If in the
definition of treewidth we restrict the tree T to be a path then we get the notion of pathwidth and denote
it by pw(G). For our purpose we need an equivalent definition of pathwidth via mixed search games.
In a mixed search game, a graph G is considered as a system of tunnels. Initially, all edges are
contaminated by a gas. An edge is cleared by placing searchers at both its end-points simultaneously
or by sliding a searcher along the edge. A cleared edge is re-contaminated if there is a path from
an uncleared edge to the cleared edge without any searchers on its vertices or edges. A search is a
sequence of operations that can be of the following types: (a) placement of a new searcher on a vertex;
(b) removal of a searcher from a vertex; (c) sliding a searcher on a vertex along an incident edge and
placing the searcher on the other end. A search strategy is winning if after its termination all edges are
cleared. The mixed search number of a graph G, denoted by ms(G), is the minimum number of searchers
required for a winning strategy of mixed searching on G. Takahashi, Ueno and Kajitani [24] obtained
the following relationship between pw(G) and ms(G), which we use for bounding the pathwidth of the
graphs obtained in reduction.
Proposition 1 ([24]). For a graph G, pw(G) ≤ ms(G) ≤ pw(G) + 1.
An instance to SAT will always consists of a boolean formula φ = C1 ∧ · · · ∧ Cm over n variables
{v1, . . . , vn} where each clause Ci is OR of one or more literals of variables. We also denote a clause Ci
by the set {ℓ1, ℓ2, . . . , ℓc} of its literals and denote by Ci the number of literals in Ci. An assignment
τ to the variables is an element of {0, 1}n, and it satisfies the formula φ if for every clause Ci there
is literal that is assigned 1 by τ . We say that a variable vi satisfies a clause Cj if there exists a literal
corresponding to vi in {ℓ1, ℓ2, . . . , ℓc} and it is set to 1 by τ . A group of variables satisfy a clause Cj if
there is a variable that satisfies the clause Cj. All the sections in this paper follows the following pattern:
definition of the problem; statement of the lower bound; construction used in the reduction; correctness
of the reduction; and the upper bound on the pathwidth of the resultant graph.
3
3 Independent Set
An independent set of a graph G is a set S ⊆ V (G) such that G[S] contains no edges. In the INDEPEN-
DENT SET problem we are given a graph G and the objective is to find an independent set of maximum
size.
Theorem 1. If INDEPENDENT SET can be solved in O∗((2 − ǫ)tw(G)) for some ǫ > 0 then SAT can be
solved in O∗((2 − δ)n) time for some δ > 0.
i
i
j
1, cp′
2 . . . cp′
i . . . p2m
c having c vertices each, and connect cpi with cp′
Construction. Given an instance φ to SAT we construct a graph G as follows. We assume that every
clause has an even number of variables, if not we can add a single variable to all odd size clauses
and force this variable to false. First we describe the construction of clause gadgets. For a clause
C = {ℓ1, ℓ2, . . . , ℓc} we make a gadget bC as follows. We take two paths, CP = cp1, cp2 . . . , cpc and
CP ′ = cp′
i for every i. For each literal ℓi
we make a vertex ℓi in bC and make it adjacent to cpi and cp′
i. Finally we add two vertices cstart and
cend, such that cstart is adjacent to cp1 and cend is adjacent to cpc. Observe that the size of the maximum
independent set of bC is c + 2. Also, since c is even, any independent set of size c + 2 in bC must contain
at least one vertex in C = {ℓ1, ℓ2, . . . , ℓc}. Finally, notice that for any i, there is an independent set of
size c + 2 in bC that contains ℓi and none of ℓj for j 6= i.
We first construct a graph G1. We make n paths P1, . . . , Pn, each path of length 2m. Let the vertices
of the path Pi be p1
. The path Pi corresponds to the variable vi. For every clause Ci of φ we
make a gadget bCi. Now, for every variable vi, if vi occurs positively in Cj, we add an edge between p2j
and the literal corresponding to vi in bCj. If vi occurs negatively in Cj, we add an edge between p2j−1
and the literal corresponding to vi in bCj. Now we construct the graph G as follows. We take n + 1 copies
of G1, call them G1, . . . Gn+1. For every i ≤ n we connect Gi and Gi+1 by connecting p2m
in Gi with
p1
j in Gi+1 for every j ≤ n. This concludes the construction of G.
Lemma 1. If φ is satisfiable, then G has an independent set of size (mn + Pi≤m Ci + 2)(n + 1).
Proof. Consider a satisfying assignment to φ. We construct an independent set I in G. For every variable
vi if vi is set to true, then pick all the vertices on odd positions from all copies of Pi, that is p1
i , p5
i
i , p6
and so on. If vi is false then pick all the vertices on even positions from all copies of Pi, that is p2
i
and so on. It is easy to see that this is an independent set of size mn(n + 1) containing vertices from all
the paths. We will now consider the gadget bCj corresponding to a clause Cj. We will only consider the
copy of bCj in G1 as the other copies can be dealt identically. Let use choose a true literal ℓa in Cj and
let vi be the corresponding variable. Consider the vertex ℓa in bCj. If vi occurs positively in Cj then vi is
true. Then I does not contain p2j
i , the only neighbour of ℓa outside of bCj. On the other hand if vi occurs
negatively in Cj then vi is false. In this case I does not contain p2j−1
, the only neighbour of ℓa outside
of bCj. There is an independent set of size Cj + 2 in bC that contains ℓa and none out of ℓb, b 6= a.
We add this independent set to I and proceed in this manner for every clause gadget. By the end of the
process (Pi≤m Ci + 2)(n + 1) vertices from clause gadgets are added to I, yielding that the size of I
is (mn + Pi≤m Ci + 2)(n + 1), concluding the proof.
Lemma 2. If G has an independent set of size (2mn + Pi≤m Ci + 2)(n + 1), then φ is satisfiable.
Proof. Consider an independent set of G of size (mn + Pi≤m Ci + 2)(n + 1). The set I can contain
at most m vertices from each copy of Pi for every i ≤ n and at most Cj + 2 vertices from each copy
of the gadget Cj. Since I must contain at least these many vertices from each path and clause gadget in
order to contain at least (mn + Pi≤m Ci + 2)(n + 1) vertices, it follows that I has exactly m vertices
in each copy of each path Pi and exactly Cj + 2 vertices in each copy of each clause gadget bCj. For a
fixed j, consider the n + 1 copies of the path Pj. Since Pj in Gi is attached to Pj in Gi+1 these n + 1
i , p3
i , p4
i
i
4
cstart
cp′
1
bCj
cp1
cpc
ℓ1
ℓc
cend
p2j−1
1
p2j
1
PSfrag replacements
P1
Pn
Figure 1: Reduction to INDEPENDENT SET: clause gadget bCj attached to the n paths representing the
variables.
p2j−1
n
p2j
n
copies of Pi together form a path P having 2m(n + 1) vertices. Since I ∩ P = m(n + 1) it follows
that I ∩ P must contain every second vertex of P , except possibly in one position where I ∩ P skips two
vertices of P . There are only n paths and n + 1 copies of G1, hence the pigeon-hole principle yields that
in some copy Gy of G1, I contains every second vertex on every path Pi. From now onwards we only
consider such a copy Gy.
In Gy, for every i ≤ n, I contains every second vertex of Pi. We make an assignment to the variables
of φ as follows.
If I contains all the odd numbered vertices of Pi then vi is set to true, otherwise
I contains all the even numbered vertices of Pi and vi is set to false. We argue that this assignment
satisfies φ. Indeed, consider any clause Cj, and look at the gadget bCj. We know that I contains Cj + 2
vertices from bCj and hence I must contain a vertex ℓa in corresponding to a literal of Cj. Suppose ℓa
is a literal of vi. Since I contains ℓa, if ℓa occurs positively in Cj, then I can not contain p2j
i and hence
vi is true. Similarly, if ℓa occurs negatively in Cj then I can not contain p2j−1
and hence vi is false. In
both cases vi satisfies Cj and hence all clauses of φ are satisfied by the assignment.
i
Lemma 3. pw(G) ≤ n + 4.
i
Proof. We give a mixed search strategy to clean G using n + 3 searchers. For every i we place a searcher
on the first vertex of Pi in G1. The n searchers slide along the paths P1, . . . Pn in m rounds. In round
j each searcher i starts on p2j−1
. Then, for every variable vi that occurs positively in Cj, the searcher i
slide forward to p2j
. Observe that at this point there is a searcher on every neighbour of the gadget bCj.
i
This gadget can now be cleaned with 3 additional searchers. After bCj is clean, the additional 3 searchers
are removed, and each of the n searchers on the paths P1, . . . Pn slide forward along these paths, such
that searcher i stands on p2(j+1)
. At that point, the next round commences. When the searchers have
cleaned G1 they slide onto the first vertex of P1 . . . Pn in G2. Then they proceed to clean G2, . . . , Gn+1
in the same way that G1 was cleaned. Now applying Proposition 1 we get that pw(G) ≤ n + 4.
i
5
PSfrag replacements
p3
1
p3
1
p1
1
p3
1
P1
g1
g′
1
p1
p
Pp
gp
p3
1
g′
p
p3
1
p3
p
xS
x′
S
x
Figure 2: Reduction to DOMINATING SET: group gadget bB. The set S is shown by the circled vertices.
The construction, together with Lemmata 1, 2 and 3 proves Theorem 1.
4 Dominating Set
A dominating set of a graph G is a set S ⊆ V (G) such that V (G) = N [S]. In the DOMINATING SET
problem we are given a graph G and the objective is to find a dominating set of minimum size.
Theorem 2. If DOMINATING SET can be solved in O∗((3 − ǫ)pw(G)) time for some ǫ > 0 then SAT can
be solved in O∗((2 − δ)n) time for some δ > 0.
i, both of which are neighbours to p1
i , p2
i . To each path Pi we attach two guards gi and g′
Construction. Given ǫ < 1 and an instance φ to SAT we construct a graph G as follows. We first
chose an integer p depending only on ǫ. Exactly how p is chosen will be discussed in the proof of
Theorem 2. We group the variables of φ into groups F1, F2, . . . , Ft, each of size at most β = ⌊log 3p⌋.
Hence t = ⌈n/β⌉. We now proceed to describe a "group gadget" bB, which is central in our construction.
To build the group gadget bB we make p paths P1, . . . , Pp, where the path Pi contains the vertices p1
i ,
p2
i and p3
i and
p3
i . When the gadgets are attached to each other, the guards will not have any neighbours outside of their
own gadget bB, and will ensure that at least one vertex out of p1
i are chosen in any minimum
size dominating set of G. Let P be a vertex set containing all the vertices on the paths P1, . . . , Pp. For
every subset S of P that picks exactly one vertex from each path Pi we make two vertices xS and x′
S,
where xS is adjacent to all vertices of P \ S (all those vertices that are on paths and not in S) and x′
S is
only adjacent to xS. We conclude the construction of bB by making all the vertices x′
S (for every set S)
adjacent to each other, that is making them into a clique, and adding a guard x adjacent to x′
S for every
set S. Essentially x′
S's together with x forms a clique and all the neighbors of x reside in this clique.
We construct the graph G as follows. For every group Fi of variables we make m(2pt + 1) copies
i for 1 ≤ j ≤ m(2pt + 1). For every fixed i ≤ t we connect the gadgets
in a path-like manner. In particular, for every j < m(2pt + 1) and every ℓ ≤ p
. Now we make two new
of the gadget bB, call them bBj
bB1
i , bB2
we make an edge between p3
i . . . , bBm(2pt+1)
i , p2
i and p3
i
ℓ in the gadget bBj
i with p1
ℓ in the gadget bBj+1
i
6
bcℓ
j
bBx
1
bBx
t
h
h′
PSfrag replacements
Figure 3: Reduction to DOMINATING SET: arranging the group gadgets. Note that x = mℓ + j, thus bcℓ
is attached to vertices in bBx
1 , . . . , bBx
t .
j
j in bBm(2pt+1)
i
j in bB1
i
in a path-like manner.
i , bB2
i . . . , bBm(2pt+1)
i for every i ≤ t, j ≤ p and to p3
vertices h and h′, with h adjacent to h′, p1
for every
i ≤ t, j ≤ p. That is, for all 1 ≤ i ≤ t, h is adjacent to first and last vertices of "long paths" obtained
after connecting the gadgets bB1
For every 1 ≤ i ≤ t, and to every assignment of the variables in the group Fi, we designate a subset
S of P in the gadget bB that picks exactly one vertex from each path Pj. Since there are at most 2β
different assignments to the variables in Fi, and there are 3p ≥ 2β such sets S, we can assign a unique
set to each assignment. Of course, the same set S can correspond to one assignment of the group F1 and
some another assignment of the group F2. Recall that the clauses of φ are C1, . . . , Cm. For every clause
Cj we make 2pt + 1 vertices bcℓ
j will be connected to the
gadgets bBmℓ+j
for every 1 ≤ i ≤ t. In particular, for every assignment of the variables in the group Fi
that satisfy the clause Cj, we consider the subset S of P that corresponds to the assignment. For every
0 ≤ ℓ < 2n + 1, we make x′
j. The best way to view this is that every clause Cj
has 2pt + 1 private gadgets, bBj
, in every group of gadgets corresponding to Fi's.
Now we have 2pt + 1 vertices corresponding to the clause Cj, one each for one gadget from each group
gadgets corresponding to Fi's. This concludes the construction of G.
j, one for each 0 ≤ ℓ < 2pt + 1. The vertex bcℓ
adjacent to bcℓ
, . . . , bBm2pt+j
S in bBmℓ+j
i , bBm+j
i
i
i
i
Lemma 4. If φ has a satisfying assignment, then G has a dominating set of size (p + 1)tm(2pt + 1) + 1.
Proof. Given a satisfying assignment to φ we construct a dominating set D of G that contains the vertex
h and exactly p + 1 vertices in each gadget bBj
i . For each group Fi of variables we consider the set S
which corresponds to the restriction of the assignment to the variables in Fi. From each gadget bBj
i we
add the set S to D and also the vertex x′
S to D. It remains to argue that D is indeed a dominating set.
Clearly the size is bounded by (p + 1)tm(2pt + 1) + 1, as the number of gadgets is tm(2pt + 1).
For a fixed i ≤ t and j consider the vertices on the path Pj in the gadgets bBℓ
i for every ℓ ≤
m(2pt + 1). Together these vertices form a path of length 3m(2pt + 1) and every third vertex of this
7
path is in S. Thus, all vertices on this path are dominated by other vertices on the path, except for the
first and last one. Both these vertices, however, are dominated by h.
Now, fix some i ≤ t and l ≤ m(2pt + 1) and consider the gadget bBℓ
i . Since D contains some vertex
on the path Pj, we have that for every j both gj and g′
j are dominated. Furthermore, for every set S∗ not
equal to S that picks exactly one vertex from each Pj, vertex xS∗ is dominated by some vertex on some
Pj-namely by all vertices in S \ S∗ 6= ∅. The last assertion follows since xS∗ is connected to all the
vertices on paths except S∗. On the other hand, xS is dominated by x′
S also dominates all the
other vertices x′
S∗ for S∗ 6= S and the guard x.
The only vertices not yet accounted for are the vertices bcℓ
j for every j ≤ m and ℓ < 2pt + 1. Fix a j
and a ℓ and consider the clause Cj. This clause contains a literal set to true, and this literal corresponds
to a variable in the group Fi for some i ≤ t. Of course, the assignment to Fi satisfies Cj. Let S be the
set corresponding to this assignment of Fi. By the construction of D, the dominating set contains x′
S in
bBmℓ+j
Lemma 5. If G has a dominating set of size (p + 1)tm(2pt + 1) + 1, then φ has a satisfying assignment.
j. This concludes the proof.
S is adjacent to bcℓ
S, and x′
and x′
i
Proof. Let D be a dominating set of G of size at most (p + 1)tm(2pt + 1) + 1. Since D must dominate
h′, hence without loss of generality we can assume that D contains h. Furthermore, inside every gadget
bBℓ
i , D must dominate all the guards, namely gj and g′
j for every j ≤ p, and also x. Thus D contains at
least p + 1 vertices from each gadget bBℓ
i which in turn implies that D contains exactly p + 1 vertices
from each gadget bBℓ
j for every j and in addition dominate x
with only p + 1 verticesis if D has one vertex from each Pj, j ≤ p and in addition contains some vertex
in N [x]. Let S be D ∩ P in bBℓ
i . Observe that xS is not dominated by D ∩ S. The only vertex in N [x]
that dominates xS is x′
i . The only way D can dominate gj and g′
S and hence D contains x′
S.
i and S′ be D ∩ P in bBℓ+1
, 1 ≤ r ≤ m. Consider a gadget bBℓ
. Observe that if S contains pa
Now we want to show that for every 1 ≤ i ≤ t there exists one 0 ≤ ℓ ≤ 2tp such that for fixed i,
D ∩ P is same in all the gadgets bBmℓ+r
. Let
S be D ∩ P in bBℓ
then
we must have b ≤ a. We call a consecutive pair bad if for some j ≤ p, D contains pa
j in
bBℓ+1
and b < a. Hence for a fixed i, we can at most have 2p consecutive bad pairs. Now we mark all
the bad pairs that occur among the gadgets corresponding to some Fi. This way we can mark only 2tp
bad pairs. Thus, by the pigeon hole principle, there exists an ℓ ∈ {0, . . . , 2tp} such that there are no bad
pairs in bBmℓ+r
i and its follower, bBℓ+1
j in bBℓ
j in bBℓ+1
We make an assignment φ by reading off D ∩ P in each gadget bBmℓ+1
. In particular, for every
group Fi, we consider S = D ∩ P in the gadget bBmℓ+1
. This set S corresponds to an assignment of Fi,
and this is the assignment of Fi that we use. It remains to argue that every clause Cr is satisfied by this
assignment.
for all 1 ≤ i ≤ t and 1 ≤ r ≤ m.
j in bBℓ
i and pb
i and pb
i
i
i
i
i
i
i
i
ℓ. We know that it is dominated by some x′
Consider the vertex bcr
. The set S
corresponds to an assignment of Fi that satisfies the clause Cr. Because D ∩ P remains unchanged
in all gadgets from bBmℓ+1
, this is exactly the assignment φ restricted to the group Fi. This
concludes the proof.
S in a gadget bBmℓ+r
to bBmℓ+r
i
i
i
j and p3
Lemma 6. pw(G) ≤ tp + O(3p)
Proof. We give a mixed search strategy to clean the graph with tp + O(3p) searchers. For a gadget bB
we call the vertices p1
j , 1 ≤ j ≤ p, as entry vertices and exit vertices respectively. We search the
graph in m(2tp + 1) rounds. In the beginning of round ℓ there are searchers on the entry vertices of the
gadgets bBℓ
i for every i ≤ t. Let 1 ≤ a ≤ m and 0 ≤ b < 2tp + 1 be integers such that ℓ = a + mb. We
place a searcher on bcb
a. Then, for each i between 1 and p in turn we first put searchers on all vertices of
bBℓ
i and then remove all the searchers from bBℓ
i except for the ones standing on the exit vertices. After all
gadgets bBℓ
a. To commence
t have been cleaned in this manner, we can remove the searcher from bcb
1 . . . bBℓ
8
the next round, the searchers slide from the exit positions of bBℓ
for every
i. In total, at most tp + V ( bB) + 1 ≤ tp + O(3p) searchers are used simultaneously. This together with
Proposition 1 give the desired upperbound on the pathwidth.
i to the entry positions of bBℓ+1
i
Proof (of Theorem 2). Suppose DOMINATING SET can be solved in O∗((3 − ǫ)pw(G))= O∗(3λpw(G))
log 3 for some δ′ < 1.
time, where λ = log3(3−ǫ) < 1. We choose p large enough such that λ·
Given an instance of SAT we construct an instance of DOMINATING SET using the above construction
and the chosen value of p. Then we solve the DOMINATING SET instance using the O∗(3λpw(G)) time
algorithm. Correctness is ensured by Lemmata 4 and 5. Lemma 6 yields that the total time taken is upper
bounded by O∗(3λpw(G)) ≤ O∗(3λ(tp+f (λ))) ≤ O∗(3λ
log 3 ) ≤ O∗(2δ′′n) =O∗((2 −
δ)n), for some δ′′, δ < 1. This concludes the proof.
⌊p log 3⌋ ) ≤ O∗(3δ′ n
np
p
⌊p log 3⌋ = δ′
5 Max Cut
A cut in a graph G is a partition of V (G) into V0 and V1. The cut-set of the cut is the set of edges whose
one end point is in V0 and the other in V1. We say that an edge is crossing this cut if it has one endpoint in
V0 and one in V1, that is, the edge is in the cut-set. The size of the cut is the number of edges in G which
are crossing this cut. If the edges of G have positive integer weights then the weight of the cut is the
sum of the weights of edges which are crossing the cut. In the MAX CUT problem we are given a graph
G together with an integer t and asked whether there is a cut of G of size at least t. In the WEIGHTED
MAX CUT problem every edge has a positive integer weight and the objective is to find a cut of weight
at least t.
Theorem 3. If MAX CUT can be solved in O∗((2 − ǫ)pw(G)) for some ǫ > 0 then SAT can be solved in
O∗((2 − δ)n) time for some δ > 0.
Construction. Given an instance φ of SAT we first construct an instance Gw of WEIGHTED MAX
CUT as follows. We later explain how to obtain an instance of unweighted MAX CUT from here.
We start with making a vertex x0. Without loss of generality, we will assume that x0 ∈ V0 in every
solution. We make a vertex bvi for each variable vi. For every clause Cj we make a gadget as follows.
We make a path bPj having 4Cj vertices. All the edges on bPj have weight 3n. Now, we make the
first and last vertex of bPj adjacent to x0 with an edge of weight 3n. Thus the path bPj plus the edges
from the first and last vertex of bPj to x0 form an odd cycle bCj. We will say that the first, third, fifth,
etc, vertices are on odd positions on bPj while the remaining vertices are on even positions. For every
variable vi that appears positively in Cj we select a vertex p at an even position (but not the last vertex)
on bPj and make bv adjacent to p and p's successor on bPj with edges of weight 1. For every variable vi
that appears negatively in Cj we select a vertex p at an odd position on bPj and make bv adjacent to p
and p's successor on bPj with edges of weight 1. We make sure that each vertex on bPj receives an edge
at most once in this process. There are more than enough vertices on bPj to accommodate all the edges
incident to vertices corresponding to variables in the clause Cj. We create such a gadget for each clause
and set t = 1 + (12n + 1)Pm
Lemma 7. If φ is satisfiable, then Gw has a cut of weight at least t.
Proof. Suppose φ is satisfiable. We put x0 in V0 and for every variable vi we put bvi in V1 if vi is true and
bvi in V0 if vi is false. For every clause Cj we proceed as follows. Let us choose a true literal of Cj and
suppose that this literal corresponds to a vertex pj on bPj. We put the first vertex on bPj in V1, the second
in V0 and then we proceed along bPj putting every second vertex into V1 and V0 until we reach pj. The
j of pj on bPj is put into the same set as pj. Then we continue along cPj putting every second
successor p′
vertex in V1 and V0. Notice that even though Cj may contain more than one literal that is set to true, we
j=1 Cj. This concludes the construction.
9
For every clause Cj all edges on the path bPj except for pjp′
only select one vertex pj from the path bPj and put pj and its successor on the same side of the partition.
It remains to argue that this cut has weight at least t.
j are crossing, and the two edges to x0
from the first and last vertex of bPj are crossing as well. These edges contribute 12nCj to the weight
of the cut. We know that pj corresponds to a literal that is set to true, and this literal corresponds to a
variable vi. If vi occurs positively in Cj then vi ∈ V1 and pj is on an even position of bPj. Thus both pj
and his successor p′
j are crossing, contributing 2 to the weight of
the cut. For each of the remaining variables vi′ appearing in Cj, one of the two neighbours of bvi′ on bPj
appear in V0 and one in V1, so exactly one edge from vi′ to bPj is crossing. Thus the total weight of the
cut is t = Pm
Lemma 8. If Gw has a cut of weight at least t, then φ is satisfiable.
j are in V0 and hence both vipj and vip′
j=1 12nCj + Cj + 1 = m + (12n + 1)Pm
j=1 Cj. This completes the proof.
Proof. Let (V0, V1) be a cut of G of maximum weight, hence the weight of this cut is at least t. Without
loss of generality, let x0 ∈ V0. For every clause Cj at least one edge of the odd cycle bCj is not crossing.
If more than one edge of this cycle is not crossing, then the total weight of the cut edges incident to the
path bPj is at most 3n(4Cj − 1) + 2n < 12Cj. In this case we could change the partition (V0, V1) such
that all edges of bPj are crossing and the first vertex of bPj is in V1. Using the new partition the weight of
the crossing edges in the cycle bCj is at least 12Cj and the edges not incident to bPj are unaffected by
the changes. This contradicts that (V0, V1) was a maximum weight cut. Thus it follows that exactly one
edge of bCj is not crossing.
Given the cut (V0, V1) we set each variable vi to true if bvi ∈ V1 and vi to false otherwise. Consider a
clause Cj and a variable vi that appears in Cj. Let uv be the edge of bC ′
j that is not crossing. If there is a
variable bvi adjacent to both u and v, then it is possible that both bviu and bviv are crossing. For every other
variable vi′ in Cj, at most one of the edges from bvi′ to bPj is crossing. Thus, the weight of the edges that
are crossing in the gadget bCj is at most (12n + 1)Cj + 1. Hence, to find a cut-set of weight at least t in
G, we need to have crossing edges in bCj with sum of their weights exactly equal to 12nCj + Cj + 1.
It follows that there is a vertex bvi adjacent to both u and v such that both bviu and bviv are crossing.
If vi occurs in Cj positively then u is on an even position and hence, u ∈ V0. Since bviu is crossing
it follows that vi is true and Cj is satisfied. On the other hand, if vi occurs in Cj negated then u is on an
odd position and hence, u ∈ V1. Since bviu is crossing it follows that vi is false and Cj is satisfied. As
this holds for each clause individually, this concludes the proof.
For every edge e ∈ E(Gw), let we be the weight of e in Gw. We construct an unweighted graph G
from Gw by replacing every edge e = uv by we paths from u to v on three edges. Let W be the sum of
the edge weights of all edges in Gw.
Lemma 9. G has a cut of size 2W + t if and only if Gw has a cut of weight at least t.
Proof. Given a partition of V (Gw) we partition V (G) as follows. The vertices of G that also are vertices
of V (G) are partitioned in the same way as in V (Gw). On each path of length 3, if the endpoints of the
path are in different sets we can partition the middle vertices of the path such that all edges are cut. If
the endpoints are in the same set we can only partition the middle vertices such that 2 out of the 3 edges
are cut. The reverse direction is similar.
Lemma 10. pw(G) ≤ n + 5.
Proof. We give a search strategy to clean G with n + 5 searchers. We place one searcher on each vertex
bvi and one searcher on x0. Then one can search the gadgets cHj one by one. In Gw it is sufficient to use 2
searchers for each cHj, whereas in G after the edges have been replaced by multiple paths on three edges,
we need 4 searchers. This combined with Proposition 1 gives the desired upper bound on the pathwidth
of the graph.
10
The construction, together with Lemmata 7, 8, 9 and 10 proves Theorem 3.
6 Graph Coloring
A q-coloring of G is a function µ : V (G) → [q]. A q-coloring µ of G is proper if for every edge
uv ∈ E(G) we have µ(u) 6= µ(v). In the q-COLORING problem we are given as input a graph G
and the objective is to decide whether G has a proper q-coloring. In the LIST COLORING problem,
every vertex v is given a list L(v) ⊆ [q] of admissible colors. A proper list coloring of G is a function
µ : V (G) → [q] such that µ is a proper coloring of G that satisfies µ(v) ∈ L(v) for every v ∈ V (G). In
the q-LIST COLORING problem we are given a graph G together with a list L(v) ⊆ [q] for every vertex
v. The task is to determine whether there exists a proper list coloring of G.
A feedback vertex set of a graph G is a set S ⊆ V (G) such that G \ S is a forest; we denote by
fvs(G) the size of the smallest such set. It is well-known that tw(G) ≤ fvs(G) + 1. Unlike in the other
sections, where we give lower bounds for algorithms parameterized by pw(G), the following theorem
gives also a lower bound for algorithms parameterized by fvs(G). Such a lower bound follows very
naturally from the construction we are doing here, but not from the constructions in the other sections.
It would be interesting to explore whether it is possible to prove tight bounds parameterized by fvs(G)
for the problems considered in the other sections.
Theorem 4. If q-COLORING can be solved in O∗((q − ǫ)fvs(G)) or O∗((3 − ǫ)pw(G)) time for some
ǫ > 0, then SAT can be solved in O∗((2 − δ)n) time for some δ > 0.
Construction. We will show the result for LIST COLORING first, and then give a simple reduction that
demonstrates that q-COLORING can be solved in O∗((q −ǫ)fvs(G)) time if and only if q-LIST COLORING
can.
Depending on ǫ and q we choose a parameter p. Now, given an instance φ to SAT we will construct
a graph G with a list L(v) for every v, such that G has a proper list-coloring if and only if φ is satisfiable.
Throughout the construction we will call color 1-red, color 2-white and color 3-black.
We start by grouping the variables of φ into t groups F1, . . . , Ft of size ⌊log qp⌋. Thus t = ⌈
⌊log qp⌋ ⌉.
We will call an assignment of truth values to the variables in a group Fi a group assignment. We will
say that a group assignment satisfies a clause Cj of φ if Cj contains at least one literal which is set to
true by the group assignment. Notice that Cj can be satisfied by a group assignment of a group Fi, even
though Cj also contains variables that are not in Fi.
For each group Fi, we make a set Vi of p vertices v1
i . The vertices in Vi get full lists,
that is, they can be colored by any color in [q]. The coloring of the vertices in Vi will encode the group
assignment of Fi. There are qp ≥ 2Fi possible colorings of Vi. Thus, to each possible group assignment
of Fi we attach a unique coloring of Vi. Notice that some colorings of Vi may not correspond to any
group assignments of Fi.
i , . . . , vp
n
For each clause Cj of φ, we make a gadget bCj. The main part of bCj is a long path bPj that has
one vertex for each group assignment that satisfies bCj. Notice that there are at most tqp possible group
assignments, and that q and p are constants independent of the input φ. The list of every vertex on bPj
to the start and end of bPj respectively,
is {red, white, black}. We attach two vertices pstart
and the two vertices are not counted as vertices of the path bPj itself. The list of pstart
is {white}. If
V ( bPj ) is even, then the list of pend
is {black}.
The intention is that to properly color bPj one needs to use the color red at least once, and that once is
sufficient. The position of the red colored vertex on the path bPj encodes how the clause Cj is satisfied.
For every vertex v on bPj we proceed as follows. The vertex v corresponds to a group assignment to
Fi that satisfies the clause Cj. This assignment in turn corresponds to a coloring of the vertices of Vi.
Let this coloring be µi. We build a connector whose role is to enforce that v can be red only if coloring
is {white}, whereas if V ( bPj ) is odd then the list of pend
and pend
j
j
j
j
j
11
w3
w4
{red, 3}
{red, 2, 3, 4}
w
v
{red, 2}
{red, 2}
vl
i
vl
i
w2
w3
w4
w2
w3
w4
{red, 3}
{red, 4}
{red, 2}
w
{red, 2, 3, 4}
{red, 2}
{red, 2}
{red, 3}
w′
2
w′
3
w′
4
{red, 2}
{red, 4}
w
{red, 2, 3, 4}
v
v
Figure 4: Reduction to q-COLORING: the way the connector connects a vertex vl
"bad color" x ∈ [q] \ {µi(vl
i with v for a particular
i)}. The left side shows the case x = red = 1, the right side x = 2 (q = 4).
µi appears on Vi. To build the connector, for each vertex vl
following.
i ∈ Vi and color x ∈ [q] \ {µi(vl
i)} we do the
• If x is red, then we add one vertex wy for every color y except for red. We make wy adjacent to vl
i
and the list of wy is {red, y}. Then we add a vertex w which is adjacent to all vertices wy and v,
and whose list is all of [q].
• If x is not red, we add two vertices wy and w′
adjacent to vl
Finally we add a vertex w adjacent to w′
y adjacent to wy. The list of wy is {x, red} while the list of w′
y for all y and to v. The list of w is all of [q].
i and w′
y for each color y except for red. We make wy
y is {y, red}.
Notice that in the above construction we have reused the names w, wy and w′
y for many different vertices:
in each connector, there is a separate vertex w for each vertex vl
i)}. Build-
ing a connector for each vertex v on bPj concludes the construction of the clause gadget bCj, and creating
one such gadget for each clause concludes the construction of G. The following lemma summarizes the
most important properties of the connector:
Lemma 11. Consider the connector corresponding a vertex v on bPj and coloring µi of Vi.
i ∈ Vi and color x ∈ [q]\{µi(vl
1. Any coloring on Vi and any color c ∈ {white, black} on v can be extended to the rest of the
connector.
2. Coloring µi on Vi and any color c ∈ {red, white, black} on v can be extended to the rest of the
connector.
3. In any coloring of the connector, if v is red, then µi appears on Vi.
Proof. 1. For each vertex vl
i ∈ Vi and color x ∈ [q] \ {µi(vl
i)} we do the following.
i, and a vertex w with list [q] adjacent to wy for every y 6= red. If vl
• If x is red then in the construction of bCj we added a vertex wy with list {y, red} for every color
y 6= red adjacent to vl
i is
colored red, then we color each vertex wy with y and w with red. Notice that w is adjacent to v,
but v is colored either white or black, so it is safe to color w red. If, on the other hand, vl
i is not
colored red, we can color wy red for every y. Then all the neighbours of w have been colored with
red, except for v which has been colored white or black. Thus it is safe to color w with the color
out of black and white which was not used to color v.
12
• If x is not red, then in the construction of bCj we added two vertices wy and w′
y for each color y
except for red, and also added a vertex w. The vertices wy are adjacent to vl
i and for every y 6= red
y is adjacent to wy. Finally w is adjacent to al the vertices w′
the vertex w′
y and to v. For every y
the list of wy is {x, red} while the list of w′
i is colored with
x, then we let wy take color red and w′
y take color y for every y 6= red. We color w with red. In
the case that vl
y be
colored red for every y 6= red. Finally, all the neighours of w except for v have been colored red,
while v is colored with either black or white. According to the color of v we can either color w
black or white.
i is colored with a color different from x, we let wy be colored with x and w′
y is {y, red}. The list of w is [q]. If vl
2. We can assume that v is red, otherwise we are done by the previous statement. For each vertex
vl
i ∈ Vi and color x ∈ [q] \ {µi(vl
i)} we do the following.
• If x is red then in the construction of bCj we added a vertex wy with list {y, red} for every color
y 6= red adjacent to vl
i′ is
not colored red by µi, we can color wy red for every y. Then all the neighbours of w including v
have been colored with red and it is safe to color w with white.
i, and a vertex w with list [q] adjacent to wy for every y 6= red. Since vl
• If x is not red, then in the construction of bCj we added two vertices wy and w′
except for red, and also added a vertex w. The vertices wy are adjacent to vl
the vertex w′
y is adjacent to wy. Finally w is adjacent to all the vertices w′
the list of wy is {x, red} while the list of w′
with a color different from x we let wy be colored with x and w′
Finally, all the neighours of w including v have been colored red so it is safe to color w white.
y for each color y
i and for every y 6= red
y and to v. For every y
y is {y, red}. The list of w is [q]. Since µi colors vl
i
y be colored red for every y 6= red.
3. Suppose for contradiction that v is red, but some vertex vl
i ∈ Vi has been colored with a color
x 6= µi(vl
i). There are two cases. If x is red, then in the construction we added vertices wy adjacent to
vl
i for every color y 6= red. Also we added a vertex w adjacent to v and to wy for each y 6= red. The
list of wy is {red, y} and hence wy must have been colored y for every y 6= red. But then w is adjacent
to v which is colored red, and to wy which is colored y for every y 6= red. Thus vertex w has all colors
in its neighborhood, a contradiction. In the case when x is not red, then in the construction we added
two vertices wy and w′
i and had {x, red} as its list. Since
vl
i is colored x, all the wy vertices must be colored red. For every y 6= red, we have that w′
y is adjacent
to wy and has {red, y} as its list. Hence for every y 6= red the vertex w′
y is colored with y. But, in the
construction we also added a vertex w adjacent to v and to w′
y for each y 6= red. Thus again, vertex w
has all colors in its neighbourhood, a contradiction.
y for each y 6= red. Each wy was adjacent to vl
Lemma 12.
If φ is satisfiable, then G has a proper list-coloring.
Proof. Starting from a satisfying assignment of φ we construct a coloring γ of G. The assignment to φ
corresponds to a group assignment to each group Fi. Each group assignment corresponds to a coloring of
Vi. For every i, we let γ color the vertices of Vi using the coloring corresponding to the group assignment
of Fi.
Now we show how to complete this coloring to a proper coloring of G. Since the gadgets bCj are
pairwise disjoint, and there are no edges going between them, it is sufficient to show that we can complete
the coloring for every gadget bCj. Consider the clause Cj. The clause contains a literal that is set to true,
and this literal belongs to a variable in the group Fi. The group assignment of Fi satisfies the clause
Cj. Thus, there is a vertex v on bPj that corresponds to this assignment. We set γ(v) as red (that is, γ
colors v red), pstart
is colored with its only admissible color, namely black
if V ( bPj) is even and white if V ( bPj) is odd. The remaining vertices of bPj are colored alternatingly
white or black. By Lemma 11(2), the coloring can be extended to every vertex of the connector between
is colored white and pend
j
j
13
pstart
j
bPj
pend
j
PSfrag replacements
v1
1
vp
1
V1
v1
t
vp
t
Vt
Figure 5: Reduction to q-COLORING. The t groups of vertices V1, . . . , Vt represent the t groups of
variables F1, . . . , Ft (each of size ⌈log qp⌉). Each vertex of the clause path bPj is connected to one group
Vi via a connector.
Vi and v: the coloring appearing on Vi is the coloring µi corresponding to the group assignment Fi. For
every other vertex u on bPj, the color of u is black or white, thus Lemma 11(1) ensures that the coloring
can be extended to any connector on u.
As this procedure can be repeated to color the gadget bCj for every clause Cj, we can complete γ to
a proper list-coloring of G.
Lemma 13.
If G has a proper list-coloring γ, then φ is satisfiable.
Proof. Given γ we construct an assignment to the variables of φ as follows. For every group Fi of
variables, if γ colors Vi with a coloring that corresponds to a group assignment of Fi then we set this
assignment for the variables in Fi. Otherwise we set all the variables in Fi to false. We need to argue
that this assignment satisfies all the clauses of φ.
Consider a clause Cj and the corresponding gadget bCj. By a simple parity argument, bPj can not be
colored using only the colors black and white. Thus, some vertex v on bPj is colored red. The vertex v
corresponds to a group assignment of some group Fi that satisfies bCj. As v is red, Lemma 11(3) implies
that Vi is colored with the coloring µi that corresponds to this assignment. The construction then implies
that our chosen assignment satisfies Cj. As this is true for every clause, this concludes the proof.
Observation 1. The vertices Si≤t Vi form a feedback vertex set of G. Furthermore, pw(G) ≤ pt + 4
Proof. Observe that after removing Si≤t Vi, all that is left are the gadgets bCj which do not have any
edges between each other. Each such gadget is a tree and hence Si≤t Vi form a feedback vertex set of
G. If we place a searcher on each vertex of Si≤t Vi it is easy to see that each gadget bCj can be searched
with 4 searchers. The pathwidth bound on G follows using Proposition 1.
Lemma 14. If q-LIST COLORING can be solved in O∗((q − ǫ)fvs(G)) time for some ǫ < 1, then SAT
can be solved in O∗((2 − δ)n) time for some δ < 1.
Proof. Let O∗((q − ǫ)fvs(G))= O∗(qλfvs(G)) time, where λ = logq(q − ǫ) < 1. We choose a sufficiently
large p such that δ′ = λ p
p−1 < 1. Given an instance φ of SAT we construct a graph G using the
construction above, and run the assumed q-LIST COLORING. Correctness follows from Lemmata 12
and 13. By Observation 1 the graph G has a feedback vertex set of size p⌈
⌊p log q⌋ ⌉. The choice of p
implies that
n
+ p ≤ δ′ n
log q
+ p ≤ δ′′n,
n
⌉ ≤ λp
λp⌈
⌊p log q⌋
n
(p − 1) log q
14
for some δ′′ < 1. Hence SAT can be solved in time O∗(2δ′′ n) =O∗((2 − δ)n), for some δ > 0.
Finally, observe that we can reduce q-LIST-COLORING to q-COLORING by adding a clique Q =
{q1, . . . , qc} on q vertices to G and making qi adjacent to v when i /∈ L(v). Any coloring of Q must
use q different colors, and without loss of generality qi is colored with color i. Then one can complete
the coloring if and only if one can properly color G using a color from L(v) for each v. We can add the
clique Q to the feedback vertex set-this increases the size of the minimum feedback vertex set by q.
Since q is a constant independent of the input, this yields Theorem 4.
7 Odd Cycle Transversal
An equivalent formulation of MAX CUT is to delete the minimum number of edges to make the graph
bipartite. We can also consider the vertex deletion version of the problem. An odd cycle transversal of a
graph G is a subset S ⊆ V (G) such that G \ S is bipartite. In the ODD CYCLE TRANSVERSAL problem
we are given a graph G together with an integer k and asked whether G has an odd cycle transversal of
size k.
Theorem 5. If ODD CYCLE TRANSVERSAL can be solved in O∗((3 − ǫ)pw(G)) time for ǫ > 0, then
SAT can be solved in O∗((2 − δ)n) time for some δ > 0.
Construction. Given ǫ > 0 and an instance φ of SAT we construct a graph G as follows. We chose an
integer p based just on ǫ. Exactly how p is chosen will be discussed at the end of this section. We start
by grouping the variables of φ into t groups F1, . . . , Ft of size at most ⌊log 3p⌋. Thus t = ⌈
⌊log 3p⌋ ⌉. We
will call an assignment of truth values to the variables in a group Fi a group assignment. We will say
that a group assignment satisfies a clause Cj of φ if Cj contains at least one literal which is set to true by
the group assignment. Notice that Cj can be satisfied by a group assignment of a group Fi, even though
Cj also contains variables that are not in Fi.
n
Now we describe an auxiliary gadget which will be very useful in our construction. For two vertices
u and v by adding an arrow from u to v we will mean adding a path ua1a2a3v on four edges starting
in u and ending in v. Furthermore, we add four vertices b1, b2, b3 and b4 and edges ub1, b1a1, a1b2,
b2a2, a2b3, b3a3, a3b4, b4v, and b4v. Denote the resulting graph A(u, v). None of the vertices in A(u, v)
except for u and v will receive any further neighbours throughout the construction of G. The graph
A(u, v) has the following properties, which are useful for our construction.
• The unique smallest odd cycle transversal of A(u, v) is {a1, a3}. We call this the passive odd
cycle transversal of the arrow.
• In A(u, v) \ {a1, a3}, u and v are in different connected components.
• The set {a2, v} is a smallest odd cycle transversal of A(u, v) \ {u}. We call this the active odd
cycle transversal of the arrow.
The intuition behind an arrow from u to v is that if u is put into the odd cycle transversal, then v can
be put into the odd cycle transversal "for free." When the active odd cycle transversal of the arrow is
picked, we say the arrow is active, otherwise we say the arrow is passive.
To construct G we make t · p paths, {Pi,j} for 1 ≤ i ≤ t, 1 ≤ j ≤ p. Each path has 3m(tp + 1)
vertices, and the vertices of Pi,j are denoted by pℓ
i,j for 1 ≤ ℓ ≤ 3m(tp + 1). For a fixed i, the paths
{Pi,j : 1 ≤ j ≤ p} correspond to the set Fi of variables. For every 1 ≤ i ≤ t, 1 ≤ j ≤ p and
1 ≤ ℓ < 3m(tp + 1) we add three vertices aℓ
i,j adjacent to each other. We also add the
edges aℓ
i,j and qℓ
i,j, bℓ
One can think of the vertices of the paths {Pi,j} layed out as rows in a matrix, where for every fixed
i,j : 1 ≤ i ≤ t, 1 ≤ j ≤ p}. We group the colums three by three.
1 ≤ ℓ ≤ 3m(tp+1) there is a column {pℓ
i,jpℓ
i,j and bℓ
i,jpℓ+1
i,j .
15
In particular, For every i ≤ t and 0 ≤ ℓ < m(tp + 1) we define the sets P ℓ
, b3ℓ+2
1 ≤ j ≤ p}, Aℓ
Qℓ
, a3ℓ+2
, a3ℓ+3
: 1 ≤ j ≤ p}.
i = {a3ℓ+1
i,j
, q3ℓ+3
: 1 ≤ j ≤ p}, Bℓ
i = {b3ℓ+1
, q3ℓ+2
i,j
i,j
i,j
i,j
i = {q3ℓ+1
i,j
i,j
i,j
i = {p3ℓ+1
, b3ℓ+3
:
i,j
: 1 ≤ j ≤ p} and
, p3ℓ+2
, p3ℓ+3
i,j
i,j
i,j
For every i ≤ t and 0 ≤ ℓ < m(tp + 1) we make two new sets Lℓ
i and Rℓ
i are independent sets of size 5p, and we add all the edges possible between Lℓ
i . We make all the vertices in Aℓ
i and from Rℓ
i, and we make all vertices in Bℓ
i adjacent to all vertices of Rℓ
i we pick brℓ
i and Rℓ
Lℓ
i we pick a special vertex blℓ
Lℓ
vertices of Lℓ
rℓ+1
i
, except for ℓ = m(tp + 1) − 1.
We will say that a subset S of P ℓ
good. The idea is that there are 3p ≥ 2h good subsets of P ℓ
correspond to good subsets of P ℓ
has length 2p + 1. We select a vertex on X ℓ
arrow from u to a vertex of X ℓ
of exactly one arrow.
i which picks exactly one vertex from Pi,j for every 1 ≤ j ≤ p is
i , so we can make group assignments of Fi
i,S. The cycle X ℓ
i,S
i \ S we add an
i,S is the endpoint
i,S. For every vertex u ∈ P ℓ
i,S. We add arrows in such a way that every vertex of X ℓ
i . For every good subset S of P ℓ
i we add a cycle X ℓ
i,S and call it xℓ
i of new vertices. Both
i and Rℓ
i. From
i adjacent to all
i adjacent to
i. We make lℓ
For every i ≤ t and 0 ≤ ℓ < m(tp + 1) we make a cycle Y ℓ
i corresponds to a good subset S of P ℓ
i of length 3p, notice that the length of
i . For each good subset S of
i,S of the cycle X ℓ
i,S to the vertex in Y ℓ
i which corresponds to S.
the cycle is odd. Every vertex of Y ℓ
P ℓ
i we add an arrow from xℓ
We say that a good subset of P ℓ
i
if for every 1 ≤ j ≤ t,
the distance along Pi,j between the vertex of S on Pi,j and the vertex of S′ on Pi,j is divisible by 3.
Informally, S and S′ are equal if they look identical when we superimpose P ℓ
i . To every group
assignment of variables Fi we designate a good subset of P ℓ
i for every ℓ. We designate good subsets in
such a way that good subsets corresponding to the same group assignment are equal.
is equal with a good subset S′ of P ℓ′
i
i onto P ℓ′
Finally, for every clause Ch, 1 ≤ h ≤ m, we will add tp + 1 cycles. That is, for every 0 ≤ r ≤ tp
we add a cycle bC r
j . The cycle contains one vertex for every i ≤ t and group assignment to Fi, and
potentially one dummy vertex to make it have odd length. Going around the cycle counterclockwise
we first encounter all the vertices corresponding to group assignments of F1, then all the vertices cor-
responding to group assignments of F2, and so on. For i ≤ t and every good subset S of P rm+j
that
corresponds to a group assignment of Fi that satisfies Cj we add an arrow from xrm+j
to the vertex on
bC r
j that corresponds to the same group assignment of Fi as S does. This concludes the construction of
G.
i,S
i
The intention behind the construction is that if φ is satisfiable, then a minimum odd cycle transversal
of G can pick:
• One vertex from each triangle {aℓ
i,j, bℓ
i,j, qℓ
i,j} for each 1 ≤ i ≤ t, 1 ≤ j ≤ p, 1 ≤ ℓ < 3m(tp + 1).
There are tp(3m(tp + 1) − 1) such triangles in total.
• One vertex from {p3ℓ+1
, p3ℓ+2
i,j
are tpm(tp + 1) such triples.
i,j
, p3ℓ+3
i,j } for each 1 ≤ i ≤ t, 1 ≤ j ≤ p, 0 ≤ ℓ < m(tp + 1). There
• Two vertices from every arrow added, without counting the starting point of the arrow. For each
i ≤ t and 0 ≤ ℓ < m(tp + 1) there are 2p3p arrows ending in some cycle X ℓ
i,S. Hence there are
2p3ptm(tp + 1) such arrows. For every i ≤ t and 0 ≤ ℓ < m(tp + 1) there are 3p arrows ending in
the cycle Y ℓ
i . Hence there are 3ptm(tp + 1) such arrows. For every clause Cj there are m arrows
added for every group assignment that satisfies that clause. Let µ be the sum over all clauses of
the number of group assignments that satisfy that clause. The total number of arrows added is then
mµ+(2p+1)3ptm(tp+1). Thus the odd cycle transversal can pick 2mµ+2(2p+1)3ptm(tp+1)
vertices from arrows.
• One vertex xℓ
i,S for every i ≤ t and 0 ≤ ℓ < m(tp + 1). There are tm(tp + 1) choices for i and ℓ.
We let the α be the value of the total budget, that is the sum of the items above.
Lemma 15. If φ is satisfiable, then G has an odd cycle transversal of size α.
16
Proof. Given a satisfying assignment γ to φ we construct an odd cycle transversal Z of G of size α
together with a partition of V (G) \ Z into L and R such that every edge of G \ Z goes between a vertex
in L and a vertex in R. The assignment to φ corresponds to a group assignment of each Fi for 1 ≤ i ≤ t.
For every 1 ≤ i ≤ t and 0 ≤ ℓ < m(tp + 1) we add to Z the good subset S of P ℓ
i that corresponds to the
group assignment of Fi. Notice that for each fixed i, the sets picked from P ℓ
i are equal for any
ℓ, ℓ′. At this point we have picked one vertex from {p3ℓ+1
i,j } for each 1 ≤ i ≤ t, 1 ≤ j ≤ p,
0 ≤ ℓ < m(tp + 1).
i and P ℓ′
, p3ℓ+3
, p3ℓ+2
i,j
i,j
For every fixed 1 ≤ i ≤ t, 1 ≤ j ≤ p there are three cases. If p1
i,j into L. If p3
i,j ∈ Z we put p1
i,j /∈ Z we put pℓ
i,j ∈ Z we put p2
i,j into L and p3
i,j
i,j into R.
i,j into the same set out of {L, R} as
i,j into L and p2
i,j ∈ Z we put p1
into R. If p2
Now, for every 4 ≤ ℓ ≤ 3m(tp + 1) such that pℓ
pℓ′
i,j where 1 ≤ ℓ′ ≤ 3 and ℓ ≡ ℓ′ mod 3.
i,j into R and p3
For every 1 ≤ i ≤ t, 0 ≤ ℓ ≤ m(tp + 1) we put Lℓ
i, b ∈ Bℓ
of pairwise adjacent vertices such that a ∈ Aℓ
a has a neighbour a′ in P ℓ
on Pi,j. Thus, there are three cases;
i and b has a neighbour b′ in P ℓ
i into L and Rℓ
i , and q ∈ Qℓ
i into R. For every triple of a, b, q
i, we proceed as follows. The vertex
i . There is a j such that b′ is the successor of a′
• a′ ∈ Z and b′ ∈ L, we put a in R, q in L and b in Z.
• a′ ∈ R and b′ ∈ Z, we put a in Z, q in R and b in L.
• a′ ∈ L and b′ ∈ R, we put a in R, q in Z and b in L.
i,S for good subsets S of P ℓ
For every 1 ≤ i ≤ t, 0 ≤ ℓ ≤ m(tp + 1) there are many arrows from vertices in P ℓ
i to vertices on cycles
X ℓ
i is in Z we add the active odd cycle
transversal of the arrow to Z, otherwise we add the passive odd cycle transversal of the arrow to Z. In
either case the remaining vertices on the arrow form a forest, and therefore we can insert the remaining
vertices of the arrow into L and R according to which sets out of {L, R, Z} u and v are in.
i . For each arrow, if its endpoint in P ℓ
i,S we put xℓ
For every 1 ≤ i ≤ t, 0 ≤ ℓ ≤ m(tp + 1) there is exactly one set S such that the cycle X ℓ
i,S only
has passive arrows pointing into it. This is exactly the set S which corresponds to the restriction of γ to
Fi. Each cycle X ℓ
i,S′ that has at least one arrow pointing into them already contain at least one vertex in
Z-the endpoint of the active arrow pointing into the cycle. Thus we can partition the remaining vertices
of X ℓ
i,S′ into L and R such that no edge has both endpoints in L or both endpoints in R. For the cycle
X ℓ
i,S into L and R such that no edge has
both endpoints in L or both endpoints in R. We add the active odd cycle transversal in the arrow from
into Z. For all other good subsets S′ we add the passive odd cycle transversal in the
xℓ
i,S to the cycle Y ℓ
i
arrow from xℓ
i contains one vertex in Z and the remaining
vertices of Y ℓ
i,S into Z and partition the remaining vertices of X ℓ
into Z. Thus each cycle Y ℓ
i,S to the cycle Y ℓ
i
i can be distributed into L and R.
For every arrow that goes from a vertex xℓ
h we add the active odd cycle transversal
of the arrow to Z if xℓ
i,S ∈ Z and add the passive odd cycle transversal to Z otherwise. Again the
remaining vertices on each arrow can easily be partitioned into L and R such that no edge has both
endpoints in L or both endpoints in R. This concludes the construction of Z. Since we have put the
vertices into Z in accordance to the budget described in the construction it follows that Z ≤ α. All that
remains to show, is that for each 1 ≤ h ≤ m and 0 ≤ r < n + 1, the cycle bC r
h has at least one active
arrow pointing into it.
i,S into a cycle bC r
The cycle bC r
h corresponds to the clause Ch. The clause Ch is satisfied by γ and hence it is satisfied
by the restriction of γ to a group Fi. This restriction is a group assignment of Fi and hence it corresponds
to a good subset S of P rm+h
, which happens to be exactly Z ∩ P rm+h
∈ Z and since the
restriction of γ to Fi satisfies Ch there is an arrow pointing from xrm+h
h. Since this arrow is
active, this concludes the proof.
. Thus xrm+h
and into bC r
i,S
i,S
i
i
Lemma 16. If G has an odd cycle transversal of size α, then φ is satisfiable.
17
Proof. Let Z be an odd cycle transversal of G of size α. Since G \ Z is bipartite, the vertices of G \ Z
can be partitioned into L and R such that every edge of G \ Z has one endpoint in L and the other in
R. Given Z, L and R, we construct a satisfying assignment to φ. Every arrow in G must contain at
least two vertices in Z, not counting the startpoint of the arrow. Let ~Z be a subset of Z containing two
vertices from each arrow, but no arrow start point. Observe that no two arrows have the same endpoint,
and therefore ~Z is exactly two times the number of arrows in G. Let Z ′ = Z \ ~Z.
We argue that for any 1 ≤ i ≤ t and 0 ≤ ℓ < m(tp+1) we have Z ′∩(Lℓ
i, Aℓ
i, Rℓ
i, Bℓ
4p. Observe that no vertices in Lℓ
contain any vertices of ~Z. Suppose for contradiction that Z ′ ∩ (Lℓ
Then there is a vertex in bl ∈ Lℓ
br ∈ R. Furthermore, there is a 1 ≤ j ≤ p such that
, a3ℓ+3
i \ Z ′, and a vertex br ∈ Rℓ
Z ′ ∩ {p3ℓ+1
i or P ℓ
, a3ℓ+2
, a3ℓ+1
, p3ℓ+2
, p3ℓ+3
, b3ℓ+1
, b3ℓ+2
i , Qℓ
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i ∪Aℓ
i ∪Rℓ
i ∪Bℓ
i ∪Qℓ
i ) ≥
i are endpoints of arrows, and hence they do not
i ) < 4p.
i \ Z ′. Without loss of generality, bl ∈ L and
i ∪ Qℓ
i ∪ Bℓ
i ∪ P ℓ
i ∪ Rℓ
i ∪ Aℓ
i ∪P ℓ
, b3ℓ+3
i,j
, q3ℓ+1
i,j
, q3ℓ+2
i,j
, q3ℓ+3
i,j } < 4.
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
i,j
, c3ℓ+1
, b3ℓ+1
, c3ℓ+3
, b3ℓ+3
, c3ℓ+2
, b3ℓ+2
, p3ℓ+3
, p3ℓ+2
i ∪ Bℓ
i ∪ Qℓ
or b3ℓ+2
Since {a3ℓ+1
i,j }, {a3ℓ+2
i,j } and {a3ℓ+3
Hence, there are two cases to consider either (1) {p3ℓ+1
i,j } form triangles and must
contain a vertex from Z ′ each, it follows that each of these triangles contain exactly one vertex from Z ′,
i,j } = ∅. Since bl ∈ L and br ∈ R, bl is adjacent to all vertices of Aℓ
and that Z ′ ∩ {p3ℓ+1
, p3ℓ+3
i,j and
i,j it follows that Aℓ
br is adjacent to all vertices of Bℓ
{p3ℓ+1
i,j } ⊆ R and p3ℓ+2
and hence either a3ℓ+2
tradiction. The second case is similar, either a3ℓ+1
or b3ℓ+1
Z ′ ∩ (Lℓ
∈ R or (2)
∈ L
have both endpoints in the same set out of {L, R}, a con-
i,j p3ℓ+1
have both endpoints in the same set out of {L, R}, a contradiction. We conclude that
i ∪ Aℓ
∈ L. In the first case observe that either a3ℓ+2
i,j p3ℓ+3
i,j \ Z ′ ⊆ R and Bℓ
, p3ℓ+3
∈ L and hence either a3ℓ+1
i,j } ⊆ L and p3ℓ+2
i,j p3ℓ+2
i,j
i ∪ Rℓ
i,j
i,j p3ℓ+2
i,j \ Z ′ ⊆ L.
∈ R or b3ℓ+1
∈ R or b3ℓ+2
i ) ≥ 4p.
For any 1 ≤ i ≤ t and 0 ≤ ℓ < m(tp + 1), Y ℓ
i contains a vertex in Z.
i
i contains no vertices of Z ′ it contains a vertex from ~Z and there is an active arrow pointing into
i . Since the arrow is
i,S ∈ Z ′. Hence for any 1 ≤ i ≤ t and
i,S ∈ Z ′ or at least one
If Y ℓ
Y ℓ
i . The starting point of this arrow is a vertex xℓ
active and xℓ
0 ≤ ℓ < m(tp + 1) we have that either there is a good subset S of P ℓ
vertex of Y ℓ
i
i,S is not the endpoint of any arrow, we know that xℓ
i,S for some good subset S of P ℓ
is an odd cycle so Y ℓ
i such that xℓ
The above arguments, together with the budget constraints, imply that for every 1 ≤ i ≤ t and
i ) = 1,
i . Let
i,S has odd length, and hence it must contain some vertex from Z. On the
i,S cannot contain any vertices from ~Z.
0 ≤ ℓ < m(tp+1), we have Z ′∩(Lℓ
i ∪Qℓ
where the union is taken over all good subsets S of P ℓ
S = Z ′ ∩ P ℓ
other hand, all the arrows pointing into X ℓ
Thus X ℓ
i,S contains a vertex from Z ′, and by the budget constraints this must be xℓ
i∪P ℓ
i . It follows Z ′ ∩ P ℓ
i ) = 4p and that Z ′∩S{xℓ
i is a good subset of P ℓ
i,S are passive, so X ℓ
i . The cycle X ℓ
i,S}∪V (Y ℓ
is in Z ′.
i ∪ P ℓ
i∪Bℓ
i ∪Rℓ
i ∪Aℓ
i,S.
i,j
i,j
i
i
i
i
i
i
i,j, pℓ+1
i,j , pℓ+2
i,j
, bl⌊ℓ/3⌋+1
nor br⌊ℓ/3⌋+1
, br⌊ℓ/3⌋
, br⌊ℓ/3⌋+1
} ⊆ R. There are two cases. Either pℓ
Now, consider three consecutive vertices pℓ
for some 1 ≤ i ≤ t, 1 ≤ j ≤ p, 1 ≤ ℓ ≤
3m(tp + 1) − 2. We prove that at least one of them has to be in Z. Suppose not. We know that neither
bl⌊ℓ/3⌋
are in Z. Thus, without loss of generality {bl⌊ℓ/3⌋
} ⊆ L and
{br⌊ℓ/3⌋
i,j ∈ L or pℓ+1
i,j ∈ R.
In the first case we obtain a contradiction since either aℓ
i,j ∈ L. In the second case we get
a contradiction since either aℓ+1
i,j ∈ L. Hence for any three consecutive vertices on Pi,j, at
least one of them is in Z. Since the budget constraints ensure that there are at most V (Pi,j)/3 vertices
in Pi,j ∩ Z it follows from the pigeon hole principle, that there is an 0 ≤ r < n + 1 such that for any
1 ≤ i ≤ t and 1 ≤ h ≤ m and 1 ≤ h′ ≤ m the set P rm+h
∩ Z. Here equality is in
the sense of equality of good subsets of P ℓ
i .
i,j ∈ R and pℓ+1
i,j ∈ R or bℓ
∩ Z equals P rm+h′
i,j ∈ L and pℓ+3
i,j ∈ R or bℓ+1
For every 1 ≤ i ≤ t, P rm+1
∩ Z corresponds to a group
assignment of Fi, then we set the variables in Fi to this assignment. Otherwise we set all the variables
∩ Z is a good subset of P rm+1
,bl⌊ℓ/3⌋+1
. If P rm+1
i
i
i
i
i
i
i
18
in Fi to false. We need to argue that every clause Ch is satisfied by this assignment. Consider the cycle
bC r
h. Since it is an odd cycle, it must contain a vertex from Z, the budget constraints and the discussion
above implies that this vertex is from ~Z. Hence there must be an active arrow pointing into bC r
h. The
starting point of this active arrow is a vertex xmr+h
. The set
S corresponds to a group assignment of Fi that satisfies Ch. Since the arrow is active xmr+h
∈ Z ′,
and by the discussion above we have that P mr+h
∩ Z ′ and S is equal to
P mr+1
∩ Z ′ and hence the assignment to the variables of Fi satisfies Ch. Since this holds for all clauses,
this concludes the proof.
for some i and good subset S of P mr+h
∩ Z ′ = S. Now, S = P mr+h
i,S
i,S
i
i
i
i
Lemma 17. pw(G) ≤ t(p + 1) + 10p3p.
i,j
Proof. We show how to search the graph using at most t(p + 1) + 10p3p searchers. The strategy consists
of m(tp + 1) rounds numbered from round 0 to round m(tp + 1) − 1. Each round has t stages, numbered
from 1 to t. In the beginning of round k there is a searcher on p3k+1
i for every 1 ≤ i ≤ t,
1 ≤ j ≤ p. Let r and 1 ≤ h ≤ m be integers such that k + 1 = rm + h.Recall, that as we go
around bC r
h counterclockwise we first encounter vertices corresponding to group assignments of F1, then
to assignments of F2 and so on. In the beginning of round k we place a searcher on the first vertex on
bC r
h that corresponds to an assignment of F1. If bC r
h contains a dummy vertex, we place a searcher on this
vertex as well. These two searchers will remain on their respective vertices throughout the round. In the
beginning of stage s of round k we will assume that the vertices on the cycle bC r
h corresponding to group
assignments of Fs′, s′ < s have already been cleaned, and in the beginning of every stage s > 1, there
is a searcher standing on the first vertex corresponding to a group assignment of Fs.
s , Qk
In stage s of round k, we place searchers on all vertices of P k
and brk
s , Bk
s , Rk
s and all
s , on all vertices of arrows starting or ending in such
h corresponding to group assignments of Fs. In total this amounts to less
s , Ak
s , Y k
s , Lk
s,S for every good subset S of P k
vertices of cycles X k
cycles, and on all vertices of bC r
than 10p3p vertices.
s
In the last part of stage s of round k, we place searchers on p3(k+1)+1
for every 1 ≤ j ≤ p and on
brk+1
. Then we remove all the searchers that were placed out in the first part of phase s except for the
searcher on the last vertex on bC r
h corresponding to a group assignment of Fs. Unless s = 1 there is
also a searcher on the last vertex on bC r
h corresponding to a group assignment of Fs−1. We remove this
searcher, and the next stage can commence. In the end of the last stage of round k we remove all the
searchers from bC r
h. Then the last stage can commence. At any point in time, at most t(p + 1) + 10p3p
searchers are placed on G.
s,j
Proof (of Theorem 5). Suppose ODD CYCLE TRANSVERSAL can be solved in O∗((3 − ǫ)pw(G)) time
for ǫ < 1. Then there is an ǫ′ < 1 such that O∗((3 − ǫ)pw(G)) ≤ O∗(3ǫ′pw(G)). We chose p large
enough such that ǫ′ · p+1
p−1 = δ′ < 1. Given an instance of SAT we construct an instance of ODD
CYCLE TRANSVERSAL using the above construction and the chosen value of p. Then we solve the ODD
CYCLE TRANSVERSAL instance using the O∗((3 − ǫ)pw(G)) time algorithm. Correctness is ensured by
Lemmata 15 and 16. Lemma 17 yields that the total time taken is upper bounded by O∗((3 − ǫ)pw(G)) ≤
O∗(3ǫ′pw(G)) ≤ O∗(3ǫ′(t(p+1)+f (ǫ′))) ≤ O∗(3ǫ′⌈
(p−1) log 3 ) ≤
O∗(3δ′ n
log 3 ) ≤ O∗(2δ′n) =. O∗((2 − δ)n) for δ < 1.
⌊p log 3⌋ ⌉(p+1)) ≤ O∗(3ǫ′ n(p+1)
⌊p log 3⌋ ) ≤ O∗(3ǫ′ n(p+1)
n
8 Partition Into Triangles
A triangle packing in a graph G is a collection of pairwise disjoint vertex sets S1, S2, . . . St in G such
If V (G) = Si≤t Si then
that Si induces a triangle in G for every i. The size of the packing is t.
the collection S1 . . . St is a partition of G into triangles. In the TRIANGLE PACKING problem we are
given a graph G and an integer t and asked whether there is a triangle packing in G of size at least
19
t. In the PARTITION INTO TRIANGLES problem we are given a graph G and asked whether G can
be partitioned into triangles. Notice that since PARTITION INTO TRIANGLES is the special case of
TRIANGLE PACKING when the number of triangles is the number of vertices divided by 3, the bound of
Theorem 6 holds for TRIANGLE PACKING as well.
Theorem 6. If PARTITION INTO TRIANGLES can be solved in O∗((2 − ǫ)pw(G)) for ǫ > 0 then SAT
can be solved in O∗((2 − δ)n) time for some δ > 0.
Construction. first show the lower bound for TRIANGLE PACKING and then modify our construction
to also work for the more restricted PARTITION INTO TRIANGLES problem. Given an instance φ of
SAT we construct a graph G as follows. For every variable vi we make a path Pi on 2m(n + 1) + 1
vertices. We denote the l'th vertex of Pi by pl
i. For every i we add a set Ti of 2m(n + 1) vertices, and
let the l'th vertex of Ti be denoted tl
i. For every 1 ≤ l ≤ 2m(n + 1) we add the edges tl
i and tl
ipl+1
ipl
.
i
For every clause Cj we add n + 1 gadgets corresponding to the clause.
0 ≤ r ≤ n we do the following. First we add the vertices bcr
j t2(mr+j)
vi that occurs in Cj positively we add the edges bcr
j t2(mr+j)−1
and bdr
occurs in Cj negated we add the edges bcr
i
clause Cj concludes the construction of G.
Lemma 18. If φ satisfiable, then G has a triangle packing of size mn(n + 1) + m(n + 1).
j and bdr
and bdr
j t2(mr+j)−1
i
i
In particular, for every
j bdr
j and the edge bcr
j. For every variable
j t2(mr+j)
. For every variable vi that
. Doing this for every r and every
i
i
i
, p2l
i , p2l
, p2l−1
i , p2l+1
Proof. Consider a satisfying assignment to φ. For every variable vi that is set to true and integer 1 ≤ l ≤
m(n + 1) we add {t2l−1
i } to the triangle packing. For every variable vi that is set to false and
integer 1 ≤ l ≤ m(n + 1) we add {t2l
} to the triangle packing. For every clause Cj there is
a literal set to true. Suppose this literal corresponds to the variable vi. Notice that if vi occurs positively
in Cj, then vi is set to true, and if it occurs negatively it is set to false. For each 0 ≤ r ≤ n, if vi occurs
positively in Cj, then t2(mr+j)
} to
the triangle packing. On the other hand, if vi occurs negated in Cj then t2(mr+j)−1
has not yet been used
in any triangle, so we can add {bcr
} to the triangle packing. In total mn(n+1)+m(n+1)
triangles are packed.
has not yet been used in any triangle, so we can add {bcr
j , t2(mr+j)−1
j , t2(mr+j)
j , bdr
j , bdr
i
i
i
i
i
j and bdr
j also contains bdr
Lemma 19. If G has a triangle packing of size mn(n + 1) + m(n + 1), then φ satisfiable.
Proof. Observe that for any j and r, every triangle that contains bcr
j and vice versa.
Furthermore, if we remove all the vertices bcr
j for every j and r from G we obtain a disconnected
graph with n connected components, G[Ti ∪ V (Pi)] for every i. Thus, the only way to pack mn(n +
1) + m(n + 1) triangles in G is to pack mn(n + 1) triangles in each component G[Ti ∪ V (Pi)] and in
addition make sure that every pair (bcr
The only way to pack mn(n + 1) triangles in a component G[Ti ∪ V (Pi)] is to use every second
triangle of the form {tl
}, except possibly at one point where two triangles on this form are
skipped. By the pigeon hole principle there is an 0 ≤ r ≤ n such that for every i, every second triangle
of the form {t2mr+l
} for 1 ≤ l ≤ 2m is used. We make an assignment to the variables
of φ as follows. For every i such that {t2mr+1
} is used, vi is set to true, and otherwise
{t2mr+2
} is used in the packing and vi is set to false. We prove that this assignment
satisfies φ.
j ) is used in some triangle in the packing.
, p2mr+l+1
, p2mr+l+1
, p2mr+1
, p2mr+2
, p2mr+3
, p2mr+l
i, pl+1
j , bdr
i, pl
i
i
i
i
i
i
i
i
i
i
j , bdr
i
or t2(mr+j)−1
for some i. If it contains t2(mr+j)
For every j, the pair (bcr
t2(mr+j)
i
since the triangle packing contains every second triangle of the form {t2mr+l
i
1 ≤ l ≤ 2m, it follows that the triangle packing contains {t2mr+1
, p2mr+1
j , bdr
set to true. By an identical argument, if the triangle containing the pair (bcr
vi occurs negated in Cj and vi is set to false. This concludes the proof.
j ) is used in some triangle in the packing. This triangle either contains
, then vi occurs positively in Cj. Furthermore,
, p2mr+l+1
} for
} and hence vi is
then
, p2mr+l
, p2mr+l+1
j ) contains t2(mr+j)−1
i
i
i
i
i
i
i
20
We now modify the construction to work for PARTITION INTO TRIANGLES instead of TRIANGLE
PACKING. Given the graph G as constructed from φ, we construct a graph G′ as follows. For every
1 ≤ i ≤ n and 1 ≤ l ≤ m(n + 1) we make a clique Ql
i are all
i and to t2l−1
. For every i < n and and 1 ≤ l ≤ m(n + 1) we make all vertices of Ql
adjacent to t2l
i
adjacent to all vertices of Ql
i+1. Suppose that 2n + 2 is p modulo 3 for some p ∈ {0, 1, 2}. We remove
p vertices from Ql
i on four vertices. The vertices of Ql
n for every l ≤ m(n + 1).
i
Lemma 20. G has a triangle packing of size t if and only if G′ can be partitioned into triangles.
Proof. In the forward direction, consider a triangle packing of size t in G as constructed in Lemma 18.
We can assume that the triangle packing has this form, because by Lemma 19 we have that φ is satisfi-
able.
i
i
i and t2l−1
For every fixed 1 ≤ l ≤ m(n + 1), we proceed as follows. We know that there exists an i such that
is used in
and
both t2l
i
i and t2l−1
the packing. For each such i′, we make a triangle containing the unused vertex out of t2l
two vertices of Ql
are used in the packing. For every i′ 6= i, exactly one out of t2l
i and t2l−1
i′. Then we "clean up" Ql
n as follows.
1, . . . , Ql
3. Continue in this fashion until arrive at Ql
In particular, we start with the yet unused vertices of Ql
2. Now Ql
1. There are two of them. Make a triangle
containing these two vertices and one vertex of Ql
2 has one unused vertex left. Make a triangle
containing this vertex and the two unused vertices of Ql
i. At
this point we have used 0, 1 or 2 vertices of Ql
i−1. The case
when we have used 0 vertices of Ql
i also covers the case that i = 1. If we only used 0 or 1 vertices of Ql
i,
then we add a triangle that contains 3 vertices of Ql
i, then their
number is either 1 or 2. We make a triangle containing these vertices and 1 or 2 of the unused vertices of
Ql
i+1. Now we proceed to Ql
n. Since the total number
of vertices in Sj≤n Ql
j is 4n − p, we know that 2n − 2 of these vertices are used for triangles with
vertices of G, and 2n + 2 − p is divisible by 3 the process described above will partition all the unused
vertices of Sj≤n Ql
i+1 and continue in this manner until we reach Ql
i a triangle containing some vertices in Ql
i. If there are still unused vertices in Ql
j into triangles.
In the reverse direction, we argue that in any partitioning of G′ into triangles, exactly t triangles
must lie entirely within G. In fact, we argue that for any l ≤ m(n + 1) exactly n − 1 vertices out of
Si≤n{t2l
} are used in triangles containing vertices from Si≤n Ql
i.
j. Furthermore, for each i ≤ n the vertex p2l
Pick 1 ≤ j ≤ m and r such that l = mr + j. Exactly one out of Si≤n{t2l
j and bdr
} is in a triangle
i must be in a triangle either containing t2l
i or
} are used in triangles containing vertices from
j or some vertex in
} are used in triangles containing vertices
i. Thus in the packing, exactly 3t vertices in G′ are contained in triangles completely inside
with bcr
i . Hence, at most n − 1 vertices out of Si≤n{t2l
i , t2l−1
t2l
Si≤n Ql
i or t2l−1
i. Furthermore, any triangle containing t2l
i. Hence exactly n − 1 vertices out of Si≤n{t2l
Si≤n Ql
from Si≤n Ql
G, and hence G has a triangle packing of size t.
} must either contain p2l
i
i , t2l−1
i , t2l−1
i , t2l−1
i , bcr
i
i
i
i
To complete the proof for PARTITION INTO TRIANGLES we need to bound the pathwidth of G′.
Lemma 21. pw(G′) ≤ n + 10.
Proof. We give a search strategy for G′ that uses n + 10 searchers. The strategy consists of m(n + 1)
rounds and each round has n stages. In the beginning of round l, 1 ≤ l ≤ m(n + 1), there are searchers
n searchers placed, one on each vertex p2l−1
for every i. Let r and 1 ≤ j ≤ m be integers such that
l = mr +j. We place one searcher on bcr
j. These two searchers will stay put throughout the
duration of this round. In stage i of round l we place searchers on all vertices of Ql
i+1. Then we
place searchers on t2l−1
i and p2l+1
i, t2l−1
,
t2l
i . We then proceed to the next stage. At the end of the round we remove the searchers from bcr
i and p2l
j
and bdr
. At the end of stage i we remove the searchers from Ql
for every i, and the next round can commence.
j and one on bdr
i and Ql
i , p2l
, t2l
i
i
i
i
j . Notice that now, there are searchers on p2l+1
Lemmata 18,19,20 and 21 prove Theorem 6.
i
21
9 Conclusion
We have showed that for a number of basic graph problems, the best known algorithms parameterized by
treewidth are optimal in the sense that base of the exponential dependence on treewidth is best possible.
Recall that for DOMINATING SET and PARTITION INTO TRIANGLES, this running time was obtained
quite recently using the new technique of fast subset sum convolutions [27]. Thus it could have been a
real possibility that the running time is improved for some other problems as well.
The results are proved under the Strong Exponential Time Hypothesis (SETH). While this hypothesis
is relatively recent and might not be accepted by everyone, our results at least make a connection between
rather specific graph problems and the very basic issue of better SAT algorithms. Our results suggest that
one should not try to find better algorithms on bounded treewidth graphs for the problems considered
in the paper: as this would disprove SETH, such an effort is better spent on trying to disprove SETH
directly in the domain of satisfiability. Finally, we suggest the following open questions for future work:
• Can we prove similar tight lower bounds under the restriction that the graph is planar? Or is it
possible to find improved algorithms on bounded treewidth planar graphs?
• Can we prove tight lower bounds for problems parameterized not by treewidth, but by something
else? Naturally, one should look at problems where the algorithm or the the running time suggests
that the best known algorithm is optimal. Possible candidates are the O(2k) time algorithm for
STEINER TREE with k terminals [2], the O(2k) time randomized algorithm for k-PATH [29],
and the O(2k) (resp., O(3k)) time algorithms for EDGE BIPARTIZATION (resp., ODD CYCLE
TRANSVERSAL) [16, 22].
• For the q-COLORING problem, we were able to prove lower bounds parameterized by the feedback
vertex set number. Can we prove such bounds for the other problems as well?
References
[1] J. Alber and R. Niedermeier. Improved tree decomposition based algorithms for domination-like
problems. In LATIN, pages 613–628, 2002.
[2] A. Bjorklund, T. Husfeldt, P. Kaski, and M. Koivisto. Fourier meets mobius: fast subset convolu-
tion. In STOC, pages 67–74, 2007.
[3] C. Calabro, R. Impagliazzo, and R. Paturi. The complexity of satisfiability of small depth circuits.
In IWPEC, pages 75–85, 2009.
[4] J. Chen, X. Huang, I. A. Kanj, and G. Xia. On the computational hardness based on linear FPT-
reductions. J. Comb. Optim., 11(2):231–247, 2006.
[5] J. Chen, X. Huang, I. A. Kanj, and G. Xia. Strong computational lower bounds via parameterized
complexity. J. Comput. Syst. Sci., 72(8):1346–1367, 2006.
[6] E. D. Demaine, F. V. Fomin, M. T. Hajiaghayi, and D. M. Thilikos. Subexponential parameterized
algorithms on bounded-genus graphs and -minor-free graphs. J. ACM, 52(6):866–893, 2005.
[7] E. D. Demaine and M. Hajiaghayi. The bidimensionality theory and its algorithmic applications.
Comput. J., 51(3):292–302, 2008.
[8] D. Eppstein. Diameter and treewidth in minor-closed graph families. Algorithmica, 27(3):275–291,
2000.
[9] S. Fiorini, N. Hardy, B. A. Reed, and A. Vetta. Planar graph bipartization in linear time. Discrete
Applied Mathematics, 156(7), 2008.
22
[10] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer, Berlin, 2006.
[11] F. Fomin, P. Golovach, D. Lokshtanov, and S. Saurabh. Algorithmic lower bounds for problems
parameterized by clique-width. In SODA, pages 493–502, 2010.
[12] F. V. Fomin, S. Gaspers, S. Saurabh, and A. A. Stepanov. On two techniques of combining branch-
ing and treewidth. Algorithmica, 54(2):181–207, 2009.
[13] R. Impagliazzo and R. Paturi. On the complexity of k-sat. J. Comput. Syst. Sci., 62(2):367–375,
2001.
[14] R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential complexity? J.
Comput. Syst. Sci., 63(4):512–530, 2001.
[15] J. Kleinberg and E. Tardos. Algorithm Design. Addison-Wesley Longman Publishing Co., Inc.,
Boston, MA, USA, 2005.
[16] D. Lokshtanov, S. Saurabh, and S. Sikdar. Simpler parameterized algorithm for oct. In IWOCA,
pages 380–384, 2009.
[17] D. Marx. Can you beat treewidth? In FOCS, pages 169–179, 2007.
[18] D. Marx. On the optimality of planar and geometric approximation schemes.
In FOCS, pages
338–348, 2007.
[19] D. Molle, S. Richter, and P. Rossmanith. Enumerate and expand: Improved algorithms for con-
nected vertex cover and tree cover. Theory Comput. Syst., 43(2):234–253, 2008.
[20] R. Niedermeier. Invitation to fixed-parameter algorithms, volume 31 of Oxford Lecture Series in
Mathematics and its Applications. Oxford University Press, Oxford, 2006.
[21] M. Patras¸cu and R. Williams. On the possibility of faster sat algorithms. In Proc. 21st ACM/SIAM
Symposium on Discrete Algorithms (SODA), 2010. To appear.
[22] B. Reed, K. Smith, and A. Vetta. Finding odd cycle transversals. Operations Research Letters,
32(4):299–301, 2004.
[23] A. D. Scott and G. B. Sorkin. Linear-programming design and analysis of fast algorithms for max
2-csp. Discrete Optimization, 4(3-4):260–287, 2007.
[24] A. Takahashi, S. Ueno, and Y. Kajitani. Mixed searching and proper-path-width. Theor. Comput.
Sci., 137(2):253–268, 1995.
[25] J. A. Telle and A. Proskurowski. Practical algorithms on partial k-trees with an application to
domination-like problems. In WADS, pages 610–621, 1993.
[26] D. M. Thilikos, M. J. Serna, and H. L. Bodlaender. Cutwidth i: A linear time fixed parameter
algorithm. J. Algorithms, 56(1):1–24, 2005.
[27] J. M. M. van Rooij, H. L. Bodlaender, and P. Rossmanith. Dynamic programming on tree decom-
positions using generalised fast subset convolution. In ESA, pages 566–577, 2009.
[28] J. M. M. van Rooij, J. Nederlof, and T. C. van Dijk.
Inclusion/exclusion meets measure and
conquer. In ESA, pages 554–565, 2009.
[29] R. Williams. Finding paths of length k in o*(2k) time. Inf. Process. Lett., 109(6):315–318, 2009.
23
|
1710.10026 | 1 | 1710 | 2017-10-27T08:37:26 | A note on faithful coupling of Markov chains | [
"cs.DS",
"math.PR"
] | One often needs to turn a coupling $(X_i, Y_i)_{i\geq 0}$ of a Markov chain into a sticky coupling where once $X_T = Y_T$ at some $T$, then from then on, at each subsequent time step $T'\geq T$, we shall have $X_{T'} = Y_{T'}$. However, not all of what are considered couplings in literature, even Markovian couplings, can be turned into sticky couplings, as proved by Rosenthal through a counter example. Rosenthal then proposed a strengthening of the Markovian coupling notion, termed as faithful coupling, from which a sticky coupling can indeed be obtained. We identify the reason why a sticky coupling could not be obtained in the counter example of Rosenthal, which motivates us to define a type of coupling which can obviously be turned into a sticky coupling. We show then that the new type of coupling that we define, and the faithful coupling as defined by Rosenthal, are actually identical. Our note may be seen as a demonstration of the naturalness of the notion of faithful coupling. | cs.DS | cs |
A note on faithful coupling of Markov chains
Debojyoti Dey1, Pranjal Dutta2, and Somenath Biswas1
1Department of Computer Science and Engineering, Indian Institute of Technology Kanpur,
2Chennai Mathematical Institute, Chennai-603103.
Kanpur-208016.
October 2017
Abstract
One often needs to turn a coupling (Xi, Yi)i≥0 of a Markov chain into a sticky
coupling where once XT = YT at some T , then from then on, at each subsequent
time step T ′ ≥ T , we shall have XT ′ = YT ′ . However, not all of what are considered
couplings in literature, even Markovian couplings, can be turned into sticky couplings,
as proved by Rosenthal [Ro97] through a counter example. Rosenthal then proposed a
strengthening of the Markovian coupling notion, termed as faithful coupling, from which
a sticky coupling can indeed be obtained. We identify the reason why a sticky coupling
could not be obtained in the counter example of Rosenthal, which motivates us to define
a type of coupling which can obviously be turned into a sticky coupling. We show then
that the new type of coupling that we define, and the faithful coupling as defined by
Rosenthal [Ro97], are actually identical. Our note may be seen as a demonstration of
the naturalness of the notion of faithful coupling.
1
Introduction
First, we recall certain basic definitions and fix our notations. For a random variable X,
we denote its distribution as D(X). Let M be a finite Markov chain on the state space
Ω, and with the transition matrix P . Following [No97], we abbreviate the Markov chain
(Xi)i≥0, with D(X0) = λ and D(Xi+1) = D(Xi)P for i ≥ 0, as Markov(λ, P ). A coupling
of the Markov chain M is a process (Xi, Yi)i≥0, with each Xi, Yi taking values in Ω and
evolving on a common probability space, where (Xi)i≥0 satisfies the property that, for i ≥ 0,
D(Xi+1) = D(Xi)P and similarly, (Yi)i≥0 satisfies, for i ≥ 0, D(Yi+1) = D(Yi)P .
Remark A: The definition of Markov chain coupling, as given in [LPW09], says that both (Xi)i≥0
and (Yi)i≥0 are Markov chains with transition matrix P . The reason we prefer the above weaker
definition is because in literature we find processes considered as couplings that do not satisfy
1
the stronger notion of [LPW09], but do satisfy our above definition. A Markov chain (Xi)i≥0
with transition matrix P , by definition, satisfies the property (π) that for i ≥ 0, conditioned on
Xi = l, Xi+1 has the distribution P (l, ·) and is indepedent of X0, . . . , Xi−1. As we shall see in the
next section, there is a Markovian coupling (Xi, Yi)i≥0, where (Xi)i≥0 does not satisfy the property
(π), although Xi's do evolve (under certain conditions) as per the transition matrix P .
Let D(X0) be µ and D(Y0) be ν. In the above, since each pair, Xi and Yi, i ≥ 0, is defined
on the same probability space, the two random variables, Xi, Yi, for each i ≥ 0, are coupled,
and so we have, from the above definition of Markov chain coupling, and Proposition 4.7 of
[LPW09]:
kµP i − νP ikTV = kD(Xi) − D(Yi)kTV ≤ Pr(Xi 6= Yi)
(1)
(Here, k · kTV denotes the total variational distance between two distributions.)
Let us suppose that a coupling of M has the 'now-equals-forever' property ([Ro97]), namely,
if Xn = Yn for some n, then for all j ≥ n, Xj = Yj. Then, if T is defined as the random
time
(2)
(3)
we get from the relation (1)
T
def
= inf{iXi = Yi}
kD(Xi) − D(Yi)kTV ≤ Pr(Xi 6= Yi) = Pr(T > i)
as the two events (Xi 6= Yi) and (T > i) imply each other, making use of the 'now-equals-
forever'propery.
As such, the applicability of the inequality (3) is limited as most couplings of Markov chains
will not have the 'now-equals-forever' property. The way that is often resorted to to tackle
this problem is to define a new process (Zi)i≥0 as follows: ([Ro97])
Zi = (cid:26) Xi
Yi
if i ≤ T
otherwise
(4)
where T is as in (2) above.
If it so happens that (Zi)i≥0 = Markov(µ, P ) then, since
(Zi, Yi)i≥0 is a coupling of the Markov chain satisfying 'now-equals-forever' property, we
have
kµP i − νP ikTV = kD(Zi) − D(Yi)kTV ≤ Pr(T > i)
The construction of (Zi) has been evocatively termed as sticking and the resultant new
coupling as a sticky coupling [HM17].
Definition 1 (Markovian coupling of a Markov chain) A coupling (Xi, Yi)i≥0 of a Markov
chain M on state space Ω, and with its transition matrix as P , where D(X0) = µ and
def=
D(Y0) = ν, for given µ and ν, is said to be a Markovian coupling if (Wi)i≥0, with Wi
(Xi, Yi) is itself a Markov chain on state space Ω × Ω, with a specified joint distribution
of D(X0) and D(Y0) as the distribution D(W0). In other words, there is a (Ω × Ω, Ω × Ω)
2
transition matrix, say Q, such that for some distribution θ0 on Ω × Ω, satisfying that its
def= θi−1Q, the two
marginals are µ = D(X0) and ν = D(Y0), we will have, for i ≥ 1, with θi
marginals of θi will be µP i, that is, D(Xi) and νP i, that is D(Yi).
Remark B: We note that the defintion above is specific to a joint distribution of (X0, Y0). The
reason for the specificity is that in the example given by Rosenthal ([Ro97]) of a Markovian coupling
for which sticking fails, as we note in the next section, the ability to evolve the two copies as per
the transition matrix P crucially depends on the initial coupling of X0 and Y0.
As noted in [HM17], it has been stated at times that sticking will work for all Markovian
couplings. However, this is not correct, Rosenthal [Ro97] provides a counter example. He
then provides a stronger version of Markovian couplings, termed as faithful couplings, for
which it is proved that sticking will provably result in a coupling of M where (Zi)i≥0, as
defined in (4) above, will indeed be Markov(µ, P ). We discuss Rosenthal's counterexample
in Section 2 and identify what we consider to be the reason why the sticking operation fails
there. This reason then motivates us in defining a stronger version of Markovian coupling
(Section 3) for which it is easy to see that the sticking will indeed work. We then prove that
the new notion of coupling that we define is actually equivalent to the notion of faithful
coupling.
2 Rosenthal's counterexample
In his example, Rosenthal [Ro97] considers the Markov chain M with state space Ω = {0, 1},
and with the transition matrix P defined as:
P =
0
1
0 1/2 1/2
1 1/2 1/2
(Xi), (Yi), i ≥ 0 are defined as follows:
1. Each of D(X0), D(Y0) is the uniform distribution on the state space {0, 1},
2. Each D(Yi), i > 0 is also the uniform distribution, whereas for i ≥ 0, Xi+1
def
= Xi ⊕ Yi,
⊕ being the exclusive-or operation.
Thus, the joint evolution of the two copies (Xi, Yi) of the chain, on the state space Ω × Ω,
that is, {(0, 0), (0, 1), (1, 0), (1, 1)} is governed by the following transition matrix Q:
Q =
(0, 0)
1/2
0
0
1/2
(0, 1)
1/2
0
0
1/2
(1, 0)
(1, 1)
0
1/2
1/2
0
0
1/2
1/2
0
(0, 0)
(0, 1)
(1, 0)
(1, 1)
3
Convention: The distribtion of a random variable X on Ω = {0, 1} is specified as the
vector
[Pr(X = 0) Pr(X = 1)]
and similarly, the distribution of a random variable W on Ω × Ω as the vector
[Pr(W = (0, 0)) Pr(W = (0, 1)) Pr(W = (1, 0)) Pr(W = (1, 1))]
We note that for any distribution σ
1 − p], Ω = {0, 1}, σP will be the uniform
distribution. We also note that for the trivial coupling θ, of [1/2
1/2],
which is [1/4 1/4 1/4 1/4], θQ will again be θ. Therefore, with the joint distribution θ
of D(X0) and D(Y0), Q indeed defines a Markovian coupling of the Markov chain M .
1/2] and [1/2
def
= [p
3/8
1/8
1/8
1/8
1/8]. The marginals of the latter distributions are [3/4
Remark C: We also note that there are other joint distributions of the same D(X0) and D(Y0),
for which Q will not be a Markovian coupling: consider θ′ def
3/8), θ′Q =
[3/8
1/2]
respectively. Thus, if we use the joint distribution θ′ of the same two initial distributions, while Q
evolves Yi's (understandably) as per P , it does not evolve Xi's as per P . Hence with the initial joint
distribution θ′, Q fails to be a Markovian coupling of M . This is why we felt that it is necessary
to make explicit the initial joint distribution in the definition of Markovian coupling of a Markov
chain.
1/4] and [1/2
= [3/8
Rosenthal proves that although Q is a Markovian coupling of M with θ as the initial joint
distribution, the result of the sticking operation will fail to evolve M correctly: with T
defined as in (2) and Zi's as defined in (4),
Pr(Z0 = 1, Z1 = 0)
= Pr(T = 0, Y0 = 1, Y1 = 0, X0 = 1) + Pr(T > 0, X0 = 1, Y0 = 0, X1 = 0)
= 1/8 + 0
= 1/8
However, had Zi's been evolving as per P with the same initial distribution as the uniform
distribution, the probability above would have been 1/4, and not 1/8. Hence, sticking fails.
We see that sticking failed with T = 0. As it happens, sticking will fail here for every value
of T (except, of course, for T = ∞). The reason is provided by the Proposition below. We
need the following definition:
Definition 2 For a state space Ω with s an element in it, δs denotes the distribution on Ω
in which the probability of s is 1, and (therefore,) all the other states have probability 0.
Proposition 3 Let M be a Markov chain with Ω as its state space and P as its transition
matrix, and let Q be a transition matrix for the state space Ω × Ω that defines a Markovian
4
coupling (Wi = (Xi, Yi))i≥0 of M where D(X0) = µ and D(Y0) = ν, where the initial
joint distribution used is θ0. This Markovian coupling of M can be turned into a sticky
coupling using the sticking operation if the following condition holds:
for every s ∈ Ω,
defining η0 as the unique joint distribution on Ω × Ω with δs and δs as its two marginals,
and further defining ηi+1 as ηiQ, for each i ≥ 0, we have that for each ηi, i ≥ 0, will be a
joint distribution that has δsP i and δsP i as its marginals. On the other hand, if Q does not
satisfy the condition, then the Q-defined Markovian coupling, in general, cannot be turned
into a sticky coupling.
Proof Sketch.
The condition ensures that both (Xi)i≥0 and (Yi)i≥0 of the Markovian
coupling of M satisfy the strong Markov property (Theorem 1.4.2, [No97]) with respect to
the stopping time T , as defined in (2). Therefore, conditioned on T = m and XT = YT = s,
each of (XT +n)n≥0 and (YT +n)n≥0 will be Markov(δs, P ) and the former will be independent
of X0, X1, . . . , XT and the latter will be independent of Y0, Y1, . . . , YT . Thus, (XT +n)n≥0
and (YT +n)n≥0 can replace each other. Because of this, sticking is guaranteed to work as
(Xi)i≥0 and (X0, X1, . . . , XT , YT +1, . . .) will be indistinguishable.
We now show that the counter example of Rosenthal does not satify the condition as stated
in the statement of the proposition. As we have seen that the Markovian coupling there
cannot be turned into a sticky coupling, the counter example proves the second part of
the Proposition. Let us suppose that T = m and XT = YT = 0. Conditioned on these
0]. The only joint distribution with these two distributions
events, D(XT ) = D(YT ) = [1
0]. The joint distribution of XT +1 and YT +1 will be given by
as marginals is [1
[1
0], which is not
[1/2 1/2] = D(XT )P . Thus, Q fails to ensure that, conditioned as above, D(XT +1) = δ0P .
On the other hand, D(YT +1) = δ0P , and so (XT +n)n≥0 and (YT +n)n≥0 are different and
cannot replace each other, as required for the sticking operation to work.
0]. This gives D(XT +1) as [1
1/2
(cid:3)
0
0
0]Q, which is [1/2
0
0
0
3 Strong Markovian coupling
We saw in Section 2 that sticking fails for Q because it does not evolve (Xi)i≥0 as per P
for certain joint distributions. Therefore, the following definition of a type of coupling of
Markov chains immediately suggests itself to correct the defect:
Definition 4 (Strong Markovian coupling) A coupling of a finite Markov chain M
with state space Ω and transition matrix P is a strong Markovian coupling if there is a
transition matrix Q from Ω × Ω to Ω × Ω such that for every pair of distributions µ and ν,
each on Ω, and for every joint distribution θ of µ and ν, θQ will be a joint distribution of
µP and νP .
From this definition it follows that:
5
Claim 5 Let Q be a strong Markovian coupling of a finite Markov chain M with state space
Ω and transition matrix P , and for two random variables X0 and Y0 on Ω let D(X0), D(Y0)
be µ and ν respectively, and let θ0 be any joint distribution of µ and ν. If we define θi+1
as θiQ, for i ≥ 0, and Xi+1 and Yi+1 as two random variables whose distributions are
respectively the two marginals of θi+1 then
1. (Xi)i≥0 will be Markov(µ, P ) and (Yi)i≥0 will be Markov(ν, P ). Consequently,
2. D(Xi) = µP i and D(Yi) = νP i, for i ≥ 0,
3. Both (Xi)i≥0 and (Yi)i≥0 will satisfy the strong Markov property, and therefore, the
coupling Q of M can be turned into a valid sticky coupling through sticking.
We see next that the strong Markovian coupling is actually equivalent to the familiar notion
of faithful coupling which is defined as:
Definition 6 (Faithful coupling [Ro97]) A Markov coupling of a finite Markov chain
M , with Ω as its state space and P as its transition matrix, is a faithful coupling given by
a Markov chain (Wi)i≥0 = (Xi, Yi)i≥0, on state space Ω × Ω, with transition matrix Q, if Q
satisfies, for all i, j, i′, j′ ∈ Ω, the following
Q((i, j), (i′, j′)) = P (i, i′), and
Xj ′∈Ω
Q((i, j), (i′, j′)) = P (j, j′)
Xi′∈Ω
Equivalently, for all t ≥ 0 and all i, j, i′, j′ ∈ Ω
Pr(Xt+1 = i′Xt = i, Yt = j) = P (i, i′), and
Pr(Yt+1 = j′Xt = i, Yt = j) = P (j, j′)
Remark D: This kind of coupling has been termed in [LPW09], as well as in [MU05], as Markovian
coupling; however, we follow the terminology of [HM17] for reasons given therein. It is to be noted
that most coupling constructions of Markov chains turn out to be faithful couplings, as the various
coupling examples in [LPW09] demonstrate.
Proposition 7 A Markov chain coupling is faithful if and only if it is a strong Markovian
coupling.
The two lemmas below prove the two directions of the above proposition.
6
Lemma 8 Every faithful coupling of a Markov chain M is a strong Markovian coupling.
Proof.
Let M be a Markov chain with state space Ω and transition matrix P . Let the
transition matrix Q, giving transition probabilities from Ω × Ω to Ω × Ω, define a faithful
coupling of the Markov chain M . Let µ and ν be two distributions on Ω, and let θ be a
joint distribution of µ and ν. We need to prove that µP and νP are the two marginals of
θQ. In particular, we need to show:
For all x ∈ Ω,Xy∈Ω
(θQ)(x, y) = (µP )(x)
For all y ∈ Ω,Xx∈Ω
(θQ)(x, y) = (νP )(y)
and,
We prove (5) as follows:
(5)
(6)
θ(u, v)Q((u, v), (x, y))
θ(u, v)Xy∈Ω
Q((u, v), (x, y))
(θQ)(x, y)
Xy∈Ω
= Xy∈Ω X(u,v)∈Ω×Ω
= X(u,v)∈Ω×Ω
= X(u,v)∈Ω×Ω
= Xu∈Ω
= Xu∈Ω
P (u, x)Xv∈Ω
θ(u, v)
θ(u, v)P (u, x), because Q defines a faithful coupling
P (u, x)µ(u), θ being the joint ditribution of µ and ν
= (µP )(x)
In a similar manner we can prove (6).
(cid:3)
For the other direction, we prove
Lemma 9 Every strong Markovian coupling of a Markov chain M is a faithful coupling of
M .
Proof.
Let M be a Markov chain with state space Ω and transition matrix P . Let the
transition matrix Q, giving transition probabilities from Ω × Ω to Ω × Ω, define a strong
Markovian coupling of the Markov of M . For any (u, v) ∈ Ω × Ω, consider the probabilty
7
distribution δ(u,v). This distribution is the joint distribution δu and δv, both defined on Ω.
(The definition of δx as in Definition 2.) δuP is the uth row of P , namely, P (u, ·). Similarly,
δvP will be the vth row of P , namely, P (v, ·), and δ(u,v)Q will be the (u, v)th row of Q, that
is, Q((u, v), ·). As Q defines a strong Markovian coupling of M , we have that Q((u, v), ·)
will be the joint distribution of P (u, ·) and P (v, ·). From this condition, we get for any
x, u, v ∈ Ω:
Similarly, for any y, u, v ∈ Ω, we get
Q((u, v), (x, y)) = P (u, x)
Xy∈Ω
Q((u, v), (x, y)) = P (v, y)
Xx∈Ω
(7)
(8)
As (7) and (8) are precisely the conditions to be met by Q to be a faithful coupling, we
conclude that Q defines a faithful coupling of M .
(cid:3)
4 Concluding remarks
We have seen that it is sufficient for a Markovian coupling to have the faithfulness property
in order to be turned to a sticky coupling.
Is the faithfulness property also a necessary
property? It may be that there is a Markovian coupling of a Markov chain M which evolves
two copies of the chain correctly for a pair of initial distributions µ, ν of M , using a joint
distribution of the pair, and satisfies the condition in the statement of Proposition 3, but it
either does not work for some other joint distribution of the same pair µ and ν, or, does not
satisfy the condition for some other pair of initial distributions of M . Such a Markovian
coupling can be turned into a sticky coupling for the µ, ν pair, but will not be strongly
Markovian, and hence will not be a faithful coupling of M . One feels that even if such
an example exists, it is unlikely to be a natural example. Faithfulness appears to be the
only natural strengthening of the Markovian coupling notion that ensures that the sticking
operation will work.
References
[HM17] Hirscher, Timo and Anders Martinsson, Segregating Markov chains, manuscript,
arXiv:1510.036661v2 [math.PR], 2017.
[MU05] Mitzenmacher, Michael and Eli Upfal, Probability and Computing: Randomized
Algorithms and Probabilistic Analysis, Cambridge University Press, 2005.
[No97] Norris, J.R., Markov Chains, Cambridge University Press, 1997. Paperbackk edition
1998, 15th printing 2009.
8
[LPW09] Levin, David A., Yuval Peres and Elizabeth L. Wilmer, Markov Chains and Mix-
ing Times, American Mathematical Society, 2009.
[Ro97] Rosenthal, Jeffrey S., Faithful Couplings of Markov Chains: Now Equals Forever,
Advances in Applied Mathematics, Vol. 18, pp. 372 -- 381, 1997.
9
|
1009.1697 | 1 | 1009 | 2010-09-09T07:41:22 | One method of storing information | [
"cs.DS"
] | Formulate the problem as follows. Split a file into n pieces so that it can be restored without any m parts (1<=m<=n). Such problems are called problems secret sharing. There exists a set of methods for solving such problems, but they all require a fairly large number of calculations applied to the problem posed above. The proposed method does not require calculations, and requires only the operations of the division of the file into equal (nearly equal) parts and gluing them in a certain order in one or more files. | cs.DS | cs | Учреждение образов ания
«Гомельский инженерный институт Министерств а по чрезвычайным ситуациям
Респ ублики Беларусь»
Об одном методе хранения информации
Автор:
кандидат физ.-мат . наук,
Титов Олег Владимирович
2010
2
3
4
5
7
8
9
11
12
Содержание
Введение
Постан овка задачи
Описание метода
Подобные схемы
Числовые характеристики схем
Приложения метода
Оптимизация метода
Литература
3
Введение
В современном мире одними из самых актуальных задач являют ся за-
дачи передачи и хранения информации.
В данном исследов ании рассматрив ается новый метод распределенного
хранения информации.
Роль и в ажность системы хранения определяются п остоянно возрас-
тающей ценностью информации в современном обществе, возможность дос-
туп а к данным и уп равления ими является необходимым условием для вы-
полнения любых задач.
Безвозвратн ая п отеря данных может привести к фат альным последст -
виям. Утраченные вычислительные ресурсы можно восст ановить, а утрачен-
ные данные, при отсут ствии грамотно сп роектиров анной и внедренной сис-
темы резервиров ания, уже не подлежат восст ановлению.
Разработка разного рода «распределенных файловых систем» в н а-
стоящее время является одной из энергично развиваемых областей информа-
тики. Но большинство т аки х систем работ ает на основе изготовления множе -
ства п олных копий данных, хранимых в разных мест ах, и обеспечения раз-
личных механизмов синхронизации этих данных.
Автор исследов ания предлагает метод хранения информации, когда
информация разделяет ся н а маленькие «кусочки» и сохраняется на различ-
ных носителях. Предложенный автором способ позволяет восстан авлив ать
исходн ую информацию имея не все «кусочки», а лишь их некоторое количе-
ство.
В отличие от современных си стем хранения информации, этот способ
не требует полного дубли ров ания всех данных н а разных носителях. Так же
мы можем хранить все «кусочки» в открытом доступе на разных носителях,
так как имея лишь один «кусочек» невозможно восст ановить все данные . По-
следний факт позволяет исп ользовать в качестве носителей информации раз-
личные службы хранения файлов в Интернет.
Такой подход гораздо н адежнее и безоп аснее, чем хранить информа-
цию на своем собственном компьютере или н а удаленном сервере, который
может повредиться в любой момент. Это особенно актуально в связи с тем,
что количество пользов ательских данных возраст ает с каждым годом в ге о-
метрической прогрессии.
4
Постановка задачи
Сформулируем задачу следующим образом.
Задача. Разбить файл н а n ч астей таким образом, чтобы его можно бы-
ло восстан овить имея любые m ч астей (1 ≤ m ≤ n).
Такие задачи назыв аются задачами разделения секрет а [1, 2, 3, 4]. Су-
ществует мн ожество методов решения таких задач, но в се они требуют дос-
таточно большого количества вычислений применительно к поставленной
выше задаче.
Наша пост ановка задачи отлич ается от классической задачи разделения
секрета. Отличия заключ аются в следующем:
1) главная н аша задача — восстановить и сходный файл, имея п о крайне
мере m частей, тогда как в классической задаче разделения секрет а главная
задач а — невозможность восст ановить секрет при наличии меньше, чем m
частей. Другими слов ами в нашей задаче не и сключена возможность восст а-
новления исходного файла п ри наличии меньшего, чем m ч астей;
2) в классической задаче разделение секрет а мы под секретом п онима-
ем число, находящееся в определенном диап азоне возможных значений, а в
нашей постан овке задачи мы имеем ввиду произвольный набор данных пред -
ст авленных как один файл.
Предложенный автором метод не требует проведения вычислений, а
требуются только операции разделения файла на равные (почти равные) час-
ти и их склеивание в определенном порядке в один или несколько файлов.
R=
)1
5
Описание метода
Пусть у нас имеет ся и сходный файл. Требует ся п олучить н абор из n
файлов (будем в дальнейшем назыв ать такие файлы модулями) таки х, что
имея любые m модулей мы можем восст ановить исходный файл.
mn
1
элементов (под эле-
Разобьём исходный фай л на:
R
НОД
mnn
,(
)1
ментами будем понимать непересекающиеся ч асти исходного файла равного
(почти равного) размера). Таким образом мы п олучаем, что н аш и сходный
файл состоит из R элементов, каждому из которых можн о присвоить поряд -
ковый номер от 1 до R.
Так как и сходный файл можно восстан овить имея любые m модулей, то
каждый из R элементов должен н аходиться в n – m +1 модулей. Тогда хотя
бы в одном из m любых модулей будет содержаться необходимый нам эле -
мент.
Наполнять модули элементами будем следующим образом: построим
mn
1
таблиц у, в которой n строк (количество модулей) и
K
mnn
НОД
)1
,(
столбцов (количество элементов н аходящихся в каждом модуле). Каждый из
R элементов должен встречаться в таблице n – m +1 раз, при этом он должен
находиться в различных строках т аблицы. Можно зап олнять т аблиц у ставя
номера элементов в каждом столбце по n – m +1 раз, н ачиная с первого
столбца сверху вниз и затем переходить к следующему столбц у. И т ак п ока
не зап олнится вся т аблица.
Пример. Разбить файл н а 5 частей таким образом, чтобы его можно
было восст ановить имея любые 3 ч асти.
Так как НОД(5, 5 – 3+1)=1, то R = 5 и K = 3. Построим т аблиц у, как
описано выше.
1 2 4
1:
2:
1 3 4
3:
1 3 5
4:
2 3 5
5: 2 4 5
Т.е. первый модуль состоит из элементов 1, 2, 4; второй модуль состоит
из элементов 1, 3, 4 и т.д.
Понятно, что общее количество ячеек в таблице делится на n и
n – m +1, зн ачит, таблица имеет НОК(n, n – m +1) ячеек. Тогда столбцов в
mnn
НОК
,(
)1
mn
1
таблице K=
элементов
количество
а
,
=
НОД
n
) 1m-n,(
n
mnn
НОК
,(
n
=
.
НОД
n
) 1m-n,(
mn
1
Легко получить формулу н ахождения н омера элемент а в т аблице по его
координат ам:
6
n
i
(
)1
*,
j
iN
,(
)
1
mn
1
где i – номер строки (1≤ i ≤ n), j – номер столбца (1 ≤ j ≤ K).
Набор из n модулей, из m любых которых можно восст ановить исход -
ный файл, будем назыв ать схемой (n, m).
j
* [x] – наибольшее целое число, не превосходящее x.
7
Подобные схемы
Пусть дан а схема (n1, m1), построим схему (n2, m2), в которой n2 = pn1 и
n2 – m2 +1 = p (n1 –m1+1), где p — рациональное , положительное число, такое
что n2 и m2 целые.
Тогда m2 =1 – p + pm1, или
n
n
.
1
2
m
m
1
1
1
2
Схемы (n1, m1) и (n2, m2), для которых выполняется равенство
n
, будем назыв ать п одобными.
2
m
1
2
Для подобных схем (n1, m1) и (n2, m2) выполняют ся следующие равен-
. Следов ательн о,
n = p и
2
n
1
n =
2
n
1
n
1
m
1
или
m
2
m
1
m
2
m
1
1
1
1
1
p
1
ства:
K
1
(
2
2
)1
R
1
))1
НОД
НОД
mn
mn
mnp
1
1
(
)1
1
1
2
2
2
2
mnn
mnn
mnp
pn
(
,
,
(
(
,
1
1
1
2
2
2
n
n
pn
1
2
2
НОД
mnn
mnp
mnn
НОД
pn
НОД
(
)1
(
,
(
))1
(
,
)1
1,1
2
2
1
2
2
2
Значит, у подобных схем используют ся одни и те же элементы и разме -
ры модулей совпадают. Более того, если n1<n2 и схемы (n1, m1) и (n2, m 2) по-
добны, то множество модулей схемы (n1, m 1) являет ся п одмножеством мно-
жеств а модулей схемы (n2, m2).
НОД
)1
K
R
2
2
2
Пример. Схемы (6, 4), (4, 3) и (2, 2) п одобны.
Схему (n, m) будем назыв ать базовой, если для любой подобной схемы
(n1,m1) имеем n≤n1. Очевидно, что схема (n, m) является базовой тогда и толь-
ко тогда, когда НОД(n, m – 1) = 1.
8
Числовые характеристики схем
Для анализа эффективности различных схем введем числовые характе -
ристики схем.
Избыточность – отношение размеров всех n модулей к размеру исход -
ного файла. Обозначим избыточность через
,( mnZ
).
KnmnZ
mn
,(
)
.1
R
Ясн о, что чем меньше избыточность схемы, тем меньше информации
(сумма размеров в сех модулей) придётся рассылать и сохранять.
Введём следующую чи словую характеристику п оказыв ающую н а
сколько маленькие модули по отношению к исходному фай лу. Т.е. отноше -
ние размера одного модуля к размеру и сходного файла. Обозн ачим эту ха-
рактери стику через
.
,( mnml
)
mnml
,(
)
K
R
mn
n
1
1
1
.
m
n
Надёжность – вероятность того, что и сходный файл будет восстанов -
лен. Будем считать, что вероятности получения каждого модуля равны p.
Обозн ачим надёжность через
.
,( mnP
)
n
mi
n
mi
.
где
, т.е.
i
i
)
)
in
in
p
p
1(
1(
pC
i
n
iP
)(
n
pC
i
n
n iP
,)(
mnP
,(
)
mnP
,(
)
Рассчитаем н аименьшее число модулей для восст ановления исходного
файла.
В каждом модуле содержится K элементов. Каждый элемент встречает -
ся в т аблице n – m +1 раз. Значит имея один модуль мы получ аем
элементов и з т аблицы. В т аблице содержится
элемен-
mnK
mnR
(
)1
)1
(
тов. Найдём отношение общего количеств а элементов в таблице к количеству
элементов, которое мы получ аем имея один модуль . Таким образом мы полу-
чим минимальн ое количество модулей. Обозн ачим это значение через
,( mnD
).
)
mnD
,(
mnR
)1
(
mnK
(
)1
Если схемы (n1, m1) и (n2, m2) подобны, то выполняют ся следующие
равенства:
n
mn
R
K
.
1
mnml
,
(
1
1
mnD
(
,
1
1
)
)
mnml
(
,
)
2
2
.
mnD
(
,
)
2
2
,
Используя описанный выше метод мы можем решить следующие зада-
Приложения метода
9
чи.
1
Задача 1. Пусть у нас имеет ся n носителей информации равного объё -
ма. Требуется сохранить на них как можно больше информации и в любой
момент времени восст ановить её, при этом мы знаем, что одновременно мо-
гут выйти из строя не более t носителей.
Самый простой способ – записать одн у и ту же информацию на в се но-
сители. Однако, зн ая что одновременно могут выйти из строя не более t но-
сителей мы можем и спользов ать схему (n, m), где m = n – t, записывая на ка-
ждый носитель по одному модулю.
Подсчитаем на сколько процентов возрастёт объём сохраняемой ин-
формации по сравнению со способом, когда мы сохраняем на каждый носи -
тель одн у и ту же информацию.
KR
Pr1
K
Пример. Пусть у нас имеет ся 3 носителя информации равного объёма.
Одновременно может выйти не более одного н осителя.
Используем схему (3, 3 – 1). Тогда записыв ая н а первый носитель пер-
вый модуль , на второй — второй, а н а третий — третий, мы всегда сможем
восст ановить исходн ую информацию при выходе из строя одного из н осите-
лей.
m
1
mn
%100
%100
.
23
1%100
.
%50%100
Pr1
2
2
Значит, в нашем примере мы можем сохранить на 50% больше инфор-
мации.
Задача 2. Пусть у н ас есть файл, который требует ся сохранить в раз-
личных файловых хранилищах в интернете. Предположим, что из n ресурсов,
на которые мы сохраняли файл, мы всегда можем получить доступ к m ре -
сурсам. Требует ся в любой момент времени иметь доступ к сохранённому
файлу.
Самый п ростой способ — сохранить исходный файл во в сех n файло-
вых хранилищах..
Так как нам в сегда доступны любые m файловых хранилищ, то мы
можем использов ать схему (n, m) и на каждый ресурс сохранять по одному
модулю.
Подсчитаем сколько трафика (количество пересылаемой информации)
мы экономим используя предложенный метод.
m
KR
Pr2
%100
n
R
%100
.
1
Пример. Пусть н ам требует ся сохранить файл в 5-ти файловых хра-
нилищах, при этом в сегда доступны любые 3 из них. Используя схему (5, 3),
в каждое хранилище запи сыв аем по одному модулю. При необходимости
восст ановить файл, мы скачиваем из любых доступных 3–х хранилищ моду-
ли и восстан авлив аем файл.
Экономия трафика составляет
%40%100
.
Pr2
10
13
5
2%100
5
Оптимизация метода
Введем следующую н умерацию модулей.
Первый модуль так и ост аётся первым, вторым модулем назовем n – m
+ 2 модуль, третьим модулем н азовем 2(n – m + 1) +1 и т.д. Общая формула
новой н умерации: f(i) = (n – m +1) (i – 1) +1 mod n (для базовой схемы). Для
произвольной схемы формула будет следующей:
11
(
n
i
(
f
i
)(
,
mn
,(
)1
mod
)1
mn
i
)1
1)1
(
НОД
n
где i =1, 2, … , n; f(i) — новый номер модуля.
Подобная н умерация даёт н ам следующие преимущества:
1) Каждый последующий модуль в сегда состоит из элементов, кото-
рых не было в предыдущем.
2) Имея D(n, m) любых последов ательных модулей, мы всегда восста-
новим исходный файл.
Т.к. R n , то перест авим элементы в модулях т аким образом, чтобы
первые элементы в сех модулей содержали все R элементов. Аналогично на
втором, третьем и т.д. месте каждого модуля, тогда даже если мы получим
только н ачало, середин у или конец каждого модуля, мы сможем восст ано-
вить исходный файл.
Пример. Преобразуем рассмотренный пример схемы (5, 3) по приве-
денным выше п равилам.
Нач альный вари ант После п реобразования
1:
1 2 4
1:
1 2 4
2:
2 3 5
2:
1 3 4
3:
3 4 1
3:
1 3 5
4:
4:
4 5 2
2 3 5
5: 5 1 3
5: 2 4 5
5
5
.
D
)3,5(
2
135
3
Таким образом, для восст ановления исходного файла нам дост аточно
получить любые 2 последов ательных модуля.
12
Литература
1. Adi Shamir, "How to share a secret" // Communicat ions of the ACM ,
22(11), p. 612–613, 1979.
2. Brickell E. F., Davenport D. M. On the classification of Ideal Secret Shar-
ing Schemes. // J. Cryptology. V. 4, 1991. P. 123–134.
3. Блейкли Г. Р., Кабатян ский Г. А. Обобщенные идеальные схемы,
разделяющие секрет, и матроиды // Проблемы передачи информации. Т. 33,
вып. 3, 1997. С. 102–110.
4. Seymour P. O. On Secret–Sharing Matroids. // J. Comb. Theory. Se r. B.
V. 56, 1992. P. 69– 73.
5. Blakley G. R. Safeguarding cryptographic keys // Proc. AFIPS 1979 Na-
tional Computer Confe rence. V. 48. N . Y., 1979. P. 313–317.
6. Мак–Вильямс Ф . Дж., Слоэн Н. Дж. А. Теория кодов, исправляющи х
ошибки. М.: Связь, 1979.
7. Csirmaz L. The size of a share must be large // J. Cryptology. V. 10, No 4,
1997. P. 223–232.
|
1804.02895 | 1 | 1804 | 2018-04-09T10:20:52 | Characterizing Star-PCGs | [
"cs.DS",
"cs.DM",
"math.CO"
] | A graph $G$ is called a pairwise compatibility graph (PCG, for short) if it admits a tuple $(T,w, d_{\min},d_{\max})$ of a tree $T$ whose leaf set is equal to the vertex set of $G$, a non-negative edge weight $w$, and two non-negative reals $d_{\min}\leq d_{\max}$ such that $G$ has an edge between two vertices $u,v\in V$ if and only if the distance between the two leaves $u$ and $v$ in the weighted tree $(T,w)$ is in the interval $[d_{\min}, d_{\max}]$. The tree $T$ is also called a witness tree of the PCG $G$. The problem of testing if a given graph is a PCG is not known to be NP-hard yet. To obtain a complete characterization of PCGs is a wide open problem in computational biology and graph theory. In literature, most witness trees admitted by known PCGs are stars and caterpillars. In this paper, we give a complete characterization for a graph to be a star-PCG (a PCG that admits a star as its witness tree), which provides us the first polynomial-time algorithm for recognizing star-PCGs. | cs.DS | cs |
Characterizing Star-PCGs
Mingyu Xiao1 and Hiroshi Nagamochi2
1 School of Computer Science and Engineering, University of Electronic Science and
2 Department of Applied Mathematics and Physics, Graduate School of Informatics,
Technology of China, China, [email protected]
Kyoto University, Japan, [email protected]
Abstract. A graph G is called a pairwise compatibility graph (PCG,
for short) if it admits a tuple (T, w, dmin, dmax) of a tree T whose leaf set
is equal to the vertex set of G, a non-negative edge weight w, and two
non-negative reals dmin ≤ dmax such that G has an edge between two
vertices u, v ∈ V if and only if the distance between the two leaves u and
v in the weighted tree (T, w) is in the interval [dmin, dmax]. The tree T
is also called a witness tree of the PCG G. The problem of testing if a
given graph is a PCG is not known to be NP-hard yet. To obtain a com-
plete characterization of PCGs is a wide open problem in computational
biology and graph theory. In literature, most witness trees admitted by
known PCGs are stars and caterpillars. In this paper, we give a com-
plete characterization for a graph to be a star-PCG (a PCG that admits
a star as its witness tree), which provides us the first polynomial-time
algorithm for recognizing star-PCGs.
Key words.
rithm; Graph Algorithm; Graph Theory
Pairwise Compatibility Graph; Polynomial-time Algo-
1
Introduction
Pairwise compatibility graph is a graph class originally motivated from com-
putational biology. In biology, the evolutionary history of a set of organisms is
represented by a phylogenetic tree, which is a tree with leaves representing known
taxa and internal nodes representing ancestors that might have led to these taxa
through evolution. Moreover, the edges in the phylogenetic tree may be assigned
weights to represent the evolutionary distance among species. Given a set of taxa
and some relations among the taxa, we may want to construct a phylogenetic
tree of the taxa. The set of taxa may be a subset of taxa from a large phyloge-
netic tree, subject to some biologically-motivated constraints. Kearney, Munro
and Phillips [12] considered the following constraint on sampling based on the
observation in [10]: the pairwise distance between any two leaves in the sample
phylogenetic tree is between two given integers dmin and dmax. This motivates
the introduction of pairwise compatibility graphs (PCGs). Given a phylogenetic
tree T with an edge weight w and two real numbers dmin and dmax, we can
construct a graph G each vertex of which is corresponding to a leaf of T so that
there is an edge between two vertices in G if and only if the corresponding two
leaves of T are at a distance within the interval [dmin, dmax] in T . The graph
G is called the PCG of the tuple (T, w, dmin, dmax). Nowadays, PCG becomes
an interesting graph class and topic in graph theory. Plenty of structural results
have been developed.
It is straightforward to construct a PCG from a given tuple (T, w, dmin, dmax).
However, the inverse direction seems a considerably hard task. Few methods
have been known for constructing a corresponding tuple (T, w, dmin, dmax) from
a given graph G. The inverse problem attracts certain interests in graph algo-
rithms, which may also have potential applications in computational biology.
It has been extensively studied from many aspects after the introduction of
PCG [3,6,7,9,19,18].
A natural question was whether all graphs are PCGs. This was proposed
as a conjecture in [12], and was confuted in [18] by giving a counterexample
of a bipartite graph with with 15 vertices. Later, a counterexample with eight
vertices and a counterexample of a planar graph with 20 vertices were found [9].
It has been checked that all graphs with at most seven vertices are PCGs [3] and
all bipartite graphs with at most eight vertices are PCGs [14]. In fact, it is even
not easy to check whether a graph with a small constant number of vertices is
a PCG or not. Whether recognizing PCGs is NP-hard or not is currently open.
Some references conjecture the NP-hardness of the problem [7,9]. A generalized
version of PCG recognition is shown to be NP-hard [9].
PCG also becomes an interesting graph class in graph theory. It contains the
well-studied graph class of leaf power graphs (LPGs) as a subset of instances such
that dmin = 0, which was introduced in the context of constructing phylogenies
from species similarity data [8,13,15]. Another natural relaxation of PCG is to
set dmax = ∞. This graph class is known as min leaf power graph (mLPG) [6],
which is the complement of LPG. Several other known graph classes have been
shown to be subclasses of PCG, e.g., disjoint union of cliques [2], forests [11],
chordless cycles and single chord cycles [19], tree power graphs [18], threshold
graphs [6], triangle-free outerplanar 3-graphs [16], some particular subclasses of
split matrogenic graphs [6], Dilworth 2 graphs [5], the complement of a forest [11]
and so on. It is also known that a PCG with a witness tree being a caterpillar
also allows a witness tree being a centipede [4]. A method for constructing PCGs
is derived [17], where it is shown that a graph G consisting two graphs G1 and
G2 that share a vertex as a cut-vertex in G is a PCG if and only both G1 and
G2 are PCGs.
How to recognize PCGs or construct a corresponding phylogenetic tree for a
PCG has become an interesting open problem in this area. To make a step toward
this open problem, we consider PCGs with a witness tree being a star in this
paper, which we call star-PCGs. One motivation why we consider stars is that:
in the literature, most of the witness trees of PCGs have simple graph structures,
such as stars and caterpillars [7]. It is also fundamental to consider the problem
of characterizing subclasses of PCGs derived from a specific topology of trees.
Although stars are trees with a rather simple topology, star-PCG recognition is
2
not easy at all. It is known that threshold graphs are star-PCGs (even in star-
LPG and star-mLPG) and the class of star-PCGs is nearly the class of three-
threshold graphs, a graph class extended from the threshold graphs [6]. However,
no complete characterization of star-PCGs and no polynomial-time recognition
of star-PCGs are known. In this paper, we give a complete characterization for a
graph to be a star-PCG, which provides us the first polynomial-time algorithm
for recognizing star-PCGs.
The main idea of our algorithm is as follows. Without loss of generality, we
always rank the leaves of the witness star TV (and the corresponding vertices
in the star-PCG G) according to the weight of the edges incident on it. When
such an ordering of the vertices in a star-PCG G is given, we can see that all the
neighbors of each vertex in G must appear consecutively in the ordering. This
motivates us to define such an ordering to be "consecutive ordering." To check
if a graph is a star-PCG, we can first check if the graph can have a consecutive
ordering of vertices. Consecutive orderings can be computed in polynomial time
by reducing to the problem of recognizing interval graphs. However, this is not
enough to test star-PCGs. A graph may not be a star-PCG even if it has a
consecutive ordering of vertices. We further investigate the structural properties
of star-PCGs on a fixed consecutive ordering of vertices. We find that three cases
of non-adjacent vertex pairs, called gaps, can be used to characterize star-PCGs.
A graph is a star-PCG if and only if it admits a consecutive ordering of vertices
that is gap-free (Theorem 3). Finally, to show that whether a given graph is gap-
free or not can be tested in polynomial time (Theorem 4), we also use a notion
of "contiguous orderings." All these together contribute to a polynomial-time
algorithm for our problem.
The paper is organized as follows. Section 2 introduces some basic notions
and notations necessary to this paper. Section 3 discusses how to test whether
a given family S of subsets of an element set V admits a special ordering on V ,
called "consecutive" or "contiguous" orderings and proves the uniqueness of such
orderings under some conditions on S. This uniqueness plays a key role to prove
that whether a given graph is a star-PCG or not can tested in polynomial time.
Section 4 characterizes the class of star-PCGs G = (V, E) in terms of an ordering
σ of the vertex set V , called a "gap-free" ordering, and shows that given a gap-
free ordering of V , a tuple (T, w, dmin, dmax) that represents G can be computed
in polynomial time. Section 5 first derives structural properties on a graph that
admits a "gap-free" ordering, and then presents a method for testing if a given
graph is a star-PCG or not in polynomial time by using the result on contiguous
orderings to a family of sets. Finally Section 6 makes some concluding remarks.
Due to the space limitation, some proofs are moved to Appendix.
2 Preliminaries
For two integers a and b, let [a, b] denote the set of integers i with a ≤ i ≤ b. For
a sequence σ of elements, let σ denote the reversal of σ. A sequence obtained by
concatenating two sequences σ1 and σ2 in this order is denoted by (σ1, σ2).
3
Families of Sets. Let V be a set of n ≥ 1 elements. We call a subset S ∈ V
trivial in V if S ≤ 1 or S = V . We say that a set X has a common element
with a set Y if X ∩ Y 6= ∅. We say that two subsets X, Y ⊆ V intersect (or X
intersects Y ) if three sets X ∩ Y , X \ Y , and Y \ X are all non-empty sets. A
partition {V1, V2, . . . , Vk} of V is defined to be a collection of disjoint non-empty
subsets Vi of V such that their union is V , where possibly k = 1.
Let S ⊆ 2V be a family of m subsets of V . A total ordering u1, u2, . . . , un of
elements in V is called consecutive to S if each non-empty set S ∈ S consists of
elements with consecutive indices, i.e., S is equal to {ui, ui+1, . . . , ui+S−1} for
some i ∈ [1, n − S − 1]. A consecutive ordering u1, u2, . . . , un of elements in V
to S is called contiguous if any two sets S, S′ ∈ S with S′ ⊆ S start from or end
with the same element along the ordering, i.e., S′ = {uj, uj+1, . . . , uj+S ′−1}
and S = {ui, ui+1, . . . , ui+S−1} satisfy j = i or j + S′ = i + S.
Graphs. Let a graph stand for a simple undirected graph. A graph (resp.,
bipartite graph) with a vertex set V and an edge set E (resp., an edge set E
between two vertex sets V1 and V2 = V \ V1) is denoted by G = (V, E) (resp.,
(V1, V2, E)). Let G be a graph, where V (G) and E(G) denote the sets of vertices
and edges in a graph G, respectively. For a vertex v in G, we denote by NG(v) the
set of neighbors of a vertex v in G, and define degree degG(v) to be the NG(v).
We call a pair of vertices u and v in G a mirror pair if NG(v)\{u} = NG(u)\{v}.
Let X be a subset of V (G). Define NG(X) to be the set of neighbors of X, i.e.,
NG(X) = {u ∈ NG(v) \ X v ∈ X}. Let G − X denote the graph obtained from
G by removing vertices in X together with all edges incident to vertices in X,
where G − {v} for a vertex v may be written as G − v. Let G[X] denote the
graph induced by X, i.e., G[X] = G − (V (G) \ X).
Let T be a tree. A vertex v in T is called an inner vertex if degeT (v) ≥ 2 and
is called a leaf otherwise. Let L(T ) denote the set of leaves. An edge incident to
a leaf in T is called a leaf edge of T . A tree T is called a star if it has at most
one inner vertex.
Weighted Graphs. An edge-weighted graph (G, w) is defined to be a pair of
a graph G and a non-negative weight function w : E(G) → ℜ+. For a subgraph
G′ of G, let w(G′) denote the sum Pe∈E(G′) w(e) of edge weights in G′.
Let (T, w) be an edge-weighted tree. For two vertices u, v ∈ V (T ), let dT,w(u, v)
denote the sum of weights of edges in the unique path of T between u and v.
PCGs. For a tuple (T, w, dmin, dmax) of an edge-weighted tree (T, w) and two
non-negative reals dmin and dmax, define G(T, w, dmin, dmax) to be the simple
graph (L(T ), E) such that, for any two distinct vertices u, v ∈ L(T ), uv ∈ E
if and only if dmin ≤ dT,w(u, v) ≤ dmax. Note that G(T, w, dmin, dmax) is not
necessarily connected.
A graph G is called a pairwise compatibility graph (PCG, for short) if there ex-
ists a tuple (T, w, dmin, dmax) such that G is isomorphic to the graph G(T, dmin, dmax),
where we call such a tuple a pairwise compatibility representation (PCR, for
short) of G, and call a tree T in a PCR of G a pairwise compatibility tree (PCT,
4
for short) of G. The tree T is called a witness tree of G. We call a PCG G a star-
PCG if it admits a PCR (T, w, dmin, dmax) such that T is a star. Fig. 1 illustrates
examples of star-PCGs and PCRs of them. Although phylogenetic trees may not
have edges with weight 0 or degree-2 vertices by some biological motivations [4],
our PCTs do not have these constraints. This relaxation will be helpful for us to
analyze structural properties of PCGs from graph theory. Furthermore, it is easy
to get rid of edges with weight 0 or degree-2 vertices in a tree by contracting an
edge.
Lemma 1. Every PCG admits a PCR (T, w, dmin, dmax) such that 0 < dmin <
dmax and w(e) > 0 for all edges e ∈ E(T ).
v1
v2
v3
v4
(a)
v1
v2
v1
v2
v3
v4
v5
v4
v5
v6
v7
v8
v3
(d)
v8
v7
v6
v5
v*
2
2 3
3
5
5
6
6
v1
v2
v3 v4
dmin=8
v5
v6
dmax=9
v7
v8
(b)
(c)
v*
1
2
3
4
5
v1
v2
dmin=4
v3
v4
dmax=8
v5
(e)
Fig. 1. Illustration of examples of star-PCG. (a) A connected and bipartite star-PCG
G1 = (V1, V2, E), where ordering σa = v1, v2, v3, v4, v8, v7, v6, v5 is not gap-free to G1.
(b) G1 in (a) with a gap-free ordering σb = v1, v2, . . . , v8. (c) A PCR (T, w, dmin =
8, dmax = 9) of G1 in (b). (d) A connected and non-bipartite star-PCG G2. (e) A PCR
(T, w, dmin = 4, dmax = 8) of G2 in (d).
3 Consecutive/Contiguous Orderings of Elements
Let S ⊆ 2V be a family of m subsets of a set V of n ≥ 1 elements in this section.
Let V (S) denote the union of all subsets in S, and π(S) denote the partition
{V1, V2, . . . , Vp} of V (S) such that u, v ∈ Vi for some i if and only if S has no set
S with {u, v} ∩ S = 1. An auxiliary graph HS for S is defined to be the graph
(S, ES ) that joins two sets S, S′ ∈ S with an edge SS′ ∈ ES if and only if S and
S′ intersect.
5
3.1 Consecutive Orderings of Elements
Observe that when S admits a consecutive ordering of V (S), any subfamily
S′ ⊆ S admits a consecutive ordering of V (S′). We call a non-trivial set C ⊆ V
a cut to S if no set S ∈ S intersects C, i.e., each S ∈ S satisfies one of S ⊇ C,
S ⊆ C and S ∩ C = ∅. We call S cut-free if S has no cut.
Theorem 1. For a set V of n ≥ 1 elements and a family S ⊆ 2V of m ≥ 1 sets,
a consecutive ordering of V to S can be found in O(nm2) time, if one exists.
Moreover if S is cut-free, then a consecutive ordering of V to S is unique up to
reversal.
3.2 Contiguous Orderings of Elements
We call two elements u, v ∈ V equivalent in S if no set S ∈ S satisfies {u, v} ∩
S = 1. We call S simple if there is no pair of equivalent elements u, v ∈ V .
Define XS to be the family of maximal sets X ⊆ V such that any two vertices
in X are equivalent and X is maximal subject to this property.
A non-trivial set S ∈ S is called a separator if no other set S′ ∈ S contains
or intersects S, i.e., each S′ ∈ S satisfies S′ ⊆ S or S′ ∩ S = ∅. We call S
separator-free in S if S has no separator.
Theorem 2. For a set V of n ≥ 1 elements and a family S ⊆ 2V of m ≥ 1
sets, a contiguous ordering of V to S can be found in O(nm2) time, if one
exists. Moreover all elements in each set X ∈ XS appear consecutively in any
contiguous ordering of V to S, and a contiguous ordering of V to S is unique
up to reversal of the entire ordering and arbitrariness of orderings of elements
in each set X ∈ XS.
4 Star-PCGs
Let G = (V, E) be a graph with n ≥ 2 vertices, not necessarily connected. Let MG
denote the set of mirror pairs {u, v} ⊆ V in G, i.e., NG(u) \ {v} = NG(v) \ {u},
where u and v are not necessarily adjacent. Let TV be a star with a center v∗ and
L(T ) = V . An ordering of V is defined to be a bijection σ : V → {1, 2, . . . , n},
and we simply write a vertex v with σ(v) = i with vi. For an edge weight
w in TV , we simply denote w(v∗vi) by wi. When G is a star-PCG of a tuple
(TV , w, dmin, dmax), there is an ordering σ of V such that w1 ≤ w2 ≤ · · · ≤ wn.
Conversely this section derives a necessary and sufficient condition for a pair
(G, σ) of a graph G and an ordering σ of V to admit a PCR (TV , w, dmin, dmax)
of G such that w1 ≤ w2 ≤ · · · ≤ wn.
For an ordering σ of V , a non-adjacent vertex pair {vi, vj} with i < j in G is
called a gap (with respect to edges e1, e2 ∈ E) if there are edges e1, e2 ∈ E that
satisfy one of the following:
(g1) e1 = vivj′ and e2 = vivj′′ such that j′ < j < j′′ (or e1 = vi′ vj and
e2 = vi′′ vj such that i′ < i < i′′), as illustrated in Fig. 2(a);
6
v1
vn
v1
e2
vi
e1
vj''
vj
vj'
vj'
vi
e1
e2
vn
vi'
v1
vi'
e1
vj
vi
e1
(a)
(b)
vi'
(c)
vn
vj'
e2
vj
e2
vj'
Fig. 2. Illustration of a gap {vi, vj } in an ordered graph G = (V = {v1, v2, . . . , vn}, E):
(a) e1 = vivj′ and e2 = vivj′′ such that j ′ < j < j ′′, (b) e1 = vivi′ and e2 = vj vj′ such
that j ′ < i and j < i′, (c) e1 = vivi′ and e2 = vj vj′ such that i′ < j and i < j ′, where
possibly j ′ ≤ i′ or i′ < j ′.
(g2) e1 = vivi′ and e2 = vjvj′ such that j′ < i and j < i′, as illustrated
in Fig. 2(b); and
(g3) e1 = vivi′ and e2 = vjvj′ such that i′ < j and i < j′, as illustrated
in Fig. 2(c).
We call an ordering σ of V gap-free in G if it has no gap. Clearly the reversal of a
gap-free ordering of V is also gap-free. We can test if a given ordering is gap-free
or not in O(n4) time by checking the conditions (a)-(c) for each non-adjacent
vertex pair {vi, vj} in G.
Fig. 1(a) and (b) illustrate the same graph G1 with different orderings σa =
u1, u2, . . . , u8 and σb = v1, v2, . . . , v8, where σa is not gap-free while σb is gap-
free.
We have the following result, which implies that a graph G = (V, E) is a
star-PCG if and only if it admits a gap-free ordering of V .
Theorem 3. For a graph G = (V, E), let σ be an ordering of V . Then there is
a PCR (TV , w, dmin, dmax) of G such that w1 ≤ w2 ≤ · · · ≤ wn if and only if σ
is gap-free.
The necessity of this theorem is relatively easy to prove (see Lemma 9 in the
Appendix). Next we consider the sufficiency of Theorem 3, which is implied by
the next lemma.
Lemma 2. For a graph G = (V, E), let σ = v1, v2, . . . , vn be an gap-free order-
ing of V . There is a PCR (TV , w, dmin, dmax) of G such that w1 ≤ w2 ≤ · · · ≤ wn.
Such a set {w1, w2, . . . , wn, dmin, dmax} of weights and bounds can be obtained in
O(n3) time.
Note that when two vertices u and v are not adjacent in a PCG G, there are
two reasons: one is that the distance between them in the PCR (T, w, dmin, dmax)
is smaller than dmin, and the other is that the distance is larger than dmax. Before
we try to assign some value to each wi, we first detect this by coloring edges
in the complete graph KV = (V, E ∪ E) on the vertex set V obtained from a
graph G by adding an edge between each non-adjacent vertex pair in G, where
E = (cid:0)V
2(cid:1) \ E.
7
For a function c : E ∪ E → {red, green, blue}, we call an edge e with
c(e) = red (resp., green and blue) a red (resp., green and blue) edge, and let
Ered (resp., Egreen and Eblue) denote the sets of red (resp., green and blue) edges.
We denote by Nred(v) the set of neighbors of a vertex v via red edges. We define
Ngreen(v) and Nblue(v) analogously.
A coloring of G = (V, E) is defined to be a function c : E∪E → {red, green, blue}
such that Egreen = E. When an ordering σ of V is fixed, we simply write
(i, j) ∈ red (resp., (i, j) ∈ green and (i, j) ∈ blue) if an edge vivj ∈ E ∪ E
is a red (resp., green and blue) edge. For (G, σ) and a coloring c of G, we wish
to determine weights wi, i − 1, 2, . . . , n and bounds dmin and dmax so that the
next holds:
w1 ≤ w2 ≤ · · · ≤ wn;
dmin ≤ wi + wj ≤ dmax for (i, j) ∈ red;
wi + wj < dmin for (i, j) ∈ green; and
wi + wj > dmax for (i, j) ∈ blue.
To have such a set {w1, . . . , wn, dmin, dmax} of values for an ordering σ and a
coloring c of G, the coloring c must satisfy the following conditions:
each vi ∈ V admits integers a(i), b(i) ∈ [1, n] such that
Nred(vi) = {vj 1 ≤ j ≤ a(i) − 1} \ {vi} and Nblue(vi) = {vj b(i) + 1 ≤ j ≤ n} \ {vi},
where a(i) = 1 if Nred(vi) = ∅; b(i) = n if Nblue(vi) = ∅; and Ngreen(vi) =
V \ (Nred(vi) ∪ Nblue(vi) ∪ {vi}) = {vj a(i) ≤ j ≤ b(i)} \ {vi} \ {vi}, and
Ngreen(vi) = ∅ if b(i) < a(i). Such a coloring c of G is called proper to (G, σ).
Lemma 3. For a graph G = (V, E) and a gap-free ordering σ of V , there is a
coloring c of G that is proper to (G, σ), which can be found in in O(n2) time.
Define integers ired and iblue as follows.
ired = (cid:26) the largest index i such that i < a(i) if Ered 6= ∅,
iblue = (cid:26) the smallest index i such that b(i) < i if Eblue 6= ∅,
if Eblue = ∅.
if Ered = ∅,
0
n + 1
In other words, ired 6= 0 is the largest i with (i, i + 1) ∈ red, and ired < n,
whereas iblue 6= n + 1 is the smallest i with (i − 1, i) ∈ blue, and iblue > 1.
Given a graph G, a gap-free ordering σ = v1, v2, . . . , vn of V , and a coloring c
proper to (G, σ), we can find the set {a(i), b(i) i = 1, 2, . . . , n} ∪ {ired, iblue} of
indices in O(n2) time. We also compute the set MG of all mirror pairs in O(n3)
time. Equipped with above results, we can prove the sufficiency of Theorem 3
by designing an O(n)-time algorithm that assigns the right values to weights
w1, w2, . . . , wn in TV . The details can be found in Appendix E.
5 Recognizing Star-PCGS
Based on Theorem 3, we can test whether a graph G = (V, E) is a star-PCG
or not by generating all n! orderings of V . In this section, we show that testing
whether a graph has a gap-free ordering of V can be tested in polynomial time.
8
Theorem 4. Whether a given graph G = (V, E) with n vertices has a gap-free
ordering of V can be tested in O(n6) time.
In a graph G = (V, E), let Et denote the union of edge sets of all cycles
G(v)
of length 3 in G, V t denote the set of end-vertices of edges in Et, and N t
denote the set of neighbors u ∈ NG(v) of a vertex v ∈ V such that uv ∈ Et.
v1
v2
vired-1
vired
vn
vn-1
viblue+1
viblue
u1
u2
u3
u4
u5
V'1
u6
u7
vired+1
viblue-1
(a) G
u19
u18
u17
u16
u15
V'2
u14
u13
u12
u8
u9
u10
u11
(b) G
Fig. 3. (a) A disconnected graph G = (V, E), (b) A connected non-bipartite graph G =
(V, E), where the edges between two vertices in V ∗ = {u7, u8, u9, u10, u11, u12, u13, u14}
are not depicted.
Lemma 4. For a graph G = (V, E) with a gap-free ordering σ = v1, v2, . . . , vn
of V and a coloring c proper to σ, let V1 = {vi 1 ≤ i ≤ ired}, V2 = {vi iblue ≤
i ≤ n}, and V ∗ = {vi ired − 1 ≤ i ≤ iblue + 1}. Then
(i) If two edges vivj and vi′ vj′ with i < j and i′ < j′ cross (i.e., i < i′ < j < j′
or i′ < i < j′ < j), then they belong to the same component of G;
(ii) It holds ired +1 ≤ iblue −1. The graph G[V ∗] is a complete graph, and G−V ∗
is a bipartite graph between vertex sets V1 and V2;
(iii) Every two vertices vi, vj ∈ V1 ∩ NG(V ∗) with i < j satisfy viblue −1 ∈ NG(vi) ∩
V ∗ ⊆ NG(vj ) ∩ V ∗ ⊆ V ∗ \ {vired+1}; and
Every two vertices vi, vj ∈ V2 ∩ NG(V ∗) with i < j satisfy vired+1 ∈ NG(vj) ∩
V ∗ ⊆ NG(vi) ∩ V ∗ ⊆ V ∗ \ {viblue−1}.
We call the complete graph G[V ∗] in Lemma 4(ii) the core of G. Based on the
next lemma, we can treat each component of a disconnected graph G separately
to test whether G is a star-PCG or not.
Lemma 5. Let G = (V, E) be a graph with at least two components.
(i) If G admits a gap-free ordering of V , then each component of G admits a
gap-free ordering of its vertex set, and there is at most one non-bipartite
component in G; and
9
(ii) Let G′ = (V ′
1 , V ′
1, v′
2 , E′) be a bipartite component of G, and G′′ = G − V (G′).
2 and G′′
Assume that G′ admits a gap-free ordering v′
admits a gap-free ordering v1, v2, . . . , vq of V2. Then there is an index k
such that {{v′
2 }. Moreover, the
ordering v′
1, v′
p of V is gap-free to G.
k, v1, v2, . . . , vq, v′
1, v′
2, . . . , v′
k+2, . . . , v′
p}} = {V ′
k+2, . . . , v′
2, . . . , v′
2, . . . , v′
k+1, v′
p of V ′
1 ∪ V ′
k+1, v′
k}, {v′
1 , V ′
Proof. (i) Let G admit a gap-free ordering of V . Any induced subgraph G such
as a component of G is a star-PCG, and a gap-free ordering of its vertex set by
Theorem 3. By Lemma 4(i), at most one component H containing a complete
graph with at least three vertices can be non-bipartite, and the remaining graph
G − V (H) must be a collection of bipartite graphs.
(ii) Immediate from the definition of gap-free orderings.
⊓⊔
We first consider the problem of testing if a given connected bipartite graph is
a star-PCG or not. We reduce this to the problem of finding contiguous ordering
to a family of sets. For a bipartite graph G = (V1, V2, E), define Si to be the
family {NG(v) v ∈ Vj} for the j ∈ {1, 2} − {i, j}, where even if there are
distinct vertices u, v ∈ Vj with NG(u) = NG(v), Si contains exactly one set
S = NG(u) = NG(v).
For the example of a connected bipartite graph G1 = (V1, V2, E) in Fig. 1(a),
we have S1 = {{v3, v4}, {v1, v2, v3, v4}}, and S2 = {{v5, v6}, {v5, v6, v7, v8}}.
Lemma 6. Let G = (V1, V2, E) be a connected bipartite graph with E ≥ 1.
Then family Si is separator-free for each i = 1, 2, and G has a gap-free ordering
of V if and only if for each i = 1, 2, family Si admits a contiguous ordering σi
of Vi. For any contiguous ordering σi of Vi, i = 1, 2, one of orderings (σ1, σ2)
and (σ1, σ2) of V is a gap-free ordering to G.
Note that S1+ S2+ V (S1)+ V (S2) = O(n). By Theorem 2, a contiguous
ordering of V (Si) for each i = 1, 2 can be computed in O(V (Si)Si2) = O(n3)
time.
Fig. 1(a) illustrates an ordering σa = v1, v2, v3, v4, v8, v7, v6, v5 of V (G1) of
a connected bipartite graph G1 = (V1, V2, E), where σa consists of a contiguous
ordering σ1 = v1, v2, v3, v4 of V1 and a contiguous ordering σ2 = v8, v7, v6, v5
of V2. Although σa is not gap-free in G, the other ordering σb of V (G1) that
consists of σ1 and the reversal of σ2 is gap-free, as illustrated in Fig. 1(b).
Finally we consider the case where a given graph G is a connected and non-
bipartite graph. Fig. 1(d) illustrates a connected and non-bipartite star-PCG
whose maximum clique is not unique.
1 , v∗
2 be two adjacent vertices in V t. Let V ∗ = {v∗
Lemma 7. For a connected non-bipartite graph G = (V, E) with V t 6= ∅, and
let v∗
2 )),
1)\V ∗. Assume that G has a gap-free ordering
V ′
1 = NG(v∗
σ of V and a proper coloring c to σ such that v∗
2 = viblue−1. Then:
2 )\V ∗, and V ′
2} ∪ (NG(v∗
2 = NG(v∗
1) ∩ NG(v∗
1, v∗
(i) A maximal clique Kv∗
2 is uniquely given
as G[V ∗]. The graph G[V ∗] is the core of the ordering σ, and G − V ∗ is a
bipartite graph (V1, V2, E′); and
2 of G that contains edge v∗
1 ,v∗
1 = vired+1, v∗
1 , v∗
10
(ii) Let Si denote the family {NG(v) v ∈ Vj } for {i, j} = {1, 2}, and S =
S1 ∪ S2 ∪ {V ∗}. Then S is a separator-free family that admits a contiguous
ordering σ of V , and any contiguous ordering σ of V is a gap-free ordering
to G.
For example, when we choose vertices v∗
2 = u14 in the connected
non-bipartite graph G = (V, E) in Fig. 3(b), we have V ∗ = {u7, u8, u9, u10, u11,
u12, u13, u14}, S1 = {{u1, u2}, {u2, u3, u4}, {u3, u4, u5}, {u3, u4, u5, u6, u7, u8},
{u5, u6, u7, u8, u9, u10, u11}}, and S2 = {{u19}, {u18, u19}, {u16, u17, u18}, {u13,
u14, u15, u16, u17}, {u10, u11, u12, u13, u14, u15, u16}}.
1 = u7 and v∗
For a fixed V ∗ in Lemma 7, we can test whether the separator-free family
S in Lemma 7(ii) is constructed from V ∗ in O(V (S)S2) = O(n3) time by
Theorem 2, since S + V (S) = O(n) holds. It takes O(n4) time to check a
given ordering is gap-free or not. To find the right choice of a vertex pair v∗
1 =
vired+1 and v∗
2 = viblue−1 of some gap-free ordering σ of V , we need to try O(n2)
combinations of vertices to construct V ∗ according to the lemma. Then we can
find a gap-free ordering of a given graph, if one exists in O(n6) time, proving
Theorem 4.
By Theorems 3 and 4, we conclude that whether a given graph with n vertices
is a star-PCG or not can be tested in O(n6) time.
6 Concluding Remarks
Pairwise compatibility graphs were initially introduced from the context of phy-
logenetics in computational biology and later became an interesting graph class
in graph theory. PCG recognition is a hard task and we are still far from a
complete characterization of PCG. Significant progresses toward PCG recogni-
tion would be interesting from a graph theory perspective and also be helpful
in designing sampling algorithms for phylogenetic trees. In this paper, we give
the first polynomial-time algorithm to recognize star-PCGs. Although stars are
trees of a simple topology, it is not an easy task to recognize star-PCGs. For
further study, it is an interesting topic to study the characterization of PCGs
with witness trees of other particular topologies.
References
1. S. Booth and S. Lueker, Testing for the consecutive ones property, interval graphs,
and graph planarity using PQ-tree algorithms, J. Comput. Syst. Sci., 13 (1976)
335–379
2. A. Brandstadt, On leaf powers, Technical report, University of Rostock, 2010.
3. T. Calamoneri, D. Frascaria, and B. Sinaimeri, All graphs with at most seven
vertices are pairwise compatibility graphs, The Computer Journal, 56(7) (2013)
882–886
4. T. Calamoneri, A. Frangioni, and B. Sinaimeri, Pairwise compatibility graphs of
caterpillars, The Computer Journal, 57(11) (2014) 1616–1623
11
5. T. Calamoneri and R. Petreschi, On pairwise compatibility graphs having Dilworth
number two, Theoret. Comput. Sci., 524 (2014) 34–40
6. T. Calamoneri, R. Petreschi, and B. Sinaimeri, On the pairwise compatibility prop-
erty of some superclasses of threshold graphs, Discrete Math. Algorithms Appl.,
5(2) (2013), article 360002
7. T. Calamoneri and B. Sinaimeri, Pairwise compatibility graphs: A survey, SIAM
Review 58(3)(2016) 445–460
8. Z.-Z. Chen, T. Jiang, and G. Lin, Computing phylogenetic roots with bounded
degrees and errors, SIAM J. Comput., 32 (2003) 864–879
9. S. Durocher, D. Mondal, and Md. S. Rahman, On graphs that are not PCGs,
Theoret. Comput. Sci., 571 (2015) 78–87
10. J. Felsenstein, Cases in which parsimony or compatibility methods will be posi-
tively misleading, Systematic Zoology, 27 (1978) 401–410
11. Md. I. Hossain, S. A. Salma, Md. S. Rahman, and D. Mondal, A necessary condition
and a sufficient condition for pairwise compatibility graphs, J. Graph Algorithms
Appl. 21(3) (2017) 341–352
12. P. E. Kearney, J. I. Munro, and D. Phillips, Efficient generation of uniform samples
from phylogenetic trees, in Algorithms in Bioinformatics, G. Benson and R. Page,
eds., LNCS 2812, Springer, Berlin, Heidelberg, 2003, 177–189
13. G. Lin, P. E. Kearney, and T. Jiang, Phylogenetic k-root and Steiner k-root, in
Algorithms and Computation, G. Goos, J. Hartmanis, J. van Leeuwen, D. T. Lee,
and S.-H. Teng, eds., LNCS 1969, Springer, Berlin, Heidelberg, 2000, 539–551
14. S. Mehnaz and Md. S. Rahman, Pairwise compatibility graphs revisited, in Pro-
ceedings of the 2013 International Conference on Informatics, Electronics Vision
(ICIEV), 2013, 1–6
15. N. Nishimura, P. Ragde, and D. M. Thilikos, On graph powers for leaf-labeled
trees, J. Algorithms, 42 (2002) 69–108
16. S. A. Salma, Md. S. Rahman, and Md. I. Hossain, Triangle-free outerplanar 3-
graphs are pairwise compatibility graphs, J. Graph Algorithms Appl., 17 (2013)
81–102
17. M. Xiao
and H. Nagamochi, Some
reduction operations
to pairwise
2017,
1,
compatibility
http://www-or.amp.i.kyoto-u.ac.jp/members/nag/Technical report/TR2017-003.pdf
2017-003, December
graphs, Technical
report
18. M. N. Yanhaona, Md. S. Bayzid, and Md. S. Rahman, Discovering pairwise com-
patibility graphs. Discrete Math., Alg. and Appl. 2(4)(2010) 607–624
19. M. N. Yanhaona, K. S. M. T. Hossain, and Md. S. Rahman, Pairwise compatibility
graphs, in: WALCOM 2008, 2008, 222–233
12
Appendix
A Proof of Lemma 1
Lemma 1 Every PCG admits a PCR (T, w, dmin, dmax) such that 0 < dmin <
dmax and for all edges e ∈ E(T ).
Proof. Since the case where a given PCG G has at most two vertices is trivial,
we assume that G has at least three vertices. Let (T, w′, d′
max) be a PCR
of G, where each path between two leaves in T contains exactly two leaf edges
since L(T ) = V (G) ≥ 3. We increase some values of w′(e), e ∈ E(T ) and d′
min
and d′
max so that the resulting tuple satisfies the lemma. For two positive reals
max, let w(e) := w′(e) + δ for each leaf edge
δ and ε < minuv6∈E dT,w′ (u, v) − d′
e ∈ E(T ), w(e) := w′(e) for all non-leaf edges e ∈ E(T ), dmin := d′
min + 2δ,
and dmax := d′
max + 2δ + ε. We observe that (T, w, dmin, dmax) is a PCR of G
satisfying the lemma because each path between two leaves in T contains exactly
⊓⊔
two leaf edges
min, d′
B Proof of Theorem 1
First we prove the time complexity of the theorem. A graph H with n ≥ 1
vertices is called an interval graph if each vertex v ∈ V (H) can represented by
an ordered pair (av, bv) of reals av ≤ bv so that two vertices u, v ∈ V (H) are
adjacent if and only if there is a real c such that two intervals au ≤ c ≤ bu
and av ≤ c ≤ bv. It is know that testing if a given graph H is an interval
graph and finding such a representation (av, bv), v ∈ V (H), if one exists can
be done in O(V (H) + E(H)) time [1]. Given a family S ⊆ 2V , we see that
S admits a consecutive ordering if and only if the auxiliary graph HS for S is
an interval graph by the definition of HS. The time to construct HS from S is
O(nm2) since we can check in O(n) time whether two sets S, S′ ∈ S intersect,
i.e., vertices S, S′ ∈ V (H) are adjacent in H. Clearly V (H) = S = m and
E(H) ≤ S2 = O(m2). Then testing if HS is an interval graph and finding a
representation (av, bv), v ∈ V (H) can be done in O(n + m2) time. In total, it
takes O(nm2) time to find a consecutive ordering for S, if one exits.
Next we prove that a consecutive ordering of V to a cut-free family S is unique
up to reversal. Assume that S admits another consecutive ordering τ 6∈ {σ, σ}
to derive a contradiction that S would have a cut. Since τ 6∈ {σ, σ}, some two
elements a, b ∈ V appear consecutively in σ, but another element c ∈ V \ {a, b}
appears in the subsequence τab of τ between a and b. Hence a = vk, b = vk+1 and
c = vp for σ = v1, v2, . . . , vn, where k+1 < p is assumed without loss of generality
and c = vp is chosen from τab so that the index p is maximized. We claim that
the set C = {vk+1, vk+2, . . . , vp} is a cut to S. To derive a contradiction, C
intersects a set S ∈ S, where S = {vs, vs+1, . . . , vt} for s ≤ k < k + 1 ≤ t < p
or k + 1 < s ≤ p < t by the consecutiveness of S in σ. In the former, S with
a, b ∈ S contains τab, implying c = vp ∈ S, a contradiction. In the latter, S is
13
contained in τab, since a, b 6∈ S and c ∈ S, implying that vt with p < t belongs
to τab, a contradiction to the choice of c = vp.
C Proof of Theorem 2
To prove Theorem 2, we show that an instance (V, S) of the problem of finding
a contiguous ordering of V to S can be modified to an instance (V, S′) of the
problem of finding a consecutive ordering of V to S′ by introducing some new
subsets of V .
A set S ∈ S is called minimal (resp., maximal) if S contains no proper subset
(resp., superset) of S. Define Smin (resp., Smax) to be the family of minimal (resp.,
maximal) sets in S.
Theorem 2 follows from (iv) and (v) of the next lemma.
Lemma 8. For a set V of n ≥ 1 elements, let S ⊆ 2V \ {∅} be a family with
m ≥ 1 sets.
(i) If some set B ∈ Smax contains at least three sets from Smin, then S cannot
have a contiguous ordering of V ;
(ii) Assume that S is separator-free. Let X be a maximal set of elements any
two of which are equivalent. Then the elements in X appear consecutively in
any contiguous ordering of V to S;
(iii) Assume that S is simple and separator-free. Let S′ denote the family obtained
from S by adding a new set SA,B = B \ A for each pair of sets A ∈ Smin
and B ∈ Smax such that A ( B and B \ A 6∈ S. Then S′ is cut-free and any
consecutive ordering of V to S′ is a contiguous ordering of V to S; and
(iv) Assume that S is separator-free. Then a contiguous ordering of V to S can
be found in O(nm2) time, if one exists. Moreover all elements in each set
X ∈ XS appear consecutively in any contiguous ordering of V to S, and a
contiguous ordering of V to S is unique up to reversal of the entire ordering
and arbitrariness of orderings of elements in each set X ∈ XS; and
(v) A contiguous ordering of V to S can be found in O(nm2) time, if one exists.
Proof. (i) Let three sets Ai ∈ Smin, i = 1, 2, 3 be contained in some set B ∈ Smax.
Note that each Ai is a proper subset of B, and no set Ai is contained in any
other set Aj with j 6= i. Hence in any contiguous ordering of V to S, at least
one of the three sets Ai, i = 1, 2, 3 cannot share the first element in B or the
last element in B. This means that S cannot have a contiguous ordering of V .
(ii) To derive a contradiction, assume that there is a contiguous ordering
u1, u2, . . . , un of V to S wherein the indices of elements X are not consecutive,
i.e., there are elements ui, uj, uk ∈ V with i < j < k and a set S ∈ S such
that {ui, uj, uk} ∩ S = {uj}. Since S is separator-free, there is a set S′ ∈ S that
intersects S or contains S. Since any two elements in X are equivalent, it holds
that either X ⊆ S′ or X ∩ S′ = ∅. If S′ intersects S, then S′ \ S 6= ∅ means
that S′ ⊇ X and S′ would contain S too since the indices of elements in S′ are
14
consecutive in the ordering. Hence S′ always contains S, and it also contains X,
where S′ \ S ⊇ X. Now S does not contain the first element or the last element
of S′ in the ordering. This contradicts that the ordering is contiguous to S.
(iii) For any two sets A ∈ Smin and B ∈ Smax such that A ⊆ B and B \A 6= ∅,
the set B \ A must consist of elements with consecutive indices in a contiguous
ordering of V to S. Hence after adding such a set B \ A to S, the contiguous
ordering of V to S is a consecutive ordering of V to S′. We show that any
consecutive ordering of V to S′ is a contiguous ordering of V to S′. Assume that,
for a consecutive ordering u1, u2, . . . , un of V to S′, there are sets X, Y ∈ S′ such
that X ⊆ Y = {ui, ui+1, . . . , ui+Y −1} but X does not contain any of ui and
ui+Y −1. Let AX ∈ Smin be a set such that AX ⊆ X and let BY ∈ Smax be a set
such that Y ⊆ BY . Note that S′ contains BY \AX , where ui, ui+Y −1 ∈ BY \AX .
However BY \ AX does not consist of elements with consecutive indices in a
the consecutive ordering u1, u2, . . . , un, a contradiction. Hence any consecutive
ordering of V to S′ is a contiguous ordering of V to S′.
We next prove that S′ is cut-free. Let C be a non-trivial subset of V , and
assume that no set S intersects C. Since C ≥ 2 and S is simple, there is a set
S ∈ S such that S ∩ C 6= ∅ 6= C \ S. If S \ C 6= ∅, then S would intersect C.
Hence S \ C. If each set S′ ∈ S with S′ ∩ (V \ C) 6= ∅ is disjoint with C, then
this would contradict that S is separator-free. Hence there is a set S′ ∈ S with
S′ ∩ (V \ C) 6= ∅ and S′ ⊇ C. After the above procedure, the resulting family S′
contains a set S′′ such that S \S′′ 6= ∅ and S′′ ⊆ S′ \S, where C ∩S′′ ⊇ C \S 6= ∅,
C \ S′′ ⊇ S \ S′′ 6= ∅, and S′′ \ C ⊇ S′ ∩ (V \ C) 6= ∅. Therefore S′′ intersects C.
This proves that S′ is cut-free.
(iv) Let S0 ⊆ 2V be a given separator-free family with m ≥ 1 subsets. First
we compute the family XS0 of all maximal subsets. This takes O(nm2) time.
By (ii) of this lemma, all elements in each X ∈ XS0 appear consecutively in
any contiguous ordering of V . Next we construct a simple family S from S0 as
follows. For each set X ∈ XS0, we choose one element vX from X and replace
X with X \ {uX } in the family. Note that the resulting family S is simple and
separator-free and S ≤ S0. If S0 admits a contiguous ordering of V then so
does S. For the family S, we then compute the families Smin and Smax and
construct a bipartite graph B = (Smin, Smax, E∗) such that for each pair of sets
A ∈ Smin, B ∈ Smax, AB ∈ E∗ if and only if A ⊆ B. This also can be done
in O(nm2) time. Now testing whether there is a set B ∈ Smax satisfying the
condition (i) of this lemma can be done in O(m2n) time, because a set B ∈ Smax
contains three sets in Smin if and only if the degree of vertex B ∈ Smax in B is
at least 3. When there exists such a set B ∈ Smax, we conclude that S0 does not
admits any contiguous ordering of V . Assume that no set B ∈ Smax satisfies the
condition (i) of this lemma, and construct from S the cut-free family S′ in (iii) of
this lemma, where we see that S′ ≤ S + 2Smax ≤ 3m. By Theorem 1 applied
to S′, we can find a consecutive ordering σ of V (S′) to the cut-free family S′
in O(m2n) time, if one exists, where σ is unique up to reversal. By (iii) of this
lemma,, the ordering σ is contiguous to S. All contiguous orderings of V to the
input separator-free family S0 can be obtained by replacing the representative
15
element vX in σ for each set X ∈ XS0 with an arbitrary ordering of X. Therefore
a contiguous ordering σ∗ of V to S0 can be found in O(nm2) time, if one exists,
and σ∗ is unique up to reversal and arbitrariness of orderings of elements in each
set X ∈ XS0 .
(v) Given a family S ⊆ 2V , we can test in O(nm) time if a set X ∈ S is a
cut of S, i.e., X does not intersect any other set S ∈ S \ {X}. Then the family
CS of all cuts of S can be found in O(nm2) time, where CS ≤ 2n since CS is a
laminar. Note that any inclusion-wise maximal set in CS is a separator of S. For
each cut X ∈ CS, let C(X) denote the family of sets Y ∈ CS such that Y ( X
and Y is maximal, i.e., no other set S ∈ CS satisfies Y ( S ( X. For each
set X ∈ CS, let S[X] denote the family of sets S ∈ S with S ⊆ X, and ShXi
denote the family obtained from S[X] by contracting each set Y ∈ C(X) into
a single element vY , ignoring all sets S ∈ S with S ⊆ Y , where V (ShXi) =
X − PY ∈C(X)(Y − 1) and ShXi = S[X] − PT ∈C(X) S[Y ]. We easily see
that PX∈CS V (ShXi) ≤ n + CS ≤ 3n, and PX∈CS ShXi ≤ m. Observe
that, for each set X ∈ CS, the family ShXi is separator-free, and S[X] admits
a contiguous ordering of V if and only if ShXi admits a contiguous ordering
of V (ShXi) and S[Y ] for each set Y ∈ C(X) admits a contiguous ordering of
V (S[Y ]). To construct a contiguous ordering of V to S, we choose each set
X ∈ CS in a non-decreasing order of size X, where contiguous orderings σ[Y ]
for families S[Y ] with sets Y ∈ C(X) are available by induction. We then find a
contiguous ordering σhXi for family ShXi in O(V (ShXi)ShXi2) time, if one
exists by (iv) of this lemma, and construct a contiguous ordering σ[X] for S[X]
by replacing the element vY in σhXi with ordering σ[Y ] for each set Y ∈ C(X) in
O(n) time. Since PX∈CS V (ShXi)ShXi2 = O(nm2), we can find a contiguous
ordering of V to S in O(nm2) time, if one exists.
⊓⊔
D The Necessity of Theorem 3
The necessity of Theorem 3 is given by the following lemma.
Lemma 9. For a graph G = (V, E), let σ = v1, v2, . . . , vn be an ordering of V .
If σ is not gap-free, then there is no PCR (TV , w, dmin, dmax) of G such that
w1 ≤ w2 ≤ · · · ≤ wn.
Proof. Let (G, σ) admit a PCR (TV , w, dmin, dmax) with w1 ≤ w2 ≤ · · · ≤ wn.
To derive a contradiction, assume that σ has a gap {vi, vj} with i < j. If it
satisfies the condition (g1) with respect to edges e1 = vivj′ and e2 = vivj′′
such that j′ < j < j′′ (or e1 = vi′ vj and e2 = vi′′ vj such that i′ < i < i′′),
then dmin ≤ wi + wj′ ≤ wi + wj ≤ wi + wj′′ ≤ dmax (or dmin ≤ wi′ + wj ≤
wi + wj ≤ wi′′ + wj ≤ dmax), which implies vivj ∈ E, i.e., vi and vj must
be adjacent in G, a contradiction. If the gap satisfies the condition (g2) with
respect to edges e1 = vivi′ and e2 = vj vj′ such that j′ < i and j < i′, then
dmin ≤ wj′ + wj ≤ wi + wj ≤ wi + wi′ ≤ dmax, again implying that vi and
vj must be adjacent, a contradiction. Analogously with the case where the gap
16
satisfies the condition (g3) with respect to edges e1 = vivi′ and e2 = vjvj′ such
that i′ < j and i < j′, where dmin ≤ wi + wi′ ≤ wi + wj ≤ wj′ + wj ≤ dmax
⊓⊔
would imply that vi and vj are adjacent in G.
E Proof of Lemma 2: The Sufficiency of Theorem 3
For the sufficiency of Theorem 3, we prove Lemma 2 by designing an O(n)-time
algorithm that assigns the right values to weights w1, w2, . . . , wn in TV .
We start with proving Lemma 3.
Lemma 3 For a graph G = (V, E), let σ = v1, v2, . . . , vn be an ordering of V .
For a graph G = (V, E) and a gap-free ordering σ of V , there is a coloring c of
G that is proper to (G, σ), which can be found in in O(n2) time.
Proof. First we assign color green all edges in E. Since (G, σ) has no gap in the
condition (g1), the neighbor set NG(vi) of each vertex vi ∈ V is given by a set
{vai, vai+1, . . . , vbi} \ {vi} of vertices with consecutive indices. Next we assign
color red to all edges vj vi ∈ E with j < ai and color blue to all edges vivk ∈ E
with bi < k. It suffices to show that no edge vivj ∈ E with i < j is assigned two
colors at the same time, in such a way that either
(i) color red from vi and color blue from vj; or
(ii) color blue from vi and color red from vj.
When (i) (resp., (ii)) occurs, there are edges vivi′ ∈ E and vj′ vj ∈ E such that
j′ < i < j < i′ (resp., i′ < j and i < j′), which means that {vi, vj} would be a gap
in the condition (g2) (resp., (g3)), a contradiction. Hence the above procedure
constructs a coloring c proper to (G, σ). We easily see that the procedure can
be implemented to run in O(n2) time.
⊓⊔
By the lemma, we consider the case where a given graph G = (V, E) with
an ordering σ of V admits a coloring c of G proper to (G, σ). By the definition
of indices a(i), b(i), ired and iblue for a coloring c of G, we easily observe the
following property.
Lemma 10. For a graph G = (V, E) with n ≥ 2, an ordering σ = v1, v2, . . . , vn
of V and a coloring c of G proper to (G, σ), the following holds.
(i) Every two indices i and j with 1 ≤ i < j ≤ n satisfy a(j) ≤ a(i) and
b(j) ≤ b(i);
(ii) It holds that ired + 1 ≤ iblue − 1, (i, j) ∈ red for i < j ≤ ired, (i, j) ∈ blue
for iblue ≤ i < j, and (i, j) ∈ green for ired < i < j < iblue.
(iii) Each index p ∈ [1, ired] satisfies p + 2 ≤ ired + 2 ≤ a(ired) ≤ a(p); and
(iv) Each index q ∈ [iblue, n] satisfies b(q) ≤ b(iblue) ≤ iblue − 2 ≤ q − 2.
Proof. (i) Let 1 ≤ i < j ≤ n. If a(i) < a(j), then (i, a(i)) ∈ green while
(j, a(i)) ∈ red, contradicting that coloring c is proper to σ. If b(i) < b(j), then
(b(j), j) ∈ green while (b(j), i) ∈ blue, contradicting that coloring c is proper
to σ.
17
(ii) By definition, ired ≤ n − 1 and iblue ≥ 2. Then if ired = 0 or iblue = n + 1,
then ired + 2 ≤ iblue holds. When Ered, Eblue 6= ∅, the index ired is the largest
i with (i, i + 1) ∈ red, while ired is the smallest i with (i − 1, i) ∈ blue. See
Fig. 5 for an illustration of a star TV . The former implies that (j, ired + 1) ∈ red
for all j ≤ ired. Hence if (i − 1, i) ∈ blue then ired + 1 ≤ i − 1, indicating that
ired + 1 ≤ iblue − 1, as required.
For every ired < i < j, it holds (i, j) 6∈ red by the definition of ired, and
for every i < j < iblue, it holds (i, j) 6∈ blue by the definition of iblue. This
means that (i, j) ∈ red for i < j ≤ ired, (i, j) ∈ blue for iblue ≤ i < j, and
(i, j) ∈ green for ired < i < j < iblue.
(iii) By choice of p, it holds that p + 2 ≤ ired + 2. By this and (i), it holds
that a(ired) ≤ a(p). By definition of ired, it holds that ired + 2 ≤ a(ired).
(iv) Analogous with (iii).
⊓⊔
v*
wired
w1
wired-1
. . .
wired+1
. . .
wiblue-1
wi
. . .
wiblue
wiblue+1
. . .
wn
v1
vired-1
vired
vired+1
vi
viblue-1
viblue
viblue+1
vn
Fig. 4. Illustration of a star TV with weights wi = w(v∗vi) of edge v∗vi ∈ E(T ).
We determine a real value to each weight wi, where the value may be nega-
tive. Recall that we can convert the weights of leaf edges into positive values if
necessary without changing the tree in a PCR.
For some p, q ∈ [1, n] with p < q, suppose that weights wi with p+1 ≤ i ≤ q−1
and bounds dmin and dmax have been determined so that
(a) wp+1 ≤ wp+2 ≤ · · · ≤ wq−1;
(b) wj < wj+1 for {vj, vj+1} 6∈ MG with j ∈ [p + 1, q − 2];
(c) wj + wk > dmin for (j, k) ∈ green with p + 1 ≤ j < k ≤ q − 1,
wj + wk < dmin for (j, k) ∈ red with p + 1 ≤ j < k ≤ q − 1,
wh + wj < dmax for (h, j) ∈ green with p + 1 ≤ h < j ≤ q − 1,
wh + wj > dmax for (h, j) ∈ blue with p + 1 ≤ h < j ≤ q − 1.
Hence for distinct i, i′ ∈ [p + 1, q − 1], it holds dmin > wi + wi′ if (i, i′) ∈ red;
dmin < wi + wi′ < dmax if (i, i′) ∈ green; and wi + wi′ > dmax if (i, i′) ∈ blue.
With the next lemma, we start with weights wi with p + 1 ≤ i ≤ q − 1 and
bounds dmin and dmax for p = ired and q = iblue.
18
Lemma 11. For a graph G = (V, E) with n ≥ 2 and an ordering σ = v1, v2, . . . , vn
of V , let c be a coloring c of G proper to (G, σ). For p = ired and q = iblue, there
is a set {wp+1, wp+2, . . . , wq−1, dmin, dmax} of weights and bounds that satisfies
condition (a)-(c).
Proof. Since (i, j) ∈ green for all ired < i < j < iblue by Lemma 10(i), we
can easily find required weights wi for all i with ired < i < iblue so that dmin <
wi + wi′ < dmax for all i, i′ with ired < i < i′ < iblue. For example, such weights
wi and bounds dmin and dmax can be obtained as follows.
dmin := 0; wired+1 := 1;
for i = ired + 2, ired + 3, . . . , iblue − 1 do
if {vi, vi+1} ∈ MG then
wi+1 := wi
else
wi := wi+1 + 1
end if;
dmax := 2wiblue−1 + 1.
Then without changing the determined values to wp+1, wp+2, . . . , wq−1, we
determine wp or wq so that the above condition holds for (p := p − 1, q) or
(p, q := q + 1).
⊓⊔
v*
v*
wp
wp+1
. . .
wa(p)-1
. . .
α
wa(p)
. . .
wp
wp+1
. . .
p+1
ired
p
a(p)-1
a(p)
q-1
q
p+1
ired
p
(a)
wa(p)-1
. . .
β
δ
wa(p)
. . .
a(p)-1 a(p)
q-1
q
(b)
Fig. 5. Illustration of a process of determining weight wp, where vertex vi is indicated
with its index i: (a) Case (ii) where {vp, vp+1} 6∈ MG and a(p + 1) = a(p) ≤ q − 1; (b)
Case (iii) where {vp, vp+1} 6∈ MG and a(p + 1) < a(p) ≤ q − 1.
We can determine values to weights wired , wired−1, . . . , w1 in this order as
follows.
Lemma 12. For a graph G = (V, E) with n ≥ 2 and an ordering σ = v1, v2, . . . , vn
of V , let c be a coloring c of G proper to (G, σ). For p ≤ ired and q ≥ iblue,
19
assume that a set {wp+1, wp+2, . . . , wq−1, dmin, dmax} of weights and bounds sat-
isfies conditions (a)-(c). When p ≥ 1, let weight wp be determined such that
wp+1
wp+1 − α/2
if {vp, vp+1} ∈ MG,
if {vp, vp+1} 6∈ MG and a(p + 1) = a(p) ≤ q − 1,
where α = wp+1 + wa(p) − dmin (> 0),
wp+1 − β − δ/2
if {vp, vp+1} 6∈ MG and a(p + 1) < a(p) ≤ q − 1,
where β = wp+1 + wa(p)−1 − dmin (> 0) and
δ = wa(p) − wa(p)−1 (> 0),
min{wp+1, dmin−wq−1}−1 otherwise, i.e., {vp, vp+1} 6∈ MG and q ≤ a(p).
wp =
Then the set {wp, wp+1, wp+2, . . . , wq−1, dmin, dmax} of weights and bounds sat-
isfies conditions (a)-(c) for p′ = p − 1 and q.
Proof. We distinguish the four cases.
(i) {vp, vp+1} ∈ MG: Since {vp, vp+1} ∈ MG and wp = wp+1, we see that
for each i ∈ [p + 2, q] with (p, i) ∈ green (resp., (p, i) ∈ red, blue) dmin <
wp + wi = wp+1 + wi < dmax (resp., wp + wi = wp+1 + wi < dmin and wp + wi =
wp+1 + wi > dmax). To show that conditions (a)-(c) hold for (p′ := p − 1, q), it
suffices to show that (p, p + 1) ∈ red and wp + wp+1 < dmin. If p = ired then
ired would be larger than the current p since {vp, vp+1} ∈ MG. Hence p < ired,
which implies that p + 2 ≤ red + 1 and (p, p + 1), (p + 1, p + 2) ∈ red. Since
red + 2 ≤ iblue by Lemma 10(ii), Then wp+1 and wp+2 have been determined
so that wp+1 + wp+2 < dmin since condition (c) holds. Therefore wp + wp+1 ≤
wp+1 + wp+2 < dmin since condition (a) holds.
(ii) {vp, vp+1} 6∈ MG and a(p + 1) = a(p) ≤ q − 1: Fig. 5(a) illustrates the
case where {vp, vp+1} 6∈ MG and a(p + 1) = a(p) ≤ q − 1. Since condition (a)
holds and a(p) ≤ q − 1, it holds that wp+1 ≤ · · · ≤ wa(q)−1 ≤ wa(q) ≤ wq−1.
Since a(p + 1) = a(p), we see that (p, a(p) − 1), (p + 1, a(p) − 1) ∈ red and
(p, a(a)), (p+1, a(p)) ∈ green. Since condition (c) holds for (p+1, a(p)) ∈ green,
we see that α = wp+1 + wa(p) − dmin is positive. To show that conditions (a)-(c)
hold for (p′ := p − 1, q), it suffices to show that wp < wp+1; wp + wa(p) > dmin;
and wp + wa(p)−1 < dmin. Since α > 0, we have wp = wp+1 − α/2 < wp+1.
We next see that wp + wa(p) = (wp+1 − α/2) + wa(p) = dmin + α/2 > dmin, as
required. Finally we observe that wp + wa(p)−1 < wp+1 + wa(p)−1 < dmin since
condition (c) holds for (p + 1, a(p) − 1) ∈ red.
(iii) {vp, vp+1} 6∈ MG and a(p + 1) < a(p) ≤ q − 1: Fig. 5(b) illustrates the
case where {vp, vp+1} 6∈ MG and a(p + 1) < a(p) ≤ q − 1 As in (ii), we see
that wp+1 ≤ · · · ≤ wa(q)−1 ≤ wa(q) ≤ wq−1. Since a(p + 1) < a(p), we see that
(p, a(p)−1) ∈ red and (p+1, a(p)−1) ∈ green. This means that {va(p)−1, va(p)} 6∈
MG, where wa(q)−1 < wa(q) by condition (b) and δ = wa(p) − wa(p)−1 is positive.
Since condition (c) holds for (p + 1, a(p) − 1) ∈ green, β = wp+1 + wa(p)−1 − dmin
is positive. To show that conditions (a)-(c) hold for (p′ := p − 1, q), it suffices
to show that wp < wp+1; wp + wa(p) > dmin; and wp + wa(p)−1 < dmin. Since
β, δ > 0, we have wp = wp+1 − β − δ/2 < wp+1. We next see that wp + wa(p) =
(wp+1 − β − δ/2) + (wa(p)−1 + δ) = dmin + δ/2 > dmin, as required. Finally we
20
observe that wp + wa(p)−1 = (wp+1 − β − δ/2) + wa(p)−1 = dmin − δ/2 < dmin,
as required.
(iv) {vp, vp+1} 6∈ MG and q ≤ a(p): Since q ≤ a(p), (p, i) ∈ red for all i ∈ [p+
1, q−1]. As in (ii), we see that wp+1 ≤ · · · ≤ wq−1. To show that conditions (a)-(c)
hold for (p′ := p−1, q), it suffices to show that wp < wp+1; and wp+wq−1 < dmin.
Obviously wp = min{wp+1, dmin − wq−1} − 1 ≤ wp+1 − 1 < wp+1, as required.
Finally we observe that wp + wq−1 ≤ (dmin − wq−1 − 1)+ wq−1 = dmin − 1 < dmin,
⊓⊔
as required.
Analogously with the process of computing weights wired , wired−1, . . . , w1 by
Lemma 12, we can choose weights wiblue , wiblue+1, . . . , wn in this order as follows.
Lemma 13. For a graph G = (V, E) with n ≥ 2 and an ordering σ = v1, v2, . . . , vn
of V , let c be a coloring c of G proper to (G, σ). For p ≤ ired and q ≥ iblue,
assume that a set {wp+1, wp+2, . . . , wq−1, dmin, dmax} of weights and bounds sat-
isfies conditions (a)-(c). When q ≤ n, let weight wq be determined such that
wq−1
wq−1 + α/2
wq−1 + β + δ/2
if {vq−1, vq} ∈ MG,
if {vq−1, vq} 6∈ MG and p + 1 ≤ b(q) = b(q − 1),
where α = dmax − (wb(p) + wq−1) (> 0),
if {vq−1, vq} 6∈ MG and p + 1 ≤ b(q) < b(q − 1),
where β = dmax − (wb(p)+1 + wq−1) (> 0) and
δ = wb(q)+1 − wb(q) (> 0),
max{wq−1, dmax −wq−1}+1 otherwise, i.e., {vq−1, vq} 6∈ MG and b(q) ≤ p.
wq =
Then the set {wp+1, wp+2, . . . , wq−1, wq, dmin, dmax} of weights and bounds sat-
isfies conditions (a)-(c) for p and q′ := q + 1.
Proof. Analogously with the proof of Lemma 12.
⊓⊔
We are ready to prove Lemma 2. Given a graph G = (V, E) with an gap-
free ordering σ = v1, v2, . . . , vn of V , a coloring c of G proper to (G, σ) can
be obtained in O(n2) time by Lemma 3. For the coloring c, we first comput
MG, indices a(i), b(i), i = 1, 2, . . . , n and ired and iblue for in O(n3) time. Based
on these, we can determine weights wi with ired < i < iblue in O(n) time by
the method in the proof of Lemma 11. Finally we determine weights wi with
1 ≤ i ≤ ired and iblue ≤ i ≤ n in O(n) time by Lemmas 12 and 13, respectively.
This gives a PCR (TV , w, dmin, dmax) of G such that w1 ≤ w2 ≤ · · · ≤ wn in
O(n3) time. This proves Lemma 2.
F Proof of Lemma 4
Lemma 4 For a graph G = (V, E) with a gap-free ordering σ = v1, v2, . . . , vn of
V and a coloring c proper to σ, let V1 = {vi 1 ≤ i ≤ ired}, V2 = {vi iblue ≤
i ≤ n}, and V ∗ = {vi ired − 1 ≤ i ≤ iblue + 1}. Then
21
(i) If two edges vivj and vi′ vj′ with i < j and i′ < j′ cross (i.e., i < i′ < j < j′
or i′ < i < j′ < j), then they belong to the same component of G;
(ii) It holds ired +1 ≤ iblue −1. The graph G[V ∗] is a complete graph, and G−V ∗
is a bipartite graph between vertex sets V1 and V2; and
(iii) Every two vertices vi, vj ∈ V1 ∩ NG(V ∗) with i < j satisfy viblue −1 ∈ NG(vi) ∩
V ∗ ⊆ NG(vj ) ∩ V ∗ ⊆ V ∗ \ {vired+1}; and
Every two vertices vi, vj ∈ V2 ∩ NG(V ∗) with i < j satisfy vired+1 ∈ NG(vj) ∩
V ∗ ⊆ NG(vi) ∩ V ∗ ⊆ V ∗ \ {viblue−1}.
Proof. (i) Let edges e = vivj , e′ = vi′ vj′ ∈ E cross, where i < i′ < j < j′. If
e and e′ belong to different components of G, then {vi′ , vj} would a gap with
respect to edges e and e′.
(ii) Immediate from Lemma 10(ii).
(iii) We show that every two vertices vi, vj ∈ V1 ∩ NG(V ∗) with i < j satisfy
viblue−1 ∈ NG(vi) ∩ V ∗ ⊆ NG(vj) ∩ V ∗ ⊆ V ∗ \ {vired+1} (the other case can be
treated symmetrically). For this, it suffices to show that, for any vertex vi ∈ V1
with NG(vi) ∩ V ∗ 6= ∅,
(a) it holds viblue−1 ∈ NG(vi); and
(b) for any vertex v ∈ NG(vj ) ∩ V ∗ with vj ∈ V1 and i < j,
it holds that v ∈ NG(vi).
In (a), otherwise {vi, viblue−1} would be a gap with respect to edges viv and
vviblue−1 for any vertex v ∈ NG(vi) ∩ V ∗. In (b), otherwise {vj, v} would be a
⊓⊔
gap with respect to edges viv and vjviblue−1.
G Proof of Lemma 6
Lemma 6 Let G = (V1, V2, E) be a connected bipartite graph with E ≥ 1. Then
family Si is separator-free for each i = 1, 2, and G has a gap-free ordering of
V if and only if for each i = 1, 2, family Si admits a contiguous ordering σi of
Vi. For any contiguous ordering σi of Vi, i = 1, 2, one of orderings (σ1, σ2) and
(σ1, σ2) of V is a gap-free ordering to G.
Proof. Since G is connected, we see that, for each i = 1, 2, the family Si is
separation-free.
The only if part: Let v1, v2, . . . , vk, vk+1, . . . , vn be a gap-free ordering of V
to G, where V1 = {v1, v2, . . . , vk} and V2 = {vk+1, . . . , vn}. Since there is no
gap, for each vertex v ∈ V2, the neighbors in NG(v) appear consecutively as
a subsequence of v1, v2, . . . , vk. Also for any two vertices u, v ∈ V2 such that
NG(u) is a proper subset of NG(v), the subsequence vi, vi+1, . . . , vj for NG(u)
must contain the first vertex or the last vertex in the subsequence vh, vh+1, . . . , vp
for NG(v). This is because otherwise vh, vp 6∈ NG(u) would imply that {u, vh}
(or {u, vp}) is a gap with respect to edges e1 = uvi and e1 = vvh (or e1 = uvj
and e1 = vvp). Therefore v1, v2, . . . , vk is a contiguous ordering of V1 to S1.
Analogously with V2 and S2.
22
The if part: Assume that for each i = 1, 2, the family Si has a contiguous
ordering σi of Vi. Note that any set X ∈ XSi, i = 1, 2 is either contained in
NG(v) or disjoint with NG(v) for each vertex v ∈ Vj , j 6= i. By Theorem 2
applied to Si, the vertices in each maximal set X ∈ XSi appear consecutively in
any contiguous ordering of V (Si). Also a contiguous ordering of V (Si) is unique
up to reversal and choice of an ordering of each set X ∈ XSi . This means that
an ordering σ = (σ1, σ2) or (σ1, σ2) of V is gap-free if and only if any ordering
obtained from σ by changing an ordering of vertices in each set X ∈ XS1 ∪ XS2 .
Therefore, to see if G admits a gap-free ordering of V , we only need to check if
⊓⊔
at least one of (σ1, σ2) and (σ1, σ2) is gap-free in G.
H Proof of Lemma 7
2 be two adjacent vertices in V t. Let V ∗ = {v∗
Lemma 7 For a connected non-bipartite graph G = (V, E) with V t 6= ∅, and let
1, v∗
v∗
2 )),
1)\V ∗. Assume that G has a gap-free ordering
V ′
1 = NG(v∗
σ of V and a proper coloring c to σ such that v∗
2 = viblue−1. Then:
2 )\V ∗, and V ′
1 = vired+1, v∗
2} ∪ (NG(v∗
2 = NG(v∗
1, v∗
1 ) ∩ NG(v∗
(i) A maximal clique Kv∗
2 is uniquely given
as G[V ∗]. The graph G[V ∗] is the core of the ordering σ, and G − V ∗ is a
bipartite graph (V1, V2, E′); and
2 of G that contains edge v∗
1 , v∗
1 ,v∗
(ii) Let Si be the family {NG(v) v ∈ Vj} for {i, j} = {1, 2}, and S = S1 ∪ S2 ∪
{V ∗}. Then S is a separator-free family that admits a contiguous ordering
σ of V , and any contiguous ordering σ of V is a gap-free ordering to G.
Proof. (i) By Lemma 5(ii), we see that {vi
ired + 1 < i < iblue − 1} ⊆
NG(vired+1) ∩ NG(viblue −1). On the other hand, by Lemma 5(ii), vertex vired+1
(resp., viblue−1) is not adjacent to any vertex in V2 (resp., V1) in G. Hence
{vired+1, viblue−1}∪(NG(vired+1)∩NG(viblue −1)) induces uniquely a maximal clique
that contains vired+1 and viblue −1. Hence the clique is the core of the gap-free or-
dering of V . By Lemma 4(ii), G − V ∗ is a bipartite graph (V1, V2, E′).
(ii) Since G is connected, we see that Si is separation-free. First we prove
that S admits a contiguous ordering of V . Any set S ∈ S with S ∩ V ∗ 6= ∅
satisfies one of the following:
- vired+1 6∈ S and viblue−1 ∈ S ∈ S2;
- viblue−1 6∈ S and vired+1 ∈ S ∈ S1; and
- vired+1, viblue−1 ∈ S = V ∗.
This means that any two sets S, S′ ∈ S with S ⊆ S′ belong to the same family
Si. Hence any gap-free ordering σ of V to G is a contiguous ordering of V to S,
as discussed in the proof of the only if part of Lemma 6.
Next we prove that any contiguous ordering σ of V to S is a gap-free ordering
of V in G. Since G is connected, each Si is separator-free in V (Si). We see that
S = S1 ∪ S2 ∪ {V ∗} is separator-free in V even if V (S1) ∩ V (S2) 6= ∅. Note that
any set X ∈ XS is either contained in NG(v) or disjoint with NG(v) for each
23
vertex v ∈ Vj, j 6= i. By Theorem 2 applied to S, the vertices in each maximal
set X ∈ XS appear consecutively in any contiguous ordering of V (S). Also a
contiguous ordering of V (S) is unique up to reversal and choice of an ordering
of each set X ∈ XS. This means that an ordering σ of V is gap-free if and only
if any ordering obtained from σ by changing an ordering of vertices in each set
X ∈ XS. Therefore any contiguous ordering σ of V to S is a gap-free ordering
⊓⊔
of V in G.
24
|
1411.6756 | 5 | 1411 | 2015-04-24T08:40:05 | Fast Algorithms for Parameterized Problems with Relaxed Disjointness Constraints | [
"cs.DS"
] | In parameterized complexity, it is a natural idea to consider different generalizations of classic problems. Usually, such generalization are obtained by introducing a "relaxation" variable, where the original problem corresponds to setting this variable to a constant value. For instance, the problem of packing sets of size at most $p$ into a given universe generalizes the Maximum Matching problem, which is recovered by taking $p=2$. Most often, the complexity of the problem increases with the relaxation variable, but very recently Abasi et al. have given a surprising example of a problem --- $r$-Simple $k$-Path --- that can be solved by a randomized algorithm with running time $O^*(2^{O(k \frac{\log r}{r})})$. That is, the complexity of the problem decreases with $r$. In this paper we pursue further the direction sketched by Abasi et al. Our main contribution is a derandomization tool that provides a deterministic counterpart of the main technical result of Abasi et al.: the $O^*(2^{O(k \frac{\log r}{r})})$ algorithm for $(r,k)$-Monomial Detection, which is the problem of finding a monomial of total degree $k$ and individual degrees at most $r$ in a polynomial given as an arithmetic circuit. Our technique works for a large class of circuits, and in particular it can be used to derandomize the result of Abasi et al. for $r$-Simple $k$-Path. On our way to this result we introduce the notion of representative sets for multisets, which may be of independent interest. Finally, we give two more examples of problems that were already studied in the literature, where the same relaxation phenomenon happens. The first one is a natural relaxation of the Set Packing problem, where we allow the packed sets to overlap at each element at most $r$ times. The second one is Degree Bounded Spanning Tree, where we seek for a spanning tree of the graph with a small maximum degree. | cs.DS | cs |
Fast Algorithms for Parameterized Problems with Relaxed
Disjointness Constraints∗
Ariel Gabizon†
Daniel Lokshtanov‡
Micha l Pilipczuk§
June 18, 2018
Abstract
In parameterized complexity, it is a natural idea to consider different generalizations of classic problems.
Usually, such generalization are obtained by introducing a "relaxation" variable, where the original problem
corresponds to setting this variable to a constant value. For instance, the problem of packing sets of size at
most p into a given universe generalizes the Maximum Matching problem, which is recovered by taking
p = 2. Most often, the complexity of the problem increases with the relaxation variable, but very recently
Abasi et al. have given a surprising example of a problem - r-Simple k-Path - that can be solved by a
randomized algorithm with running time O∗(2O(k log r
r )). That is, the complexity of the problem decreases
with r.
In this paper we pursue further the direction sketched by Abasi et al. Our main contribution is a
derandomization tool that provides a deterministic counterpart of the main technical result of Abasi et al.:
the O∗(2O(k log r
r )) algorithm for (r, k)-Monomial Detection, which is the problem of finding a monomial
of total degree k and individual degrees at most r in a polynomial given as an arithmetic circuit. Our
technique works for a large class of circuits, and in particular it can be used to derandomize the result of
Abasi et al. for r-Simple k-Path. On our way to this result we introduce the notion of representative sets
for multisets, which may be of independent interest.
Finally, we give two more examples of problems that were already studied in the literature, where the
same relaxation phenomenon happens. The first one is a natural relaxation of the Set Packing problem,
where we allow the packed sets to overlap at each element at most r times. The second one is Degree
Bounded Spanning Tree, where we seek for a spanning tree of the graph with a small maximum degree.
1
Introduction
Many of the combinatorial optimization problems studied in theoretical computer science are idealized math-
ematical models of real-world problems. When the simplest model is well-understood, it can be enriched to
better capture the real-world problem one actually wants to solve. Thus it comes as no surprise that many
of the well-studied computational problems generalize each other: the Constraint Satisfaction Problem
generalizes Satisfiability, the problem of finding a spanning tree with maximum degree at most d generalizes
Hamiltonian Path, while the problem of packing sets of size 3 generalizes packing sets of size 2, also known
as the Maximum Matching problem.
By definition, the generalized problem is computationally harder than the original. However it is sometimes
the case that the most difficult instances of the generalized problem are actually instances of the original
problem. In other words, the "further away" an instance of the generalized problem is from being an instance of
the original, the easier the instance is. Abasi et. al [1] initiated the study of this phenomenon in parameterized
complexity (we refer the reader to the textbooks [6, 7, 9, 19] for an introduction to parameterized complexity).
In particular, they study the r-Simple k-Path problem. Here the input is a graph G, and integers k and r, and
∗The research leading to these results has received funding from the European Community's Seventh Framework Programme
(FP7/2007-2013) under grant agreement number 257575 and 240258. Mi. Pilipczuk is currently holding a post-doc position
at Warsaw Center of Mathematics and Computer Science and is supported by Polish National Science Centre grant DEC-
2013/11/D/ST6/03073.
†Department of Computer Science, Technion, Israel, [email protected].
‡Department of Informatics, University of Bergen, Norway, [email protected].
§Institute of Informatics, University of Warsaw, Poland, [email protected].
1
the objective is to determine whether there is an r-simple k-path in G, where an r-simple k-path is a sequence
v1, v2, . . . , vk of vertices such that every pair of consecutive vertices is adjacent and no vertex of G is repeated
more than r times in the sequence. Observe that for r = 1 the problem is exactly the problem of finding a
simple path of length k in G. On the other hand, for r = k the problem is easily solvable in polynomial time,
as one just has to look for a walk in G of length k. Thus, gradually increasing r from 1 to k should provide a
sequence of computational problems that become easier as r increases. Abasi et al. [1] confirm this intuition
by giving a randomized algorithm for r-Simple k-Path with running time O(r2k/rnO(1)).
In this paper we continue the investigation of algorithms for problems with a relaxation parameter r that
interpolates between an NP-hard and a polynomial time solvable problem. We show that in several interesting
cases one can get a sequence of algorithms with better and better running times as the relaxation parameter
r increases, essentially providing a smooth transition from the NP hard to the polynomial time solvable case.
Our main technical contribution is a new algorithm for the (r, k)-Monomial Detection problem. Here
the input is an arithmetic circuit C that computes a polynomial f of degree k in n variables x1, . . . , xn. The task
is to determine whether the polynomial f has a monomial Πn
i , such that 0 ≤ ai ≤ r for every i ≤ n. The
main result of Abasi et al. [1] is a randomized algorithm for (r, k)-Monomial Detection with running time
O(r2k/rCnO(1)), and their algorithm for r-Simple k-Path is obtained using a reduction to (r, k)-Monomial
Detection. We give a deterministic algorithm for the problem with running time rO(k/r)CnO(1) in the case
when the circuit C is non-canceling. Formally, this means that the circuit contains only variables at its leaves
(i.e., no constants) and only addition and multiplication gates (i.e, no subtraction gates).
Informally, all
monomials of the polynomials computed at intermediate gates of C contribute to the polynomial computed
by C.
i=1xai
i=1xai
i=1xai
i
is defined as Pn
Comparing our algorithm with the algorithm of Abasi et al. [1], our algorithm is slower by a constant
factor in the exponent of r, and only works for non-cancelling circuits. However our algorithm is deterministic
(while the one by Abasi et al. is randomized) and also works for the weighted variant of the problem, while
the one by Abasi et al. does not. In the weighted variant each variable xi has a non-negative integer weight
wi, and the weight of a monomial Πn
i=1 wiai. The task is to determine whether there
exists a monomial Πn
i , such that 0 ≤ ai ≤ r for every i ≤ n, and if so, to return one of minimum weight.
As a direct consequence we obtain the first deterministic algorithm for r-Simple k-Path with running time
rO(k/r)CnO(1), and the first algorithm with such a running time for weighted r-Simple k-Path.
The significance of an in-depth study of (r, k)-Monomial Detection, is that it is the natural "relaxation
parameter"-based generalization of the fundamental Multi-linear Monomial Detection problem. The
Multi-linear Monomial Detection problem is simply (r, k)-Monomial Detection with r = 1. A
multitude of parameterized problems reduce to Multi-linear Monomial Detection [4, 6, 11, 14, 24]. Thus,
obtaining good algorithms (r, k)-Monomial Detection is an important step towards efficient algorithms for
relaxation parameter-variants of these problems. For some problems, such as k-Path, efficient algorithms
for the relaxation parameter variant (i.e r-Simple k-Path) follow directly from the algorithms for (r, k)-
Monomial Detection. In this paper we give two more examples of fundamental problems for which efficient
algorithms for (r, k)-Monomial Detection lead to efficient algorithms for their "relaxation parameter"-
variant.
Our first example is the (r, p, q)-Packing problem. Here the input is a family F of sets of size q over a
universe of size n, together with integers r and p. The task is to find a subfamily A ⊆ F of size at least p such
that every element of the universe is contained in at most r sets in A. Observe that (r, p, q)-Packing is the
relaxation parameter variant of the classic Set Packing problem ((r, p, q)-Packing with r = 1). We give an
algorithm for (r, p, q)-Packing with running time 2O(pq· log r
r )FnO(1). For r = 1 our algorithm matches the
best known algorithm [4] for Set Packing, up to constants in the exponent, and when r grows our algorithm
is significantly faster than 2pqFnO(1). Just as for r-Simple k-Path, our algorithm also works for weighted
variants of the problem. We remark that (r, p, q)-Packing was also studied by Fernau et al. [8] from the
perspective of kernelization.
Our second example is the Degree-Bounded Spanning Tree problem. Here, we are given as input a
graph G and integer d, and the task is to determine whether G has a spanning tree T whose maximum degree
does not exceed d. For d = 2 this problem is equivalent to Hamiltonian Path, and hence the problem is
NP-complete in general, but for d = n − 1 it boils down to checking the connectedness of G. Thus, Degree-
Bounded Spanning Tree can be thought of as a relaxation parameter variant of Hamiltonian Path. The
problem has received significant attention in the field of approximation algorithms: there are classic results of
Goemans [13] and of Singh and Lau [22] that give additive approximation algorithms for the problem and its
2
weighted variant. From the point of view of exact algorithms, the currently fastest exact algorithm, working
for any value of d, is due to Fomin et al. [10] and has running time O(2n+o(n)).
In this work, we give a
randomized algorithm for Degree-Bounded Spanning Tree with running time 2O(n log d
d ), by reducing the
problem to an instance of (r, k)-Monomial Detection. Thus, our algorithm significantly outperforms the
algorithm of Fomin et al. [10] for all super-constant d, and runs in polynomial time for d = Ω(n). Interestingly,
the instance of (r, k)-Monomial Detection that we create crucially uses subtraction, since the constructed
circuit computes the determinant of some matrix. Thus we are not able to apply our algorithm for non-
cancelling circuits, and have to resort to the randomized algorithm of Abasi et al. [1] instead. Obtaining a
deterministic algorithm for Degree-Bounded Spanning Tree that would match the running time of our
algorithm, or extending the result to the weighted setting, remains as an interesting open problem.
Our methods. The starting point for our algorithms is the notion of representative sets. If A is a family
of sets, with all sets in A having the same size p, we say that a subfamily A′ ⊆ A q-represents A if for every
set B of size q, whenever there exists a set A ∈ A such that A is disjoint from B, then there also exists a set
A′ ∈ A′ such that A′ is disjoint from B.
Representative sets were defined by Monien [16], and were recently used to give efficient parameterized
algorithms for a number of problems [11, 12, 20, 21, 25, 26], including k-Path [11, 12, 21], Set Packing [21, 25,
26] and Multi-linear Monomial Detection [11]. It is therefore very tempting to try to use representative
sets also for the relaxation parameter variants of these problems. However, it looks very hard to directly
use representative sets in this setting. On a superficial level the difficulty lies in that representative sets are
useful to guarantee disjointedness, while the solutions to the relaxation parameter variants of the considered
problems may self-intersect up to r times.
We overcome this difficulty by generalizing the notion of representative sets to multisets. When taking the
union of two multisets A and B, an element that appears a times in A and b times in B will appear a + b times
in the union A + B. Thus, if two regular sets A and B are viewed as multisets, they are disjoint if and only if
no element appears more than once in A + B. We can now relax the notion of disjointedness and require that
no element appears more than r times in A + B. Specifically, if A is a family of multisets, with all multisets in
A having the same size p (counting duplicates), we say that a subfamily A′ ⊆ A q-represents A if the following
condition is satisfied. For every multiset B of size q, whenever there exists an A ∈ A such that no element
appears more than r times in A + B, there also exists an A′ ∈ A′ such that no element appears more than
r times in A′ + B. The majority of the technical effort in the paper is spent on proving that every family
A of multisets has a relatively small q-representative family A′ in this new setting, and to give an efficient
algorithm to compute A′ from A. The formal statement of this result can be found in Corollary 3.8.
On the way to develop our algorithm for computing representative sets of multisets, we give a new con-
struction of a pseudo-random object called lopsided universal sets. Informally speaking, an (n, p, q)-lopsided
universal set is a set of strings such that, when focusing on any k , p + q locations, we see all patterns of
hamming weight p. These objects have been of interest for a while in mathematics and theoretical computer
science under the name Cover Free Families (Cf.
[5]). We give, for the first time, an explicit construction
of an (n, p, q)-lopsided universal set whose size is only polynomially larger than optimal for all p and q. See
Theorem 4.9 in Section 4 for a formal statement. Both our algorithm for computing representative sets of
multisets, and the new construction of lopsided universal sets may be of independent interest.
In Section 2 we give the necessary definitions and set up notational conventions. In
Outline of the paper.
Section 3 we give our construction of representative sets for multisets. This construction requires an auxiliary
tool called minimal separating families. The construction of representative sets for multisets in Section 3
assumes that an appropriate construction of minimal separating families is given as a black box, while the
construction of minimal separating families is deferred to Section 4. Our new construction of lopsided universal
sets is a corollary of the construction of minimal separating families, and is also explained in Section 4. In
Section 5 we use the new construction for representative sets of multisets to give efficient algorithms for (r, k)-
Monomial Detection, (r, p, q)-Packing and r-Simple k-Path. In Section 6 we present our algorithm for
Degree-Bounded Spanning Tree. Finally, we conclude by discussing open problems and directions for
future research in Section 7.
3
2 Preliminaries
Notation. Throughout the paper, we use the notation Ok to hide kO(1) terms. We denote [n] = {1, 2, . . . , n}.
For sets A and B, by { A → B } we denote the set of all functions from A to B. The notation , is used to
introduce new objects defined by formulas on the right hand side.
Hashing families. Recall that, for an integer t ≥ 1, we say that a family of functions H ⊆ { [n] → [m]} is
a t-perfect hash family, if for every C ⊆ [n] of size C = t there is f ∈ H that is injective on T . Alon, Yuster
and Zwick [3] used a construction of Moni Naor (based on ideas from Naor et al. [18]) to hash a subset of size
t into a world of size t2 using a very small set of functions:
Theorem 2.1 ([3] based on Naor). For integers 1 ≤ t ≤ n, a t-perfect hash family H ⊆ { [n] → [t2]} of size
tO(1) · log n can be constructed in time O(tO(1) · n · log n)
We will also use the following perfect hash family given by Naor, Schulman and Srinivasan [18].
Theorem 2.2 ([18]). For integers 1 ≤ t ≤ n, a t-perfect hash family H ⊆ { [k2] → [t]} of size et+O(log2 t)·log k
can be constructed in time O(et+O(log2 t) · k · log k).
Separating families. We will be interested in constructing families of perfect hash functions that, in ad-
dition to being injective on a set C, have the property of sending another large set D to an output disjoint
from the image of C. We call such a family of functions a separating family.
Definition 2.3 (Separating family). Fix integers t, k, s, n such that 1 ≤ t ≤ n. For disjoint subsets C, D ⊆ [n],
we say that a function h : [n] → [s] separates C from D if
• h is injective on C; and
• there are no collisions between C and D. That is, h(C) ∩ h(D) = ∅.
A family of functions H ⊆ { [n] → [s]} is (t, k, s)-separating if for every disjoint subsets C, D ⊆ [n] with
We say that H is (t, k, s)-separating with probability γ if for any fixed C and D with sizes as above, a
C = t and D ≤ k − t, there is a function h ∈ H that separates C from D.
function h chosen uniformly at random from H separates C from D with probability at least γ.
For us, the most important case of separating families is when the range size is C + 1. In this case we use
the term minimal separating family. It will also be convenient to assume in the definition that C is mapped
to the first C elements in the range.
Definition 2.4 (Minimal separating family). A family of functions H ⊆ { [n] → [t + 1]} is (t, k)-minimal
separating if for every disjoint subsets C, D ⊆ [n] with C = t and D ≤ k − t, there is a function h ∈ H such
that
• h(C) = [t].
• h(D) ⊆ {t + 1}.
3 Multiset separators and representative sets
The purpose of this section is to formally define and construct representative sets for multisets. We will use,
as an auxiliary result, an efficient construction of a small separating family.
Theorem 3.1. Fix integers n, t, k such that 1 ≤ t ≤ min(n, k). Then a (t, k)-minimal separating family of
size Ok((k/t)2t · 2O(t) · log n) can be constructed in time Ok((k/t)2t · 2O(t) · n · log n).
We defer the proof of Theorem 3.1 to Section 4, and now we shall use it as a blackbox in order to provide a
construction of representative sets for multisets. The primary tool for this is what we call a multiset separator
(see Definition 3.2). Informally, a multiset separator is a not too large set of 'witnesses' for the fact that two
multisets of bounded size do not jointly contain too many repetitions per element.
4
Notation for multisets. Fix integers integers n, r, k ≥ 1. We use [r]0 to denote {0, . . . , r}. An r-set is a
multiset A where each element of [n] appears at most r times. It will be convenient to think of A as a vector
in [r]n
0 , where Ai denotes the number of times i appears in A. We denote by A the number of elements in
A counting repetitions. That is, A = Pn
i=1 Ai. We refer to A as the size of A. An (r, k)-set is an r-set
A ∈ [r]n
0 , where the number of elements with repetitions is at most k. That is, A ≤ k. For two multisets
A, B over [n],
0 we denote the
"complement" of r-set A, that is, Ai = r − Ai for all i ∈ [n]. By A + B we denote the "union" of A and B,
that is, (A + B)i = Ai + Bi for all i ∈ [n]. Suppose now that A and B are (r, k)-sets. We say that A and B
are (r, k)-compatible if A + B is also an (r, k)-set, and A + B = k. That is, the total number of elements with
repetitions in A and B together is k and any specific element i ∈ [n] appears in A and B together at most r
times. With the notation above at hand, we can define the central object needed for our algorithms.
0 . We say that A ≤ B when Ai ≤ Bi for all i ∈ [n]. By A ∈ [r]n
Fix r-sets A, B ∈ [r]n
0 that are (r, k)-compatible, there exists F ∈ F such that A ≤ F ≤ B.
Definition 3.2 (Multiset separator). Let F be a family of r-sets. We say that F is an (r, k)-separator if for
any (r, k)-sets A, B ∈ [r]n
Construction of multiset separators. The following theorem shows how an (r, k)-separator can be con-
structed from a minimal separating family.
Theorem 3.3. Fix integers n, r, k such that 1 < r ≤ k, and let t , ⌊2k/r⌋. Suppose a (t, k)-minimal
separating family H ⊆ { [n] → [t + 1]} can be constructed in time f (r, k, n). Then an (r, k)-separator F of
size H · (r + 1)t can be constructed in time Ok(f (r, k, max(n, t)) · (r + 1)t).
Proof. In the proof we will assume that n ≥ t, since otherwise we can apply the same construction for n
increased to t, and at the end remove from each multiset of the obtained F all the elements from [t]\ [n]. Note
that thus we apply the construction of a minimal separating family to the set of size max(n, t), rather than
n. Also, observe that from the assumption that r > 1 it follows that t ≤ k.
h : [n] → [t + 1] in H, and for each w = (w1, . . . , wt) ∈ [r]t
all j ∈ [t] and all i ∈ [n] with h(i) = j, we put F h,w
Let F consist of all the constructed r-sets F h,w
Let H be the constructed (t, k)-minimal separating family of functions from [n] to [t + 1]. For each
0 . For
i = r/2.
0, we construct the following r-set F h,w ∈ [r]n
i = wj . For all i ∈ [n] with h(i) = t + 1, we put F h,w
. Thus we have that
i
F = H · (r + 1)t,
Fix (r, k)-compatible (r, k)-sets A, B ∈ [r]n
and F clearly can be constructed in time as claimed in the theorem statement. We are left with proving that
F is indeed an (r, k)-separator.
0 . Let U be the set of all elements that appear in A or in B,
that is, U = {i ∈ [n] Ai > 0 or Bi > 0}. Denote by C0 ⊆ U the sets of elements in A and B that appear
more than r/2 times in one of the sets, that is, C0 = {i ∈ [n] Ai > r/2 or Bi > r/2}. Note that since A
and B are (r, k)-sets, we have that C0 ≤ ⌊2k/r⌋ = t. Let C be a superset of C0 of size exactly t, constructed
by augmenting C0 with arbitrary elements of U \ C0 up to size t, and if there is not enough of them, then
by additionally augmenting it with the remaining number of arbitrary elements of [n] \ U . Note that this is
always possible since t ≤ n. Let D = U \ C. Since A and B are (r, k)-compatible, we have that U ≤ k and
from the construction of C it follows that D ≤ k − t.
Therefore, there exists some h ∈ H that separates C from D. For j ∈ [t], let ij be the unique element of
C mapped to j under h. Choose wj ∈ [r]0 such that Aij ≤ wj ≤ r − Bij ; such wj exists since A and B are
(r, k)-compatible. Let w = (w1, . . . , wt). We claim that A ≤ F h,w ≤ B. For i ∈ C, the choice of w guarantees
that Ai ≤ F h,w
i = ⌊r/2⌋, and from the definition of D we have that D∩ C0 = ∅.
So for such i it holds that Ai ≤ ⌊r/2⌋ ≤ Bi. Finally, for i /∈ C ∪ D we have that Ai = 0 and Bi = r, so surely
Ai ≤ F h,w
i ≤ Bi.
i ≤ Bi. For i ∈ D we have F h,w
By combining Theorems 3.1 and 3.3 we immediately obtain the following construction.
Corollary 3.4. Fix integers n, r, k such that 1 < r ≤ k. Then an (r, k)-separator F of size Ok(r6k/r · 2O(k/r) ·
log n) can be constructed in time Ok(r6k/r · 2O(k/r) · n · log n)
5
Proof. Let t , ⌊2k/r⌋. By Theorem 3.1, a (t, k)-minimal separating family H ⊆ { [n] → [t + 1]} of size
Ok((k/t)2t · 2O(t) · log n) can be constructed in time Ok((k/t)2t · 2O(t) · n · log n) from Theorem 3.1. Using this
construction in Theorem 3.3, we obtain an (r, k)-separator F such that
F = H · (r + 1)t = Ok(r6k/r · 2O(k/r) · log n).
Moreover, from Theorem 3.3 it follows that F can be constructed in time Ok(r6k/r · 2O(k/r) · n · log n).
Multisets over a weighted universe. Before proceeding, we discuss the issue of how the considered
multisets will be equipped with weights. For simplicity, we assume that the universe {1, . . . , n} is weighted,
that is, each element i ∈ {1, . . . , n} is assigned an integer weight w(i). We define the weight of a multiset as
the sum of the weights of its elements counting repetitions. Formally, for A ∈ [r]n
0 we have
w(A) =
n
X
i=1
Ai · w(i).
Whenever we talk about a weighted family of multisets, we mean that the universe [n] is equipped with a
weight function and the weights of the multisets are defined as in the formula above.
Let us remark that the results to follow can be also extended to a more general setting where each multiset
is assigned its own weight that is not directly related to its elements. However, for concreteness we now focus
on the simpler case. We discuss briefly the generalization in Section 5, where we argue that our tools can be
also used to solve the edge-weighted variant of r-Simple k-Path.
Representative sets for multisets. We are ready to define the notion of a representative set for a family
of multisets.
Definition 3.5 (Representative sets for multisets). Let P be a weighted family of (r, k)-sets. We say that
a subfamily P ⊆ P represents P if for every (r, k)-set Q the following holds. If there exists some P ∈ P
of weight w that is (r, k)-compatible with Q, then there also exists some P ′ ∈ P of weight w′ ≤ w that is
(r, k)-compatible with Q.
The following definition and lemma show that having an (r, k)-separator is sufficient for constructing
representative sets.
Definition 3.6. Let P be a weighted family of r-sets and let F be a family of (r, k)-sets. The weighted family
TrimF (P) ⊆ P is defined as follows: For each F ∈ F, and for each 1 ≤ i ≤ k, check if there exists some
P ∈ P with P = i and P ≤ F . If so, insert into TrimF (P) some P ∈ P that is of minimal weight among
those with P = i and P ≤ F .
Lemma 3.7. Let F be an (r, k)-separator and let P be a weighted family of (r, k)-sets. Then TrimF (P)
represents P.
Proof. Fix an (r, k)-set Q and suppose there is an (r, k)-set P ∈ P that is (r, k)-compatible with P .
In
particular, we have P = k − Q. Since F is an (r, k)-separator, there exists F ∈ F such that P ≤ F ≤ Q.
As P ≤ F , when constructing TrimF (P) we must have inserted into it an (r, k)-set P ′ ∈ P of size k − Q
such that P ′ ≤ F and w(P ′) ≤ w(P ). Therefore P ′ ≤ Q, which implies that P ′ + Q is an (r, k)-set. As
P ′ + Q = k, we have that P ′ and Q are (r, k)-compatible.
We can now combine Lemma 3.7 with the construction of an (r, k)-separator from Corollary 3.4, and thus
obtain a construction of a small representative family for a weighted family of multisets.
Corollary 3.8. There exists a deterministic algorithm that, given a weighted family P or (r, k)-sets, runs
in time Ok(P · r6k/r · 2O(k/r) · n log n) and returns returns a family P ⊆ P that represents P and has size
Ok(r6k/r · 2O(k/r) · log n).
Proof. Let F be the (r, k)-separator of size Ok(r6k/r·2O(k/r)·log n) given by Corollary 3.4; recall that F can be
computed in time Ok(r6k/r · 2O(k/r) · n log n). We compute P = TrimF (P) which represents P by Lemma 3.7.
The construction of TrimF (P) amounts to going over all pairs of r-sets P ∈ P and F ∈ F and checking
whether P ≤ F . Thus, the computation takes time at most Ok(P · F· n) = Ok(P · r6k/r · 2O(k/r) · n log n).
6
4 Construction of a separating family
The purpose of this section is to prove Theorem 3.1, that is, to construct a small separating family of functions.
First, we need to introduce some auxiliary results.
Hitting combinatorial rectangles. We first recall the notion of a hitting set for combinatorial rectangles.
For a sequence of integers (m1, m2, . . . , mt), by Qi∈[t][mi] we denote [m1] × [m2] × . . . × [mt].
Definition 4.1. Let R ⊆ Qi∈[t][mi] be a set of the form R1 × . . .× Rt, where Ri ⊆ [mi] for all i ∈ [t]. We say
that R is a combinatorial rectangle with sidewise density γ, if for every i ∈ [t] we have that Ri ≥ γ · mi. A
set H ⊆ Qi∈[t][mi] is a hitting set for rectangles with sidewise density γ if for every set R ⊆ Qi∈[t][mi] that
is a combinatorial rectangle of sidewise density γ, it holds that R ∩ H 6= ∅.
Linial et al. [15] gave the following construction of a hitting set for combinatorial rectangles.
Theorem 4.2 ([15]). A hitting set H ⊆ [m]t for rectangles with sidewise density 1/3 of size H = 2O(t)·mO(1)
can be constructed in time 2O(t) · mO(1).
We need a hitting set for combinatorial rectangles in a universe where the coordinates are from domains
of different sizes. We show that Theorem 4.2 can be adapted to this setting.
Corollary 4.3. Suppose m1, . . . , mt ≤ m. Then a hitting set H ⊆ Qi∈[t][mi] for rectangles with sidewise
density 1/2 of size H = 2O(t) · mO(1) can be constructed in time 2O(t) · mO(1).
Proof. For the purpose of the proof it will be convenient to redefine [a] as {0, . . . , a− 1}. Let m′ = 3m. Define
mapping π : [m′]t → Qi∈[t][mi] as follows:
π(a1, a2, . . . , at) = (a1 mod m1, a2 mod m2, . . . , at mod mt).
Observe that if Ri ⊆ [mi] is such that Ri ≥ mi/2, and R′
i ⊆ [m′] is the set of all elements of [m′] whose
i ≥ m′/3. This follows from the fact that m′ = 3m ≥ 3mi.
remainders modulo mi belong to Ri, then R′
Therefore, if R ⊆ Qi∈[t][mi] is a combinatorial rectangle with sidewise density 1/2, then R′ , π−1(R) ⊆ [m′]t
is a combinatorial rectangle with sidewise density 1/3. Let H ′ ⊆ [m′]t be the hitting set for combinatorial
rectangles with sidewise density 1/3 given by Theorem 4.2. Moreover, let H , π(H ′); hence we have that
H ≤ H ′ = 2O(t) · mO(1). We have that H ′ ∩ R′ 6= ∅, so also H ∩ R = π(H ′) ∩ π(R′) ⊇ π(H ′ ∩ R′) 6= ∅.
Since R was chosen arbitrarily, we infer that H is a hitting set for combinatorial rectangles with sidewise
density 1/2.
Pairwise independent families. Another component in our construction is a family of ǫ-pairwise inde-
pendent functions.
Definition 4.4. A family of functions H ⊆ { [n] → [m] } is ǫ-pairwise independent if for all x, y ∈ [n] and
a, b ∈ [m], it holds that
Pf ←H(f (x) = a ∧ f (y) = b) − 1/m2 ≤ ǫ.
The constructions of [17] and [2] imply the following construction of an ǫ-pairwise independent family of
functions (cf. [3, Section 4]).
Theorem 4.5 ([2, 17]). Fix any m ≤ n. Then a 1/m2-pairwise independent family H ⊆ { [n] → [m]} of size
mO(1) · log n can be constructed in time O(mO(1) · n · log n).
We now use Theorem 4.5 to construct a small family that separates one element from k other elements
with non-negligible probability.
Lemma 4.6. Fix integers k, n with 1 ≤ k ≤ n. Then a family Hk,4k ⊆ { [n] → [4 · k]} that is (1, k, 4 · k)-
separating with probability 1/2 and has size kO(1) · log n can be constructed in time kO(1) · n · log n.
7
Proof. Let H ⊆ { [n] → [4k]} be the family of 1/(4k)2-pairwise independent functions, given by Theorem 4.5.
Fix sets C = {a} and D = {b1, . . . , bk−1} we want to separate. For j ∈ [k−1], let Xj be a random variable that
is equal to one if f (a) = f (bj) and to zero otherwise, where f is chosen uniformly at random from H. From
the 1/4k-pairwise independence of H, we have that E(Xj ) = P(Xj = 1) ≤ 4k · (1/(4k)2 + 1/(4k)2) ≤ 1/(2k).
Let us define X = Pk−1
j=1 Xj, so that we have E(X) ≤ (k− 1)/(2k) < 1/2. Note that X is precisely the number
of collisions between C and D. From Markov's inequality, we have that P(X ≥ 1) ≤ 1/2. So with probability
at least 1/2 over the choice of f ∈ H, C and D are separated by f .
The next claim follows from the well-known theorem that the geometric mean of non-negative numbers is
never larger than their arithmetic mean.
Proposition 4.7. Let k1, . . . , kt be non-negative real numbers such that k1 + . . . + kt ≤ k. Then Qt
(k/t)t.
i=1 ki ≤
The main construction. We are ready to proceed to the main construction.
Proof of Theorem 3.1. Fix disjoint subsets C, D ⊆ [n] with C = t and D ≤ k − t. Recall that we want to
construct a family of functions H ⊆ { [n] → [t + 1]}, such that for any C and D as above, there exists h ∈ H
that separates C from D. It will be convenient to present the family by constructing h adaptively given C
and D. That is, for arbitrarily chosen C and D, we will adaptively construct a function h that separates C
from D. Function h will be constructed by taking a number of choices, where each choice is taken among a
number of possibilities. The final family H will comprise all h that can be obtained using any such sequence
of choices; thus, the product of the numbers of possibilities will limit the size of H. As C and D are taken
arbitrarily, it immediately follows that such H separates every pair (C, D).
1. Let H0 ⊆ { [n] → [k2]} be the k-perfect hash family given by Theorem 2.1. Choose f0 ∈ H0 that is
injective on C ∪ D - there are kO(1) · log n choices for this stage.
From now on, we identify C and D with their images in [k2] under f0.
2. Let H1 ⊆ { [k2] → [t]} be the t-perfect hash family given by Theorem 2.2. Choose a function f1 ∈ H1
that is injective on C - there are et+O(log2 t) · log k choices for this stage.
For i ∈ [t], we denote from now on by ci the (unique) element of C with f1(ci) = i. Moreover, elements
mapped to i under f1 will be denoted by Bi, and we will think of them as the i-th bucket.
3. Choose non-negative integers k1, . . . , kt such that ki is the number of elements a ∈ D with f1(a) = i.
Note that k1 + . . . + kt ≤ k − t, so the number of choices for this stage is at most (cid:0)k
t(cid:1) ≤ (ek/t)t.
i
· log k.
4. For i ∈ [t], let Hki,4ki ⊆ { [k2] → [4 · ki]} be family given by Lemma 4.6. That is, Hki,4ki is (1, ki, 4· ki)-
separating with probability 1/2 and has size mi = kO(1)
Let R be the set of all vectors (h1, . . . , ht) such that for all i ∈ [t], hi is an element of Hki,4ki that
separates ci from Di. Identify [mi] with Hki,4ki by ordering the functions in Hki,4ki arbitrarily. Observe
that R is a combinatorial rectangle of sidewise density 1/2 in Qi∈[t][mi]. Note that for all i ∈ [t], mi ≤ m
for some m = kO(1). Let H be the hitting set for combinatorial rectangles with sidewise density 1/2
given by Corollary 4.3. Choose an element (h1, . . . , ht) ∈ H ∩ R. As H = 2O(t) · kO(1), there are
2O(t) · kO(1) choices for this stage.
For i ∈ [t], apply hi to partition bucket Bi into buckets Bi,1, . . . ,Bi,4ki; that is, element a ∈ [n] is put
into bucket Bf1(a),hf1(a)(a). Since (h1, . . . , ht) ∈ R, we thus have that each ci is mapped to a separate
bucket from all elements of Di.
5. For each i ∈ [t], choose the unique index ji ∈ [4ki] such that ci was mapped to bucket Bi,ji . Construct
the final function h by mapping all the elements of Bi,ji to i, and mapping all the elements of Bi,j ′ for
j ′ 6= ji to t + 1, for all i ∈ [t]. The number of choices for this stage is 4t · Qt
i=1 ki ≤ (4k/t)t, where the
inequality follows from Proposition 4.7.
8
To sum up, the number of different functions enumerated in all the above stages is at most
kO(1) · log n · et+O(log2 t) · log k · (ek/t)t · 2O(t) · kO(1) · (4k/t)t = (k/t)2t · 2O(t) · kO(1) · log n
= Ok((k/t)2t · 2O(t) · log n).
Similarly, going over the construction times of the different components used in the above stages, we get that
the family can be constructed in time
O(kO(1) · n · log n · et+O(log2 t) · k · log k · (ek/t)t · 2O(t) · kO(1) · (4k/t)t)
= O((k/t)2t · 2O(t) · kO(1) · n · log n) = Ok((k/t)2t · 2O(t) · n · log n).
Informally speaking, an (n, p, q)-lopsided universal set is a set of strings such that,
Lopsided universal sets.
when focusing on any k , p + q locations, we see all patterns of hamming weight p. Universal set families, and
in particular the lopsided ones, have important applications in the determinization of parameterized algorithms
that use the technique of color coding; cf. [12].
Definition 4.8. An (n, p, q)-lopsided universal set is a family of subsets F ⊆ [n] such that for every disjoint
subsets A, B ⊆ [n] with A = p and B = q there exists F ∈ F with A ⊆ F and F ∩ B = ∅.
p(cid:1)· √p · q · k · log n
A probabilistic argument shows the existence of (n, p, q)-lopsided universal sets of size (cid:0)k
(cf. Lemma 52 in [5]). Fomin, Lokshtanov and Saurabh [12] used the technique of [18] to construct an (n, p, q)-
log log(k)(cid:1) · log n. Consider the case p = o(q): For simplicity of discussion,
lopsided universal set of size (cid:0)k
let us focus on the (cid:0)k
p(cid:1) ≤ (e· k/p)p term in the probablistic construction and omit log n terms. In this case the
2O(cid:0)
p(cid:1). Bshouty [5] gives a better construction for such p achieving
size roughly kp+2 (see Lemma 53 in [5]). We remark that Bshouty [5] referred to this object as a cover free
family. As a corollary of the last section we obtain an explicit construction of size (k/p)O(p).
log log(k)(cid:1) term of [12] is much larger than (cid:0)k
k
k
p(cid:1) · 2O(cid:0)
Theorem 4.9. Fix integers p, q and n such that k = p + q ≤ n. Then an (n, p, q)-lopsided universal set of
size Ok((k/p)2·p · 2O(p) · log n) can be constructed in time Ok((k/p)2·p · 2O(p) · n · log n).
Proof. Let H ⊆ { [n] → [p + 1]} be the (p, k)-minimal separating family of size Ok((k/p)2p · 2O(p) · log n) given
by Theorem 3.1. For each h ∈ H, construct set Fh = h−1([p]), and let F , {Fh h ∈ H}. The definition
of a minimal separating family immediately implies that F is an (n, p, q)-lopsided universal set, and the time
needed for the construction follows from Theorem 3.1.
5 Algorithmic applications
In this section we present applications of our construction of representative sets for multisets to the design of
deterministic algorithms for problems with relaxed disjointness constraints. We first give some useful auxiliary
tools, and then present algorithms for three problems: r-Simple k-Path, (r, p, q)-Packing, (r, k)-Monomial
Detection. Let us remark that the algorithms for the first two problems follow from the algorithm for the
(more general) last problem. Nevertheless, we find it more didactic to present the applications in increasing
order of difficulty. Throughout this section, r always denotes an integer in the range 1 < r < k.
5.1 Algorithmic preliminaries
The following simple observation will be used in all of our algorithmic applications.
Lemma 5.1. Let P be a family of (r, k)-sets that constains a multiset P of size k and weight w. Then any
P ⊆ P that represents P contains a multiset P ′ of size k and weight w′ ≤ w.
Proof. Take B = ∅. Then P and B are (r, k)-compatible. So there must exist P ′ ∈ P that is (r, k)-compatible
with B and has weight at most w. As P ′ + B = k, we have P ′ = k.
9
The following simple fact is immediate and will be used implicitly.
Proposition 5.2. If A′′ represents A′ and A′ represents A, then A′′ represents A.
We now show that representative sets behave robustly under union of families.
Lemma 5.3. Suppose A and B are two weighted families of (r,k)−sets, and suppose further that A represents
A and B represents B. Then A ∪ B represents A ∪ B.
Proof. Take any (r, k)-set D such that there exists C ∈ A∪B that is (r, k)-compatible with D. If C ∈ A, then
there exists C ′ ∈ A with w(C ′) ≤ w(C) such that C ′ is (r, k)-compatible with D. Then also C ′ ∈ A ∪ B and
we are done. The reasoning for the case when C ∈ B is symmetric.
Finally, we introduce operation • that extends the considered weighted family of (r, k)-sets. This is the
crucial ingredient in the classic technique of iteratively extending the current family by a new object, and then
shrinking the size of the family by taking a representative subfamily.
Definition 5.4. We say that (r, k)-sets A and B are (r, k)-consistent, if A + B ≤ k and Ai + Bi ≤ r for all
i ∈ [n]. (Note that this is a weaker definition than being (r, k)-compatible, where it is required that A+B = k.)
Let A and B be weighted families of (r, k)-sets. We define the weighted family A • B as follows:
A • B , {A + B A ∈ A, B ∈ B, A and B are (r, k)-consistent}.
In particular, for an element i ∈ [n], we denote
A • i , {A + {i} A ∈ A,A < k, Ai < r}.
Lemma 5.5. Suppose A and B are two weighted families of (r,k)−sets, and suppose further that A represents
A and B represents B. Then A • B represents A • B.
Proof. Fix an (r, k)-set D such that there exists C ∈ A • B that is (r, k)-compatible with D. Hence, we have
that C = A + B for some A ∈ A and B ∈ B that are (r, k)-consistent. In particular, A is (r, k)-compatible with
B + D. Hence, there exists A′ ∈ A with w(A′) ≤ w(A) that is (r, k)-compatible with B + D. In other words,
A′ + B + D is an (r, k)-set of size A′ + B + D = k. It follows that B is (r, k)-compatible with A′ + D. Hence,
there exists B′ ∈ B with w(B′) ≤ w(B) that is (r, k)-compatible with A + D. Again, for this B′ we have that
A′ + B′ + D is an (r, k)-set of size k. Therefore, A′ + B′ is an (r, k)-set of size at most k and weight not larger
than the weight of A + B. In other words, A′ and B′ are (r, k)-consistent, and therefore A′ + B′ ∈ A′ • B′. As
A′ + B′ is (r, k)-compatible with D and w(A′ + B′) ≤ w(A + B), we are done.
5.2 r-Simple k-Path
We first introduce some notation for defining the r-Simple k-Path problem. Let G be a directed graph on n
vertices. A k-path in G is a simple walk P = v1 → v2 → . . . → vk in G, that is, all vertices traversed by the
walk are pairwise different An r-simple k-path is a walk P = v1 → v2 → . . . → vk in G such that no vertex
is repeated more than r times. In case the vertex set of G is equipped with a weight function, the weight of
an r-simple k-path P is defined as the weight of the multiset of vertices traversed by P , where each vertex is
taken the number of times equal to the number of its visits on P .
r-Simple k-Path
Input: Directed graph G = (V, E), weight function w : V → R, integers r and k.
Parameters: r, k
Question: Determine whether there exists an r-simple k-path in G, and if so then return one of minimal
possible weight.
Theorem 5.6. r-Simple k-Path can be solved in deterministic time Ok(r12k/r · 2O(k/r) · n3 · log n).
Proof. Identify the vertex set V with [n]. For u ∈ [n] and i ∈ [k], let Pi,u denote the set of r-simple paths in
G that have length i and end in u. Note that an element P of such a set Pi,u can be viewed as an (r, k)-set
when we ignore the order of vertices in the path. We can thus view the sets Pi,u as families of (r, k)-sets.
10
We will compute, using a dynamic programming algorithm, for each u ∈ [n] and i ∈ [k], a set Pi,u ⊆ Pi,u
that represents Pi,u and has size at most Ok(r6k/r · 2O(k/r) · log n). By Lemma 5.1, a minimal-weight r-simple
k-path in G can be recovered by taking a minimum-weight multiset among families Pk,u, for u ∈ [n].
In
particular, if all these sets are empty, then no r-simple k-path exists in G. Thus, such an algorithm can be
used to solve r-Simple k-Path.
We proceed with the algorithm description. For i = 1 and u ∈ [n], family P1,u simply contains only
the length-one path u. Assume that we have computed families Pi,u for every u ∈ [n]. We describe the
computation of Pi+1,v for a fixed v ∈ [n], together with its running time.
1. We compute P ′ , S(u,v)∈E
Pi,u • v. Clearly, we have that P ′ ≤ n · Ok(r6k/r · 2O(k/r) · log n). We
now claim that P ′ represents Pi+1,v. Note first that Pi+1,v = S(u,v)∈E Pi,u • v. By Lemma 5.5, for any
Pi,u•v represents S(u,v)∈E Pi,u•v = Pi+1,v.
u ∈ [n], Pi,u•v represents Pi,u•v. Therefore, P ′ = S(u,v)∈E
Computing P ′ directly from the definition requires time Ok(Pu∈[n] Pi,u) = Ok(n· r6k/r · 2O(k/r) · log n).
2. Now, using the algorithm of Corollary 3.8, compute a set Pi+1,v that represents P ′ and has size Pi+1,v =
Ok(r6k/r · 2O(k/r) · log n). This takes time Ok(P ′· r6k/r · 2O(k/r)· n log n) = Ok(n2· r12k/r · 2O(k/r)· log n).
Since during each of k steps of the algorithm, this procedure is applied to every vertex v ∈ [n], we conclude
that the running time of the algorithm is Ok(r12k/r · 2O(k/r) · n3 · log n), as claimed.
We have considered the r-Simple k-Path problem where the weights are on the vertices. One could also
consider the edge weighted variant where every edge has a weight and the weight of a walk is the sum of the
weight of the edges on the walk (counting multiplicities of edges). One can reduce this variant to the vertex
weighted variant as follows. Let G be the input graph, we make a new graph G′ from G. Here each edge uv of
G is replaced by a new vertex xuv with u being the only in-neighbor and v being the only out-neighbor of xuv.
The weight of xuv is set to the weight of the edge uv. The weight of the vertices of G′ that correspond to the
vertices of G is set to 0. An r-simple k-path in G corresponds to an r-simple (2k + 1)-path in G′ of the same
weight. On the other hand, a minimum weight r-simple (2k + 1)-path in G′ must start and end in vertices
corresponding to vertices of G, since these have weight 0. Such r-simple (2k + 1)-paths in G′ correspond to
an r-simple k-path in G of the same weight. Thus we can run the algorithm for Theorem 5.6 on G′, obtaining
an algorithm for the edge weighted variant with running time rO(k/r)nO(1).
5.3 (r, p, q)-Packing
We say that P ⊆ [n] is a q-set if P = q. We say that a family of q-sets A = {P1, . . . , Pt} is an r-packing if
every element j ∈ [n] appears in at most r of the q-sets in A. The (r, p, q)-Packing problem asks whether a
given family of q-sets contains an r-packing. Again, in case universe [n] is equipped with a weight function,
we can say that a family of q-sets is weighted by setting the weight of a q-set to be the total weight of its
elements. The weight of an r-packing is defined as the sum of weights of its sets.
(r, p, q)-Packing
Input: Weighted family F of q-sets, integers r, p, q
Parameters: r, p, q
Question: Determine whether there exists a subfamily A ⊆ F with A = p that is an r-packing, and if
so then return such A of minimal total weight.
Theorem 5.7. Let k , p · q. Then (r, p, q)-Packing can be solved in deterministic time Ok(F · r12k/r ·
2O(k/r) · n log2 n).
Proof. Let F be the input family of q-sets. For 0 ≤ i ≤ p, denote by Pi the set of r-packings of size i in F .
Note that an element A = {P1, . . . , Pi} of Pi can be viewed as an (r, k)-set: We think of A as the multiset
A = P1 + . . . + Pi. We know that this multiset has size at most p · q = k and every element j ∈ [n] appears
in A at most r times. We will compute, for each 0 ≤ i ≤ p, a subfamily Pi ⊆ Pi that represents Pi and has
size at most Ok(r6k/r · 2O(k/r) · log n). By Lemma 5.1, a minimum-weight r-packing in F can be recovered by
11
taking a minimum-weight element of Pp. In particular, if Pp is empty, then no r-packing exists in F . Thus,
such an algorithm can be used to solve (r, p, q)-Packing.
We proceed with the algorithm description. For i = 0, we take P0 = P0 = {∅}. Assume we have computed
Pi for some 0 ≤ i < p. We describe the computation of Pi+1.
1. We compute P ′ , Pi•F . Clearly, we have that P ′ ≤ F· Pi ≤ F·Ok(r6k/r ·2O(k/r)·log n). We claim
that P ′ represents Pi+1. Note first that Pi+1 = Pi • F . Hence, by Lemma 5.5, P ′ = Pi • F represents
Pi • F = Pi+1.
Computing P ′ directly from the definition requires time Ok(F· Pi·n) = Ok(F·r6k/r·2O(k/r)·n log n).
2. Now, using the algorithm of Corollary 3.8, compute a set Pi+1 that represents P ′ and has size Pi+1 =
Ok(r6k/r·2O(k/r)·log n). This takes time Ok(P ′·r6k/r·2O(k/r)·n log n) = Ok(F·r12k/r·2O(k/r)·n log2 n).
Since we repeat this procedure p times, the claimed running time follows
In Theorem 5.7 we gave an algorithm for (r, p, q)-Packing where every element has a weight and the weight
of a set is equal to the sum of the weights of the elements. A related variant is one where a weight function
w : F → N is given as input. That is, every set P ∈ F has its own weight w(P ), and the weight of a subfamily
A ⊆ F is the sum of the weights of the sets in A. We will call this the set weighted (r, p, q)-Packing problem.
The set weighted problem can be reduced to the original by adding for every set P ∈ F a new element eP of
weight w(P ), inserting eP into P and giving eP weight w(P ). Note that P is the only set in the new instance
that contains eP . All elements corresponding to the elements of the instance of set weighted (r, p, q)-Packing
are given weight 0. All sets in the new instance have size q + 1, and r-packings in the new instance correspond
to r-packings in the original instance with the same weight. Thus we can apply the algorithm of Theorem 5.7
on the new instance in order to solve the instance of set weighted (r, p, q)-Packing. This yields an algorithm
for set weighted (r, p, q)-Packing with running time FrO(k/r)nO(1).
5.4 (r, k)-Monomial Detection
1 ··· X dn
In this subsection we consider polynomials from ring Z[X1, . . . , Xn]. We say a monomial X d1
n is an
if for all i ∈ [n], we have di ≤ r. We say the monomial is an (r, k)-monomial
if the above
r-monomial
holds and the total degree of the monomial is k. Let C be an arithmetic circuit computing a polynomial
f (X1, . . . , Xn) ∈ Z[X1, . . . , Xn]. We say C is non-canceling if it contains only variables at its leaves (i.e., no
constants), and only addition and multiplication gates (i.e., no substractions). For a non-canceling circuit C,
we define C to be the number of multiplication and addition gates plus the number of leaves (each containing
a variable). We assume the fan-in of a non-canceling circuit is two, i.e., each multiplication and addition gate
has at most two wires coming in. (This will simply be convenient for bounding the running time as a function
of C.)
(r, k)-Monomial Detection
Input: A non-canceling circuit C computing a polynomial f (X1, . . . , Xn) ∈ Z[X1, . . . , Xn], integers r, k.
Parameters: r, k
Question: Determine if there exists an (r, k)-monomial in f with non-zero coefficient and if so, then return
such a monomial.
Theorem 5.8. Given a non-canceling circuit C computing a polynomial f (X1, . . . , Xn) ∈ Z[X1, . . . , Xn],
(r, k)-Monomial Detection can be solved in deterministic time Ok(C · r18k/r · 2O(k/r) · n log3 n).
Proof. For each gate s of C, let fs be the polynomial computed at s. Define Ps to be the set of r-monomials of
total degree at most k that appear in fs with nonzero coefficient. We can view Ps as a family of (r, k)-sets: An
r-monomial M = X d1
n of total degree at most k corresponds to the (r, k)-set where each j ∈ [n] appears
dj times (in this case, we put uniform weights on the elements of universe [n]). We present an algorithm that
computes, for every gate s of C, a subfamily Ps ⊆ Ps of size Ok(r6k/r · 2O(k/r) · log n) that represents Ps. Let
sout be the output gate of C. By Lemma 5.1, if f contains an r-monomial of total degree k, then Psout will
contain such a monomial. Hence, to solve (r, k)-Monomial Detection it suffices to check whether Psout is
nonempty.
1 ··· X dn
12
We compute the sets Ps from bottom to top. For s being an input gate containing variable Xi, we simply
put Ps = Ps = {{i}}. Take then any non-input gate s, and suppose we have computed Ps1 and Ps2 for the
gates s1 and s2 having wires into s.
1. If s is an addition gate, then we define P ′ , Ps1 ∪ Ps2 . We claim that P ′ represents Ps: Since C is
non-canceling, the set of monomials that appear in fs with a nonzero coefficient is simply the union of the
sets of monomials of appearing in fs1 and in fs2. In particular, Ps = Ps1 ∪Ps2 . Therefore, by Lemma 5.3
we infer that P ′ = Ps1 ∪ Ps2 represents Ps. Note that P ′ ≤ Ps1 + Ps2 = Ok(r6k/r · 2O(k/r) · log n)
and P ′ can be computed in time Ok( Ps1 · Ps2 · n) = Ok(r12k/r · 2O(k/r) · n log2 n).
2. If s is a multiplication gate, then we define P ′ , Ps1 • Ps2 . Since C is non-canceling, the set of
monomials appearing in fs is exactly the set of all products of a monomial appearing in fs1 and a
monomial appearing in fs2 . In particular, Ps is exactly the set of these products that are also (r, k)-
monomials. When viewed as a multiset, the product of monomials M1 and M2 is the multiset M1 + M2.
Thus, we have that Ps = Ps1 • Ps2 and therefore, using Lemma 5.5, we infer that P ′ = Ps1 • Ps2
represents Ps. Note that P ′ ≤ Ps1 · Ps2 = Ok(r12k/r · 2O(k/r) · log2 n) and P ′ can be computed in
time Ok( Ps1 · Ps2 · n) = Ok(r12k/r · 2O(k/r) · n log2 n).
3. Now, using the algorithm of Corollary 3.8, compute subfamily Ps of size Ok(r6k/r · 2O(k/r) · log n) that
represents P ′. This takes time Ok(P ′ · r6k/r · 2O(k/r) · n log n) = Ok(r18k/r · 2O(k/r) · n log3 n).
Since the above procedure is applied to every gate of C, the claimed running time follows.
We remark that, after a trivial modification, the algorithm above can equally easily solve also a weighted
variant of (r, k)-Monomial Detection, where each variable is equipped with a weight and we are interested
in extracting a monomial of minimum total weight, defined as the sum of the weights of variables times their
degrees. It is not hard to reduce the problems r-Simple k-Path and (r, p, q)-Packing, considered in the
previous sections, to this variant; we leave the details to the reader.
6 Low degree monomials and low degree spanning trees
Let G be a simple, undirected graph with n vertices. Let T be the family of spanning trees of G; In particular,
if G is not connected then T = ∅. With every edge e ∈ E(G) we associate a variable ye. The Kirchhoff's
polynomial of G is defined as:
KG((ye)e∈E(G)) = X
T ∈T
Y
e∈E(T )
ye.
Thus, KG is a polynomial in Z[(ye)e∈E(G)]. Let v1, v2, . . . , vn be an arbitrary ordering of V (G). The Laplacian
of G is an n × n matrix LG = [aij ], where
aij =
ye
Pe incident to vi
−yvivj
0
if i = j,
if i 6= j and vivj ∈ E(G),
if i 6= j and vivj /∈ E(G).
Observe that LG is symmetric and the entries in every column and in every row of LG sum up to zero. Then
it can be shown that all the first cofactors of LG, i.e., the determinants of matrix LG after removing a row
and a column with the same indices, are equal. Let NG be this common value; then NG is again a polynomial
over variables (ye)e∈E(G). The Kirchhoff's Matrix Tree Theorem, in its general form, states that these two
polynomials coincide.
Theorem 6.1 (Matrix Tree Theorem). KG = NG.
We remark that the Matrix Tree Theorem is usually given in the more specific variants, where all variables ye
are replaced with 1; then the theorem expresses the number of spanning trees of G in terms of the first cofactors
of LG. However, the proof can be easily extended to the above, more general form; see for instance [23].
Observe that Theorem 6.1 provides a polynomial-time algorithm for evaluating KG over a given field and
vector of values of variables (ye)e∈E(G). Indeed, we just need to construct matrix LG, remove, say, the first
row and the first column, and compute the determinant. We now present how this observation can be used
to design a fast exact algorithm for the Degree-Bounded Spanning Tree problem.
13
Theorem 6.2. The Degree-Bounded Spanning Tree problem can be solved in randomized time O∗(dO(n/d))
with false negatives.
Proof. Associate every vertex v of the given graph G with a distinct variable xv. Let KG ∈ Z[(xv)v∈V (G)] be
a polynomial defined as KG with every variable yuv, for uv ∈ E(G), evaluated to xuxv. Then it follows that
KG((xv)v∈V (G)) = X
T ∈T
Y
v∈V (G)
xdegT (v)
v
.
Observe also that K G is 2(n− 1)-homogenous, that is, all the monomials of K G have their total degrees equal
to 2(n − 1). Thus, graph G admits a spanning tree with maximum degree at most d if and only if polynomial
KG contains a (d, 2(n − 1))-monomial. Using Theorem 6.1 we can construct a nO(1)-sized circuit evaluating
KG. Hence, verifying whether KG contains a (d, 2(n − 1))-monomial boils down to applying the algorithm
of Abasi et al. [1] for (r, k)-Monomial Detection with r = d and k = 2(n − 1). This algorithm runs in
randomized time O∗(dO(n/d)) and can only produce false negatives.
Let us repeat that in the proof of Theorem 6.2 we could not have used Theorem 5.8 instead of the result
of Abasi et al. [1], because the constructed circuit is not non-cancelling. Derandomizing the algorithm and
extending it to the weighted setting remains hence open.
Interestingly, the running time of the algorithm of Theorem 6.2 is essentially optimal, up to the log d factor
in the exponent. A similar lower bound for r-Simple k-Path was given by Abasi et al. [1].
Theorem 6.3. Unless ETH fails, there exists a constant s > 0 such that for no fixed integer d ≥ 2 the
Degree-Bounded Spanning Tree problem with the degree bound d can be solved in time O∗(2sn/d).
Proof. It is known (see e.g. [6]) that, assuming ETH, there exists a constant s > 0 such that the Hamiltonian
Path problem cannot be solved in time O∗(2sn) on n-vertex graphs. Consider the following reduction from
Hamiltonian Path to Degree-Bounded Spanning Tree with the degree bound d: given an instance G
of Hamiltonian Path, create G′ by attaching to every vertex v ∈ V (G) a set of d − 2 degree-1 vertices,
adjacent only to v. Since the new vertices have to be leaves in every spanning tree of G′, it follows that every
spanning tree T ′ of G′ is in fact a spanning tree of G with all the vertices of V (G′) \ V (G) attached as leaves.
In particular, G′ admits a spanning tree with maximum degree d if and only if G admits a spanning tree with
maximum degree 2, i.e., a hamiltonian path.
Observe that V (G′) = (d − 1) · V (G). Hence, if there was an algorithm solving Degree-Bounded
Spanning Tree with the degree bound d in time O∗(2sn/d), then by composing the reduction with the
algorithm we would be able to solve Hamiltonian Path on an n-vertex graph in time O∗(2s(d−1)n/d) ≤
O∗(2sn), thus contradicting ETH.
7 Conclusions
In this paper we considered relaxation parameter variants of several well studied problems in parameterized
complexity and exact algorithms. We proved, somewhat surprisingly, that instances with moderate values of
the relaxation parameter are significantly easier than instances of the original problems. We hope that our
work, together with the result of Abasi et al. [1] breaks the ground for a systematic investigation of relaxation
parameters in parameterized complexity and exact algorithms. We conclude with mentioning some of the
most natural concrete follow up questions to our work.
• We gave a determinstic algorithm for non-cancelling (r, k)-Monomial Detection with running time
2O(k log r
r )CnO(1), while Abasi et al. [1] gave a randomized algorithm with such a running time for (r, k)-
Monomial Detection without the non-cancellation restriction. Is there a determinstic algorithm for
(r, k)-Monomial Detection with running time 2O(k log r
r )CnO(1)?
• Does there exist a deterministic algorithm with running time 2O(n log r
Tree? Note that a deterministic algorithm with running time 2O(k log r
Detection immediately imply such an algorithm for Degree-Bounded Spanning Tree.
r ) for Degree-Bounded Spanning
r )CnO(1) for (r, k)-Monomial
14
• Is there a 2O(k log r
r )nO(1) time algorithm for the problem where we are given as input a graph G, integers
k and d, and asked whether G contains a subtree T on at least k vertices, such that the maximum degree
of T is at most d? Observe that for k = n this is exactly the Degree-Bounded Spanning Tree
problem.
• Is it possible to obtain polynomial kernels for problems with relaxation parameters with smaller and
smaller size bounds as the relaxation parameter increases?
• Are the log r (or log d) factors in the exponents of our running time bounds necessary? For example, is
there an algorithm for r-Simple k-Path with running time 2O(k/r)nO(1)? Or a 2O(n/d) time algorithm
for Degree Bounded Spanning Tree?
Acknowledgements
The first author thanks Fedor V. Fomin for hosting him at a visit in the University of Bergen where this
research was initiated.
References
[1] H. Abasi, N. H. Bshouty, A. Gabizon, and E. Haramaty. On r-Simple k-Path. In MFCS 2014 (Part II),
pages 1–12, 2014.
[2] N. Alon, O. Goldreich, J. Hastad, and R. Peralta. Simple construction of almost k-wise independent
random variables. Random Struct. Algorithms, 3(3):289–304, 1992.
[3] N. Alon, R. Yuster, and U. Zwick. Color coding. In Encyclopedia of Algorithms. 2008.
[4] A. Bjorklund, T. Husfeldt, P. Kaski, and M. Koivisto. Narrow sieves for parameterized paths and packings.
CoRR, abs/1007.1161, 2010.
[5] N. H. Bshouty. Testers and their applications. In ITCS 2014, pages 327–352, 2014.
[6] M. Cygan, F. V. Fomin, D. Lokshtanov, L. Kowalik, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh.
Parameterized Algorithms. Springer, 2015. In press.
[7] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Texts in Computer Science.
Springer, 2013.
[8] H. Fernau, A. L´opez-Ortiz, and J. Romero. Kernelization algorithms for packing problems allowing
overlaps. In TAMC 2015, pages 415–427, 2015.
[9] J. Flum and M. Grohe. Parameterized Complexity Theory. Texts in Theoretical Computer Science. An
EATCS Series. Springer-Verlag, Berlin, 2006.
[10] F. V. Fomin, F. Grandoni, D. Lokshtanov, and S. Saurabh. Sharp separation and applications to exact
and parameterized algorithms. Algorithmica, 63(3):692–706, 2012.
[11] F. V. Fomin, D. Lokshtanov, F. Panolan, and S. Saurabh. Representative sets of product families. In
ESA 2014, pages 443–454, 2014.
[12] F. V. Fomin, D. Lokshtanov, and S. Saurabh. Efficient computation of representative sets with applications
in parameterized and exact algorithms. In SODA 2014, pages 142–151, 2014.
[13] M. X. Goemans. Minimum bounded degree spanning trees. In FOCS 2006, pages 273–282, 2006.
[14] I. Koutis. Faster algebraic algorithms for path and packing problems. In ICALP 2008, pages 575–586,
2008.
[15] N. Linial, M. Luby, M. E. Saks, and D. Zuckerman. Efficient construction of a small hitting set for
combinatorial rectangles in high dimension. Combinatorica, 17(2):215–234, 1997.
15
[16] B. Monien. How to find long paths efficiently.
In Analysis and design of algorithms for combinato-
rial problems (Udine, 1982), volume 109 of North-Holland Math. Stud., pages 239–254. North-Holland,
Amsterdam.
[17] J. Naor and M. Naor. Small-bias probability spaces: Efficient constructions and applications. SIAM J.
Comput., 22(4):838–856, 1993.
[18] M. Naor, L. J. Schulman, and A. Srinivasan. Splitters and near-optimal derandomization. In FOCS 1995,
pages 182–191, 1995.
[19] R. Niedermeier. Invitation to fixed-parameter algorithms, volume 31 of Oxford Lecture Series in Mathe-
matics and its Applications. Oxford University Press, Oxford, 2006.
[20] R. Y. Pinter, H. Shachnai, and M. Zehavi. Deterministic parameterized algorithms for the graph motif
problem. In MFCS 2014 (II), pages 589–600, 2014.
[21] H. Shachnai and M. Zehavi. Representative families: A unified tradeoff-based approach. In ESA 2014,
pages 786–797, 2014.
[22] M. Singh and L. C. Lau. Approximating minimum bounded degree spanning trees to within one of
optimal. J. ACM, 62(1):1:1–1:19, 2015.
[23] W. T. Tutte. Graph Theory. Cambridge University Press, 2001.
[24] R. Williams. Finding paths of length k in O∗(2k) time. Inf. Process. Lett., 109(6):315–318, 2009.
[25] M. Zehavi. Deterministic parameterized algorithms for matching and packing problems. CoRR,
abs/1311.0484, 2013.
[26] M. Zehavi.
Solving parameterized problems by mixing color coding-related techniques. CoRR,
abs/1410.5062, 2014.
16
|
1312.0194 | 1 | 1312 | 2013-12-01T09:05:14 | Some Combinatorial Problems on Binary Matrices in Programming Courses | [
"cs.DS",
"math.CO"
] | The study proves the existence of an algorithm to receive all elements of a class of binary matrices without obtaining redundant elements, e. g. without obtaining binary matrices that do not belong to the class. This makes it possible to avoid checking whether each of the objects received possesses the necessary properties. This significantly improves the efficiency of the algorithm in terms of the criterion of time. Certain useful educational effects related to the analysis of such problems in programming classes are also pointed out. | cs.DS | cs |
Some Combinatorial Problems on Binary Matrices in
Programming Courses
Krasimir Yankov Yordzhev
South-West University "N. Rilsky", Blagoevgrad, Bulgaria
e-mail: [email protected]; [email protected]
Abstract
The study proves the existence of an algorithm to receive all elements of a class
of binary matrices without obtaining redundant elements, e. g. without obtaining
binary matrices that do not belong to the class. This makes it possible to avoid
checking whether each of the objects received possesses the necessary properties.
This significantly improves the efficiency of the algorithm in terms of the criterion
of time. Certain useful educational effects related to the analysis of such problems
in programming classes are also pointed out.
Key words: stimulation of students' interest, motivation to study, education in pro-
gramming, binary matrix, S-permutation matrices, combinatorial algorithms
2010 Mathematics Subject Classification: 97P50, 68R05, 05B20
1 Introduction
The stimulation of students' interest and motivation a certain discipline or branch of
science is a matter of combining a variety of methods. The present study discusses two
techniques which experience has shown to be quite efficient in programming classes:
1. Emphasizing the fact that the solution to a mathematical problem for all values
of the parameters is yet to be discovered (for example problems 1 and 2 below),
students can be given an assignment to write a program solving the problem in
certain specific cases, e. g. not for all the possible values of the parameters. In
section 2 two problems such as the ones described above will be formulated so that
the students can write a program for relatively small values of the parameters n
and k.
2. It is also possible to point out that a particular problem may be solved by applying
an algorithm that is more efficient as compared with standard algorithms which
excelling students do not normally find difficult to apply. Section 3 offers a specific
example how to apply this technique.
The present study is thus especially useful for students educated to become program-
mers as well as for their instructors and lecturers.
1
The paper discusses certain combinatorial problems on binary matrices.
For the classification of all non defined concepts and notations as well as for common
assertion which have not been proved here, we recommend sources [1, 6, 7, 10, 14].
2 Some mathematical problems whose solution for all
values of the parameters has not been discovered and
certain results related to these problems
A binary (or boolean, or (0,1)-matrix ) is a matrix whose all elements belong to the set
B = {0, 1}. With Bn we will denote the set of all n × n binary matrices.
Using the notation from [15], we will call Λk
n-matrices all n × n binary matrices in
each row and each column of which there are exactly k in number 1's.
Problem 1 Find out the number of all n × n binary matrices containing exactly k ele-
ments equal to 1 in each row and each column, e.g. the number of all Λk
n-matrices.
n-matrices with λ(n, k).
Let us denote the number of all Λk
Problem 1 has not been solved for all values of the parameters. That is there is no
known formula to calculate the λ(n, k) for all n and k. There are formulas for the calcu-
lation of the function λ(n, k) for each n for relatively small values of k; more specifically,
for k = 1, k = 2 and k = 3. We do not know formula to calculate the function λ(n, k)
for k > 3 and for all positive integer n.
It is easy to prove the following well-known formula:
In [13] is offered the formula:
λ(n, 1) = n!
λ(n, 2) =
X
2x2+3x3+···+nxn=n
(n!)2
xr!(2r)xr
n
Y
r=2
(1)
(2)
One of the first recursive formulas for the calculation of λ(n, 2) appeared in [2] (see
also [5, p. 763]):
λ(n, 2) =
1
2
n(n − 1)2 (cid:2)(2n − 3)λ(n − 2, 2) + (n − 2)2λ(n − 3, 2)(cid:3) ; n ≥ 4
λ(1, 2) = 0, λ(2, 2) = 1, λ(3, 2) = 6
Another recursive formula for the calculation of λ(n, 2) occurs in [4] :
λ(n, 2) = (n − 1)nλ(n − 1, 2) +
λ(1, 2) = 0,
λ(2, 2) = 1
(n − 1)2n
2
λ(n − 2, 2) ; n ≥ 3
2
(3)
(4)
The following recursive system for the calculation of λ(n, 2) is put forward in [11]:
λ(n + 1, 2) = n(2n − 1)λ(n, 2) + n2λ(n − 1, 2) − π(n + 1) ; n ≥ 2
π(n + 1) =
= n2(n−1)2
λ(1, 2) = 0,
π(1) = π(2) = π(3) = 0,
λ(2, 2) = 1,
π(4) = 9
[8(n − 2)(n − 3)λ(n − 2, 2) + (n − 2)2λ(n − 3, 2) − 4π(n − 1)] ; n ≥ 4
4
(5)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where π(n) identifies the number of a special class of Λ2
n-matrices.
The following formula in an explicit form for the calculation of λ(n, 3) is offered in
[8].
λ(n, 3) =
n!2
6n X (−1)β(β + 3γ)!2α3β
α!β!γ!26γ
(6)
where the sum is done as regard all (n+2)(n+1)
solutions in nonnegative integers of the
equation α+β+γ = n. As it is noted in [7] formula (6) does not give us good opportunities
to study behavior of λ(n, 3).
2
Let n be a positive integer and let A ∈ B
n2 is a n2 × n2 binary matrix. With the help
of n − 1 horizontal lines and n − 1 vertical lines A has been divided into n2 of number
non-intersecting n × n square sub-matrices Akl, 1 ≤ k, l ≤ n, e. g.
A =
A11 A12
A21 A22
...
...
An1 An2
· · · A1n
· · · A2n
...
...
· · · Ann
.
(7)
The sub-matrices Akl, 1 ≤ k, l ≤ n will be called blocks.
Adding one more condition, we can make the problem 1 more complicated:
Problem 2 Find out the number µ(n, k) of all n2 × n2 binary matrices that have k
elements equal to 1 in each row, each column, and each n × n block.
As demonstrated in [3], problem 2 has to do with the solution of a variety of combi-
natorial problems associated with the Sudoku riddles.
Problem 2 is solved in [3] for k = 1 and in [9] other methods are used to prove that
No formula has been put forward for the calculation of the function µ(n, k) when
µ(n, 1) = (n!)2n
(8)
k > 1.
3 S-permutation matrices
A matrix A ∈ B
n2 is called S-permutation if in each row, each column, and each block,
of A there is exactly one 1. Let the set of all n2 × n2 S-permutation matrices be denoted
by Σn2.
3
Two matrices A = (aij) ∈ Σn2 and B = (bij ) ∈ Σn2, 1 ≤ i, j ≤ n2 will be called
disjoint, if there are not elements with one and the same indices aij and bij such that
aij = bij = 1.
The following obvious proposition is given in [3]:
Proposition 1 [3] Square n2 × n2 matrix P with elements of Z
n2 = {1, 2, . . . , n2} is
Sudoku matrix if and only if there are matrices A1, A2, . . . , An2 ∈ Σn2 , each two of them
are disjoint and such that P can be given in the following way:
P = 1 · A1 + 2 · A2 + · · · + n2 · An2
✷
Let us analyze the following programming task:
Task 1 Write a program to obtain all S-permutation n2 × n2 matrices for a specific
positive integer n.
Experience shows that the majority of students do not fined it difficult to solve a task
such as the one offered above. Unfortunately, the solutions they normally suggest are
not very efficient. Below we present of the most common solutions given by students:
It is easy to observe that if we remove the condition to have only one 1 for each block
of the n2 × n2 binary matrices, the task above can be transformed into a task for the
obtaining of all permutations of the integers from 1 to n2. This combinatorial task is
often discussed in programming classes and a clear-cut solution can be found in a number
of study books, such as [12]. Let π = hp1, p2, . . . , pmi be a permutation of the integers
from 1 to m. Then we obtain the m × m binary matrix B = (bij) ∈ Bm, such that
bij = 1 if and only if pi = j, 1 ≤ i, j ≤ m. It is clear that the matrix B obtained in this
case has one 1 in each row and each column. This is where the name of such matrices
comes from: permutation matrices. This gives us the following algorithm for the solution
to task 1:
Algorithm 1 Obtaining all S-permutation matrices.
1. Obtain all the permutations of integers from 1 to n2;
2. For each permutation π = hp1, p2, . . . , pn2 i obtained in step 1, obtain the binary
n2 , such that αij = 1 if and only if pi = j. In all other cases
matrix A = (αij ) ∈ B
αij = 0, 1 ≤ i, j ≤ n2;
3. For each matrix obtained in step 2, check whether each block has only one 1. If
(true) then the matrix is S-permutation, if (false) then we remove this matrix
from the list.
Unfortunately, algorithm 1 entails the obtaining of a variety of redundant matrices
and a lot of time is wasted to check whether these meet the conditions (step 3). The total
number of the permutation matrices obtained in algorithm 1 is n2!, while according to
formula (8) the number of the S-permutation matrices is (n!)2n. But it is not difficult to
see that n2! > (n!)2n when n ≥ 2. Importantly, the program implementation in step 3 of
4
algorithm 1 is also significantly aggravated and requires certain efforts and mathematical
competence.
When n = 2, we have 22! = 24, (2!)2·2 = 16; when n = 3, we have 32! = 362 880,
(3!)2·3 = 46 656; when n = 4, we have 42! = 20 922 789 888 000, (4!)2·4 = 110 075 314 176,
n2!
(n!)2n increases
etc. It is possible to prove that as n increases, the value of expression
as well. This proves the inadequate efficiency of algorithm 1.
The present study will demonstrate that there is an algorithm for the obtaining of
all S-permutation matrices which bypasses the redundant, non S-permutation matrices.
This reduces the iterations in the algorithm to the absolute minimum and each iteration
bypasses the checking whether the matrix obtained is S-permutation. Such an algorithm
is obviously more efficient and takes less time to apply than algorithm 1. It is based on
theorem 1 proven below:
Let denote with Πn the set of all (2n)×n matrices, which we shortly call Πn matrices,
in which every row is a permutation of all elements of Zn = {1, 2, . . . , n}. It is obvious
that
Πn = (n!)2n
(9)
We will give a little bit more complicated definition of the term disjoint about Πn
matrices. Let C = (cij) and D = (dij ) be two Πn matrices. We say that C and D
are disjoint, if there are not natural numbers s, t ∈ {1, 2, . . . n} such that ordered pair
hcst, cn+t si has to be equal to the ordered pair hdst, dn+t si.
Theorem 1 There is a bijective map from Πn to Σn2 and the pair of disjoint matrices
of Πn corresponds to the pair of disjoint matrices of Σn2 .
Proof. Let P = (pij )2n×n ∈ Πn. We obtain the unique matrix of Σn2 from P with
the help of the following algorithm:
Algorithm 2 Obtaining just one matrix of Σn2 if there is given P = (pij)2n×n ∈ Πn.
1. for s = 1, 2, . . . , n do
2. for t = 1, 2, . . . , n do
begin
3. k := pst;
4. l := pn+t s;
5. Obtain n × n matrix Ast = (aij)n×n such that akl = 1 и aij = 0 in all other
occasions;
end;
6. Obtain matrix A according to formula (7);
5
Let s ∈ Zn = {1, 2, . . . , n}. Since ordered n-tuple hps1, ps2, . . . , psni which is s-th row
of the matrix P is a permutation, then in every row of n × n2 matrix
Rs = (cid:2) As1 As2
· · · Asn (cid:3)
there is only one 1. For every j = 1, 2, . . . , n Asj is binary n × n matrix in this case.
Analogously for every t ∈ Zn because ordered n-tuple hpn+t 1, pn+t 2, . . . , pn+t ni which
is (n + t)-th row of P is a permutation, then in every column of n2 × n matrix
Ct =
A1t
A2t
...
Ant
there is only one 1, where Ait, i = 1, 2, . . . , n is a binary n × n matrix. Hence, the matrix
A which is obtained with the help of algorithm 2 is Σn2 matrix.
Since for every P ∈ Πn with the help of algorithm 2 is obtained unique element of
Σn2, then this algorithm is a description of the map ϕ : Πn → Σn2. It is easy to see that
if there are given different elements of Πn, with the help of algorithm 2 we can obtain
different elements of Σn2. Hence, ϕ is an injection. But according to formulas (8) and
(9) Σn2 = Πn, from where it follows that ϕ is a bijection.
Analyzing algorithm 2 we take the conclusion that P and Q are disjoint matrices of
Πn if and only if ϕ(P ) and ϕ(Q) are disjoint matrices of Σn2 according to the above
given definitions. The theorem is proved.
✷
As an entailment of theorem 1, the following algorithm is received for the obtaining
of all S-permutation matrices, which, based on the arguments above, can be claimed to
be considerably more efficient than algorithm 1 for the same problem.
Algorithm 3 Getting of all S-permutation matrices.
1. Obtain all the permutations of integers from 1 to n2;
2. With the help of all permutations obtained in step 1, obtain all Πn matrices;
3. From each Πn matrix obtained in step 2, obtain the next S-permutation matrix with
the help of algorithm 2.
References
[1] M. Aigner Combinatorial theory. Springer-Verlag, 1979.
[2] H. Anand, V. C. Dumir, H. Gupta A combinatorial distribution problem. Duke
Math. J. 33 (1966), 757-769.
[3] G. Dahl Permutation Matrices Related to Sudoku. Linear Algebra and its Appli-
cations, 430 (2009), 2457-2463.
6
[4] I. Good, J. Grook The enumeration of arrays and generalization related to con-
tingency tables. Discrete Math, 19 (1977), 23-45.
[5] H. Gupta, G. L. Nath Enumeration of stochastic cubes. Notices of the Amer.
Math. Soc. 19 (1972) A-568.
[6] P. Lancaster Theory of Matrices. Academic Press, NY, 1969.
[7] R. P. Stanley Enumerative combinatorics. V.1, Wadword & Brooks, California,
1986.
[8] M. L. Stein, P. R. Stein Enumeration of stochastic matrices with integer ele-
ments. Los Alamos Scientific Laboratory Report LA-4434, 1970.
[9] K. Yordzhev On a Relationship Between the S-permutation Matrices and the
Bipartite Graphs. (to appear).
[10] В. И. Баранов, Б. С. Стечкин Экстремальные номбинаторные задачи и их
приложения. Москва, Наука, 1989
[11] К. Я. Йорджев Комбинаторни задачи над бинарни матрици. Математика и
математическо образование, 24 (1995), 288-296.
[12] П. Наков, П. Добриков Програмиране=++Алгоритми. Трето издание,
София 2005, ISBN 954-9805-06-X.
[13] В. Е. Тараканов Комбинаторные
задачи
на
бинарных матрицах.
Комбинаторный анализ, Москва, изд-во МГУ, 1980, вып.5, 4-15.
[14] В. Е. Тараканов Комбинаторные задачи и (0,1)-матрицы. Москва, Наука,
1985.
[15] В. С. Шевелев Редуцированные латинские прямоугольники и квадратные
матрицы с одинаковыми суммами в строках и столбцах. Дискретная мате-
матика, том 4, вып. 1, 1992, 91-110.
Associate Professor Dr. Krasimir Yankov Yordzhev
South-West University "N. Rilsky"
2700 Blagoevgrad, Ivan Mihaylov Str. 66
Bulgaria
e-mail: [email protected]; [email protected]
Некоторые комбинаторные задачи с бинарными матрицами на курсах
программирования
Исследование доказывает существование алгоритма для получения всех элемен-
тов класса бинарных матриц без получения избыточных элементов, т.е. без получе-
ния бинарных матриц, которые не принадлежат к этому классу. Это дает возмож-
ность избежать проверки, обладает ли каждый из полученных объектов требуемы-
ми свойствами. Так во много раз улучшается эффективность алгоритма в связи с
7
критерием времени. Обращается внимание на выгоды из рассмотреных задач для
обучения по программированию.
Ключевые слова: стимулирование интерес студентов, мотивация к обуче-
нию, подготовка в области программирования, бинарная матрица, S-матрица пе-
рестановок, комбинаторный алгоритм
Доц. д-р Красимир Я. Йорджев
Юго-западный университет "Неофит Рилски"
2700 Благоевград, ул. "Иван Михайлов" № 66
Болгария
e-mail: [email protected]; [email protected]
8
|
1601.05557 | 2 | 1601 | 2016-05-09T06:55:09 | A New Approach for Testing Properties of Discrete Distributions | [
"cs.DS",
"cs.IT",
"cs.IT",
"math.ST",
"math.ST"
] | In this work, we give a novel general approach for distribution testing. We describe two techniques: our first technique gives sample-optimal testers, while our second technique gives matching sample lower bounds. As a consequence, we resolve the sample complexity of a wide variety of testing problems.
Our upper bounds are obtained via a modular reduction-based approach. Our approach yields optimal testers for numerous problems by using a standard $\ell_2$-identity tester as a black-box. Using this recipe, we obtain simple estimators for a wide range of problems, encompassing most problems previously studied in the TCS literature, namely: (1) identity testing to a fixed distribution, (2) closeness testing between two unknown distributions (with equal/unequal sample sizes), (3) independence testing (in any number of dimensions), (4) closeness testing for collections of distributions, and (5) testing histograms. For all of these problems, our testers are sample-optimal, up to constant factors. With the exception of (1), ours are the {\em first sample-optimal testers for the corresponding problems.} Moreover, our estimators are significantly simpler to state and analyze compared to previous results.
As an application of our reduction-based technique, we obtain the first {\em nearly instance-optimal} algorithm for testing equivalence between two {\em unknown} distributions. Moreover, our technique naturally generalizes to other metrics beyond the $\ell_1$-distance.
Our lower bounds are obtained via a direct information-theoretic approach: Given a candidate hard instance, our proof proceeds by bounding the mutual information between appropriate random variables. While this is a classical method in information theory, prior to our work, it had not been used in distribution property testing. | cs.DS | cs |
A New Approach for Testing Properties of Discrete Distributions
Ilias Diakonikolas∗
Daniel M. Kane
University of Southern California
University of California, San Diego
[email protected].
[email protected].
May 10, 2016
Abstract
We study problems in distribution property testing: Given sample access to one or more
unknown discrete distributions, we want to determine whether they have some global property
or are ǫ-far from having the property in ℓ1 distance (equivalently, total variation distance, or
"statistical distance"). In this work, we give a novel general approach for distribution testing.
We describe two techniques: our first technique gives sample–optimal testers, while our sec-
ond technique gives matching sample lower bounds. As a consequence, we resolve the sample
complexity of a wide variety of testing problems.
Our upper bounds are obtained via a modular reduction-based approach. Our approach
yields optimal testers for numerous problems by using a standard ℓ2-identity tester as a black-
box. Using this recipe, we obtain simple estimators for a wide range of problems, encompassing
most problems previously studied in the TCS literature, namely: (1) identity testing to a fixed
distribution, (2) closeness testing between two unknown distributions (with equal/unequal sam-
ple sizes), (3) independence testing (in any number of dimensions), (4) closeness testing for
collections of distributions, and (5) testing histograms. For all of these problems, our testers
are sample-optimal, up to constant factors. With the exception of (1), ours are the first sample-
optimal testers for the corresponding problems. Moreover, our estimators are significantly sim-
pler to state and analyze compared to previous results.
As an important application of our reduction-based technique, we obtain the first nearly
instance-optimal algorithm for testing equivalence between two unknown distributions. The
sample complexity of our algorithm depends on the structure of the unknown distributions – as
opposed to merely their domain size – and is much better compared to the worst-case optimal
ℓ1-tester in most natural instances. Moreover, our technique naturally generalizes to other
metrics beyond the ℓ1-distance. As an illustration of its flexibility, we use it to obtain the first
near-optimal equivalence tester under the Hellinger distance.
Our lower bounds are obtained via a direct information-theoretic approach: Given a candi-
date hard instance, our proof proceeds by bounding the mutual information between appropriate
random variables. While this is a classical method in information theory, prior to our work, it
had not been used in distribution property testing. Previous lower bounds relied either on the
birthday paradox, or on moment-matching and were thus restricted to symmetric properties.
Our lower bound approach does not suffer from any such restrictions and gives tight sample
lower bounds for the aforementioned problems.
∗Part of this work was performed while the author was at the University of Edinburgh. Research supported by
EPSRC grant EP/L021749/1, a Marie Curie Career Integration Grant, and a SICSA grant.
1
Introduction
1.1 Background The problem of determining whether an unknown object fits a model based
on observed data is of fundamental scientific importance. We study the following formalization
of this problem: Given samples from a collection of probability distributions, can we determine
whether the distributions in question satisfy a certain property? This is the prototypical question in
statistical hypothesis testing [NP33, LR05]. During the past two decades, this question has received
considerable attention by the TCS community in the framework of property testing [RS96, GGR98],
with a focus on discrete probability distributions.
The area of distribution property testing [BFR+00, BFR+13] has developed into a mature re-
search field with connections to information theory, learning and statistics. The generic inference
problem in this field is the following: given sample access to one or more unknown distributions,
determine whether they have some global property or are "far" (in statistical distance or, equiva-
lently, ℓ1 norm) from having the property. The goal is to obtain statistically and computationally
efficient testing algorithms, i.e., algorithms that use the information-theoretically minimum sample
size and run in polynomial time. See [GR00, BFR+00, BFF+01, Bat01, BDKR02, BKR04, Pan08,
Val11, DDS+13, ADJ+11, LRR11, ILR12, CDVV14, VV14, DKN15b, DKN15a, ADK15, CDGR16]
for a sample of works and [Rub12, Can15] for two recent surveys.
In this work, we give a new general approach for distribution testing. We describe two novel
techniques: our first technique yields sample–optimal testers, while our second technique gives
matching sample lower bounds. As a consequence, we resolve the sample complexity of a wide
variety of testing problems.
All our upper bounds are obtained via a collection of modular reductions. Our reduction-based
method provides a simple recipe to obtain optimal testers under the ℓ1-norm (and other metrics),
by applying a randomized transformation to a basic ℓ2-identity tester. While the ℓ2-norm has been
used before as a tool in distribution testing [BFR+00], our reduction-based approach is conceptually
and technically different than previous approaches. We elaborate on this point in Section 1.4. We
use our reduction-based approach to resolve a number of open problems in the literature (see
Section 1.3). In addition to pinning–down the sample complexity of a wide range of problems, a
key contribution of our algorithmic approach is methodological. In particular, the main conceptual
message is that one does not need an inherently different statistic for each testing problem.
In
contrast, all our testing algorithms follow the same pattern: They are obtained by applying a
simple transformation to a basic statistic – one that tests the identity between two distributions in
ℓ2-norm – in a black-box manner. Following this scheme, we obtain the first sample-optimal testers
for many properties. Importantly, our testers are simple and, in most cases, their analysis fits in a
paragraph.
As our second main contribution, we provide a direct, elementary approach to prove sample
complexity lower bounds for distribution testing problems. Given a candidate hard instance, our
proof proceeds by bounding the mutual information between appropriate random variables. Our
analysis leads to new, optimal lower bounds for several problems, including testing closeness (under
various metrics), testing independence (in any dimension), and testing histograms. Notably, proving
sample complexity lower bounds by bounding the mutual information is a classical approach in
information theory. Perhaps surprisingly, prior to our work, this method had not been used in
distribution testing. Previous techniques were either based on the birthday paradox or on moment-
matching [RRSS09, Val11], and were thus restricted to testing symmetric properties. Our technique
circumvents the moment-matching approach, and is not restricted to symmetric properties.
1
1.2 Notation We write [n] to denote the set {1, . . . , n}. We consider discrete distributions over
i=1 pi = 1. We use the notation pi to denote the
probability of element i in distribution p.
[n], which are functions p : [n] → [0, 1] such thatPn
A pseudo-distribution over a finite set S is any function p : S → [0, 1]. For a distribution
p : [n] → [0, 1] and a set S ⊆ [n], we will denote by (pS) the conditional distribution on S and by
p[S] the pseudo-distribution obtained by restricting p on the set S.
The ℓ1 (resp. ℓ2) norm of a (pseudo-)distribution is identified with the ℓ1 (resp. ℓ2) norm of
i . The ℓ1 (resp. ℓ2) distance
between (pseudo-)distributions p and q is defined as the the ℓ1 (resp. ℓ2) norm of the vector of
the corresponding vector, i.e., kpk1 =Pn
their difference, i.e., kp − qk1 =Pn
i=1 pi and kpk2 =qPn
i=1 pi − qi and kp − qk2 =pPn
i=1 p2
i=1(pi − qi)2.
1.3 Our Contributions The main contribution of this paper is a reduction–based framework
to obtain testing algorithms, and a direct approach to prove lower bounds. We do not aim to
exhaustively cover all possible applications of our techniques, but rather to give some selected
results that are indicative of the generality and power of our methods. More specifically, we obtain
the following results:
1. We give an alternative optimal ℓ1-identity tester against a fixed distribution, with sample com-
plexity O(√n/ǫ2), matching the recently obtained tight bound [VV14, DKN15b]. The main advan-
tage of our tester is its simplicity: Our reduction and its analysis are remarkably short and simple
in this case. Our tester straightforwardly implies the "χ2 versus ℓ1" guarantee recently used as the
main statistical test in [ADK15].
2. We design an optimal tester for ℓ1-closeness between two unknown distributions in the standard
and the (more general) unequal-sized sample regimes. For the standard regime (i.e., when we draw
the same number of samples from each distribution), we recover the tight sample complexity of
O(max(n2/3/ǫ4/3, n1/2/ǫ2)), matching [CDVV14]. Importantly, our tester straightforwardly extends
to unequal-sized samples, giving the first optimal tester in this setting. Closeness testing with
unequal sized samples was considered in [AJOS14] that gives sample upper and lower bounds with
a polynomial gap between them. Our tester uses m1= Ω(max(n2/3/ǫ4/3, n1/2/ǫ2)) samples from one
/ǫ2,√n/ǫ2)) from the other. This tradeoff is sample-optimal
distribution and m2 = O(max(nm−1/2
(up to a constant factor) for all settings, and improves on the recent work [BV15] that obtains the
same tradeoff under the additional assumption that ǫ > n−1/12. In sharp contrast to [BV15], our
algorithm is extremely simple and its analysis fits in a few lines.
1
3. We study the problem of ℓ1-testing closeness between two unknown distributions in an instance-
optimal setting, where the goal is to design estimators whose sample complexity depends on the
(unknown) structure of the sampled distributions – as opposed to merely their domain size. We
obtain the first algorithm for this problem: Our tester uses
O(min
m>0
(m + kq<1/mk0 · kq<1/mk2/ǫ2 + kqk2/3/ǫ2))
samples from each of the distributions p, q on [n]. Here, q<1/m denotes the pseudo-distribution
obtained from q by removing the domain elements with mass ≥ 1/m, and kq<1/mk0 is the number
of elements with mass < 1/m. (Observe that since kq<1/mk2 ≤ 1/√m, taking m = min(n, n2/3/ǫ4/3)
attains the complexity of the standard ℓ1-closeness testing algorithm to within logarithmic factors.)
An important distinction between our algorithm and the instance-optimal identity testing algorithm
of [VV14] is that the sample complexity of the latter depends on the structure of the explicitly known
2
distribution, while ours depends on parameters of the unknown distributions. Roughly speaking,
the [VV14] algorithm knows a priori when to stop drawing samples, while ours "discovers" the
right sample complexity adaptively while obtaining samples. As an illustration of our technique,
we give an alternative algorithm for the identity testing problem in [VV14] whose analysis fits in
two short paragraphs. The sample complexity of our alternative algorithm is O(kqk2/3/ǫ2), where
q is the explicit distribution, matching the bound of [VV14] up to logarithmic factors.
4. We show that our framework easily generalizes to give near-optimal algorithms and lower
bounds for other metrics as well, beyond the ℓ1-norm. As an illustration of this fact, we describe
an algorithm and a nearly-matching lower bound for testing closeness under Hellinger distance,
H 2(p, q) = (1/2)k√p − √qk2
2, one of the most powerful f -divergences. This question has been
studied before: [GMV09] gave a tester for this problem with sample complexity O(n2/3/ǫ4). The
sample complexity of our algorithm is O(min(n2/3/ǫ4/3, n3/4/ǫ)), and we prove a lower bound of
Ω(min(n2/3/ǫ4/3, n3/4/ǫ)). Note that the second term of n3/4/ǫ in the sample complexity differs
from the corresponding ℓ1 term of n1/2/ǫ2.
5. We obtain the first sample-optimal algorithm and matching lower bound for testing indepen-
dence over ×d
i=1[ni]. Prior to our work, the sample complexity of this problem remained open,
even for the two-dimensional case. We prove that the optimal sample complexity of indepen-
dence testing (upper and lower bound) is Θ(maxj((Qd
i=1 ni)1/3/ǫ4/3)). Previ-
ous testers for independence were suboptimal up to polynomial factors in n and 1/ǫ, even for
d = 2. Specifically, Batu et al. [BFF+01] gave an independence tester over [n] × [m] with sam-
ple complexity eO(n2/3m1/3) · poly(1/ǫ), for n ≥ m. On the lower bound side, Levi, Ron, and
Rubinfeld [LRR11] showed a sample complexity lower bound of Ω(√nm) (for all n ≥ m), and
Ω(n2/3m1/3) (for n = Ω(m log m)). More recently, Acharya et al. [ADK15] gave an upper bound
of O(((Qd
i=1 ni)/ǫ2), which is optimal up to constant factors for the very special case
that all the ni's are the same. In summary, we resolve the sample complexity of this problem in
any dimension d, up to a constant factor, as a function of all relevant parameters.
i=1 ni)1/2 +Pd
(Qd
i=1 ni)1/2/ǫ2, n1/3
j
In fact, in the unknown-weights case, the problem is identical.
6. We obtain the first sample-optimal algorithms for testing equivalence for collections of distribu-
tions [LRR11] in the sampling and the oracle model, improving on [LRR11] by polynomial factors.
In the sampling model, we observe that the problem is equivalent to (a variant of) two-dimensional
independence testing.
In the
known-weights case, the problem is equivalent to two-dimensional independence testing, where the
algorithm is given explicit access to one of the marginals (say, the marginal on [m]). For this
setting, we give a sample-optimal tester with sample size O(max(√nm/ǫ2, n2/3m1/3/ǫ4/3))1. In the
query model, we give a sample-optimal closeness tester for m distributions over [n] with sample
complexity O(max(√n/ǫ2, n2/3/ǫ4/3)). This bound is independent of m and matches the worst-case
optimal bound for testing closeness between two unknown distributions.
7. As a final application of our techniques, we study the problem of testing whether a distribution
belongs in a given "structured family" [ADK15, CDGR16]. We focus on the property of being a
k-histogram over [n], i.e., that the probability mass function is piecewise constant with at most
k known interval pieces. This is a natural problem of particular interest in model selection. For
k = 1, the problem is tantamount to uniformity testing, while for k = Ω(n) it can be seen to be
equivalent to testing closeness between two unknown distributions over a domain of size Ω(n). We
1It should be noted that, while this is the same form as the sample complexity for independence testing in two
dimensions, there is a crucial difference. In this setting, the parameter m represents the support size of the marginal
that is explicitly given to us, rather than the marginal with smaller support size.
3
design a tester for the property of being a k-histogram (with respect to a given set of intervals)
with sample complexity O(max(√n/ǫ2, n1/3k1/3/ǫ4/3)) samples. We also prove that this bound is
information-theoretically optimal, up to constant factors.
1.4 Prior Techniques and Overview of our Approach In this section, we provide a detailed
intuitive explanation of our two techniques, in tandem with a comparison to previous approaches.
We start with our upper bound approach. It is reasonable to expect that the ℓ2-norm is useful
as a tool in distribution property testing.
Indeed, for elements with "small" probability mass,
estimating second moments is a natural choice in the sublinear regime. Alas, a direct ℓ2-tester will
often not work for the following reason: The error coming from the "heavy" elements will force the
estimator to draw too many samples.
In their seminal paper, Batu et al. [BFR+00, BFR+13] gave an ℓ2-closeness tester and used it
to obtain an ℓ1-closeness tester. To circumvent the aforementioned issue, their ℓ1-tester has two
stages: It first explicitly learns the pseudo-distribution supported on the heavy elements, and then it
applies the ℓ2-tester on the pseudo-distribution over the light elements. This approach of combining
learning (for the heavy elements) and ℓ2-closeness testing (for the light elements) is later refined by
Chan et al. [CDVV14], where it is shown that it inherently leads to a suboptimal sample complexity
for the testing closeness problem. Motivated by this shortcoming, it was suggested in [CDVV14]
that the use of the ℓ2-norm may be insufficient, and that a more direct approach may be needed to
achieve sample-optimal ℓ1-testers. This suggestion led researchers to consider different approaches
to ℓ1-testing (e.g., appropriately rescaled versions of the chi-squared test [ADJ+12, CDVV14, VV14,
BV15, ADK15]) that, although shown optimal for a couple of cases, lead to somewhat ad-hoc
estimators that come with a highly-nontrivial analysis.
Our upper bound approach postulates that the inefficiency of [BFR+00, BFR+13] is due to the
explicit learning of the heavy elements and not to the use of the ℓ2-norm. Our approach provides a
simple and general way to essentially remove this learning step. We achieve this via a collection of
simple reductions: Starting from a given instance of an ℓ1-testing problem A, we construct a new
instance of an appropriate ℓ2-testing problem B, so that the answers to the two problems for these
instances are identical. Here, problem A can be any of the testing problems discussed in Section 1.3,
while problem B is always the same. Namely, we define B to be the problem of ℓ2-testing closeness
between two unknown distributions, under the promise that at least one of the distributions in
question has small ℓ2-norm. Our reductions have the property that a sample-optimal algorithm for
problem B implies a sample-optimal algorithm for A. An important conceptual consequence of our
direct reduction-based approach is that problem B is of central importance in distribution testing,
since a wide range of problems can be reduced to it with optimal sample guarantees. We remark
that sample-optimal algorithms for problem B are known in the literature: a natural estimator
from [CDVV14], as well as a similar estimator from [BFR+00] achieve optimal bounds.
The precise form of our reductions naturally depends on the problem A that we start from.
While the details differ based on the problem, all our reductions rely on a common recipe: We
randomly transform the initial distributions in question (i.e., the distributions we are given sample
access to) to new distributions (over a potentially larger domain) such that at least one of the
new distributions has appropriately small ℓ2-norm. Our transformation preserves the ℓ1-norm, and
is such that we can easily simulate samples from the new distributions. More specifically, our
transformation is obtained by drawing random samples from one of the distributions in question
to discover its heavy bins. We then artificially subdivide each heavy bin into multiple bins, so that
the resulting distribution becomes approximately flat. This procedure decreases the ℓ2-norm while
increasing the domain size. By balancing these two quantities, we obtain sample-optimal testers
for a wide variety of properties.
4
In summary, our upper bound approach provides reductions of numerous distribution testing
problems to a specific ℓ2-testing problem B that yield sample-optimal algorithms. It is tempting to
conjecture that optimal reductions in the opposite direction exist, which would allow translating
lower bounds for problem B to tight lower bounds for other problems. We do not expect optimal
reductions in the opposite direction, roughly because the hard instances for many of our problems
are substantially different from the hard instances for problem B. This naturally brings us to our
lower bound approach, explained below.
Our lower bounds proceed by constructing explicit distributions D and D′ over (sets of) distribu-
tions, so that a random distribution p drawn from D satisfies the property, a random distribution
p from D′ is far from satisfying the property (with high probability), and it is hard to distin-
guish between the two cases given a small number of samples. Our analysis is based on classical
information-theoretic notions and is significantly different from previous approaches in this context.
Instead of using techniques involving matching moments [RRSS09, Val11], we are able to directly
prove that the mutual information between the set of samples drawn and the distribution that p
was drawn from is small. Appropriately bounding the mutual information is perhaps a technical
exercise, but remains quite manageable only requiring elementary approximation arguments. We
believe that this technique is more flexible than the techniques of [RRSS09, Val11] (e.g., it is not
restricted to symmetric properties), and may prove useful in future testing problems.
Remark 1.1. We believe that our reduction-based approach is a simple and appealing framework
to obtain tight upper bounds for distribution testing problems in a unifying manner. Since the
dissemination of an earlier version of our paper, Oded Goldreich gave an excellent exposition of our
approach in the corresponding chapter of his upcoming book [Gol16b].
1.5 Organization The structure of this paper is as follows:
In Section 2, we describe our
reduction-based approach and exploit it to obtain our optimal testers for a variety of problems.
In Section 3, we describe our lower bound approach and apply it to prove tight lower bounds for
various problems.
2 Our Reduction and its Algorithmic Applications
In Section 2.1, we describe our basic reduction from ℓ1 to ℓ2 testing. In Section 2.2, we apply our
reduction to a variety of concrete distribution testing problems.
2.1 Reduction of ℓ1-testing to ℓ2-testing The starting point of our reduction-based approach
is a "basic tester" for the identity between two unknown distributions with respect to the ℓ2-
norm. We emphasize that a simple and natural tester turns out to be optimal in this setting.
More specifically, we will use the following simple lemma (that follows, e.g., from Proposition 3.1
in [CDVV14]):
Lemma 2.1. Let p and q be two unknown distributions on [n]. There exists an algorithm that on
input n, ǫ > 0, and b ≥ max{kpk2,kqk2} draws O(bn/ǫ2) samples from each of p and q, and with
probability at least 2/3 distinguishes between the cases that p = q and kp − qk1 > ǫ.
Remark 2.2. We remark that Proposition 3.1 of [CDVV14] provides a somewhat stronger guar-
antee than the one of Lemma 2.1. Specifically, it yields a robust ℓ2-closeness tester with the
following performance guarantee: Given O(bn/ǫ2) samples from distributions p, q over [n], where
b ≥ max{kpk2,kqk2}, the algorithm distinguishes (with probability at least 2/3) between the cases
that kp − qk2 ≤ ǫ/(2√n) and kp − qk2 ≥ ǫ/√n. The soundness guarantee of Lemma 2.1 follows
from the Cauchy-Schwarz inequality.
5
Observe that if kpk2 and kqk2 are both small, the algorithm of Lemma 2.1 is in fact sample-
efficient. For example, if both are O(1/√n), its sample complexity is an optimal O(√n/ǫ2). On
the other hand, the performance of this algorithm degrades as kpk2 or kqk2 increases. Fortunately,
there are some convenient reductions that simplify matters. To begin with, we note that it suffices
that only one of kpk2 and kqk2 is small. This is essentially because if there is a large difference
between the two, this is easy to detect.
Lemma 2.3. Let p and q be two unknown distributions on [n]. There exists an algorithm that on
input n, ǫ > 0, and b ≥ min{kpk2,kqk2} draws O(bn/ǫ2) samples from each of p and q and, with
probability at least 2/3, distinguishes between the cases that p = q and kp − qk1 > ǫ.
Proof. The basic idea is to first test if kpk2 = Θ(kqk2), and if so to run the tester of Lemma 2.1.
To test whether kpk2 = Θ(kqk2), we estimate kpk2 and kqk2 up to a multiplicative constant factor.
It is known [GR00, BFF+01] that this can be done with O(√n) = O(min(kpk2,kqk2)n) samples. If
kpk2 and kqk2 do not agree to within a constant factor, we can conclude that p 6= q. Otherwise, we
use the tester from Lemma 2.1, and note that the number of required samples is O(kpk2n/ǫ2).
In our applications of Lemma 2.3, we take the parameter b to be equal to our upper bound on
min{kpk2,kqk2}. In all our algorithms in Section 2.2 this upper bound will be clear from the
context. If both our initial distributions have large ℓ2-norm, we describe a new way to reduce them
by splitting the large weight bins (domain elements) into pieces. The following key definition is the
basis for our reduction:
S that are equal to i. Thus,Pn
Definition 2.4. Given a distribution p on [n] and a multiset S of elements of [n], define the split
distribution pS on [n +S] as follows: For 1 ≤ i ≤ n, let ai denote 1 plus the number of elements of
i=1 ai = n +S. We can therefore associate the elements of [n +S]
to elements of the set B = {(i, j) : i ∈ [n], 1 ≤ j ≤ ai}. We now define a distribution pS with
support B, by letting a random sample from pS be given by (i, j), where i is drawn randomly from
p and j is drawn randomly from [ai].
We now show two basic facts about split distributions:
Fact 2.5. Let p and q be probability distributions on [n], and S a given multiset of [n]. Then: (i)
We can simulate a sample from pS or qS by taking a single sample from p or q, respectively. (ii) It
holds kpS − qSk1 = kp − qk1.
Fact 2.5 implies that it suffices to be able to test the closeness of pS and qS, for some S. In
particular, we want to find an S so that kpSk2 and kqSk2 are small. The following lemma shows
how to achieve this:
Lemma 2.6. Let p be a distribution on [n]. Then: (i) For any multisets S ⊆ S′ of [n], kpS ′k2 ≤
kpSk2, and (ii) If S is obtained by taking Poi(m) samples from p, then E[kpSk2
Proof. Let ai equal one plus the number of copies of i in S, and a′
copies of i in S′. We note that pS = (i, j) with probability pi/ai. Therefore, for (i) we have that
i equal one plus the number of
2] ≤ 1/m.
(pi/ai)2 =
p2
i /ai ≥
p2
i /a′
i = kpS ′k2
2.
nXi=1
For claim (ii), we note that the expected squared ℓ2-norm of pS is Pn
]. We note that
ai is distributed as 1 + X where X is a Poi(mpi) random variable. Recall that if Y is a random
variable distributed as Poi(λ), then E[zY ] = eλ(z−1). Taking an integral we find that
i=1 p2
i
E[a−1
i
kpSk2
2 =
nXi=1
aiXj=1
nXi=1
6
E [1/(1 + X)] = E(cid:20)Z 1
Therefore, we have that E[kpSk2
proof.
0
zX dz(cid:21) =Z 1
2] ≤ Pn
0
E[zX ]dz =Z 1
i /(mpi) = (1/m)Pn
0
i=1 p2
eλ(z−1)dz = (1 − e−λ)/λ ≤ 1/λ.
i=1 pi = 1/m. This completes the
2.2 Algorithmic Applications
2.2.1 Testing Identity to a Known Distribution We start by applying our framework to
give a simple alternate optimal identity tester to a fixed distribution in the minimax sense. In this
case, our algorithm is extremely easy, and provides a much simpler proof of the known optimal
bound [VV14, DKN15b]:
Proposition 2.7. There exists an algorithm that given an explicit distribution q supported on [n]
and O(√n/ǫ2) independent samples from a distribution p over [n] distinguishes with probability at
least 2/3 between the cases where p = q and kp − qk1 ≥ ǫ.
Proof. Let S be the multiset where S contains ⌊nqi⌋ copies of i. Note that S ≤ Pn
i=1 nqi = n.
Note also that qS assigns probability mass at most 1/n to each bin. Therefore, we have that
kqSk2 = O(1/√n). It now suffices to distinguish between the cases that pS = qS and the case that
kpS − qSk1 ≥ ǫ. Using the basic tester from Lemma 2.3 for b = O(1/√n), we can do this using
O(2nb/ǫ2) = O(√n/ǫ2) samples from pS. This can be simulated using O(√n/ǫ2) samples from p,
which completes the proof.
Remark 2.8. We observe that the identity tester of Proposition 2.7 satisfies a stronger guarantee:
More specifically, it distinguishes between the cases that χ2(p, q) :=Pn
i=1(pi−qi)2/qi ≤ ǫ2/10 versus
kp− qk1 ≥ ǫ. Hence, it implies Theorem 1 of [ADK15]. This can be seen as follows: As explained in
Remark 2.2, the basic tester of Lemma 2.1 from [CDVV14] is a robust tester with respect to the ℓ2-
norm. Thus, the tester of Proposition 2.7 distinguishes between the cases that kpS−qSk2 ≤ ǫ/(2√n)
and kpS − qSk2 ≥ ǫ/√n. The desired soundness follows from the fact kp − qk1 = kpS − qSk1 and
the Cauchy-Schwarz inequality. The desired "chi-squared" completeness property follows from the
easily verifiable (in)equalities χ2(p, q) = χ2(pS, qS) and χ2(pS, qS) ≥ n · kpS − qSk2
2.
Remark 2.9. After the dissemination of an earlier version of this paper, inspired by our work,
Goldreich [Gol16a] reduced testing identity to a fixed distribution to its special case of uniformity
testing, via a refinement of the above idea. Unfortunately, this elegant idea does not seem to
generalize to other problems considered here.
2.2.2 Testing Closeness between two Unknown Distributions We now turn to the prob-
lem of testing closeness between two unknown distributions p, q. The difficulty of this case lies in
the fact that, not knowing q, we cannot subdivide into bins in such a way as to guarantee that
kqSk2 = O(1/√n). However, we can do nearly as well by first drawing an appropriate number of
samples from q, and then using them to provide our subdivisions.
Proposition 2.10. There exists an algorithm that given sample access to two distributions p and
q over [n] distinguishes with probability 2/3 between the cases p = q and kp − qk1 > ǫ using
O(max(n2/3/ǫ4/3,√n/ǫ2)) samples from each of p and q.
Proof. The algorithm is as follows:
7
Algorithm Test-Closeness
Input: Sample access to distributions p and q supported on [n] and ǫ > 0.
Output: "YES" with probability at least 2/3 if p = q, "NO" with probability at least 2/3
if kp − qk1 ≥ ǫ.
1. Let k = min(n, n2/3ǫ−4/3).
2. Define a multiset S by taking Poi(k) samples from q.
3. Run the tester from Lemma 2.3 to distinguish between pS = qS and kpS − qSk1 ≥ ǫ.
To show correctness, we first note that with high probability we have S = O(n). Furthermore,
by Lemma 2.6 it follows that the expected squared ℓ2 norm of qS is at most 1/k. Therefore, with
probability at least 9/10, we have that S = O(n) and kqSk2 = O(1/√k).
The tester from Lemma 2.3 distinguishes between pS = qS and kpS−qSk1 ≥ ǫ with O(nk−1/2/ǫ2)
samples. By Fact 2.5, this is equivalent to distinguishing between p = q and kp− qk1 ≥ ǫ. Thus, the
total number of samples taken by the algorithm is O(k + nk−1/2/ǫ2) = O(max(n2/3ǫ−4/3,√n/ǫ2)).
We consider a generalization of testing closeness where we have access to different size samples
from the two distributions, and use our technique to provide the first sample-optimal algorithm for
the entire range of parameters:
Proposition 2.11. There exists an algorithm that given sample access to two distributions, p and
q over [n] distinguishes with probability 2/3 between the cases p = q and kp − qk1 > ǫ given m1
/ǫ2,√n/ǫ2)) samples from each of p and q.
samples from q and an additional m2 = O(max(nm−1/2
1
Proof. The algorithm is as follows:
Algorithm Test-Closeness-Unequal
Input: Sample access to distributions p and q supported on [n] and ǫ > 0.
Output: "YES" with probability at least 2/3 if p = q, "NO" with probability at least 2/3
if kp − qk1 ≥ ǫ.
1. Let k = min(n, m1).
2. Define a multiset S by taking Poi(k) samples from q.
3. Run the tester from Lemma 2.3 to distinguish between pS = qS and kpS − qSk1 ≥ ǫ.
To show correctness, we first note that with high probability we have S = O(n). Furthermore,
by Lemma 2.6 it follows that the expected squared ℓ2-norm of qS is at most 1/k. Therefore, with
probability at least 9/10, we have that S = O(n) and kqSk2 = O(1/√k).
The tester from Lemma 2.3 distinguishes between pS = qS and kpS−qSk1 ≥ ǫ with O(nk−1/2/ǫ2)
samples. By Fact 2.5, this is equivalent to distinguishing between p = q and kp−qk1 ≥ ǫ. In addition
to the m1 samples from q, we had to take O(nk−1/2/ǫ2) = O(m2) samples from each of p and q.
8
2.2.3 Nearly Instance–Optimal Testing In this subsection, we provide near-optimal testers
for identity and closeness in the instance-optimal setting. We start with the simpler case of testing
identity to a fixed distribution. This serves as a warm-up for the more challenging case of two
unknown distributions.
Note that the identity tester of Proposition 2.7 is sample-optimal only for a worst-case choice
of the explicit distribution q. (It turns out that the worst case corresponds to q being the uniform
distribution over [n].) Intuitively, for most choices of q, one can actually do substantially better.
This fact was first formalized and shown in [VV14].
In the following proposition, we give a very simple tester with a compact analysis whose sample
complexity is essentially optimal as a function of q. The basic idea of our tester is the following:
First, we partition the domain into categories based on the approximate mass of the elements of q,
and then we run an ℓ2-tester independently on each category.
Proposition 2.12. There exists an algorithm that on input an explicit distribution q over [n],
a parameter ǫ > 0, and O(kqk2/3/ǫ2) samples from a distribution p over [n] distinguishes with
probability at least 2/3 between the cases where p = q and kp − qk1 ≥ ǫ.
Proof. For j = 0, . . . , k with k = ⌈2 log2(10n/ǫ)⌉, let Sj ⊆ [n] be the set of coordinates i ∈ [n] so
that qi ∈ (2−j−1, 2−j]. Let S∞ be the set of coordinates i ∈ [n] so that qi < 2−k−1 ≤ ǫ/(10n). We
note that if kp − qk1 > ǫ, then kp[Sj] − q[Sj]k1 ≫ ǫ/ log(n/ǫ) for some j. We claim that it suffices
to design a tester that distinguishes between the cases that p[Sj] = q[Sj] and kp[Sj] − q[Sj]k1 ≫
ǫ/ log(n/ǫ) with probability at least 2/3. Repeating such a tester O(log log(n/ǫ)) times amplifies
its probability of success to 1 − 1/ log2(n/ǫ). By a union bound over j, all such testers are correct
with probability at least 2/3. To show completeness, note that if p = q, then p[Sj] = q[Sj] for all j,
and therefore all testers output "YES". For soundness, if kp − q 1 ≥ ǫ, there exists a j such that
kp[Sj] − q[Sj]k1 ≫ ǫ/ log(n/ǫ), and therefore the corresponding tester returns "NO".
To test whether kp[Sj] − q[Sj]k1 ≫ ǫ/ log(n/ǫ) versus p[Sj] = q[Sj], we proceed as follows: We
first use O(log2(n/ǫ)/ǫ2) samples to approximate kp[Sj]k1 to within additive error ǫ/(10 log(n/ǫ)).
If kq[Sj]k1 is not within the range of possible values, we determine that p[Sj] 6= q[Sj]. Otherwise, we
consider the conditional distributions (pSj), (qSj) and describe a tester to distinguish between the
cases that (pSj) = (qSj) and k(pSj)− (qSj)k1 ≫ ǫ/(log(n/ǫ)kq[Sj]k1). Note that we can assume
that kq[Sj]k1 ≫ ǫ/ log(n/ǫ), otherwise there is nothing to prove. We note that this necessarily fails
to happen if j = ∞.
Let mj = Sj. We have that kq[Sj]k1 = Θ(mj2−j) and k(qSj)k2 = Θ(m−1/2
). Therefore, using
the tester from Lemma 2.3, we can distinguish between (pSj) = (qSj) and k(pSj) − (qSj)k1 ≫
ǫ′ := ǫ/(log(n/ǫ)kq[Sj]k1) using O(k(qSj)k2 · mj/ǫ′2) = O(m5/2
j 4−j log2(n/ǫ)/ǫ2). samples from
(pSj). The probability that a sample from p lies in Sj is kp[Sj]k1 ≫ kq[Sj]k1 ≫ mj2−j. Using
rejection sampling, we can get a sample from (pSj) using O(2j/mj) samples from p. Therefore,
the number of samples from p needed to make the above determination is O(m3/2
j 2−j log2(n/ǫ)/ǫ2).
In summary, we have described a tester that distinguishes between p = q and kp− qk1 > ǫ with
j
sample complexity polylog(n/ǫ) · O((1 + maxj(m3/2
j 2−j))/ǫ2). We note that
kqk2/3 ≥ max
j Pi∈Sj
i !3/2
q2/3
≥ max
j
(mj2−2j/3)3/2 = max
j
(m3/2
j 2−j).
Therefore, the overall sample complexity is O(kqk2/3polylog(n/ǫ)/ǫ2) as desired.
9
We now show how to use our reduction-based approach to obtain the first nearly instance-
optimal algorithm for testing closeness between two unknown distributions. Note that the algorithm
of Proposition 2.12 crucially exploits the a priori knowledge of the explicit distribution.
In the
setting where both distributions are unknown, this is no longer possible. At a high-level, our
adaptive closeness testing algorithm is similar to that of Proposition 2.12: We start by partitioning
[n] into categories based on the approximate mass of one of the two unknown distributions, say
q, and then we run an ℓ2-tester independently on each category. A fundamental difficulty in our
setting is that q is unknown. Hence, to achieve this, we will need to take samples from q and create
categories based on the number of samples coming from each bin.
To state our result, we need the following notation:
Definition 2.13. Let q be a discrete distribution and x > 0. We denote by q<x the pseudo-
distribution obtained from q by setting the probabilities of all domain elements with probability at
least x to 0.
The main result of this subsection is the following:
Proposition 2.14. Given sample access to two unknown distributions, p, q over [n] and ǫ > 0,
there exists a computationally efficient algorithm that draws an expected
O(min
m>0
(m + kq<1/mk0kq<1/mk2/ǫ2 + kqk2/3/ǫ2))
samples from each of p and q, and distinguishes with probability 2/3 between p = q and kp−qk1 ≥ ǫ.
Before we proceed with the proof of Proposition 2.14 some comments are in order. First, note
that since kq<1/mk2 ≤ 1/√m, taking m = min(n, n2/3/ǫ4/3) attains the complexity of the standard
ℓ1-closeness testing algorithm to within logarithmic factors. (It should be noted that the logarithmic
factors in the above proposition can be removed by combining the required ℓ2-testers into a single
tester, using as a test statistic a linear combination of the individual test statistics.)
We now illustrate with a number of examples that the algorithm of Proposition 2.14 performs
substantially better than the worst-case optimal ℓ1-closeness tester in a number of interesting cases.
First, consider the case that the distribution q is essentially supported on relatively heavy bins. It
is easy to see that the sample complexity of our algorithm will then be roughly proportional to
kqk2/3/ǫ2. We remark that this bound is essentially optimal, even for the easier setting that q had
been given to us explicitly. As a second example, consider the case that q is roughly uniform. In this
case, we have that kqk2 will be small, and our algorithm will have sample complexity O(√n/ǫ2).
Finally, consider the case that the bins of the distribution q can be partitioned into two classes:
they have mass either approximately 1/n or approximately x > 1/n. For this case, our above
algorithm will need
O(min(x−1 + √n/ǫ2, nx−1/2/ǫ2))
samples.
(This follows by taking m = 2/x in the first case, and m = 1 in the second case.)
We remark that this sample bound can be shown to be optimal for such distributions (up to the
logarithmic factor in the O). Also note that the aforementioned sample upper bound is strictly
better than the worst-case bound of n2/3/ǫ4/3, unless x equals n−2/3ǫ4/3.
We are now ready to give the proof of Proposition 2.14.
Proof. We begin by describing and analyzing a testing algorithm that attains the stated sample
complexity for a given value of m. We then show how to adapt this algorithm to complete the
proof.
10
For the case of fixed m, our algorithm is given in the following pseudo-code:
Algorithm Test-Closeness-Adaptive
Input: Sample access to distributions p and q supported on [n] and m, ǫ > 0.
Output: "YES" with probability at least 2/3 if p = q, "NO" with probability at least 2/3
if kp − qk1 ≥ ǫ.
1. Let C be a sufficiently large constant. Draw Cm log2(n) independent samples from q.
2. Divide [n] into B
def
= O(log(m log(n))) categories in the following way: A bin is in
category S−∞ if at most log(n) of the samples from the previous step landed in the
bin. Otherwise, if a samples landed in the bin, place it in category S⌊log2(a)⌋.
3. Let pS and qS be the distributions over categories for p and q. Use the standard
ℓ1-tester to test whether pS = qS versus kpS − qSk1 ≥ ǫ/C with error probability at
most 1/10. If they are unequal, return "NO".
4. Approximate the probability mass q(S−∞) up to additive accuracy ǫ/C. If it is less
than 2ǫ/C, ignore this category in the following.
5. For each category Sa that is non-empty (and not S−∞ thrown out by the last step):
(a) Approximate q(Sa) to within a factor of 2 with error probability at most
1/(100B).
(b) Verify that p(Sa) is within a factor of 2 of this approximation. Otherwise, return
"NO".
(c) Approximate k(qSa)k2 to within a factor of 2 with error probability 1/(100B).
(d) Draw samples from p and q until they each have at least CSak(qSa)k2B3/ǫ2
many samples from Sa.
(e) Use these samples along with the ℓ2-tester of Lemma 2.3 to distinguish between
the cases (pSa) = (qSa) and k(pSa) − (qSa)k1 ≥ ǫ/(CBq(Sa)) with error
probability at most 1/(100B).
6. In the latter case, return "NO". Otherwise, return "YES".
To analyze the above algorithm, we note that it suffices to assume that all of the intermediate
tests for which the hypotheses are satisfied, return the correct results. We also note that it suffices to
consider the case that m ≤ n/ǫ2 (as otherwise the empirical distribution is already an O(m)-sample
algorithm).
We start by noting that with high probability over the samples taken in Step 1, for any bin i
with q(i) = x the number of samples drawn from this bin is at most xCm log3(n), and if x > 1/m
the number of samples is at least xCm log(n). Furthermore, if x < 1/(Cm), i lies in category S−∞
with high probability. We assume throughout the rest of this analysis that this event holds for all
bins.
We now prove correctness. Assuming that all intermediate tests are correct, it is easy to see
that if the algorithm returns "NO", then p and q must be unequal. We need to argue the converse.
If our algorithm returns "YES", it must be the case that
11
• kpS − qSk1 < ǫ/C.
• For each a with Sa non-empty, except for possibly S−∞, we have that k(pSa) − (qSa)k1 <
ǫ/(CBq(Sa)).
We claim that if this holds, then kp − qk1 < ǫ. This is because, after modifying p by at most ǫ/C,
we may keep the restrictions to each category the same and make it so that pS = qS. Once this is
true, it will be the case that
kp − qk1 =Xa
q(Sa)k(pSa) − (qSa)k1 < ǫ/C.
This completes the proof of correctness.
It remains to analyze the sample complexity. Step 1 clearly takes at most an appropriate
number of samples. Step 3 requires at most O(B/ǫ2) samples, which is sufficient for our purposes.
For Step 5, we may analyze the number of samples required for each category, Sa separately.
Note that approximating k(qSa)k2 to within a factor of 2 requires at most O(kq[Sa]k−1
log(B))
samples. If a ≥ 0 and Sa in non-empty, it will consist of at least one bin of mass at least (1/(Cm)),
so the sample complexity will be sufficiently small. For a = −∞, kq[Sa]k−1
2 ≤ kq[Sa]k0kq[Sa]k2,
and this is within our desired sample complexity bound.
We note that CSak(qSa)k2B3/ǫ2 samples are indeed enough to run our ℓ2-tester. We can
2
obtain this many samples by taking at most
polylog(n/ǫ)Sak(qSa)k2/q(Sa)/ǫ2 = polylog(n/ǫ)Sakq[Sa]k2/ǫ2
samples from each of p and q. Now, if a > 0, all bins in Sa have mass xpolylog(n) for some
appropriate value of x. We will then have that
Sakq[Sa]k2 = xSa3/2polylog(n) ≤ kq[Sa]k2/3polylog(n) ≤ kqk2/3polylog(n),
which is sufficient.
Finally, for a = −∞, the number of samples required is
polylog(n/ǫ)Sakq[Sa]k2/ǫ2 ≤ polylog(n/ǫ) · kq<1/mk0 · kq<1/mk2/ǫ2.
This completes the proof fo the case of a fixed value of m.
In order to obtain the minimum over all m we proceed as follows: First, since the product
kq<1/mk0 · kq<1/mk2 is decreasing in m, the minimum is attained (up to a constant factor) by
taking m to be the smallest power of 2 so that m > kqk2/3/ǫ2 + kq<1/mk0kq<1/mk2/ǫ2. Therefore,
it suffices to iterate the above procedure taking m to be increasingly large powers of 2 until the
algorithm terminates with O(mpolylog(n/ǫ)) samples. This completes the proof.
2.2.4 Testing Closeness in Hellinger Distance In this subsection, we use our reduction-
based approach to obtain a nearly sample-optimal algorithm for testing closeness of two unknown
distributions with respect to the Hellinger distance. We prove:
Proposition 2.15. There exists an algorithm that given sample access to two distributions p and q
supported on [n] draws O(min(n2/3/ǫ4/3, n3/4/ǫ)) samples from each and distinguishes between the
cases p = q and H 2(p, q) ≥ ǫ with probability at least 2/3.
12
Proof. First, note that an O(n2/3/ǫ4/3) upper bound follows immediately from the upper bound
on ℓ1-testing and the fact that kp − qk1 ≥ H 2(p, q). To prove the O(n3/4/ǫ) lower bound, we use
ideas similar to those in our adaptive closeness tester from the previous subsection. In particular,
let m = n3/4/ǫ. We take O(m) samples from q to divide [n] into O(log(m)) categories so with
high probability we have: (i) in all but one category, each bin in the category has the same mass
under q up to polylog(m) factors, and (ii) all bins in the remaining category have mass at most
1/m. We can then verify using O(n3/4/ǫ) samples that either pS 6= qS or that kpS − qSk1 < ǫ/100.
In the former case, we output "NO", while in the latter case it suffices to distinguish for each
category between the cases that p = q on that category, and that the contribution of that category
to H 2(p, q) is at least ǫ/(C log(m)), for C some sufficiently large constant.
Suppose that we have a category S so that for each bin in S the mass of this bin under q is
within a polylog factor of x. Then, H 2(p[S], q[S]) < polylog(m)kp[S] − q[S]k2
2/x. Therefore, it
suffices to distinguish between the cases p[S] = q[S] and kp[S] − q[S]k2
2 ≥ xǫ/polylog(m). This can
be done as follows: We define the distribution p′ by taking a sample from p, leaving it where it is
if the sample lies in S, and randomly and uniformly placing it in one of N new bins (for some very
large N ) otherwise. Defining q′ similarly, we note that kp′ − q′k2 = kp[S] − q[S]k2 + O(1/√N ).
Therefore, we can distinguish between p = q and kp[S] − q[S]k2 > δ using O(kq[S]k2/δ2) samples
with our standard ℓ2-tester. In summary, the number of samples required to perform this test is
polylog(m)kq[S]k2/(xǫ) ≤ O(n1/2x/(xǫ)) = O(n1/2/ǫ).
Finally, we need to consider the case of the last category. Here, we use that kp[S] − q[S]k1 ≥
H 2(p[S], q[S]), and therefore it suffices to distinguish between p[S] = q[S] and kp[S] − q[S]k1 > ǫ.
Equivalently, it suffices to distinguish between (pS) = (qS) and k(pS) − (qS)k1 > ǫ/q(S). This
can be achieved using
O(max(n2/3q(S)4/3/ǫ4/3, n1/2q(S)2/ǫ2))
samples from the conditional distributions. This is at most
O(max(n2/3q(S)1/3/ǫ4/3, n1/2q(S)/ǫ2))
samples from the original distribution. Since q(S) < n/m = n1/4ǫ, this quantity is at most
which completes the proof.
O(n3/4/ǫ) ,
Independence Testing In this subsection we study the problem of testing independence
2.2.5
of a d-dimensional discrete distribution p. More specifically, we want to design a tester that distin-
guishes between the case that p is a product distribution versus ǫ-far from any product distribution,
in ℓ1-norm. We start by giving an optimal independence tester for the two-dimensional case, and
then handle the case of arbitrary dimension.
Our algorithm for testing independence in two dimensions is as follows:
13
Algorithm Test-Independence-2D
Input: Sample access to a distribution p on [n] × [m] with n ≥ m and ǫ > 0.
Output:"YES" with probability at least 2/3 if the coordinates of p are independent, "NO"
with probability at least 2/3 if p is ǫ-far from any product distribution on [n] × [m].
1. Let k = min(n, n2/3m1/3ǫ−4/3).
2. Let S1 be a multiset in [n] obtained by taking Poi(k) samples from p1 = π1(p). Let
S2 be a multiset in [m] obtained by taking Poi(m) samples from p2 = π2(p). Let S
be the multiset of elements of [n] × [m] so that
1 + {Number of copies of (a, b) in S} =
(1 + {Number of copies of a in S1})(1 + {Number of copies of b in S2}).
3. Let q be the distribution on [n]× [m] obtained by taking (x1, y1), (x2, y2) independent
samples from p and returning (x1, y2). Run the tester from Lemma 2.3 to distinguish
between the cases pS = qS and kpS − qSk1 ≥ ǫ.
For correctness, we note that by Lemma 2.6, with probability at least 9/10 over our samples
from S1 and S2, all of the above hold: (i) S1 = O(n) and S2 = O(m), and (ii) k(p1)S1k2
2 = O(1/k),
k(p2)S2k2
2 = O(1/m). We henceforth condition on this event. We note that the distribution q is
exactly p1 × p2. Therefore, if the coordinates of p are independent, then p = q. On the other
hand, since q has independent coordinates, if p is ǫ-far from any product distribution, kp− qk1 ≥ ǫ.
Therefore, it suffices to distinguish between p = q and kp − qk1 ≥ ǫ. By Fact 2.5, this is equivalent
to distinguishing between pS = qS and kpS − qSk1 ≥ ǫ. This completes correctness.
We now analyze the sample complexity. We first draw samples when picking S1 and S2. With
high probability, the corresponding number of samples is O(m+k) = O(max(n2/3m1/3ǫ−4/3,√nm/ǫ2)).
Next, we note that qS = (p1)S1 × (p2)S2. Therefore, by Lemma 2.3, the number of samples drawn
in the last step of the algorithm is at most
O(nmkqSk2/ǫ2) = O(nmk(p1)S1 × (p2)S2k2/ǫ2) = O(nmk(p1)S1k2k(p2)S2k2/ǫ2)
= O(nmk−1/2m−1/2/ǫ2) = O(max(n2/3m1/3ǫ−4/3,√nm/ǫ2)).
Drawing a sample from q requires taking only two samples from p, which completes the analysis.
In the following proposition, we generalize the two-dimensional algorithm to optimally test
independence in any number of dimensions.
Proposition 2.16. Let p be a distribution on ×d
i=1[ni]. There is an algorithm that draws
Omax
j dYi=1
ni!1/2
/ǫ2, n1/3
j dYi=1
ni!1/3
/ǫ4/3
samples form p and with probability at least 2/3 distinguishes between the coordinates of p being
independent and p being ǫ-far from any such distribution.
Roughly speaking, our independence tester in general dimension uses recursion to reduce to the
2-dimensional case, in which case we may apply Test-Independence-2D. For the details, see the full
version.
14
j
(Qd
Proof. We can assume that all ni ≥ 2, for otherwise removing that term does not affect the
problem. We first note that the obvious generalization of Test-Independence-2D (that is, draw
min(ni, maxj n1/3
i=1 ni)1/3/ǫ4/3) samples from the i-th marginal and use them to subdivide the
domain in that dimension; then run the basic ℓ2-closeness tester between pS and the product of the
marginals) allows us for any constant d to distinguish between p having independent coordinates
and kp − p∗k1 > ǫ, where p∗ is the product of the marginals of p, with arbitrarily small constant
probability of failure. This generalization incurs an additional 2O(d) factor in the sample complexity,
hence is not optimal for super-constant d. To obtain the optimal sample complexity, we will use
the aforementioned algorithm for d = 2, 3 along with a careful recursion to reduce the dimension.
Our sample-optimal independence tester in d dimensions is as follows: First, let us assume for
simplicity that the maximum in the sample complexity is attained by the second term with j = 1.
Then, we use the algorithm Test-Independence-2D to distinguish between the cases that the first
coordinate is independent of the others from the case that p is at least ǫ/2-far from the product of
the distributions on the first coordinate and the distribution on the remaining coordinates. If it is
not, we return "NO". Otherwise, we recursively test whether or not the coordinates (p2, . . . , pd) are
independent versus at least ǫ/2-far from the product of their marginals, and return the result. We
note that if (p2, . . . , pd) is ǫ/2-close to the product distribution on p2, . . . , pd, and if p is ǫ/2-close
to the product distribution on its first coordinate with the remaining coordinates, then p is ǫ-close
to the product of its marginals.
We next deal with the remaining case. We let N =Qd
i=1 ni. We first partition [N ] into sets Si
nj ≤ √N . We do this by greedily adding elements to a single set S1
for 1 ≤ i ≤ 3 so that Qj∈Si
until the product is more than √N . We then remove the most recently added element, place it in
S2, and place all remaining elements in S3. This clearly satisfies the desired property. We let pSi
be the distribution of p ignoring all but the coordinates in Si. We use the obvious independence
tester in three dimensions to distinguish whether the pSi are independent versus p differing from
the product by at least ǫ/4. In the latter case, we return "NO". In the former, we recursively
distinguish between pSi having independent coordinates versus being ǫ/4-far from the product of
its marginals for each i and return "NO" unless all three pass.
i=1 ni)1/2/ǫ2, n1/3
In order to analyze the sample complexity, we note that our d-dimensional independence tester
i=1 ni)1/3/ǫ4/3)) samples on the highest level call to the 2 or 3-
dimensional version of the tester. It then needs to make O(1) recursive calls to the high-dimensional
i=1 ni)1/2 and error at
i=1 ni)1/3/ǫ2) samples, which is well
uses O(max((Qd
version of the algorithm on distributions with support of size at most (Qd
most ǫ/4. These recursive calls take a total of at most O((Qd
within our desired bounds.
j
(Qd
2.2.6 Testing Properties of Collections of Distributions In this subsection, we consider
the model of testing properties of collections of distributions [LRR11] in both the sampling and
query models.
We begin by considering the sampling model, as this is closely related to independence testing.
In fact, in the unknown-weights case, the problem is identical.
In the known-weights case, the
problem is equivalent to independence testing, where the algorithm is given explicit access to one
of the marginals (say, the distribution on [m]). For this setting, we give a tester with sample
complexity O(max(√nm/ǫ2, n2/3m1/3/ǫ4/3)). We also note that this bound can be shown the be
optimal. Formally, we prove the following:
Proposition 2.17. There is an algorithm that given sample access to a distribution p on [n] × [m]
and an explicit description of the marginal of p on [m] distinguishes between the cases that the
coordinates of p are independent and the case where p is ǫ-far from any product distribution on
15
[n] × [m] with probability at least 2/3 using O(max(√nm/ǫ2, n2/3m1/3/ǫ4/3)) samples.
Proof. The algorithm is as follows:
Algorithm Test-Collection-Sample-Model
Input: Sample access to a distribution p on [n]× [m] with ǫ > 0, and an explicit description
of the marginal of p on [m].
Output:"YES" with probability at least 2/3 if the coordinates of p are independent, "NO"
with probability at least 2/3 if p is ǫ-far from any product distribution on [n] × [m].
1. Let k = min(n, n2/3m1/3ǫ−4/3).
2. Let S1 be a multiset in [n] obtained by taking Poi(k) samples from p1 = π1(p). Let S2
be a multiset in [m] obtained by taking ⌊m(p2)i⌋ copies of i. Let S be the multiset of
elements of [n] × [m] so that
1 + {Number of copies of (a, b) in S} =
(1 + {Number of copies of a in S1})(1 + {Number of copies of b in S2}).
3. Let q be the distribution on [n]× [m] obtained by taking (x1, y1), (x2, y2) independent
samples from p and returning (x1, y2). Run the tester from Lemma 2.3 to distinguish
between the cases pS = qS and kpS − qSk1 ≥ ǫ.
For the analysis, we note that k(p2)S2k2 = O(1/√m) and with probability at least 9/10, it holds
k(p1)S1k2 = O(1/√k). Therefore, we have that k(p1 × p2)Sk2 = O(1/√km). Thus, the ℓ2-tester of
Lemma 2.3 draws O(nm1/2k−1/2/ǫ2) = O(max(√nm/ǫ2, n2/3m1/3/ǫ4/3)) samples and the sample
complexity is bounded as desired.
Next, we consider the query model. In this model, we are essentially guaranteed that the distri-
bution on [m] is uniform, but are allowed to extract samples conditioned on a particular value of the
second coordinate. Equivalently, there are m distributions q1, . . . , qm on [n].. We wish to distinguish
between the cases that the qi's are identical and the case where there is no distribution q so that
1
i=1 kq− qik1 ≤ ǫ. We show that we can solve this problem with O(max(√n/ǫ2, n2/3/ǫ4/3)) sam-
ples for any m. This is optimal for all m ≥ 2, even if we are guaranteed that q1 = q2 = . . . = q⌊m/2⌋
and q⌊m/2+1⌋ = . . . = qm.
mPm
Proposition 2.18. There is an algorithm that given sample access to distributions q1, . . . , qm on [n]
distinguishes between the cases that the qi's are identical and the case where there is no distribution
q so that 1
i=1 kq−qik1 ≤ ǫ with probability at least 2/3 using O(max(√n/ǫ2, n2/3/ǫ4/3)) samples.
mPm
Proof. The algorithm is as follows:
16
Algorithm Test-Collection-Query-Model
Input: Sample access to a distribution q1, . . . , qm on [n] with ǫ > 0.
Output:"YES" with probability at least 2/3 if the qi are identical, "NO" with probability
at least 2/3 if there is no distribution q so that 1
mPm
i=1 kq − qik1 ≤ ǫ.
1. Let C be a sufficiently large constant.
2. Let q∗ denote the distribution obtained by sampling from a uniformly random qi.
3. For k from 0 to ⌈log2(m)⌉:
(a) Select 25k/4C uniformly random elements i ∈ [m].
(b) For each selected i, use the ℓ1-closeness tester to distinguish between q∗ = qi and
kq∗ − qik1 > 2k−1ǫ with failure probability at most C −26−k.
(c) If any of these testers returned "NO", return "NO".
4. Return "YES".
To analyze this algorithm, we note that with probability 9/10 all the testers we call whose
hypotheses are satisfied output correctly. Therefore, if all qi are equal, they are equal to q∗, and
thus our algorithm returns "YES" with appropriately large probability. On the other hand, if for
any q we have that 1
i=1 kq − qik1 > ǫ, then in particular 1
i=1 kq∗ − qik1 > ǫ. Note that
1
m
mPm
kq∗ − qik1 ≤ ǫ/2 + O Xk
mXi=1
mPm
mPm
m
{i : kq∗ − qik ≥ 2k−1ǫ}2kǫ
! .
Therefore, since 1
i=1 kq∗ − qik1 > ǫ, we have that for some k it holds {i : kq∗ − qik ≥ 2k−1ǫ} =
Ω(m2−5k/4). For this value of k, there is at least a 9/10 probability that some i with this property
was selected as one of our C25k/4 that were used, and then assuming that the appropriate tester
returned correctly, our algorithm will output "NO". This establishes correctness. The total sample
O(max(√n/ǫ2, n2/3/ǫ4/3)).
complexity of this algorithm is easily seen to be Pk 25k/4k · O(√n/ǫ24−k + n2/3/ǫ4/32−4k/3) =
2.2.7 Testing k-Histograms Finally,
sample-optimal algorithm for the property of being a k-histogram with known intervals.
in this subsection we use our framework to design a
Let I be a partition of [n] into k intervals. We wish to be able to distinguish between the cases
where a distribution p has constant density on each interval versus the case where it is ǫ-far from
any such distribution. We show the following:
Proposition 2.19. Let I be a partition of [n] into k intervals. Let p be a distribution on [n].
There exists an algorithm which draws O(max(√n/ǫ2, n1/3k1/3/ǫ4/3)) independent samples from p
and distinguishes between the cases where p is uniform on each of the intervals in I from the case
where p is ǫ-far from any such distribution with probability at least 2/3.
Proof. First, we wish to guarantee that each of the intervals has reasonably large support. We can
achieve this as follows: For each interval I ∈ I we divide each bin within I into ⌈n/(kI)⌉ bins.
Note that this increases the number of bins in I by at most n/k, hence doing this to each interval
in I at most doubles the total size of the domain. Therefore, after applying this operation we get
a distribution over a domain of size O(n), and each of the k intervals in I is of length Ω(n/k).
17
Next, in order to use an ℓ2-closeness tester, we want to further subdivide bins using our ran-
domized transformation. To this end, we let m = min(k, n1/3k1/3/ǫ4/3) and take Poi(m) samples
from p. Then, for each interval Ii ∈ I, we divide each bin in Ii into ⌊nai/(kIi)⌋+ 1 new bins, where
ai is the number of samples that were drawn from Ii. Let I ′
i denote the new interval obtained from
Ii. Note that after this procedure the total number of bins is still O(n) and that the number of bins
i is now Ω((n/k)(ai + 1)). Let p′ be the distribution obtained from p under this transformation.
in I ′
Let q′ be the distribution obtained by sampling from p′ and then returning a uniform random bin
i as the sample. We claim that the ℓ2-norm of q′ is small. In particular the
from the same interval I ′
squared ℓ2-norm will be the sum over intervals I ′ in our new partition (that is, after the subdivisions
described above) of O(p(I ′)2/((n/k)(ai + 1))). Recall that 1/(ai + 1) has expectation at most
2 = O(k/(nm)).
We can now apply the tester from Lemma 2.3 to distinguish between the cases where p′ = q′
1/(mp(I ′)). This implies that the expected squared ℓ2-norm of q′ is at mostPI ′ O(p(I ′)/(nm/k)) =
O(k/(nm)). Therefore, with large constant probability, we have that kq′k2
and kp′ − q′k1 > ǫ with O(n1/2k1/2m−1/2/ǫ2) = O(max(√n/ǫ2, n1/3k1/3/ǫ4/3)) samples. We have
that p′ = q′ if and only if p is flat on each of the intervals in I, and kp′ − q′k1 > ǫ if p is ǫ-far from
any distribution which is flat on I. This final test is sufficient to make our determination.
3 Sample Complexity Lower Bounds
We illustrate our lower bound technique by proving tight information-theoretic lower bounds for
testing independence (in any dimension), testing closeness in Hellinger distance, and testing his-
tograms.
3.1 Lower Bound for Two-Dimensional Independence Testing
Theorem 3.1. Let n ≥ m ≥ 2 be integers and ǫ > 0 a sufficiently small universal constant. Then,
any algorithm that draws samples from a distribution p on [n] × [m] and, with probability at least
2/3, distinguishes between the case that the coordinates of p are independent and the case where p
is ǫ-far from any product distribution must use Ω(max(√nm/ǫ2, n2/3m1/3/ǫ4/3)) samples.
We split our argument into two parts proving each of the above lower bounds separately.
3.1.1 The Ω(√nmǫ−2) Lower Bound We start by proving the easier of the two bounds. It
should be noted that this part of the lower bound can essentially be obtained using known results.
We give a proof using our technique, in part as a guide to the somewhat more complicated proof
in the next section, which will be along similar lines.
First, we note that it suffices to consider the case where n and m are each sufficiently large since
Ω(ǫ−2) samples are required to distinguish the uniform distribution on [2]×[2] from the distribution
which takes value (i, j) with probability (1 + (2δi,j − 1)ǫ)/2.
Our goal is to exhibit distributions D and D′ over distributions on [n] × [m] so that all dis-
tributions in D have independent coordinates, and all distributions in D′ are ǫ-far from product
distributions, so that for any k = o(√nm/ǫ2), no algorithm given k independent samples from a
random element of either D or D′ can determine which family the distribution came from with
greater than 90% probability.
Although the above will be our overall approach, we will actually analyze the following gen-
eralization in order to simplify the argument. First, we use the standard Poissonization trick. In
particular, instead of drawing k samples from the appropriate distribution, we will draw Poi(k)
samples. This is acceptable because with 99% probability, this is at least Ω(k) samples. Next, we
relax the condition that elements of D′ be ǫ-far from product distributions, and simply require that
18
they are Ω(ǫ)-far from product distributions with 99% probability. This is clearly equivalent upon
accepting an additional 1% probability of failure, and altering ǫ by a constant factor.
Finally, we will relax the constraint that elements of D and D′ are probability distributions.
Instead, we will merely require that they are positive measures on [n] × [m], so that elements of D
are product measures and elements of D′ are Ω(ǫ)-far from being product measures with probability
at least 99%. We will require that the selected measures have total mass Θ(1) with probability
at least 99%, and instead of taking samples from these measures (as this is no longer as sensible
concept), we will use the points obtained from a Poisson process of parameter k (so the number
of samples in a given bin is a Poisson random variable with parameter k times the mass of the
bin). This is sufficient, because the output of such a Poisson process for a measure µ is identical to
the outcome of drawing Poi(kµk1k) samples from the distribution µ/kµk1. Moreover, the distance
from µ to the nearest product distribution is kµk1 times the distance from µ/kµk1 to the nearest
product distribution.
We are now prepared to describe D and D′ explicitly:
• We define D to deterministically return the uniform distribution µ with µ(i, j) = 1
nm for all
(i, j) ∈ [n] × [m].
• We define D′ to return the positive measure ν so that for each (i, j) ∈ [n] × [m] the value
nm each with probability 1/2 and independently over different pairs
nm or 1−ǫ
ν(i, j) is either 1+ǫ
(i, j).
It is clear that kµk1,kνk1 = Θ(1) deterministically. We need to show that the relevant Poisson
processes return similar distributions. To do this, we consider the following procedure: Let X be a
uniformly random bit. Let p be a measure on [n] × [m] drawn from either D if X = 0 or from D′
if X = 1. We run a Poisson process with parameter k on p, and let ai,j be the number of samples
drawn from bin (i, j). We wish to show that, given access to all ai,j's, one is not able to determine
the value of X with probability more than 51%. To prove this, it suffices to bound from above the
mutual information between X and the set of samples (ai,j)(i,j)∈[n]×[m]. In particular, this holds
true because of the following simple fact:
Lemma 3.2. If X is a uniform random bit and A is a correlated random variable, then if f is any
function so that f (A) = X with at least 51% probability, then I(X : A) ≥ 2 · 10−4.
Proof. This is a standard result in information theory, and the simple proof is included here for
the sake of completeness. We begin by showing that I(X : f (A)) ≥ 2 · 10−4. This is because the
conditional entropy, H(Xf (A)), is the expectation over f (A) of h(q) = −q log(q)−(1−q) log(1−q),
where q is the probability that X = f (A) conditional on that value of f (A). Since E[q] ≥ 51% and
since h is concave, we have that H(Xf (A)) ≤ h(0.51) < log(2) − 2 · 10−4. Therefore, we have that
I(X : f (A)) = H(X) − H(Xf (A)) ≥ log(2) − (log(2) − 2 · 10−4) = 2 · 10−4.
The lemma now follows from the data processing inequality, i.e., the fact that I(X : A) ≥ I(X :
f (A)).
In order to bound I(X : {ai,j}) from above, we note that the ai,j's are independent conditional
on X, and therefore that
I(X : (ai,j)(i,j)∈[n]×[m]) ≤ X(i,j)∈[n]×[m]
I(X : ai,j).
(1)
By symmetry, it is clear that all of the ai,j's are the same, so it suffices to consider I(X : a) for a
being one of the ai,j. We prove the following technical lemma:
19
Lemma 3.3. For all (i, j) ∈ [n] × [m], it holds I(X : ai,j) = O(k2ǫ4/(m2n2)).
The proof of this lemma is technical and is deferred to Appendix A. The essential idea is that
we condition on whether or not λ := k/(nm) ≥ 1. If λ < 1, then the probabilities of seeing 0 or 1
samples are approximately the same, and most of the information comes from how often one sees
exactly 2 samples. For λ ≥ 1, we are comparing a Poisson distribution to a mixture of Poisson
distributions with the same average mean, and we can deal with the information theory by making
a Gaussian approximation.
By Lemma 3.3, (1) yields that I(X : (ai,j)(i,j)∈[n]×[m]) = O(k2ǫ4/mn) = o(1). In conjunction
with Lemma 3.2, this implies that o(√mn/ǫ2) samples are insufficient to reliably distinguish an
element of D from an element of D′. To complete the proof, it remains to show that elements of
D are all product distributions, and that most elements of D′ are far from product distributions.
The former follows trivially, and the latter is not difficult. We show:
Lemma 3.4. With 99% probability a sample from D′ is Ω(ǫ)-far from being a product distribution.
Proof. For this, we require the following simple claim:
Claim 3.5. Let µ be a measure on [n] × [m] with marginals µ1 and µ2. If kµ − µ1 × µ2/kµk1k1 >
ǫkµk1, then µ is at least ǫkµk1/4-far from any product measure.
Proof. By normalizing, we may assume that kµk1 = 1. Suppose for the sake of contradiction that
for some measures ν1, ν2 it holds kµ − ν1 × ν2k1 ≤ ǫ/4. Then, we must have that kµi − νik1 ≤ ǫ/4.
This means that
kµ − µ1 × µ2k1 ≤ kµ − ν1 × ν2k1 + kν1 × ν2 − µ1 × ν2k1 + kµ1 × ν2 − µ1 × µ2k1
≤ ǫ/4 + kν2k1kµ1 − ν1k1 + kµ1k1kµ2 − ν2k1
≤ ǫ/4(3 + ǫ/4) ≤ ǫ ,
which yields the desired contradiction.
In light of the above claim, it suffices to show that with 99% probability over the choice of ν
from D′ we have kν − ν1 × ν2/kνkk1 = Ω(ǫ). For this, we note that when n and m are sufficiently
large constants, with 99% probability we have that: (i) kνk1 − 1 ≤ ǫ/10, (ii) ν1 has mass in the
range [(1 − ǫ/10)/n, (1 + ǫ/10)/n] for at least half of its points, and (iii) ν2 has mass in the range
[(1 − ǫ/10)/m, (1 + ǫ/10)/m] for at least half of its points. If all of these conditions hold, then for
at least a quarter of all points the mass assigned by ν1 × ν2/kνk1 is between (1 − ǫ/2)/(nm) and
(1 + ǫ/2)/(nm). In such points, the difference between this quantity and the mass assigned by ν is
at least ǫ/(2mn). Therefore, under these conditions, we have that
This completes the proof.
kν − ν1 × ν2/kνk1k1 ≥ (nm/4)(ǫ/(2mn)) = ǫ/8 = Ω(ǫ).
3.1.2 The Ω(n2/3m1/3ǫ−4/3) Lower Bound In this subsection, we prove the other half of the
lower bound. As in the proof of the previous subsection, it suffices to exhibit a pair of distributions
D,D′ over measures on [n] × [m], so that with 99% probability each of these measures has total
mass Θ(1), the measures from D are product measures and those from D′ are Ω(ǫ)-far from being
product measures, and so that if a Poisson process with parameter k = o(n2/3m1/3/ǫ−4/3) is used
to draw samples from [n] × [m] by way of a uniformly random measure from either D or D′, it is
impossible to reliably determine which distribution the measure came from.
We start by noting that it suffices to consider only the case where k ≤ n/2, since otherwise the
bound follows from the previous subsection. We define the distributions over measures as follows:
20
• When generating an element from either D or D′, we generate a sequence c1, . . . , cn, where
ci is 1/k with probability k/n and 1/n otherwise. Furthermore, we assume that the ci's are
selected independently of each other.
• Then D returns the measure µ where µ(i, j) = ci/m.
• The distribution D′ generates the measure ν, where ν(i, j) = 1/(km) if ci = 1/k and otherwise
ν(i, j) is randomly either (1 + ǫ)/(nm) or (1 − ǫ)/(nm).
It is easy to verify that with 99% probability that kµk1,kνk1 = Θ(1). It is also easy to see that
D only generates product measures. We can show that D′ typically generates measures far from
product measures:
Lemma 3.6. With 99% probability a sample from D′ is Ω(ǫ)-far from being a product distribution.
Proof. If ν is a random draw from D′ and ν2 is second marginal distribution, it is easy to see
that with high probability it holds ν2(j)/kνk1 ∈ [(1 − ǫ/3)/m, (1 + ǫ/3)/m] for at least half of
the j ∈ [m]. Also, with high probability, for at least half of the i ∈ [n] we have that ν1(i) ∈
[(1 − ǫ/3)/n, (1 + ǫ/3)/n]. For such pairs (i, j), we have that ν(i, j) − ν1(i)ν2(j)/kνk1 ≥ ǫ/(4nm),
and thus
The lemma follows by Claim 3.5.
kν − ν1 × ν2/kνk1k1 ≥ (nm/4)(ǫ/4nm) ≥ ǫ/16 = Ω(ǫ).
It remains to show that the Poisson process in question is insufficient to distinguish which
distribution the measure came from with non-trivial probability. As before, we let X be a uniformly
random bit, and let µ be a measure drawn from either D if X = 0 or D′ if X = 1. We run the
Poisson process and let ai,j be the number of elements drawn from bin (i, j). We let Ai be the
vector (ai,1, ai,2, . . . , ai,m). It suffices to show that the mutual information I(X : A1, A2, . . . , An) is
small.
Note that the Ai's are conditionally independent on X (though that ai,j's are not, because ai,1
and ai,2 are correlated due to their relation to ci). Therefore, we have that
I(X : A1, A2, . . . , An) ≤
nXi=1
I(X : Ai) = nI(X : A) ,
(2)
by symmetry where A = Ai. To complete the proof, we need the following technical lemma:
Lemma 3.7. We have that I(X : A) = O(k3ǫ4/(n3m)).
Morally speaking, this lemma holds because if all the c's were 1/n, we would get a mutual
information of roughly O(k2ǫ4/(n2m)), by techniques from the last section. However, the possibility
that c = 1/k adds sufficient amount of "noise" to somewhat decrease the amount of available
information. The formal proof is deferred to Appendix A. Combining (2) and the above lemma,
we obtain that
I(X : A1, . . . , An) = O(k3ǫ4/(n2m) = o(1).
This completes the proof.
21
3.2 Lower Bound for Hellinger Closeness Testing We prove that our upper bound for
Hellinger distance closeness is tight up to polylogarithmic factors.
Proposition 3.8. Any algorithm that given sample access to distributions p and q on [n] that distin-
guishes between p = q and H 2(p, q) > ǫ with probability at least 2/3 must take Ω(min(n2/3/ǫ4/3, n3/4/ǫ))
samples.
Proof. The proof of this lower bound follows our direct information-theoretic approach. We let X
be randomly either 0 or 1. We then describe a distribution of pairs of pseudo-distributions p, q on
[n] so that if X = 0 then p = q and if X = 1, H 2(p, q) ≫ ǫ with 99% probability, and so that
kpk1,kqk1 = Θ(1) with 99% probability. We then show that the mutual information between X
and the output of Poi(k) samples from each of p and q is o(1) for k = o(min(n2/3/ǫ4/3, n3/4/ǫ)).
We begin by describing this distribution. For i = 1, . . . , n − 1 with probability min(k/n, 1/2)
we set pi = qi = 1/(2k), otherwise if X = 0, we set pi = qi = ǫ/n, and if X = 1 randomly set either
pi = 2ǫ/n, qi = 0 or pi = 0, qi = 2ǫ/n. pn = qn = 1/3.
First, we note that X = 0, we have p = q and if X = 1, H 2(p/kpk1, q/kqk1) ≫ ǫ with high
Let ai, bi be the number of samples drawn from bin i under p and q respectively. We wish to
probability. Furthermore, kpk1,kqk1 = Θ(1) with high probability.
bound I(X : a1, b1, . . . , an, bn) from below. By conditional independence, this is
I(X : ai, bi) .
Xi
Note that I(X : an, bn) = 0, otherwise the distribution on (X, ai, bi) is independent of i, so we will
analyze it ignoring the subscript i.
It is not hard to see that if k = o(n/ǫ),
I(X : a, b) =Xi,j
=Xi,j
Pr((a, b) = (i, j)X = 0) + Pr((a, b) = (i, j)X = 1) (cid:19)
O(cid:18)(Pr((a, b) = (i, j)X = 0) − Pr((a, b) = (i, j)X = 1))2
(O(kǫ/n)max(2,i+j))2
i!j!Ω(min(k/n, 1)(1/(2))i+j )
= O(max(1, n/k))(ǫk/n)4.
If k = o(n), this is O(ǫ4k3/n3), so the total mutual information with the samples is O(ǫ4k3/n2),
which is o(1) if k = o(n2/3/ǫ4/3). Note that n2/3/ǫ4/3 is smaller than n3/4/ǫ if and only if it is less
than n, so the lower bound in proved in this case. Otherwise, if k > n and k = o(n3/4/ǫ), then the
mutual information is O(n(ǫk/n)4) = O(ǫ4k4/n3) = o(1), proving the other case of our bound.
i=1 ni/ǫ2 and n1/3
i (cid:16)Qd
j=1 nj(cid:17)1/3
lower bounds, namelyqQd
N =Qd
3.3 Lower Bound for High Dimensional Independence Testing We need to show two
/ǫ4/3. We can obtain both of these from
the lower bound constructions from the 2-variable case.
In particular, for the first bound, we
have shown that it takes this many samples to distinguish between the uniform distribution on
i=1 ni inputs (which is a product distribution), from a distribution that assigns probability
(1 ± ǫ)/N randomly to each input (which once renormalized is probably Ω(ǫ)-far from being a
product distribution). For the latter bound, we think of [n1]×···× [nd] as [ni]× ([n1]×··· [ni−1]×
[ni+1] × ··· × [nd]), and consider the lower bound construction for 2-variable independence testing.
It then takes at least n1/3
i N 1/3/ǫ4/3 samples to reliably distinguish a "YES" instance from a "NO"
instance. Note that in a "YES" instance the first and second coordinates are independent and
22
the distribution on the second coordinate is uniform. Therefore, in a "YES" instance we have a
d-dimensional product distribution. On the other hand, a "NO" instance is likely Ω(ǫ)-far from any
distribution that is a product distribution over just this partition of the coordinates, and therefore
Ω(ǫ) far from any d-dimensional product distribution. This completes the proof.
In particular,
3.4 Lower Bound for k-Histograms We can use the above construction to show that our
upper bound for k-histograms is in fact tight.
if we rewrite [n] as [k] × [n/k]
and let the intervals be given by the subsets [n/k] × {i} for 1 ≤ i ≤ k, we need to show that
Ω(max(√n/ǫ2, n1/3k1/3/ǫ4/3)) samples are required from a distribution p on [k] × [n/k] to dis-
tinguish between the coordinates of p being independent with the second coordinate having the
uniform distribution, and p being ǫ-far from any such distribution. We note that in the lower bound
distributions given for each part of our lower bound constructions for the independence tester, the
"YES" distributions all had uniform marginal over the second coordinate. Therefore, the same hard
distributions give a lower bound for testing k-histograms of Ω(max(√n/ǫ2, k2/3(n/k)1/3/ǫ4/3)) =
Ω(max(√n/ǫ2, n1/3k1/3/ǫ4/3)). This completes the proof.
Acknowledgment. We would like to thank Oded Goldreich for numerous useful comments and
insightful conversations that helped us improve the presentation of this work. We are grate-
ful to Oded for his excellent exposition of our reduction-based technique in his recent lecture
notes [Gol16b].
References
[ADJ+11] J. Acharya, H. Das, A. Jafarpour, A. Orlitsky, and S. Pan. Competitive closeness
testing. Journal of Machine Learning Research - Proceedings Track, 19:47–68, 2011.
[ADJ+12] J. Acharya, H. Das, A. Jafarpour, A. Orlitsky, S. Pan, and A. Suresh. Competitive
classification and closeness testing. In COLT, 2012.
[ADK15]
J. Acharya, C. Daskalakis, and G. Kamath. Optimal testing for properties of distribu-
tions. CoRR, abs/1507.05952, 2015.
[AJOS14] J. Acharya, A. Jafarpour, A. Orlitsky, and A. T. Suresh. Sublinear algorithms for outlier
detection and generalized closeness testing. In 2014 IEEE International Symposium on
Information Theory, pages 3200–3204, 2014.
[Bat01]
T. Batu. Testing Properties of Distributions. PhD thesis, Cornell University, 2001.
[BDKR02] T. Batu, S. Dasgupta, R. Kumar, and R. Rubinfeld. The complexity of approximating
entropy. In ACM Symposium on Theory of Computing, pages 678–687, 2002.
[BFF+01] T. Batu, E. Fischer, L. Fortnow, R. Kumar, R. Rubinfeld, and P. White. Testing
random variables for independence and identity. In Proc. 42nd IEEE Symposium on
Foundations of Computer Science, pages 442–451, 2001.
[BFR+00] T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White. Testing that distri-
In IEEE Symposium on Foundations of Computer Science, pages
butions are close.
259–269, 2000.
[BFR+13] T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White. Testing closeness of
discrete distributions. J. ACM, 60(1):4, 2013.
23
[BKR04] T. Batu, R. Kumar, and R. Rubinfeld. Sublinear algorithms for testing monotone and
unimodal distributions. In ACM Symposium on Theory of Computing, pages 381–390,
2004.
[BV15]
[Can15]
B. B. Bhattacharya and G. Valiant. Testing closeness with unequal sized samples.
CoRR, abs/1504.04599, 2015.
C. L. Canonne. A survey on distribution testing: Your data is big. but is it blue?
Electronic Colloquium on Computational Complexity (ECCC), 22:63, 2015.
[CDGR16] C. L. Canonne, I. Diakonikolas, T. Gouleakis, and R. Rubinfeld. Testing shape restric-
tions of discrete distributions. In 33rd Symposium on Theoretical Aspects of Computer
Science, STACS, pages 25:1–25:14, 2016.
[CDVV14] S. Chan, I. Diakonikolas, P. Valiant, and G. Valiant. Optimal algorithms for testing
closeness of discrete distributions. In SODA, pages 1193–1203, 2014.
[DDS+13] C. Daskalakis, I. Diakonikolas, R. Servedio, G. Valiant, and P. Valiant. Testing k-modal
distributions: Optimal algorithms via reductions. In SODA, pages 1833–1852, 2013.
[DKN15a]
I. Diakonikolas, D. M. Kane, and V. Nikishkin. Optimal algorithms and lower bounds
for testing closeness of structured distributions. In 56th Annual IEEE Symposium on
Foundations of Computer Science, FOCS 2015, 2015.
[DKN15b] I. Diakonikolas, D. M. Kane, and V. Nikishkin. Testing Identity of Structured Distribu-
tions. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, 2015.
[GGR98] O. Goldreich, S. Goldwasser, and D. Ron. Property testing and its connection to
learning and approximation. Journal of the ACM, 45:653–750, 1998.
[GMV09] S. Guha, A. McGregor, and S. Venkatasubramanian. Sublinear estimation of entropy
and information distances. ACM Trans. Algorithms, 5(4):35:1–35:16, November 2009.
[Gol16a] O. Goldreich. The uniform distribution is complete with respect to testing identity
to a fixed distribution. Electronic Colloquium on Computational Complexity (ECCC),
23:15, 2016.
[Gol16b] O. Goldreich. Lecture Notes on Property Testing of Distributions. Available at
http://www.wisdom.weizmann.ac.il/ oded/PDF/pt-dist.pdf, March, 2016.
[GR00]
[ILR12]
[LR05]
O. Goldreich and D. Ron. On testing expansion in bounded-degree graphs. Technical
Report TR00-020, Electronic Colloquium on Computational Complexity, 2000.
P. Indyk, R. Levi, and R. Rubinfeld. Approximating and Testing k-Histogram Distri-
butions in Sub-linear Time. In PODS, pages 15–22, 2012.
E. L. Lehmann and J. P. Romano. Testing statistical hypotheses. Springer Texts in
Statistics. Springer, 2005.
[LRR11] R. Levi, D. Ron, and R. Rubinfeld. Testing properties of collections of distributions.
In ICS, pages 179–194, 2011.
24
[NP33]
J. Neyman and E. S. Pearson. On the problem of the most efficient tests of statis-
tical hypotheses. Philosophical Transactions of the Royal Society of London. Series
A, Containing Papers of a Mathematical or Physical Character, 231(694-706):289–337,
1933.
[Pan08]
L. Paninski. A coincidence-based test for uniformity given very sparsely-sampled dis-
crete data. IEEE Transactions on Information Theory, 54:4750–4755, 2008.
[RRSS09] S. Raskhodnikova, D. Ron, A. Shpilka, and A. Smith. Strong lower bounds for approxi-
mating distribution support size and the distinct elements problem. SIAM J. Comput.,
39(3):813–842, 2009.
[RS96]
R. Rubinfeld and M. Sudan. Robust characterizations of polynomials with applications
to program testing. SIAM J. on Comput., 25:252–271, 1996.
[Rub12]
R. Rubinfeld. Taming big probability distributions. XRDS, 19(1):24–28, 2012.
[Val11]
[VV14]
P. Valiant. Testing symmetric properties of distributions.
40(6):1927–1968, 2011.
SIAM J. Comput.,
G. Valiant and P. Valiant. An automatic inequality prover and instance optimal identity
testing. In FOCS, 2014.
A Omitted Proofs from Section 3
A.1 Proof of Lemma 3.3 We note that
I(X : a) =Xℓ
O Pr(a = ℓ)(cid:18)1 −
Pr(a = ℓX = 0)
Pr(a = ℓX = 1)(cid:19)2! .
A simple computation yields that
Pr(a = ℓX = 0) = e−k/mn (k/mn)ℓ
ℓ!
,
Pr(a = ℓX = 1) =(cid:18)e−k/mn (k/mn)ℓ
ℓ!
(cid:19) e−kǫ/mn(1 + ǫ)ℓ + ekǫ/mn(1 − ǫ)ℓ
2
! .
Expanding the above out as a Taylor series in ǫ we note that the odd degree terms cancel.
We condition based on the size of k/mn. First, we analyze the case that k/mn ≤ 1.
Therefore, we can see that if ℓ ≤ 2
and for 2ǫ−1 ≥ ℓ ≥ 2,
e−kǫ/mn(1 + ǫ)ℓ + ekǫ/mn(1 − ǫ)ℓ
2
e−kǫ/mn(1 + ǫ)ℓ + ekǫ/mn(1 − ǫ)ℓ
2
(3)
! = 1 + O(cid:16)ǫ2(k/mn)2−ℓ(cid:17) ,
! = 1 + O(cid:0)ǫ2ℓ2(cid:1) .
25
Hence, we have that
I(X : a) ≤ O(ǫ4(k/mn)2) +
2ǫ−1
Xℓ=2
Pr(a = ℓ)O(ǫ4ℓ4) + Pr(a > 2ǫ−1)
= O(ǫ4(k/mn)2) + O(ǫ4E[a(a − 1) + a(a − 1)(a − 2)(a − 3)]) + (ǫk/(mn))1/ǫ
= O(ǫ4(k/mn)2),
where in the last step we use that E[a(a − 1) + a(a − 1)(a − 2)(a − 3)] = (k/mn)2 + (k/mn)4, and
the last term is analyzed by case analysis based on whether or not ǫ > (mn)−1/8.
For λ = k/mn ≥ 1, we note that the probability that a − λ > √λ log(mn) is o(1/(mn)). So,
it suffices to consider only ℓ at least this close to λ. We note that for ℓ in this range,
e±λǫ(1 ∓ ǫ)ℓ = exp(±ǫ(λ − ℓ) + O(λǫ2)) = 1 ± ǫ(λ − ℓ) + O(λǫ2).
This implies that
which completes the proof.
(cid:18)1 −
Pr(a = ℓX = 0)
Pr(a = ℓX = 1)(cid:19)2
= O(λ2ǫ4),
A.2 Proof of Lemma 3.7 As before,
I(X : A) =Xv
O Pr(A = v)(cid:18)1 −
Pr(A = vX = 0)
Pr(A = vX = 1)(cid:19)2! .
(4)
We break this sum up into pieces based on whether or not v1 ≥ 2.
If v1 < 2, note that Pr(A = vci = 1/k, X = 0) = Pr(A = vci = 1/k, X = 1). Therefore,
(cid:18)1 −
Pr(A = vX = 0)
Pr(A = vX = 1)(cid:19)2
≤(cid:18)1 −
Pr(A = vX = 0, ci = 1/n)
Pr(A = vX = 1, ci = 1/n)(cid:19)2
.
Hence, the contribution coming from all such v is at most
Pr(A1 = 1ci = 1/n)) max
v1=1
O(cid:18)1 −
Pr(A = vX = 0, ci = 1/n)
Pr(A = vX = 1, ci = 1/n)(cid:19)2
+O(cid:18)1 −
Pr(A = 0X = 0, ci = 1/n)
Pr(A = 0X = 1, ci = 1/n)(cid:19)2
.
Note that the ai,j are actually independent of each other conditionally on both X and ci. We have
by Equation (3) that
if v1 = 1, and
Pr(A = vX = 0, ci = 1/n)
Pr(A = vX = 1, ci = 1/n)
= exp(cid:0)O(ǫ2k2n−2m−1 + ǫ2kn−1m−1)(cid:1) ,
Pr(A = 0X = 0, ci = 1/n)
Pr(A = 0X = 1, ci = 1/n)
= exp(cid:0)O(ǫ2k2n−2m−1)(cid:1) .
We have that Pr(A1 = 1ci = 1/n) is the probability that a Poisson statistic with parameter
k/n(1 + O(ǫ)) gives 1, which is O(k/n). Therefore, the contribution to I(X : A) coming from these
terms is
O(ǫ4k4n−4m−2 + ǫ4k3n−3m−2) = o(kn−2) + o(n−1) = o(n−1).
26
Next, we consider the contribution coming from terms with v1 ≥ 2. We note that
Pr(A = v, ci = 1/kX = x)
is (for either x = 0 or x = 1)
ke−1
n
(m)−v1
However, Pr(A = v, ci = 1/nX = x) is at most
1
vi!
.
mYi=1
((1 + ǫ)k/nm)v1
1
vi!
.
mYi=1
This is at most 2k/n times Pr(A = v, ci = 1/kX = x). Therefore, we have that
Pr(A = v)(cid:18)1 −
Pr(A = vX = 0)
Pr(A = vX = 1)(cid:19)2
Pr(A = vX = 0, ci = 1/n)
Pr(A = vX = 1, ci = 1/n)(cid:19)2
.
= O(k/n) Pr(A = vci = 1/n)(cid:18)1 −
mYi=1
1
vi!
Pr(A = vX = 0, ci = 1/n) = e−k/n(k/nm)v1
Note that
and
Pr(A = vX = 1, ci = 1/n) ≥ e−k(1+ǫ)/n(k/nm)v1
1
vi! ≫ Pr(A = vX = 0, ci = 1/n).
mYi=1
Therefore,
Pr(A = vci = 1/n)(cid:18)1 −
Pr(A = vX = 0, ci = 1/n)
Pr(A = vX = 1, ci = 1/n)(cid:19)2
Xv
= Θ(I(A : Xci = 1/n)).
Finally, we have that
I(X : Aci = 1/n) ≤
mXj=1
I(X : ai,jci = 1/n) = O(k2ǫ4/(n2m)) ,
by Equation (4). This means that the contribution from these terms to I(X : A) is at most
and the proof is complete.
O(k3ǫ4/(n3m)) ,
27
|
1503.02413 | 1 | 1503 | 2015-03-09T09:59:58 | Stochastic Service Placement | [
"cs.DS",
"cs.NI",
"cs.PF"
] | Resource allocation for cloud services is a complex task due to the diversity of the services and the dynamic workloads. One way to address this is by overprovisioning which results in high cost due to the unutilized resources. A much more economical approach, relying on the stochastic nature of the demand, is to allocate just the right amount of resources and use additional more expensive mechanisms in case of overflow situations where demand exceeds the capacity. In this paper we study this approach and show both by comprehensive analysis for independent normal distributed demands and simulation on synthetic data that it is significantly better than currently deployed methods. | cs.DS | cs |
Stochastic Service Placement
Galia Shabtai, Danny Raz, and Yuval Shavitt
[email protected], [email protected], [email protected]
Tel-Aviv University and Technion
Abstract. Resource allocation for cloud services is a complex task due
to the diversity of the services and the dynamic workloads. One way to
address this is by overprovisioning which results in high cost due to the
unutilized resources. A much more economical approach, relying on the
stochastic nature of the demand, is to allocate just the right amount
of resources and use additional more expensive mechanisms in case of
overflow situations where demand exceeds the capacity. In this paper
we study this approach and show both by comprehensive analysis for
independent normal distributed demands and simulation on synthetic
data that it is significantly better than currently deployed methods.
1
Introduction
The recent rapid development of cloud technology gives rise to "many-and-
diverse" services being deployed in datacenters across the world. The allocation
of the available resources in the various locations to these services has a critical
impact on the ability to provide a ubiquitous cost-effective high quality service.
There are many challenges associated with optimal service placement due to
the large scale of the problem, the need to obtain state information, and the
geographical spreading of the datacenters and users.
One intriguing problem is the fact that the service resource requirement
changes over time and is not fully known at the time of placement. A popular
way of addressing this important problem is over-provisioning, that is allocating
resources for the peak demand. Clearly, this is not a cost effective approach as
much of the resources are unused most of the time. An alternative approach is
to model the service requirement as a stochastic process with known parameters
(these parameters can be inferred from historical data).
Much of the previous work that used stochastic demand modeling [1,2,3,4]
attempted to minimize the probability of an overflow event and not the cost of
this overflow. Their solution was based on solving the Stochastic Bin Packing
(SBP) problem, where the goal is to pack a given set of random items to a
minimum number of bins such that each bin overflow probability will not exceed
a given value.1
Thus, in these works there is no distinction between cases with marginal and
substantial overflow of the demand, and under this modeling one again prepares
1 [3,4] also looked at the online SBP problem, where the items (services in our case)
are assigned to some bin (datacenter) as they arrive.
for the worst and tries to prevent an overflow. However, in reality it is not cost
effective to allocate resources according to the worst case scenario. Instead, one
often wishes to minimize the cost associated with periods of insufficient resource
availability, which is commonly dealt with by either diverting the service request
to a remote location or dynamically buying additional resources. In both cases,
the cost associated with such events is proportional to the amount of unavailable
resources. Thus, our aim is to minimize the expected overflow of the demand over
time.
We deviate from previous work in two ways. First, We look at a stochastic
packing problem where the number of bins is given, e.g., a company already
has two datacenters where services can be placed, and one would like to place
services in these two datacenters optimally. Second, as we mentioned before, we
look at a more practical optimality criterion where we do not wish to optimize
the probability of an overflow, but instead its expected deviation.
To better understand this, assume that the resource we are optimizing is the
bandwidth consumption of the service in a datacenter where we have a prebooked
bandwidth for each datacenter. Since we are dealing with stochastic bandwidth
demand, with some probability the traffic will exceed the prebooked bandwidth
and will result with expensive overpayment that depends on the amount of
oversubscribed traffic. Obviously, in such a case minimizing the probability of
oversubscription may not give the optimal cost, and what we need to minimize is
the expected deviation from the prebooked bandwidth, thus we term our problem
SP-MED (stochastic packing with minimum expected deviation).
We analyze the case of independent normal distribution and develop algo-
rithms for the optimal partition between two or more datacenters that need not
be identical. We prove the correctness of our algorithm by separating the prob-
lem into two: one dealing with the continuous characteristics of the stochastic
objective function and the other dealing with the discrete nature of the combi-
natorial algorithm. This approach results in a clean and elegant algorithm and
proof.
In fact, our proof technique reveals that the algorithms we developed hold for
a large family of optimization criteria, and thus may be applicable for other opti-
mization problems. In particular we study two other natural cost functions that
correspond to minimizing overflow probability (rather than expected overflow
deviation). We show that these cost functions also fall into our general frame-
work, and therefore the same algorithm works for them as well! As will become
clear later the requirements from the cost functions we have are quite natural,
and we believe our methods will turn useful for many other applications.
2 Related work
Early work on VM placement (e.g., [5,6,7]) models the problem as a deter-
ministic bin packing problem, namely, for every service there is an estimate of
its deterministic demand. Stochastic bin packing was first suggested by Klein-
berg, Rabani and Tardos [1] for statistical multiplexing. [1] mostly considered
Bernoulli-type distributions. Goel and Indyk [2] further studied Poisson and ex-
ponential distributions. Wang, Meng and Zhang [3] suggested to model real data
with the normal distribution. Thus, the input to their stochastic packing problem
is n independent services each with demand requirement distributed according
to a distribution X (i) that is normal with mean µ(i) and variance V (i). The
output is some partition of the services to bins in a way that minimizes a target
function that differs from problem to problem.
A naive approach to such a problem is to reduce it to classical bin packing as
follows: for the i'th service define the effective size as the number e(i) such that
the probability that X (i) is larger than e(i) is small; then solve the classical bin
packing problem (or a variant of it) with item sizes e(1), . . . , e(n). However, [3]
showed this approach can be quite wasteful, mostly because it adds extra space
per service and not per bin. To demonstrate the issue, think about unbiased, in-
dependent coin tosses. The probability one coin toss significantly deviates from
its mean is 1, while the probability 100 independent coin tosses significantly
deviate from the mean is exponentially small. When running independent trials
there is a smoothing effect that considerably reduces the chance of high devi-
ations. This can also be seen from the fact that the standard deviation of n
n times the standard deviation of one
independent, identical processes is only
process, and so the standard deviation grows much slower than the number of
processes.
√
Breitgand and Epstein [4], building on [3], suggest an algorithm for stochas-
tic bin packing that takes advantage of this smoothing effect. The algorithm
assumes all bins have equal capacity. The algorithm first sorts the processes by
their variance to mean ratio (VMR), i.e., V (1)
µ(n) . Then the
algorithm finds the largest prefix of the sorted list such that allocating that set
of services to the first bin makes the probability the first bin overflows at most
p. The algorithm then proceeds bin by bin, each time allocating a prefix of the
remaining services on the sorted list to the next bin. [4] show that if we allow
fractional solutions, i.e., we allow splitting services between bins, the algorithm
finds an optimal solution, and also show an online, integral version that gives a
2-approximation to the optimum.
µ(2) ≤ . . . ≤ V (n)
µ(1) ≤ V (2)
3 The risk unbalancing principle: A bird's overview of
our technique
In this bird's overview we focus on the SP-MED problem, and in the next section
we generalize it to a much wider class of cost functions. As mentioned above
we are given k bins with capacities c1, . . . , ck. We are looking for a solution
that minimizes the expected deviation. As before, we study the case where the
stochastic demands are independent and normally distributed. We develop a
new general framework to analyze this problem that also sheds light on previous
work.
We first observe that an optimal fractional solution to SP-MED is also op-
timal for any two bins. This is true because the cost function is the sum of the
deviation of the two bins.
i∈I µ(i) and variance V1 =(cid:80)
distributed with mean µ1 =(cid:80)
µ = (cid:80)n
i=1 µ(i) and V = (cid:80)n
expected deviation of all the bins, and changing the internal allocation of two
bins only affects them and not the other bins. Thus, one possible approach to
achieve an optimal solution, is by repeatedly improving the internal division of
two bins. However, offhand, there is no reason to believe such a sequence of local
improvements efficiently converges to the optimal solution. Surprisingly, this is
indeed the case. Therefore, we first focus on the two bin case. Later (in Appendix
A) we will see that solving the k = 2 case implies a solution to the general case
of arbitrary k.
We recall that if we have n independent normally distributed services with
mean and variance (µ(i), V (i)), and we allocate the services with indices in I ⊆ [n]
to the first bin, and the rest to the second bin, then the first bin is normally
i∈I V (i), while the
second one is normally distributed with mean µ− µ1 and variance V − V1 where
i=1 V (i). We wish to minimize the total expected
Let us define the function Dev : [0, 1] × [0, 1] → R such that Dev(a, b) is the
total expected deviation when bin one is distributed with mean aµ and variance
bV and bin two with mean (1− a)µ and variance (1− b)V . Figure 1 (left) depicts
Dev(a, b) for two bins with equal capacity. Dev has a reflection symmetry around
the a = b line, that corresponds to the fact that we can change the order of the
bins as long as they have equal capacity and remain with a solution of the
same cost (see Appendix G.2 for a proof of a similar symmetry for the case of
different capacity bins). The points (0, 0) and (1, 1) correspond to allocating all
the services to a single bin, and indeed the expected deviation there is maximal.
2 ) is a saddle point (see Appendix G.3 for a proof). Figure 1
(middle) shows a zoom in around this saddle point. There are big mountains in
the lower left and upper right quarters that fall down to the saddle point, and
also valleys going down from the saddle point to the bottom right and top left.
Going back to Figure 1 (left) we see that for every fixed b the function is convex
in a, with a single minimum point, and all these minimum points form the two
valleys mentioned above.
The point ( 1
2 , 1
µ , b(i) = V (i)
service i with the pair (a(i) = µ(i)
We now go back and consider the input(cid:8)(µ(i), V (i))(cid:9)n
and [n] \ I, the expected deviation of this partition is Dev((cid:80)
Thus, the n input points induce 2n possible solutions PI = ((cid:80)
i=1. We can represent
V ). If we split the input to the sets I
i∈I b(i)).
i∈I b(i))
for each I ⊆ [n], and our task is to find the partition I that minimizes Dev.
Sorting the services by their VMR, is equivalent to sorting the vectors P (i) =
(a(i), b(i)) by their angle with the a axis. If the sorted list is P (1), . . . , P (n) with
increasing angles, looking for a solution in a partition that breaks the sorted
sequence to two consecutive parts finds a solution among (0, 0), P (1), P (1) +
P (2), . . . , P (1) + . . . + P (n) = (1, 1). If we connect a line between two consecutive
points in the above sequence we get the sorted path, connecting (0, 0) with (1, 1).
A crucial observation is that all the 2n possible solutions PI lie on or above
that sorted path. See Lemma 1 for a rigorous statement and proof and Figure
1 (right) for an example. The fact that the sorted path always lies beneath all
i∈I a(i),(cid:80)
i∈I a(i),(cid:80)
Fig. 1. The left figure depicts Dev(a, b) when µ = 160, V = 6400 and c1 = c2 = 100.
The middle figure is a zoom in around the saddle point ( 1
2 ). The white lines represent
the zero level sets of the partial derivatives. The right figure depicts Dev(a, b) with
the 210 possible partitions represented in orange. The bottom sorted path (that sorts
services by their VMR in increasing order) and the upper sorted path (that sorts
services by their VMR in decreasing order), together with their integral points, are in
black. Notice that all partition points are confined by both sorted paths.
2 , 1
possible solutions is independent of the target function, and only depends on the
fact that we deal with the normal distribution. Actually, this is true for any
distribution where allocating a subset of services to a bin amounts to adding the
corresponding means and variances.
A second crucial observation, proved in Theorem 1, is that the optimal frac-
tional solution lies on the sorted path. In fact, the proof shows this is a general
phenomenon that holds for many possible target functions. What we need from
the cost function Dev we are trying to minimize is:
-- Dev has a symmetry of reflection around a line that passes through a saddle
point (a(cid:48), 1
2 ).
-- For every fixed b, Dev is strictly uni-modal in a (i.e., has a unique minimum
in a). We call the set of solutions {(m(b), b)} the valley. The saddle point
lies on the valley.
2 and b ≥ 1
-- Dev restricted to the valley is strictly monotone for b ≤ 1
maximum at the saddle point that is obtained when b = 1
2 .
2 with a
To see why these conditions suffice, consider an arbitrary fractional solution
(a, b). By the symmetry property we may assume without loss of generality that
b ≤ 1
2 . If (a, b) is left to the valley, we can improve the solution by keeping b and
moving a towards the valley, until we either hit the sorted path or the valley. If
we hit the valley first, we can improve the expected deviation, by going down
the valley until we hit the sorted path. A similar argument applies when (a, b) is
right to the valley. The conclusion is that we can always find another fractional
solution that is better than (a, b) and lies on the sorted path.
Thus, quite surprisingly, we manage to decouple the question to two separate
and almost orthogonal questions. The first question is what is the behavior of
the function Dev over its entire domain. Notice that Dev is a function of only
(cid:8)(µ(i), V (i))(cid:9) at all. We study Dev using analytic tools. The second question is
two variables, independent of n. This question does not depend on the input
what is the geometric structure of the set of feasible points, and here we inves-
tigate the input, completely ignoring the target function Dev. For this question
we use geometric intuition and combinatorial tools.
In Appendix B we prove our three cost functions fall into the above frame-
work. The discussion above also gives an intuitive geometric interpretation of
the results in [4,3]. We believe the framework is also applicable to many other
target functions.
There is an intuitive explanation why the optimal fractional solution lies on
the sorted path. Sorting services by their VMR essentially sorts them by their
risk, where risk is the amount of variance per one unit of expectation. Partition-
ing the sorted list corresponds to putting all the low risk services in one bin, and
all the high risk services in the other. We call this the risk unbalancing principle.
Intuitively, we would like to give the high risk services as much spare capacity
as possible, and we achieve that by grouping all low risk services together and
giving them less spare capacity. In contrast, balancing risk amounts to taking
the point ( 1
2 ) (for the case where the two bins have equal capacities). Having
the geometric picture in mind, we immediately see that this saddle point is not
optimal, and can be improved by taking the point on the valley that intersects
the sorted path.
2 , 1
The technique also applies to bins having different capacities that previous
work did not analyze, and reveals that we should allocate low risk services to
the bin with the lower capacity. This follows from a neat geometric argument
that when c1 < c2 the optimal fractional solution lies on the sorted path (that
sorts services by their VMR in increasing order) and not the upper sorted path
(that sorts services by their VMR in decreasing order). See Section 5.2 and
Theorem 1 for more details. If we have k bins, we should sort the bins by their
capacity and the services by their VMR, and then allocate consecutive segments
of the sorted list of services to the bins, with lower risk services allocated to
smaller capacity bins. This double-sorting again intuitively follows from the risk
unbalancing principle, trying to preserve as much spare capacity to the high risk
services.2 We prove the double-sorting algorithm in Appendix A.
Finally, we are left with the question of finding the right partition of the
sorted list. With two bins one can simply try all n partition points. However,
(cid:1) = Θ(nk−1) possible partition points. We show a
with k bins there are (cid:0) n
k−1
dynamic programming algorithm that finds the best partition in poly(n) time.
The optimal solution can probably be found much faster and we have candidate
2 In a similar vein if we have total capacity c to split between two bins, it is always
better to make the bin capacities unbalanced as much as possible, i.e., the minimum
expected deviation decreases as c2 − c1 increases. In particular the best choice is
having a single bin with capacity c and the worst choice is splitting the capacities
evenly between the two bins. Obviously, in practice there might be other reasons to
have several bins, but if there is tolerance in each bin capacity it is always better to
minimize the number of bins. We give a precise statement and proof in Appendix J.
algorithms that work well in practice and we are currently working on formally
proving their correctness.
4 Formal treatment
i=1, specifying the bin capacities, and values(cid:8)(µ(i), V (i))(cid:9)n
The input to the problem consists of k and n, specifying the number of bins and
services, integers {ci}k
where the demand distribution X (i) of service i is normal with mean µ(i) and
variance V (i). X (i) are independent. The output is a partition of [n] to k disjoint
sets S1, . . . , Sk ⊆ [n], where Sj includes indices of services that needs to be allo-
cated to bin j. Our goal is to find a partition minimizing a given cost function
D(S1, . . . , Sk). The optimal (integral) partition is the partition with minimum
cost among all possible partitions.
variance, i.e., c =(cid:80)k
j=1 cj, µ =(cid:80)n
i=1 µ(i) and V =(cid:80)n
i=1,
√
i∈Sj
i∈Sj
V (i). The
We let c denote the total capacity, µ the total demand and V the total
i=1 V (i). The value c − µ
represents the total spare capacity we have. The (total) standard deviation is
V which represents the standard deviation of the input had all services
σ =
been put into a single bin. We let ∆ denote the spare capacity in units of the
standard deviation, i.e., ∆ = c−µ
with mean µ(i) and variance V (i) is normal with mean (cid:80) µ(i) and variance
(cid:80) V (i). Consider a partition S1, . . . , Sk. Then, the demand distribution of bin
j is normal with mean µj = (cid:80)
standard deviation of bin j is σj = (cid:112)Vj and its spare capacity in units of its
µ(i) and variance Vj = (cid:80)
We remind the reader that the sum of independent normal distributions
σ . We assume that c (cid:62) µ.
standard deviation is ∆j = cj−µj
. Following previous work, we assume that for
every i, µ(i) ≥ 0 and V (i) is small enough compared with (µ(i))2 so that the
probability of getting negative demand in service i is negligible.
j=1 aj ≤ 1 and (cid:80)k−1
j=1 aj and bk = 1 −(cid:80)k−1
s.t., (cid:80)k−1
let ak = 1 −(cid:80)k−1
j = 1, . . . , k and we define σj =(cid:112)Vj =(cid:112)bjσ (where σ =
(cid:80)k−1
j=1 aj ≤ 1 and(cid:80)k−1
Next we normalize everything with regard to the total mean µ and the total
variance V . The input to the function D are two vectors a1, . . . , ak−1, b1, . . . , bk−1
j=1 bj ≤ 1. We think of the vectors a1, . . . , ak−1 and
b1, . . . , bk−1 as representing the fraction of mean and variance each bin takes. We
j=1 bj be the fraction of the last bin (that
takes all the remaining mean and variance). We let µj = ajµ and Vj = bjV for
V ) and ∆j = cj−µj
.
The function D is defined over all tuples a1, . . . , ak−1, b1, . . . , bk−1 ∈ [0, 1] s.t.,
j=1 bj ≤ 1. There are kn possible partitions that correspond
to the kn (possibly) different tuple values. We call a tuple integral if it is induced
by some partition. Our goal is to approximate the integral solution minimizing
D over all integral inputs.
In this notation the cost function is D(a1, . . . , ak−1; b1, . . . , bk−1).
σj
√
σj
4.1 What do we require form a cost function?
We can handle a wide variety of cost functions. Specifically, we require the fol-
lowing from the cost function:
1. In any solution to a k-bin problem, the allocation for any two bins is also
optimal. Formally, if S1, . . . , Sk is optimal for a k-bin problem, then for any
j and j(cid:48), the partition Sj, Sj(cid:48) is optimal for the two-bin problem defined by
the services in Sj ∪Sj(cid:48) and capacities cj, cj(cid:48). We remark that this is a natural
condition that is true for almost any cost function we know.
For k = 2 we require that:
2. D(a, b) = D(1 − a − c2−c1
µ , 1 − b). When c1 = c2 this simply translates to
D(a, b) = D(1 − a, 1 − b) and there is no difference between allocating the
set S1 to the first bin or to the second one.
3. For every fixed b ∈ [0, 1], D(a, b) has a unique minimum on a ∈ [0, 1], at
some point a = m(b), i.e. D is decreasing at a < m(b) and increasing at
a > m(b). We call the points on the curve {(m(b), b)} the valley.
2 ), 1
4. D has a unique maximum over the valley at the point (m( 1
the symmetry above this point is ( 1
is increasing for b ≤ 1
2 ). In fact by
2µ , 1
2 ). This means that D(m(b), b)
2 and decreasing for b ≥ 1
2 .
2 − c2−c1
4.2 Three cost functions
We have already seen SP-MED, which minimizes the expected deviation. Previ-
ous work on SBP (stochastic bin packing) studied the worst overflow probability
of a bin. We define SP-MWOP to be the problem that minimizes the worst
overflow probability of a bin. Another natural problem is what we call SP-MOP
that minimizes the probability that some bin overflows. The three cost functions
are related but behave differently. For example, SP-MWOP is not differentiable
but the other two are. It is relatively simple to find the valley in SP-MED and
SP-MWOP but we are not aware of any explicit description of the valley in
SP-MOP.
Nevertheless, it is remarkable that all three cost functions fall into our frame-
work and we prove that in Appendix B. The proof is not always simple, and we
have used Lagrange multipliers and the log concavity of the cumulative distribu-
tion function of the normal distribution for proving this for SP-MOP. It follows
that all the machinery we develop in the paper holds for these cost functions.
5 The two bin case
We first consider the two bin case (k = 2). We shall later see (in Appendix A)
that solving the k = 2 case is key for solving the general problem for any k > 2.
As explained before we decouple the question to one about the function D and
another about the structure of the integral points.
5.1 The sorted path
We now consider the input (µ(1), V (1)), . . . , (µ(n), V (n)). We represent service
i with the pair (a(i) = µ(i)
V ). If we split the input to the sets I
µ , b(i) = V (i)
and [n] \ I, then the first bin is normally distributed with mean µ(cid:80)
and variance V (cid:80)
i∈I a(i),(cid:80)
PI = ((cid:80)
i∈I a(i)
i∈I b(i). Thus, the n input points induce 2n possible solutions
i∈I b(i)) for each I ⊆ [n] and we call each such point an
integral point. Sorting the services by their VMR, is equivalent to sorting the
vectors P (i) = (a(i), b(i)) by the angle they make with the a axis.
Definition 1. (The sorted path) Sort the services by their VMR in increasing
order and calculate the P (1), P (2), . . . , P (n) vectors. For i = 1, . . . , n define
P [i]
bottom = P (1) + P (2) + . . . + P (i) and,
up = P (n) + P (n−1) + . . . + P (n−i+1),
P [i]
and also define P [0]
bottom = P [0]
up = (0, 0).
The bottom sorted path is the curve that is formed by connecting P [i]
bottom with a line, for i = 0, . . . , n− 1. The upper sorted path is the curve
with a line, for i = 0, . . . , n− 1. We
and P [i+1]
that is formed by connecting P [i]
sometimes abbreviate the bottom sorted path and call it the sorted path.
up and P [i+1]
bottom
up
The integral point P [i]
bottom on the bottom sorted path corresponds to allocat-
ing the i services with the lowest VMR to the first bin and the rest to the second.
On the other hand, the integral point P [i]
up on the upper sorted path corresponds
to allocating the i services with the highest VMR to the first bin and the rest
to the second. A crucial, yet simple, observation (proven in Appendix D):
Lemma 1. All the integral points lie within the polygon confined by the bottom
sorted path and the upper sorted path.
Lemma 1 is a key lemma of the paper. Unfortunately, we had to omit the
proof because of lack of space. We recommend the reader to look at the simple
proof that appears in Appendix D. In fact,
Lemma 2. The set of fractional points coincides with the polygon confined by
the bottom sorted path and the upper sorted path.
5.2 The sorting algorithm
n,(cid:8)(µ(i), V (i))(cid:9)n
We now present the algorithm for the k = 2 case. The algorithm for the general
case of arbitrary k is presented in Appendix A. The algorithm gets as input
i=1 , c1, c2 and outputs a partition S1, S2 of [n] minimizing cost.
We keep all notation as before. The algorithm works as follows:
The sorting algorithm: 2 bins
-- Sort the bins by their capacity such that c1 ≤ c2.
-- Sort the services by their VMR such that v(1)
-- Calculate the points P [0] = (0, 0), P [1], . . . , P [n] = (1, 1) on the sorted path.
-- Calculate D(P [i]) for each 0 ≤ i ≤ n and find the index i∗ such that the
-- Let S1 = {1, . . . , i∗} and S2 = [n] \ S1. Output (S1, S2).
point P [i∗] achieves the minimal cost among all points P [i].
µ(2) ≤ ··· ≤ v(n)
µ(n) .
µ(1) ≤ v(2)
The optimal fractional point is the fractional point that minimizes cost. In
Section 3 we said, and soon will prove, that the optimal solution allocates low
risk services to one bin and the rest to the other. However, offhand, it is not
clear whether to allocate the smaller risk services to the lower capacity bin or
the higher capacity bin. In fact, offhand, it is not clear whether the optimal
solution is on the lower sorted path or the upper sorted path, and it might even
depend on the input. Remarkably, we prove that it always lies on the bottom
sorted path, meaning that it is always better to allocate low risk services to the
smaller capacity bin and high risk services to the higher capacity bin. We gave
an intuitive explanation to this phenomenon in Section 3. We prove Theorem 1
in Appendix E.
Theorem 1. The optimal fractional point lies on the bottom sorted path. The
optimal fractional solution splits at most one service between two bins.
6 Simulation results
In this section we present our simulation results for the two bin and k bin cases.
We run our simulation on bins with equal capacity. We compare the sorting
algorithms for SP-MED, SP-MWOP and SP-MOP to an algorithm we call BM
(Balanced Mean). The algorithm goes through the list, item by item, and allo-
cates each item to the bin which is less occupied, which is a natural benchmark
and also much better than other naive solutions like first-fit and first-fit decreas-
ing. 3
We show simulation results on synthetic data. We first generate our stochastic
i=1 for n = 100 and for n = 500. Our sample space is a
mixture of three populations: all items have the same mean (we fixed it at
input (cid:8)((cid:101)µ(i),(cid:101)σ(i))(cid:9)n
(cid:101)µ(i) = 500) but 50% had standard deviation picked uniformly from [0, 0.1 ·(cid:101)µ(i)],
25% had standard deviation picked uniformly from [0.1 ·(cid:101)µ(i), 0.5 ·(cid:101)µ(i)] and 25%
had standard deviation picked uniformly from [0.5 ·(cid:101)µ(i), 0.9 ·(cid:101)µ(i)].
1 ≤ l ≤ 500 using the normal distribution N [(cid:101)µ(i),(cid:101)V (i)] and from this we inferred
parameters µ(i), V (i), best explaining the sample as a normal distribution. Both
We then randomly generated 500 sample values x(i)
for each 1 ≤ i ≤ n and
l
3 At first, we also wanted to compare our algorithm with variants of the algorithms
considered in [3,4] for the SBP problem. In both papers, the authors consider the
algorithms First Fit and First Fit Decreasing [8] with item size equal to the effective
size, which is the mean value of the item plus an extra value that guarantees an
overflow probability be at most some given value p. Their algorithm chooses an
existing bin when possible, and otherwise opens a new bin. However, when the
number of bins is fixed in advance, taking effective size rather than size does not
change much. For a new item (regardless of its size or effective size) we keep choosing
the bin that is less occupied, but this time we measure occupancy with respect to
effective size rather than size. Thus, if elements come in a random order, the net
outcome of this is that the two bins are almost balanced and a new item is placed
in each bin with almost equal probability.
Fig. 2. Average cost of the sorting algorithm and the BM algorithm for SP-MED,
SP-MWOP and SP-MOP with two bins. The x axis measures c
µ .
Fig. 3. Average cost of the BM algorithm divided by average cost of the sorting algo-
rithm for SP-MED, SP-MWOP and SP-MOP with 2 bins. The x axis measures c
µ .
the sorting algorithm and the BM algorithm got as input (cid:8)(µ(i), V (i))(cid:9)n
i=1, as
well as c1 = . . . = ck = c
k and output their partition.
To check the suggested partitions, we viewed each sample x(i)
as represent-
ing an item instantiation in a different time slot. We then computed the cost
function. For example, for SP-MED, the deviation value for bin j at time slot l
l
(cid:26)
(cid:27)
(cid:80)
(cid:80)n
i∈Sj
l −cj
x(i)
i=1 µ(i)
is: max
0, 100
, i.e., the deviation is measured as a percent of the
total mean value µ. We generated 20 such lists and calculated the average cost
for these 20 input lists for each algorithm. We run the experiment for different
values of c.
Figure 2 shows the 2 bins average cost of both algorithms for SP-MED, SP-
MWOP and SP-MOP as a function of c
µ . As expected, the average cost decreases
as the value c
µ increases, i.e. as the total spare capacity increases. We also see
that the sorting algorithm out-performs BM. Figure 3 shows the average cost of
the BM algorithm divided by the average cost of the sorting algorithm for the
three problems, again as a function of c
µ .
Figure 4 shows the 4 bins average cost and cost ratio of both algorithms for
SP-MED as a function of c
µ for n = 100 and n = 500. As we saw in the two bins
case, the average cost decreases as the value c
µ increases. We can also see that the
average deviation for n = 100 is higher than the deviation for n = 500. However,
the ratio between the BM algorithm average cost and the sorting algorithms
average cost for n = 100 is lower than the ratio we get for n = 500. Still, in both
n values, the sorting algorithm out-performs the BM.
Fig. 4. The left figure depicts the average cost of the sorting algorithm and the BM
algorithm for SP-MED with four bins, both for n = 100 and for n = 500. The right
figure depicts the average cost of the BM algorithm divided by the average cost of the
sorting algorithm for n = 100 and for n = 500. The x axis measures c
µ .
References
1. Kleinberg, J., Rabani, Y., Tardos, ´E.: Allocating bandwidth for bursty connections.
SIAM Journal on Computing 30 (2000) 191 -- 217
2. Goel, A., Indyk, P.: Stochastic load balancing and related problems. In: Foundations
of Computer Science, 1999. 40th Annual Symposium on, IEEE (1999) 579 -- 586
3. Wang, M., Meng, X., Zhang, L.: Consolidating virtual machines with dynamic
bandwidth demand in data centers. In: INFOCOM, 2011 Proceedings IEEE, IEEE
(2011) 71 -- 75
4. Breitgand, D., Epstein, A.: Improving consolidation of virtual machines with risk-
aware bandwidth oversubscription in compute clouds. In: INFOCOM, 2012 Pro-
ceedings IEEE, IEEE (2012) 2861 -- 2865
5. Ajiro, Y., Tanaka, A.: Improving packing algorithms for server consolidation. In:
Int. CMG Conference. (2007) 399 -- 406
6. Bobroff, N., Kochut, A., Beaty, K.: Dynamic placement of virtual machines for
managing sla violations. In: Integrated Network Management, 2007. IM'07. 10th
IFIP/IEEE International Symposium on, IEEE (2007) 119 -- 128
7. Mehta, S., Neogi, A.: Recon: A tool to recommend dynamic server consolidation in
multi-cluster data centers. In: Network Operations and Management Symposium,
2008. NOMS 2008. IEEE, IEEE (2008) 363 -- 370
8. Johnson, D.: Computers and intractability-a guide to the theory of np-completeness
(1979)
A The general case
We now analyze the general k bin case using the results we have obtained for
the two-bin case. We first show the optimal factional solution is obtained by
first sorting the bins by their capacities and the services by their VMR ratio,
and then partitioning the sorted list to k consecutive parts.
Assume the bins are sorted by their capacity, c1 ≤ c2 ≤ . . . ≤ ck. A fractional
solution is called sorted if for every j < j(cid:48) and every two services i and i(cid:48) such
that service i (resp. i(cid:48)) is allocated to bin j (resp. j(cid:48)) it holds that the VMR of
service i is at most that of service i(cid:48).
Theorem 2. The optimal fractional solution is sorted.
Proof. Assume there is an optimal solution S1, . . . , Sk that is not sorted. I.e.,
there exist j < j(cid:48) and i, i(cid:48) such that service i (resp. i(cid:48)) is allocated to bin j
(resp. j(cid:48)) and the VMR of service i is strictly larger than that of service i(cid:48). By
Theorem 1 the sorted fractional solution for the problem defined by bins j and
j(cid:48) is strictly better than the one offered by the solution S1, . . . , Sk. This means
the solution is not optimal for bins j and j(cid:48) - in contradiction to item (1) in
Section 4.1.
The fact that with two bins and different capacities c1 < c2, low risk services
should be allocated to the smaller capacity bin (See Theorem 1 and the discussion
before it) implies also that for k bins lower risk services should be allocated to
lower capacity bins.
Next we present an algorithmic framework for the problem:
The double sorting algorithm framework
-- Sort the bins by their capacity c1 ≤ c2 ≤ . . . ≤ ck.
-- Sort the services by their VMR v(1)
-- Use a partitioning algorithm to find k − 1 partition points (cid:96)0 = 0 ≤ (cid:96)1 <
. . . (cid:96)k−1 ≤ (cid:96)k = n and output S1, . . . , Sk where Sj = {(cid:96)j−1 + 1, . . . , (cid:96)j}.
µ(2) ≤ ··· ≤ v(n)
µ(n) .
µ(1) ≤ v(2)
-- Allocate the services in Sj to bin number j (with capacity cj).
The algorithmic framework is not complete since it does not specify how to
find the partition points on the sorted path. With two bins there is only one
partition point, and we can check all possible n − 1 integral partition points.
(cid:1) possible partition points, and checking all partition
With k bins there are(cid:0) n
k−1
points may be infeasible.
We now describe a dynamic programming algorithm that solves the problem,
and works under very general conditions.
Finding an optimal integral partition
For each 1 ≤ t ≤ log k, 1 ≤ i < i(cid:48) ≤ n and j = 1, . . . , k keep the two values
P artition(2t, i, i(cid:48), j) and D(2t, i, i(cid:48), j) that are defined as follows:
-- Base case t = 1: P artition(2, i, i(cid:48), j) is the best integral partition point
for the two bin problem with inputs Xi, . . . , Xi(cid:48) and capacities cj, cj+1.
D(2, i, i(cid:48), j) is its corresponding expected deviation.
-- Induction step. Suppose we have built the tables for t, we show how to build
the tables for t + 1. We let Dev(2t+1, i, i(cid:48), j) be
(cid:8)D(2t, i, i(cid:48)(cid:48) − 1, j) + D(2t, i(cid:48)(cid:48), i(cid:48), j + 2t)(cid:9)
min
i≤i(cid:48)(cid:48)≤i(cid:48)
and we let P artition(2t+1, i, i(cid:48), j) be the partition point that obtains the
minimum in the above equation.
The optimal partition to k bins (assuming k is a power of two) can be recovered
from the tables. For example for 4 bins P artition(4, 1, n, 1) returns the middle
point (cid:96)2 of the partition, and the partition points within each half are obtained
by (cid:96)1 = P artition(2, 1, (cid:96)2 − 1, 1) and (cid:96)3 = P artition(2, (cid:96)2, n, 3).
It can be seen (by a simple induction) that P artition(2t, i, i(cid:48), j) gives the
middle point of the optimal integral solution on the sorted path to the problem
of dividing Xi, . . . , Xi(cid:48) to 2t bins with capacities cj, cj+1, . . . , cj+2t−1. It can also
be verified that the running time of the algorithm is O(n3k log k).
We analyze the error the integral solution we find has compared with the
optimal fractional solution. The error analysis builds upon the error analysis of
the two bin case. Start with an optimal fractional solution π. By Theorem 2 we
know π is sorted. π defines some sorted fractional solution to the problem defined
by the first two bins. By the remark after Theorem 3 we know that one of the
integral points to the left or to the right of the fractional solution has a very close
deviation to that of π, and we fix the first partition point accordingly to make it
integral. Now assume after fixing j stages we hold some fractional solution π(j)
that is integral on the first j partition points and possibly fractional on the rest.
We look at bins j + 1 and j + 2 and find an integral partition point (to the left or
to the right of the next fractional partition point) that is almost as good as the
fractional one. Doing the above for j = 1, . . . , k − 1 we end up with an integral
solution whose deviation is close to that of the optimal fractional solution, where
the error bound in Theorem 3 is multiplied by a factor of k − 1. We remark that
this also shows an efficient way of converting an optimal fractional solution to
an integral solution close to it.
The algorithm is quite general and works whenever we can solve the k = 2
case, and therefore may be seen as a reduction from arbitrary k to the k =
2 case. The main point we want to stress is that even in this generality the
complexity is polynomial. The main disadvantage of the algorithm is that it
runs in about n3 time. For specific cases (such as SP-MED and SP-MOP and
probably other natural variants) the optimal solution may be probably found
much faster, and we are currently working on proving the correctness of an
almost linear algorithm.
B Three cost functions
B.1 Stochastic Packing with Minimum Expected Deviation
(SP-MED)
In Appendix G.1 we prove the expected deviation of bin j (j = 1, . . . , k), denoted
by DevSj , is
DevSj = σj[φ(∆j) − ∆j(1 − Φ(∆j))],
where φ is the probability density function (pdf) of the standard normal distri-
bution and Φ is its cumulative distribution function (CDF).4 Denoting g(∆) =
φ(∆) − ∆(1 − Φ(∆)) we see that DevSj = σj g(∆j). With two bins Dev is a
function from [0, 1]2 to R and Dev(a, b) = σ1g(∆1) + σ2g(∆2) where the first bin
has mean aµ and variance bV , the second bin has mean (1 − a)µ and variance
(1 − b)V and σj, ∆j are defined as above.
Lemma 3. Dev respects conditions (1)-(4) of Section 4.1.
Proof. We go over the conditions one by one.
item (1): Let S1, . . . , Sk be an optimal solution. Suppose there are j and j(cid:48) for
which Sj, Sj(cid:48) are not the optimal solution for the two-bin problem defined
by the services in Sj ∪ Sj(cid:48) and capacities cj, cj(cid:48). Change the allocation of the
solution S1, . . . , Sk on bins j and j(cid:48) to an optimal solution for the two bin
problem. This change improves the expected total deviation of the two bins
while not affecting the expected deviation of any other bin. In total we get
a better solution, in contradiction to the optimality of S1, . . . , Sk.
item (2): In Appendix G.2 we prove Dev(a, b) = Dev(1− a− c2−c1
µ , 1− b). We
remark that for c1 = c2 this simply says we can switch the names of the first
and second bin.
] ≥ 0. It
follows that for any 0 < b < 1, Dev(a) is convex and has a unique minimum.
The unique point (m(b), b) on the valley is the one where ∆1 = ∆2.
item (3): In Appendix G.3 we calculate ∂2Dev
∂a2 = µ2 [ φ(∆2)
σ2
+ φ(∆1)
σ1
item (4): We first explicitly determine what Dev restricted to the valley is as
a function D(b) = Dev(m(b), b) of b. As Dev(a, b) = σ1g(∆1) + σ2g(∆2) and
on the valley ∆1 = ∆2 we see that on the valley Dev(a, b) = (σ1 + σ2)g(∆1).
However, σ1 + σ2 also simplifies to c1−aµ
. Altogether,
+ c2−(1−a)µ
= c−µ
∆1
∆2
∆1
4 A quick background on the normal distribution is given in Appendix F.
∆1
∂∆1
∂b = ∂Dev
∂∆1
is a function of ∆1
= −(c− µ) φ(∆1)
we conclude that on the valley Dev(a, b) = (c − µ) g(∆1)
alone.
Now it is a straight forward calculation that ∂Dev(∆1)
< 0.
2 and positive when b ≥ 1
is negative when b ≤ 1
Also ∂∆1
2 (see Appendix
∂b
· ∂∆1
∂b , we see that D(b) is increasing for b ≤ 1
G.3 and G.3). As ∂D
2
2 − c2−c1
and decreasing for b ≥ 1
2µ , b = 1
2 )
lies on the valley and is a maximum point for Dev restricted to the valley.
We remark that we could simplify the proof by using Lagrange multipliers
(as we do in Lemma 8). However, since here it is easy to explicitly find Dev
restricted to the valley we prefer the explicit solution. Later, we will not
be able to explicitly find the restriction to the valley and we use instead
Lagrange multipliers that solves the problem with an implicit description of
the valley.
2 as claimed. The saddle point (a = 1
∆2
1
Lemma 4. The difference between the expected deviation in the integral point
found by the sorting algorithm and the optimal integral (or fractional) point for
SP-MED is at most
n · µ. In particular, when L = o(cid:0)n(cid:1) and α is a
constant, the error is o(cid:0)µ(cid:1).
· 1
α · L
1√
2πe
Proof. We know from Theorem 3 that the difference is at most
min{∇D(ξ1),∇D(ξ2)} L
n
,
where ξ1 ∈ [O1, OP Tf ], ξ2 ∈ [OP Tf , O2] and O1 and O2 are the two points on
the bottom sorted path between which OP Tf lies. Using the partial derivatives
calculated in Appendix G.3 we see that
∇(σ2g)(∆2) (cid:54) µ (1 − Φ(∆2)) +
√
2
σ
1 − b
φ(∆2) ≤ µ +
√
2
σ
1 − b
φ(∆2).
Moreover,
√
σ
1−b
2
√
φ(∆2) = σ
2
1−b
1
∆2
∆2φ(∆2) and a simple calculation shows
. By our
1√
that the function ∆φ(∆) maximizes at ∆ = 1 with value at most
assumption that ∆j ≥ 0 for every j, we get that
2πe
√
2
σ
1 − b
φ(∆2) (cid:54)
√
2
σ
1 − b
√
σ
1 − b
c2 − (1 − a)µ
1√
2πe
(cid:54)
√
V
2πe
2
1
c2 − (1 − a)µ
.
1
c1−aµ .
√
V
2πe
2
Applying the same argument on O2 shows the error can also be bounded by
However, (c1 − aµ) + (c2 − (1− a)µ) = c− µ which is the total spare capacity,
and at least one of the bins takes spare capacity that is at least half of that,
namely c−µ
2 . Since the error is bounded by either term, we can choose the one
where the spare capacity is at least c−µ
2 and we therefore see that the error is at
c−µ . Since we assume c− µ ≥ αµ for some constant α > 0, the error
most
µ ≤ µ which completes the proof.
V√
αµ . As we assume V ≤ µ2, V
is at most
√
V
2πe
2
2
1
2πe
This shows the approximation factor goes to 1 and linearly (in the number of
services) fast. Thus, from a practical point of view the theorem is very satisfying.
B.2 Stochastic Packing with Min Worst Overflow Probability
(SP-MWOP)
ues(cid:8)(µ(i), V (i))(cid:9)n
The SP-MWOP problem gets as input integers k and n, specifying the number
of bins and services, integers c1, . . . , ck, specifying the bin capacities and val-
i=1, specifying that the demand distribution X (i) of service i is
normal with mean µ(i) and variance V (i). A solution to the problem is a parti-
tion of [n] to k disjoint sets S1, . . . , Sk ⊆ [n] that minimizes the worst overflow
probability.
The SP-MWOP problem is a natural variant of SBP. For a given partition
let OF Pj (for j = 1, . . . , k) denote the overflow probability of bin j. Let W OF P
j=1 {OF Pj}. In the
denote the worst overflow probability, i.e., W OF P = maxk
SBP problem we are given n normal distributions and wish to pack them into
few bins such that the OF P ≤ k for some given parameter p. Suppose we
solve the SBP problem for a given p and know that k bins suffice. We now ask
ourselves what is the minimal W OF P achieved with the k bins (this probability
is clearly at most p but can also be significantly smaller). We also ask what is
the partition that achieves this minimal worst overflow probability. The problem
SP-MOP does exactly that.
OF Pj = 1 − Φ(∆j) where ∆j = cj−µj
In Appendix H.1 we prove the overflow probability of bin j (j = 1, . . . , k), is
. Thus,
σj
W OF P =
{1 − Φ(∆j)} .
k
max
j=1
With two bins W OF P is a function from [0, 1]2 to R and W OF P (a, b) =
max{1 − Φ(∆1), 1 − Φ(∆2)} where the first bin has mean aµ and variance bV ,
the second bin has mean (1 − a)µ and variance (1 − b)V and σj, ∆j are defined
as above.
Lemma 5. W OF P respects conditions (1)-(4) of Section 4.1.
Proof. We go over the conditions one by one.
item (1): Let S1, . . . , Sk be an optimal solution. Suppose there are j and j(cid:48) for
which Sj, Sj(cid:48) are not the optimal solution for the two-bin problem defined
by the services in Sj ∪ Sj(cid:48) and capacities cj, cj(cid:48). Change the allocation of
the solution S1, . . . , Sk on bins j and j(cid:48) to an optimal solution for the two
bin problem. This change improves the worst overflow probability of the two
bins while not affecting the overflow probability of any other bin. In total
we get a better solution, in contradiction to the optimality of S1, . . . , Sk.
item (2): The same proof as in Appendix G.2 shows W OF P (a, b) = W OF P (1−
a − c2−c1
µ , 1 − b).
∂a
(a, b) = µ√
bσ
item (3): Fix b. Denote OF P1(a, b) = OF PS1 (aµ, bV ) = 1− Φ(∆1). It is a sim-
· φ(∆1) > 0. Similarly, if OF P2(a, b)
ple calculation that ∂OF P1
denotes the overflow probability in the second bin when the first bin has to-
· φ(∆2) < 0. Thus,
tal mean aµ and total variance bV , then ∂OF P2
OF P1 is monotonically increasing in a and OF P2 is monotonically decreas-
ing in a, and therefore there is a unique minimum for OF P (a, b) (when b
is fixed and a is free) that is obtained when OF P1(a, b) = OF P2(a, b), i.e.,
when ∆1 = ∆2.
−µ√
1−bσ
∂a =
item (4): We first explicitly determine what W OF P restricted to the valley is
as a function D(b) = W OF P (m(b), b) of b. From before we know that on the
valley ∆1 = ∆. Therefore, following the same reasoning as in the SP-MED
case,
D(b) =
c − µ
σ
1√
√
b +
.
1 − b
It follows that D(b) is monotonically decreasing in b for b ≤ 1
2 and increasing
otherwise. The maximal point is obtained in the saddle point that is the
center of the symmetry.
Lemma 6. The difference between minimal worst overflow probability in the
integral point found by the sorting algorithm and the optimal integral (or frac-
tional) point for SP-MWOP is at most O( L
αn ). In particular, when L = o(cid:0)n(cid:1)
and α is a constant, the difference is o(cid:0)1(cid:1).
Proof. We know from Theorem 3 that the difference is at most
min{∇D(ξ1),∇D(ξ2)} L
n
,
where ξ1 = (a1, b1) ∈ [O1, OP Tf ], ξ2 = (a2, b2) ∈ [OP Tf , O2] and O1 and O2 are
the two points on the bottom sorted path between which OP Tf lies, and notice
that even though W OF P is not differentiable when ∆1 = ∆2, it is differentiable
everywhere else. We now use the partial derivatives calculated in Appendix G.3.
We also replace φ(∆2)
c2−(1−a)µ and similarly for the other term. We get:
with ∆2φ(∆2)
σ2
(cid:26)
(cid:27) L
.
,
)
1
2b2
µ
1
∆2φ(∆2) · (
),∆1φ(∆1) · (
µ
,
min
2(1 − b1)
c2 − (1 − a1)µ
∆φ(∆) maximizes at ∆ = 1 with value at most
c1 − a2µ
n
. Also, (c1 − a2µ) +
2πe
(c2 − (1 − a1)µ) = c − µ − (a2 − a1)µ ≥ c − µ L
2 µ, where α is the total
space capacity, and a constant by our assumption. Hence, at least one of the
α . Also, for that term, the spare capacity
terms
is maximal, and therefore it takes at least half of the variance. Altogether, the
difference is at most O( L
c1−a2µ is at most 4
1√
n ≥ α
αn ) which completes the proof.
µ
c2−(1−a1)µ ,
µ
B.3 Stochastic Packing with Minimum Overflow Probability
(SP-MOP)
(cid:8)(µ(i), V (i))(cid:9)n
The total overflow probability is OF P = 1 −(cid:81)k
The SP-MOP problem gets as input integers k and n, specifying the number
of bins and services, integers c1, . . . , ck, specifying the bin capacities and values
i=1, specifying that the demand distribution X (i) of service i is
normal with mean µ(i) and variance V (i). A solution to the problem is a partition
of [n] to k disjoint sets S1, . . . , Sk ⊆ [n] that minimizes the overflow probability.
j=1(1 − OF Pj) where as we
computed before (in Appendix H.1) OF Pj = 1 − Φ(∆j). With two bins OF P is
a function from [0, 1]2 to R and OF P (a, b) = 1− Φ(∆1)Φ(∆2) where the first bin
has mean aµ and variance bV , the second bin has mean (1 − a)µ and variance
(1 − b)V and σj, ∆j are defined as above.
Lemma 7. OF P respects conditions (1)-(3) of Section 4.1.
Proof. We go over the conditions one by one.
item (1): Let S1, . . . , Sk be an optimal solution. Suppose there are j and j(cid:48) for
which Sj, Sj(cid:48) are not the optimal solution for the two-bin problem defined
by the services in Sj ∪ Sj(cid:48) and capacities cj, cj(cid:48). Change the allocation of the
solution S1, . . . , Sk on bins j and j(cid:48) to an optimal solution for the two bin
problem. This change improves (1 − OF Pj)(1 − OF Pj(cid:48)) while not affecting
the overflow probability of any other bin. In total we get a better solution,
in contradiction to the optimality of S1, . . . , Sk.
item (2): The same proof as in Appendix G.2 shows OF P (a, b) = OF P (1 −
a − c2−c1
µ , 1 − b).
item (3): Fix b.
∂2OF P
∂2a
= µ2[
φ(∆1)Φ(∆2) +
∆1
σ2
1
In particular ∂2OF P
a ∈ [0..1] and has a unique minimum a = m(b).
∆2
σ2
2
φ(∆2)Φ(∆1) +
2
σ1σ2
φ(∆1)φ(∆2)].
∂2a > 0 and for every fixed b, OF P (a, b) is convex over
.
Proving there exists a unique maximum over the valley is more challenging.
We wish to find all extremum points of the cost function D (OF P in our case)
over the valley {(m(b), b)}. Define m(a, b) = a−m(b). Then we wish to maximize
D(a, b) subject to m(a, b) = 0. Before, we computed the restriction D(b) of the
cost function over the valley and found its extremum points. However, here we
do not know how to explicitly find D(b). Instead, we use Lagrange multipliers
that allow working with the implicit form m(a, b) = 0 without explicitly finding
D(b). We prove a general result:
Lemma 8. If a cost function D is differentiable twice over [0, 1] × [0, 1] and
respects conditions (1)-(4) of Section 4.1, and if for every b ∈ [0, 1], D(a, b) is
strictly convex for a ∈ [0, 1], then any extremum point of D over the valley must
have zero gradient at Q, i.e., ∇(D)(Q) = 0.
Proof. The valley is defined by the equation m(a, b) = ∂D
multipliers we find that at any extremum point Q of D over the valley,
∂a (a, b). Using Lagrange
∇(D)(Q) = λ∇m(Q).
For some real value λ. However,
∇(D)(Q) = (
∂D
∂a
(Q),
∂D
∂b
(Q)) = (0,
∂D
∂b
(Q)),
because Q is on the valley. H
Also, as for every fixed b, D(a, b) is strictly convex, we have that
∂m
∂a
(Q) =
∂2D
∂2a
(Q) > 0.
Therefore, we conclude that λ = 0. This implies that ∂D
∇(D)(Q) = 0.
∂b (Q) = 0. Hence,
With that we prove:
Lemma 9. OF P respects condition (4) of Section 4.1.
Proof. Let Q = (a, b) be an extremum point of OF P over the valley. We look at
the range b ∈ [0.. 1
2 is obtained by the symmetry. Then, by Lemma 8:
2 ), b ≥ 1
φ(∆1)Φ(∆2)
φ(∆1)Φ(∆2)
∂∆1
∂a
∂∆1
∂b
= −φ(∆2)Φ(∆1)
= −φ(∆2)Φ(∆1)
∂∆2
∂a
∂∆2
∂b
.
, and
Dividing the two equations we get
∂∆1
∂a
∂∆2
∂b
=
∂∆2
∂a
∂∆1
∂b
.
Plugging the partial derivatives of ∆i by a and b, we get the equation
(cid:114) b
∆1
∆2
=
1 − b
.
As b ≤ 1
2 , b < 1 − b and we conclude that at Q ∆1 < ∆2. However, using the
log-concavity of the normal c.d.f function Φ we prove in Appendix I that:
Lemma 10. ∂OF P
∂a = 0 at a point Q = (a, b) with b ≤ 1
2 implies ∆1 ≥ ∆2.
Together, this implies that the only extremum point of OF P over the valley
is at b = 1
2 . However, at b = 0, the best is to fill the largest bin to full capacity
with variance 0, and thus, OF P (m(0), 0) = 1 − Φ(∆) where ∆ = c−µ
σ . On the
other hand, at b = 1
). As
(c1 − aµ) + (c2 − (1− a)µ) = c− µ, either c1 − aµ or c2 − (1− a)µ is at most c−µ
√
) ≤ Φ( c−µ
and therefore Φ(
σ ).
)Φ(
2 ) ≥ OF P (m(0), 0) and there is a unique maximum
We conclude that OF P (a, 1
point on the valley and it is obtained at b = 1
2 .
2 ) = 1 − Φ( c1−aµ√
2 c−µ
2σ ) = Φ( c−µ√
2 , OF P (a = m( 1
)Φ( c2−(1−a)µ√
2 c2−(1−a)µ
) ≤ Φ(
2 ), 1
√
2 c1−aµ
σ
1
2 σ
1
2 σ
σ
2σ
√
2
C Bounding the approximation error of the sorting
algorithm
√
(cid:80)
i P (i) ≥ (1, 1) =
2 (by the triangle inequality) and(cid:80)
We assume that no input service is too dominant. Recall that we represent service
i with the point P (i) = (a(i), b(i)) and P (1) + P (2) + . . . + P (n) = (1, 1). Thus,
i P (i) ≤ 2 (because
the length of the longest increasing path from (0, 0) to (1, 1), is obtained by the
path going from (0, 0) to (1, 0) and then to (1, 1)). Hence, the average length
of an input point P (i) is somewhere between
n . Our assumption states
that no element takes more than L times its "fair" share, i.e., that for every i,
P (i) ≤ L
√
n and 2
n . With that we prove:
2
Theorem 3. Let OP Tf be the fractional optimal solution. If D is differentiable,
the difference between the cost on the integral point found by the sorting al-
gorithm and the cost on the optimal integral (or fractional) point is at most
min{∇D(ξ1),∇D(ξ2)} L
n , where ξ1 ∈ [O1, OP Tf ], ξ2 ∈ [OP Tf , O2] and O1
and O2 are the two points on the bottom sorted path between which OP Tf lies.
Proof. Suppose we run the sorting algorithm on some input. Let OP Tint be the
integral optimal solution, OP Tf the fractional optimal solution and OP Tsort
the integral point the sorting algorithm finds on the bottom sorted path. We
wish to bound D(OP Tsort)− D(OP Tint) and clearly it is at most D(OP Tsort)−
D(OP Tf ). We now look at the two points O1 and O2 on the bottom sorted path
between which OP Tf lies (and notice that as far as we know it is possible that
OP Tsort is none of these points). Since D(OP Tf ) ≤ D(OP Tsort) ≤ D(O1) and
D(OP Tf ) ≤ D(OP Tsort) ≤ D(O2) the error the sorting algorithm makes is at
most
min{D(O1) − D(OP Tf ), D(O2) − D(OP Tf )} .
We now apply the mean value theorem and use our assumption that for every i,
P (i) ≤ L
n .
We define a new system constant, relative spare capacity, denoted by α which
equals c−µ
µ , i.e., it expresses the spare capacity as a fraction of the total mean.
We assume that the system has some constant (possibly small) relative spare ca-
pacity. Also, we only consider solutions where each bin is allocated services with
total mean not exceeding its capacity. Equivalently, we only consider solutions
where ∆j ≥ 0 for every 1 ≤ j ≤ k. We will later see that under these conditions
the sorting algorithm solves all three cost functions with a small error going fast
to zero with n.
We remark that in fact the proof shows something stronger: the deviation of
any (not necessarily optimal) fractional solution on the bottom sorted path, is
close to the deviation of the integral solution to the left or to the right of it on
the bottom sorted path.
D Proof of Lemma 1
is(cid:80)i
τ where P [i]
τ
Proof. We introduce some notation. Let τ = τ1, . . . , τn be a sequence of n ele-
ments that is a reordering of {1, . . . , n}. We associate with τ the n partial sums
P [1]
τ , . . . , P [n]
is the integral point that is the
sum of the first i points according to the sequence τ . We also define P [0]
and P [n]
necting P [i]
τ = (0, 0)
τ = (1, 1). The curve connecting τ is the curve that is formed by con-
with a line, for i = 0, . . . , n − 1.
j=1 P (τj ), i.e., P [i]
τ
τ and P [i+1]
τ
τ
Assume that in the sequence τ = τ1, . . . , τn there is some index i such that
the VMR of P (τi) is larger than the VMR of P (τi+1). Consider the sequence
τ(cid:48) that is the same as τ except for switching the order of τi and τi+1. I.e.,
τ(cid:48) = τ1, . . . , τi−1, τi+1, τi, τi+2, . . . , τn. We claim that the curve connecting τ(cid:48) lies
beneath the curve connecting τ . To see that notice that both curves are the same
up to the point P [i−1]
. There, the two paths split. τ adds P (τi) and then P (τi+1)
while τ(cid:48) first adds P (τi+1) and then P (τi). Then the two curves coincide and
overlap all the way to (1, 1). In the section where the two paths differ, the two
different paths form a parallelogram with P (τi) and P (τi+1) as two neighboring
edges of the parallelogram. As the angle P (τi+1) has with the a axis is smaller
than the angle P (τi) has with the a axis, the curve connecting τ(cid:48) goes beneath
that of τ .
To finish the argument, let PI be an arbitrary integral point for some I ⊆ [n].
Look at the sequence τ that starts with the elements of I followed by the elements
of [n]\I in an arbitrary order. Notice that PI lies on the curve connecting τ . Now
run a bubble sort on τ , each time ordering a pair of elements by their VMR.
Notice that the process terminates with the sequence that sorts the elements
by their VMR and the curve connecting the final sequence is the bottom sorted
path. Thus, we see that the bottom sorted path lies beneath the curve connecting
τ , and in particular PI lies above the bottom sorted path. A similar argument
shows PI lies underneath the upper sorted path.
E Proof of Theorem 1: The optimal solution is on the
bottom sorted path
Proof. Consider an arbitrary fractional point (a0, b0) lying strictly inside the
polygon confined by the upper and bottom sorted paths. If b0 ≤ 1
2 , then by
keeping b = b0 constant and changing a till it reaches the valley we strictly
decrease cost (because D is strictly monotone in this range). Now, when changing
a we either hit the bottom sorted path or the valley. If we hit the bottom sorted
path, we found a point on the bottom sorted path with less cost and we are
done. If we hit the valley, we can go down the valley until we hit the bottom
sorted path and again we are done (as D is strictly monotone on the valley).
2 we recall two facts that we already know:
If b0 ≥ 1
-- The point (1− a0, 1− b0) = ϕ(a0, b0) is fractional (since (a0, b0) is fractional
-- By the reflection symmetry we know that D(a0, b0) = D(1 − a0 − ζ, 1 − b0)
and ϕ maps fractional points to fractional points), and,
where ζ = c2−c1
µ ≥ 0.
Now, (1−a0−ζ, 1−b0) has b coordinate that is at most 1
2 . Also (1−a0−ζ, 1−b0)
lies to the left of the fractional point (1 − a0, 1 − b0) (since ζ > 0) and therefore
it lies above the bottom sorted path. We therefore see that the point (a0, b0) has
a corresponding fractional point with the same cost and with b coordinate at
most 1
2 . Applying the argument that appears in the first paragraph of the proof
we conclude that there exists some point on the bottom sorted path with less
cost, and conclude the proof.
F Standard Normal Distribution
The probability density function of the standard normal distribution (that has
mean 0 and variance 1) is φ(x) = 1√
2π
tion is Φ(x) = 1√
2π
−xφ(x). The second derivative is φ(cid:48)(cid:48)(x) = −φ(x) + x2φ(x) = (x2 − 1)φ(x).
2 dt. Clearly, Φ(cid:48)(x) = φ(x). Also, φ(cid:48)(x) = − x√
(cid:82) x
−∞ e− t2
2 and the cumulative distribution func-
e− x2
e− x2
2π
2 =
G SP-MED
G.1 Expected Deviation of a single bin
√
By definition the expected deviation of a single bin is DevSj = 1
σj
j dx. Doing the variable change t = x−µj
and then the variable change
− (x−µj )2
2σ2
cj)e
y = −t2
2 we get:
(cid:82) ∞
cj
2π
(x −
σj
(cid:90) −∞
cj − µj
DevSj = (µj − cj)[1 − Φ(
σj
= −σj∆j[1 − Φ(∆j)] +
)] − σj√
2π
j = σj[φ(∆j) − ∆j(1 − Φ(∆j))].
e− 1
eydy
− 1
2 (
cj−µj
2 ∆2
)2
σj
σj√
2π
Denoting g(∆) = φ(∆) − ∆[1 − Φ(∆)] we see that DevSj = σjg(∆j).
G.2 Symmetry
Claim. Dev(a, b) = Dev(1 − a − c2−c1
µ , 1 − b).
√
√
1 − b σ, ∆1(a, b) = c1−aµ
σ2(b)
b σ, σ2(b) =
. We know that Dev(a, b) = σ1(b)g(∆1(a, b))+σ2(b)g(∆2(a, b)).
Proof. Let us define σ1(b) =
∆2(a, b) = c2−(1−a)µ
To prove the claim it is enough to show that the following four equations hold:
σ1(b) = σ2(1 − b), σ2(b) = σ1(1 − b), ∆1(a, b) = ∆2(1 − a + c1−c2
∆2(a, b) = ∆1(1 − a + c1−c2
µ , 1 − b) and
µ , 1 − b).
√
1 − b σ = σ2(b) and similarly σ2(1 − b) = σ1(b). Also,
Indeed, σ1(1 − b) =
σ1(b) and
∆2(1 − a − c2 − c1
µ
, 1 − b) =
c2 − (1 − (1 − a + c1−c2
µ ))µ
σ2(1 − b)
A similar check shows that ∆1(1 − a − c2−c1
µ , 1 − b) = ∆2(a, b).
G.3 The partial derivatives of Dev
√
√
√
= c1−aµ
1 − bσ. We let ∆1 = c1−µ1
The system has absolute constants µ and V , σ =
bσ and
. In
σ2 =
this notation, Dev(a, b) = σ1 g(∆1) + σ2 g(∆2).
We calculate the g derivatives. Since g(∆) = φ(∆)+∆Φ(∆)−∆ we have that
g(cid:48)(∆) = φ(cid:48)(∆)+Φ(∆)+∆φ(∆)−1 = −∆φ(∆)+Φ(∆)+∆φ(∆)−1 = −(1−Φ(∆))
and g(cid:48)(cid:48)(∆) = φ(∆). As Φ(∆) ≤ 1, g(cid:48)(∆) ≤ 0 and g is monotonically decreasing.
As g(cid:48)(cid:48)(∆) > 0, g is convex.
and ∆2 = c2−µ2
= c2−(1−a)µ
V . We let σ1 =
σ1
σ2
σ1
σ2
Deriving by a We have: ∂∆1
∂a = − µ
σ1
and ∂∆2
∂a = µ
σ2
. Also,
∂Dev
∂a
∂g(∆1)
(a, b) = σ1
= −σ1
∂∆1
µ
σ1
∂∆1
∂a
+ σ2
∂g(∆2)
∂∆2
∂a
∂∆2
µ
σ2
[Φ(∆1) − 1] + σ2
[Φ(∆2) − 1] = µ [Φ(∆2) − Φ(∆1)]
Since Φ is monotonic, a zero value is achieved when ∆1 = ∆2. Deriving again
by a:
∂2Dev
∂a2 = µ [
∂Φ(∆2)
∂∆2
∂∆2
∂a
− ∂Φ(∆1)
∂∆1
∂∆1
∂a
] = µ2 [
φ(∆2)
σ2
+
φ(∆1)
σ1
]
Since ∂2Dev
∂a2 (cid:62) 0, it follows that for any 0 < b < 1, Dev(a) is convex and
hence ∆1 = ∆2 is a minimum point in this range.
Deriving by b We have: ∂σ1
∂b = ∆2
∂∆2
2(1−b) . Now,
√
∂b = σ
2
and ∂σ2
∂b = − σ
√
1−b
2
b
. Also, ∂∆1
∂b = − ∆1
2b and
∂Dev
∂b
=
=
∂g(∆1)
∂∆1
∂σ1
g(∆1) + σ1
∂b
φ(∆1)√
σ
2
b
− φ(∆2)√
1 − b
[
] =
Deriving a second time we get:
∂∆1
∂b
V
2
[
+
∂σ2
∂b
φ(∆1)
σ1
g(∆2) + σ2
− φ(∆2)
]
σ2
∂g(∆2)
∂∆2
∂∆2
∂b
∂2Dev
∂b2 =
=
σ
2
σ
4
[
[
1√
b
√
φ(∆1)
b
b
∂φ(∆1)
∂∆1
∂∆1
∂b
1 − 1) +
√
− 1
2b
b
√
φ(∆2)
(1 − b)
(∆2
φ(∆1) −
(∆2
1 − b
1√
1 − b
2 − 1)]
∂φ(∆2)
∂∆2
∂∆2
∂b
−
√
1
2(1 − b)
1 − b
φ(∆2) ]
The Mixed Derivative
∂2Dev
∂b∂a
=
∂
∂b
∂Dev
∂a
=
[ µ[Φ(∆2) − Φ(∆1)] ] = µ [
∂Φ(∆2)
∂∆2
− ∂Φ(∆1)
∂∆1
∂∆1
∂b
]
= µ [φ(∆2)
2(1 − b)
+ φ(∆1)
∆1
2b
] = µ [
1
2b
∆1φ(∆1) +
∆2φ(∆2)]
∂∆2
∂b
1
2(1 − b)
∂
∂b
∆2
The saddle point The Hessian of the function Dev contains the second order
partial derivatives of Dev, i.e., it is a 2 × 2 matrix H,
(cid:32) ∂2Dev
(cid:33)
H(a, b) =
∂a2 (a, b) ∂2Dev
∂a∂b (a, b) ∂2Dev
∂2Dev
∂b∂a (a, b)
∂b2 (a, b)
H is symmetric and therefore at any point (a, b) the matrix H(a, b) has two real
eigenvalues. Now, assume the first order derivatives of Dev vanish at some point
(a0, b0). If the Hessian at that point (a0, b0) has negative determinant, then the
product of these two eigenvalues is negative, and hence one is positive and the
other negative. It then follows that the point (a0, b0) is a saddle point of Dev.
2 ) which is
the center of the reflection symmetry. We know that at this point ∆1 = ∆2 and
σ1 = σ2. If then follows that:
We now compute the Hessian at the point (a0, b0) = ( 1
2D , 1
2 − c2−c1
(cid:19)
(cid:18) 2
√
H(a0, b0) =
2 µ2 φ(∆1)
2 µ∆1φ(∆1)
σ
√
2 µ∆1φ(∆1)
2 σ φ(∆1) [∆2
1 − 1]
We can now compute the determinant:
det[H(a0, b0)] =
∂2Dev
(a0, b0)
∂a2
√
2 µ2 φ(∆1)
∂2Dev
∂b2
√
= 2
= −4µ2[φ(∆1)]2 < 0.
σ
(a0, b0) − [
∂2Dev
∂a∂b
(a0, b0)]2
1 − 1] − 4µ2∆2
1 [φ(∆1)]2
2 σ φ(∆1) [∆2
As det[H(a0, b0)] < 0 and the partial derivatives of the first order vanish at
(a0, b0) we conclude that (a0, b0) is a saddle point.
H SP-MWOP
H.1 Expected overflow probability of a single bin
√
The overflow probability of bin j, denoted by OF PSj is OF PSj (µj, Vj) = 1
σj
Substituting t = x−µj
we get
σj
OF PSj (µj, Vj) =
1√
2π
e− t2
2 dt = 1 − Φ(
cj − µj
σj
) = 1 − Φ(∆j).
(cid:90) ∞
cj−µj
σj
(cid:82) ∞
cj
e
2π
− (x−µj )2
2σ2
j dx.
I SP-MOP
I.1 Proof of Lemma 10
Proof. The condition ∂OF P
∂a = 0 is equivalent to
σ1
σ2
2 , b < 1 − b and σ1 < σ2. Hence,
φ(∆1)
Φ(∆1)
=
As b < 1
· φ(∆2)
Φ(∆2)
φ(∆1)
Φ(∆1)
<
φ(∆2)
Φ(∆2)
.
Denote h(∆) = φ(∆)
Φ(∆) . We will prove that h is monotone decreasing, and this
implies that ∆1 > ∆2.
To see that h is monotone decreasing define H(∆) = ln(Φ(∆)). Then h = H(cid:48).
Therefore, h(cid:48) = H(cid:48)(cid:48). However, Φ is log-concave, hence H(cid:48)(cid:48) < 0. We conclude that
h(cid:48) < 0 and h is monotone decreasing.
J Unbalancing bin capacities is always better
Suppose we are given a capacity budget c and we have the freedom to choose
capacities c1, c2 that sum up to c for two bins. Which choice is the best? Off-
hand, it is possible that for each input there is a different choice of c1 and c2
that minimizes the expected deviation. In contrast, we show that the minimum
expected deviation always decreases as the difference c2 − c1 increases.
Lemma 11. Given a capacity budget c, the minimum expected deviation de-
creases as c2 − c1 increases. In particular the best choice is having a single bin
with capacity c and the worst choice is splitting the capacities evenly between the
two bins.
Proof. Recall that ∆1(a, b) = c1−aµ
√
reduce c1 by c and increase c2 by c, we get
and ∆2(a, b) = c2−(1−a)µ
1−b
. Therefore, if we
√
σ
b
σ
c1 − c − aµ
√
c1 − (a − c
µ )µ
√
σ
b
=
= ∆1(a − c
µ
, b).
∆1(a, b) =
b
Similarly, ∆2(a, b) = ∆2(a− c
σ
µ , b). Let Devc1,c2(a, b) denote the expected devi-
ation with bin capacities c1, c2. As Dev(a, b) = σ1(b)g(∆1(a, b))+σ2(b)g(∆2(a, b))
we see that
Devc1−c,c2+c(a, b) = Devc1,c2(a − c
µ
, b),
i.e., the graph is shifted left by c
µ .
Notice that the bottom sorted path does not depend on the bin capacities
and is the same in both cases. Let (a, b) be the optimal fractional solution for bin
capacities c1, c2. We know that (a, b) is on the bottom sorted path. Let a = a− c
µ .
We know that Devc1−c,c2+c(a, b) = Devc1,c2 (a, b). The point (a, b) lies to the left
of the bottom sorted path and therefore above it. As the optimal solution for
bin capacities c1 − c, c2 + c is also on the bottom sorted path and is strictly
better than any internal point, we conclude that the expected deviation for bin
capacities c1 − c, c2 + c is strictly smaller than the expected deviation for bin
capacities c1, c2.
An immediate corollary is the trivial fact that putting all the capacity budget
in one bin is best. Obviously, this is not always possible nor desirable, but if there
is tolerance in each bin capacity, we recommend minimizing the number of bins.
|
1301.2626 | 1 | 1301 | 2013-01-11T23:01:15 | Active Self-Assembly of Algorithmic Shapes and Patterns in Polylogarithmic Time | [
"cs.DS",
"cs.CC",
"cs.CG"
] | We describe a computational model for studying the complexity of self-assembled structures with active molecular components. Our model captures notions of growth and movement ubiquitous in biological systems. The model is inspired by biology's fantastic ability to assemble biomolecules that form systems with complicated structure and dynamics, from molecular motors that walk on rigid tracks and proteins that dynamically alter the structure of the cell during mitosis, to embryonic development where large-scale complicated organisms efficiently grow from a single cell. Using this active self-assembly model, we show how to efficiently self-assemble shapes and patterns from simple monomers. For example, we show how to grow a line of monomers in time and number of monomer states that is merely logarithmic in the length of the line.
Our main results show how to grow arbitrary connected two-dimensional geometric shapes and patterns in expected time that is polylogarithmic in the size of the shape, plus roughly the time required to run a Turing machine deciding whether or not a given pixel is in the shape. We do this while keeping the number of monomer types logarithmic in shape size, plus those monomers required by the Kolmogorov complexity of the shape or pattern. This work thus highlights the efficiency advantages of active self-assembly over passive self-assembly and motivates experimental effort to construct general-purpose active molecular self-assembly systems. | cs.DS | cs | Active Self-Assembly of Algorithmic Shapes and Patterns
in Polylogarithmic Time∗
Damien Woods1,2 Ho-Lin Chen1,2,7 Scott Goodfriend3
Nadine Dabby4 Erik Winfree1,4,5,6 Peng Yin1,5,6,8
Computer Science1 , Center for Mathematics of Information2 ,
Department of Chemical Engineering3 , Department of Computation and Neural Systems4 ,
Department of Bioengineering5 , Center for Biological Circuit Design6 ,
Caltech, Pasadena, CA 91125, U.S.A.
Present address: Department of Electrical Engineering, National Taiwan University7
Present address: Wyss Institute for Biologically Inspired Engineering, 3 Blackfan Circle,
Boston, MA 024458
Abstract
We describe a computational model for studying the complexity of self-assembled
structures with active molecular components. Our model captures notions of growth
and movement ubiquitous in biological systems. The model is inspired by biology’s
fantastic ability to assemble biomolecules that form systems with complicated structure
and dynamics, from molecular motors that walk on rigid tracks and proteins that
dynamically alter the structure of the cell during mitosis, to embryonic development
where large-scale complicated organisms efficiently grow from a single cell. Using this
active self-assembly model, we show how to efficiently self-assemble shapes and patterns
from simple monomers. For example, we show how to grow a line of monomers in time
and number of monomer states that is merely logarithmic in the length of the line.
Our main results show how to grow arbitrary connected two-dimensional geometric
shapes and patterns in expected time that is polylogarithmic in the size of the shape,
plus roughly the time required to run a Turing machine deciding whether or not a
given pixel is in the shape. We do this while keeping the number of monomer types
logarithmic in shape size, plus those monomers required by the Kolmogorov complexity
of the shape or pattern. This work thus highlights the efficiency advantages of active
self-assembly over passive self-assembly and motivates experimental effort to construct
general-purpose active molecular self-assembly systems.
∗Supported by NSF grants CCF-1219274, CCF-1162589, and 0832824—the Molecular Programming
Pro ject, an NSF Graduate Fellowship, and The Caltech Center for Biological Circuit Design.
3
1
0
2
n
a
J
1
1
]
S
D
.
s
c
[
1
v
6
2
6
2
.
1
0
3
1
:
v
i
X
r
a
1
Figure 1: (a) Increase in mass in embryonic mouse [29], cow [49], and human [45]. Gray lines highlight
a time interval of exponential growth in each species. (b) An abstract example of exponential growth
showing insertion and state change.
1 Introduction
We propose a model of computation that takes its main inspiration from biology. Embryonic
development showcases the capability of molecules to compute efficiently. A human zygote
cell contains within it a compact program that encodes the geometric structure of an
organism with roughly 1014 cells, that have a wide variety of forms and roles, with each
such cell containing up to tens of thousands of proteins and other molecules with their own
intricate arrangement and functions. As shown in Figure 1(a), early stages of embryonic
development demonstrate exponential [45] growth rates in mass and number of cells over
time, showing remarkable time efficiency. Not only this, but the developmental path from
a single cell to a functioning organism is an intricately choreographed, and incredibly
robust, temporal and spatial manufacturing process that operates at the molecular scale.
Development is probably the most striking example of the ubiquitous process of molecular
self-assembly, where relatively simple components (such as nucleic and amino acids, lipids,
carbohydrates) organize themselves into larger systems with extremely complicated structure
and dynamics (cells, organs, humans). To capture the potential for exponential growth
during development, an abstract model must allow for growth by insertion and rearrangement
of elements as well as for local state changes that store information that guides the process,
as shown for example in Figure 1(b).
Molecular programming, where nanoscale engineering is thought of as a programming
task, provides our second motivation. The field has progressed to the stage where we can
design and synthesize a range of programable self-assembling molecular systems, all with
relatively simple laboratory techniques. For example, short DNA strands that form ‘tiles’
can be self-assembled into larger DNA tile crystals [79] that are algorithmically-patterned
as counters or Sierpinski triangles [61, 8, 31, 9]. This form of passive self-assembly is
theoretically capable of growing arbitrarily complex algorithmically described shapes and
patterns. Indeed, the theory of algorithmic self-assembly serves as a guide to tell us what
kinds of structures are possible, and interesting, to implement with DNA tiles. DNA origami
can be used to create uniquely addressable shapes and patterns upon which ob jects can be
localized within six nanometer resolution [62]. These DNA tile and origami systems are
2
0510152025303540455010−810−710−610−510−410−310−210−1100time (in days)log(mass in grams) mouse embryo eyecow embryohuman embryo(a)(b) static, in the sense that after formation their structure is essentially fixed. However, DNA
nanotechnology has seen increased interest in the fabrication of active nanostructures that
have the ability to dynamically change their structure [11]. Examples include DNA-based
walkers [10, 34, 35, 46, 52, 66, 67, 71, 81, 82], DNA origami that reconfigure [6, 36, 48],
and simple structures called molecular motors that transition between a small number of
discrete states [15, 22, 28, 33, 42, 44, 47, 72, 73, 83, 84]. In these systems the interplay
between structure and dynamics leads to behaviors and capabilities that are not seen in
static structures, nor in other well-mixed chemical reaction network type systems.
Here we suggest a model to motivate engineering of molecular structures that have
complicated active dynamics of the kind seen in living biomolecular systems. Our model
combines features seen in passive DNA-based tile self-assembly, molecular motors and other
active systems, molecular circuits that evolve according to well-mixed chemical kinetics, and
even reaction-diffusion systems. The model is designed to capture the interplay between
molecular structure and dynamics.
In our model, simple molecular components form
assemblies that can grow and shrink, and individual components undergo state changes
and move relative to each other.
The model consists of a two-dimensional grid of monomers. A specified set of rules,
or a program, directs adjacent monomers to interact in a pairwise fashion. Monomers
have internal states, and a pair of adjacent monomers can change their state with a single
rule application. Monomers can appear and disappear from the grid. So far, the model
can be thought of as a cellular automaton of a certain kind (that is, a cellular automaton
where rules are applied asynchronously and are nondeterministic, and there is a notion of a
growth front). An additional rule type allows monomers to move relative to each other.
The movement rule is locally applied but propagates movement throughout the system in
very non-local fashion. This geometric and mechanical feature distinguishes our model, the
nubot model, from previous molecular models and cellular automata, and, as we show in this
paper, crucially underlies its construction efficiency. The system evolves as a continuous
time Markov process, with rules being applied to the grid asynchronously and in parallel
using standard chemical kinetics, modeling the distributed nature of molecular reactions.
The model can carry out local state changes on a grid, so it can easily simulate Turing
machines, walkers and cellular automata-like systems. We show examples of other simple
programs such as robotic molecular arms that can move distance n in expected time O(log n),
something that can not be done by cellular automata. By using a combination of monomer
movement, appearance, and state change we show how to build a line of monomers in time
that is merely logarithmic in the length of the line, something that is provably impossible in
the (passive) abstract tile assembly model [2]. We go on to efficiently build a binary counter
(a program that counts builds an n × log2 n rectangle while counting from 0 to n − 1),
within time and number of monomer states that are both logarithmic in n, where n is a
power of 2. We build on these results to show that the model is capable of building wide
classes of shapes exponentially faster than passive self-assembly. We show how to build
a computable shape of size ≤ n × n in time polylogarithmic in n, plus roughly the time
3
needed to simulate a Turing machine that computes whether or not a given pixel is in the
final shape. Our constructions are not only time efficient, but efficient in terms of their
program-size: requiring at most polylogarithmic monomer types in terms of shape size, plus
that required by the Kolmogorov complexity of the shape.
For shapes of size ≤ n × n that require a lot (i.e. > n) of time and space to compute
their pixels on a Turing machine, the previous computable shape construction requires
(temporary) growth well beyond the shape boundary. One can ask if there are interesting
structures that we can build efficiently, but yet keep all growth “in-place”, that is, confined
within the boundary. It turns out that colored patterns, where the color of each pixel in the
pattern is computable by a polynomial time Turing machine, can be computed extremely
efficiently in this way. More precisely, we show that n × n colored patterns are computable
in expected time O(log(cid:96)+1 n) and using O(s + log n) states, where each pixel’s color is
computable by a program-size s Turing machine that runs in polynomial time in the length
of the binary description of pixel indices (specifically, in time O(log(cid:96) n) where (cid:96) is O(1)).
Thus the entire pattern completes with only a linear factor slowdown in Turing machine
time (i.e. log n factor slowdown). Furthermore this entire construction is initiated by a
single monomer and is carried out using only local information in an entirely distributed
and asynchronous fashion.
Essentially, our constructions serve to show that shapes and patterns that are expo-
nentially larger than their description length can be fabricated in polynomial time, with a
linear number of states, in their description length, besides the states that are required by
the Kolmogorov complexity of the shape.
Our active self-assembly model is intentionally rather abstract, however our results
show that it captures some of the features seen in the biological development of organisms
(exponential growth—with and without fast communication over long distances, active yet
simple components) as well those seen in many of the active molecular systems that are
currently under development (for example, DNA walkers, motors and a variety of active
systems that exploit DNA strand displacement). Also, the proof techniques we use, at a very
abstract level, are informed by natural processes. In the creation of a line, a simple analog of
cell division is used; division is also used in the construction of a binary counter along with
a copying (with minor modifications) process to create new counter rows; in the assembly
of arbitrary shapes we use analogs of protein folding and scaffold-based tissue engineering;
for the assembly of arbitrary patterns we were inspired by biological development where
growth and patterning takes places in a bounded region (e.g. an egg or womb) and where
many parts of the development of a single embryo occur in an independent and seemingly
asynchronous fashion.
Section 2 contains the detailed definition of our nubot model, which is followed by a
number of simple and intuitive examples in Section 3. Section 4 gives a polynomial time
algorithm for simulating nubots rule applications, showing that the non-local movement
rule is in a certain sense tractable to simulate and providing a basis for a software simulator
we have developed. In Section 5 we show how to grow lines and squares in time logarithmic
4
in their size. This section includes a number of useful programming techniques and analysis
tools for nubots. In Sections 6 and 7 we give our main results: building arbitrary shapes and
patterns, respectively, in polylogarithmic time in shape/pattern size (plus the worst-case
time for a Turing machine to compute a single pixel) and using only a logarithmic number
of states (plus the Kolmogorov complexity of the shape/pattern). Section 8 contains a
number of directions for future work.
1.1 Related work
Although our model takes inspiration from natural and artificial active molecular processes, it
posesses a number of features seen in a variety of existing theoretical models of computation.
We discuss some such related models here.
The tile assembly model [75, 76, 63] formalizes the self-assembly of molecular crystals [61]
from square units called tiles. This model takes the Wang tiling model [74], which is a
model of plane-tilings with square tiles, and restricts it by including a natural mechanism
for the growth of a tiling from a single seed tile. Self-assembly starts from a seed tile that
is contained in a soup of other tiles. Tiles with matching colors on their sides may bind to
each other with a certain strength. A tile can attach itself to a partially-formed assembly
of tiles if the total binding strength between the tile and the assembly exceeds a prescribed
system parameter called the temperature. The assembly process proceeds as tiles attach
themselves to the growing assembly, and stops when no more tile attachments can occur.
Tile assembly is a computational process: the exposed edges of the growing crystal encode
the state information of the system, and this information is modified as a new tiles attach
themselves to the crystal. In fact, tile assembly can be thought of as a nondeterministic
asynchronous cellular automaton, where there is a notion of a growth starting from a seed.
Tile assembly formally couples computation with shape construction, and the shape can be
viewed as the final output of the tile assembly “program”. The model is capable of universal
computation [75] (Turing machine simulation) and even intrinsic universality [26] (there
is a single tile set that simulates any tile assembly system). Tiles can efficiently encode
shapes, in the sense that there is a close link between the Kolmogorov complexity of an
arbitrary, scaled, connected shape and the number of tile types required to assemble that
shape [69]. There have been a wide selection of results on clarifying what shapes can and
can not be constructed in the tile assembly model, with and without resource constraints
on time and the number of required tile types (e.g. see surveys [55, 24]), showing that the
tile assembly model demonstrates a wide range of computational capabilities.
There have been a number of interesting generalizations of the tile assembly model.
These include the two-handed, or hierarchical, tile assembly model where whole assemblies
can bind together in a single step [3, 18, 14], the geometric model where the sides of
tiles have jigsaw-piece like geometries [30], rotatable and flipable polygon and polyomino
tiles [19], and models where temperature [3, 38, 70], concentration [12, 16, 39, 23] or mixing
stages [18, 20] are programmed. All of these models could be classed as passive in the sense
5
that after some structure is grown, via a crystal-like growth process, the structure does
not change. Active tile assembly models include the activatable tile model where tiles pass
signals to each other and can change their internal ‘state’ based on other tiles that join
to the assembly [53, 54, 37]. This interesting generalization of the tile assembly model is
essentially an asynchronous nondeterministic cellular automaton of a certain kind, and may
indeed be implementable using DNA origami tiles with strand displacement signals [53, 54].
There are also models of algorithmic self-assembly where after a structure grows, tiles can
be removed, such as the kinetic tile-assembly model [77, 78, 17] (which implements the
chemical kinetics of tile assembly) and the negative-strength glue model [25, 59, 56], or the
RNAse enzyme model [1, 57, 21]. Although these models share properties of our model
including geometry and the ability of tile assemblies to change over time, they do not share
our notion of movement, something that is fundamental to our model and sets it apart from
models that can be described as asynchronous nondeterministic cellular automata.
As a model of computation, stochastic chemical reaction networks are computationally
universal in the presence of non-zero, but arbitrarily low, error [7, 68]. Our model crucially
uses ideas from stochastic chemical reaction networks. In particular our update rules are all
unimolecular or bimolecular (involving at most 1 or 2 monomers per rule) and our model
evolves in time as a continuous time Markov process; a notion of time we have borrowed
from the stochastic chemical reaction network model [32].
Our active-self assembly model is capable of building and reconfiguring structures. The
already well-established field of reconfigurable robotics is concerned with systems composed
of a large number of (usually) identical and relatively simple robots that act cooperatively
to perform a range of tasks beyond the capability of any particular special-purpose robot.
Individual robots are typically capable of independent physical movement and collectively
share resources such as power and computation [51, 64, 80]. One standard problem is
to arrange the robots in some arbitrary initial configuration, specify some desired target
configuration, and then have the robots collectively carry out the required reconfiguration
from source to target. Crystalline reconfigurable robots [13, 64] have been studied in this
way. This model consists of unit-square robots that can extend or contract arms on each
of four sides and attach or detach from neighbors. For this model, Aloupis et al [5] give
an algorithm for universal reconfiguration between a pair of connected shapes that works
in time merely logarithmic in shape size. As pointed out by Aloupis et al in a subsequent
paper [4], this high-speed reconfiguration can lead to strain on individual components, but
they show that if each robot can displace at most a constant number of other robots, and
reach at most constant velocity, then there is an optimal Θ(n) parallel time reconfiguration
√
algorithm (which also has the desired property of working “in-place”). Reif and Slee [60]
also consider physicaly constrained movement, giving an optimal Θ(
n) reconfiguration
algorithm where at most constant acceleration, but up to linear speed, is permitted. Like our
nubot model, these models and algorithms implement fast parallel reconfiguration. However,
here we intentionally focus on growth and reconfiguration algorithms that use very few
states (typically logarithmic in assembly size) to model the fact that molecules are simple
6
and ‘dumb’—however, in reconfigurable robotics it is typically the case that individual
robots can store the entire desired configuration. Our model also has other properties that
don’t always make sense for macro-scale robots such as growth and shrinkage (large numbers
of monomers can be created and destroyed), a presumed unlimited fuel source, asynchronous
independent rule updates and Brownian motion-style agitation. Nevertheless we think it
will make interesting future work to see what ideas and results from reconfigurable robotics
apply to a nanoscale model and what do not.
Klavins et al [40, 41] model active self-assembly using conformational switching [65] and
graph grammars [27]. In such work, an assembly is a graph, which can be modified by graph
rewriting rules that add or delete vertices or edges. Thus the model focuses attention on the
topology of assemblies. Besides permitting the assembly of static graph structures, other
more dynamic structures—such as a walker subgraph that moves around on a larger graph—
can be expressed. Our model also has the ability to change structure and connectivity
in a way that takes inspiration from such systems, but additionally includes geometric
constraints by virtue of the fact that it lives on a two-dimensional grid. Lindenmayer
systems [43] are another model where a graph-like structure is modified via insertion and
addition of nodes, and where it is possible to generate beautiful movies of the growth of
ferns and other plants [58]. Although it is indeed a model of (potentially fast) growth via
insertion of nodes, it is quite different in a number of ways from our own model.
2 Nubot model description
Figure 2: (a) A monomer in the triangular grid coordinate system. (b) Examples of monomer
interaction rules, written formally as follows: r1 = (1, 1, null, (cid:126)x) → (2, 3, null, (cid:126)x), r2 = (1, 1, null, (cid:126)x) →
(1, 1, flexible, (cid:126)x), r3 = (1, 1, rigid, (cid:126)x) → (1, 1, null, (cid:126)x), r4 = (1, 1, rigid, (cid:126)x) → (2, 3, flexible, (cid:126)x),
r5 = (b, empty, null, (cid:126)x) → (1, 1, flexible, (cid:126)x), r6 = (1, a, rigid, (cid:126)x) → (1, empty, null, (cid:126)x), and r7 =
(1, 1, rigid, (cid:126)x) → (1, 2, rigid, (cid:126)y). For rule r7, the two potential symmetric movements are shown
corresponding to two choices for arm and base, one of which is nondeterministically chosen.
This section contains the model definition. We also give a number of simple example
constructions in Section 3, which may aid the reader here.
The model uses a two-dimensional triangular grid with a coordinate system using axes
x and y as shown in Figure 2(a). For notational convenience we define a third axis w, that
7
11(0,0)xyw(1,0)(2,0)(0,1)(0,2)(1,1)pp + yp + wp + xp - xp - wp - yabChange states1123Make a flexible bond111111Break a rigid bond2311Change a rigid bond to a flexible bond and change states1211Movement in the w directionwBaseArm11Appearanceb1aDisappearance1ABAB1211Movement in the -w direction-wBaseArmABAB r1 r2 r3 r4 r5 r6 r7 r7 runs through the origin and parallel to the vector (cid:126)w in Figure 2(a), but only axes x and y
are used to define coordinates. In the vector space induced by this coordinate system, the
axial directions D = { (cid:126)w, − (cid:126)w, (cid:126)x, −(cid:126)x, (cid:126)y , −(cid:126)y} are the unit vectors along the grid axes. A
pair (cid:126)p ∈ Z2 is called a grid point and has the set of six neighbors {(cid:126)p + (cid:126)u (cid:126)u ∈ D}.
The basic assembly unit is called a nubot monomer. A monomer is a pair X = (sX , (cid:126)p(X ))
where sX ∈ Σ is one of a finite set of states and (cid:126)p(X ) ∈ Z2 is a grid point. Each grid
point contains zero or one monomers. Monomers are depicted as state-labelled disks of unit
diameter centered on a grid point. In general, a nubot is a collection of nubot monomers.
Two neighboring monomers (i.e. residing on neighboring grid points) are either connected
by a flexible bond or a rigid bond, or else have no bond between them (called a nul l bond).
A flexible bond is depicted as a small empty circle and a rigid bond is depicted as a solid
disk. Flexible and rigid bonds are described in more detail in Definition 2.2.
A connected component is a maximal set of adjacent monomers where every pair of
monomers in the set is connected by a path consisting of monomers bound by either flexible
or rigid bonds.
A configuration C is defined to be a finite set of monomers at distinct locations and
the bonds between them, and is assumed to define the state of an entire grid at some time
instance. A configuration C can be changed either by the application of an interaction rule
or by a random agitation, as we now describe.
2.1 Rules
Two neighboring monomers can interact by an interaction rule, r = (s1, s2, b, (cid:126)u) →
(s1(cid:48) , s2(cid:48) , b(cid:48) , (cid:126)u(cid:48) ). To the left of the arrow, s1, s2 ∈ Σ ∪ {empty} are monomer states, at
most one of s1, s2 is empty (denotes lack of a monomer), b ∈ {flexible, rigid, null} is the
bond type between them, and (cid:126)u ∈ D is the relative position of the s2 monomer to the
s1 monomer. If either of s1, s2 is empty then b is null, also if either or both of s1(cid:48) , s2(cid:48) is
empty then b(cid:48) is null. The right is defined similarly, although there are some restrictions
on the rules (involving (cid:126)u(cid:48) ) which are described below. The left and right hand sides of the
arrow respectively represent the contents of the two monomer positions before and after the
application of rule r . In summary, via suitable rules, adjacent monomers can change states
and bond type, one or both monomers can disappear, one can appear in an empty space, or
one monomer can move relative to another by unit distance. A rule is only applicable in the
orientation specified by (cid:126)u, and so rules are not rotationally invariant. The rule semantics
are defined next, and a number of examples are shown in Figure 2b.
A rule may involve a movement (translation), or not. First, let us consider the case
where there is no movement, that is, where (cid:126)u = (cid:126)u(cid:48) . Thus we have a rule of the form
r = (s1, s2, b, (cid:126)u) → (s1(cid:48) , s2(cid:48) , b(cid:48) , (cid:126)u). From above, we have the restriction that at most one of
s1, s2 is the special empty state (hence we disallow spontaneous generation of monomers
from completely empty space). State change and bond change occurs in a straightforward
way and a few examples are shown in Figure 2b. If si ∈ {s1, s2} is empty and s(cid:48)
i is not,
8
then the rule induces the appearance of a new monomer. If one or both monomers go from
being non-empty to being empty, the rule induces the disappearance of monomer(s).
A movement rule is an interaction rule where (cid:126)u (cid:54)= (cid:126)u(cid:48) . More precisely, in a movement
rule d((cid:126)u, (cid:126)u(cid:48) ) = 1, where d(u, v) is Manhattan distance on the triangular grid, and none
If we fix (cid:126)u ∈ D, then there are exactly two (cid:126)u(cid:48) ∈ D that
of s1, s2, s1(cid:48) , s2(cid:48) is empty.
satisfy d((cid:126)u, (cid:126)u(cid:48) ) = 1. A movement rule is applied as follows. One of the two monomers is
nondeterministically chosen to be the base (which remains stationary), the other is the
arm (which moves). If the s2 monomer, denoted X , is chosen as the arm then X moves
from its current position (cid:126)p(X ) to a new position (cid:126)p(X ) − (cid:126)u + (cid:126)u(cid:48) . After this movement (and
potential state change), (cid:126)u(cid:48) is the relative position of the s2(cid:48) monomer to the s1(cid:48) monomer,
as illustrated in Figure 2b. If the s1 monomer, Y , is chosen as the arm then Y moves
from (cid:126)p(Y ) to (cid:126)p(Y ) + (cid:126)u − (cid:126)u(cid:48) . Again, (cid:126)u(cid:48) is the relative position of the s2(cid:48) monomer to the
s1(cid:48) monomer. Bonds and states can change during the movement, as dictated by the rule.
However, we are not done yet, as during a movement, the translation of the arm monomer A
by a unit vector may cause the translation of a collection of monomers, or may in fact be
impossible; to describe this phenomenon we introduce two definitions.
The (cid:126)v -boundary of a set of monomers S is defined to be the set of grid locations located
unit distance in the (cid:126)v direction from the monomers in S .
Definition 2.1 (Agitation set) Let C be a configuration containing monomer A, and let
(cid:126)v ∈ D be a unit vector. The agitation set A(C, A, (cid:126)v) is defined to be the minimal monomer
set in C containing A that can be translated by (cid:126)v such that: (a) monomer pairs in C that
are joined by rigid bonds do not change their relative position to each other, (b) monomer
pairs in C that are joined by flexible bonds stay within each other’s neighborhood, and (c)
the (cid:126)v-boundary of A(C, A, (cid:126)v) contains no monomers.
It can be seen that for any non-empty configuration the agitation set is always non-empty.
Using this definition we define the movable set M(C, A, B , (cid:126)v) for a pair of monomers
A, B , unit vector (cid:126)v and configuration C . Essentially, the movable set is the minimal set that
can be moved without disrupting existing bonds or causing collisions with other monomers.
Definition 2.2 (Movable set) Let C be a configuration containing adjacent monomers
A, B , let (cid:126)v ∈ D be a unit vector, and let C (cid:48) be the same configuration as C except that C (cid:48)
omits any bond between A and B . The movable set M(C, A, B , (cid:126)v) is defined to be the
agitation set A(C (cid:48) , A, (cid:126)v) if B (cid:54)∈ A(C (cid:48) , A, (cid:126)v), and the empty set otherwise.
Figure 3 illustrates this definition with two examples.
Now we are ready to define what happens upon application of a movement rule. If
M(C, A, B , (cid:126)v) (cid:54)= {}, then the movement where A is the arm (which should be translated
by (cid:126)v) and B is the base (which should not be translated) is applied as follows: (1) the
movable set M(C, A, B , (cid:126)v) moves unit distance along (cid:126)v ; (2) the states of, and the bond
between, A and B are updated according to the rule; (3) the states of all the monomers
besides A and B remain unchanged and pairwise bonds remain intact (although monomer
positions and bond orientations may change).
9
Figure 3: Two movable set examples. (a) The movement rule is (a, a, rigid, − (cid:126)w) → (c, c, rigid, (cid:126)x),
where the rightmost of the two monomers is nondeterministically chosen as the arm. (b) Example
with a non-empty movable set. (b1) Current configuration of the system. (b2) Computation of
the movable set. The movable set is highlighted in green, and numbers in parentheses indicate the
sequence of incorporation of monomers into the movable set via Algorithm 4.1. Monomers move
as shown in (b3). Finally, (b4) shows the configuration of the system after the application of the
movement rule. (c) Example with an empty movable set. (c1) A configuration that is identical
to the configuration in b1, except for a single rigid bond highlighted in black. (c2) Algorithm 4.1
completes at the pink monomer, which is blocked by the base monomer B , and hence returns an
empty movable set. Thus the movement rule can not be applied.
If M(C, A, B , (cid:126)v) = {}, the movement rule is inapplicable (the pair of monomers A, B
are “blocked” and thus A is prevented from translating).
Section 4 describes a greedy algorithm for computing the movable set M(C, A, B , (cid:126)v) in
time linear in assembly size.
We note that flexible bonds are not required for any of the constructions in this paper
(they are used in the construction in Figure 11 but they can be removed if we add extra
monomers and rules), however we retain this aspect of the model because we anticipate
that flexible bonds will be useful for future studies.
2.2 Agitation
Agitation is intended to model movement that is not a direct consequence of a rule application,
but rather results from diffusion, Brownian motion, turbulent flow or other undirected inputs
of energy. An agitation step, applies a unit vector movement to a monomer. This monomer
10
aArmBaseuPosition change rule:aaccMovable set emptyA(1)(2)(3)(4)(5)(6)(7)B(4)b2A(1)(2)(3)(4)(5)(6)(7)B(4)c2b3A(1)(2)(3)(4)(5)(6)(7)B(4)b4ccbbbbbbbbbbbbbbbbbbbbbbaaCurrent configurationbbbbbbbbbbbbbbbbbbbbbbb1c1aaCurrent configurationbbbbbbbbbbbbbbbbbbbbbb=> rule is not applied then moves, possibly along with many other monomers in a way that does not break rigid
nor flexible bonds. More precisely, applying a (cid:126)v ∈ D agitation step to monomer A causes
the agitation set A(C, A, (cid:126)v) (Definition 2.1) to move by vector (cid:126)v . During agitation, the only
change in the system configuration is in the positions of the constituent monomers in the
diffusing component, and all the monomer states and bond types remain unchanged.
None of the constructions in this paper exploit agitation, and all work correctly regardless
of its presence or absence (due to the fact that our constructions are stable, see below).
However, we feel it is important enough to be considered as part of the model definition.
We leave open the possibility of designing interesting systems that exploit agitation (e.g. by
having components interact by drifting into each other as is typical in a molecular-scale
environment).
2.3 Stability
The following definition is useful for proving correctness of many of our constructions.
Definition 2.3 (Stable) A configuration C is stable if for al l monomers A in C and for
al l 6 unit vectors (cid:126)v ∈ D, the agitation set A(C, A, (cid:126)v) is the entire set of monomers in C .
In other words, translating any monomer by any of the 6 unit vectors in D results in the
translation of the entire set. This happens when monomers have a bond structure such
that under agitation all monomers move together unit distance, and their relative positions
remain unchanged. Hence, stable configurations have a bond structure that allows the
entire structure to be “pushed” or “pulled” around by the movement of any individual
monomer, without changing the relative location of any monomer to any other monomer in
the configuration. Essentially, this can be used as a tool to show that a structure does not
unintentionally become disconnected or end up in an unintended configuration. This is a
very useful property when proving the correctness and carrying out time analysis of our
constructions, and is extensively used in this paper.
2.4 System evolution
An assembly system T = (C0 , R) is a pair where C0 is the initial configuration, and R is
the set of interaction rules. Consider two configurations Ci and Cj . If Cj can be derived
from Ci by a single step agitation, we write Ci (cid:96)A Cj . Let the relation (cid:96)∗
A be the reflexive
transitive closure of (cid:96)A . If Ci (cid:96)∗
A Cj , then Ci and Cj are called isomorphic configurations,
∼= Cj . (Notice that the relative position of monomers may differ for
which is denoted as Ci
two isomorphic configurations, although their connectivity is identical.) If Cj can be derived
from Ci by a single application of a rule r ∈ R, we write Ci (cid:96)r Cj . If Ci transitions to Cj by
either a single agitation step or a single application of some rule r ∈ R, we write Ci (cid:96)T Cj .
Let the relation (cid:96)∗
T be the reflexive transitive closure of (cid:96)T . The set of configurations that
can be produced by an assembly system T = (C0 , R) is Prod(T ) = {C C0 (cid:96)∗
T C }. The set
D}.
of terminal configurations are Term(T ) = {C C ∈ Prod(T ) and (cid:64) D (cid:29) C s.t. C (cid:96)∗
T
11
An assembly system T uniquely produces C if ∀ D ∈ Term(T ), C ∼= D. A trajectory is a
finite sequence of configurations C1 , C2 , . . . , Ck where Ci (cid:96)T Ci+1 and 1 ≤ i ≤ k − 1.
An assembly system evolves as a continuous time Markov process. For simplicity, when
counting the number of applicable rules for a configuration, a movement rule is counted
twice, to account for the two symmetric choices of arm and base. If there are k applicable
transitions for a configuration Ci (i.e. k is the sum of the number of rule and agitation steps
that can be applied to all monomers), then the probability of any given transition being
applied is 1/k , and the time until the next transition is an exponential random variable
with rate k (i.e. the expected time is 1/k). The rate for each rule application and agitation
step is 1 in this paper (although more sophisticated rate choices can be accommodated by
of the expected times of each transition in the tra jectory. Thus, (cid:80)
the model). The probability of a tra jectory is then the product of the probabilities of each
of the transitions along the tra jectory, and the expected time of a tra jectory is the sum
t∈T Pr[t]time(t) is the
expected time for the system to evolve from configuration Ci to configuration Cj , where T
is the set of all tra jectories from Ci to any configuration isomorphic to Cj , that do not pass
through any other configuration isomorphic to Cj , and time(t) is the expected time for
tra jectory t.
The following lemma is useful for the time analysis of the constructions in this paper.
Lemma 2.4 Given an assembly system T in which al l configurations in Prod(T ) have at
most one stable connected component, the expected time from configuration Ci to configura-
tion Cj is the same with or without agitation steps.
Proof: Assembly system T has a corresponding Markov process M . We consider another
Markov process M (cid:48) . Each state SC in M (cid:48) corresponds to all configurations in Prod(T ) that
are isomorphic to a particular configuration C , which by hypothesis are all stable. For each
transition in M from configuration Cx to Cy , there is also a transition from SCx to SCy
with the same transition rate. The expected time from configuration Ci to configuration Cj
in the assembly system T is the same as the expected time from state SCi to first reach SCj
in Markov process M (cid:48) since every tra jectory in M corresponds to a tra jectory in M (cid:48) .
Furthermore, all agitation steps in T corresponds to self-loops in M (cid:48) . Therefore, removing
the agitation steps in T is equivalent to removing some self-loops in M (cid:48) and does not change
2
the expected time to transition from one state to another.
Since all of our constructions satisfy the assumption that all producible configurations
have only a single connected stable component, we ignore agitation steps in the time analysis
of the constructions in this paper.
3 Examples
Figures 4 to 8 show a number of examples. Figure 4 depicts a simple assembly system
that grows a line of monomers. Figure 4b shows the initial configuration C0 , which consists
12
Figure 4: Growth of a linear chain. (a) Rule set: Rk = {ri ri = (i, empty, null, (cid:126)x) → (0, i −
1, rigid, (cid:126)x), where 0 < i ≤ k}. (b) Initial configuration of the system. (c) Evolution of a system
tra jectory over time.
of a single monomer with state k ∈ N. This system evolves as shown in Figure 4c, and
uniquely produces a linear chain with k + 1 monomers, with each monomer having a final
state of 0. Since the system involves k sequential events each with unit expected time, it
takes expected time k to complete.
Figure 5: Autonomous unidirectional motion of a walker along a linear track.
(a) Rule set:
r1 = (1, 1, null, − (cid:126)w) → (2, 1, rigid, − (cid:126)w), r2 = (1, 2, rigid, (cid:126)y) → (1, 1, null, (cid:126)y), and r3 = (1, 1, rigid, (cid:126)w) →
(1, 1, rigid, (cid:126)y). (b) Evolution of an example initial configuration. (c) A subroutine abstraction.
Figure 5 depicts the autonomous unidirectional motion of a single-monomer “walker”
along a linear track of monomers. It takes O(n) expected time for a monomer to move n
steps. The rules depicted in Figure 5a implement the subroutine depicted in Figure 5c.
Figure 6 describes insertion of a single monomer between two other monomers, which
occurs in constant expected time.
Figure 7 describes the rotation of a long arm. As the motion of each monomer is
independent of the motion of the other monomers, it takes O(log n) expected time to
complete the rotation of an arm of n monomers. Interestingly, this example captures a kind
of movement and speed that is impossible to achieve with cellular automata, but has an
analog in reconfigurable robotics [5].
Figure 8 shows a simple Turing machine example. The Turing machine program is
stored in the rule set and directs the tape head monomer to walk left or right, and to update
the relevant tape monomer. The Turing machine state is stored in the tape monomer,
currently the Turing machine is in state q . If the Turing machine requires a longer tape,
new tape cell monomers are created to the left or right of the end of tape marker B . This
example shows that the model is capable of algorithmically directed behavior.
13
Initial configurationRules:ki0i-1r_iabwhere i > 0System evolutionkc0k-1r_kr_k-100k-2r_k-2r10000Unique terminal configurationA linear chain of k+1 monomers1111121111111111Rules:Execution:a111111Subroutine:cu1211r22111r11111ur3r2r1r3bInitial configurationUnique terminal configuration1111r1r3 Figure 6: Insertion of a single monomer between two others. (a) Rule set: r1 = (1, empty, null, (cid:126)y) →
(1.1, 0, rigid, (cid:126)y), r2 = (0, x, null, − (cid:126)w) → (0, x.1, rigid, − (cid:126)w), r3 = (1.1, x.1, rigid, (cid:126)x) → (1.1, x.1, null, (cid:126)x),
r4 = (0, x.1, rigid, − (cid:126)w) → (0.1, x.1, rigid, (cid:126)x), r5 = (1.1, 0.1, rigid, (cid:126)y) → (2, 2, rigid, (cid:126)x), and r6 =
(2, x.1, rigid, (cid:126)x) → (2, x, rigid, (cid:126)x). (b) Evolution of an example initial configuration. (c) A subroutine
abstraction.
Figure 7: Rotation of a long arm. (a) Rule set: r1 = (1, 1, rigid, (cid:126)w) → (1, 1, rigid, (cid:126)y). (b) Evolution
of an example initial configuration to a terminal configuration. (c) A subroutine abstraction.
4 System simulation
In this section we show that there is an algorithm that simulates one of the tra jectories
of an assembly system in time O(tn2 ) where t is the number of configuration transitions
in the tra jectory and n is the maximum number of monomers of any configuration in the
tra jectory. This shows that simulation of individual tra jectories is tractable in terms of the
number of rule applications, and in terms of the number of monomers.1 In fact, it is not
difficult to imagine other variations on the model where simulation of a single, non-local,
rule is an intractable problem. The existence of our algorithm gives evidence that our rules,
in particular the movement rule, are in some sense reasonable.
1This forms the basis of our software simulator for the model.
14
Subroutine:Rules:1x1.1x01.1x.101.1x.101.1x.10.122x.122x1x22xr1r2r3r4r5r6x0x.10r22x.12xr611.10r11.1x.11.1x.1r3-w1.10.122r5x.10x.10.1r4yExecutionacbInitial configurationUnique terminal configurationRules:Subroutinec11111111111111111111xr1aExecution:br11111111111111111r1r1r11111111111111111Initial configurationUnique terminal configuration Figure 8: Turing machine example. The Turing machine program is stored in the rule set. New
tape monomers can be created as needed.
4.1 Simulation of a single step
The continuous-time Markov process that describes a tra jectory is simulated using a discrete
time-algorithm. Essentially, the algorithm examines the grid contents and using a local
neighborhood of radius 2, a list of potentially applicable rules are generated. The list also
contains 6n potentially applicable agitation steps (6 directions for each of the n monomers).
All of this can be done in O(n) time. The algorithm then, uniformly at random, picks an
event from the list to apply, if the rule or agitation step can indeed be applied the grid
contents are updated accordingly. Besides the movement rules and agitation steps, the
other rule types can be easily simulated in time O(n), since at most two grid sites are
affected. As described below, movement is simulated in time O(n2 ). Hence a single step is
simulated in time O(n2 ).
Due to its non-local nature, the movement rule is the most complicated rule type to
simulate. Algorithm 4.1 below calculates the moveable set for a given movement rule in
time O(n). We may have to try this algorithm < n times before we can find a non-empty
movable set and can apply a movement rule, or decide that there is no applicable movement
rule. Applying a rule simply involves translating < n monomers by unit distance, which
can be done in O(n) time. Hence movement can be simulated in O(n2 ) time. Agitation is
simulated similarly.
4.2 Computing the movable set
We describe a greedy algorithm for computing the movable set M(C, A, B , (cid:126)v) of monomers
for a movement rule where C is a configuration, A is an arm monomer, B is a base monomer
and (cid:126)v ∈ D the unit vector describing the translation of A. The algorithm takes time linear
in the number of monomers. Figure 3 shows two examples of computing the movable set.
Algorithm 4.1 Compute movable set M(C, A, B , (cid:126)v).
• Step 1. Let M ← {A}, F ← {A}, B ← {}.
• Step 2. Compute the blocking set B for the frontier set F along (cid:126)v , as fol lows.
For each monomer X ∈ F , do:
15
qBB1110000tape headtapeend of tape symboltape cell 1. If (cid:126)p(X ) + (cid:126)v is occupied by Y /∈ M, then B = B ∪ {Y };
2. If X is bonded to Y /∈ M, and if translating X by (cid:126)v without translating Y would
disrupt the bond between X and Y , then B = B ∪ {Y }; (Ignore the special case
where X = A, Y = B ).
• Step 3. Inspect the blocking set:
1. If B ∈ B, return {};
2. If B = {}, return M;
3. Otherwise, let M ← M ∪ B, F ← B, B ← {}, and go to Step 2.
Lemma 4.2 Algorithm 4.1 identifies the movable set M(C, A, B , (cid:126)v) in time O(n), where
n is the number of monomers in C .
Proof: To argue that the algorithm identifies the movable set we consider two cases: the
algorithm completes with (1) a non-empty set, and (2) an empty set.
In Case (1) when the algorithm completes with a non-empty set M, we only need to
prove the following claims: (1.1) M contains A but not B , (1.2) M can move unit distance
along (cid:126)v without causing monomer collision nor bond disruption, and (1.3) M is the minimal
set that satisfies (1.1) and (1.2).
Claim (1.1) follows directly from step 1 and step 3.1.
To prove Claim (1.2), assume, for contradiction, that there exists a monomer Y /∈ M
that blocks a monomer X ∈ M. Therefore, when X gets first incorporated into M, Y must
be incorporated into M in the next round of execution of step 2. This contradicts Y /∈ M.
Therefore, Claim (1.2) must be true.
To prove Claim (1.3), assume, for contradiction, that there exists a set C ⊂ M, A ∈ C
such that C can move by (cid:126)v . The first monomer in M \ C that gets incorporated in M
must block some monomer in C , which contradicts that C can move a unit distance along (cid:126)v .
Therefore, Claim (1.3) holds.
In Case (2), we know from Claim (1.3) that any movable set that contains A must
contain every monomer in M and thus contain B . Therefore, the movable set must be
2
empty.
We note that the nondeterministic choice of which monomer is the arm and which is
the base can make a difference in the resulting configuration: for example switching the
arm and base monomers in Figure 3a will induce a different movable set. In this paper we
do not exploit such asymmetric nondeterministic choices.
5 Efficient growth of simple shapes: lines and squares
In this section, we show how to efficiently construct lines and squares in time and number
of monomer types logarithmic in shape size. We give a fast (logarithmic time) method to
16
Figure 9: Building a line of length n ∈ N by decomposing into O(log n) lines whose lengths are
distinct powers of 2.
synchronize a line of monomers: the procedure detects in logarithmic time in line length
whether all monomers in the line are in the same state. We also give a Chernoff bound
lemma that aids in the time-analysis of these and other systems.
5.1 Line
Theorem 5.1 A line of monomers of length n ∈ N can be uniquely produced in expected
time O(log n) and with O(log n) states.
Proof: We first describe the construction, then prove correctness and conclude with a time
analysis.
{k1 , k2 , . . . , kp}, where (cid:80)
Description: To build a line of length n, from the start monomer s.n, we first (se-
quentially) generate a short line of p = O(log n) monomers with respective states K =
k∈K 2k = n. Figure 9 illustrates this first step. Then, each
monomer with state k ∈ K efficiently builds a line of length 2k as described below.
Figure 10 gives an overview of the main construction, as well as many of the rules. The
idea is to quickly build a line of length 2k , by having the start monomer, s.k , create 2
monomers (one of which is in state k − 1), which in turn create 4 monomers (2 of which
are in state k − 2), and so on until there are 2k monomers with state 0. An overview
of a possible tra jectory of the system is given in Figure 10a3. However, as the model is
asynchronous, most tra jectories are not of this simple form.
The construction can be described using the two subroutines shown in Figure 10a2.
Subroutine (1) consists of a single rule that is applied only once, to the seed monomer.
Subroutine (2) consists of k − 1 sets of rules, one for each x where k > x > 0. A schematic of
one of these k − 1 rule sets is given in Figure 10b, with an example execution of Subroutines
(1) and (2) in Figure 10c.
Subroutine (2) begins with a single pair of monomers with states x, 0 and ends with four
monomers in states x − 1, 0, x − 1, 0. Monomers are shown as left (purple), right (blue) pairs
to aid readability. The rules for Subroutine (2) are given in Figure 10b and an example
can be seen in Figure 10c. The subroutine works as follows. Each monomer on the line
has a left/right component to its state: left is colored purple, right is colored blue. The
initial xleft , 0right monomers send themselves to state x − 1left , 0right while inserting two new
monomers to give the pattern x − 1left , 0right , x − 1left , 0right , as indicated in Subroutine (2).
To achieve this, the initial pair of monomers create a “bridge” of 2 monomers on top, and
17
cRuless.ns.k1s.kis.kis.ki+1where ∑i ki = n, and all ki are distinct powers of 2bs.naSubroutines.k1s.k2s.kps.3s.1n = 11 = 23+21+20 s.0Examples.11 Figure 10: Construction that builds a length 2k line in expected time O(k). (a) Overview: 1
monomer in state x ∈ N creates 2 in state x − 1, and this happens independently in parallel along
the entire line as it is growing. The blue/purple colors are for readability purposes only. (b) Rules
for subroutines (1) & (2), for all x where k > x > 0. (c) Example execution of 13 steps, starting
with a left-right pair of monomers. The “stick and dot” illustrations emphasize bond structure,
representing bonds and monomers respectively. (d) Example configuration, in stick and dot notation,
that emphasizes parallel asynchronous rule applications. Each gray box shows the application of a
subroutine.
18
aConstruction overview0s.4a2s.k a3Overview of one possible trajectory where k = 4Subroutinesa1Initial configurationbRules for Subroutines (1) & (2)k > x > 0 d 302020101010100000000000000000k-1s.k0x0x-10x-1 k > x > 0 r5x-1’x-10’xr2b.1xxExample execution of Subroutine (2)r3b.1b.2b.3r40’0b.3b.4r6x-1’x-1’b.4b.5r70x-1’’x-1’r8b.5b.6x-1’’x-1’’00b.6b.7r9r11r10b.7b.8x-1’’x-1’’x-1’’0.’x-1r14b.2b.8cr10k-1s.kr2b.1x0r3b.2b.3x0r4b.2b.4x0’r5b.2b.4x-1x-1’r120.’x-10.’’x-1r130.’’0r6x-1’b.2b.5x-1r70x-1’’b.2b.5x-1r80x-1’’b.2b.6x-1r90x-1’’b.2b.7x-1r100x-1’’b.2b.8x-1r110x-1b.2b.8x-10.’0x-1b.2b.8x-10.’’r120x-1b.2x-100x-1x-10r13r14Example configuration(2)(1)0x by using movement and appearance rules two new monomers are inserted. The bridge
monomers are then deleted and we are left with four monomers. Throughout execution, all
monomers are connected by rigid bonds so the entire structure is stable. Subroutine (2)
completes in constant expected time 13.
Subroutine (2) has the following properties: (i) during the application of its rules to
an initial pair of monomers xleft , 0right it does not interact with any monomers outside
of this pair, and (ii) a left-right pair creates two adjacent left-right pairs.
Intuitively,
these properties imply that along a partially formed line, multiple subroutines can execute
asynchronously and in parallel, on disjoint left-right pairs, without interfering with each
other.
Correctness: We argue by induction that the line completes with 2k monomers in state
0. The initial rule (Subroutine (1) in Figure 10a2) creates a left-right pair of monomers
k − 1left , 0right ; the base case. For the inductive case, consider an arbitrary, even length, line
of monomers (cid:96)j where all monomers are arranged in left-right pairs of the form xleft , 0right ,
where either x ∈ N. For any left-right pairs of the form 0left , 0right , no rules are applicable
(they have reached their final state). For all other left-right pairs xleft , 0right Subroutine
(2) is applicable. Choose any such left-right pair, and consider the new line (cid:96)j+1 created
after applying Subroutine (2). The new line (cid:96)j+1 is identical to (cid:96)j except that our chosen
pair xleft , 0right has been replaced by x − 1left , 0right , x − 1left , 0right . Line (cid:96)j+1 shares the
following property with line (cid:96)j : all monomers are in left-right pairs. Hence, except for
(already completed) 0left , 0right pairs, Subroutine (2) is applicable to every left-right pair
of (cid:96)j+1 , and so by induction we maintain the property that rules can be correctly applied.
Furthermore, application of Subroutine (2) leaves one 0 state untouched, creates a new 0
state, and creates two new x − 1 states. Hence, eventually we get a line where all states are
0 and no rules are applicable. The fact that the line grown from the monomer with state
s.k has length 2k follows from a straightforward counting argument.
Time analysis: Consider any pair of adjacent monomers 0left , 0right in a final line of
length 2k . The number of rule applications from the start monomer s.k to this pair is k (i.e.
using Figure 10a2, apply Subroutine (2) k times to get this pair). Given that these k rule
applications are applied independently (without interference from other rules that are acting
on other monomers on the line) and in sequence the expected time to generate our chosen
pair is O(k). There are 2k−1 of these 0left , 0right pairs in a final line of length 2k , giving an
O(k log 2k−1 ) = O(k2 ) bound on the expected time for the line to finish. In a line of length
n ∈ N we have O(log n) lines, each of length a power of 2, being generated in parallel (using
the technique in Figure 9), giving an expected time of O(log2 n log log n) for the length n
line to complete. This analysis can be improved using Chernoff bounds. Specifically, in
Lemma 5.2 we choose m = n/2 and a1 , a2 , . . . , am to be the n/2 rule applications that
generate the n/2 pairs of 0left , 0right monomers in the final configuration. Each ai requires
k ≤ 2 log2 n insertions to happen before it. Therefore, the expected time for the line to
finish is O(k) = O(log n).
19
Number of states: Subroutine (1) has 1 rule. Subroutine (2) has O(1) states for each
x ∈ 1, . . . , k − 1, hence the total number of states is O(k).
2
Lemma 5.2 In an assembly system, if there are m rule applications a1 , a2 , . . . , am that
must happen, and
1. the desired configuration is reached as soon as al l m rule applications happen,
2. for any specific rule application ai among those m rule applications, there exist at
most k rule applications r1 , r2 , . . . , rk such that ai = rk and for al l j , rj can be
applied directly after r1 , r2 , . . . , rj−1 have been applied, regard less of whether other
rule applications have happened or not,
3. m ≤ ck for some constant c,
then the expected time to reach the desired configuration is O(k).
Proof: From the assumptions, the time Ti at which the rule application ai happens is
upper bounded by the sum of k mutually independent exponential variables, each with
Prob[Ti > k(1 + δ)] ≤ (cid:0) 1 + δ
(cid:1)k .
mean 1 for every k . Using Chernoff bounds for exponential variables [50], it follows that
eδ
Prob[T > k(1 + δ)] ≤ m(cid:0) 1 + δ
(cid:1)k ≤ ck e− kδ
(cid:1)k ≤ (cid:0) c(1 + δ)
Let T be the time to first reach the desired configuration. From the union bound, we know
that
2 , for all δ ≥ 3.
eδ
eδ
Therefore, the expected time is E [T ] = O(k).
2
For some constructions it is useful to have a procedure to efficiently (in time and states
logarithmic in n) detect when a large number (n) of monomers are in a certain state. Here
we give such a procedure. Specifically, we show that after growing a line, we can use a fast
signaling mechanism to synchronize the states of all the monomers in the completed line.
Theorem 5.3 (Synchronized line) In time O(log n), with O(log n) states, a line of
length n ∈ N can be uniquely produced in such a way that each monomer switches to a
prescribed final state, but only after al l insertions on the line have finished.
Proof: The basic idea is to build a synchronization row of monomers below the line. This
row is grown only in regions of the line that have finished growth (are in state 0), and takes
time O(log n) time to grow. The bond structure of the synchronization row is such that
upon the completion of the entire line, a relative shift of the synchronization row to the line
occurs (in O(1) time), informing the line monomers to switch to a prescribed final state.
Finally, the synchronization row is deleted, in O(log n) time.
20
Figure 11: Synchronization mechanism for Theorem 5.3 that quickly, in O(log n) expected time,
sends a signal to n monomers in a line. Stick and dot notation is used to emphasize the bond
structure throughout. A single movement, or shift, between configurations (7) and (8) sends the
signal to all monomers. The structure maintains stability throughout execution.
Figure 11 describes the synchronization mechanism. Starting from a seed monomer we
grow a line (1), as each monomer on the line reaches state 0 (and so has finished inserting) it
grows a synchronization monomer (in black) below it (2), joined to the line with a rigid bond.
In the following we want to always ensure that the entire structure is stable. Neighboring
synchronization monomers form horizontal rigid bonds (3). Any synchronization monomer
that is joined with its two horizontal neighbors changes its bond to the line from rigid
to flexible (the rightmost monomer is a special case, it changes to flexible immediately
upon bonding with a horizontal neighbor). According to this rule we eventually get to
configuration (7) where the entire synchronization row is bonded to the backbone row by
flexible bonds, except for the leftmost pair. At this time, the leftmost pair is, for the first
time, able to execute a movement rule (8) which shifts the synchronization row to the right,
relative to the backbone row. Backbone monomers can detect this shift.
21
Rigid bondFlexible bondShift monomerInsertion completeSynchronized(1)(2)(3)(4)(5)InsertionShift does not applyShift can now applyBackbone rowSynchronization row(1)(2)(3)(4)(5)(8)(9)(10)(15)(16)Unsynchronized(7)(5)(6)(11)(12)(13)(14)Synchronized From here on the aim is to delete the synchronization row, while maintaining the property
of stability. This is achieved in a manner inversely analogous to before: synchronization
monomers create rigid bonds with the backbone, then delete their bonds to their horizontal
neighbors only when the neighbors have formed vertical rigid bonds.
From Theorem 5.1 the expected time to grow the unsynchronized line, and to then
grow the additional synchronization row is O(log n) (the addition of the synchronization
row requires O(1) for each monomer in the unsynchronized line). The movement rule
that underlies the synchronization then takes expected time O(1), and a further O(log n)
expected time is required to delete the length n synchronization row. the number of states
is dominated by the O(log n) states to build a line (Theorem 5.1), as the synchronization
2
mechanism itself can be executed using O(1) states (Figure 11).
5.2 Square
Square building is a common benchmark problem in self-assembly.
Figure 12: Building a square in O(log n) time, using O(log n) states.
Theorem 5.4 An n × n square can be constructed in time O(log n) using O(log n) states.
Proof: Figure 12 contains an overview. Using the construction in Theorem 5.3, we first
assemble a horizontal backbone line of length n. When a given monomer pair on the line
finishes insertion (i.e. reach states 0, 0), the line then expands by a factor of 3 at that
location. Every third monomer in the expanded backbone grows a vertical line of length n,
the previous expansion ensures that each vertical line has sufficient space to grow. Each
vertical line synchronizes upon completion. This synchronization signals to the backbone
line to contract by factor of 3, essentially bringing all n vertical lines into contact. Adjacent
vertical lines form rigid bonds, so that the final shape is fully connected.
To analyze the expected time to completion we first consider the expected time for
the n vertical rows to be a situation where all synchronization rows have grown and they
are about to apply the synchronization steps. Each synchronization monomer can grow
independently of all others, and depends only on ≤ 4(cid:100)log2 n(cid:101) prior insertion events to
happen. By setting m = n2 − n in Lemma 5.2 and k = O(log n) the expected time is
O(log n). The synchronizations then apply in expected time O(log n), as do the final folding
and bonding steps. The number of states to build and synchronize each line is O(log n), a
2
constant number of other states are used in the construction.
22
Seed(1) Backbone assembly (2) Expansion(3) Growth(4) Contraction(5) Bonding 6 Computable shapes
Let n = (cid:100)log2 n(cid:101) be the length of binary string encoding n ∈ N.
Theorem 6.1 An arbitrary connected computable 2D shape of size ≤ √
n × √
n can be
constructed in expected time O(log2 n + t(n)) using O(s + log n) states. Here, t(n) is the
time required for a program-size s Turing machine to compute, given the index of a pixel n,
whether the pixel is present in the shape.
The remainder of this section contains the proof.
6.1 Construction overview
Figure 13 gives an overview of the construction. We first assemble a binary counter that
writes out the n binary numbers {0, 1 . . . , n − 1}, and where each row of the counter
√
n × √
represents a pixel location in the
n square that contains the final shape. The counter
completes in expected time O(log2 n). The counter has an additional backbone column of
monomers of length n. After the counter is complete, each row of the counter acts as a finite
Turing machine tape: the binary string on the tape represents an input i ∈ {0, . . . n − 1} to
the Turing machine. For each such i, there is a monomer that encodes the Turing program
and acts as a tape head. If the head needs to increase the length of the tape, new monomers
are created beyond the end of the counter row as needed. Eventually the simulated Turing
machine finishes its computation on input i, and the head transmits the yes/no answer to a
single backbone monomer. All Turing machines complete their computation in expected
time O(t(n)), where t(n) is the worst case time for a single Turing machine to finish
on an input of length n = O(log n). The Turing machine head monomers then cause the
deletion of the counter rows. A synchronization on the backbone occurs after all backbone
monomers encode either yes or no. The entire backbone then “folds” into a square, using a
number of parallel “arm rotation” movements. Folding runs in expected time O(log2 n).
After folding, the “no” pixels (monomers) are deleted from the shape in a process called
“carving”, which happens in O(1) expected time. After carving is complete we are left with
the desired connected shape, all in expected time O(log2 n + t(n)) and using O(s + log n)
states.
6.2 Binary counter
Figure 14 gives an overview of a binary counter that efficiently writes out the binary
strings that represent 0 to n − 1 ∈ N in O(log2 n) time, and O(log n) states. The counter
construction builds upon the line construction in Theorem 5.1 (and Figure 10). The
essential idea is to build a line in one direction while simultaneously building counter rows
in an orthogonal direction. The counter begins with a single seed monomer as shown in
Figure 14(0) and ends with a configuration of the form shown in Figure 14(8), growing in
an unsynchronized manner.
23
Figure 13: Construction of an arbitrary connected computable shape.
24
000000000001001000001001110110110111111110111111tttttttt100000000000100100000100111011011011111110111111tttttttttttttttt(2) Binary counter assembly(4) Shape computationSeedS000000000001001000001001110110110111111110111111(6) Prepare for folding(10) Change bond structure}}}}}}......}}}}(1) (11) Carve out shape by deleting monomers(9) Folded structure (8) Folding & synchronization (7) Folding (5) Yes/no pixels (3) Create Turing machine tape heads Figure 14(1)–(6) illustrates the construction by making the (unlikely, but valid) as-
sumption that at configuration (1) we have created a counter that has already counted
the set {0, 1, 2, 3}, and then gives a number configurations along a tra jectory to compute
the set {0, 1, . . . , 8}. (Note that the system is asynchronous so very few tra jectories are of
this nice form). A line is efficiently grown using the technique in Theorem 5.1, but where
the 0 monomers from the line are denoted with a monomer with no state name in the
counter (to simplify the presentation). Starting from configuration (1), insertion events
independently take place across the entire line. Each counter row in (1) is separated by
unit distance, which enables multiple insertion routines to act independently (each uses
a pair of monomers to form “bridge” monomers b.∗, similar to Figure 10). Insertion of a
pair of monomers triggers copying of a counter row, as seen in configuration (4). While a
row is being copied, a 0 monomer is appended to one copy and a 1 monomer is appended
to another copy. Insertion of two monomers and the growth of the new counter row takes
O(log n) expected time: this comes from the fact that insertion works in constant time,
and that copying of the O(log n) monomers takes O(log n) expected time. After copying
and generation of the new 0 or 1 monomers is complete, a signal is sent to the line. The
line can then continue the insertion process. Growing a line takes expected time O(log n),
but we’ve replaced each O(1) time insertion event with an expected time O(log n) copying
process, hence the overall expected time is O(log2 n).
From Theorem 5.1, the number of states to build the backbone line is O(log n), a further
O(1) states can be used to carry out counter row growth and copying.
Correctness for the counter essentially follows from that of the line: We can consider
that each insertion on the line is paused while counter row copying completes. After the
copying the line monomers can continue their insertions, and eventually will complete. The
copying and bit-flipping mechanism guarantees that the rows of the counter encode the
correct bit sequences.
6.2.1 Counter synchronization
After the counter is complete, the backbone synchronizes. More precisely, after the backbone
has grown to its full length n, and all counter rows have finished their final copying (i.e.
configuration (8) in Figure 14 is reached), the synchronization routine from Theorem 5.1 is
executed to inform all backbone monomers that the counter is complete, in expected time
O(log n).
6.3 Turing machine computations
The next part of the construction involves simulating n2 Turing machines (in parallel) in
order to determine which of the n2 pixels in the n × n canvas are in the desired shape, and
which are not.
We begin with the synchronized counter described in the previous section. After
25
Figure 14: A binary counter that efficiently writes out the binary strings that represent all integers
from 0 to n − 1, in O(log2 n) time and O(log n) states. The counter grows in an unsynchronized
fashion, some example configurations are shown in (0)–(8), although many other tra jectories are
possible. A backbone line is efficiently grown using the technique in Theorem 5.1. The configuration
shown in (1) represents a count of 0 to 3. In this example the backbone then expands (6) by a factor
of 2, with each expansion a row of the counter is copied, with a 0 appended to one copy and a 1
appended to another copy (4). To save space the grid is rotated anti-clockwise.
synchronization the counter changes its bond structure so that it is of that shown in
Figure 13(2). Each row can do this independently, by a straightforward application of
Lemma 5.2, this takes O(log n) time.
The counter then expands by a factor of 2 (in the (cid:126)y direction), and each row grows a
single “head” monomer on top as shown in Figure 13(3). The “head” monomer acts as a
Turing machine tape head, as in Figure 8, and treats the counter row as a finite Turing
machine tape. The input to the Turing machine is the binary number i ∈ {0, 1, . . . , n2 − 1}
stored on the tape. The Turing machine program executed by the head is of size s and is
stored explicitly in the rules that are applicable to the head monomer, this gives the O(s)
term of monomer states in the statement of Theorem 6.1. If the Turing machine requires
more than log2 n tape space (the length of the counter row), then new monomers are
grown to the left as required (we assume that the turing machine runs on a single, one-way
infinite tape). When the Turing machine enters an accept or reject state, the head moves
to the right, deleting each tape monomer along the way, and communicates this bit to the
backbone row, as shown in Figure 13(5). Each tape head then deletes itself, and n2 backbone
monomers undergo a synchronization when all tape head monomers are deleted. We are
left with n2 monomers as shown in Figure 13(6), each of which stores a bit representing
whether or not it represents a pixel in the final desired shape.
We have n2 Turing machines to simulate in parallel. In the folding step below we modify
the Turing machine program so each machine first computes a simple inequality, which
causes each simulated Turing machine to run for time t(n) = Ω(log2 n). Then, by setting
26
110101k-2k-2k-2k-2000000101110k-2k-2k-2k-200b.1b.3b.2b.4b.2101110k-2k-200b.4b.200’0k-3b.2b.5k-3’k-3b.2b.5k-3’1110k-2k-3k-300b.2b.2b.500copy, flip last bitk-301b.2b.601b.1b.500k-3’’110k-3’k-3’’00copy, flip last bitk-2b.4b.200010010101100k-311b.2b.611010k-3’’k-3b.2b.2k-3k-30k-300.’’00b.8copy, flip last bits.k(0)(1)(2)(3)(4)(5) (6) (7) (8)1111111111000000000000011101111000010101100111111000000011111111100000000000001000010110001000010110001111111111110000000000001100111110101100111110000111copy, flip last bitcopy, flip last bit0’k-300101110111k-31k-3k-3k-300001010110k-30k-3k-300000000 m = n2 , and k = t(n) in Lemma 5.2, all Turing machines finish their computations in
expected time O(t(n)).
6.4 Folding
In this part of the construction the line of n2 monomers folds itself into an n × n square.
The folding process is outlined in Figure 13(7)–(9), and is accomplished as follows.
In the previous section, we used a Turing machine computation on the ith row of a
counter was to decide whether or not pixel i ∈ {0, 1, . . . n2} is in the final shape. We
use these same Turing machines to carry out an additional computation. We can see in
Figure 13(6)–(9) that the line of monomers is divided into alternating segments, each of
length n, that fold in one of two ways. In particular, the monomers highlighted in green
each carry out a sequence of 3 clockwise rotations with respect to their left neighbor. This
can be done using rules similar to the single rule in the arm rotation example shown in
Figure 7. For each monomer to know whether it should rotate (green) or not (blue), we have
the Turing machine check if i satisfies the inequality (2j − 1)n ≤ i < 2jn, and if so this bit
is communicated to the ith backbone monomer (along with the yes/no pixel information).
Then after the synchronization described in the previous section, monomers with indices i
that satisfy the inequality rotate as shown in Figure 13(6)–(9). A synchronization takes place
for each rotating arm, after the second rotation of each monomer, as shown in Figure 13(8).
The expected time for the Turing machine monomer head to compute the inequality is
Ω(log n), as it requires reading the entire input (given our orientation of binary strings),
and this is already accounted for in Section 6.3 above. The expected time to rotate an arm
(twice) so that it is in the position shown in Figure 13(8) is O(log n), for each individual arm.
By letting m = n2/2 and k = log n in Lemma 5.2, the expected time for all n arms to rotate
to the configuration shown in Figure 13(8) is O(log n). Then we apply n synchronization
steps, each of which runs in expected time O(log n), giving an expected time of O(log2 n).
As with the first rotations, the final rotations occur in expected time O(log n), to give a
total expected time of O(log2 n) for the folding step. The number of states used for folding
is O(1).
6.5 Bonding and carving
The goal here is to delete those pixels that should not be in the final stage, and to do this
in a way such that the shape does not become unintentionally disconnected (otherwise if
we delete too early, e.g. before all “arms” have folded or all bonds have formed, part of the
shape could become disconnected). In Figure 13 “yes” pixels belonging to the final shape
are highlighted in gray, and “no” pixels that should be deleted are in white. After folding,
i.e. when green and blue outlined monomers begin to come into horizontal ((cid:126)x) contact, as
shown in Figure 13(9), these monomers bond to all of their neighbors. Then the following
distributed carving algorithm is executed, a “no” pixel is deleted if and only if either (a) it
27
Figure 15: Construction of an n × n pattern in expected time polylogarithmic in n, without growing
outside the pattern borders. No long range communication (synchronization) is used and so the
computation happens in a completely asynchronous and distributed fashion. Final configuration
adapted from M. C. Escher, Sky & Water I, woodcut, 1938.
is bonded to its 6 neighbors and they have the maximum number of bonds to each other
(b) all neighbors that have not been deleted already are bonded to it and to each other.
This procedure prevents the shape from becoming disconnected by the following argu-
ment. Any two adjacent regions, one containing as yet unbonded monomers, and the other
containing deleted monomers are bordered by a connected path of rigid bonds. The only
way for monomers on this rigid path to be deleted are if neighboring unbonded monomers
become bonded, thereby maintaining the connectedness of the boundary and making more
monomers available for safe deletion.
Assuming folding has completed, bonding is a local operation that for each individual
monomer runs in time O(1). The same is true for carving. The expected time for all
monomers to bonding and carving is O(log n).
This completes the construction for Theorem 6.1.
7 Efficient computation of patterns
Let n = (cid:100)log2 n(cid:101).
Theorem 7.1 An arbitrary finite computable 2D pattern of size ≤ n × n, where n = 2p , p ∈
N, with pixels whose color is computable on a polynomial time O(log(cid:96) n) (inputs are of
length O(log n)), linear space O(n), program-size s Turing machine, can be constructed in
expected time O(log(cid:96)+1 n), with O(s + log n) monomer states and without growing outside
the pattern borders. Moreover, this can be done without explicitly using synchronization.
The construction is described in the remainder of this section. Figure 15 gives an
overview. For simplicity, the figure is drawn on a square grid. It should be noted that the
construction occurs in a completely asynchronous distributed fashion. For ease of exposition,
28
S0 0 ... 00 0 ... 10 0 ... 01 1 ... 1......i0 0...0......0 0...11 1...11i2}log nnlog n...i, j...0 0...00 0...11 1...11i2}log n 1 2 . . . . . . nnlog n11 . . . . . . nn. . . . . . (5)(4)......i, j0 0...00 0...11 1...11i2}log n nnlog n...i , j 1 1i , j 2 2i , j nlog nnlog n(2)(1)(3) we first describe a construction where n is a double power of 2, i.e. n = 22p for p ∈ N, and
then in Section 7.3 we describe how to modify it for n = 2p .
Figure 16: Sketch of a procedure to efficiently reconfigure a counter into a straight line.
7.1 Vertical and horizontal binary counters
The construction begins by growing a counter, shown in Figure 15(1), that contains all
binary numbers from 1 to n/ log2 n. The counter is described in Section 6.2, and illustrated
in Figure 14. The only difference here is that we omit the final synchronization steps from
Section 6.2. Notice that such a counter contains exactly n monomers; i.e. n/ log2 n rows
each of length log2 (n/ log2 n). The counter rows then expand so that they are of length
log2 n (we omit the details, but this is easy to achieve with the number of states permitted
in the theorem statement). As counter rows finish, we want to rotate them so that they
stand in a column as shown in Figure 15(2), although due to the asynchronous nature of
the computation they will most likely not do this as shown. As can be seen in Figure 14,
the counter is stable throughout its entire construction.
Figure 16 shows how this is accomplished. After a counter row completes, the backbone
monomer that is attached to that counter row expands vertically by log2 n monomers. Then
the counter row rotates from a horizontal, to a vertical position as shown in Figure 16.
Finally the binary string stored in the counter row is copied to the expanded backbone
monomers, and the counter row deletes itself. All of this is carried out while maintaining
stability (Definition 2.3). By the time all rows rotate we have a backbone of height n.
After row i rotates to the vertical orientation, it immediately initiates growth of a second,
horizontal, counter which counts while copying the binary number i. Note that i can not
be stored in single monomer (as this would require Ω(n) states for the entire construction),
instead the entire strip of log2 n monomers encoding i is copied as the second counter grows.
29
010101100101011000001111iii010101100101011000001111iiii010101100101011000001111iiiiiiiiiiiiiiiiiiiiiiii010101100101011000001111iiiiiiiiiiiiii011001010110101010000111(4)(2)(1)(3)(5) Figure 17: A strip of monomers of length log2 n. (1) The entire strip acts as an input tape and work
tape for a Turing machine that uses log2 n, whose head is denote in green. (2) After the first Turing
machine has completed its computation the second machine is initialized, and so on until machine
log2 n completes its computation. (4) The output, yes or no, from each machine is denoted as solid
green or white.
A sketch is shown in row i of Figure 15(3). This second counter works as follows. Its
backbone line grows horizontally, starting from the top monomer of the i strip. During
each insertion event, the log2 n monomers that encode i are copied. At the same time
these log2 n monomers are used to encode the values being generated by second counter
(essentially, while copying i we execute the copying and bit flip idea seen in Figure 14 to get
a new value j ). By the time row i finishes it contains n vertical strips of log2 n monomers,
and each strip encodes a distinct pair (i, j ) where j ∈ {0, 1, . . . n − 1}.
To find the expected time to complete all strips notice that each strip (i, j ) can grow
independently from all others, and each strip depends on only O(log2 n) events to happen in
order for that strip to be complete. Hence we can apply Lemma 5.2 by setting m = n2/ log n
(the number of strips), and k = O(log2 n), to get an expected time of O(log2 n) for all strips
to complete.
7.2 Turing machine computations
We will treat each strip (i, j ) as a Turing machine tape. After a strip completes, a signal is
sent from the top to the bottom of the strip, successively writing out one of the integers
p ∈ {1, 2, . . . , log2 n} in each of the log2 n monomers. This signal also tells the topmost
monomer that it now encodes a Turing machine tape head, as shown in green in Figure 17(1).
The encoded tape head moves up and down the strip as required. By the theorem hypothesis,
the Turing machine requires at most log2 n space, which is exactly the strip length. These
Turing machine are assumed to takes as input three integers (i, j, p) which provide a unique
coordinate for each monomer in the entire n × n pattern. In time polynomial in its input
30
...i, j 1,0,1, * 1,1,5, * 1,0,4, * 0,1,3, * 1,1,2, * 0,1,6, * 1,0,log n, * 0,0,8, * 1,1,7, * ...i, j 1,0,1, YES 1,1,5, * 1,0,4, * 0,1,3, * 1,1,2, * 0,1,6, * 1,0,log n, * 0,0,8, * 1,1,7, * ...i, j 1,0,1, YES 1,1,5, YES 1,0,4, YES 0,1,3, NO 1,1,2, NO 0,1,6, YES 1,0,log n, YES 0,0,8, NO 1,1,7, YES ...i, j .........(4)(2)(1)(3) length n = log2 n, the first Turing machine decides whether pixel (i, j, 1) is black or white,
communicates this bit to the topmost monomer, and moves on to the second from top
monomer, which has coordinate (i, j, 2). This process continues until all log2 n monomers
in the strip are colored either black or white.
Each Turing machine runs in time O(log(cid:96) n), which takes expected time O(log(cid:96) n) to
simulate. We can apply the Chernoff bound in Lemma 5.2 by setting m = n2/ log n (the
number of strips) and k = O(log(cid:96)+1 n) (the expected time for all log2 n Turing machines
to finish on a single strip) to get an expected running time of O(log(cid:96)+1 n) for all Turing
machine computations to complete.
The overall time bound for the entire construction is O(log2 n + log(cid:96)+1 n) = O(log(cid:96)+1 n)
since we know (cid:96) ≥ 1 (because our Turing machines are required to read their entire input).
7.3 Patterns with diameter a power of 2
The previous construction works for when n is a double power of 2. The following text shows
how to modify the construction so it works for n being a single power of 2, i.e. n = 2p ∈ N.
In this case, the counter that builds Figure 15(1) is modified so that it does not produce
the final (bottom) row with index (cid:100)n/ log n(cid:101). Let k = n − (cid:98)n/ log n(cid:99) log n, and observe
that k < log n. Now, the bottoms row has index b = (cid:98)n/ log n(cid:99). Row b triggers growth of
a counter as before, however when each column (of length log2 n) of the counter finishes,
it grows an additional k monomers, each with a unique id i ∈ {1, . . . , k}. This results in
a counter of size n × (k + log n). When all counters complete in the entire construction,
the canvas is of size n × (k + (cid:98)n/ log n(cid:99) log n) = n × n. The Turing machine computations
proceed as follows. As each column of the counter in row b completes, the log n Turing
machine computations in that column proceed as before. When they are finished, an
additional k Turing machine computations are triggered, which run one after the other, and
use k + log n workspace in the column (which is more than enough), and the unique IDs of
the k extra monomers in order to decide whether or not each of those k pixels belong in
the pattern. These additional aspects of the construction merely add a constant factor to
the time analysis.
8 Discussion and future work
We have introduced a model of computation called the nubot model, and explored its ability
to construct shapes and patterns. We have shown that the model is capable of efficiently
generating a wide variety of shapes and patterns exponentially quickly. The intention for
our model is to explore the abilities of molecules to compute in ways that are seen in nature
and that we are starting to see in the laboratory. This perspective leaves a lot of room for
future work.
One interesting direction is to explore the algorithmic limits of dynamic structures. In
this paper we have used our active self-assembly model to grow shapes and patterns that
31
are ultimately static. One can also consider shapes and patterns that are forever dynamic.
Cellular automata are a well-studied model from this point of view, although they are
incapable of expressing our notion of movement and active self-assembly. What kinds of
dynamic structural systems can, and can not, be modeled by nubots?
One possible ob jection to our model, on physical grounds, is the lack of any realistic
notion of persistence length (which is also absent from many models of self-assembly). On
the one hand, it is clear that there are natural and artificial structures of high aspect ratio
coupled with high tensile strength or persistence length (hair, actin filaments, microtubules).
On the other hand, ‘high’ does not mean ‘infinite’ ! One could introduce complexity measures
of tensile strength or persistence length and analyze the capabilities of the nubots model
with respect to these resources. The important point here would be to appropriately
define these measures so that they capture what is observed at the molecular scale in the
laboratory.
For the topic of tensile strength, one could take inspiration from cellular migration and
adhesion in developmental biology: nubot monomers could have variable strength bonds
which break if there is enough movement in one direction. For example, bonds could have
strength s ∈ [0, 1] ⊆ R where 0 is not bonded, 1 is fully bonded, and other values are of
intermediate strength. Ob jects are pulled apart if enough movement rules are applicable
and so that their bond strength, or tensile strength, can not overcome the strength of
movement. What are the classes of systems that can and can not be built under such
constraints? If we have to pay for bond strength, in general, is it possible to place bounds
on this cost in terms of the shapes, patterns or dynamics we wish to model?
As noted in the introduction, the field of reconfigurable robotics considers a wide range
of models that share a number of features with our nubot self-assembly model. In particular,
reconfigurable modular robots with a similar long-range movement primitive to ours can
achieve arbitrary reconfiguration in time logarithmic in shape size [5], and are capable of
linear parallel time reconfiguration with more realistic physical constraints [4]. It remains
as future work to compare such models to ours. What would be particularly interesting
would be to explore the differences in model capabilities that are solely due to the inherent
differences in macro-scale and molecular-scale self-assembly (in molecular systems we are
typically unconcerned with gravity and friction; energy in the form of fuel molecules may
be freely available in the environment; temperature, brownian motion and other forms of
agitation play a ma jor role).
There are also some more technical questions arising from our work. We use a Chernoff
bound in Lemma 5.2 to simplify the time analysis of our constructions. However, many
assembly systems that are expressible in our model do not satisfy the conditions of this
lemma. It would be nice to find other tools, perhaps more general, to aid in the time
analysis of nubots systems. For example, starting from a single monomer, if the longest
sequence of rule applications that can lead to some ‘terminal’ monomer type is k , then is
O(k) the expected time for all rules to complete?
The pattern construction in Section 7 terminates with an assembly that is not completely
32
connected (however, it is stable and connected). In that construction we intentionally
did not use synchronization over long distances, but by using synchronization it is indeed
possible (and relatively easy) to modify the final pattern so that it is completely connected.
However, given that synchronization has such power, it is interesting to ask what can
be done in its absence. Without using synchronization, or any similar form of rapid
communication over long (> log n) distances, is it possible to deterministically assemble an
n × n completely-connected square in time polylogarithmic in n?
Without the movement rule, the nubot model is a kind of asynchronous and nonde-
terministic cellular automaton. Thus in the absence of movement in our model, cellular
automata are a good starting point to assess its computational complexity (how efficiently
can problems be solved?). However, with movement, it is clear that our model can carry
out certain tasks that cellular automata, or indeed Turing machines, can not (even under
reasonable encodings). What are the upper bounds and lower bounds on the computa-
tional complexity of our model? For example, Section 4 gives a polynomial (in nubots
time plus number of monomers) time algorithm for simulating a nubots tra jectory. If
nubots time is merely polylogarithmic in the maximum number of monomer types, then
is it possible to simulate nubots in polylogarithmic time on a parallel computer (e.g. on
polylogarithmic-depth Boolean circuits)?
One could consider variations on the rules. Already it is the case that movement
rules facilitate non-local interactions. One could take this even further, as we now discuss.
Consider a variant on the movement rule where the application of a movement rule r
remotely triggers the application of another movement rule r (cid:48) . More precisely, let A be
an arm monomer bonded to base monomer B , and let M(C, A, B , (cid:126)v) (cid:54)= {} be the movable
set for the application of rule r to monomers A, B . Also, in the same configuration C ,
there is another pair of bonded monomers D , E , with an applicable movement rule r (cid:48) , and
where D ∈ M(C, A, B , (cid:126)v) and E /∈ M(C, A, B , (cid:126)v), and where r (cid:48) translates D by (cid:126)v . When
r is applied it triggers the application of r (cid:48) : i.e. when monomer D moves because of the
remote application of r, both D and E change state as if rule r (cid:48) was applied; hence r and
r (cid:48) are applied at the same time. We do not consider this kind of remote rule triggering in
this paper, but we mention it here as a way for potential future work to model non-local
interactions that can occur due to movement. This is one possible way to model systems
where interactions occur in a decentralized manner.
The model uses both rigid and flexible bonds. Besides simple examples given early in
the paper, the only construction that uses flexible bonds is synchronization, and it turns
out that one can design a synchronization routine that does not use flexible bonds that
works in the presence of agitation (by building a rigid row immediately below, and parallel
to, the synchronization row—see Figure 11—that stops monomers floating away). What
classes of (perhaps scale-invariant) shapes can be assembled by exploiting flexible bonds
that can not be assembled otherwise?
Finally, since the model is directly inspired by a wide election of natural and artificial
molecular systems, this begs the question: can nubots be implemented at the molecular
33
scale in the laboratory?
Acknowledgments
Many thanks to Niles Pierce and Patrick Mullen for valuable discussions and input. We
also thank Moya Chen, Doris Xin, Joseph Schaefer, and Andrew Winslow for stimulating
and fruitful discussions.
References
1. Z. Abel, N. Benbernou, M. Damian, E. Demaine, M. Demaine, R. Flatland, S. Komin-
ers, and R. Schweller. Shape replication through self-assembly and RNase enzymes.
In SODA 2010: Proceedings of the Twenty-first Annual ACM-SIAM Symposium on
Discrete Algorithms, Austin, Texas, 2010.
2. L. M. Adleman, Q. Cheng, A. Goel, and M.-D. Huang. Running time and program
size for self-assembled squares. In STOC 2001: Proceedings of the thirty-third annual
ACM Symposium on Theory of Computing, pages 740–748, Hersonissos, Greece, 2001.
ACM.
3. G. Aggarwal, Q. Cheng, M. H. Goldwasser, M.-Y. Kao, P. M. de Espanes, and R. T.
Schweller. Complexities for generalized models of self-assembly. SIAM Journal on
Computing, 34:1493–1515, 2005.
4. G. Aloupis, S. Collette, M. Damian, E. Demaine, R. Flatland, S. Langerman,
J. O’rourke, V. Pinciu, S. Ramaswami, V. Sacrist´an, and S. Wuhrer. Efficient
constant-velocity reconfiguration of crystalline robots. Robotica, 29(1):59–71, 2011.
5. G. Aloupis, S. Collette, E. D. Demaine, S. Langerman, V. Sacrist´an, and S. Wuhrer.
Reconfiguration of cube-style modular robots using O(log n) parallel moves. In Pro-
ceedings of the 19th Annual International Symposium on Algorithms and Computation
(ISAAC 2008), pages 342–353, Gold Coast, Australia, December 15–17 2008.
6. E. Andersen, M. Dong, M. Nielsen, K. Jahn, R. Subramani, W. Mamdouh, M. Golas,
B. Sander, H. Stark, C. Oliveira, et al. Self-assembly of a nanoscale DNA box with a
controllable lid. Nature, 459(7243):73–76, 2009.
7. D. Angluin, J. Aspnes, and D. Eisenstat. Fast computation by population protocols
with a leader. Distributed Computing, pages 61–75, 2006.
8. R. D. Barish, P. W. K. Rothemund, and E. Winfree. Two computational primitives
for algorithmic self-assembly: Copying and counting. Nano Lett., 5:2586–2592, 2005.
9. R. D. Barish, R. Schulman, P. W. K. Rothemund, and E. Winfree. An information-
bearing seed for nucleating algorithmic self-assembly. Proceedings of the National
Academy of Sciences, 106(15):6054, 2009.
10. J. Bath, S. Green, and A. Turberfield. A free-running DNA motor powered by a
nicking enzyme. Angewandte Chemie International Edition, 44:4358–4361, 2005.
34
11. J. Bath and A. Turberfield. DNA nanomachines. Nature Nanotechnology, 2:275–284,
2007.
12. F. Becker, E. Remila, and I. Rapaport. Self-assemblying classes of shapes, fast and
with minimal number of tiles. In Proceedings of the 26th Conference on Foundations
of Software Technology and Theoretical Computer Science (FSTTCS 2006), volume
4337 of LNCS, pages 45–56. Springer, Dec. 2006.
13. Z. Butler, R. Fitch, and D. Rus. Distributed control for unit-compressible robots: goal-
recognition, locomotion, and splitting. IEEE/ASME Transactions on Mechatronics,
7:418–430, 2002.
14. S. Cannon, E. Demaine, M. Demaine, S. Eisenstat, M. Patitz, R. Schweller, S. Summers,
and A. Winslow. Two hands are better than one (up to constant factors). In STACS
2013: Proceedings of the 30th International Symposium on Theoretical Aspects of
Computer Science, To appear.
15. B. Chakraborty, R. Sha, and N. Seeman. A DNA-based nanomechanical device with
three robust states. Proceedings of the National Academy of Sciences, 105(45):17245,
2008.
16. H. Chandran, N. Gopalkrishnan, and J. Reif. Tile complexity of approximate squares.
Algorithmica, pages 1–17, 2012.
17. H. Chen and M. Kao. Optimizing tile concentrations to minimize errors and time for
dna tile self-assembly systems. DNA Computing and Molecular Programming, pages
13–24, 2011.
18. E. Demaine, M. Demaine, S. Fekete, M. Ishaque, E. Rafalin, R. Schweller, and
D. Souvaine. Staged self-assembly: nanomanufacture of arbitrary shapes with O(1)
glues. Natural Computing, 7(3):347–370, 2008.
19. E. Demaine, M. Demaine, S. Fekete, M. Patitz, R. Schweller, A. Winslow, and
D. Woods. One tile to rule them all: Simulating any Turing machine, tile assembly
system, or tiling system with a single puzzle piece. Arxiv preprint arXiv:1212.4756,
Dec. 2012.
20. E. D. Demaine, S. Eisenstat, M. Ishaque, and A. Winslow. One-dimensional staged
self-assembly. In Proceedings of the 17th International Conference on DNA Computing
and Molecular Programming (DNA 2011), volume 6937 of Lecture Notes in Computer
Science, pages 100–114, Pasadena, California, Sept. 2011.
21. E. D. Demaine, M. J. Patitz, R. T. Schweller, and S. M. Summers. Self-assembly
of arbitrary shapes using RNAse enzymes: Meeting the Kolmogorov bound with
small scale factor. In Proceedings of the 28th International Symposium on Theoretical
Aspects of Computer Science (STACS 2011), pages 201–212, March 2011.
22. B. Ding and N. Seeman. Operation of a DNA robot arm inserted into a 2D DNA
crystalline substrate. Science, 314(5805):1583, 2006.
23. D. Doty. Randomized self-assembly for exact shapes. SIAM Journal on Computing,
39:3521, 2010.
24. D. Doty. Theory of algorithmic self-assembly. Communications of the ACM, 55:78–88,
35
2012.
25. D. Doty, L. Kari, and B. Masson. Negative interactions in irreversible self-assembly.
In DNA Computing and Molecular Programming, Lecture Notes in Computer Science,
pages 37–48. Springer, 2011.
26. D. Doty, J. Lutz, M. Patitz, R. Schweller, S. Summers, and D. Woods. The tile
assembly model is intrinsically universal.
In FOCS 2012: Proceedings of the 53th
Annual IEEE Symposium on Foundations of Computer Science, pages 302–310, New
Brunswick, New Jersey, Oct. 2012.
27. H. Ehrig. Introduction to the algebraic theory of graph grammars (a survey). Lecture
Notes in Computer Science, 73:1–69, 1979.
28. L. Feng, S. H. Park, J. H. Reif, and H. Yan. A two-state DNA lattice switched by
DNA nanoactuator. Angew. Chem. Int. Ed., 42:4342–4346, 2003.
29. F. Foster, M. Zhang, A. Duckett, V. Cucevic, and C. Pavlin.
In vivo imaging of
embryonic development in the mouse eye by ultrasound biomicroscopy. Investigative
ophthalmology & visual science, 44(6):2361, 2003.
30. B. Fu, M. Patitz, R. Schweller, and B. Sheline. Self-assembly with geometric tiles.
In The 39th International Col loquium on Automata, Languages and Programming
(ICALP 2012), volume 7391 of Lecture Notes in Computer Science, pages 714–725.
Springer, 2012.
31. K. Fujibayashi, R. Hariadi, S. Park, E. Winfree, and S. Murata. Toward reliable
algorithmic self-assembly of DNA tiles: A fixed-width cellular automaton pattern.
Nano Letters, 8(7):1791–1797, 2007.
32. D. Gillespie. A rigorous derivation of the chemical master equation. Physica A:
Statistical Mechanics and its Applications, 188(1):404–425, 1992.
33. R. Goodman, M. Heilemann, S. Doose, C. Erben, A. Kapanidis, and A. Turber-
field. Reconfigurable, braced, three-dimensional DNA nanostructures. Nature
nanotechnology, 3(2):93–96, 2008.
34. S. J. Green, J. Bath, and A. J. Turberfield. Coordinated chemomechanical cy-
cles: A mechanism for autonomous molecular motion. Physical Review Letters,
101(23):238101, 2008.
35. H. Gu, J. Chao, S. Xiao, and N. Seeman. A proximity-based programmable DNA
nanoscale assembly line. Nature, 465(7295):202–205, 2010.
36. D. Han, S. Pal, Y. Liu, and H. Yan. Folding and cutting DNA into reconfigurable
topological nanostructures. Nature nanotechnology, 5(10):712–717, 2010.
37. N. Jonoska and D. Karpenko. Active tile self-assembly, self-similar structures and
recursion. arXiv preprint arXiv:1211.3085, 2012.
38. M. Kao and R. Schweller. Reducing tile complexity for self-assembly through
temperature programming. In Proceedings of the seventeenth annual ACM-SIAM
symposium on Discrete algorithm, pages 571–580. ACM, 2006.
39. M. Kao and R. Schweller. Randomized self-assembly for approximate shapes. Au-
tomata, Languages and Programming, pages 370–384, 2008.
36
40. E. Klavins. Directed self-assembly using graph grammars.
In Foundations of
Nanoscience: Self Assembled Architectures and Devices, Snowbird, UT, 2004.
41. E. Klavins, R. Ghrist, and D. Lipsky. Graph grammars for self-assembling robotic
systems. In Proceedings of the International Conference on Robotics and Automation,
pages 5293–5300, 2004.
42. T. Liedl and F. Simmel. Switching the conformation of a DNA molecule with a
chemical oscillator. Nano letters, 5(10):1894–1898, 2005.
43. A. Lindenmayer. Mathematical models for cellular interactions in development I.
Filaments with one-sided inputs. Journal of theoretical biology, 18(3):280–299, 1968.
44. D. Lubrich, J. Lin, and J. Yan. A contractile DNA machine. Angewandte Chemie
International Edition, 47(37):7026–7028, 2008.
45. R. Luecke, W. Wosilait, and J. Young. Mathematical modeling of human embryonic
and fetal growth rates. Growth Development and Aging, 63:49–59, 1999.
46. K. Lund, A. Manzo, N. Dabby, N. Michelotti, A. Johnson-Buck, J. Nangreave,
S. Taylor, R. Pei, M. Sto janovic, N. Walter, E. Winfree, and H. Yan. Molecular
robots guided by prescriptive landscapes. Nature, 465(7295):206–210, 2010.
47. C. Mao, W. Sun, Z. Shen, and N. Seeman. A nanomechanical device based on the
B-Z transition of DNA. Nature, 397(6715):144–146, 1999.
48. M. Marini, L. Piantanida, R. Musetti, A. Bek, M. Dong, F. Besenbacher, M. Lazzarino,
and G. Firrao. A revertible, autonomous, self-assembled DNA-origami nanoactuator.
Nano Letters, 2011.
49. D. Morris, M. Grealy, H. Leese, and G. R. Centre. Cattle embryo growth, development
and viability. Teagasc, Ireland, 2001. Pro ject No. 4388, Beef Production Series No.
36.
50. R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press,
1995.
51. S. Murata and H. Kurokawa. Self-reconfigurable robots. Robotics & Automation
Magazine, IEEE, 14(1):71–78, 2007.
52. T. Omabegho, R. Sha, and N. Seeman. A bipedal DNA Brownian motor with
coordinated legs. Science, 324(5923):67, 2009.
53. J. Padilla, W. Liu, and N. Seeman. Hierarchical self assembly of patterns from the
Robinson tilings: DNA tile design in an enhanced tile assembly model. Natural
Computing, pages 1–16, 2011.
54. J. Padilla, M. Patitz, R. Pena, R. Schweller, N. Seeman, R. Sheline, S. Summers,
and X. Zhong. Asynchronous signal passing for tile self-assembly: Fuel efficient
computation and efficient assembly of shapes. Arxiv preprint arXiv:1202.5012, 2012.
55. M. Patitz. An introduction to tile-based self-assembly.
In Unconventional Com-
putation and Natural Computation, volume 7445 of LNCS, pages 34–62. Springer,
2012.
56. M. Patitz, R. Schweller, and S. Summers. Exact shapes and turing universality
In DNA Computing and Molecular
at temperature 1 with a single negative glue.
37
Programming, Lecture Notes in Computer Science, pages 175–189. Springer, 2011.
57. M. Patitz and S. Summers. Identifying shapes using self-assembly. In Proceedings of
the Twenty First International Symposium on Algorithms and Computation (ISAAC
2010), volume 6507, pages 458–469. Springer, 2010.
58. P. Prusinkiewicz and A. Lindenmayer. The algorithmic beauty of plants. Springer-
Verlag, 1990.
59. J. Reif, S. Sahu, and P. Yin. Complexity of graph self-assembly in accretive systems
and self-destructible systems.
In Eleventh International Meeting on DNA Based
Computers, volume 3892 of Lecture Notes in Computer Science, pages 257–274.
Springer, 2006.
60. J. Reif and S. Slee. Optimal kinodynamic motion planning for 2d reconfiguration of
self-reconfigurable robots. Robot. Sci. Syst, 2007.
61. P. Rothemund, N. Papadakis, and E. Winfree. Algorithmic self-assembly of DNA
Sierpinski triangles. PLoS Biology, 2:2041–2053, 2004.
62. P. W. Rothemund. Folding DNA to create nanoscale shapes and patterns. Nature,
440(7082):297–302, 2006.
63. P. W. K. Rothemund and E. Winfree. The program-size complexity of self-assembled
squares (extended abstract). In Proceedings of the thirty-second annual ACM sympo-
sium on Theory of computing, pages 459–468. ACM Press, 2000.
64. D. Rus and M. Vona. Crystalline robots: Self-reconfiguration with compressible unit
modules. Autonomous Robots, 10(1):107–124, 2001.
65. K. Saitou. Conformational switching in self-assembling mechanical systems. IEEE
Transactions on Robotics and Automation, 15(3):510–520, 1999.
66. W. Sherman and N. Seeman. A precisely controlled DNA biped walking device. Nano
Letters, 4(7):1203–1207, 2004.
67. J. S. Shin and N. A. Pierce. A synthetic DNA walker for molecular transport. Journal
of the American Chemical Society, 126:10834–10835, 2004.
68. D. Soloveichik, M. Cook, E. Winfree, and J. Bruck. Computation with finite stochastic
chemical reaction networks. Natural Computing, 7(4):615–633, 2008.
69. D. Soloveichik and E. Winfree. Complexity of self-assembled shapes. SIAM Journal
on Computing, 36(6):1544–1569, 2007.
70. S. Summers. Reducing tile complexity for the self-assembly of scaled shapes through
temperature programming. Algorithmica, pages 1–20, 2012.
71. Y. Tian, Y. He, Y. Chen, P. Yin, and C. Mao. A DNAzyme that walks processively
and autonomously along a one-dimensional track. Angewandte Chemie International
Edition, 44(28):4355–4358, 2005.
72. A. J. Turberfield, J. C. Mitchell, B. Yurke, A. P. Mills, Jr., M. I. Blakey, and
F. C. Simmel. DNA fuel for free-running nanomachines. Physical Review Letters,
90(11):118102–1–4, 2003.
73. S. Venkataraman, R. Dirks, P. Rothemund, E. Winfree, and N. Pierce. An autonomous
polymerization motor powered by DNA hybridization. Nature Nanotechnology, 2:490–
38
494, 2007.
74. H. Wang. Proving theorems by pattern recognition II. Bel l Systems Technical Journal,
40:1–41, 1961.
75. E. Winfree. On the computational power of DNA annealing and ligation.
In
R. Lipton and E. Baum, editors, DNA Based Computers, pages 199–221. American
Mathematical Society, Providence, RI, 1996.
76. E. Winfree. Algorithmic Self-Assembly of DNA. Ph.D. thesis, California Institute of
Technology, 1998.
77. E. Winfree.
Simulations of computing by self-assembly. Technical Report CS-
TR:1998.22, Caltech, 1998.
78. E. Winfree and R. Bekbolatov. Proofreading tile sets: Error correction for algorithmic
self-assembly.
In DNA Computing: 9th International Workshop on DNA Based
Computers, DNA9, volume 2943 of Lecture Notes in Computer Science, pages 126–144.
Springer, 2004.
79. E. Winfree, F. Liu, L. Wenzler, and N. C. Seeman. Design and self-assembly of
two-dimensional DNA crystals. Nature, 394:539–544, 1998.
80. M. Yim, W. Shen, B. Salemi, D. Rus, M. Moll, H. Lipson, E. Klavins, and G. Chirikjian.
Modular self-reconfigurable robot systems. Robotics & Automation Magazine, IEEE,
14(1):43–52, 2007.
81. P. Yin, H. M. T. Choi, C. R. Calvert, and N. A. Pierce. Programming biomolecular
self-assembly pathways. Nature, 451:318–322, 2008.
82. P. Yin, H. Yan, X. Daniell, A. J. Turberfield, and J. Reif. A unidirectional DNA walker
that moves autonomously along a track. Angew. Chem. Int. Ed., 43:4906–4911,
2004.
83. B. Yurke, A. J. Turberfield, A. P. Mills, Jr., F. C. Simmel, and J. L. Nuemann. A
DNA-fuelled molecular machine made of DNA. Nature, 406:605–608, 2000.
84. Z. Zhang, E. Olsen, M. Kryger, N. Voigt, T. Tørring, E. Gultekin, M. Nielsen, R. Mo-
hammadZadegan, E. Andersen, M. Nielsen, J. Kjems, V. Birkedal, and K. Gothelf.
A DNA tile actuator with eleven discrete states. Angewandte Chemie International
Edition, 50(17):3983–3987, 2011.
39
|
1811.02457 | 2 | 1811 | 2019-05-29T12:39:32 | Tunneling on Wheeler Graphs | [
"cs.DS"
] | The Burrows-Wheeler Transform (BWT) is an important technique both in data compression and in the design of compact indexing data structures. It has been generalized from single strings to collections of strings and some classes of labeled directed graphs, such as tries and de Bruijn graphs. The BWTs of repetitive datasets are often compressible using run-length compression, but recently Baier (CPM 2018) described how they could be even further compressed using an idea he called tunneling. In this paper we show that tunneled BWTs can still be used for indexing and extend tunneling to the BWTs of Wheeler graphs, a framework that includes all the generalizations mentioned above. | cs.DS | cs |
Tunneling on Wheeler Graphs
Jarno Alanko1, Travis Gagie2,3, Gonzalo Navarro2,4, Louisa Seelbach Benkner5
1Dept. of Computer Science, University of Helsinki, Finland, [email protected]
2CeBiB -- Center for Biotechnology and Bioengineering, Chile
3EIT, Diego Portales University, Chile, [email protected]
4Dept. of Computer Science, University of Chile, Chile, [email protected]
5Dept. of Electrical Engineering and Computer Science, University of Siegen, Germany,
[email protected]
Abstract
The Burrows-Wheeler Transform (BWT) is an important technique both in data
compression and in the design of compact indexing data structures.
It has been
generalized from single strings to collections of strings and some classes of labeled
directed graphs, such as tries and de Bruijn graphs. The BWTs of repetitive datasets
are often compressible using run-length compression, but recently Baier (CPM 2018)
described how they could be even further compressed using an idea he called tunnel-
ing. In this paper we show that tunneled BWTs can still be used for indexing and
extend tunneling to the BWTs of Wheeler graphs, a framework that includes all the
generalizations mentioned above.
Introduction
The Burrows-Wheeler transform (BWT) is a cornerstone of data compression and
succinct text indexing. It is a reversible permutation on a string that tends to com-
press well with run-length coding, while simultaneously facilitating pattern matching
against the original string. Recently, practical data structures have been designed on
top of the run-length compressed BWT to support optimal-time text indexing within
space bounded by the number of runs of the BWT [1].
However, run-length coding does not necessarily exploit all the available redun-
dancy in the BWT of a repetitive string. To this end, Baier recently introduced the
concept of tunneling to compress the BWT by exploiting additional redundancy not
yet captured by run-length compression [2]. While his representation can achieve
better compression than run-length coding, no support for text indexing is given.
Meanwhile, the concept of Wheeler graphs was introduced by Gagie et al. as an
alternative way to view Burrows-Wheeler type indices [3]. The framework can be used
to derive a number of existing index structures in the Burrows-Wheeler family, like
the classical FM-index [4] including its variants for multiple strings [5] and alignments
[6], the XBWT for trees [7], the GCSA for directed acyclic graphs [8], and the BOSS
data structure for de Bruijn graphs [9].
In this work, we show how Baier's concept of tunneling can be neatly explained
in terms of Wheeler graphs. Using the new point of view, we show how to support
FM-index style pattern searching on the tunneled BWT. We also describe a sam-
pling strategy to support pattern counting and locating and character extraction,
making our set of data structures a fully-functional FM-index. Supporting FM-index
operations was posed as an open problem by Baier [2].
We also use the generality of the Wheeler graph framework to generalize the
concept of tunneling to any Wheeler graph. This result can be used to compress any
Wheeler graph while still supporting basic pattern searching to decide if a pattern
exists as a path label in the graph. This has applications for all index structures that
can be explained in terms of Wheeler graphs.
Preliminaries
Let G = (V, E, λ) denote a directed edge-labeled multi-graph, in which V denotes
the set of nodes, E denotes the multiset of edges and λ : E → A denotes a function
labeling each edge of G with a character from a totally-ordered alphabet A. Let ≺
denote the ordering among A's elements. We follow the definition of Gagie et al. [3].
Definition 1 (Wheeler graph). The graph G = (V, E, λ) is called a Wheeler graph
if there is an ordering on the set of nodes such that nodes with in-degree 0 precede
those with positive in-degree and for any two edges (u1, v1), (u2, v2) labeled with
λ((u1, v1)) = a1 and λ((u2, v2)) = a2, we have
(i) a1 ≺ a2 ⇒ v1 < v2,
(ii) (a1 = a2) ∧ (u1 < u2) ⇒ v1 ≤ v2.
We note that the definition implies that all edges arriving at a node have the
same label. We call such an ordering a Wheeler ordering of nodes, and the rank of
a node within this ordering the Wheeler rank of the node. Given a pattern P ∈ A∗,
we call a node v ∈ V an occurrence of P if there is a path in G ending at v such that
the concatenation of edge labels on the path is equal to P . The Wheeler ranks of all
occurrences of P form a contiguous range of integers, which we call the Wheeler range
of P . Finding such a range is called path searching for P .
Given a Wheeler ordering of nodes, we define the corresponding Wheeler ordering
of edges such that for a pair of edges e1 = (u1, v1) and e2 = (u2, v2), we have e1 < e2
iff λ(e1) ≺ λ(e2) or (λ(e1) = λ(e2) and u1 < u2). When referring to edges, the term
Wheeler rank refers to the rank of the edge in the Wheeler ordering of edges, and
Wheeler range refers to a set of edges whose Wheeler ranks form a contiguous interval.
We use a slightly modified version of the representation of Wheeler graphs pro-
posed by Gagie et al. [3]. Suppose we have a Wheeler graph with n nodes and m
edges. Then we represent the graph with the following data structures:
• A string L[1..m] = L1 · · · Ln where Li is the concatenation of the labels of the
Li edges going out from the node with Wheeler rank i such that the labels
from a node are concatenated in their relative Wheeler order.
• An array C[1..A] such that C[i] is the number of edges in E with a label
smaller than the ith smallest symbol in A.
• A binary string I[1..n + m + 1] = X1, · · · , Xn · 1 where Xi = 1 · 0ki and ki is the
indegree of the node with Wheeler rank i.
• A binary string O[1..n + m + 1] = X1, · · · , Xn · 1 where Xi = 1 · 0li and li is the
outdegree of the node with Wheeler rank i.
Given these data structures, we can traverse the Wheeler graph with the following
two operations: First, given the Wheeler rank i of a node, find the Wheeler rank
of the k-th out-edge labeled c from the node using Eq. (1) below; second, given the
Wheeler rank j of an edge, find the Wheeler-rank r of its target node with Eq. (2):
C[c] + rankc(L, select1(O, i) − i) + k,
rank1(I, select0(I, j)),
(1)
(2)
where, given a string S, rankc(S, i) denotes the number of occurrences of character c
in S[1..i] and selectc(S, i) denotes the position of the i-th occurrence of c in S. For
the binary strings I and O, these operations can be carried out in constant time using
o(I + O) extra bits [10]. For string L, operation rank can be carried out in time
O(log logw A) on a w-bit RAM machine using o(L log A) extra bits [11, Thm. 8].
Given a Wheeler range [i, i′] of nodes, we can find the Wheeler rank of the
first and last edge labeled c leaving from a node in [i, i′] with a variant of Eq. (1):
C[c] + rankc(L, select1(O, i) − i) + 1 and C[c] + rankc(L, select1(O, i′ + 1) − (i′ + 1)),
respectively; then we apply Eq. (2) on both edge ranks to obtain the corresponding
node range (the result is a range by the path coherence property of Wheeler graphs
[3]). This operation enables path searches on Wheeler graphs.
Tunneling
We adapt Baier's concept of blocks on the BWT [2] to Wheeler graphs.
Definition 2 (Block). A block B of a Wheeler graph G = (V, E, λ) of size s and
width w is a sequence of w-tuples (v1,1, . . . , vw,1), . . . , (v1,s, . . . , vw,s) of pairwise dis-
tinct nodes of G such that
(i) For 1 ≤ i ≤ w − 1 and 1 ≤ j ≤ s, the node vi+1,j is the immediate successor of
vi,j with respect to the Wheeler ordering on V .
(ii) For 1 ≤ i ≤ w, let Vi = {vi,j 1 ≤ j ≤ s}, Ei = E ∩(Vi ×Vi), and λi = λ(cid:12)(cid:12)Ei
. The
subgraphs ti = (Vi, Ei, λi) are isomorphic subtrees of G, preserving topology
and labels. For 1 ≤ i ≤ w − 1, let fi : ti → ti+1 denote the corresponding
isomorphisms, thus vi+1,j = fi(vi,j) for all 1 ≤ j ≤ s.
(iii) For 1 ≤ i ≤ w, let vi,1 denote the root node of ti. In particular, vi,1 is the only
node of indegree 0 in ti. All edges leading to a node in {vi,1 1 ≤ i ≤ w} are
labeled with the same character. The indegrees of these nodes may differ.
(iv) For 1 ≤ i ≤ w and 2 ≤ j ≤ s, the nodes vi,j are of indegree 1 in G (and by (i)
and (ii), of indegree 1 in the corresponding subtree ti, that is, the only edge in
G leading to such a node vi,j belongs to the subtree ti).
b
a
24
b
c
b
63
c
74
41
aa
c a
31
9
4
a
a
c
b
b
b
a
c
25
64
c
b
a
b
b
32
a
a
b
75
42
10
a
b
b
b
a
a
a
b
c
b
c
a
b
b
a
a
b
b
c
c
a
b
b
a
b
5
a
c
c
c
c
a
a
c
c
Figure 1: Tunneling a block of size 7 and width 2 in a Wheeler graph.
(v) For every integer 1 ≤ j ≤ s and character c ∈ A, exactly one of the following
conditions holds:
(a) For every 1 ≤ i ≤ w, there is exactly one out-edge of vi,j labeled with
c, which is contained in Ei. There are no out-edges of vi,j labeled with c
leading to non-block nodes.
(b) For every 1 ≤ i ≤ w, there is no out-edge of vi,j labeled with c contained
in Ei. There may be out-edges of vi,j labeled with c leading to non-block
nodes. The number of such out-edges for each node may differ.
A block B = ((v1,1, . . . , vw,1), . . . , (v1,s, . . . , vw,s)), abbreviated B = (vi,j)1≤i≤w,1≤j≤s,
is called maximal in size if, for any choice of nodes v1, . . . , vw ∈ V , the sequences
of size s + 1 of w-tuples ((vi)1≤i≤w, (vi,1)1≤i≤w, . . . , (vi,s)1≤i≤w) and ((vi,1)1≤i≤w, . . . ,
(vi,s)1≤i≤w, (vi)1≤i≤w) do not form a block. The block is called maximal in width if, for
any choice of nodes v1, . . . , vs ∈ V , the sequences of (w+1)-tuples ((vj, v1,j, . . . , vw,j))1≤j≤s
and ((v1,j, . . . , vw,j, vj))1≤j≤s do not form a block. A block is called maximal if it is
maximal in both width and size.
Let G = (V, E, λ) denote a Wheeler graph containing a maximal block. We then
obtain a directed edge-labeled (multi-)graph Gt from G as follows:
(i) We merge the corresponding nodes and edges of the isomorphic subtrees ti, with
1 ≤ i ≤ w, in order to obtain a subtree t of G. In particular, for 1 ≤ j ≤ s,
we collapse the nodes of the w-tuple (v1,j, . . . , vw,j) to obtain a node xj of the
graph t. The labels of the merged edges coincide and stay the same.
(ii) All edges leading from a non-block node u to the root node vi,1 of the subtree
ti are redirected to lead to the node x1 of t, preserving their labels.
(iii) All edges leading from a node vi,j of subgraph ti to a non-block node u are
redirected to leave from the node xj of t, preserving their labels.
Formally, the graph Gt = (Vt, Et, λt) is defined as follows: The set of nodes of Gt
is defined as Vt = (V \ {vi,j 1 ≤ i ≤ w, 1 ≤ j ≤ s}) ∪ {xj 1 ≤ j ≤ s}. We define a
function ϕ : V → Vt mapping a node in G to its corresponding node in Gt by
ϕ(v) =(v
xj
if v /∈ {vi,j 1 ≤ i ≤ w, 1 ≤ j ≤ s}
if v = vi,j for some integers 1 ≤ i ≤ w, 1 ≤ j ≤ s.
The multiset of edges Et is defined as the difference of multisets {(ϕ(u), ϕ(v))
(u, v) ∈ E} \ {(ϕ(u), ϕ(v)) (u, v) ∈ Ei, 2 ≤ i ≤ w}. For every edge (x, y) ∈ Et, there
is a corresponding edge (u, v) ∈ E such that (x, y) = (ϕ(u), ϕ(v)). Thus, we define
λt((x, y)) = λ((u, v)). This is well-defined because only edges with the same label are
merged. We call tunneling the process of obtaining the graph Gt from G. See Fig. 1.
Lemma 3. Let G = (V, E, λ) denote a Wheeler graph containing a maximal block
B = (vi,j)1≤i≤w,1≤j≤s of width w and size s and let Gt = (Vt, Et, λt) denote the graph
obtained from G by tunneling. Then Gt is a Wheeler graph.
Proof. To show that Gt is a Wheeler graph, we must define a Wheeler ordering on Vt.
As only consecutive nodes of V are merged in order to obtain Gt, this induces a canon-
ical ordering on the nodes of Vt: Pick two nodes x 6= y of Vt. Let ϕ−1(x), ϕ−1(y) ⊆ V
denote the corresponding preimages under ϕ. By property (i) of Def. 2, the nodes of
ϕ−1(x) (respectively, ϕ−1(y)) are consecutive in Wheeler order. Thus, we either have
u < v for every u ∈ ϕ−1(x), v ∈ ϕ−1(y), or vice versa. We set x < y in the first case
and x > y in the second. This yields an ordering on the nodes of Vt such that for any
two nodes u 6= v of V , we have ϕ(u) < ϕ(v) ⇒ u < v and u < v ⇒ ϕ(u) ≤ ϕ(v).
First, take two nodes x 6= y of Gt, such that x has in-degree 0 and y is of positive
in-degree. As in the process of tunneling the in-degree of a node is not decreased,
every node u ∈ ϕ−1(x) is of in-degree 0 as well. Moreover, as y is of positive in-degree,
there is a node v ∈ ϕ−1(y), such that v is of positive in-degree. As G is a Wheeler
graph, we have u < v and thus x = ϕ(u) ≤ ϕ(v) = y. As by assumption, x 6= y, this
yields x < y. Therefore, nodes with in-degree 0 precede those with positive in-degree.
Now, take two edges (x, y) and (x′, y′) of G labeled with a and a′, respectively.
Without loss of generality, assume a (cid:22) a′. Choose u ∈ ϕ−1(x), v ∈ ϕ−1(y), u′ ∈
ϕ−1(x′) and v′ ∈ ϕ−1(y′), such that (u, v) and (u′, v′) are edges of G. By definition
of Gt, the label on the edge (u, v) (resp., (u′, v′)) is a (resp., a′). We then have
(x, y) = (ϕ(u), ϕ(v)) and (x′, y′) = (ϕ(u′), ϕ(v′)). Consider the two cases of Def. 1:
(i) Let a ≺ a′. As G is a Wheeler graph, we have v < v′ and thus ϕ(v) ≤ ϕ(v′).
If ϕ(v) < ϕ(v′) we are done, so assume ϕ(v) = ϕ(v′). We then have v = vi,j and
v′ = vk,j for some nodes vi,j 6= vk,j of the block. By properties (ii) - (iv) of Def. 2, the
labels of the incoming edges of vi,j and vk,j are the same, contradicting a ≺ a′.
(ii) Let a = a′ and without loss of generality assume u < u′. This yields ϕ(u) ≤ ϕ(u′).
As G is a Wheeler graph, we obtain v ≤ v′, and thus ϕ(v) ≤ ϕ(v′).
Two blocks B1 = (vi,j)1≤i≤w1,1≤j≤s1 and B2 = (ui,j)1≤i≤w2,1≤j≤s2 of a Wheeler graph
G = (V, E, λ) are called disjoint if their corresponding node sets {vi,j 1 ≤ i ≤ w1, 1 ≤
j ≤ s1} ⊂ V and {ui,j 1 ≤ i ≤ w2, 1 ≤ j ≤ s2} ⊂ V are disjoint. Since, by Lemma
3, a graph obtained from a Wheeler graph by tunneling is still a Wheeler graph, we
can tunnel iteratively with disjoint blocks.
Definition 4 (Tunneled Graph). Let G be a Wheeler graph containing k pairwise
disjoint maximal blocks B1, . . . , Bk. The tunneled graph Gt of G corresponding to
those blocks is defined as the Wheeler graph obtained from G by iteratively tunneling
all the blocks Bi, for 1 ≤ i ≤ k. Each maximal block Bi is also called a tunnel.
Note that tunnels more general than the tree form given in Def. 2 would break
the Wheeler graph rules of Def. 1 before or after tunneling, except that the nodes
of t1 and tw could be connected with outside nodes, all of Wheeler ranks smaller
and larger, respectively, than their corresponding tunnel edges. We could also handle
forests, but those can be seen as a set of disjoint tunnels.
Path searching on a tunneled Wheeler graph
Wheeler graphs can be searched for the existence of paths whose concatenated labels
yield a given string P [3], generalizing the classical backward search on strings [4].
We now show that those searches can also be performed on tunneled Wheeler graphs.
Given Gt, we can simulate the traversal of G as follows. A node v ∈ V is represented
by a pair hϕ(v), off(v)i, where ϕ(v) is the corresponding node in Gt as defined in the
previous section and off(v) is the tunnel offset of v.
If the node ϕ(v) does not
belong to a tunnel, then it must be that off(v) = 1 and the pair represents just the
node v. Otherwise, ϕ(v) corresponds to multiple nodes (vi,j)1≤i≤w of some tunnel
B = (vi,j)1≤i≤w,1≤j≤s in G and the tunnel offset represents which of the original nodes
we are currently at, in Wheeler rank order. That is, if v = vi,j, then off(v) = i.
The idea is that, when our traversal enters a tunnel B, we remember which orig-
inal subgraph ti we actually entered, and use that information to exit the tunnel
accordingly. We mark the nodes of Gt that are tunnel entrances in a bitvector, and
the other tunnel nodes in another bitvector. We then distinguish three cases.
If we are at a pair hi, 1i not in a tunnel, compute j and
Keeping out of tunnels.
r with Eqs. (1) and (2), respectively, and it turns out that r is not marked as a tunnel
entrance, then we stay out of any tunnel and our new pair is hr, 1i.
Entering a tunnel. Assume we are at a pair hi, 1i, where i = ϕ(v) for some non-
tunnel node v, compute j and r with Eqs. (1) and (2), respectively, and then it turns
out that r = ϕ(u) is marked as a tunnel entrance. Then we have entered a tunnel
and the new pair must be hr, oi, for some offset o we have to find out.
The w nodes (up)1≤p≤w ∈ V that were collapsed to form ϕ(u) are of indegree 0
within the subgraphs tp, but may receive a number of edges from non-tunnel nodes
(indeed, we are traversing one). Since all those edges are labeled by the same symbol
c, the Wheeler rank of all the sources of edges that lead to up must precede the
Wheeler ranks of all the sources of edges that lead to up′ for any 1 ≤ p < p′ ≤ w.
The problem is, knowing that we are entering by the edge with Wheeler rank j,
and that the edges that enter into r start at Wheeler rank select1(I, r) − r + 1, how
to determine the index o of the subgraph to we have entered. For this purpose, we
store a bitvector I ′[1..m], where m = Et, so that I ′[j] = 1 iff the jth edge of Et, in
Wheeler order, corresponds to the first edge leading to its target in G. Said another
way, I ′ marks, in the area of I corresponding to the edges that reach ϕ(u), which
were the first edges arriving at each copy up that was collapsed to form ϕ(u). We can
then compute o = rank1(I ′, j) − rank1(I ′, select1(I, r) − r).
Moving in a tunnel. Assume we are at a pair hi, oi inside a tunnel, for i = ϕ(u),
and want to traverse the kth edge labeled c leaving the pair. We then use the formulas
given after Eqs. (1) and (2) to compute the first and last edge labeled c leaving node
i. These form a Wheeler range [j1, j2]. We then apply Eq. (2) from j = j1 to find
the first target node r. If r is marked as an in-tunnel node, then we know that letter
c keeps us inside the tunnel, j1 = j2 (that is, there is exactly one out-edge by letter
c from each of the nodes (up)1≤p≤w that were collapsed to form ϕ(u), by part (a) of
item (v) of Definition 2), and thus our new node is simply hr, oi.
If, instead, r is not an in-tunnel node, then letter c takes us out of the tunnel,
and we must compute the appropriate target node. Each of the nodes us may have
zero or more outgoing edges labeled c. The Wheeler ranks of the edges leaving up
precede those of the edges leaving up′, for any 1 ≤ p < p′ ≤ w. Thus, we use a
bitvector O′[1..m] analogous to I ′, where O′[j] = 1 iff the jth edge of Et in Wheeler
order corresponds to the first edge leaving a node in G by some letter c.
In the
range O′[j1, j2], then, each 1 marks the first edge leaving from each up. The kth
edge labeled c leaving from uo is thus j = select1(O′, rank1(O′, j1) + o − 1) + k − 1.
If, for pattern searching, we want the last edge labeled c leaving from uo, this is
j = select1(O′, rank1(O′, j1) + o) − 1. We then compute the correct target node rank
r using Eq. (2).
Finally, there are two possibilities.
If the new node r is marked as a tunnel
entrance, then we have left our original tunnel to enter a new one. We then apply the
method described to enter a tunnel from the values j and r we have just computed.
Otherwise, r is not a tunnel node and we just return the pair hr′, 1i.
Therefore, we can simulate path searching in G by using just our representation
of Gt. Given a character c and the Wheeler range [i, i′] of a string S, we can find
the Wheeler range of the string Sc by following the first and last edge labeled with c
leaving from [i, i′], as described after Eqs. (1) and (2), and operate as described from
the corresponding ranks j and r. Thus, after P steps, we have the Wheeler range
of the nodes that can be reached by following the characters in P .
Theorem 5. We can represent a Wheeler graph G with labeled edges from the alphabet
[1..σ] in nt log σ + o(nt log σ) + O(nt) bits of space, where nt is the number of edges
in a tunneled version of G, such that we can decide if there exists a path labeled with
P in time O(P log logw σ) in a w-bit RAM machine.
Wheeler graphs of strings
We now focus on a particular type of Wheeler graphs, which corresponds to the
traditional notion of Burrows-Wheeler Transform (BWT) [12] and FM-index [4] (only
that our arrows go forward in the text, not backwards), and show that the full self-
index functionality on strings can still be supported after tunneling. This is close to
the original tunneling concept developed by Baier [2], to which we now add search
and traversal capabilities.
Definition 6 (Wheeler graph of a string). Let T be a string over an alphabet A. The
Wheeler graph of the string T is defined as G = (V, E, λ) with V := {v1, . . . , vT +1},
E = {(vi, vi+1) 1 ≤ i ≤ T + 1} and λ : E → A with λ((vi, vi+1)) = T [i], where T [i]
denotes the ith character in the string T .
In other words, the Wheeler graph of a string T is a path of length T + 1, where
the ith edge is labeled with the ith character of T . There is exactly one valid Wheeler-
ordering, which is given by the colexicographic order of prefixes of T , i.e., node vi
comes before node vj iff the reverse of T [1..i − 1] is lexicographically smaller than the
reverse of T [1..j − 1]. There is a close connection to the BWT: the Wheeler order is
given by the suffix array of the reverse of T and therefore the L-array corresponds to
the BWT of the reverse of T .
For this special case of Wheeler graphs, Def. 2 simplifies as follows: A block B in
G of width w and length s is a sequence of length s + 1 of w-tuples (v1,1, . . . , vw,1), . . . ,
(v1,s+1, . . . , vw,s+1) of pairwise distinct nodes of G satisfying
(i) For 1 ≤ i ≤ w − 1 and 1 ≤ j ≤ s + 1, the immediate successor of the node vi,j
with respect to the Wheeler ordering on V is vi+1,j.
(ii) For 1 ≤ i ≤ w and 1 ≤ j ≤ s, (vi,j, vi,j+1) is an edge of E.
(iii) For 1 ≤ j ≤ s, all the edges leading to the nodes in {vi,j 1 ≤ i ≤ w} have the
same label.
The process of tunneling in a Wheeler graph G of a string then consists of collapsing
the nodes of each w-tuple (v1,j, . . . , vw,j) into a single node xj and collapsing the edges
in {(vi,j, vi,j+1) 1 ≤ i ≤ w} into a single edge (xj, xj+1). Furthermore, all edges
leading to a node vi,1 for some integer 1 ≤ i ≤ w are redirected to lead to the node
x1 and all edges leaving from a node vi,s for some integer 1 ≤ i ≤ w are redirected to
leave from the node xs. The labels of the edges stay the same. Note that we ensure
that every path of the tunnel is followed by a non-tunnel node.
Suppose we have the Wheeler graph G = (V, E, λ) of a string T . Denote T = n.
Let Gt = (Vt, Et, λt) be a tunneled version of G with Vt = nt. We represent Gt with
the data structures L, C, I and O described in the preliminaries. We can do path
searches without the bitvectors I ′ and O′ used in the previous section, as in these
particular graphs they are all 1s. We now describe how to implement the operations
count, locate and extract, analogous to the operations in a regular FM-index, by
using sampling schemes that extend those of the standard FM-index solution.
First, for each tunnel of length at least log nt in Gt, we store a pointer and the
distance to the end of the tunnel for every (log nt)th consecutive node in the tunnel.
This information takes O(nt) bits of space and lets us skip to the end of the tunnel
in O(log nt) steps by walking forward until the end of the tunnel is found, or until we
hit a node with a pointer to the end. The stored distance value tells us how many
nodes we have skipped over.
Locating. We define a graph Gc, called the contracted graph, that is identical to Gt
except that tunnels have been contracted into single nodes. Let ψ : Gt → Gc be the
mapping such that nodes in a tunnel in Gt map to the corresponding contracted node
in Gc. Take the path Q = (q1, . . . , qn) of all nodes in G in the order of the path from
the source to the sink. Let Qc = ((ψ ◦ ϕ)(q1), . . . , (ψ ◦ ϕ)(qn)) be the corresponding
path in Gc. Let Q′
c be the same sequence as Qc except that every run of the same
node is contracted to length 1. Note that this path traverses all edges of Gc exactly
once, i.e., it is an Eulerian path. We store a sample for every (log n)th node in the
path on Q′
c, except that if a node represents a contracted tunnel, we sample the next
node (our definition of blocks guarantees that the next node is not in a tunnel). The
value associated with the sample is the text position corresponding to the node. This
takes space O(nc), where nc is the number of nodes in Gc.
We can then locate the text position of a node by walking to the next sample in
text order, using at most log n graph traversal operations in Gt. In this walk we may
have to skip tunnels, which is done in O(log nt) time using their stored pointers when
necessary. In the end, we subtract the travelled distance from the text position of
the sampled node to get the text position of the original node. We can view such
a search as a walk in Gc, where traversing a contracted-tunnel node takes O(log nt)
graph traversal steps and traversing a non-tunneled node takes just one. The worst
case time, dominated by the time to traverse tunnels, is O(log n log nt) steps.
Counting. Efficiently counting the number of occurrences of a pattern given its
Wheeler-range requires a sampling structure different from that used for locating. The
Wheeler-range could span many tunnels, whose widths are not immediately available.
Let us define w(v) as the width of the tunnel v belongs to, or 1 if v does not belong
to a tunnel. We can then afford to sample the cumulative sum of the values w(v) for
all the nodes v up Wheeler-rank k for every k multiple of log nt nodes, using O(nt)
bits of space. Within this space we can also mark which nodes belong to tunnels.
This allows us to compute the sum of values w(v) for any Wheeler range with
endpoints that are multiples of log nt, which leaves us to compute the width of only
O(log nt) nodes at the ends of the range. For these nodes, we add 1 if they are not
in a tunnel; otherwise we go to the end of the tunnel using the stored pointers and
compute the width of the tunnel by looking at the out-degree of the exit of the tunnel.
The total counting time is then O(log2 nt) graph traversal steps.
Extracting. To extract characters from T , we use a copy of the samples hWheeler
rank of graph node v, text position of node vi we store for locating, but sorted by
text position. Also, at the end of every tunnel of length at least log nt, we store
backpointers to the nodes storing pointers to the end of the tunnel.
Suppose we want to extract T [i..j].
If we know the node u of Gt representing
position i, we can simply walk forward from that node to find the j − i + 1 desired
characters by accessing L at each position. Therefore it is enough to show how to
find the node u. We binary search our sample pairs to find the Wheeler rank of the
closest sample before text position i. This sample is at most log n nodes away (in
Gc) from u, so we can reach u in O(log n) steps in Gc, or equivalently, O(log n log nt)
steps in Gt. Note, however, that our target node u might be in a tunnel. If the tunnel
is of length less than log nt, we walk towards it normally. If u is inside a tunnel of
length at least log nt, instead, we use its pointers to skip to the end of the tunnel,
and from there take the backpointer to the nearest position before u; we then walk
the (at most) log nt nodes until reaching u.
The time needed to reach u is again dominated by the time to skip over and within
tunnels, so the total time complexity is O(log n log nt) graph traversal steps.
We note that, in all cases, our graph traversal steps are of a particular form,
because all the edges leaving from the current node are labeled by the same symbol.
That is, the c in Eq. (1) is always L[select1(O, i)]. This particular form of rank
is called partial rank and it can be implemented in constant time using o(L log A)
further bits [13, Lem. 2]. The following theorem summarizes the results in this section.
Theorem 7. We can store a text T [1..n] over alphabet [1..σ] in nt log σ +o(nt log σ)+
O(nt) bits of space, such that in a w-bit RAM machine we can decide the existence of
any pattern P in time O(P log logw σ), and then report the text position of any occur-
rence in time O(log n log nt) or count the number of occurrences in time O(log2 nt).
We can also extract any k consecutive characters of T in time O(k + log n log nt),
where nt ≤ n is the number of nodes in a tunneled Wheeler graph of T .
Future work
Open problems are: How to find the optimal blocks that minimize space? Can we
still support path searching if blocks are overlapping? Can the O(log2 n) times of
counting, locating, and extracting be reduced to O(log n), as in the basic sampling
scheme on non-tunneled BWTs? How to extend those operations to more complex
graphs, like trees? And can we count paths instead of path endpoints?
Acknowledgements. Funded in part by EU's Horizon 2020 research and innova-
tion programme under Marie Sk lodowska-Curie grant agreement No 690941 (project
BIRDS). T.G. and G.N. partially funded with Basal Funds FB0001, Conicyt, Chile.
T.G. partly funded by Fondecyt grant 1171058. L.S.B. supported by the EU project
731143 - CID and the DFG project LO748/10-1 (QUANT-KOMP). Operation extract
was designed in StringMasters 2018. We thank Uwe Baier for helpful discussions.
References
[1] T. Gagie, G. Navarro, and N. Prezza, "Optimal-time text indexing in BWT-runs
bounded space," in Proc. 29th SODA, 2018, pp. 1459 -- 1477.
[2] U. Baier, "On undetected redundancy in the Burrows-Wheeler transform," in Proc.
29th CPM, 2018, pp. 3.1 -- 3.15.
[3] T. Gagie, G. Manzini, and J. Sir´en, "Wheeler graphs: A framework for BWT-based
data structures," Theoretical Computer Science, vol. 698, pp. 67 -- 78, 2017.
[4] P. Ferragina and G. Manzini, "Indexing compressed texts," Journal of the ACM, vol.
52, no. 4, pp. 552 -- 581, 2005.
[5] S. Mantaci, A. Restivo, G. Rosone, and M. Sciortino, "An extension of the Burrows --
Wheeler transform," Theoretical Computer Science, vol. 387, no. 3, pp. 298 -- 312, 2007.
[6] J. C. Na, H. Kim, H. Park, T. Lecroq, M. L´eonard, L. Mouchard, and K. Park, "FM-
index of alignment: A compressed index for similar strings," Theoretical Computer
Science, vol. 638, pp. 159 -- 170, 2016.
[7] P. Ferragina, F. Luccio, G. Manzini, and S. Muthukrishnan, "Structuring labeled trees
for optimal succinctness, and beyond," in Proc. 46th FOCS, 2005, pp. 184 -- 193.
[8] J. Sir´en, N. Valimaki, and V. Makinen, "Indexing graphs for path queries with appli-
cations in genome research," IEEE/ACM Transactions on Computational Biology and
Bioinformatics, vol. 11, no. 2, pp. 375 -- 388, 2014.
[9] A. Bowe, T. Onodera, K. Sadakane, and T. Shibuya, "Succinct de bruijn graphs," in
Proc. WABI, 2012, pp. 225 -- 235.
[10] D. R. Clark, Compact PAT Trees, Ph.D. thesis, University of Waterloo, Canada, 1996.
[11] D. Belazzougui and G. Navarro, "Optimal lower and upper bounds for representing
sequences," ACM Transactions on Algorithms, vol. 11, no. 4, pp. article 31, 2015.
[12] M. Burrows and D. Wheeler, "A block sorting lossless data compression algorithm,"
Tech. Rep. 124, Digital Equipment Corporation, 1994.
[13] D. Belazzougui and G. Navarro, "Alphabet-independent compressed text indexing,"
ACM Transactions on Algorithms, vol. 10, no. 4, pp. article 23, 2014.
|
1507.07789 | 1 | 1507 | 2015-07-28T14:46:48 | Duality and Nonlinear Graph Laplacians | [
"cs.DS"
] | We present an iterative algorithm for solving a class of \\nonlinear Laplacian system of equations in $\tilde{O}(k^2m \log(kn/\epsilon))$ iterations, where $k$ is a measure of nonlinearity, $n$ is the number of variables, $m$ is the number of nonzero entries in the graph Laplacian $L$, $\epsilon$ is the solution accuracy and $\tilde{O}()$ neglects (non-leading) logarithmic terms. This algorithm is a natural nonlinear extension of the one by of Kelner et. al., which solves a linear Laplacian system of equations in nearly linear time. Unlike the linear case, in the nonlinear case each iteration takes $\tilde{O}(n)$ time so the total running time is $\tilde{O}(k^2mn \log(kn/\epsilon))$. For sparse graphs where $m = O(n)$ and fixed $k$ this nonlinear algorithm is $\tilde{O}(n^2 \log(n/\epsilon))$ which is slightly faster than standard methods for solving linear equations, which require approximately $O(n^{2.38})$ time. Our analysis relies on the construction of a nonlinear "energy function" and a nonlinear extension of the duality analysis of Kelner et. al to the nonlinear case without any explicit references to spectral analysis or electrical flows. These new insights and results provide tools for more general extensions to spectral theory and nonlinear applications. | cs.DS | cs | Duality and Nonlinear Graph Laplacians
International Computer Science Institute
W. M. Keck Science Department of Claremont
Eric J. Friedman
[email protected]
Adam S. Landsberg
McKenna, Pitzer, and Scripps Colleges
[email protected]
5
1
0
2
l
u
J
8
2
]
S
D
.
s
c
[
1
v
9
8
7
7
0
.
7
0
5
1
:
v
i
X
r
a
ABSTRACT
We present an iterative algorithm for solving a class of
nonlinear Laplacian system of equations in O(k2m log(kn/ǫ))
iterations, where k is a measure of nonlinearity, n is the
number of variables, m is the number of nonzero entries in
the graph Laplacian L, ǫ is the solution accuracy and O()
neglects (non-leading) logarithmic terms. This algorithm is
a natural nonlinear extension of the one by of Kelner et.
al., which solves a linear Laplacian system of equations in
nearly linear time. Unlike the linear case, in the nonlinear
case each iteration takes O(n) time so the total running time
is O(k2mn log(kn/ǫ)). For sparse graphs where m = O(n)
and fixed k this nonlinear algorithm is O(n2 log(n/ǫ)) which
is slightly faster than standard methods for solving linear
equations, which require approximately O(n2.38) time. Our
analysis relies on the construction of a nonlinear "energy
function" and a nonlinear extension of the duality analysis
of Kelner et. al to the nonlinear case without any explicit
references to spectral analysis or electrical flows. These new
insights and results provide tools for more general extensions
to spectral theory and nonlinear applications.
General Terms
Graph Laplacian, Spectral analysis, Nonlinear equations
INTRODUCTION
1.
The goal of solving linear systems of equations in nearly lin-
ear time, spearheaded by Vaidya [10], has propelled a rev-
olution in spectral graph theory and several related fields.
In addition to many important results about graph sparsi-
fication (see [6, 11] for overviews and references) this the-
ory has led to two main types of methods of solving sets of
symmetrical diagonally dominant linear equations in nearly
linear time -- the first using graph sparsification (and ul-
tra sparsifiers) (e.g., [8]) combined with classical iterative
methods, and the second which relies on a underlying low
stretch spanning tree to provide an algorithm that relies on
cycle updates for "electrical flows" in the underlying graph
[3]. In addition to having important applications in matrix
computations, these results have led to a large number theo-
retical results, reducing the complexity of several important
graphical algorithms.
The key insights underlying these approaches are the re-
duction of a set of symmetrical diagonally dominant linear
equations to a system of equations Lx = b arising from a
graph Laplacian, L. Previous analyses (e.g., [8]) rely on
the linearity of the underlying system of equations and the
use of a carefully chosen spanning tree on the underlying
graph. Thus, one might expect that linearity is required
for the analyses; however, as we will show, linearity is not
necessary and there appears to be a more general struc-
ture underlying these approaches which allows a certain de-
gree of nonlinearity. More precisely, we show that one can
generalize the graph Laplacian in a nonlinear manner and
apply a nonlinear extension of the Algorithm in Kelner et
[3] to solve them in O(k2m log(kn/ǫ)) iterations and
al.
O(k2mn log(kn/ǫ)) time. Our analysis relies on the con-
struction of an "nonlinear energy function" and extension
of their duality analysis to nonlinear Laplacian systems of
equations. While providing new insights into the linear
problem, this analysis shows that one can generalize graph
Laplacians without losing much of the underlying structure
and suggests further nonlinear generalizations.
2. RELATED WORK
The idea of using a spanning tree to precondition a set of
linear equations comes from an unpublished presentation by
Vaidya [10]. That paper led to a large body of work over
many years that led to Spielman and Teng's seminal paper
[8] which showed the theoretical existence of nearly linear
time algorithms for systems of equations, using very sophis-
ticated machinery [4, 5, 7, 9].
Our algorithm is based on Kelner et. al's algorithm [3],
which dramatically reduces the amount of machinery re-
quired. Its origins are clearly implicitly based on the more
sophisticated machinery in the earlier papers, such as low-
stretch spanning trees (e.g., [1] ) and ultra-sparsifiers (e.g.,
[8]), but does not require them directly.
3. GENERALIZED MODEL AND RELAX-
ATION
We begin with an undirected weighted graph G =< V, E, w >
with node set V , edges E and weights wij > 0 where wij =
wji. The graph Laplacian of G is given by L where Lij =
a Laplacian system of equations is simply Lx = b. Such a
−wij for i 6= j and Lii = −Pj6=i wij . Given a vector b ∈ ℜn
system always has a solution if Pi bi = 0, which we denote
x∗ and is unique up to an additive constant. Throughout
this paper whenever we talk about solutions, we will ignore
the multiplicity of solutions due to the nullspace of L. Note
that for notational clarity we will always assume that G is
connected.
In the case that G is not connected one can
simply solve on each connected component separately.
Let N (i) be the neighbors of node i in the graph G. It is
well known that one can express the action of the Laplacian
as
(Lx)i = Xj∈N(i)
wij(xi − xj),
which only depends on xi −xj for (i, j) ∈ E. In the following
we will consider a generalization of this to non-linear graph
Laplacians where
(Lx)i = Xj∈N(i)
wij hij (xi − xj)
(∗)
where the nonlinear function hij (·) is anti-symmetric, satis-
fies hij (−v) = −hij (v), is continuous, the derivative h′
ij (v)
is left-continuous, vh′
ij (v) is strictly increasing and 1/k ≤
h′
ij (v) < k for a chosen k ≥ 1.
Note that the running time of our algorithm will depend on
k and the computational complexity of hij (·). In order to
simplify the analysis with respect to hij (·), we assume that
we have constant time oracles for computing hij (·), h−1
ij (·),
and h′
ij (·). The more natural version where these oracles
require O(log(1/ǫ)) computations, would only increase our
running times by O(log(1/ǫ)) factors, where ǫ is the accuracy
of the solution.
A simple example of this model arises when h′
ij (v) = 1/k for
v ≤ 1 and h′
ij (v) = 1 otherwise. This example arises when
the cost is small for small flows, in contrast to the linear case
where the "cost" of a flow between edges is linear in the flow.
More generally, we can choose h′
ij (v) to be constant except
at a discrete set of points where it increases, corresponding
to a piecewise linear hij (v) with discrete increases in slope,
bounded by slopes 1/k and k. Surprisingly, one can also use
an hij (v) that is not convex in the positive quadrant. For ex-
ample the function given by hij (v) = v + arctan(v) satisfies
the requirements and is not convex in the positive quad-
rant, leading to"economies of scale" where the incremental
cost can decrease as a flow increases.
Following Kelner et. al [3] we can introduce new variables,
gij , which for generalized Laplacians allow us to rewrite
Lx = b as:
Formally, we will only have a single variable for each edge,
written ge for e ∈ E but for notational convenience we will
write gij = −gji where one of these two variables as canon-
ical and the other is simply a notational convenience. For
example if P(i,j)∈E gij corresponds to a sum where each
edge appears once, while in Pi∈V Pj∈N(i) gij each ge ap-
pears twice.
Note that our variables differ from those in Kelner et. al
[3] and most of the literature in the linear case where in the
electrical model our gij variables correspond to the voltage
drop between nodes i and j while their hij corresponds to
the current flow between those nodes. Algebraically, gij =
hij /wij .
4. CYCLE UPDATES
A key idea in the algorithm by Kelner et al [3] is to update
the gij's using cycles to maintain b-feasibility. Let C be an
oriented set of edges in G that forms a cycle. Let α(C, t) be
defined such that for all (i, j) ∈ C,
wij hij (gij + αij (C, t)) = wijhij (gij ) + t,
and αij (C, t) = 0 for (i, j) 6∈ C and note that α(C, 0) = 0 by
construction. It can be easily checked that this procedure
maintains b-feasibility for any value of t, since the increase
from an incoming edge of the cycle through any node is
canceled by the decrease in the sum from the outgoing edge
that follows it. When hij (·) is the identity, αij (C, t) = t/wij
as in the linear analysis.
The key idea in the linear analysis is to iteratively update cy-
cles by updating cycles using α(C, t) in order to approach p-
feasibility while always maintaining b-feasibility. The anal-
ysis [3] relies on duality related to "electrical systems" with
an energy function. In the nonlinear case we can also define
an analogous energy function
Φ(g) = Xij
φij(gij )
where g is the vector of gij's and
φij (gij) = wij Z gij
0
sh′
ij (s) ds.
In the linear case where hij (v) is the identity, this reduces
to the standard energy function of [3], after a change of
variables.
The key insight in our analysis is that for any cycle C the
"cycle adjustment", α(C, t) that minimizes Φ(g) satisfies p-
feasibility around the cycle.
∀i ∈ V : Xj∈N(i)
wij hij (gij) = bi
(BCi)
Lemma 1. If for some cycle C
t∗ = argmintΦ(g + α(C, t))
∀(i, j) ∈ E :
gij = xi − xj
(P Ci)
then
which decomposes the problem in a natural way into b-
constraints (BCi) and p-constraints (P Ci), where x is a
'potential function' for g. We say that g is b-feasible if g
satisfies the b-constraints and p-feasible for p-constraints.
(gij + αij (C, t∗)) = 0.
X(i,j)∈C
Proof. To see this note that Φ(g) is convex by the defi-
nition of the h's and that the derivative is
Summing over C and integrating over t yields
dΦ(g + α(C, t))
dt
= X(i,j)∈C
dφij(gij + αij (C, t))
dt
X(i,j)∈C
αij (C, t) ≤ kt/WC
= X(i,j)∈C
φ′
ij(gij + αij (C, t))
dαij(C, t)
dt
.
Since GC + P(i,j)∈C αij (C, t∗) = 0 this implies that t∗ ≥
−WCGC /k. Combining this with (***) yields
Inserting the definition of φij into the right hand side of the
equation yields
Φ(g) − Φ(g + α(C, t∗)) ≥ Z 0
−WC GC /k
(GC + kt/WC )dt
= X(i,j)∈C
wij (gij+αij (C, t))h′
ij(gij+αij (C, t))
dαij(C, t)
dt
.
Implicitly differentiating the defining equation for α(C, t)
with respect to t gives
wij h′
ij (gij + αij (C, t))
dαij(C, t)
dt
= 1,
(∗∗)
which combined with the previous equation and the defini-
tion of t∗ gives the desired result after setting the derivative
to 0. ✷
This lemma implies our first theorem:
Theorem 1. If g is a b-feasible flow that minimizes Φ(g)
over all "cycle updates" in G, then g is also p-feasible.
Proof. Clearly if there are no cycle updates that de-
crease the energy then by the preceding Lemma the flows
around those cycles must sum to 0. ✷
We next derive a quantitative measure of the gain in energy
from a cycle update. This will depend on both the initial
flow around the cycle
GC = X(i,j)∈C
gij
and the harmonic sum of the w's around the cycle
WC = ( X(i,j)∈C
1/wij )−1.
Lemma 2. For a cycle C the decrease in energy from a
cycle update satisfies:
Φ(g) − Φ(g + α(C, t∗)) ≥ WCG2
C/(2k).
Proof. We assume that GC ≤ 0.
(If not then simply
traverse the cycle in the opposite direction.) As seen in the
proof in Lemma 1
dΦ(g + α(C, t))
dt
= X(i,j)∈C
(gij + αij (C, t)).
Integrating from t = t∗ to t = 0 yields
Φ(g)−Φ(g+α(C, t∗)) = Z 0
t∗
(GC + X(i,j)∈C
αij (C, t))dt
(∗∗∗).
Note that the integrand is 0 at t = t∗, GC at t = 0 and is
strictly decreasing over the range of integration. Using (**)
we see that
dαij(C, t)
dt
= (wij h′
ij (gij + αij (C, t)))−1 ≤ k/wij
and integrating completes the proof. ✷
To compute the cycle update we can use Newton's method
on the derivative Φ(g + α(C, t)), written in the form.
GC + X(i,j)∈C
αij (C, t).
Note that the sum is bounded by (t/WC)/k and (t/WC )k
so we can use a binary search on the interval [0, (t/WC)k].
Since the function is convex in t, for any t ∈ [t∗/4, 3t∗/4]
the value of the function will be reduced by at least a factor
of 2. This will require O(log(k2)) iterations so the total
time required is O(n log(k)) since each evaluation of the sum
requires O(n) time.
Thus we have proven an approximate version of the previous
lemma.
Lemma 3. One can compute an approximate cycle update
in O(n log(k)) such that for any cycle C the decrease in en-
ergy satisfies:
Φ(g) − Φ(g + α(C, t∗)) ≥ WCG2
C /(4k).
5. SPANNING TREES
Given a spanning tree, T ⊆ E one can analyze a simple set
of cycles that span the cycle space of all cycles [2]. Given
an edge (i, j) ∈ E \ T define the tree path P(i,j) to be the
minimal oriented set of edges from node j to node i and let
the tree cycle to be C(i,j) = P(i,j) S(i, j). Thus, instead of
minimizing Φ(g) over all cycles one only needs to minimize
it over each of the tree cycles.
As in [3], we will choose edges e ∈ E \ T at random with
probability pe and then minimize Φ(g) over the tree cycle
Ce. The main difference here is that in the linear case one
can minimize the energy over a cycle analytically while in
the nonlinear case one needs to rely on an iterative method,
which we described earlier.
For later use, we define the tree-values of x, denoted x(T, g)
to be those which satisfy gij = xi − xj on the tree. These
are easily computed, up to an additive constant, using using
a breadth first search from any leaf of the spanning tree of
G.
6. LAGRANGIAN DUALITY
In order to analyze the progress made by each iteration of
our algorithm we need an estimate of the distance to opti-
mality. To do this we compute the Lagrangian dual for our
minimization problem.
First recall that our primal problem is to compute
Proof. Since GAP ≤ T GAP we can focus on TGAP,
min
g
Φ(g)
s.t. Xj∈N(i)
wij hij (gij) = bi. ∀i ∈ V
∂TGAPij (g)
∂gij
= φ′
ij (gij) − wij(xi − xj)h′
ij (gij)
The dual function is given by
substituting for φ′
ij gives
Θ(x) = min
g
{Φ(g) − Xi∈V
xi[( Xj∈N(i)
wij hij (gij)) − bi]}
where x are the dual variables, and as we will see correspond
to the x in our original problem. (As a reminder, we only
consider one of gij, gji as a variable, say gij and implicitly
replace all occurrences of gji with −gij.) The first order
condition for this minimization when differentiating by gij
is:
0 = φ′
ij(gij) − (xi − xj)wijh′
ij (gij)
Using the definition of φij yields:
0 = wij h′
ij (gij)gij − (xi − xj)wij h′
ij (gij)
which reduces to gij = xi − xj . Plugging this into Θ(x)
yields
Θ(x) = X(i,j)∈E
φij(xi−xj)−Xi∈V
xi[( Xj∈N(i)
wijhij (xi−xj))−bi]
To simplify this we define gij = xi − xj and use the b-
feasibility of gij to write this as
Θ(x) = X(i,j)∈E
φij(gij) − X(i,j)∈E
wij gij(hij (gij) − hij (gij)).
Using this we can approximate the distance to optimality
since Φ(g) ≥ Θ(x) for any b-feasible g and Φ(g∗) = Θ(x∗)
where g∗, x∗ are the optimizers of their respective problems.
Thus, following Kelner et. al if we define
GAP (g) = Φ(g) − Φ(g∗)
then for any b-feasible g we have
GAP (g) ≤ Φ(g) − Θ(x)
∂TGAPij(g)
∂gij
= wij (gij − (xi − xj))h′
ij (gij).
Integrating yields
TGAPij (g) ≤ Z gij
(xi−xj )
wij(sij − (xi − xj))k dsij
which after noting that GCij = (gij − (xi − xj)) yields
TGAPij (g) ≤ wij G2
Cij k
which implies the result after summing. ✷
7. TREE CONDITION NUMBERS
In order to guarantee convergence in the linear case, Kelner
et. al.
[3] require that the underlying spanning tree have
nice properties. As we will show the same condition suffices
for the nonlinear case as well.
Define the tree condition number of spanning tree T to be
τ (T ) = X(i,j)∈E\T
wij /WCij
where Cij is the tree cycle over edge (i, j). Now, as discussed
in Kelner et. al, the tree condition number is closely related
to the stretch of a tree, a well studied topic. The stretch of
a tree T is defined to be st(T ) = P(i,j)∈E st(i, j) where
st(i, j) = wij X(r,s)∈P(i,j)
1/wrs
and it is easy to show that τ (T ) = st(T ) + m − 2n + 2.
Thus, we can use the following result due to Abraham et. al
[1] to construct a tree with a good condition number
We can use the tree values for x to provide this estimate
yielding
Theorem 2
(Abraham et. al [1]). In O(m log n log log n)
time we can compute a spanning tree T with total stretch
st(T ) = O(m log n log log n).
GAP (g) ≤ TGAP = X(i,j)∈E\T
TGAPij(g)
where we define TGAP by
TGAPij (g) = φij (gij) − φij(xi − xj)
+ wij (xi − xj)(hij(xi − xj) − hij (gij))
where the sums in the gap formula need only be over non-
tree edges since xi − xj = gij for all (i, j) ∈ T .
8. LINEARIZATION
For static analyses we can linearize the Laplacian at some g
by considering the adjusted weights, wg
ij = wij hij (gij)/gij,
creating a new Laplacian Lg. Then we can write
( Lgx)i = Xj∈N(i)
wg
ij (xi − xj).
Next we define the linearized energy to be the energy of the
linearized Laplacian:
Lemma 4. Given a b-feasible g,
ξ g(g) = X(i,j)∈E
wg
ij g2
ij/2
GAP (g) ≤ X(i,j)∈E\T
wijG2
Cij k/2
and it is easy to see that ξ g(g) = Φ(g). Then, using the
bounds on h′
ij (gij) we see that
where Cij is the tree cycle passing through edge (i, j).
wg
ij ∈ [(1/k)wij , kwij ],
Φ(g) ∈ [(1/k) ξ g(g), k ξ g(g)]
Proof. Lemma 6.1 in [3] shows that ξ(g0) ≤ k2ξ(g∗)st(T ).
and
Φ(g)/Φ(g∗) ≥ (1/k2) ξ g(g)/ ξ g(g∗).
This will allow us to use results from the linear case to an-
alyze some static properties of our algorithm.
9. ACCURACY
In the linear case, most spectral methods consider the ac-
curacy in solution [3]. Thus we will say that a solution is
ǫ-accurate if
x − x∗Lq∗
x∗q∗
L
≤ ǫ,
where Lq∗
is the linearization of L at q∗, as used in the pre-
vious section. By combining our bounds on the linearization
of the energy function (***) with Lemma 6.2 in [3] for the
linearization we can easily prove the following:
Lemma 5. If
Φ(g)/Φ(g∗) ≤ 1 + ǫ2/(τ (T )k4)
then
x − x∗q∗
L
x∗q∗
L
≤ ǫ
where x is defined by its tree values on T .
Proof. It is easy to see from Lemma 6.2 in [3] that if
Φ(g)/Φ(g∗) ≤ (1 + ǫ2/k4) then x − x∗ Lg ≤ ǫ/(k2τ (T )).
To complete the proof we need the following lemma relating
v Lg and v Lg∗ . ✷
Lemma 6. v Lg ≤ k2v Lg∗ .
Proof. This follows from the definition:
wg(vi − vj )2
v L(g) = X(i,j)∈E
and noting that wg
ij ∈ [(1/k)wij , kwij]. ✷
INITIALIZATION
10.
To initiate our algorithm we start with a b-feasible flow g0
with support only on the spanning tree. This can be com-
puted recursively with O(n) operations by noting that for
any leaf node i of the tree the value of gij for j = N (i) is
given by wij hij (gij) = bi.
However, it will be useful to note that this can be computed
by first computing the solution for the linearization of L
denoted g0, where we set all hij (·) to be identity functions,
using the above recursive algorithm. Then convert g0 to g0
using the relationship g0
ij = h−1
ij (g0
ij ).
The result follows from the fact that
Φ(g) ∈ [(1/k) ξ g(g), k ξ g(g)].
✷
11. ALGORITHM AND MAIN RESULT
We now present the algorithm which is a slight modification
of that in Kelner et. al [3].
Algorithm: Simple Nonlinear Solver
1. INPUT: Input G = (V, E, w), ǫ ∈ ℜ+, k ≥ 1, hij (·) ∀(i, j) ∈
E.
2. OUTPUT: g ∈ ℜE, x ∈ ℜV .
3. Construct T : a low stretch spanning tree of G.
4. Construct g0: a b-feasible solution with support only
on the spanning tree.
5. Set probabilities pij = wij/(WCij τ (T )) for all (i, j) ∈
E \ T .
6. Set S = ⌈2k2τ (T ) log(k6st(T )τ (T )/ǫ2)⌉.
7. For ℓ = 1 to S:
(a) Pick a random (i, j) ∈ E \ T with probability pij.
(b) Apply a cycle update to gℓ−1 using tree cycle
Cij to get gℓ. Compute this to relative accuracy
1/(2k2) using binary search.
8. RETURN g and its tree induced x.
In order to show the correctness of this algorithm, we ex-
tend the analysis in [3]. Since the GAP (gℓ) conditional on
GAP (gℓ−1) is a random variable we focus on expected val-
ues.
Lemma 8. For each iteration in the loop of "Algorithm:
Simple Nonlinear Solver"
E[GAP (gℓ)] − E[GAP (gℓ+1)GAP (gℓ)]
E[GAP (gℓ)]
≥
1
2k2τ (T )
.
Proof. First note that the change in GAP is the same
Cij /(2k)
as the change in Φ(g) which is greater than WCij G2
for an update of Cij , thus the expected decrease is
E[GAP (gℓ)]−E[GAP (gℓ+1)GAP (gℓ)] ≥ X(i,j)∈E\T
pij WCij G2
Cij /(2k).
Substituting in pij = wij /(τ (T )WCij ) yields
E[GAP (gℓ)]−E[GAP (gℓ+1)GAP (gℓ)] ≥ X(i,j)∈E\T
(wij /τ (T ))G2
Cij /(2k)
using Lemma 4 for the RHS completes the proof. ✷
The following bound on the initial energy follows directly
from Lemma 6.1 in [3].
Using this we can prove our main theorem.
Lemma 7. Φ(g0) ≤ k2Φ(g∗)st(T )
Theorem 3. Given a nonlinear graph Laplacian L, "Al-
gorithm: Simple Nonlinear Solver", computes an expected
ǫ-accurate solution of Lx = b in O(mnk2 log(kn/ǫ)) time.
13. ACKNOWLEDGMENTS
This material is based upon work supported by the National
Science Foundation under Grants No. 1216073 and 1161813.
14. REFERENCES
[1] I. Abraham and O. Neiman. Using
petal-decompositions to build a low stretch spanning
tree. In Proceedings of the forty-fourth annual ACM
symposium on Theory of computing, pages 395 -- 406.
ACM, 2012.
[2] B. Bollob´as. Modern graph theory, volume 184.
Springer Science & Business Media, 1998.
[3] J. A. Kelner, L. Orecchia, A. Sidford, and Z. A. Zhu.
A simple, combinatorial algorithm for solving sdd
systems in nearly-linear time. In Proceedings of the
forty-fifth annual ACM symposium on Theory of
computing, pages 911 -- 920. ACM, 2013.
[4] I. Koutis, G. L. Miller, and R. Peng. A fast solver for
a class of linear systems. Communications of the
ACM, 55(10):99 -- 107, 2012.
[5] J. H. Reif. Efficient approximate solution of sparse
linear systems. Computers & Mathematics with
Applications, 36(9):37 -- 58, 1998.
[6] D. A. Spielman. Algorithms, graph theory, and linear
equations in laplacian matrices. In Proceedings of the
International Congress of Mathematicians, volume 4,
pages 2698 -- 2722, 2010.
[7] D. A. Spielman and N. Srivastava. Graph
sparsification by effective resistances. SIAM Journal
on Computing, 40(6):1913 -- 1926, 2011.
[8] D. A. Spielman and S.-H. Teng. Nearly-linear time
algorithms for graph partitioning, graph sparsification,
and solving linear systems. In Proceedings of the
thirty-sixth annual ACM symposium on Theory of
computing, pages 81 -- 90. ACM, 2004.
[9] D. A. Spielman and J. Woo. A note on
preconditioning by low-stretch spanning trees. arXiv
preprint arXiv:0903.2816, 2009.
[10] P. M. Vaidya. Solving linear equations with symmetric
diagonally dominant matrices by constructing good
preconditioners. A talk based on this manuscript,
1991.
[11] N. K. Vishnoi. Laplacian solvers and their algorithmic
applications. Theoretical Computer Science,
8(1-2):1 -- 141, 2012.
Proof. Rewriting the expression in Lemma 5 shows that
in order to attain ǫ-accuracy we need
Φ(gS) − Φ(g∗) ≤ ǫ2/(τ (T )k4)Φ(g∗).
Iterating Lemma 8, implies that
Φ(gS) − Φ(g∗) ≤ (1 −
1
2k2τ (T )
)S(Φ(g0) − Φ(g∗)).
Combining this with Lemma 7 on the initial energy yields
Φ(gS) − Φ(g∗) ≤ (1 −
1
2k2τ (T )
)S(k2st(T ) − 1)Φ(g∗).
Plugging in S = ⌈2k2τ (T ) log(k6st(T )τ (T )/ǫ2)⌉ satisfies the
requirement, proving the accuracy of the Algorithm. By
Theorem 2
st(T ) = O(m log n log log n) = O(m)
the required number of iterations is given by
r = O(k2m log(kn/ǫ)).
Since each iteration takes O(n) time by Lemma 3 we get the
stated result. ✷
Note that instead of an expected ǫ-accurate solution, one
can compute an actual ǫ-accurate solution with high proba-
bility with the same time bound using standard probabilistic
techniques.
12. CONCLUDING REMARKS
This paper has shown that one can extend some of the tech-
niques from the linear theory of systems of equations arising
from linear graph Laplacians to ones with nonlinear graph
Laplacians. We view our analysis as suggestive that even
more general extensions, perhaps even asymmetric ones re-
lated to directed graphs, may be possible, extending the
range of spectral analysis. Perhaps as in Kelner et. al [3]
our algorithm has a representation as an alternating projec-
tions algorithm, showing that nonlinear extensions are also
possible in that setting.
There are many open directions for improving our algorithm.
First, note that we have not optimized the k dependence.
Next, note that one could use the ideas of tree scaling and
a randomized stopping time as in [3] to reduce the running
time. In addition, one might be able to use algorithms for
the linear problem to speed up the nonlinear algorithm, ei-
ther by providing a warm start or by speeding up conver-
gence near optimality when the linear problem might pro-
vide a good approximation for the nonlinear one.
While this paper raises many open problems relating to non-
linear spectral theory, the main problem with the current
algorithm is the dependence on n due to the complexity of
computing a nonlinear cycle update which unlike the linear
version does not have a simple analytic solution which can be
computed in logarithmic time using a special data structure
for keeping track of the flows on the tree. Nonetheless, it
might be possible to reduce this complexity either through
a specially designed data structure for the spanning tree,
or perhaps through the use of a spanning tree with small
diameter. Also, the direct application of our algorithm to
nonlinear graph problems might be fruitful.
|
1601.04233 | 1 | 1601 | 2016-01-17T00:02:37 | Sublinear-Time Algorithms for Counting Star Subgraphs with Applications to Join Selectivity Estimation | [
"cs.DS"
] | We study the problem of estimating the value of sums of the form $S_p \triangleq \sum \binom{x_i}{p}$ when one has the ability to sample $x_i \geq 0$ with probability proportional to its magnitude. When $p=2$, this problem is equivalent to estimating the selectivity of a self-join query in database systems when one can sample rows randomly. We also study the special case when $\{x_i\}$ is the degree sequence of a graph, which corresponds to counting the number of $p$-stars in a graph when one has the ability to sample edges randomly.
Our algorithm for a $(1 \pm \varepsilon)$-multiplicative approximation of $S_p$ has query and time complexities $\O(\frac{m \log \log n}{\epsilon^2 S_p^{1/p}})$. Here, $m=\sum x_i/2$ is the number of edges in the graph, or equivalently, half the number of records in the database table. Similarly, $n$ is the number of vertices in the graph and the number of unique values in the database table. We also provide tight lower bounds (up to polylogarithmic factors) in almost all cases, even when $\{x_i\}$ is a degree sequence and one is allowed to use the structure of the graph to try to get a better estimate. We are not aware of any prior lower bounds on the problem of join selectivity estimation.
For the graph problem, prior work which assumed the ability to sample only \emph{vertices} uniformly gave algorithms with matching lower bounds [Gonen, Ron, and Shavitt. \textit{SIAM J. Comput.}, 25 (2011), pp. 1365-1411]. With the ability to sample edges randomly, we show that one can achieve faster algorithms for approximating the number of star subgraphs, bypassing the lower bounds in this prior work. For example, in the regime where $S_p\leq n$, and $p=2$, our upper bound is $\tilde{O}(n/S_p^{1/2})$, in contrast to their $\Omega(n/S_p^{1/3})$ lower bound when no random edge queries are available. | cs.DS | cs |
Sublinear-Time Algorithms for Counting Star Subgraphs
with Applications to Join Selectivity Estimation
Maryam Aliakbarpour ∗
John Peebles §
Amartya Shankha Biswas †
Themistoklis Gouleakis ‡
Ronitt Rubinfeld ¶
Anak Yodpinyanee (cid:107)
We study the problem of estimating the value of sums of the form Sp (cid:44) (cid:80)(cid:0)xi
Abstract
(cid:1) when one has the
ability to sample xi ≥ 0 with probability proportional to its magnitude. When p = 2, this problem is
equivalent to estimating the selectivity of a self-join query in database systems when one can sample rows
randomly. We also study the special case when {xi} is the degree sequence of a graph, which corresponds
to counting the number of p-stars in a graph when one has the ability to sample edges randomly.
Our algorithm for a (1 ± ε)-multiplicative approximation of Sp has query and time complexities
). Here, m =(cid:80) xi/2 is the number of edges in the graph, or equivalently, half the number
O( m log log n
p
2S
1/p
p
of records in the database table. Similarly, n is the number of vertices in the graph and the number of
unique values in the database table. We also provide tight lower bounds (up to polylogarithmic factors)
in almost all cases, even when {xi} is a degree sequence and one is allowed to use the structure of the
graph to try to get a better estimate. We are not aware of any prior lower bounds on the problem of
join selectivity estimation.
For the graph problem, prior work which assumed the ability to sample only vertices uniformly gave
algorithms with matching lower bounds [Gonen, Ron, and Shavitt. SIAM J. Comput., 25 (2011), pp.
1365-1411]. With the ability to sample edges randomly, we show that one can achieve faster algorithms
for approximating the number of star subgraphs, bypassing the lower bounds in this prior work. For
example, in the regime where Sp ≤ n, and p = 2, our upper bound is O(n/S1/2
), in contrast to their
Ω(n/S1/3
) lower bound when no random edge queries are available.
p
p
In addition, we consider the problem of counting the number of directed paths of length two when
the graph is directed. This problem is equivalent to estimating the selectivity of a join query between
two distinct tables. We prove that the general version of this problem cannot be solved in sublinear
time. However, when the ratio between in-degree and out-degree is bounded—or equivalently, when the
ratio between the number of occurrences of values in the two columns being joined is bounded—we give
a sublinear time algorithm via a reduction to the undirected case.
1065125, and CCF-1420692
1065125, and CCF-1420692.
∗CSAIL, MIT, Cambridge MA 02139. E-mail: [email protected]. Research supported by NSF grants CCF-1217423, CCF-
†MIT, Cambridge MA 02139. E-mail: [email protected].
‡CSAIL, MIT, Cambridge MA 02139. E-mail: [email protected]. Research supported by NSF grants CCF-1217423, CCF-
§CSAIL, MIT, Cambridge MA 02139. E-mail: [email protected]. Research supported by NSF grants CCF-1217423,
¶CSAIL, MIT, Cambridge MA 02139 and the Blavatnik School of Computer Science, Tel Aviv University. E-mail:
[email protected]. Research supported by NSF grants CCF-1217423, CCF-1065125, CCF-1420692, and ISF grant 1536/14.
(cid:107)CSAIL, MIT, Cambridge MA 02139. E-mail: [email protected]. Research supported by NSF grants CCF-1217423,
CCF-1065125, CCF-1420692, and CCF-1122374.
CCF-1065125, and CCF-1420692, and the DPST scholarship, Royal Thai Government.
1
Introduction
We study the problem of approximately estimating Sp (cid:44)(cid:80)n
i=1
(cid:0)xi
p
(cid:1) when one has the ability to sample xi ≥ 0
with probability proportional to its magnitude. To solve this problem we design sublinear-time algorithms,
which compute such an approximation while only looking at an extremely tiny fraction of the input, rather
than having to scan the entire data set in order to determine this value.
We consider two primary motivations for this problem. The first is that in undirected graphs, if xi is the
degree of vertex i then Sp counts the number of p-stars in the graph. Thus, estimating Sp when one has the
ability to sample xi with probability proportional to its magnitude corresponds to estimating the number of
p-stars when one has the ability to sample vertices with probability proportional to their degrees (which is
equivalent to having the ability to sample edges uniformly). This problem is an instance of the more general
subgraph counting problem in which one wishes to estimate the number of occurrences of a subgraph H in
a graph G. The subgraph counting problem has applications in many different fields, including the study
of biological, internet and database systems. For example, detecting and counting subgraphs in protein
interaction networks is used to study molecular pathways and cellular processes across species [SIKS06].
The second application of interest is that the problem of estimating S2 corresponds to estimating the
selectivity of join and self-join operations in databases when one has the ability to sample rows of the tables
uniformly. For example, note that if we set xi as the number of occurrences of value i in the column
being joined, then S2 is precisely the number of records in the join of the table with itself on that column.
When performing a query in a database, a program called a query optimizer is used to determine the most
efficient way of performing the database query. In order to make this determination, it is useful for the query
optimizer to know basic statistics about the database and about the query being performed. For example,
queries that return a very larger number of records are usually serviced most efficiently by doing simple
linear scans over the data whereas queries that return a smaller number of records may be better serviced by
using an index [HILM09]. As such, being able to estimate selectivity (number of records returned compared
to the maximum possible number) of a query can be useful information for a query optimizer to have. In the
more general case of estimating the selectivity of a join between two different tables (which can be modeled
with a directed graph), the query optimizer can use this information to decide on the most efficient order to
execute a sequence of joins which is a common task.
In the “typical” regime in which we wish to estimate S2 given that n ≤ S2 ≤ n2, our algorithm has a
running time of O(
n) which is very small compared to than the total amount of data. Furthermore, in
the case of selectivity estimation, this number can be much less than the number of distinct values in the
column being joined on, which results in an even smaller number of queries than would be necessary if one
were using an index to compute the selectivity.
√
We believe that our query-based framework can be realized in many systems. One possible way to
implement random edge queries is as follows: because edges normally take most of the space for storing
graphs, an access to a random memory location where the adjacency list is stored, would readily give a
random edge. Random edge queries allow us to implement a source of weighted vertex samples, where
a vertex is output with probability proportional to its weight (magnitude). Weighted sampling is used in
[MPX07, BBS09] to find sublinear algorithms for approximating the sum of n numbers (allowing only uniform
sampling, results in a linear lower bound). We later use this as a subroutine in our algorithm.
Throughout the rest of the paper, we will mostly use graph terminology when discussing this problem.
However, we emphasize that all our results are fully general and apply to the problem of estimating Sp even
when one does not assume that the input is a graph.
1.1 Our Contribution
Prior theoretical work on this problem only considered the version of this problem on graphs and assumed the
ability to sample vertices uniformly rather than edges. Specifically, prior studies of sublinear-time algorithms
for graph problems usually consider the model where the algorithm is allowed to query the adjacency list
1
representation of the graph: it may make neighbor queries (by asking “what is the ith neighbor of a vertex
v”) and degree queries (by asking “what is the degree of vertex v”).
We propose a stronger model of sublinear-time algorithms for graph problems which allows random edge
queries. Next, for undirected graphs, we construct an algorithm which uses only degree queries and random
edge queries. This algorithm and its analysis is discussed in Section 3. For the problem of computing an
approximation Sp satisfying (1 − )Sp ≤ Sp ≤ (1 + )Sp, our algorithm has query and time complexities
O(m log log n/2S1/p
). Although our algorithm is described in terms of graphs, it also applies to the more
(cid:1) without any assumptions about graph structure.
general case when one wants to estimate Sp = (cid:80)
(cid:0)xi
p
i
p
Thus, it also applies to the problem of self-join selectivity estimation.
We then establish some relationships between m and other parameters so that we may compare the
performance of this algorithm to a related work by Gonen et al. more directly ([GRS11]). We also provide
lower bounds for our proposed model in Section 4, which are mostly tight up to polylogarithmic factors. This
comparison is given in Table 1. We emphasize that even though these lower bounds are stated for graphs,
they also apply to the problem of self-join selectivity estimation.
To understand this table, first note that these algorithms require more samples when Sp is small (i.e., stars
are rare). As Sp increases, the complexity of each algorithm decreases until—at some point—the number of
required samples drops to O(n1−1/p). Our algorithm is able to obtain this better complexity of O(n1−1/p)
for a larger range of values of Sp than that of the algorithm given in [GRS11]. Specifically, our algorithm
is more efficient for Sp ≤ n1+1/p, and has the same asymptotic bound for Sp up to np. Once Sp > np,
it is unknown whether the degree and random edge queries alone can provide the same query complexity.
Nonetheless, if we have access to all three types of queries, we may combine the two algorithms to obtain
the best of both cases as illustrated in the last column.
permitted types of queries
range of Sp
Sp ≤ n
n < Sp ≤ n1+1/p
n1+1/p < Sp ≤ np
np < Sp
neighbor, degree
n
S1/(p+1)
([GRS11])
(cid:19)
(cid:18)
(cid:101)Θ
(cid:101)Θ(cid:0)n1−1/p(cid:1)
(cid:18)
(cid:19)
(cid:101)Θ
np−1/p
S1−1/p
p
p
degree, random edge
n
(this paper)
(cid:18)
(cid:19)
(cid:101)Θ
(cid:101)Θ(cid:0)n1−1/p(cid:1)
(cid:19)
S1/p
p
(cid:18)
Ω
np−1/p
S1−1/p
p
,(cid:101)O(cid:0)n1−1/p(cid:1)
all types of queries
p
n
S1/p
(this paper)
(cid:18)
(cid:19)
(cid:101)Θ
(cid:101)Θ(cid:0)n1−1/p(cid:1)
(cid:18)
(cid:19)
(cid:101)Θ
np−1/p
S1−1/p
p
Table 1: Summary of the query and time complexities for counting p-stars on undirected graphs, given a
different set of allowed queries. is assumed to be constant. Adjacent cells in the same column with the
same contents have been merged.
We also consider a variant of the counting stars problem on directed graphs in Appendix D. If one only
needs to count “stars” where all edges are either pointing into or away from the center, this is essentially
still the undirected case. We then consider counting directed paths of length two, and discover that allowing
random edge queries does not provide an efficient algorithm in this case. In particular, we show that any
constant factor multiplicative approximation of Sp requires Ω(n) queries even when all three types of queries
are allowed. However, when the ratio between the in-degree and the out-degree on every vertex is bounded,
we solve this special case in sublinear time via a reduction to the undirected case where degree queries and
random edge queries are allowed.
directed graph, we aim at estimating the quantity (cid:80)
database context, we wish to compute the quantity (cid:80)n
This variant of the counting stars problem can also be used for approximating join selectivity. For a
(v) · deg+(v). On the other hand in the
v∈V (G) deg
i=1 xi · yi, where xi and yi denote the number of
occurrences of a label i in the column we join on, from the first and the second table, respectively. Thus,
−
2
applying simple changes in variables, the algorithms from Appendix D can be applied to the problem of
estimating join selectivity as well.
1.2 Our Approaches
In order to approximate the number of stars in the undirected case, we convert the random edge queries into
weighted vertex sampling, where the probability of sampling a particular vertex is proportional to its degree.
We then construct an unbiased estimator that approximates the number of stars using the degree of the
sampled vertex as a parameter. The analysis of this part is roughly based on the variance bounding method
used in [AMS96], which aims to approximate the frequency moment in a streaming model. The number of
samples required by this algorithm depends on Sp, which is not known in advance. Thus we create a guessed
value of Sp and iteratively update this parameter until it becomes accurate.
To demonstrate lower bounds in the undirected case, we construct new instances to prove tight bounds
for the case in which our model is more powerful than the traditional model. In other cases, we provide
a new proof to show that the ability to sample uniformly random edges does not necessarily allow better
performance in counting stars. Our proof is based on applying Yao’s principle and providing an explicit
construction of the hard instances, which unifies multiple cases together and greatly simplifies the approach
of [GRS11].1
For the directed case, we prove the lower bound using a standard construction and Yao’s principle. As
for the upper bound when the in-degree and out-degree ratios are bounded, we use rejection sampling to
adjust the sampling probabilities so that we may apply the unbiased estimator method from the undirected
case.
1.3 Related Work
Motivated by applications in a variety of areas, the subgraph detection and counting problem and its vari-
ations have been studied in many different works, often under different terminology such as network motif
counting or pathway querying (e.g., [MSOI+02, PCJ04, Wer06, SIKS06, SSRS06, GK07, HBPS07, HA08,
ADH+08]). As this problem is NP-hard in general, many approaches have been developed to efficiently count
subgraphs more efficiently for certain families of subgraphs or input graphs (e.g., [DLR95, AYZ97, FG04,
KIMA04, ADH+08, AG09, VW09, Wil09, GS09, KMPT10, AGM12, AFS12, FLR+12]). As for applications
to database systems, the problem of approximating the size of the resulting table of a join query or a self-join
query in various contexts has been studied in [SS94, HNSS96, AGMS99]. Selectivity and query optimization
have been considered, e.g., in [PI97, LKC99, GTK01, MHK+07, HILM09].
Other works that study sublinear-time algorithms for counting stars are [GRS11] that aims to approximate
the number of stars, and [Fei06, GR08] that aim to approximate the number of edges (or equivalently,
the average degree). Note that [GRS11] also shows impossibility results for approximating triangles and
paths of length three in sublinear time when given uniform edge sampling, limiting us from studying more
sophisticated subgraphs. Recent work by Eden, Levi and Ron ([ELR15]) and Seshadhri ([Ses15]) provide
sublinear time algorithms to approximate the number of triangles in a graph. However, their model uses
adjacency matrix queries and neighbor queries. The problem of counting subgraphs has also been studied
in the streaming model (e.g., [BYKS02, BFL+06, BBCG08, MMPS11, KMSS12]). There is also a body of
1One useful technique for giving lower bounds on sublinear time algorithms, pioneered by [BBM12], is to make use of
a connection between lower bounds in communication complexity and lower bounds on sublinear time algorithms. More
specifically, by giving a reduction from a communication complexity problem to the problem we want to solve, a lower bound on
the communication complexity problem yields a lower bound on our problem. In the past, this approach has led to simpler and
cleaner sublinear time lower bounds for many problems. Attempts at such an approach for reducing the set-disjointness problem
in communication complexity to our estimation problem on graphs run into the following difficulties: First, as explained in
[Gol13], the straightforward reduction adds a logarithmic overhead, thereby weakening the lower bound by the same factor.
Second, the reduction seems to work only in the case of sparse graphs. Although it is not clear if these difficulties are
insurmountable, it seems that it will not give a simpler argument than the approach that we present in this work.
3
work on sublinear-time algorithms for approximating various graph parameters (e.g., [PR07, NO08, YYI09,
HKNO09, ORRR12]).
Abstracting away the graphical context of counting stars, we may view our problem as finding a parameter
of a distribution: edge or vertex sampling can be treated as sampling according to some distribution. In
vertex sampling, we have a uniform distribution and in edge sampling, the probabilities are proportional
to the degree. The number of stars can be written as a function of the degrees. Aside from our work,
there are a number of other studies that make use of combined query types for estimating a parameter of a
distribution. Weighted and uniform sampling are considered in [MPX07, BBS09]. Their algorithms may be
adapted to approximate the number of edges in the context of approximating graph parameters when given
weighted vertex sampling, which we will later use in this paper. A closely related problem in the context of
distributions, is the task of approximating frequency moments, mainly studied in the streaming model (e.g.,
[AMS96, CK04, IW05, BGKS06]). On the other hand, the combination of weighted sampling and probability
distribution queries is also considered (e.g., [CR14]).
2 Preliminaries
In this paper, we construct algorithms to approximate the number of stars in a graph under different types
of query access to the input graph. As we focus on the case of simple undirected graphs, we explain this
model here and defer the description for the directed case to Appendix D.
2.1 Graph Specification
Let G = (V, E) be the input graph, assumed to be simple and undirected. Let n and m denote the number
of vertices and edges, respectively. The value n is known to the algorithm. Each vertex v ∈ V is associated
with a unique ID from [n] def= {1, . . . , n}. Let deg(v) denote the degree of v.
Let p ≥ 2 be a constant integer. A p-star is a subgraph of size p + 1, where one vertex, called the center,
is adjacent to the other p vertices. For example, a 2-star is an undirected path of length 2. Note that a
vertex may be a center for many stars, and a set of p + 1 vertices may form multiple stars. Let Sp denote
the number of occurrences of distinct stars in the graph.
Our goal is to construct a randomized algorithm that outputs a value that is within a (1±)-multiplicative
factor of the actual number of stars Sp. More specifically, given a parameter > 0, the algorithm must give
an approximated value (cid:98)Sp satisfying the inequality (1 − )Sp ≤ (cid:98)Sp ≤ (1 + )Sp with success probability at
least 2/3.
2.2 Query Access
The algorithm may access the input graph by querying the graph oracle, which answers for the following
types of queries. First, the neighbor queries: given a vertex v ∈ V and an index 1 ≤ i < n, the ith neighbor
of v is returned if i ≤ deg(v); otherwise, ⊥ is returned. Second, the degree queries: given a vertex v ∈ V , its
degree deg(v) is returned. Lastly, the random edge queries: a uniformly random edge {u, v} ∈ E is returned.
The query complexity of an algorithm is the total number of queries of any type that the algorithm makes
throughout the process of computing its answer.
Combining these queries, we may implement various useful sampling processes. We may perform a
uniform edge sampling using a random edge query, and a uniform vertex sampling by simply picking a
random index from [n]. We may also perform a weighted vertex sampling where each vertex is obtained with
probability proportional to its degree as follows: uniformly sample a random edge, then randomly choose
one of the endpoints with probability 1/2 each. Since any vertex v is incident with deg(v) edges, then the
probability that v is chosen is exactly deg(v)/2m, as desired.
4
2.3 Queries in the Database Model
Now we explain how the above queries in our graph model have direct interpretations in the database model.
Consider the column we wish to join on. For each valid label i, let xi be the number of rows containing this
label. We assume the ability to sample rows uniformly at random. This gives us a label i with probability
proportional to xi, which is a weighted sample from the distribution of labels. We also assume that we
can also quickly compute the number of other rows sharing the same label with a given row (analogous to
making a degree query). For example, this could be done quickly using an index on the column. Note that
if one has an index that is augmented with appropriate information, one can compute the selectivity of a
self-join query exactly in time roughly O(k log n) where k is the number of distinct elements in the column.
However, our methods can give runtimes that are asymptotically much smaller than this.
3 Upper Bounds for Counting Stars in Undirected Graphs
In this section we establish an algorithm for approximating the number of stars, Sp, of an undirected input
graph. We focus on the case where only degree queries and random edge queries are allowed. This illustrates
that even without utilizing the underlying structure of the input graph, we are still able to construct a
sublinear approximation algorithm that outperforms other algorithms under the traditional model in certain
cases.
3.1 Unbiased Estimator Subroutine
Our algorithm uses weighted vertex sampling to find stars. Intuitively, the number of samples required by
the algorithms should be larger when stars are rare because it takes more queries to find them. While the
query complexity of the algorithm depends on the actual value of Sp, our algorithm does not know this value
in advance. In order to overcome this issue, we devise a subroutine which—given a guess (cid:101)Sp for the value of
Sp—will give a (1 ± ) approximation of Sp if (cid:101)Sp is close enough to Sp or tell us that (cid:101)Sp is much larger than
Sp. Then, we start with the maximum possible value of Sp and guess multiplicatively smaller and smaller
values for it until we find one that is close enough to Sp, so that our subroutine is able to correctly output
a (1 ± ) approximation.
Our subroutine works by computing the average value of an unbiased estimator to Sp after drawing
enough weighted vertex samples. To construct the unbiased estimator, notice first that the number of
p-stars centered at a vertex v is(cid:0)deg(v)
(cid:1).2 Thus, Sp =(cid:80)
(cid:0)deg(v)
(cid:1).
(cid:1) so that Y is an unbiased estimator for Sp; that is,
(cid:18)deg(v)
(cid:19)(cid:19)
(cid:18) 2m
Next, we define the unbiased estimator and give the corresponding algorithm. First, let X be the random
variable representing the degree of a random vertex obtained through weighted vertex sampling, as explained
in Section 2.2. Recall that a vertex v is sampled with probability deg(v)/2m. We define the random variable
Y = 2m
X
(cid:0)X
p
(cid:18)deg(v)
(cid:19)
(cid:88)
(cid:88)
v∈V
p
p
E[Y ] =
deg(v)
v∈V
2m
deg(v)
p
=
v∈V
p
= Sp.
Clearly, the output ¯Y of Algorithm 1 satisfies E[ ¯Y ] = Sp. We claim that the number of samples k in
Algorithm 1 is sufficient to provide two desired properties: the algorithm returns an (1 ± )-approximation
of Sp if (cid:101)Sp is in the correct range; or, if (cid:101)Sp is too large, the anomaly will be evident as the output ¯Y will be
much smaller than (cid:101)Sp. In particular, we may distinguish between these two cases by comparing ¯Y against
(1 − ) Sp, as specified through the following lemma.
Lemma 3.1 For 0 < ≤ 1/2, with probability at least 2/3:
2For our counting purpose, if x < y then we define(cid:0)x
(cid:1) = 0.
y
5
Algorithm 1 Subroutine for Computing Sp given (cid:101)Sp with success probability 2/3
1: procedure Unbiased-Estimate((cid:101)Sp, )
v ← weighted sampled vertex obtained from a random edge query
d ← deg(v) obtained from a degree query
Yi ← 2m
p
for i = 1 to k do
k ← 36m / p2(cid:101)S1/p
(cid:0)d
(cid:1)
(cid:80)k
p
d
i=1 Yi
¯Y ← 1
return ¯Y
k
2:
3:
4:
5:
6:
7:
8:
1. If 1
2 Sp ≤ (cid:101)Sp ≤ 6Sp, then Algorithm 1 outputs ¯Y such that (1 − )Sp ≤ ¯Y ≤ (1 + )Sp;
moreover, if Sp < Sp then ¯Y ≥ (1 − )(cid:101)Sp.
2. If (cid:101)Sp > 6Sp, then Algorithm 1 outputs ¯Y such that ¯Y < 1
2(cid:101)Sp ≤ (1 − )(cid:101)Sp.
The first item of Lemma 3.1 can be proved by bounding the variance of Y using various Chebyshev’s
Inequality and identities of binomial coefficients, while the second item is a simple application of Markov’s
Inequality. Detailed proofs for these statements can be found in Appendix B.
3.2 Full Algorithm
Our full algorithm proceeds by first setting (cid:101)Sp to n(cid:0)n−1
(cid:1), the maximum possible value of Sp given by the
complete graph. We then use Algorithm 1 to check if (cid:101)Sp > 6Sp; if this is the case, we reduce (cid:101)Sp then
proceed to the next iteration. Otherwise, Algorithm 1 should already give an (1 ± )-approximation to Sp
(with constant probability). We note that if > 1/2, we may replace it with 1/2 without increasing the
asymptotic complexity.
p
Since the process above may take up to O(log n) iterations, we must amplify the success probability
of Algorithm 1 so that the overall success probability is still at least 2/3. To do so, we simply make
(cid:96) = O(log log n) multiple calls to Algorithm 1 then take the median of the returned values. Our full algorithm
can be described as Algorithm 2 below.
p
loop
for i = 1 to (cid:96) do
(cid:101)Sp ← n(cid:0)n−1
Algorithm 2 Algorithm for Approximating Sp
1: procedure Count-Stars()
2:
3:
4:
(cid:1), (cid:96) ← 40(log p + log log n)
Zi ← Unbiased-Estimate((cid:101)Sp, )
if Z ≥ (1 − )(cid:101)Sp then
(cid:101)Sp ← (cid:101)Sp/2
Z ← median{Z1,··· , Z(cid:96)}
Sp ← Z
return Sp
5:
6:
7:
8:
9:
10:
Theorem 3.2 Algorithm 2 outputs Sp such that (1 − )Sp ≤ Sp ≤ (1 + )Sp with probability at least 2/3.
The query complexity of Algorithm 2 is O
m log n log log n
.
(cid:18)
(cid:19)
2S1/p
p
6
(cid:16)
n(cid:0)n−1
(cid:1)(cid:17)(cid:101) ≤
Proof: If we assume that the events from Lemma 3.1 hold, then the algorithm will take at most (cid:100)log
(p + 1) log n iterations. By choosing (cid:96) = 40(log p + log log n), Chernoff bound (Theorem A.3) implies that
excepted for probability 1/3(p + 1) log n, more than half of the return values of Algorithm 1 satisfy the
desired property, and so does the median Z. By the union bound, the total failure probability is at most
1/3.
Now it is safe to assume that the events from the two lemmas hold. In case (cid:101)Sp > 6Sp, our algorithm will
detect this event because Z ≤ (1−)(cid:101)Sp, implying that we never stop and return an inaccurate approximation.
On the other hand, if (cid:101)Sp < Sp, our algorithm computes Z ≥ (1 − )(cid:101)Sp and must terminate. Since we only
halve (cid:101)Sp on each iteration, when (cid:101)Sp < Sp first occurs, we have (cid:101)Sp ≥ 1
terminate with the desired approximation before the value (cid:101)Sp is halved again. Thus, Algorithm 2 returns
Recall that the number of samples required by Algorithm 1 may only increase when (cid:101)Sp decreases. Thus
we may use the number of samples in the last round of Algorithm 2, where (cid:101)Sp = Θ(Sp), as the upper bound
Sp satisfying (1 − )Sp ≤ Sp ≤ (1 + )Sp with probability at least 2/3, as desired.
2 Sp. As a result, our algorithm must
p
for each previous iteration. Therefore, each of the O(log n) iterations takes O(m log log n / 2S1/p
achieving the claimed query complexity.
p
) samples,
3.3 Removing the Dependence on m
As described above, Algorithm 1 picks the value k and defines the unbiased estimator based on m, the
number of edges. Nonetheless, it is possible to remove this assumption of having prior knowledge of m by
instead computing its approximation. Furthermore, we will bound m in terms of n and Sp, so that we can
also relate the performance of our algorithm to previous studies on this problem such as [GRS11], as done
in Table 1.
3.3.1 Approximating m
probability xi/(cid:80)n
We briefly discuss how to apply our algorithm when m is unknown by first computing an approximation of
√
m. Using weighted vertex sampling, we may simulate the algorithm from [MPX07] or [BBS09] that computes
an (1 ± )-approximation to the sum of degrees using O(
n) weighted samples. More specifically, we cite
the following theorem:
Theorem 3.3 ([MPX07]) Let x1, . . . , xn be n variables, and define a distribution D that returns (i, xi) with
i=1 xi
i=1 xj. There exists an algorithm that computes a (1 ± )-approximation of S = (cid:80)n
√
using O(
Thus, we simulate the sampling process from D by drawing a weighted vertex sample v, querying its degree,
and feeding (v, deg(v)) to this algorithm. We will need to decrease used in this algorithm and our algorithm
√
by a constant factor to account for the additional error. Below we show that our complexities are at least
O(n1−1/p) which is already O(
n) for p = 2, and thus this extra step does not affect our algorithm’s
performance asymptotically.
n) samples from D.
3.3.2 Comparing m to n and Sp
For comparison of performances, we will now show some bounds relating m to n and Sp. Notice that the
function(cid:0)deg(v)
3We may use the binomial coefficients (cid:0)x
(cid:1) is convex with respect to deg(v).3 Then by applying Jensen’s inequality (Theorem A.4) to
(cid:1) for non-integral value x in the inequalities. These can be interpreted through
p
y
alternative formulations of binomial coefficients using falling factorials or analytic functions.
7
this function, we obtain
(cid:18)deg(v)
(cid:19)
p
(cid:88)
v∈V
(cid:18)(cid:80)
≥ n
Sp =
(cid:19)
(cid:18)2m/n
(cid:19)
p
.
= n
v∈V deg(v)/n
p
First, let us consider the case where the stars are very rare, namely when Sp ≤ n. The inequality above
implies that m ≤ np/2. Substituting this formula back into the bound from Theorem 3.2 yields the query
complexity O(n / 2S1/p
).
p
Now we consider the remaining case where Sp > n. If m < np/2 = O(n), then the query complexity from
Theorem 3.2 becomes O(n1−1/p / 2). Otherwise we have 2m/n ≥ p, which allows us to apply the following
bound on our binomial coefficient:
(cid:18)2m/n
(cid:19)
p
≥ n
(cid:18) 2m
(cid:19)p
.
np
Sp ≥ n
This inequality implies that m ≤ pn1−1/pS1/p
p /2, also yielding the query complexity O(n1−1/p / 2).
Compared to [GRS11], our algorithm achieves a better query complexity when Sp ≤ n1+1/p, where the
rare stars are more likely to be found via edge sampling rather than uniform vertex sampling or traversing the
graph. Our algorithm also performs no worse than their algorithm does for any Sp as large as np. Moreover,
due to the simplicity of our algorithm, the dependence on of our query complexity is only 1/2 for any
value of Sp, while that of their algorithm is as large as 1/10 in certain cases. This dependence on may
be of interest to some applications, especially when stars are rare whilst an accurate approximation of Sp is
crucial.
3.4 Allowing Neighbor Queries
We now briefly discuss how we may improve our algorithm when neighbor queries are allowed (in addition to
degree queries and random edge queries). For the case when Sp > np, it is unknown whether our algorithm
alone achieves better performance than [GRS11] (see table 1). However, their algorithm has the same basic
framework as ours, namely that it also starts by setting (cid:101)Sp to the maximum possible number of stars, then
iteratively halves this value until it is in the correct range, allowing the subroutine to correctly compute a
(1 ± )-approximation of Sp. As a result, we may achieve the same performance as them in this regime by
simply letting Algorithm 2 call the subroutine from [GRS11] when Sp ≥ np. We will later show tight lower
bounds (up to polylogarithmic factors) to the case where all three types of queries are allowed, which is a
stronger model than the one previously studied in their work.
4 Lower Bounds for Counting Stars in Undirected Graphs
In this section, we establish the lower bounds summarized in the last two columns of Table 1. We give lower
bounds that apply even when the algorithm is permitted to sample random edges. Our first lower bound is
proved in Section 4.1; While this is the simplest case, it provides useful intuition for the proofs of subsequent
bounds. In order to overcome the new obstacle of powerful queries in our model, for larger values of Sp we
create an explicit scheme for constructing families of graphs that are hard to distinguish by any algorithm
even when these queries are present. Using this construction scheme, our approach obtains the bounds for
all remaining ranges for Sp as special cases of a more general bound, and the general bound is proved via
the straightforward application of Yao’s principle and a coupling argument. Our lower bounds are tight (up
to polylogarithmic factors) for all cases except for the bottom middle cell in Table 1.
4.1 Lower Bound for Sp ≤ n
Theorem 4.1 For any constant p ≥ 2, any (randomized) algorithm for approximating Sp to a multiplicative
factor via neighbor queries, degree queries and random edge queries with probability of success at least 2/3
8
) total number of queries for any Sp ≤ np.
p
requires Ω(n/S1/p
Proof: We now construct two families of graphs, namely F1 and F2, such that any G1 and G2 drawn from
each respective family satisfy Sp(G1) = 0 and Sp(G2) = Θ(s) for some parameter s > (p + 1)p = O(1). We
construct G1 as follows: for a subset S ⊆ V of size (cid:100)s1/p(cid:101) + 1, we create a union of a (p − 1)-regular graph
on S and a (p− 1)-regular graph on V \ S, and add the resulting graph G1 to F1. To construct all graphs in
F1, we repeat this process for every subset S of size (cid:100)s1/p(cid:101) + 1. F2 is constructed a little differently: rather
than using a (p − 1)-regular graph on S, we use a star of size (cid:100)s1/p(cid:101) on this set instead. We add a union
between a star on S and a (p − 1)-regular graph on V \ S of any possible combination to F2.
By construction, every G1 ∈ F1 contains no p-stars, whereas every G2 ∈ F2 has(cid:0)O(s1/p)
(cid:1) = Θ(s) p-stars.
For any algorithm to distinguish between F1 and F2, when given a graph G2 ∈ F2, it must be able to
detect some vertex in S with probability at least 2/3. Otherwise, if we randomly generate a small induced
subgraph according to the uniform distribution in F2 conditional on not having any vertex or edge in S,
the distribution would be identical to the uniform in F1. Furthermore, notice that S cannot be reached via
traversal using neighbor queries as it is disconnected from V \ S. The probability of sampling such vertex
or edge from each query is O(s1/p/n). Thus, Ω(n/s1/p) samples are required to achieve a constant factor
approximation with probability 2/3.
p
4.2 Overview of the Lower Bound Proof for Sp > n
√
Since graphs with large Sp contain many edges, we must modify our approach above to allow graphs from the
first family to contain stars. We construct two families of graphs F1 and F2 such that the number of stars of
graphs from these families differ by some multiplicative factor c > 1; any algorithm aiming to approximate
Sp within a multiplicative factor of
c must distinguish between these two families with probability at
least 2/3. We create representations of graphs that explicitly specify their adjacency list structure. Each
G1 ∈ F1 contains n1 vertices of degree d1, while the remaining n2 = n − n1 vertices are isolated. For
each G2 ∈ F2, we modify our representation from F1 by connecting each of the remaining n2 vertices to
d2 (cid:29) d1 neighbors, so that these vertices contribute sufficient stars to establish the desired difference in Sp.
We hide these additional edges in carefully chosen random locations while ensuring minimal disturbance to
the original graph representation; our representations are still so similar that any algorithm may not detect
them without making sufficiently many queries. Moreover, we define a coupling for answering random edge
queries so that the same edges are likely to be returned regardless of the underlying graph.
While the proof of [GRS11] also uses similar families of graphs, our proof analysis greatly deviates from
their proof as follows. Firstly, we apply Yao’s principle which allows us to prove the lower bounds on ran-
domized algorithms by instead showing the lower bound on deterministic algorithms on our carefully chosen
distribution of input instances.4 Secondly, rather than constructing two families of graphs via random pro-
cesses, we construct our graphs with adjacency list representations explicitly, satisfying the above conditions
for each lower bound we aim to prove. This allows us to avoid the difficulties in [GRS11] regarding the gen-
eration of potential multiple edges and self-loops in the input instances. Thirdly, we define the distribution
of our instances based on the permutation of the representations of these two graphs, and the location we
place the edges in G2 that are absent in G1. We also apply the coupling argument, so that the distribution
of these permutations we apply on these graphs, as well as the answers to random edge queries, are as
similar as possible. As long as the small difference between these graphs is not discovered, the interaction
between the algorithm and our oracle must be exactly the same. We show that with probability 1 − o(1),
the algorithm and our oracle behave in exactly the same way whether the input instance corresponds to G1
or G2. Simplifying the arguments from [GRS11], we completely bypass the algorithm’s ability to make use
of graph structures. Our proof only requires some conditions on the parameters n1, d1, n2, d2; this allows us
to show the lower bounds for multiple ranges of Sp simply by choosing appropriate parameters.
We provide the full details in Section C. The main results of our constructions are given as the following
4See e.g., [MR10] for more information on Yao’s principle.
9
theorems. We note that lower bounds apply when only subsets of these three types of queries are provided.
This concludes all of our lower bounds in Table 1.
Theorem 4.2 For any constant p ≥ 2, any (randomized) algorithm for approximating Sp to a multiplicative
factor via neighbor queries, degree queries and random edge queries with probability of success at least 2/3
requires Ω(n1−1/p) total number of queries for any Sp = O(np).
Theorem 4.3 For any constant p ≥ 2, any (randomized) algorithm for approximating Sp to a multiplicative
factor via neighbor queries, degree queries and random edge queries with probability of success at least 2/3
(cid:19)
(cid:18)
np−1/p
S1−1/p
p
requires Ω
total number of queries for any Sp = Ω(np).
5 Acknowledgements
This material is based upon work supported by the National Science Foundation Graduate Research Fellow-
ship under Grant No. CCF-1217423, CCF-1065125, CCF-1420692, and CCF-1122374. Any opinion, findings,
and conclusions or recommendations expressed in this material are those of the authors(s) and do not nec-
essarily reflect the views of the National Science Foundation. We thank Peter Haas and Samuel Madden for
helpful discussions.
References
[ADH+08] Noga Alon, Phuong Dao, Iman Hajirasouliha, Fereydoun Hormozdiari, and S Cenk Sahinalp.
Biomolecular network motif counting and discovery by color coding. Bioinformatics, 24(13):i241–
i249, 2008.
[AFS12]
[AG09]
Omid Amini, Fedor V Fomin, and Saket Saurabh. Counting subgraphs via homomorphisms.
SIAM Journal on Discrete Mathematics, 26(2):695–717, 2012.
Noga Alon and Shai Gutner. Balanced hashing, color coding and approximate counting.
Parameterized and Exact Computation, pages 1–16. Springer, 2009.
In
[AGM12] Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Graph sketches: sparsification, spanners,
and subgraphs. In Proceedings of the 31st symposium on Principles of Database Systems, pages
5–14. ACM, 2012.
[AGMS99] Noga Alon, Phillip B Gibbons, Yossi Matias, and Mario Szegedy. Tracking join and self-join
In Proceedings of the eighteenth ACM SIGMOD-SIGACT-SIGART
sizes in limited storage.
symposium on Principles of database systems, pages 10–20. ACM, 1999.
[AMS96]
Noga Alon, Yossi Matias, and Mario Szegedy. The space complexity of approximating the
frequency moments. In Proceedings of the twenty-eighth annual ACM symposium on Theory of
computing, pages 20–29. ACM, 1996.
[AYZ97]
Noga Alon, Raphael Yuster, and Uri Zwick. Finding and counting given length cycles. Algorith-
mica, 17(3):209–223, 1997.
[BBCG08] Luca Becchetti, Paolo Boldi, Carlos Castillo, and Aristides Gionis. Efficient semi-streaming al-
gorithms for local triangle counting in massive graphs. In Proceedings of the 14th ACM SIGKDD
international conference on Knowledge discovery and data mining, pages 16–24. ACM, 2008.
[BBM12]
Eric Blais, Joshua Brody, and Kevin Matulef. Property testing lower bounds via communication
complexity. Computational Complexity, 21(2):311–358, 2012.
10
[BBS09]
Tugkan Batu, Petra Berenbrink, and Christian Sohler. A sublinear-time approximation scheme
for bin packing. Theoretical Computer Science, 410(47):5082–5092, 2009.
[BFL+06] Luciana S Buriol, Gereon Frahling, Stefano Leonardi, Alberto Marchetti-Spaccamela, and Chris-
tian Sohler. Counting triangles in data streams.
In Proceedings of the twenty-fifth ACM
SIGMOD-SIGACT-SIGART symposium on Principles of database systems, pages 253–262.
ACM, 2006.
[BGKS06] Lakshminath Bhuvanagiri, Sumit Ganguly, Deepanjan Kesh, and Chandan Saha. Simpler al-
gorithm for estimating frequency moments of data streams. In Proceedings of the seventeenth
annual ACM-SIAM symposium on Discrete algorithm, pages 708–713. ACM, 2006.
[BYKS02] Ziv Bar-Yossef, Ravi Kumar, and D Sivakumar. Reductions in streaming algorithms, with
an application to counting triangles in graphs. In Proceedings of the thirteenth annual ACM-
SIAM symposium on Discrete algorithms, pages 623–632. Society for Industrial and Applied
Mathematics, 2002.
[CK04]
Don Coppersmith and Ravi Kumar. An improved data stream algorithm for frequency moments.
In Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete algorithms, pages 151–
156. Society for Industrial and Applied Mathematics, 2004.
[CR14]
Cl´ement Canonne and Ronitt Rubinfeld. Testing probability distributions underlying aggregated
data. arXiv preprint arXiv:1402.3835, 2014.
[DLR95]
Richard A Duke, Hanno Lefmann, and Vojtech Rodl. A fast approximation algorithm for com-
puting the frequencies of subgraphs in a given graph. SIAM Journal on Computing, 24(3):598–
620, 1995.
[ELR15]
Talya Eden, Amit Levi, and Dana Ron. Approximately counting triangles in sublinear time.
arXiv preprint arXiv:1504.00954, 2015.
[Fei06]
[FG04]
Uriel Feige. On sums of independent random variables with unbounded variance and estimating
the average degree in a graph. SIAM Journal on Computing, 35(4):964–984, 2006.
Jorg Flum and Martin Grohe. The parameterized complexity of counting problems. SIAM
Journal on Computing, 33(4):892–922, 2004.
[FLR+12] Fedor V Fomin, Daniel Lokshtanov, Venkatesh Raman, Saket Saurabh, and BV Rao. Faster
algorithms for finding and counting subgraphs. Journal of Computer and System Sciences,
78(3):698–706, 2012.
[GK07]
[Gol13]
Joshua A Grochow and Manolis Kellis. Network motif discovery using subgraph enumeration and
symmetry-breaking. In Research in Computational Molecular Biology, pages 92–106. Springer,
2007.
Oded Goldreich. On the communication complexity methodology for proving lower bounds on
the query complexity of property testing. Electronic Colloquium on Computational Complexity
(ECCC), 20:73, 2013.
[GR08]
Oded Goldreich and Dana Ron. Approximating average parameters of graphs. Random Struc-
tures & Algorithms, 32(4):473–493, 2008.
[GRS11] Mira Gonen, Dana Ron, and Yuval Shavitt. Counting stars and other small subgraphs in
sublinear-time. SIAM Journal on Discrete Mathematics, 25(3):1365–1411, 2011.
[GS09]
Mira Gonen and Yuval Shavitt. Approximating the number of network motifs. Internet Mathe-
matics, 6(3):349–372, 2009.
11
[GTK01]
Lise Getoor, Benjamin Taskar, and Daphne Koller. Selectivity estimation using probabilistic
models. In ACM SIGMOD Record, volume 30, pages 461–472. ACM, 2001.
[HA08]
David Hales and Stefano Arteconi. Motifs in evolving cooperative networks look like protein
structure networks. Networks and Heterogeneous Media, 3(2):239, 2008.
[HBPS07] Fereydoun Hormozdiari, Petra Berenbrink, Natasa Przulj, and S Cenk Sahinalp. Not all scale-free
networks are born equal: the role of the seed graph in ppi network evolution. PLoS computational
biology, 3(7):e118, 2007.
[HILM09] Peter J Haas, Ihab F Ilyas, Guy M Lohman, and Volker Markl. Discovering and exploiting
statistical properties for query optimization in relational databases: A survey. Statistical Analysis
and Data Mining: The ASA Data Science Journal, 1(4):223–250, 2009.
[HKNO09] Avinatan Hassidim, Jonathan A Kelner, Huy N Nguyen, and Krzysztof Onak. Local graph
partitions for approximation and testing. In Foundations of Computer Science, 2009. FOCS’09.
50th Annual IEEE Symposium on, pages 22–31. IEEE, 2009.
[HNSS96] Peter J Haas, Jeffrey F Naughton, S Seshadri, and Arun N Swami. Selectivity and cost estimation
for joins based on random sampling. Journal of Computer and System Sciences, 52(3):550–569,
1996.
[IW05]
Piotr Indyk and David Woodruff. Optimal approximations of the frequency moments of data
streams. In Proceedings of the thirty-seventh annual ACM symposium on Theory of computing,
pages 202–208. ACM, 2005.
[KIMA04] Nadav Kashtan, Shalev Itzkovitz, Ron Milo, and Uri Alon. Efficient sampling algorithm for
estimating subgraph concentrations and detecting network motifs. Bioinformatics, 20(11):1746–
1758, 2004.
[KMPT10] Mihail N Kolountzakis, Gary L Miller, Richard Peng, and Charalampos E Tsourakakis. Efficient
triangle counting in large graphs via degree-based vertex partitioning. In Algorithms and Models
for the Web-Graph, pages 15–24. Springer, 2010.
[KMSS12] Daniel M Kane, Kurt Mehlhorn, Thomas Sauerwald, and He Sun. Counting arbitrary subgraphs
in data streams. In Automata, Languages, and Programming, pages 598–609. Springer, 2012.
[LKC99]
Ju-Hong Lee, Deok-Hwan Kim, and Chin-Wan Chung. Multi-dimensional selectivity estimation
using compressed histogram information. In ACM SIGMOD Record, volume 28, pages 205–214.
ACM, 1999.
[MHK+07] Volker Markl, Peter J Haas, Marcel Kutsch, Nimrod Megiddo, Utkarsh Srivastava, and
Tam Minh Tran. Consistent selectivity estimation via maximum entropy. The VLDB jour-
nal, 16(1):55–76, 2007.
[MMPS11] Madhusudan Manjunath, Kurt Mehlhorn, Konstantinos Panagiotou, and He Sun. Approximate
counting of cycles in streams. In Algorithms–ESA 2011, pages 677–688. Springer, 2011.
[MPX07] Rajeev Motwani, Rina Panigrahy, and Ying Xu. Estimating sum by weighted sampling.
In
Automata, Languages and Programming, pages 53–64. Springer, 2007.
[MR10]
Rajeev Motwani and Prabhakar Raghavan. Randomized algorithms. Chapman & Hall/CRC,
2010.
[MSOI+02] Ron Milo, Shai Shen-Orr, Shalev Itzkovitz, Nadav Kashtan, Dmitri Chklovskii, and Uri Alon.
Network motifs: simple building blocks of complex networks. Science, 298(5594):824–827, 2002.
12
[NO08]
Huy N Nguyen and Krzysztof Onak. Constant-time approximation algorithms via local im-
provements. In Foundations of Computer Science, 2008. FOCS’08. IEEE 49th Annual IEEE
Symposium on, pages 327–336. IEEE, 2008.
[ORRR12] Krzysztof Onak, Dana Ron, Michal Rosen, and Ronitt Rubinfeld. A near-optimal sublinear-time
algorithm for approximating the minimum vertex cover size. In Proceedings of the Twenty-Third
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1123–1131. SIAM, 2012.
[PCJ04]
N Przulj, Derek G Corneil, and Igor Jurisica. Modeling interactome: scale-free or geometric?
Bioinformatics, 20(18):3508–3515, 2004.
[PI97]
[PR07]
[Ses15]
[SIKS06]
Viswanath Poosala and Yannis E Ioannidis. Selectivity estimation without the attribute value
independence assumption. In VLDB, volume 97, pages 486–495, 1997.
Michal Parnas and Dana Ron. Approximating the minimum vertex cover in sublinear time and
a connection to distributed algorithms. Theoretical Computer Science, 381(1):183–196, 2007.
C Seshadhri. A simpler sublinear algorithm for approximating the triangle count. arXiv preprint
arXiv:1505.01927, 2015.
Jacob Scott, Trey Ideker, Richard M Karp, and Roded Sharan. Efficient algorithms for de-
tecting signaling pathways in protein interaction networks. Journal of Computational Biology,
13(2):133–144, 2006.
[SS94]
Arun Swami and K Bernhard Schiefer. On the estimation of join result sizes. Springer, 1994.
[SSRS06] Tomer Shlomi, Daniel Segal, Eytan Ruppin, and Roded Sharan. Qpath: a method for querying
pathways in a protein-protein interaction network. BMC bioinformatics, 7(1):199, 2006.
[VW09]
[Wer06]
[Wil09]
[YYI09]
Virginia Vassilevska and Ryan Williams. Finding, minimizing, and counting weighted subgraphs.
In Proceedings of the forty-first annual ACM symposium on Theory of computing, pages 455–464.
ACM, 2009.
Sebastian Wernicke. Efficient detection of network motifs. IEEE/ACM Transactions on Com-
putational Biology and Bioinformatics (TCBB), 3(4):347–359, 2006.
Ryan Williams. Finding paths of length k in o ∗ (2k) time. Information Processing Letters,
109(6):315–318, 2009.
Yuichi Yoshida, Masaki Yamamoto, and Hiro Ito. An improved constant-time approximation
algorithm for maximum.
In Proceedings of the 41st annual ACM symposium on Theory of
computing, pages 225–234. ACM, 2009.
A Useful Inequalities
This section provides standard equalities that we use throughout our paper. These inequalities exist in many
variations, but here we only present the formulations which are most convenient for our purposes.
Theorem A.1 (Chebyshev’s Inequality) For any random variable X and a > 0,
Theorem A.2 (Markov’s Inequality) For any non-negative random variable X and a > 0,
P[X − E[X] ≥ a] ≤ Var[X]
a2
P[X ≥ a] ≤ E[X]
a
.
13
Theorem A.3 (Chernoff Bound) Let X1,··· , Xn be independent Poisson random variables such that P[Xi =
1] = p for all i ∈ [n], and let X = 1
Theorem A.4 (Jensen’s Inequality) For any real convex function f with x1,··· , xn in its domain,
(cid:80)n
i=1 Xi. Then for any 0 < δ ≤ 1,
P[X > (1 + δ)p] < e−δ2pn/3.
n
n(cid:88)
f (xi) ≥ nf
i=1
i=1
(cid:32) n(cid:88)
(cid:33)
xi
1. If 1
B Proof of Lemma 3.1
Lemma 3.1 For 0 < ≤ 1/2, with probability at least 2/3:
2 Sp ≤ (cid:101)Sp ≤ 6Sp, then Algorithm 1 outputs ¯Y such that (1 − )Sp ≤ ¯Y ≤ (1 + )Sp;
moreover, if Sp < Sp then ¯Y ≥ (1 − )(cid:101)Sp.
2. If (cid:101)Sp > 6Sp, then Algorithm 1 outputs ¯Y such that ¯Y < 1
Proof: Let us first consider the first item. Since Var[Y ] ≤ E[Y 2], we will focus on establishing an upper
bound of E[Y 2]. We compute
2(cid:101)Sp ≤ (1 − )(cid:101)Sp.
(cid:18) 2m
(cid:18)deg(v)
(cid:19)(cid:19)2
(cid:88)
(cid:32)(cid:88)
(cid:19)2−1/p ≤ 2m
(cid:18)deg(v)
(cid:18)deg(v)
where the first inequality holds because (deg(v))p ≥(cid:0)deg(v)
(cid:1). Rearranging the terms, we have the following
(cid:18)deg(v)
(cid:19)2
(cid:19)(cid:33)2−1/p
= 2mS2−1/p
(cid:88)
(cid:88)
≤ 2m
E[Y 2] =
deg(v)
deg(v)
deg(v)
= 2m
v∈V
v∈V
v∈V
v∈V
2m
p
p
p
p
p
,
1
relationship:
Now let us consider our average ¯Y . Since Yi are identically distributed, we have
Var[ ¯Y ] = Var
Yi
=
1
k2 Var
Yi
=
1
k
Var[Y ] ≤ 1
k
E[Y 2].
p
E[Y 2]
S2
p
≤ 2m
pS1/p
p
.
(cid:35)
(cid:34)
k(cid:88)
i=1
1
k
(cid:35)
(cid:34) k(cid:88)
i=1
By Chebyshev’s inequality (Theorem A.1), we have
In order to achieve the desired value ¯Y such that (1 − )Sp ≤ ¯Y ≤ (1 + )Sp with error probability 1/3, it
is sufficient to take 6m / p2S1/p
the number of required samples to achieve such bound with probability 1/3 is
2 Sp ≤ (cid:101)Sp ≤ 6Sp. Thus,
Pr[ ¯Y − E[ ¯Y ] ≥ Sp] ≤ Var[ ¯Y ]
2S2
p
samples. Recall the assumption that (cid:101)Sp satisfying 1
p2S1/p
≤ 1
k
2m
·
p
p
.
For the second item, we apply Markov’s Inequality (Theorem A.2) to the given condition to obtain
(cid:20)
P
Y ≥ 1
2
(cid:21)
(cid:101)Sp
k =
36m
p2(cid:101)S1/p
p
.
≤ E[Y ]
2(cid:101)Sp
1
2(cid:101)Sp
Sp
1
<
=
14
6(cid:101)Sp
2(cid:101)Sp
1
1
=
1
3
,
implying the desired success probability.
Lastly, we substitute < 1/2 to obtain the relationship between ¯Y and (1 − ) Sp, which establishes the
condition for deciding whether the given Sp is much larger than Sp, as desired.
C Proof of Lower Bounds for Undirected Graphs with Sp > n
In this section we provide the proof of lower bounds claimed in Section 4.2. Firstly, to properly describe the
adjacency list representation of the input graphs, we introduce the notion of graph representation. Next, we
state a main lemma (Lemma C.1) that establishes the constraints of parameters n1, d1, n2, d2 that allows us
to create hard instances. We then move on to describe our constructions, including both the distribution
for applying Yao’s principle, and the implementation of the oracle for answering random edge queries. We
prove our main lemma for our construction, and lastly, we give the appropriate parameters that complete
the proof of our lower bounds.
C.1 Graph Representations
Consider the following representation L of an adjacency list for an undirected graph G. Let us say that each
vertex vi has deg(vi) ports numbered 1, . . . , deg(v) attached, where the jth port of vertex vi is identify as
a pair (i, j), which is used as an index for L. L imposes a perfect matching between these ports; namely,
L(i1, j1) = (i2, j2) indicates that ports (i1, j1) and (i2, j2) are matched to each other, and this implies
L(i2, j2) = (i1, j1) as well. We use L to define the adjacency list of our graph; that is, if L(i1, j1) = (i2, j2)
then the jth
1 neighbor of vi1 is vi2 (and vice versa). Note that there can be many such representations of G,
and some perfect matchings between ports may yield graphs parallel edges or self-loops. Furthermore, each
edge e is associated with a unique pair of matched cells.
C.2 Main Lemma
Our proof proceeds in two steps. First, we show the following lemma that applies to certain parameters of
graphs.
Lemma C.1 Let n1, d1, n2, d2 be positive parameters satisfying the following properties: d1 and n2 are even,
n2 ≤ d1 ≤ 2d2 and d1 + 2d2 < n1. Let n = n1 + n2, and define the following two families of graphs on n
vertices:
• F1: all graphs containing n1 vertices of degree d1 and n2 isolated vertices;
• F2: all graphs containing n1 vertices of degree d1 and n2 vertices of degree d2.
d1n1
and q = o(1/r). Then, there exists a distribution D of representations of graphs from
Let r = (d1+d2)n2
F1 ∪ F2 such that for any deterministic algorithm A that makes at most q total neighbor queries, degree
queries and random edge queries, on the graph representation randomly drawn from D, A cannot correctly
identify whether the given representation is of a graph from F1 or F2 with probability at least 2/3.
(cid:0)d1
p
By applying Yao’s principle, the following corollary is implied.
(cid:1) and s2 = n1
Corollary C.2 Let n1, d1, n2, d2 be parameters satisfying the properties specified in Lemma C.1. Let s1 =
n1
(randomized) algorithm for approximating Sp to a multiplicative factor via neighbor queries, degree queries
and random edge queries with probability of success at least 2/3 requires Ω(q) queries for Sp = Θ(f (n, p)).
(cid:1). If s1 = Θ(f (n, p)) and s2 ≥ c · s1 for some constant c > 1, then any
(cid:1) + n2
(cid:0)d2
p
(cid:0)d1
p
As a second step, we propose a few sets of parameters for different ranges of Sp. Applying Corollary C.2,
this yields lower bounds for the remaining ranges of Sp.
15
Figure 1: first few columns of L1
C.3 Our Constructions
C.3.1 Construction of D
We prove this lemma by explicitly constructing the distribution.
Construction of graph representations for F1. We now define the representation L1 for the graph
G1 ∈ F1 as follows. We let v1, . . . , vn1 be the vertices with degree d1. Let us refer to the jth pair of
consecutive columns (with indices 2j − 1 and 2j) as the jth slab. Then, in the jth slab, we match each cell
on the left column with the cell at distance j below on the right column. Figure 1 illustrates the matching
of cells in the first few columns of L1. More formally, for each integer i ∈ [n1] and j ∈ [d1/2], we match the
cells (i, 2j − 1) and (i + j mod n1, 2j) in L1.
Since d1 is even, this construction fills the entire table of L1. We wish to claim that we do not create
any parallel edges with this construction. Clearly, this is true within a slab. For different slabs, recall that
we map cells in the jth slab with those at vertical distance j away. Thus, it suffices to note that no pair of
slabs uses the same distance mod n1. Equivalently, we can note that as the maximum distance is d1/2 and
d1/2 < n1/2 by our assumption, the set of distances {j, n1 − j} for j ∈ [d1/2] are all disjoint. That is, our
construction creates no parallel edges or self-loops.
Construction of graph representations for F2. Next, for each integer x ∈ [n1] and y ∈ [d/2], we define
a graph Gx,y
by modifying L1 as follows. First, recall that we need
to add neighbors to the previously isolated vertices vn1+1, . . . , vn. These neighbors are represented as a table
of size n2 × d2 in Lx,y
(a) which is not
present in L1. We match the cells in this new table to a subtable of size d2 × n2, which is shown as the
(a). The top-left cell of this subtable corresponds to the index (x, 2y − 1) in
yellow rectangle in Figure Lx,y
2 , and note that if x + d2 > n1 or 2y + n2 > d1, this subtable may wrap around as shown in Figure Lx,y
Lx,y
(b). Since n2 ≤ d1 and d2 < n1, the dimensions of this yellow rectangle does not exceed the original table in
L1.
2 ; in Figure 2, it is represented as the green rectangle in Figure Lx,y
2
2 with corresponding representation Lx,y
2
2
2
Now we explain how we match the cells. Between the yellow and green subtables, we map them in a
transposed fashion. That is, the cell with index (i, j) (relative to the green table) is mapped to the yellow
cell with index (j, i) (relative to the yellow subtable), as shown in Figure 3 (a). This method guarantees
that no two rows contain two pair of matched cells between them. As a result, we do not create any parallel
edges or self-loops.
As we place the yellow subtable, some edges originally in L1 may now have only one endpoint in the
yellow subtable. We refer to the cells in the table that correspond to such edges as unmatched. Since n2 is
even and we set our offset to (x, 2y − 1), then every slab either does not overlap with the yellow subtable, or
16
123456……𝑖𝑖−1𝑖−2𝑖+1𝑖+2⋮⋮⋮⋮ Figure 2: Comparison between tables L1 and L2. Lx,y
depending on the values of x and y.
2
(a) and (b) show two different possibilities for Lx,y
2
Figure 3: matchings in Lx,y
2
17
𝑛1𝑑2𝐿1𝐿2𝑥,𝑦(a)𝑥,2𝑦−1𝑛2𝑑1𝑛2𝑑2𝑛1𝑑1𝑛2𝐿2𝑥,𝑦(b)𝑥,2𝑦−1(a)≤𝑑1/2≤ 𝑑12+𝑑2𝑑2(b) overlaps in the exact same rows for both columns of the slab. Thus, the only edges that have one endpoint
in the yellow subtable are those that go from a cell above it to one in it. Roughly speaking, we still map the
cells in the same way but ignore the distance it takes to skip over the yellow subtable. More formally, in the
jth slab, we pair each unmatched cell from the left and right respectively that are at vertical distance j + d2
away (instead of j), as shown with the red edges in Figure 3 (b).
Now the set of distances between the cells corresponding to an edge in the jth slab are {j, j + d2, n1 −
j, n1 − (j + d2)}, since distances can be measured both by going down and by going up and looping around.
From our assumption, d1/2 ≤ d2 and d1/2 + d2 < n1/2, and thus no distance is shared by multiple slabs,
and thus there are no parallel edges or self-loops.
Permutation of graph representations. Let π be a permutation over [n].5 Given a graph representation
L, we define π(L) as a new presentation of the same underlying graph, such that the indices of the vertices
are permuted according to π. We may consider this operation as an interface to the original oracle. Namely,
any query made on a vertex index i is translated into a query for index π(j) to the original oracle. If a vertex
index j is an answer from the oracle, then we return π−1(j) instead.
The distribution D. Let Sn denote the set of all n! permutations over [n]. We define D formally as follows:
for any permutation π ∈ Sn, the representation π(L1) corresponding to G1 is drawn from D with probability
1/(2n!), and each representation π(Lx,y
is drawn with probability 1/(n1d1n!) for
every (x, y) ∈ [n1] × [d1/2]. In other words, to draw a random instance from D, we flip an unbiased coin
to choose between families F1 and F2. We obtain a representation L1 if we choose F1; otherwise we pick a
random representation Lx,y
for F2. Lastly, we apply a random permutation π to such representation.
2 ) corresponding to Gx,y
2
2
C.3.2 Answering Random Edge Queries
Notice that Yao’s principle allows us to remove randomness used by the algorithm, but the randomness
of the oracle remains for the random edge queries. For any representation we draw from D, the oracle
must return an edge uniformly at random for each random edge query. Nonetheless, we may choose our
own implementation of the oracle as long as this condition is ensured. We apply a coupling argument that
imposes dependencies between the behaviors of our oracle between when the underlying graph is from F1
or F2. Let m1 = d1n1/2 and m2 = (d1n1 + d2n2)/2 denote the number of edges of graphs from F1 and F2,
respectively.
Our oracle works differently depending on which family the graph comes from. The following describes
the behavior of our oracle for a single query, and note that all queries should be evaluated independently.
Query to L1. We simply return an edge chosen uniformly at random. That is, we pick a random matched
pair of cells in L1, and return the vertices corresponding to the rows of those cells.
Query to Lx,y
mx,y
random from the set of edges in Lx,y
2 . With probability
s /m2, we return the same edge we choose for L1. Otherwise, we return an edge chosen uniformly at
denote the number of edges shared by both L1 and Lx,y
2 . Let mx,y
s
2 but not in L1.
Our oracle clearly returns an edge chosen uniformly at random from the corresponding representation.
The benefit of using this coupling oracle is that we increase the probability that the same edge is returned to
mx,y
are fully contained within the
subtable of size (d1 + d2)n2 obtained by extending the yellow subtable to include d1/2 more rows above and
s ≥ (d1n1 − (d1 + d2)n2)/2. Thus, our oracle may only return a different edge with probability
below. mx,y
s /m2. By our construction, the cells in L1 that are modified to obtain Lx,y
2
=
d1n2
d1n1 + d2n2
≤ r.
1 − mx,y
s
m2
= 1 − d1n1 − (d1 + d2)n2
d1n1 + d2n2
C.4 Proof of Lemma C.1
Recall that we consider a deterministic algorithm A that makes at most q = o(1/r) queries. We may
describe the behavior between A and the oracle with its query-answer history. Notice that since A is
5A permutation π over [n] is a bijection π : [n] → [n].
18
deterministic, if every answer that A receives from the oracle is the same, then A must return the same
answer, regardless of the underlying graph. Our general approach is to show that for most permutations
π, running A with instance π(L1) will result in the same query-answer history as running with π(Lx,y
2 ) for
most random parameters π and (x, y). If these histories are equivalent, then A may answer correctly for
only roughly half of the distribution.
Throughout this section, we refer to our indices before applying π to the representation. We bound the
probability that the query-answer histories are different using an inductive argument as follows. Suppose
that at some point during the execution of A, the history only contains vertices of indices from [n1], and all
cells in the history are matched in the same way in both L1 and Lx,y
2 . This inductive hypothesis restricts
the possible parameters π and (x, y) to those that yield same history up to this point. We now consider the
probability that the next query-answer pair differs, and aim to bound this probability by O(r).
Firstly, we consider a degree query. By our hypothesis, for a vertex of index outside [n1] to be queried,
A must specify a vertex it has not chosen before. Notice that A may learn about up to 2 vertices from
each query-answer pair, so at least n − 2q vertices have never appeared in the history. Since we pick a
random permutation π for our construction, the probability that the queried vertex has index outside [n1]
is n2/(n − 2q). As r ≥ n2/n1 ≥ 1/n1, we have q = o(n1) and our probability simplifies to at most
n2
n − 2q
=
n2
(n1 + n2) − 2 · o(n1)
≤
n2
n1(1 − o(1))
= O(r).
Next, we consider a neighbor query. From the argument above, with probability 1 − O(r), the queried
vertex given by A has an index from [n1]. Similarly, A may learn about up to 2 cells from each query-answer
pair. Notice that there are (d1 +d2)n2 different possible (x, y) for which each of these cells could be located in
the yellow subtable or the two (d1/2)× n2 strips above and below it. As a result, out of d1n1− ((d1 + d2)n2)q
remaining possible locations for the yellow subtable, the queried cell and the corresponding answer may be
in at most 2(d1 + d2)n2 of them. As (x, y) is randomly chosen, the probability that this next query-answer
pair is different is at most
2(d1 + d2)n2
d1n1 − ((d1 + d2)n2)q
=
2r
1 − rq
=
2r
1 − o(1)
= O(r).
Lastly, we consider a random edge query. From the construction in Section C.3.2 above, the probability
that the returned random edge differs is O(r), regardless of the parameters.
From this inductive argument, the probability that the history differs at each step is at most O(r). As A
only make q queries, the probability that the history differs is at most q · O(r) = o(1). Thus with probability
1 − o(1), it is impossible for A to distinguish whether the underlying graph is from F1 or F2. Since each
family is included in D with probability density 1/2, as A is deterministic, the answer given by A for these
cases is correct for only half of them. Thus, the probability of A correctly distinguish between the two graph
families is only 1 − 1
2 (1 − o(1)) = 1
2 + o(1), as required.
C.5 Establishing Lower Bounds
Now we propose the feasible asymptotic parameters according to Lemma C.1 and Lemma C.2 in order to
establish our lower bounds through the following claim.
Claim C.3 There exists parameters n1, d1, n2, d2 satisfying the properties specified in Lemma C.1, yielding
values s1, s2 satisfying the properties in Lemma C.2, for each of the following cases:
1. n1 = Θ(n), d1 = Θ((s/n)1/p), n2 = Θ(1), d2 = Θ(s1/p) for f (n, p) = O(np)
2. n1 = Θ(n), d1 = Θ((s/n)1/p), n2 = Θ(s/np), d2 = Θ(n) for f (n, p) = Ω(np)
We omit the proof of this claim; our proof only requires straightforward calculation, and a very similar
analysis can be found in [GRS11]. By computing the value r for each case and applying Lemma C.2, we
obtain Theorem 4.2 and Theorem 4.3, respectively.
19
D Extension to Directed Graphs
In this section, we extend our model to the directed case. Firstly, we formally give the specification of this
new model. Since most of the specification from the undirected graph model given in Section 2 still applies
to the directed case, we only explain the differences between these models. We assume separate adjacency
lists for in-neighbors and out-neighbors, allowing for a neighbor about either type of neighbor. Similarly, a
degree query may ask for either of the in-degree or the out-degree. Random edge queries now return directed
edges (u, v); the algorithm knows both the endpoints and the direction. We focus on the simplest case of
stars with mixed directions: approximately counting the number of paths of length two.
Notice the number of stars where all edges point inward or outward can be computed easily by modifying
the weighted vertex sampling to sample using in-degree or out-degree respectively and then applying the
algorithm from Section 3. This works because the numbers of such stars only depend on the in-degrees and
the out-degrees, respectively. Thus, we turn to the problem of counting directed paths of length two as the
next simplest case.
D.1 Lower Bound
By constructing hard instances similar to those of Lemma 4.1, we obtain a lower bound of Ω(n). More
formally, letting L(G) denote the number of paths of length two in the directed graph G, we prove the
following theorem.
Theorem D.1 Any (randomized) algorithm for approximating L(G) to a multiplicative factor via neighbor
queries, degree queries and random edge queries requires Ω(n) total number of queries. In particular, this
number of queries is necessary to distinguish the case where L(G) = 0 and the case where L(G) = n with
probability 2/3.
Proof: Without loss of generality, we assume n is even. Now, we partition the vertex set V into S and T
such that S = T = n/2. Let G1 be the family of graphs that contains only G1, the complete bipartite
graph where every vertex in S has an edge pointing to every vertex in T . Let G2 be the family of graphs
G(t,s) constructed by taking the graph from G1 and adding one extra back edge (t, s) ∈ T × S. Notice that
there can be many adjacency list representations of each graph, and this affect the answers to neighbor
queries. We associate each possible adjacency list representation to each graph, and include all possible such
representations in the family.
Clearly, L(G1) = 0, whereas L(G(t,s)) = n for every G(t,s) ∈ G2. For any algorithm to distinguish between
G1 and G2, when given a graph G(t,s) from G2, it must be able to detect the vertex s or t, the endpoints
of the extra edge, with probability at least 2/3. Otherwise, if neither s nor t is discovered, the subgraph
induced by vertices that the algorithm sees from both families would be exactly the same. The probability
of sampling vertices s or t from a vertex sampling, as well as their incident edges from an edge sampling, is
O(1/n). Similarly, in order to reach s or t from one of their neighbors, the algorithm must provide the index
of s or t in order to make such neighbor query, which may only succeed with probability O(1/n). Thus, Ω(n)
samples are required in order to find s or t with probability 2/3, which establishes our lower bound.
as L =(cid:80)
D.2 Upper Bound
(v) · deg+(v), which represents the number of length two paths whose
For each v ∈ V , define l(v) = deg
middle vertex is v. Thus the number of paths of length two, which we aim to approximate, can be written
v∈V l(v). Notice that 2n degree queries suffice for exactly computing the number of such paths,
already matching the lower bound. We explore this problem further by making an assumption in attempt
to obtain an algorithm that requires o(n) queries. To this end, we restrict to direct graphs such that there
exists a bound on the ratio of in-degree to out-degree. More specifically, we assume that there exists a value
−
20
r ≥ 1 such that 1
in G.
r ≤ deg−(v)
deg+(v) ≤ r, limiting the ratio between the in-degree and the out-degree of any vertex
of each vertex closer to the number of paths centered at that vertex by the rejection sampling method. Then
Under this additional assumption, we obtain a sublinear time algorithm by reduction to what is essentially
the undirected case. Our approach is to modify the weighted vertex sampling process via rejection sampling so
that the probability of sampling a vertex v becomes proportional to(cid:112)l(v), bringing the sampling probability
v∈V ((cid:112)l(v))2, which requires some modification to the algorithm,
we use Algorithm 2 to approximate (cid:80)
a random vertex such that each vertex v is returned with probability (cid:112)l(v)/(cid:80)
(cid:112)l(v(cid:48)) by making O(r)
Claim D.2 In the directed graph model, given weighted vertex sampling and degree queries, we may generate
explained later. First, we explain the details of our rejection sampling method.
v(cid:48)∈V
queries in expectation given the aforementioned assumption.
keep u with probability 1√
r
Thus, O(r) queries are required to generate one such sample.
deg+(u) . Otherwise, discard u and repeat the process.
Each vertex u is chosen from a random edge sampling is proportional to its in-degree, deg
(cid:113) deg−(u)
deg+(u) = (cid:112)l(u), as desired. Since 1
Proof: We draw a random edge sample (u, v), and query for u’s in-degree and out-degree. We return u
with probability 1√
r
(u). We only
(cid:113) deg−(v)
deg+(u) , so the probability that any vertex u is actually returned is proportional
deg+(v) ≤ 1.
(cid:113) deg−(u)
to deg+(u) ·(cid:113) deg−(u)
(cid:112)l(v). We now have a method to sample a vertex v with probability (cid:112)l(v)/L(cid:48) by
Define L(cid:48) =(cid:80)
the new distribution given above. Second, we redefine X = (cid:112)l(v) and Y = X · L(cid:48) = (cid:112)l(v) · L(cid:48), so that
increasing the time or query complexities only asymptotically by a factor of r. Now we make the following
changes to Algorithm 1 so that it approximates L. First, the algorithm should draw random vertices from
E[Y ] = L. Note that the value L(cid:48) can be approximated via essentially the same method as in Section 3.3.1.
√
nL2).
The proof of the variance bound (Lemma 3.1) can be subsequently modified to obtain Var[Y ] = O(
√
(This is essentially the problem of approximating the second frequency moment: see [AMS96] for more
details.) That is, O(
n) queries, suffice to
obtain a (1 ± )-approximation of L. This concludes the proof of the following theorem.
r ≤ deg−(v)
deg+(v) ≤ r, we have that 1
n) samples from this new distribution, or equivalently O(r
v∈V
√
−
r ≤ 1√
r
Theorem D.3 Assuming there exists some value r such that 1
graph G, then there exists an algorithm that, using degree queries and random edge queries, computes a
(1 ± )-approximation of the number of paths of length two in G with success probability 2/3 using O(r
n)
queries.
deg+(v) ≤ r for every v ∈ V in the directed
√
r ≤ deg−(v)
Corollary D.4 Assuming that the ratio between the in-degree and the out-degree of every vertex in the
directed graph G is bounded above and below by a constant, then there exists an algorithm that, using degree
queries and random edge queries, computes a (1 ± )-approximation of the number of paths of length two in
√
G with success probability 2/3 using O(
n) queries.
21
|
1712.06690 | 2 | 1712 | 2018-04-23T15:32:39 | An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline | [
"cs.DS"
] | Previous work has suggested that the structural restrictions of graphs from classes of bounded expansion--locally dense pockets in a globally sparse graph--naturally coincide with common properties of real-world networks such as clustering and heavy-tailed degree distributions. As such, fixed-parameter tractable algorithms for bounded expansion classes may offer a promising framework for network analysis where other approaches have struggled to scale. However, there has been little work done in implementing and evaluating the performance of these structure-based algorithms. To this end we introduce CONCUSS, a proof-of-concept implementation of a generic algorithmic pipeline for classes of bounded expansion. In particular, we focus on using CONCUSS for subgraph isomorphism counting (also called motif or graphlet counting), which has been used extensively as a tool for analyzing biological and social networks. Through a broad set of experiments we first evaluate the interactions between implementation/engineering choices at multiple stages of the pipeline and their effects on overall run time. From there, we establish viability of the bounded expansion framework by demonstrating that in some scenarios CONCUSS achieves run times competitive with a popular algorithm for subgraph isomorphism counting that does not exploit graph structure. Finally, we empirically identify two particular ways in which future theoretical advances could alleviate bottlenecks in the algorithmic pipeline. | cs.DS | cs |
An Experimental Evaluation of a Bounded
Expansion Algorithmic Pipeline
Michael P. O'Brien and Blair D. Sullivan
North Carolina State University, (mpobrie3blair_sullivan)@ncsu.edu
Abstract
Previous work has suggested that the structural restrictions of graphs from classes of bounded
expansion-locally dense pockets in a globally sparse graph-naturally coincide with common
properties of real-world networks such as clustering and heavy-tailed degree distributions. As
such, fixed-parameter tractable algorithms for bounded expansion classes may offer a promising
framework for network analysis where other approaches have struggled to scale. However, there
has been little work done in implementing and evaluating the performance of these structure-
based algorithms. To this end we introduce CONCUSS, a proof-of-concept implementation of a
generic algorithmic pipeline for classes of bounded expansion. In particular, we focus on using
CONCUSS for subgraph isomorphism counting (also called motif or graphlet counting), which
has been used extensively as a tool for analyzing biological and social networks. Through a
broad set of experiments we first evaluate the interactions between implementation/engineering
choices at multiple stages of the pipeline and their effects on overall run time. From there, we
establish viability of the bounded expansion framework by demonstrating that in some scenarios
CONCUSS achieves run times competitive with a popular algorithm for subgraph isomorphism
counting that does not exploit graph structure. Finally, we empirically identify two particular
ways in which future theoretical advances could alleviate bottlenecks in the algorithmic pipeline.
Digital Object Identifier 10.4230/LIPIcs...
Introduction
1
The analysis of complex networks has proven useful for understanding relationships between
entities in large data sets. For example, networks modeling social interactions, protein func-
tionality, and accessible transportation routes have been studied extensively in the network
science community. However, since many network properties that might yield interesting
insights are NP-hard to compute, the prevailing strategies for detecting such properties gen-
erally depend on sampling, estimation, and/or heuristics whose reasons for success or failure
are not theoretically well understood.
At the same time, the structural graph theory and parameterized algorithms communities
have a large body of literature on exactly solving NP-hard problems in polynomial time
on sparse graphs by exploiting the underlying graph structure. Unfortunately, many of
these results either rely on structure like low treewidth, which is unlikely to occur in many
complex networks, or hide enormous constants in big-O notation that in practice negate the
theoretical gains from the polynomial run times. Recent work has identified a structural class
of graphs known as bounded expansion that gives significant algorithmic advantages [21, 22,
11] while accommodating the structural characteristics of complex networks [10]. Though
there are some algorithms for bounded expansion that appear to be suitable for practical
use, they have yet to be implemented and tested in order to understand their behavior.
In this work, we seek to establish the use of bounded expansion structure as a viable
pathway to creating scalable software tools for network analysis. As a starting point, we
focus specifically on the problem of subgraph isomorphism counting in which the number of
occurrences of a small subgraph H (sometimes called a motif or graphlet) in a larger host
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
graph G is counted. Because previous research suggests that subgraph isomorphism counting
can be used as a building block for understanding network dynamics [29] and aligning similar
graphs [20], an improved counting algorithm would have many practical applications. Our
main contribution is a set of computational experiments using CONCUSS [24], an open source
software tool written in pure Python that implements the best known algorithm for subgraph
isomorphism counting in bounded expansion graphs [10]. On a high level, CONCUSS finds
a coloring of the vertices to identify the relevant structure, uses the coloring to decompose
the graph into small pieces, counts the isomorphisms on each small piece, and combines
the subsolutions to get a final count. Notably, the theoretical description of this pipeline
leaves room for multiple choices for implementing certain subroutines; we implement several
variants of these routines and discuss the associated algorithm engineering decisions.
Our experimental design is oriented towards understanding not only the strengths of the
algorithmic pipeline of CONCUSS but also the ways in which further theoretical research
could alleviate bottlenecks. Theoretical advances to this pipeline would not only enable
faster subgraph isomorphism counting, but also provide an improved framework for solving
other combinatorial problems in graphs of bounded expansion (see Section 3). As such, we
answer the following four questions:
Q1. What choices of subroutines give the best performance in CONCUSS?
Q2. How does CONCUSS compare to other algorithms for subgraph isomorphism counting?
Q3. How does the distribution of color class sizes influence downstream performance?
Q4. Would alternative colorings with weaker structural guarantees incur significant penalties?
After giving the necessary background and notation in Section 2, we briefly describe
the algorithmic pipeline CONCUSS implements in Section 3. Section 4 details the general
experimental setup.
In Section 5 we show the performance of various configurations of
CONCUSS compared to each other (Q1) and to the NetworkX [13] implementation of the
popular VF2 algorithm [9] (Q2). We then investigate in Section 6 ways in which theoretical
improvements to the coloring algorithms might improve CONCUSS (Q3 and Q4).
Background
2
All graphs in this paper are assumed to be simple and undirected unless otherwise specified.
For a graph G, let V (G) and E(G) denote the vertices and edges of G, respectively, and let
uv denote an edge with endpoints u and v. For convenience, we define G = V (G). We
will write Pn and Sn for the path and star on n vertices, respectively. A (vertex) coloring φ
is a mapping of V (G) to a set of k colors; we say that k is the size of φ. A color class C of
φ is a maximal set of vertices that for all u, v ∈ C, φ(u) = φ(v).
The remainder of this section gives background on subgraph isomorphism counting,
bounded expansion, and a related structural class called bounded treedepth.
2.1 Subgraph Isomorphism Counting
(cid:73) Definition 1. Given graphs G, H, a subgraph isomorphism from H to G is a mapping
ψ : V (H) → V (G) such that uv ∈ E(H) ⇐⇒ ψ(u)ψ(v) ∈ E(G).
The subgraph isomorphism counting problem is to count the number of distinct isomor-
phisms from V (H) to V (G); solving this problem is #W[1]-hard1 [16]. The most successful
1 There is likely no O(f(H) · nc) time algorithm for any computable function f and constant c.
Michael P. O'Brien and Blair D. Sullivan
XX:3
algorithms to date used for subgraph isomorphism counting have relied on backtracking [31,
4] or expanding partial mappings [8, 9, 6]. Other subclasses of algorithms have focused on
indexing graph databases [28, 14] or creating data structures for parallel computation [30].
Subgraph isomorphism counting is important as a building block for two other network
analysis techniques, described below.
Motif Counting: Originally used in the study of cellular biology [29], motif counting
computes the number subgraph isomorphisms expected to occur "by chance" and compares
it to the observed subgraph isomorphism count. If H represents some functional unit in the
domain of the data-i.e., cliques in social networks corresponding to people with common
interests-an unexpected abundance of H in G gives insight into the domain dynamics that
caused these motifs to occur.
Graphlet Degree Distribution: Subgraph isomorphism counting is also a building block
in the graphlet degree distribution [25], which measures the number of times a vertex occurs
in distinct "positions" of multiple small subgraphs. Graphlet degree distributions give a
more robust "fingerprint" of a graph's structure which has in turn been used to uncover
underlying domain knowledge [20]. They also have been used in network alignment [19],
using the knowledge that two vertices are more likely to be mapped to one another if they
occupy the same positions at similar frequencies.
2.2 Bounded Expansion
The algorithms implemented in CONCUSS are designed to exploit bounded expansion net-
work structure. While classes of bounded expansion can be characterized by their r-shallow
minors [21] or their neighborhood complexities [27], we focus on the algorithmically relevant
characterization using p-centered colorings [21].
(cid:73) Definition 2. A p-centered coloring of graph G is a vertex coloring such that every con-
nected subgraph H has a unique color or uses at least p colors. The minimum size of a
p-centered coloring of G is denoted χp(G).
(cid:73) Proposition 1.
function f, every graph G ∈ C satisfies χp(G) ≤ f(p) for all p ≥ 1.
Intuitively, p-centered colorings identify overlapping, locally well-structured regions of the
graph; being in a class of bounded expansion implies the graph can be covered by a small
number of such regions. More specifically we note that by Definition 2, for any p-centered
coloring φ of graph G, if there is a subgraph H for which φH uses fewer than p colors, then
every subgraph of H has a vertex of unique color. This relates classes of bounded expansion
to a much more restricted class of graphs described below.
[21] A class of graphs C has bounded expansion if and only if for some
2.3 Centered Colorings and Treedepth
(cid:73) Definition 3. A centered coloring of graph G is a vertex coloring such that every connected
subgraph H has a vertex of unique color, called a center. The minimum number of colors
needed for a centered coloring of G is its centered coloring number, denoted χcen(G).
Note that a centered coloring is also proper, or else there would be a connected subgraph of
size two with no center. Centered colorings are closely related to treedepth decompositions.
(cid:73) Definition 4. A treedepth decomposition of a graph G is an injective mapping ψ : V (G) →
V (F), where F is a rooted forest and uv ∈ E(G) =⇒ ψ(u) is an ancestor or descendant
XX:4
An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
of ψ(v). The depth of a treedepth decomposition is the height of F. The treedepth of G,
denoted td(G) is the minimum depth of a treedepth decomposition of G.
In other words, a treedepth decomposition arranges the vertices of G in such a way that
no edge joins vertices from different branches of the tree. Given a centered coloring with k
colors, we can generate a treedepth decomposition of depth at most k by choosing center
v to be the root and recursing on the components of G\{v}, as detailed in Algorithm 1
(Appendix A). In this way, Definition 2 is equivalent to the fact that every small set of
colors in a p-centered coloring induces components with small treedepth.
CONCUSS
3
Broadly speaking, the algorithmic pipeline in CONCUSS executes four modules to count the
number of isomorphisms of graph H in G.
1. Color:
2. Decompose:
Find a (H + 1)-centered coloring of G.
Compute a treedepth decomposition of the subgraph induced by each set of H colors.
Using dynamic programming, count the isomorphisms in the treedepth decompositions.
3. Compute:
4. Combine:
Combine the counts from previous step to get the total number of isomorphisms in G.
In Appendix B, we briefly describe these modules and the algorithm engineering aspects
considered as part of the implementation. Many of the subroutines in the modules can
achieve the same high-level functionality with different implementation details, e.g., heuris-
tics, data structures, etc. We draw particular attention to the following areas of algorithm
engineering: (1) post-processing the p-centered coloring to merge color classes and reduce
the coloring size (Color), (2) saving partial treedepth decompositions when enumerating
sets of colors (Decompose), and (3) representing partial isomorphisms as bitvectors to
enable fast checking of consistency (Compute). CONCUSS allows users to specify their
choices for subroutines in a configuration file; for more details, see the CONCUSS source and
documentation [24]. These alternatives are designated with teletype.
We note that the four-stage algorithmic workflow (Color, Decompose, Compute,
Combine) can be useful for solving problems other than subgraph isomorphism counting.
In particular, this approach is amenable to problems in which local subsolutions can be effi-
ciently computed on graphs of low treedepth and then combined to create a global solution.
For such a problem, it is possible to reuse the Color and Decompose modules and simply
replace the Compute and Combine modules with the appropriate alternatives. Dvořák et
al. [12] proved that any problem expressible in first-order logic can be solved in this way,
such as finding small dominating sets or small independent sets.
Experimental Design
4
We conducted four experiments, one to answer each question posed in Section 1. The first
two (Section 5) identify an "optimal" configuration for CONCUSS and test its scalability
against an existing subgraph isomorphism counting algorithm. The other two experiments
(Section 6) investigate ways in which properties of the coloring affect downstream compu-
tation to guide future theoretical research.
Michael P. O'Brien and Blair D. Sullivan
XX:5
To compare the effects of varying multiple options, we normalize two measures a, b using
a+b. This ratio will be close to 1 when a (cid:29) b and close to
the difference to sum ratio, i.e., a−b
−1 when b (cid:29) a. When comparing times, a positive ratio indicates b is faster than a.
4.1 Data
Since the theoretical advantages of the algorithms implemented in CONCUSS rely on exploit-
ing structure, it was necessary to ensure our input graphs belonged to classes of bounded
expansion. Consequently, we chose three random graph models which asymptotically al-
most surely produce classes of bounded expansion [10]: the stochastic block model (SB) [15]
and the Chung-Lu model, with households (CLH) [3] and without (CL) [7]. The first two
models both exhibit clustering commonly found in real-world networks; the CL model was
included as a baseline. For each model, we selected configuration parameters known to pro-
duce bounded expansion classes. In the CL and CLH models, we used a degree distribution
with exponential decay and household size four; for SB we used
.40
.30 .20 .10
− .50 .13 .05
− − .35 .11
− − − .45
as the probability matrix. For each model, three random instances with 1024 vertices and
average degree 6 were generated.
4.2 Hardware
The experiments were run on identical machines with four-core, 3.0 GHz Intel Xeon E5 v3
processors with a 10 MB cache and 64 GB of memory (4 × 16 GB). Each run of CONCUSS
received dedicated resources to avoid interference from other processes.
Engineering Evaluation
5
We first engineered the implementation of CONCUSS by testing different configurations
against each other before evaluating its practicality via comparison with an existing algo-
rithm.
5.1 Configuration Testing
To measure the run times of various configurations, we created one configuration file for
each combination of options described in Sections B.1 and B.4. For all possible pairings
of random instances and configuration options we ran CONCUSS three times. In each run
we counted the number of isomorphisms of P4; the particular motif did not vary because
the dynamic programming algorithm has theoretically comparable run time for all motifs
of the same size. The metrics of interest in each run were the number of colors used, the
total elapsed time, and the time specifically spent in the Color module. To compare the
performance between two implementations of a subroutine, we did a pairwise comparison
between configurations in which all other subroutines were held constant. That is to say,
to differentiate between A1 and A2, we would compare (A1, B1, C1) against (A2, B1, C1),
(A1, B2, C1) against (A2, B2, C1), etc.
Though there were minor variations across graph models, the effects of the different con-
figuration options (Figure 1) were relatively consistent. Orienting the edges using Sandpiling
XX:6
An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
Figure 1 Average difference/sum ratio in coloring size (top), Color time (middle), and De-
compose, Compute, and Combine time (bottom) between paired configuration options. Note the
differing y-axes.
(a) Chung-Lu + households
(b) Stochastic block
Figure 2 Distribution of time spent in the different submodules of the Color module. Results for
the Chung-Lu model without households were comparable to those for the model with households.
and Degeneracy led to comparable time and coloring sizes. Surprisingly, using DTFA, which
is more prudent in augmenting edges, did not outperform TFA in either time or coloring
size. We attribute this result to the fact that the average number of augmentation steps
needed to reach a p-centered coloring was smaller while using TFA (2.9 vs. 4.8), indicating
that it may be worth making "suboptimal" choices for the augmentations as long as it leads
to a p-centered coloring sooner. The Low-degree coloring prioritization generally yielded
fewer colors and a shorter run time than High-degree, while the more complicated DSATUR
prioritization was somewhere in-between. Taken as a whole, the best coloring routine used
Degeneracy orientation, TFA, and Low-degree prioritization.
The general distribution of time spent within the Color module also did not vary much
-0.100.1Coloring sizeCLHSBCL-1.001.0Color TimeHigh-DSATLow-DSATLow-HighSand-DgenTFA-DTFA-0.300.3Count TimeHigh-DSATLow-DSATLow-HighSand-DgenTFA-DTFAHigh-DSATLow-DSATLow-HighSand-DgenTFA-DTFA0.00.20.40.60.81.0Dgen,DTFA,DSATDgen,DTFA,HighDgen,DTFA,LowDgen,TFA,DSATDgen,TFA,HighDgen,TFA,LowSand,DTFA,DSATSand,DTFA,HighSand,DTFA,LowSand,TFA,DSATSand,TFA,HighSand,TFA,LowAugmentColorCheck tdOptimizeMerge0.00.20.40.60.81.0Dgen,DTFA,DSATDgen,DTFA,HighDgen,DTFA,LowDgen,TFA,DSATDgen,TFA,HighDgen,TFA,LowSand,DTFA,DSATSand,DTFA,HighSand,DTFA,LowSand,TFA,DSATSand,TFA,HighSand,TFA,LowAugmentColorCheck tdOptimizeMerge Michael P. O'Brien and Blair D. Sullivan
XX:7
Figure 3 Relationship between number of
Inclusion-exclusion.
colors and total
execution time using
with the coloring configuration (Figure 2). Computing the augmentations and coloring the
vertices took a very small fraction of the total run time; upwards of 80% of the time was
spent ensuring merging color classes kept the coloring p-centered, with the pre-merge opti-
mization consuming most of the rest of the time. Nearly all of the computation in these two
post-processing routines is in deciding whether a smaller coloring is indeed p-centered. As
such, small improvements to the efficiency of checking the treedepth of all color sets would
pay large dividends in the overall time.
Although the post-processing reduced the coloring size by at least 50% in 90% of the
colorings and the total execution time correlated positively with coloring size (Figure 3), we
cannot assess whether those optimizations were a net benefit without extrapolating beyond
our observed data. Because the time spent merging is dependent on the coloring size, we
believe there is a coloring size threshold below which the post-processing is worthwhile; it
would be useful to empirically identify this threshold in future work.
The effects of the configurations on the computation downstream from the coloring (Fig-
ure 1) were weaker and more varied. In particular, we note that Low-degree prioritization
often led to slower downstream computation, but this was offset by faster coloring times.
The only configuration option varied outside of the Color module was the method of
correcting double counting in the Combine module. We were once again surprised to see
that Hybrid, the "intelligent" method of circumventing enumerating many additional color
classes, was not successful at improving the run time over the naïve Inclusion-exclusion.
To the contrary, Hybrid increased the total time of the Decompose, Compute, and Com-
bine modules by at least a factor of two, getting up to a factor of ten when the number of
colors was very large. We ultimately concluded the preferred configuration is Degeneracy,
TFA, Low-degree, and Inclusion-exclusion.
5.2 Comparison with NXVF2
After determining the behavior of various configurations of CONCUSS, we wanted to assess
its performance compared to algorithmic approaches that do not exploit bounded expansion
structure. Recognizing that CONCUSS is a proof-of-concept of an algorithmic pipeline for
classes of bounded expansion, we were primarily interested in establishing it has comparable
run times to other algorithms and good scaling behavior; in future work, we hope to incor-
porate additional theoretical advances into CONCUSS, i.e., those identified in Section 6, to
8090100110120130Colors25005000750010000125001500017500Time (s)SBCLHCL XX:8
An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
increase its performance. To make a comparison in the same programming language, we
selected the NetworkX [13] implementation of the popular VF2 algorithm [9], which we will
refer to as NXVF2. Although there exists an updated version of this algorithm known as
VF3 [6], the authors' publicly available implementation [1] only accepts graphs with 216 or
fewer vertices, which is insufficient for the data sets described below.
5.2.1 Additional Data
Because of the large p-centered coloring sizes of the Chung-Lu and stochastic block model
graphs generated for the previous experiment, CONCUSS requires around 45 minutes to
count in graphs where NXVF2 terminates in 1 second. We hypothesized, however, that in
large graphs with small coloring sizes and many isomorphisms of H into G, CONCUSS would
outperform NXVF2. We included the restriction on number of isomorphisms because the
run time of CONCUSS does not depend on the count, while NXVF2 requires more time for
each additional isomorphism. To test this hypothesis we generated graphs that met the
criteria with the following procedure. First, we generated a complete binary tree of depth
d. Then, we selected a tree vertex uniformly at random and attached to it the endpoint of
a P'. We continued randomly selecting tree vertices with replacement and adding a P' for
a total of s · 2d times. That is, we made sure that each tree vertex had an average of s P's
as neighbors. We denote this graph Td,s,'.In this way, we could effectively vary the size of
the graph and the number of isomorphisms of small stars and paths.
'
s
Figure 4 Td,s,'.
d
5.2.2 Results
We observed (Figure 5) that our hypothesis holds when counting stars in Td,s,' for sufficiently
large s. In particular, CONCUSS has more than a fourteen-fold speed advantage over NXVF2
in counting S5s in T12,16,1. We believe the difference in time would diverge even more for
counting S5s in T14,16,1, but NXVF2 already takes nearly 80 hours in T12,16,1.
It is also important to note that when ' = 1, the number of paths does not increase in
the same way that the number of stars increases. This is because every path of length at
least four must contain multiple tree vertices, of which each tree vertex is adjacent to at
most three. Moreover, Pn has exactly two automorphisms ("forwards" and "backwards") for
every n, while Sn has (n−1)! automorphisms (any ordering of the leaves). We observed that
there were insufficiently many P4s and P5s in Td,s,1 to see the same performance benefits,
and NXVF2 was consistently faster than CONCUSS.
As predicted, CONCUSS can outperform NXVF2 when counting in Td,s,4, which has
many more P4s and P5s (Figure 9). Similar to the results shown in Figure 5 (Appendix C),
CONCUSS counts P3s faster in all graph sizes, while the advantage in counting P5s is only
realized in the largest graphs.
Michael P. O'Brien and Blair D. Sullivan
XX:9
Figure 5 Average difference/sum ratio between CONCUSS and NXVF2 on Td,s,1 as a function of
s, the average number of P1s per tree vertex. Each small plot shows a fixed motif and value of d.
Negative ratios indicate CONCUSS is faster.
6
Bottleneck Identification
Having identified configurations and graph instances in which CONCUSS outperforms NXVF2,
we proceeded to identify bottlenecks in CONCUSS that ought to receive further theoretical
consideration. We focused on how changing properties of the output of the Color module
would influence performance in the other three modules.
6.1 Color Class Distribution
Because a graph may admit many distinct p-centered colorings of small size, it is important
to know whether two colorings of the same size lead to observable differences in the down-
stream computation. Though the effects the coloring has on the rest of the pipeline almost
certainly depend on complex and subtle interactions, we hypothesized that the distribution
of sizes of color classes is an important factor. The intuition behind this hypothesis is that
different distributions in sizes will lead to different "shapes" of treedepth decompositions,
i.e., one large, wide decomposition vs. many small, thin decompositions. To test it, we
created colorings of equal size that have unequal distributions of color class sizes.
The stochastic block and Chung-Lu with households graphs described previously were
used again as inputs for this experiment; we excluded the Chung-Lu graphs without house-
holds due to their long run times. For each graph instance, we found the largest coloring
from the configuration experiment, then iteratively split in half the existing color classes
(Figure 6) in colorings generated by the other configurations until all colorings had the same
1.00.50.00.51.0d=10d=12d=141.00.50.00.51.014s=161.00.50.00.51.01/414s=161/161/41s=4s2d210212214216 XX:10 An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
min
max
med
Figure 6 Creating a new color class (black) using the three splitting heuristics (min, med, max).
size. To vary the distributions of color class sizes, we used three splitting heuristics: max, med,
and min, which target the maximum-, median-, and minimum-sized color classes2. When
splitting color classes, we chose a random subset of vertices to move into the new color class;
to account for this randomness, we repeated each splitting heuristic three times for each
coloring. We ran the Decompose, Compute, and Combine modules of CONCUSS three
times for each graph instance and associated new coloring, using Inclusion-Exclusion in
the Combine module.
We observed (Table 1) that max led to the fastest performance, while med and min had
very similar run times. The dynamic programming algorithm counts isomorphisms from
the leaves of the treedepth decomposition upwards using two operations: forgetting, which
moves from children to their parent, and joining which combines results from "siblings"
(vertices with the same parent). Since siblings often belong to the same color class, ensuring
that no color class has too many vertices limits the number of children, which in turn limits
the number of joins.
In Table 1, we report that when using max the average number of
joins was 33% lower than when splitting the median or minimum-sized classes. Thus future
coloring algorithms should attempt to make the sizes of the color classes more uniform.
Model Method Avg time (s)
856
+19%
+20%
1327
+21%
+21%
SB
SB
SB
CLH
CLH
CLH
max
med
min
max
med
min
Joins
5.12 × 105
+48%
+49%
7.63 × 105
+48%
+50%
Forgets
2.00 × 108
+13%
+13%
3.01 × 108
+4%
+5%
Table 1 Average run time and dynamic programming operations used for each color splitting
heuristic. The statistics for med and min are reported as percent increases over max.
6.2 Color Set Treedepth
The dynamic programming (DP) in the Compute module counts isomorphisms of a sub-
graph of size h in an arbitrary treedepth decomposition of depth t. While this algorithm
runs in linear time with respect to the size of the graph, it takes O(th) time with respect to
t. One potential direction for further theoretical improvements to the bounded expansion
pipeline is to find alternatives to p-centered colorings in which subgraphs with at most i < p
colors have treedepth that is small but greater than i. Ideally, relaxing treedepth require-
ment would result in a significant reduction in the size of the coloring without dramatically
increasing the time spent in DP in the Compute module. Though we cannot measure the
2 Color classes with exactly one vertex were ignored for calculating the minimum and median sizes.
Michael P. O'Brien and Blair D. Sullivan
XX:11
size of a hypothetical coloring, we can evaluate how the time per DP operation increases as
we count the same subgraph in larger treedepth decompositions. To do this, we computed
a 6-centered coloring, enumerated sets of 3, 4, and 5 colors, and counted the number of
isomorphisms of P3 in each set.
The run time of the DP algorithm is only dependent on the treedepth of the decomposi-
tions insofar as it increases the number of labelings3 of the ancestors of each vertex. There
are t3 + 3t2 + t such labelings for a P3, each of which requires a constant number of DP
operations, so the amount of time spent counting in one color set should grow proportionally
to this function.
Figure 7 Observed vs. predicted average execution time per operation in the stochastic block
graphs. Results for the graphs from the CL models are similar and are omitted for space. We only
include those operations incurred in counting in color sets of exactly t colors.
Model
t
CL
CL
CL
CLH
CLH
CLH
SB
SB
SB
3
4
5
3
4
5
3
4
5
leaf
Time per
7.5 × 10−7
7.7 × 10−7
7.9 × 10−7
8.0 × 10−7
8.0 × 10−7
8.1 × 10−7
8.1 × 10−7
7.9 × 10−7
8.0 × 10−7
join
Time per
7.3 × 10−7
8.1 × 10−7
9.0 × 10−7
7.4 × 10−7
8.1 × 10−7
9.3 × 10−7
7.3 × 10−7
7.9 × 10−7
8.5 × 10−7
forget
Time per
1.4 × 10−6
1.7 × 10−6
2.0 × 10−6
1.4 × 10−6
1.7 × 10−6
2.1 × 10−6
1.4 × 10−6
1.5 × 10−6
1.8 × 10−6
Avg.
depth
2.0
2.4
2.8
2.0
2.4
2.8
1.9
2.3
2.7
Table 2 Average execution time per operation and average depth of vertices in the treedepth
decomposition. Rows for treedepth t only include those operations incurred in counting in color
sets of exactly t colors.
To compare our observations to the predicted growth of t3 +3t2 + t, we assumed that the
average time per operation for t = 3 aligned correctly with this theoretical time and extrap-
olated outwards for t ∈ {4, 5}. As reported in Figure 7, the observed times grow significantly
slower than predicted for all operations. The time spent on each operation is dependent on
the depth of the vertex in the treedepth decomposition, but the average depth of vertices
does not grow at a one-to-one ratio with the treedepth (Table 2). This means that the ad-
ditional color classes add vertices at many depths of the decomposition, rather that simply
3 These are the k-patterns mentioned in Appendix B.3; see [10].
345Treedepth12345Time (s)1e6Leaf (observed)Leaf (predicted)Join (observed)Join (predicted)Forget (observed)Forget (predicted) XX:12 An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
increasing the depth of every branch uniformly, and thus we seldom "pay" for the cost of the
larger depths. Consequently, we conclude that colorings producing deeper decompositions
with fewer colors ought to be a viable strategy for improving the overall speed of CONCUSS.
Conclusion
7
We demonstrated that exploiting bounded expansion structure is a promising methodology
for scaling subgraph isomorphism counting to larger classes of sparse graphs. We identified
through testing various configurations of CONCUSS that orienting with a degeneracy order-
ing, augmenting with TFA, coloring low-degree vertices first, and combining counts using
the inclusion-exclusion principle led to fast run times and small colorings. On sufficiently
sparse graphs with high isomorphism counts, CONCUSS was able to significantly outperform
NXVF2. In order to get the bounded expansion pipeline to outperform other methods in a
broader set of graphs, it is important to understand in future work whether the large color-
ing sizes of the random graph models were an artifact of the heuristics used in generating
the p-centered colorings, or whether their asymptotically bounded values are actually large.
Moreover, reimplementing CONCUSS in a lower-level language like C++ would allow us to
make meaningful comparisons to other C++ implementations, e.g., the RI algorithm in [17].
Our experiments also identified three areas for future theoretical research. The first is to
reduce the time needed to verify a coloring is p-centered, which would yield large dividends
in the Color module. The second is to design p-centered coloring algorithms that balance
the sizes of the color classes, consequently reducing the number of join operations. Finally,
our results suggest using colorings that trade fewer colors for larger treedepth per color set
could be faster than using p-centered colorings.
Acknowledgments
This work was supported in part by the DARPA GRAPHS Program and the Gordon & Betty
Moore Foundation's Data-Driven Discovery Initiative through Grants SPAWAR-N66001-14-1-4063
and GBMF4560 to Blair D. Sullivan.
References
1 VF3. http://mivia.unisa.it/datasets-request/.
2
P. Bak et al. Self-organized criticality: An explanation of the 1/f noise. PRL, 59(4):381,
1987.
F. Ball et al. Threshold behaviour and final outcome of an epidemic on a random network
with household structure. Advances in Applied Probability, 41(3):765–796, 2009.
3
4 V. Bonnici, R. Giugno, A. Pulvirenti, D. Shasha, and A. Ferro. A subgraph isomorphism
algorithm and its application to biochemical data. BMC bioinformatics, 14(7):S13, 2013.
5 D. Brélaz. New methods to color the vertices of a graph. CACM, 22(4):251–256, 1979.
6 V. Carletti, P. Foggia, A. Saggese, and M. Vento. Introducing vf3: A new algorithm for
In International Workshop on Graph-Based Representations in
subgraph isomorphism.
Pattern Recognition, pages 128–139. Springer, 2017.
F. Chung and L. Lu. The average distances in random graphs with given expected degrees.
PNAS, 99(25):15879–15882, 2002.
L. P. Cordella et al. Performance evaluation of the vf graph matching algorithm. In Image
Analysis and Processing, pages 1172–1177. IEEE, 1999.
L. P. Cordella et al. A (sub) graph isomorphism algorithm for matching large graphs.
Pattern Analysis and Machine Intelligence, 26(10):1367–1372, 2004.
7
8
9
Michael P. O'Brien and Blair D. Sullivan
XX:13
10
11
12
E. D. Demaine et al. Structural sparsity of complex networks: Random graph models and
linear algorithms. CoRR, abs/1406.2587, 2015.
Z. Dvořák et al. Algorithms for classes of graphs with bounded expansion.
Theoretic Concepts in Computer Science, pages 17–32. Springer, 2009.
Z. Dvořák et al. Testing first-order properties for subclasses of sparse graphs. JACM,
60(5):36:1–36:24, October 2013.
In Graph-
13 A. A. Hagberg et al. Exploring network structure, dynamics, and function using NetworkX.
In SciPy2008, pages 11–15, Pasadena, CA USA, August 2008.
14 H. He and A. K. Singh. Graphs-at-a-time: query language and access methods for graph
databases. In SIGMOD, pages 405–418. ACM, 2008.
P. W. Holland et al. Stochastic blockmodels: First steps. Social networks, 5(2):109–137,
1983.
16 M. Jerrum and K. Meeks. The parameterised complexity of counting connected subgraphs.
CoRR, abs/1308.1575, 2013.
Jure Leskovec and Rok Sosič. Snap: A general-purpose network analysis and graph-mining
library. ACM Transactions on Intelligent Systems and Technology (TIST), 8(1):1, 2016.
15
17
18 D. W. Matula and L. L. Beck. Smallest-last ordering and clustering and graph coloring
algorithms. JACM, 30(3):417–427, 1983.
19 T. Milenkovic et al. Optimal network alignment with graphlet degree vectors. Cancer
informatics, 9:121, 2010.
20 T. Milenkovic and N. Pržulj. Uncovering biological network function via graphlet degree
signatures. Cancer informatics, 6:257, 2008.
J. Nešetřil and P. Ossona de Mendez. Grad and classes with bounded expansion i. decom-
positions. European Journal of Combinatorics, 29(3):760 – 776, 2008.
J. Nešetřil and P. Ossona de Mendez. Grad and classes with bounded expansion ii. algo-
rithmic aspects. European Journal of Combinatorics, 29(3):777 – 791, 2008.
J. Nešetřil and P. Ossona de Mendez. Grad and classes with bounded expansion iii. re-
stricted graph homomorphism dualities. European Journal of Combinatorics, 29(4):1012 –
1024, 2008. Homomorphisms: Structure and Highlights.
21
22
23
26
27
29
30
31
24 M. P. O'Brien et al. CONCUSS, v2.0. http://dx.doi.org/10.5281/zenodo.30281, June 2016.
25 N. Pržulj. Biological network comparison using graphlet degree distribution. Bioinformat-
ics, 23(2):e177–e183, 2007.
F. Reidl. Structural sparseness and complex networks. Dr., Aachen, Techn. Hochsch.,
Aachen, 2015. Aachen, Techn. Hochsch., Diss., 2015.
F. Reidl et al. Characterising bounded expansion by neighbourhood complexity. CoRR,
abs/1603.09532, 2016.
28 H. Shang et al. Taming verification hardness: an efficient algorithm for testing subgraph
isomorphism. PVLDB, 1(1):364–375, 2008.
S. S. Shen-Orr et al. Network motifs in the transcriptional regulation network of escherichia
coli. Nature genetics, 31(1):64–68, 2002.
Z. Sun et al. Efficient subgraph matching on billion node graphs. PVLDB, 5(9):788–799,
2012.
J. R. Ullmann. An algorithm for subgraph isomorphism. JACM, 23(1):31–42, 1976.
XX:14 An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
A
Computing Treedepth Decompositions
Algorithm 1 Treedepth decomposition implied by centered coloring
φ ← centered coloring of G
T ← treedepth decomposition of G
if G = 0 then
return
end if
v ← center(φG)
root(T) ← v
for all components C ∈ G\{v} do
T 0 ← treedepth decomposition of C implied by φC
parent(root(T 0)) ← v
end for
Algorithmic Description of CONCUSS
B
B.1 Color
Algorithm 2 Sketch of p-centered coloring
G0 ← G
repeat
Orient the edges of G0 to make it a directed acyclic graph of low indegree
Augment G0 with additional edges based on their orientation
Color the vertices greedily
Check whether each set of vertices with p − 1 colors has low treedepth in G
until the coloring is p-centered
The workflow in the Color module is outlined in Algorithm 2. The objective is to
produce a p-centered coloring by greedily coloring the vertices of the graph. Because a
simple greedy coloring may not immediately be p-centered, additional edges are added to
the graph to impose further constraints. After a bounded number of iterations of the loop
in Algorithm 2, a greedy coloring is guaranteed to be p-centered [21, 23, 26]. Since each
substep runs in linear time with respect to G, the entire Color module also runs in linear
time.
Orient: The later stages of the Color module operate on directed acyclic graphs. In
essence, directed edges capture the ancestor relationships in treedepth decompositions of
small sets of colors by having edges point from an ancestor to its descendants. In this repre-
sentation, vertices with large indegree have many ancestors, indicating subgraphs with large
treedepth. A low-indegree orientation can be found either with a Degeneracy ordering [18]
or by Sandpiling [2].
Augment: After the edges are oriented, additional directed edges are augmented to the
graph to impose more constraints. The placement of the augmenting edges is dependent
on the direction of edges in the orientation. Two vertices a, b are said to be transitive if
there is some vertex c such that ac and cb are both directed edges in the graph, while a, b
are said to be fraternal if there is a vertex c such that ac and bc are both directed edges.
Michael P. O'Brien and Blair D. Sullivan
XX:15
The transitive-fraternal augmentation (TFA) algorithm [22] adds the edge ab for every pair
of transitive vertices a, b and either the edge xy or yx for every pair of fraternal vertices
x, y. The orientation of edges between fraternal vertices is chosen to preserve acyclicity and
minimize indegree. Distance-truncated transitive fraternal augmentation (DTFA) [26] works
similarly, but only requires that some transitive and fraternal edges are added. If ac was
added in the ith iteration4 of the loop in Algorithm 2 and cb was added in the jth, the edge
ab need not be added until the (i + j)th iteration.
Color: The acyclic orientation of the edges defines an ordering over which to greedily
color the vertices. To minimize the number of colors, CONCUSS uses of one of three different
heuristics: prioritizing high-degree vertices (High-degree), prioritizing low-degree vertices
(Low-degree), or prioritizing vertices with many colors already represented among their
neighbors (DSATUR) [5].
Check: Finally, we check whether the coloring is p-centered by testing whether each
connected subgraph that uses p − 1 or fewer colors has a unique color. This can be done
by keeping track of components and their respective color multiplicities using union-find
structures; when two components merge, we ensure that that there is still a unique color.
We perform this check at every iteration in order to prevent the algorithm from adding any
more colors than necessary.
Other Optimizations: We include additional pre- and post-processing steps to obtain
smaller colorings. Vertices of high degree can cause many additional edges to be augmented
at each iteration of the loop, which in turn causes more colors to be used. For this reason,
removing all high-degree vertices, e.g., degree at least 4√
n, finding a coloring of the smaller
graph, and then giving each removed vertex a new unique color can reduce the total number
of colors. Likewise, in any component of size larger than two, vertices of degree one can
be removed and all given the same color in post-processing. After obtaining a p-centered
coloring, we can also potentially reduce the number of colors by randomly adding transitive
and fraternal edges, computing a new greedy coloring, and seeing if this greedy coloring is
p-centered. The final post-processing step is to check whether pairs of color classes can be
merged without violating the p-centered property.
B.2 Decompose
After the coloring is computed, CONCUSS finds the components induced by each set of H
colors. To save computation, CONCUSS processes overlapping color sets sequentially in a
depth-first search manner. For example, if H = 4 and the coloring has 9 colors, the sets
would be processed as: . . . , {1, 2, 3}, {1, 2, 3, 4}, {1, 2, 3, 5}, . . . , {1, 2, 3, 9}, {1, 2, 4},
{1, 2, 4, 5}, etc. When processing color set {1, 2, 3, 4}, the components from {1, 2, 3} are
saved to be reused for {1, 2, 3, 5}.
Given a component induced by a particular color set, a treedepth decomposition is cre-
ated using Algorithm 1. In order to save memory, these treedepth decompositions are fed
into the subsequent Count and Compute stages before a new color set is processed.
B.3 Compute
The Compute module uses the dynamic programming algorithm of Demaine et al. [10] to
count the isomorphisms of H in a treedepth decomposition.
It has exponential time de-
pendence on the small treedepth, but runs in linear time with respect to G. Briefly, this
4 Edges present originally in G are considered to have been added in the 1st iteration.
XX:16 An Experimental Evaluation of a Bounded Expansion Algorithmic Pipeline
algorithm exploits the fact that no edges in the treedepth decomposition join vertices in
different branches. Therefore, partial pieces of H in one subtree have limited ways of inter-
acting with those in other subtrees. By tracking which pieces occur in each subtree, these
counts of partial pieces can be combined to count all isomorphisms in the decomposition.
Integral to the dynamic programming are k-patterns, also called boundaried graphs.
These (sub)graphs have a subset of vertices labeled uniquely with integers on the interval
[1, k], which in turn determine how partial pieces of H can "glue" together. We store the k-
pattern labelings as bitvectors to enable fast checking of whether two patterns are compatible
to be "glued", i.e., their labelings are consistent with one another.
B.4 Combine
The goal in the Combine module is to sum the counts from each treedepth decomposition
to get the count for the whole graph. However, because the color sets contain overlapping
subgraphs, a simple summation of counts may count some isomorphisms more than once.
For example, if H = P4, the subgraph in Figure 8 will be counted in the color sets {blue,
orange, purple, yellow}, {blue, orange, purple, green}, etc. There are multiple ways to rectify
Figure 8 The highlighted P4 only uses three colors and thus will be counted in multiple color
sets of size four.
this issue. Inclusion-exclusion simply enumerates all color sets with fewer than H
colors, counts the isomorphisms in those subgraphs, and then applies the inclusion-exclusion
principle to correct the counts appropriately. We can also incorporate the combinations into
the dynamic programming itself by ignoring counts of color sets that have already been seen
(Hybrid).
Michael P. O'Brien and Blair D. Sullivan
XX:17
C
Additional Figures
Figure 9 Average difference/sum ratio between CONCUSS and NXVF2 on Td,s,4 as a function of
s, the average number of P4s per tree vertex. Each small plot shows a fixed motif and value of d.
Negative ratios indicate CONCUSS is faster.
1.00.50.00.51.0d=10d=12d=141.00.50.00.51.014s=161.00.50.00.51.01/414s=161/161/41s=4s2d210212214216 |
1606.04763 | 3 | 1606 | 2018-09-25T11:52:39 | A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching | [
"cs.DS"
] | The pattern matching problem with swaps is to find all occurrences of a pattern in a text while allowing the pattern to swap adjacent symbols. The goal is to design fast matching algorithm that takes advantage of the bit parallelism of bitwise machine instructions and has only streaming access to the input. We introduce a new approach to solve this problem based on the graph theoretic model and compare its performance to previously known algorithms. We also show that an approach using deterministic finite automata cannot achieve similarly efficient algorithms. Furthermore, we describe a fatal flaw in some of the previously published algorithms based on the same model. Finally, we provide experimental evaluation of our algorithm on real-world data. | cs.DS | cs |
A Simple Streaming Bit-parallel Algorithm
for Swap Pattern Matching⋆
V´aclav Blazej⋆⋆, Ondrej Such´y⋆ ⋆ ⋆, and Tom´as Valla†
Faculty of Information Technology, Czech Technical University in Prague,
Prague, Czech Republic
Abstract. The pattern matching problem with swaps is to find all occur-
rences of a pattern in a text while allowing the pattern to swap adjacent
symbols. The goal is to design fast matching algorithm that takes advan-
tage of the bit parallelism of bitwise machine instructions and has only
streaming access to the input. We introduce a new approach to solve
this problem based on the graph theoretic model and compare its perfor-
mance to previously known algorithms. We also show that an approach
using deterministic finite automata cannot achieve similarly efficient al-
gorithms. Furthermore, we describe a fatal flaw in some of the previously
published algorithms based on the same model. Finally, we provide ex-
perimental evaluation of our algorithm on real-world data.
1
Introduction
In the Pattern Matching problem with Swaps (Swap Matching, for short), the
goal is to find all occurrences of any swapped version of a pattern P in a text T ,
where P and T are strings of length p and t over an alphabet Σ, respectively. By
the swapped version of a pattern P we mean a string of symbols created from P
by swapping adjacent symbols while ensuring that each symbol is swapped at
most once (see Section 2 for formal definitions). The solution of Swap Match-
ing is a set of indices which represent where occurrences swapped version of P
in T begin. Swap Matching is intensively studied due to its use in practical ap-
plications such as text and music retrieval, data mining, network security and
biological computing [7].
The swap of two consecutive symbols is one of the most typical typing errors.
It also represent a simpler version of swaps that appear in nature. In particular,
⋆ An extended abstract of this work appeared in the Proceedings of the 7th Interna-
tional Conference on Mathematical Aspects of Computer and Information Sciences,
MACIS 2017 [8].
⋆⋆ Supported
OP
VVV
by
the
project
CZ.02.1.01/0.0/0.0/16 019/0000765 "Research Center for Informatics" and by
the SGS CTU project SGS17/209/OHK3/3T/18.
funded
MEYS
⋆ ⋆ ⋆ Supported by grant 17-20065S of the Czech Science Foundation.
† Supported by the Centre of Excellence -- Inst. for Theor. Comp. Sci. 79 (project
P202/12/G061 of the Czech Science Foundation.)
2
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
the phenomenon of swaps occurs in gene mutations and duplications such as in
the region of human chromosome 5 that is implicated in the disease called spinal
muscular Atrophy, a common recessive form of muscular dystrophy [18]. While
the biological swaps occur at a gene level and have several additional constraints
and characteristics, which make the problem much more difficult, they do serve as
a convincing pointer to the theoretical study of swaps as a natural edit operation
for the approximation metric [2]. Indeed Lowrance and Wagner [21] suggested
to add the swap operation when considering the edit distance of two strings.
1
Swap Matching was introduced in 1995 as an open problem in non-standard
string matching [20]. The first result was reported by Amir et al. [2] in 1997,
3 log p)-time solution for alphabets of size 2, while also
who provided an O(tp
showing that alphabets of size exceeding 2 can be reduced to size 2 with a little
overhead. Amir et al. [5] came up with solution with O(t log2 p) time complexity
for some very restrictive cases. Several years later Amir et al. [3] showed that
Swap Matching can be solved by an algorithm for the overlap matching achieving
the running time of O(t log p log Σ). This algorithm as well as all the previous
ones is based on fast Fourier transformation (FFT).
In 2008 Iliopoulos and Rahman [17] introduced a new graph theoretic ap-
proach to model the Swap Matching problem and came up with the first efficient
solution to Swap Matching without using FFT (we show it to be incorrect). Their
algorithm based on bit parallelism runs in O((t + p) log p) time if the pattern
length is similar to the word-size of the target machine. One year later Cantone
and Faro [10] presented a dynamic programming algorithm named Cross Sam-
pling solving Swap Matching in O(t) time and O(Σ) space, assuming that the
pattern length is similar to the word-size in the target machine. In the same
year Campanelli et al. [9] enhanced the Cross Sampling algorithm using notions
from Backward directed acyclic word graph matching algorithm and named the
new algorithm Backward Cross Sampling. This algorithm also assumes short
pattern length. Although Backward Cross Sampling has O(Σ) space and O(tp)
time complexity, which is worse than that of Cross Sampling, it improves the
real-world performance.
In 2013 Faro [14] presented a new model to solve Swap Matching using re-
active automata and also presented a new algorithm with O(t) time complexity
assuming short patterns. The same year Chedid [12] reformulated the dynamic
programming solution by Cantone and Faro [10] which results in more intuitive
algorithms. In 2014 a minor improvement by Fredriksson and Giaquinta [15] ap-
peared, yielding slightly (at most factor Σ) better asymptotic time complexity
(and also slightly worse space complexity) for special cases of patterns. The same
year Ahmed et al. [1] took ideas of the algorithm by Iliopoulos and Rahman [17]
and devised two algorithms named Smalgo-I and Smalgo-II which both run
in O(t) for short patterns, but bear the same error as the original algorithm.
Another remarkable effort related to Swap Matching is to actually count the
number of swaps needed to match the pattern at the location [6]. This is more
often studied with an extra operation of character change allowed [4,13,19].
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
3
Our Contribution. We design a simple algorithm which solves the Swap
Matching problem. The goal is to design a streaming algorithm, which is given
one symbol per each execution step until the end-of-input arrives, and thus does
not need access to the whole input. This algorithm has O(⌈ p
w ⌉(Σ + t) + p)
time and O(⌈ p
w ⌉Σ) space complexity where w is the word-size of the machine.
We would like to stress that our solution, as based on the graph theoretic ap-
proach, does not use FFT. Therefore, it yields a much simpler non-recursive
algorithm allowing bit parallelism and is not suffering from the disadvantages of
the convolution-based methods. While our algorithm matches the best asymp-
totic complexity bounds of the previous results [10,15] (up to a Σ factor), we
believe that its strength lies in the applications where the alphabet is small and
the pattern length is at most the word-size, as it can be implemented using only
7 + Σ CPU registers and few machine instructions. This makes it practical for
tasks like DNA sequences scanning. Also, as far as we know, our algorithm is
currently the only known streaming algorithm for the swap matching problem.
We continue by proving that any deterministic finite automaton that solves
Swap Matching has number of states exponential in the length of the pattern.
We also describe the Smalgo (swap matching algorithm) by Iliopoulos and
Rahman [17] in detail. Unfortunately, we have discovered that Smalgo and
derived algorithms contain a flaw which cause false positives to appear. We
have prepared implementations of Smalgo-I, Cross Sampling, Backward Cross
Sampling and our own algorithm, measured the running times and the rate of
false positives for the Smalgo-I algorithm. All of the sources are available for
download.1
This paper is organized as follows. First we introduce all the basic definitions,
and also recall the graph theoretic model introduced in [17] and its use for
matching in Section 2. In Section 3 we show our algorithm for Swap Matching
problem and follow it in Section 4 with the proof that Swap Matching cannot be
solved efficiently by deterministic finite automata. Then we describe the Smalgo
algorithms in detail in Section 5 and finish with the experimental evaluation of
the algorithms in Section 6.
2 Basic Definitions and the Graph Theoretic Model
In this section we state the basic definitions, present the graph theoretic model
and show a basic algorithm that solves Swap Matching using the model.
2.1 Notations and Basic Definitions
We use the word-RAM as our computational model. That means we have access
to memory cells of fixed capacity w (e.g., 64 bits). A standard set of arithmetic
and bitwise instructions include And (&), Or (), Left bitwise-shift (LShift or
≪ 1) and Right bitwise-shift (RShift or ≫ 1). Each of the standard operations on
1 http://users.fit.cvut.cz/blazeva1/gsm.html
4
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
1
−1
✚
0
1
a
b
2
a
b
c
3
b
c
b
4
c
b
b
5
b
b
a
6
b
a
c
7
a
c
✚
Fig. 1. P -graph PP for the pattern P = abcbbac
words takes single unit of time. In order to compare to other existing algorithms,
which are not streaming, we define the access to the input in a less restrictive
way -- the input is read from a read-only part of memory and the output is
written to a write-only part of memory. However, it will be easy to observe that
our algorithm accesses the input sequentially. We do not include the input and
the output into the space complexity analysis.
A string S over an alphabet Σ is a finite sequence of symbols from Σ and S
is its length. By Si we mean the i-th symbol of S and we define a substring
S[i,j] = SiSi+1 . . . Sj for 1 ≤ i ≤ j ≤ S, and prefix S[1,i] for 1 ≤ i ≤ S.
String P prefix matches string T k symbols on position i if P[1,k] = T[i,i+k−1].
Next we formally introduce a swapped version of a string.
Definition 1 (Campanelli et al. [9]). A swap permutation for S is a permu-
tation π : {1, . . . , n} → {1, . . . , n}, where n = S, such that:
(i) if π(i) = j then π(j) = i (symbols at positions i and j are swapped),
(ii) for all i, π(i) ∈ {i − 1, i, i + 1} (only adjacent symbols are swapped),
(iii) if π(i) 6= i then Sπ(i) 6= Si (identical symbols are not swapped).
For a string S a swapped version π(S) is a string π(S) = Sπ(1)Sπ(2) . . . Sπ(n)
where π is a swap permutation for S.
Now we formalize the version of matching we are interested in.
Definition 2. Given a text T = T1T2 . . . Tt and a pattern P = P1P2 . . . Pp,
the pattern P is said to swap match T at location i if there exists a swapped
version π(P ) of P that matches T at location i, i.e., π(P ) = T[i,i+p−1].
2.2 A Graph Theoretic Model
The algorithms in this paper are based on a model introduced by Iliopoulos and
Rahman [17]. In this section we briefly describe this model.
For a pattern P of length p we construct a labeled graph PP = (V, E, σ) with
vertices V , edges E, and a vertex labeling function σ : V → Σ (see Fig. 1 for
an example). Let V = V ′ \ {m−1,1, m1,p} where V ′ = {mr,c r ∈ {−1, 0, 1}, c ∈
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
5
Algorithm 1 The basic matching algorithm (BMA)
Input: Labeled directed acyclic graph G = (V, E, σ), set Q0 ⊆ V of starting
vertices, set F ⊆ V of accepting vertices, text T , and position k.
1 := Q0.
1: Let D′
2: for i = 1, 2, 3, . . . , p do
3:
4:
5:
6:
i, σ(x) = Tk+i−1}.
Let Di := {x x ∈ D′
if Di = ∅ then finish.
if Di ∩ F 6= ∅ then we have found a match and finish.
Define the next iteration set D′
D′
i+1 := {d ∈ V (PP ) (v, d) ∈ E(PP ) for some v ∈ Di}.
i+1 as vertices which are successors of Di, i.e.,
{1, 2, . . . , p}}. For mr,c ∈ V we set σ(mr,c) = Pr+c. Each vertex mr,c is identified
with an element of a 3 × p grid. We set E′ := E′
p−1, where
E′
j := {(mk,j, mi,j+1) k ∈ {−1, 0}, i ∈ {0, 1}} ∪ {(m1,j, m−1,j+1)}, and let
E = E′ ∩ V × V . We call PP the P -graph. Note that PP is directed acyclic
graph, V (PP ) = 3p − 2, and E(PP ) = 5(p − 1) − 4.
2 ∪ · · · ∪ E′
1 ∪ E′
The idea behind the construction of PP is as follows. We create vertices V ′
and edges E′ which represent every swap pattern without unnecessary restric-
tions (equal symbols can be swapped). We remove vertices m−1,1 and m1,p which
represent symbols from invalid indices 0 and p + 1.
The P -graph now represents all possible swap permutations of the pattern P
in the following sense. Vertices m0,j represent ends of prefixes of swapped version
of the pattern which end by a non-swapped symbol. Possible swap of symbols Pj
and Pj+1 is represented by vertices m1,j and m−1,j+1. Edges represent symbols
which can be consecutive. Each path from column 1 to column p represents
a swap pattern and each swap pattern is represented this way.
Definition 3. For a given Σ-labeled directed acyclic graph G = (V, E, σ) ver-
tices s, e ∈ V and a directed path f = v1, v2, . . . , vk from v1 = s to vk = e, we
call S = σ(f ) = σ(v1)σ(v2) . . . σ(vk) ∈ Σ∗ a path string of f .
2.3 Using Graph Theoretic Model for Matching
In this section we describe an algorithm called Basic Matching Algorithm (BMA)
which can determine whether there is a match of pattern P in text T on a
position k using any graph model which satisfies the following conditions.
-- It is a directed acyclic graph,
-- V = V1 ⊎ V2 ⊎ · · · ⊎ Vp (we can divide vertices to columns),
-- E ⊆ {(u, w) u ∈ Vi, w ∈ Vi+1, 1 ≤ i < p} (edges lead to next column).
Let Q0 = V1 be the starting vertices and F = Vp be the accepting vertices.
BMA is designed to run on any graph which satisfies these conditions. Since
P -graph satisfies these assumptions we can use BMA for PP .
The algorithm runs as follows (see also Algorithm 1). We initialize the algo-
1 now holds information about vertices
1 := Q0 (Step 1). D′
rithm by setting D′
6
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Algorithm 2 BMA in terms of prefix match signals
1: Let I 0(v) := 1 for each v ∈ Q0 and I 0(v) := 0 for each v /∈ Q0.
2: for i = 0, 1, 2, 3, . . . , p − 1 do
3:
4:
5:
6:
Filter signals by a symbol Tk+i.
if I i(v) = 0 for every v ∈ PP then finish.
if I i(v) = 1 for any v ∈ F then we have found a match and finish.
Propagate signals along the edges.
which are the end of some path f starting in Q0 for which σ(f ) possibly prefix
matches 1 symbol of T[k,k+p−1]. To make sure that the path f represents a pre-
fix match we need to check whether the label of the last vertex of the path f
matches the symbol Tk (Step 3). If no prefix match is left we did not find a
match (Step 4). If some prefix match is left we need to check whether we already
have a complete match (Step 5). If the algorithm did not stop it means that we
have some prefix match but it is not a complete match yet. Therefore we can
try to extend this prefix match by one symbol (Step 6) and check whether it is
a valid prefix match (Step 3). Since we extend the matched prefix in each step,
we repeat these steps until the prefix match is as long as the pattern (Step 2).
Having vertices in sets is not handy for computing so we present another way
to describe this algorithm. We use their characteristic vectors instead.
Definition 4. A Boolean labeling function I : V → {0, 1} of vertices of PP is
called a prefix match signal.
The algorithm can be easily divided into iterations according to the value
of i in Step 2. We denote the value of the prefix match signal in j-th iteration
as I j and we define the following operations:
-- propagate signal along the edges, is an operation which sets I j(v) := 1 if and
only if there exists an edge (u, v) ∈ E with I j−1(u) = 1,
-- filter signal by a symbol x ∈ Σ, is an operation which sets I (v) := 0 for
each v where σ(v) 6= x,
-- match check, is an operation which checks whether there exists v ∈ F such
that I(v) = 1 and if so reports a match.
With these definitions in hand we can describe BMA in terms of prefix match
signals as Algorithm 2. See Fig. 2 for an example of use of BMA to figure out
whether P = acbab swap matches T = babcabc at a position 2.
2.4 Shift-And Algorithm
The following description is based on [11, Chapter 5] describing the Shift-Or
algorithm.
For a pattern P and a text T of length p and t, respectively, let R be a
bit array of size p and Rj its value after text symbol Tj has been processed. It
contains information about all matches of prefixes of P that end at the position j
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
7
a
b5
a
c
b2
c3
b
a
b
a4
b
a1
c
Fig. 2. BMA of T[2,6] = abcab on a P -graph of the pattern P = acbab. The prefix
match signal propagates along the dashed edges. Index j above a vertex v represent
that I j(v) = 1, otherwise I j(v) = 0.
in the text. For 1 ≤ i ≤ p, Rj
i = 1 if P[1,i] = T[j−i+1,j] and 0 otherwise. The
vector Rj+1 can be computed from Rj as follows. For each positive i we have
Rj+1
i+1 = 0 otherwise. Furthermore,
Rj+1
p = 1 then a complete match can
be reported.
i+1 = 1 if Rj
1 = 1 if P1 = Tj+1 and 0 otherwise. If Rj+1
i = 1 and Pi+1 = Tj+1, and Rj+1
The transition from Rj to Rj+1 can be computed very fast as follows. For each
x ∈ Σ let Dx be a bit array of size p such that for 1 ≤ i ≤ p, Dx
i = 1 if and only
if Pi = x. The array Dx denotes the positions of the symbol x in the pattern P .
Each Dx can be preprocessed before the search. The computation of Rj+1 is then
reduced to three bitwise operations, namely Rj+1 = (LShift(Rj) 1) & DTj+1 .
When Rj
p = 1, the algorithm reports a match on a position j − p + 1.
3 Our Algorithm
In this section we will show an algorithm which solves Swap Matching. We call
the algorithm GSM (Graph Swap Matching). GSM uses the graph theoretic
model presented in Section 2.2 and is based on the Shift-And algorithm from
Section 2.4.
The basic idea of the GSM algorithm is to represent prefix match signals
(see Definition 4) from the basic matching algorithm (Section 2.3) over PP in bit
vectors. The GSM algorithm represents all signals I in the bitmaps RX formed
by three vectors, one for each row. Each time GSM processes a symbol of T ,
it first propagates the signal along the edges, then filters the signal and finally
checks for matches. All these operations can be done very quickly thanks to
bitwise parallelism.
First, we make the concept of GSM more familiar by presenting a way to
interpret the Shift-And algorithm by means of the basic matching algorithm
(BMA) from Section 2.3 to solve the (ordinary) Pattern Matching problem. Then
we expand this idea to Swap Matching by using the graph theoretic model.
3.1 Graph Theoretic View of the Shift-And Algorithm
Let T and P be a text and a pattern of lengths t and p, respectively. We create
the T -graph TP = (V, E, σ) of the pattern P .
8
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Definition 5. Let S be a string. The T -graph of S is a graph TS = (V, E, σ)
where V = {vi 1 ≤ i ≤ S}, E = {(vi, vi+1) 1 ≤ i ≤ S − 1} and σ : V → Σ
such that σ(vi) = Si.
Note that the T -graph is directed acyclic graph which can be divided into
columns Vi, 1 ≤ i ≤ p (each of them containing one vertex vi) such that the
edges lead from Vj to Vj+1. This means that the T -graph satisfies all assump-
tions of BMA. We apply BMA to TP to figure out whether P matches T at
a position j. We get a correct result because for each i ∈ {1, . . . , p} we check
whether Tj+i−1 = σ(vi) = Pi.
To find every occurrence of P in T we would have to run BMA for each
position separately. This is basically the naive approach to solve the pattern
matching. We can improve the algorithm significantly when we parallelize the
computations of p runs of BMA in the following way.
The algorithm processes one symbol at a time starting from T1. We say that
the algorithm is in the j-th step when a symbol Tj has been processed. BMA
represents a prefix match as a prefix match signal I : V → {0, 1}. Its value in
the j-th step is denoted I j. Since one run of the BMA uses only one column of
the T -graph at any time we can use other vertices to represent different runs of
the BMA. We represent all prefix match indicators in one vector so that we can
manipulate them easily. To do that we prepare a bit vector R. Its value in j-th
step is denoted Rj and defined as Rj
i = I j(vi).
First operation which is used in BMA (propagate signal along the edges) can
be done easily by setting the signal of vi to value of the signal of its predeces-
sor vi−1 in the previous step. I.e., for i ∈ {1, . . . , p} we set I j(vi) = 1 if i = 1 and
I j(vi) = I j−1(vi−1) otherwise. In terms of Rj this means just Rj = LSO(Rj−1),
where LSO is defined as LSO(x) = LShift(x) 1.
We also need a way to set I (vi) := 0 for each vi for which σ(vi) 6= Tj+i which
is another basic BMA operation (filter signal by a symbol). We can do this
using the bit vector Dx from Section 2.4 and taking R & Dx. I.e., the algorithm
computes Rj as Rj = LSO(Rj−1) & DTj+1 .
The last BMA operation we have to define is the match detection. We do
p = 1 and if this is the case then a match starting at
this by checking whether Rj
position j − p + 1 occurred.
3.2 Our Algorithm for Swap Matching Using the Graph Theoretic
Model
Now we are ready to describe the GSM algorithm.
We again let PP = (V, E, σ) be the P -graph of the pattern P , apply BMA
to PP to figure out whether P matches T at a position j, and parallelize p runs
of BMA on PP .
Again, the algorithm processes one symbol at a time and it is in the j-th
step when symbol Tj is being processed. We again denote the value of the prefix
match signal I : V → {0, 1} of BMA in the j-th step by I j. I.e., the semantic
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
9
Algorithm 3 The graph swap matching (GSM)
Input: Pattern P of length p and text T of length t over alphabet Σ.
Output: Positions of all swap matches.
DPi
i
:= 1
1: Let RU 0 := RM 0 := RD0 := 0p.
2: Let Dx := 0p, for all x ∈ Σ.
3: for i = 1, 2, 3, . . . , p do
4:
5: for j = 1, 2, 3, . . . , t do
6:
7:
8:
9:
10:
11:
12:
13:
RU ′j := LSO(RDj−1).
RM ′j := LSO(RM j−1 RU j−1).
RD′j := LSO(RM j−1 RU j−1).
RU j := RU ′j & LShift(DTj ).
RM j := RM ′j & DTj .
RDj := RD′ j & RShift(DTj ).
if RU j
p = 1 then
p = 1 or RM j
report a match on position j − p + 1.
meaning of I j(mr,c) is that I j(mr,c) = 1 if there exists a swap permutation π
such that π(c) = c + r and π(P )[1,c] = T[j−c+1,j]. Otherwise I j(mr,c) is 0.
We want to represent all prefix match indicators in vectors so that we can
manipulate them easily. We can do this by mapping the values of I for rows
r ∈ {−1, 0, 1} of the P -graph to vectors RU , RM , and RD, respectively. We
denote value of the vector RX ∈ {RU , RM , RD} in j-th step as RX j. We define
values of the vectors as RU j
i = I j(m−1,i), RM j
i = I j(m1,i),
where the value of I j(v) = 0 for every v /∈ V .
i = I j(m0,i), and RDj
We define BMA propagate signal along the edges operation as setting the sig-
nal of mr,c to 1 if at least one of its predecessors have signal set to 1. I.e., we set
I j+1(m−1,i) := I j(m1,i−1), I j+1(m0,i) := I j(m−1,i−1) I j(m0,i−1), I j+1(m0,1) :=
1, I j+1(m1,i) := I j(m−1,i−1) I j(m0,i−1), and I j+1(m1,1) := 1. We can perform
the above operation using the LSO(R) operation. We obtain the propagate sig-
nal along the edges operation in the form RU ′j+1 := LSO(RDj), RM ′j+1 :=
LSO(RM j RU j), and RD′j+1 := LSO(RM j RU j).
The operation filter signal by a symbol can be done by first constructing a bit
vector Dx for each x ∈ Σ as Dx
i = 0 otherwise. Then we use
these vectors to filter signal by a symbol x by taking RU j := RU ′j &LShift(DTj ),
RM j := RM ′j & DTj , and RDj := RD′j & RShift(DTj ).
i = 1 if x = Pi and Dx
The last operation we define is the match detection. We do this by checking
p = 1 and if this is the case, then a match starting at
whether RU j
a position j − p + 1 occurred.
p = 1 or RM j
The final GSM algorithm (Algorithm 3) first prepares the D-masks Dx for
every x ∈ Σ and initializes RU 0 := RM 0 := RD0 := 0 (Steps 1 -- 4). Then the
algorithm computes the value of vectors RU j, RM j, and RDj for j ∈ {1, . . . , t}
by first using the above formula for signal propagation (Steps 6 -- 8) and then the
formula for signal filtering (Steps 9 -- 11) and checks whether RU j
p = 1
and if this is the case the algorithm reports a match (Steps 12 and 13).
p = 1 or RM j
10
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Observe that Algorithm 3 accesses the input sequentially and thus it is a
streaming algorithm. We now prove correctness of our algorithm. To ease the
notation let us define Rj(mr,c) to be RU j
c if r = 0, and RDj
c
if r = 1. We define R′j(mr,c) analogously. Similarly, we define Dx(mr,c) as
(LShift(Dx))c = Dx
c+1 if
r = 1. By the way the masks Dx are computed on lines 2 -- 4 of Algorithm 3, we
get the following observation.
c if r = 0, and (RShift(Dx))c = Dx
c if r = −1, RM j
c−1 if r = −1, Dx
Observation 1 For every mr,i ∈ V and every j ∈ {i, . . . t} we have DTj (mr,i) =
1 if and only if Tj = Pr+i.
The following lemma constitutes the crucial part of the correctness proof.
Lemma 1. For every mr,i ∈ V and every j ∈ {i, . . . t} we have Rj(mr,i) = 1
if and only if there exists a swap permutation π such that π(P )[1,i] = T[j−i+1,j]
and π(i) = i + r.
Proof. Let us start with the "if" part. We prove the claim by induction on i. If
i = 1 and there is a swap permutation π such that π(1) = 1 + r and P1+r = Tj,
then the algorithm sets R′j(mr,1) to 1 on line 6, 7, or 8 (recall the definition of
LSO). As P1+r = Tj, we have DTj (mr,1) = 1 by Observation 1 and, therefore,
by lines 9 -- 11, also Rj(mr,1).
Now assume that i > 1 and that the claim is true for every smaller i. Assume
that there exists a swap permutation π such that π(P )[1,i] = T[j−i+1,j] and
π(i) = i + r. By induction hypothesis we have that Rj−1(mr′,i−1) = 1, where
r′ = i−1−π(i−1). Since r equals −1 if and only if r′ equals +1 by Definition 1, we
have (r, r′) ∈ {(−1, 1), (0, −1), (0, 0), (1, −1), (1, 0)}. Therefore the algorithm sets
R′j(mr,i) to 1 on line 6, 7, or 8. Moreover, since Pi+r = Tj, we have DTj (mr,i) = 1
by Observation 1 and the algorithm sets Rj(mr,i) to 1 on one of the lines 9 -- 11.
Now we prove the "only if" part again by induction on i. If i = 1 and
Rj(mr,i) = 1, then we must have DTj (mr,1) = 1 and, by Observation 1, also
P1+r = Tj. We obtain π by setting π(1) = 1 + r, π(2) = 2 − r and π(i′) = i′ for
every i′ ∈ {2, . . . , p}. It is easy to verify that this is a swap permutation for P
and has the desired properties.
Now assume that i > 1 and that the claim is true for every smaller i. Assume
that Rj(mr,i) = 1. Then, due to lines 9 -- 11 we must have DTj (mr,i) = 1 and,
hence, by Observation 1, also Pi+r = Tj. Moreover, we must have R′j(mr,i) = 1
and, hence, by lines 6 -- 8 of the algorithm also Rj−1(mr′,i−1) = 1 for some r′ with
(r, r′) ∈ {(−1, 1), (0, −1), (0, 0), (1, −1), (1, 0)}. By induction hypothesis there
exists a swap permutation π′ for P such that π′(P )[1,i−1] = T[j−i+1,j−1] and
π′(i − 1) = i − 1 + r′. If π′(i) = i + r, then setting π = π′ finishes the proof.
Otherwise we have either r = 0 or r = 1 and i < p. In the former case we let
π(i′) = i′ for every i′ ∈ {i, . . . , p} and in the later case we let π(i) = i + 1,
π(i + 1) = i and π(i′) = i′ for every i′ ∈ {i + 2, . . . , p}. In both cases we let
π(i′) = π′(i′) for every i′ ∈ {1, . . . , i − 1}. It is again easy to verify that π is a
swap permutation for P with the desired properties.
⊓⊔
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
11
Theorem 2. The GSM algorithm is correct.
Proof. Our GSM algorithm reports a match on position j − p + 1 if and only if
Rj(mp,−1) = 1 or Rj(mp,0) = 1. However, by Lemma 1, this happens if and only
if there is a swap match of P on position j − p + 1 in T . Hence, the algorithm
is correct.
Theorem 3. The GSM algorithm runs in O(⌈ p
w ⌉(Σ + t) + p) time and uses
O(⌈ p
w ⌉Σ) memory cells (not counting the input and output cells), where t is
the length of the input text, p length of the input pattern, w is the word-size of
the machine, and Σ size of the alphabet.2
Proof. The initialization of RX and Dx masks (lines 1 and 2) takes O(⌈ p
w ⌉Σ)
time. The bits in Dx masks are set according to the pattern in O(p) time (lines 3
and 4). The main cycle of the algorithm (lines 5 -- 13) makes t iterations. Each
iteration consists of computing values of RX in 13 bitwise operations, i.e., in
O(⌈ p
w ⌉) machine operations, and checking for the result in O(1) time. This gives
O(⌈ p
w ⌉(Σ + t) + p) time in total. The algorithm saves 3 RX masks (using
the same space for all j and also for RX ′ masks), Σ Dx masks, and constant
number of variables for other uses (iteration counters, temporary variable, etc.).
Thus, in total the GSM algorithm needs O(⌈ p
⊓⊔
w ⌉Σ) memory cells.
Corollary 1. If p = cw for some constant c, then the GSM algorithm runs in
O(Σ + p + t) time and has O(Σ) space complexity. Moreover, if p ≤ w, then
the GSM algorithm can be implemented using only 7 + Σ memory cells.
Proof. The first part follows directly from Theorem 3. Let us show the second
part. We need Σ cells for all D-masks, 3 cells for R vectors (reusing the space
also for R′ vectors), one pointer to the text, one iteration counter, one constant
for the match check and one temporary variable for the computation of the more
complex parts of the algorithm. Alltogether, we need only 7 + Σ memory cells
to run the GSM algorithm.
⊓⊔
From the space complexity analysis we see that for some sufficiently small
alphabets (e.g. DNA sequences) the GSM algorithm can be implemented in
practice using solely CPU registers with the exception of text which has to be
loaded from the RAM.
4 Limitations of the Finite Deterministic Automata
Approach
Many of the string matching problems can be solved by finite automata. The
construction of a non-deterministic finite automaton that solves Swap Matching
can be done by a simple modification of the P -graph. An alternative approach
2 To simplify the analysis, we assume that log t < w, i.e., the iteration counter fits
into one memory cell.
12
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Table 1. An example of the construction from proof of Theorem 4 for k = 3.
P = T0 acabcabcabc
T1 acabcabcbac
T2 acabcbacabc
T3 acabcbacbac
T4 acbacabcabc
T5 acbacabcbac
T6 acbacbacabc
T7 acbacbacbac
to solve the Swap Matching would thus be to determinize and execute this
automaton. The drawback is that the determinization process may lead to an
exponential number of states. We show that in some cases it actually does, con-
tradicting the conjecture of Holub [16], stating that the number of states of this
determinized automaton is O(p).
Theorem 4. There is an infinite family F of patterns such that any deter-
ministic finite automaton AP accepting the language LS(P ) = {uπ(P ) u ∈
Σ∗, π is a swap permutation for P } for P ∈ F has 2Ω(P ) states.
k−1, bi
k−2 . . . bi
j = abc if bi
j = 0 and let Bi
j = bac if bi
k−1Bi
k−2 . . . Bi
Proof. For any integer k we define the pattern Pk := ac(abc)k. Note that the
length of Pk is Θ(k). Suppose that the automaton AP recognizing language L(P )
has s states such that s < 2k. We consider a set of strings T0, . . . , T2k−1 where Ti
is defined as follows. Let bi
0 be the binary representation of the
number i. Let Bi
j = 1. Then, let
Ti := acBi
0. See Table 1 for an example. Note that each Ti, i ∈
{0, . . . , 2k − 1} is a swapped version of P = T0. Since s < 2k, there exist
0 ≤ i < j ≤ 2k − 1 such that both Ti and Tj are accepted by the same
accepting state q of the automaton A. Let m be the minimum number such
m = 1. Now we define T ′
that bi
i =
Ti(abc)(m+1) and T ′
i )[3(m+1)+1,3(m+1+k)+2] and
Y = (T ′
j both of
length 3k + 2. Note that X begins with bc . . . and Y begins with ac . . . and that
block abc or bac repeats for k times in both. Therefore pattern P swap matches Y
and does not swap match X. Since for the last symbol of both Ti and Tj the
automaton is in the same state q, the computation for T ′
j must end in the
same state q′. However as X should not be accepted and Y should be accepted
we obtain contradiction with the correctness of the automaton A. Hence, we may
define the family F as F = {P1, P2, . . . }, concluding the proof.
⊓⊔
m = 0 and bj
j = Tj(abc)(m+1). Let X = (T ′
k−1−m. Note that bi
j)[3(m+1)+1,3(m+1+k)+2] be the suffices of the strings T ′
i and T ′
k−1−m 6= bj
i and T ′
This proof shows the necessity for specially designed algorithms which solve the
Swap Matching. We presented one in the previous section and now we reiterate
on the existing algorithms.
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
13
5 Smalgo Algorithm
In this section we discuss how Smalgo by Iliopoulos and Rahman [17] and
Smalgo-I and Smalgo-II by Ahmed et al. [1] work. Since Smalgo-I is bitwise
inverse of Smalgo, we will introduce them both in terms of operations used in
Smalgo-I. After that we will describe and analyze Smalgo-II.
Before we show how these algorithms work, we need one more definition.
Definition 6. A degenerate symbol w over an alphabet Σ is a nonempty set of
symbols from alphabet Σ. A degenerate string S is a string built over an alphabet
of degenerate symbols. We say that a degenerate string eP matches a text T at a
position j if Tj+i−1 ∈ ePi for every 1 ≤ i ≤ p.
5.1 Smalgo-I
The Smalgo-I [1] algorithm is a modification of the Shift-And algorithm from
Section 2.4 for Swap Matching. The algorithm uses the graph theoretic model
introduced in Section 2.2.
First let eP = {P1, P2} . . . {Px−1, Px, Px+1} . . . {Pp−1, Pp} be a a degenerate
version of pattern P . The symbol on position i in eP represents the set of symbols
of P which can swap to that position. To accommodate the Shift-And algorithm
to match degenerate patterns we need to change the way the Dx masks are
defined. For each x ∈ Σ let eDx
i be the bit array of size p such that for 1 ≤ i ≤
p, eDx = 1 if and only if x ∈ ePi.
While a match of the degenerate pattern eP is a necessary condition for a swap
match of P , it is clearly not sufficient. The way the Smalgo algorithms try to fix
this is by introducing P-mask P (x1, x2, x3) which is defined as P (x1, x2, x3)i = 1
if i = 1 or if there exist vertices u1, u2, and u3 and edges (u1, u2), (u2, u3) in PP
for which u2 = mr,i for some r ∈ {−1, 0, 1} and σ(un) = xn for 1 ≤ n ≤ 3, and
P (x1, x2, x3)i = 0 otherwise. One P -mask called P (x, x, x) is used to represent
the P -masks for triples (x1, x2, x3) which only contain 1 in the first column.
Now, whenever checking whether P prefix swap matches T k + 1 symbols
at position j we check for a match of eP in T and we also check whether
P (Tj+k−1, Tj+k, Tj+k+1)k+1 = 1. This ensures that the symbols are able to swap
to respective positions and that those three symbols of the text T are present in
some π(P ).
With the P-masks completed we initialize R1 = 1& eDT1 . Then for every j = 1
to t we repeat the following. We compute Rj+1 as Rj+1 = LSO(Rj) & eDTj+1 &
RShift( eDTj+2 ) & P (Tj, Tj+1, Tj+2). To check whether or not a swap match oc-
curred we check whether Rj
p−1 = 1. This is claimed to be sufficient because
during the processing we are in fact considering not only the next symbol Tj+1
but also the symbol Tj+2.
5.2 The Flaw in the Smalgo, Smalgo-I and Smalgo-II
We shall see that for a pattern P = abab and a text T = aaba all Smalgo
versions give false positives.
14
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
a
b
a
b
a
b
a
b
a
b
Fig. 3. Smalgo flaw represented in the P -graph for P = abab
The concept of Smalgo is based on the assumption that we can find a path
in PP by searching for consecutive paths of length 3 (triplets), where each two
consecutive share two columns and can partially overlap. However, this only
works if the consecutive triplets actually share the two vertices in the common
columns. If the assumption is not true then the found substring of the text might
not match any swapped version of P .
The above input gives such a configuration (see Fig. 3) and therefore the
assumption is false. The Smalgo-I algorithm actually reports match of pattern
P = abab on a position 1 of text T = aaba. This is obviously a false positive, as
the pattern has two b symbols while the text has only one.
The reason behind the false positive match is as follows. The algorithm checks
whether the first triplet of symbols (a, a, b) matches. It can match the swap
pattern aabb. Next it checks the second triplet of symbols (a, b, a), which can
match baba. We know that baba is not possible since it did not appear in the
previous check, but the algorithm cannot distinguish them since it only checks
for triplets existence. Since each step gave us a positive match the algorithm
reports a swap match of the pattern in the text.
In the Fig. 3 we see the two triplets which Smalgo assumes have two vertices
in common. The Smalgo-II algorithm saves space by maintaining less informa-
tion, however it simulates how Smalgo-I works and so it contains the same
flaw.
5.3 The Run of Smalgo-I Resulting in the False Positive
In Tables 2 and 3 we can see the step by step execution of Smalgo-I algorithm
on pattern P = abab and text T = aaba. In Table 3 we see that R3 has 1
in the 3-rd row which means that the algorithm reports a pattern match on a
position 1. This is a false positive, because it is not possible to swap match the
pattern with two b symbols in the text with only one b symbol.
5.4 Description of Smalgo-II
To explain the Smalgo-II algorithm in more detail, we first introduce a notion
of change. An upward change corresponds to (the BMA) going to vertex m−1,i
for some i, a downward change corresponds to going to vertex m+1,i, and a
middle-ward change corresponds to going to vertex m0,i.
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
15
Table 2. eD-masks and P-masks for P = abab. A column xyz contains values P (x, y, z)i.
ePi
i
1 [ab]
2 [ba]
3 [ab]
4 [ba]
eDa
i
1
1
1
1
eDb
i aaa aab aba baa abb bab bba bbb
1
1
0
1
1
0
0
1
1
1
1
0
1
1
1
0
1
1
1
0
1
0
1
0
1
0
0
0
1
1
1
0
1
1
0
0
Table 3. Smalgo-I algorithm execution for P = abab and T = aaba. The column RDx
denotes the values of RShift( eDx).
i R1 LSO(R1) eDa RDb P (a, a, b) R2 LSO(R2) eDb RDa P (a, b, a) R3
1
1
1
2
3
1
0
4
1
1
0
0
1
1
1
0
1
0
0
0
1
1
1
1
1
1
1
0
1
1
1
0
1
1
0
0
1
1
1
0
1
1
1
1
1
1
1
0
If a downward change has occurred, then we have to check whether an upward
change occurs at the next position. If an upward change has occurred, then we
have to check whether a downward or middle-ward change occurs at the next
position. The main problem here is how to tell whether the changes actually
occur.
To this end, the authors of the algorithm introduce three new types of masks,
namely up-masks up(x,y), down-masks down(x,y), and middle-masks middle(x,y),
which express whether an upward, a downward, and a middle-ward change can
occur at the particular position, respectively, with the endpoints of the edge
having labels x and y.
The authors of the algorithm now claim that to perform the above checks, it
is enough to save the previous down-mask and match its value with current up-
mask and Rj, or to save the previous up-mask and match its value with current
down-mask, middle-mask, and Rj, respectively. However, this way in both cases
we only check whether the change can occur, not whether it actually occurred.
This would lead not only to false positives (as shown in Section 5), but also to
false negatives.
Unfortunately, no more details are available about the algorithm in the orig-
inal paper. The pseudocode of Smalgo-II (which contains numerous errors)
performs something different and we include its analysis in the next section for
completeness. Nevertheless, the example presented in the Section 5 and in the
previous section still makes the pseudocode (with the small errors corrected)
report a false positive.
5.5 Analysis of the Pseudocode of Smalgo-II
In this section we analyze the pseudocode of the Smalgo-II algorithm as given
by Ahmed et al. in [1], we will perform equivalent transformation on the pseu-
16
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
docode in order to understand the meaning of the checks the pseudocode actually
performs.
The original pseudocode is as follows.
Algorithm 4 Smalgo-II
Require: Text T, up-mask up, down-mask down, middle-mask middle,
P-mask pmask, D-mask D for given pattern p
1: R0 ← 2patternLength−1
2: checkUp ← checkDown ← 0
3: R0 ← R0 & DT0
4: R1 ← R0 ≫ 1
5: for j = 0 to (n − 2) do
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
Rj+1 ← Rj ≫ 1
checkUp ← checkUp ≫ 1
checkDown ← checkDown ≫ 1
Rj ← Rj & pmask(Tj ,Tj+1 ) & DTj+1
temp ← prevCheckUp ≫ 1
checkUp ← checkUp up(Tj ,Tj+1)
checkUp ← checkUp & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1 )
prevCheckUp ← checkUp
Rj ← ∼ (temp & checkUp) & Rj
temp ← prevCheckDown ≫ 1
checkDown ← checkDown down(Tj ,Tj+1)
checkDown ← checkDown & ∼ up(Tj ,Tj+1)
prevCheckDown ← checkDown
Rj ← ∼ (temp & checkDown) & Rj
if (Rj & 1) = 1 then
Match found ending at position (j − 1)
The pseudocode has several problems. First, in the first iteration of the cy-
cle, the algorithm uses the value of the variable prevCheckUp which was never
initialized. Second, the algorithm never adds new ones to the variable R and,
hence, can never report a match after position patternLength of the text. Third,
if the text is of the same length as the pattern, the algorithm only applies
the shift patternLength − 2 times to the original value of 2patternLength−1 (note
that in the first iteration it uses R0 and overwrites the value of R1) before the
last match check. Therefore, at the last check, the value could only drop to
2patternLength−1−patternLength+2 = 21 = 2 and the match check cannot be success-
ful. Also the reported position of the match does not make much sense.
Let us first correct all these easy problems.
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
17
Algorithm 5 Smalgo-II
1: R0 ← 2patternLength−1
2: prevCheckUp ← prevCheckDown ← checkUp ← checkDown ← 0
3: R0 ← R0 & DT0
4: R1 ← R0 ≫ 1
5: for j = 0 to (n − 2) do
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
Rj+1 ← Rj+1 & pmask(Tj ,Tj+1) & DTj+1
temp ← prevCheckUp ≫ 1
checkUp ← checkUp up(Tj ,Tj+1)
checkUp ← checkUp & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1 )
prevCheckUp ← checkUp
Rj+1 ← ∼ (temp & checkUp) & Rj+1
temp ← prevCheckDown ≫ 1
checkDown ← checkDown down(Tj ,Tj+1)
checkDown ← checkDown & ∼ up(Tj ,Tj+1)
prevCheckDown ← checkDown
Rj+1 ← ∼ (temp & checkDown) & Rj+1
if (Rj+1 & 1) = 1 then
Rj+2 ← (Rj+1 ≫ 1) 2patternLength−1
checkUp ← checkUp ≫ 1
checkDown ← checkDown ≫ 1
Match found ending at position (j + 1)
If we now move the line setting prevCheckUp to checkUp after the line where
the check with the temp variable is performed and similarly with prevCheckDown,
we do not need the temp variable anymore. We also move the shifts of checkUp
and checkDown closer to where this variables are used. We only show the impor-
tant part of the algorithm.
Algorithm 6 Smalgo-II
. . .
Rj+1 ← Rj+1 & pmask(Tj ,Tj+1) & DTj+1
checkUp ← checkUp up(Tj ,Tj+1)
checkUp ← checkUp & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1 )
Rj+1 ← ∼ (prevCheckUp ≫ 1 & checkUp) & Rj+1
prevCheckUp ← checkUp
checkUp ← checkUp ≫ 1
checkDown ← checkDown down(Tj ,Tj+1)
checkDown ← checkDown & ∼ up(Tj ,Tj+1)
Rj+1 ← ∼ (prevCheckDown ≫ 1 & checkDown) & Rj+1
prevCheckDown ← checkDown
checkDown ← checkDown ≫ 1
if (Rj+1 & 1) = 1 then
5: for j = 0 to (n − 2) do
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
Rj+2 ← (Rj+1 ≫ 1) 2patternLength−1
Match found ending at position (j + 1)
18
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Now we swap the order of setting prevCheckUp to checkUp and the shift of
checkUp. As this makes prevCheckUp shifted by one, we remove the additional
shift in the check. Similarly for checkDown.
Algorithm 7 Smalgo-II
. . .
7: checkUp ← checkUp up(Tj ,Tj+1)
8: checkUp ← checkUp & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1 )
9: Rj+1 ← ∼ (prevCheckUp & checkUp) & Rj+1
10: checkUp ← checkUp ≫ 1
11: prevCheckUp ← checkUp
12: checkDown ← checkDown down(Tj ,Tj+1)
13: checkDown ← checkDown & ∼ up(Tj ,Tj+1 )
14: Rj+1 ← ∼ (prevCheckDown & checkDown) & Rj+1
15: checkDown ← checkDown ≫ 1
16: prevCheckDown ← checkDown
. . .
Now we institute checkUp into the check and move its computation after the
check.
Algorithm 8 Smalgo-II
. . .
6: Rj+1 ← Rj+1 & pmask(Tj ,Tj+1 ) & DTj+1
7: Rj+1 ← ∼ (prevCheckUp & (checkUp up(Tj ,Tj+1)) & ∼ down(Tj ,Tj+1) & ∼
middle(Tj ,Tj+1)) & Rj+1
8: checkUp ← (checkUp up(Tj ,Tj+1)) & ∼ down(Tj ,Tj+1 ) & ∼ middle(Tj ,Tj+1)
9: checkUp ← checkUp ≫ 1
10: prevCheckUp ← checkUp
11: Rj+1 ← ∼ (prevCheckDown & (checkDown down(Tj ,Tj+1 )) & ∼ up(Tj ,Tj+1)) &
Rj+1
12: checkDown ← (checkDown down(Tj ,Tj+1)) & ∼ up(Tj ,Tj+1)
13: checkDown ← checkDown ≫ 1
14: prevCheckDown ← checkDown
. . .
Now note that during the check, the content of prevCheckUp is exactly the
same as the content of checkUp, so we can remove prevCheckUp completely.
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
19
Algorithm 9 Smalgo-II
1: R0 ← 2patternLength−1
2: checkUp ← checkDown ← 0
3: R0 ← R0 & DT0
4: R1 ← R0 ≫ 1
5: for j = 0 to (n − 2) do
6:
7:
Rj+1 ← Rj+1 & pmask(Tj ,Tj+1) & DTj+1
Rj+1 ← ∼ (checkUp & (checkUp up(Tj ,Tj+1 )) & ∼ down(Tj ,Tj+1) & ∼
middle(Tj ,Tj+1)) & Rj+1
checkUp ← (checkUp up(Tj ,Tj+1)) & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1 )
checkUp ← checkUp ≫ 1
Rj+1 ← ∼ (checkDown & (checkDown down(Tj ,Tj+1)) & ∼ up(Tj ,Tj+1 )) &
Rj+1
checkDown ← (checkDown down(Tj ,Tj+1 )) & ∼ up(Tj ,Tj+1)
checkDown ← checkDown ≫ 1
if (Rj+1 & 1) = 1 then
Match found ending at position (j + 1)
Rj+2 ← (Rj+1 ≫ 1) 2patternLength−1
8:
9:
10:
11:
12:
13:
14:
15:
Now we modify the expressions by laws of logic to arrive at the following
formulation.
Algorithm 10 Smalgo-II
. . .
7: Rj+1 ← Rj+1 & ( ∼ checkUp down(Tj ,Tj+1) middle(Tj ,Tj+1))
8: checkUp ← (checkUp & ∼ down(Tj ,Tj+1) & ∼ middle(Tj ,Tj+1)) (up(Tj ,Tj+1 ) &
∼ down(Tj ,Tj+1 ) & ∼ middle(Tj ,Tj+1))
9: checkUp ← checkUp ≫ 1
10: Rj+1 ← Rj+1 & ( ∼ checkDown up(Tj ,Tj+1))
11: checkDown ← (checkDown & ∼ up(Tj ,Tj+1)) (down(Tj ,Tj+1) & ∼ up(Tj ,Tj+1))
12: checkDown ← checkDown ≫ 1
. . .
Now, if the first subexpression in the logical OR setting the new value of
checkUp is true, then the appropriate bit of Rj+1 was just set to 0 on the
previous line and filtrating this bit again in future is useless. Hence, we can omit
this part of the expression. We arrive at the following resulting pseudocode.
20
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Algorithm 11 Smalgo-II
1: R0 ← 2patternLength−1
2: checkUp ← checkDown ← 0
3: R0 ← R0 & DT0
4: R1 ← R0 ≫ 1
5: for j = 0 to (n − 2) do
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
Rj+1 ← Rj+1 & pmask(Tj ,Tj+1) & DTj+1
Rj+1 ← Rj+1 & ( ∼ checkUp down(Tj ,Tj+1 ) middle(Tj ,Tj+1))
checkUp ← up(Tj ,Tj+1) & ∼ down(Tj ,Tj+1 ) & ∼ middle(Tj ,Tj+1)
checkUp ← checkUp ≫ 1
Rj+1 ← Rj+1 & ( ∼ checkDown up(Tj ,Tj+1 ))
checkDown ← down(Tj ,Tj+1) & ∼ up(Tj ,Tj+1)
checkDown ← checkDown ≫ 1
if (Rj+1 & 1) = 1 then
Match found ending at position (j + 1)
Rj+2 ← (Rj+1 ≫ 1) 2patternLength−1
Now it is easy to see, that checkUp stores the information on whether an
upward-change must have occurred in the previous step (provided that there was
a prefix match) and this is compared with the information whether downward-
change or middle-change can occur. Similarly for the downward-change. This is
not sufficient to avoid false positives since sometimes both upward-change and
downward-change can occur (e.g, as in our counterexample), in which case no
filtration is performed at all.
5.6 Why the Flaw is Not Easily Repairable
Consider the following attempt to fix the Smalgo-I or Smalgo-II. After each
reported match we check for the validity of the result using a single linear-time
algorithm. This approach would rule out false positives but it ruins the time
complexity of the algorithms, since there are texts of arbitrary length t with
Θ(t) of reported occurrences.
Namely consider the text T = aa(baa)n for some positive n, pattern P = abab,
and let t = T . Note that n = (t − 2)/3 = Θ(t). Text T contains string aaba
on positions 1, 4, 7, . . . , 3(n − 1) + 1 (n occurrences in total) and string baab on
positions 3, 6, . . . , 3(n − 1) (n − 1 occurrences in total). Thus there are 2n − 1
occurrences which need to be checked since the n occurrences of aaba are reported
by the algorithms although they are not valid matches. Even if the checking for
correctness was done in linear time O(p), the algorithms will report up to Θ(t)
occurrences which means we have to run the checking algorithm Θ(t) times.
Therefore the time complexity of a version of the Smalgo algorithm corrected
this way is O(tp) even for a pattern length similar to the word-size of the target
machine.
Also, the flaw cannot be resolved by checking for subpaths of length 4 or any
larger constant, due to the following. Consider a pattern P = (ab)n and a text
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
21
T = aa(ba)n−1 for any positive n. Obviously P does not swap match T , as they
are of the same length 2n, but T contains more a's than P . However, there is a
swap permutation π for P such that (π(P ))[1...(2n−1)] = T[1...(2n−1)] and also a
swap permutation π′ for P such that (π′(P ))[2...(2n)] = T[2...(2n)]. For example if
we have P = abab and a text T = aabaabaabaa both Smalgo algorithms report
swap matches on positions {1, 3, 4, 6, 7} while the correct output would be {3, 6}.
6 Experiments
We implemented our Algorithm 3 (GSM), described in Section 3.2, the Bitwise
Parallel Cross Sampling (BPCS) algorithm by Cantone and Faro [10], the Bitwise
Parallel Backward Cross Sampling (BPBCS) algorithm by Campanelli et al. [9],
and the faulty SMALGO algorithm by Iliopoulos and Rahman [17]. All these
implementations are available online.3
We tested the implementations on three real-world datasets. The first dataset
(CH) is the 7th chromosome of the human genome4 which consists of 159 M char-
acters from the standard ACTG nucleobases and N as for non-determined. Second
dataset (HS) is a partial genome of Homo sapiens from the Protein Corpus5 with
3.3 M characters representing proteins encoded in 19 different symbols. The last
dataset (BIB) is the Bible text of the Cantenbury Corpus6 with 4.0 M characters
containing 62 different symbols. For each length from 3, 4, 5, 6, 8, 9, 10, 12, 16, and
32 we randomly selected 10,000 patterns from each text and processed each of
them with each implemented algorithm.
All measured algorithms were implemented in C++ and compiled with -O3 in
gcc 6.3.0. Measurements were carried on an Intel Core i7-4700HQ processor
with 2.4 GHz base frequency and 3.4 GHz turbo with 8 GiB of DDR3 memory at
1.6 GHz. Time was measured using std::chrono::high resolution clock::now()
from the C++ chrono library. The resulting running times, shown in Table 4, were
averaged over the 10,000 patterns of the given length.
The results show, that the GSM algorithm runs approximately 23% faster
than Smalgo (ignoring the fact that Smalgo is faulty by design). Also, the
performance of GSM and BPCS is almost indistinguishable and according to
our experiments, it varies in the span of units of percents depending on the
exact CPU, cache, RAM and compiler setting. The seemingly superior average
performance of BPBCS is caused by the heuristics BPBCS uses; however, while
the worst-case performance of GSM is guaranteed, the performance of BPBCS for
certain patterns is worse than that of GSM. Also note that GSM is a streaming
algorithm while the others are not.
Table 5 visualizes the accurateness of Smalgo-I with respect to its flaw by
comparing the number of occurrences found by the respective algorithms. The
3 http://users.fit.cvut.cz/blazeva1/gsm.html
4 ftp://ftp.ensembl.org/pub/release-90/fasta/homo_sapiens/dna/
5 http://www.data-compression.info/Corpora/ProteinCorpus/
6 http://corpus.canterbury.ac.nz/descriptions/large/bible.html
22
V´aclav Blazej, Ondrej Such´y, and Tom´as Valla
Table 4. Comparison of the running times in milliseconds. Each value is the average
over 10,000 patterns randomly selected from the text.
Data
(Σ)
CH
(5)
HS
(19)
BIB
(62)
Algor.
3
SMALGO 426
398
824
394
BPCS
BPBCS
GSM
Pattern Length
4
376
353
675
354
5
355
335
555
338
6
350
332
472
333
8
347
329
366
332
9
347
329
328
331
10
344
326
297
329
12
347
328
257
333
16
345
329
199
331
32
345
327
112
333
SMALGO 4.80
4.43
7.16
4.42
BPCS
BPBCS
GSM
4.72
4.74
4.73
4.36 4.36 4.36
5.80
4.38
4.70
4.34
4.79 4.05 3.03 2.70 2.44 2.06 1.62 0.95
4.41
4.48
4.71
4.34
4.72
4.35
4.71
4.34
4.70
4.34
4.71
4.33
4.44
4.53
4.45
4.46
4.45
4.45
8.29
SMALGO 8.60
7.53
12.43 10.03
7.37
7.52
BPCS
BPBCS
GSM
8.34
8.38
7.36 7.28 7.29
8.32
7.26
8.33
7.25
8.26 7.03 5.44 4.93 4.52 3.93 3.19 1.88
7.31
7.40
8.35
7.28
8.35
7.29
8.30
7.26
8.33
7.27
7.42
7.44
7.35
7.38
7.40
7.38
Table 5. Found occurrences across datasets: The value is simply the sum of occurrences
over all the patterns.
Algorithm
CH
Dataset
HS
BIB
SMALGO 86243500784 51136419 315612770
84411799892 51034766 315606151
rest
ratio of false positives to true positives for the Smalgo-I was: CH 2.17%, HS
0.20% and BIB 0.002%.
References
1. Ahmed, P., Iliopoulos, C.S., Islam, A.S., Rahman, M.S.: The swap matching prob-
lem revisited. Theoretical Computer Science 557, 34 -- 49 (2014)
2. Amir, A., Aumann, Y., Landau, G.M., Lewenstein, M., Lewenstein, N.: Pattern
matching with swaps. Journal of Algorithms 37(2), 247 -- 266 (2000)
3. Amir, A., Cole, R., Hariharan, R., Lewenstein, M., Porat, E.: Overlap matching.
Information and Computation 181(1), 57 -- 74 (2003)
4. Amir, A., Eisenberg, E., Porat, E.: Swap and mismatch edit distance. Algorithmica
45(1), 109 -- 120 (2006)
5. Amir, A., Landau, G.M., Lewenstein, M., Lewenstein, N.: Efficient special cases
of pattern matching with swaps. Information Processing Letters 68(3), 125 -- 132
(1998)
6. Amir, A., Lewenstein, M., Porat, E.: Approximate swapped matching. Information
Processing Letters 83(1), 33 -- 39 (2002)
A Simple Streaming Bit-parallel Algorithm for Swap Pattern Matching
23
7. Antoniou, P., Iliopoulos, C.S., Jayasekera, I., Rahman, M.S.: Implementation of
a swap matching algorithm using a graph theoretic model. In: Bioinformatics Re-
search and Development, BIRD 2008, CCIS, vol. 13, pp. 446 -- 455. Springer (2008)
8. Blazej, V., Such´y, O., Valla, T.: A simple streaming bit-parallel algorithm for
swap pattern matching. In: Mathematical Aspects of Computer and Information
Sciences, MACIS 2017. LNCS, vol. 10693, pp. 333 -- 348. Springer (2017)
9. Campanelli, M., Cantone, D., Faro, S.: A new algorithm for efficient pattern
matching with swaps. In: International Workshop on Combinatorial Algorithms,
IWOCA 2009, LNCS, vol. 5874, pp. 230 -- 241. Springer (2009)
10. Cantone, D., Faro, S.: Pattern matching with swaps for short patterns in linear
time. In: International Conference on Current Trends in Theory and Practice of
Computer Science, SOFSEM 2009, LNCS, vol. 5404, pp. 255 -- 266. Springer (2009)
11. Charras, C., Lecroq, T.: Handbook of Exact String Matching Algorithms. King's
College Publications (2004)
12. Chedid, F.: On pattern matching with swaps. In: IEEE/ACS International Con-
ference on Computer Systems and Applications, AICCSA 2013. pp. 1 -- 5. IEEE
(2013)
13. Dombb, Y., Lipsky, O., Porat, B., Porat, E., Tsur, A.: The approximate swap and
mismatch edit distance. Theoretical Computer Science 411(43), 3814 -- 3822 (2010)
14. Faro, S.: Swap matching in strings by simulating reactive automata. In: Proceedings
of the Prague Stringology Conference 2013, pp. 7 -- 20. CTU in Prague (2013)
15. Fredriksson, K., Giaquinta, E.: On a compact encoding of the swap automaton.
Information Processing Letters 114(7), 392 -- 396 (2014)
16. Holub, J.: Personal communication (2015)
17. Iliopoulos, C.S., Rahman, M.S.: A new model to solve the swap matching problem
and efficient algorithms for short patterns. In: International Conference on Current
Trends in Theory and Practice of Computer Science, SOFSEM 2008, LNCS, vol.
4910, pp. 316 -- 327. Springer (2008)
18. Lewin, B.: Genes for SMA: Multum in parvo. Cell 80(1), 1 -- 5 (1995)
19. Lipsky, O., Porat, B., Porat, E., Shalom, B.R., Tzur, A.: String matching with
up to k swaps and mismatches. Information and Computation 208(9), 1020 -- 1030
(2010)
20. Muthukrishnan, S.: New results and open problems related to non-standard
stringology. In: Annual Symposium on Combinatorial Pattern Matching, CPM 95,
LNCS, vol. 937, pp. 298 -- 317. Springer (1995)
21. Wagner, R.A., Lowrance, R.: An extension of the string-to-string correction prob-
lem. Journal of the ACM 22(2), 177 -- 183 (1975)
|
1512.06283 | 2 | 1512 | 2016-09-12T12:23:35 | Chinese Postman Problem on Edge-Colored Multigraphs | [
"cs.DS",
"math.CO"
] | It is well-known that the Chinese postman problem on undirected and directed graphs is polynomial-time solvable. We extend this result to edge-colored multigraphs. Our result is in sharp contrast to the Chinese postman problem on mixed graphs, i.e., graphs with directed and undirected edges, for which the problem is NP-hard. | cs.DS | cs |
Chinese Postman Problem on Edge-Colored
Multigraphs
Gregory Gutin1, Mark Jones1, Bin Sheng1, Magnus Wahlstrom1, and
Anders Yeo2,3
1Department of Computer Science, Royal Holloway, University of
London, TW20 0EX, Egham, Surrey, UK
2Engineering Systems and Design, Singapore University of Technology
and Design, 8 Somapah Road 487372, Singapore
3Department of Mathematics, University of Johannesburg, Auckland
Park, 2006 South Africa
September 13, 2016
Abstract
It is well-known that the Chinese Postman Problem on undirected
and directed graphs is polynomial-time solvable. We extend this result
to edge-colored multigraphs. Our result is in sharp contrast to the Chi-
nese Postman Problem on mixed graphs, i.e., graphs with directed and
undirected edges, for which the problem is NP-hard.
1 Introduction
In this paper, we consider edge-colored multigraphs.
In such multigraphs,
each edge is assigned a color; a multigraph G is called k-edge-colored if only
colors from [k] := {1, 2, . . . , k} are used in G. A walk1 W in an edge-colored
multigraph is called properly colored (PC) if no two consecutive edges of W
have the same color. PC walks are of interest in graph theory applications,
e.g., in genetic and molecular biology [18, 20, 21], in design of printed circuit
and wiring boards [22], and in channel assignment in wireless networks [3, 19].
They are also of interest in graph theory itself as generalizations of walks in
undirected and directed graphs.
Indeed, if we assign different colors to all
edges of an undirected multigraph, every walk not traversing the same edge
1 Terminology on walks used in this paper is given in the next section.
1
twice becomes PC. Also, consider the standard transformation from a directed
graph D into a 2-edge-colored graph G by replacing every arc uv of D by a
path with a blue edge uwuv and a red edge wuvv, where wuv is a new vertex
[4]. Clearly, every directed walk in D corresponds to a PC walk in G (with
end-vertices in V (G)) and vice versa. There is an extensive literature on PC
walks: for a detailed survey of pre-2009 publications, see Chapter 16 of [4],
more recent papers include [2, 8, 13, 14, 15].
A walk is closed if it starts and ends in the same vertex. (A closed walk
W has no last edge, every edge in W has a following edge; if W is PC, each
edge of W is of different color to the following edge.) An Euler trail in a
multigraph G is a closed walk which traverses each edge of G exactly once.
PC Euler trails were one of the first types of PC walks studied in the literature
and the first papers that studied PC Euler trails were motivated by theoretical
questions [7, 12] as well as questions in molecular biology [18]. To formulate
a characterization of edge-colored graphs with PC Euler trails by Kotzig [12],
we introduce additional terminology. A vertex in an edge-colored multigraph
is balanced if no color appears on more than half of the edges incident with
the vertex, and even if it is of even degree. We say that an edge-colored graph
is PC Euler if it contains a PC Euler trail.
Theorem 1. [12] An edge-colored multigraph G is PC Euler if and only if G
is connected and every vertex of G is balanced and even.
Benkouar et al.
[6] described a polynomial-time algorithm to find a PC
Euler trail in an edge-colored multigraph, if it contains one. Studying DNA
physical mapping, Pevzner [17] came up with a simpler polynomial-time algo-
rithm solving the same problem.
In this paper, we consider the Chinese Postman Problem on edge-colored
graphs (CPP-ECG): given a connected edge-colored multigraph G with non-
negative weights on its edges, find a PC closed walk in G which traverses all
edges of G and has the minimum weight2 among such walks.
Observe that to solve CPP-ECG, it is enough to find a PC Euler edge-
colored multigraph G∗ of minimum weight such that V (G∗) = V (G) and for
every pair of distinct vertices u, v and color i, G∗ has p∗ > 0 parallel edges
between vertices u and v of color i if and only if G has at least one and at
most p∗ edges of color i between u and v. (To find the actual walk, we can
use the algorithm from [6] or [18].)
CPP-ECG is a generalization of the PC Euler trail problem as an instance
G has a PC Euler trail if and only if G∗ = G. CPP-ECG is also a generaliza-
tion of the Chinese Postman Problem (CPP) on both undirected and directed
multigraphs (the arguments are the same as for PC walks above). However,
2The weight of a walk is the sum of the weights of its edges.
2
while CPP on both undirected and directed multigraphs has a solution on
every connected multigraph G, it is not the case for CPP-ECG. Indeed, there
is no solution on any connected edge-colored multigraph containing a vertex
incident to edges of only one color.
It is not hard to solve CPP on undirected and directed multigraphs [11].
For a directed multigraph G, we construct a flow network N by assigning
lower bound 1, upper bound ∞ and cost ω(uv) to each arc uv, where ω(uv) is
the weight of uv in G. A minimum-cost circulation in N viewed as an Euler
directed multigraph corresponds to a CPP solution and vice versa. For an
undirected multigraph G, we construct an edge-weighted complete graph H
whose vertices are odd degree vertices of G and the weight of an edge xy in H
equals the minimum weight path between x and y in G. Now find a minimum-
weight perfect matching M in H and add to G a minimum-weight path of G
between x and y for each edge xy of M . The resulting Euler multigraph
corresponds to a CPP solution and vice versa.
We will prove that CPP-ECG is polynomial-time solvable as well. Note
that our proof is significantly more complicated than that for CPP on undi-
rected and directed graphs. As in the undirected case, we construct an aux-
iliary edge-weighted complete graph H and seek a minimum-weight perfect
matching M in it. However, the construction of H and the arguments justi-
fying the appropriate use of M are significantly more complicated. This can
partially be explained by the fact that CPP-ECG has no solution on many
edge-colored multigraphs.
Note that there is another generalization of CPP on both undirected and
directed multigraphs, namely, CPP on mixed multigraphs, i.e., multigraphs
that may have both edges and arcs. However, CPP on mixed multigraphs
is NP-hard [16]. It is fixed-parameter tractable when parameterized by both
number of edges and arcs [5, 9] and W[1]-hard when parameterized by path-
width [10]. For more information on the classical and parameterized complex-
ity of CPP and its generalizations, see an excellent survey by van Bevern et
al. [5].
2 Preliminaries
Walks. A walk in a multigraph is a sequence W = v1e1v2 . . . vp−1ep−1vp of
alternating vertices and edges such that vertices vi and vi+1 are end-vertices of
edge ei for every i ∈ [p − 1]. A walk W is closed (open, respectively) if v1 = vp
( v1 6= vp, respectively). A trail is a walk in which all edges are distinct.
For technical reasons we will consider walks with fixed end vertices and
call them fixed end-vertex (FEV) walks. Note that an open walk is necessarily
an FEV walk since the end-vertices are predetermined, whereas any vertex
3
in a closed walk can be viewed as its two end-vertices and thus fixing such a
vertex is somewhat similar to assigning a root vertex in a tree. An FEV walk
W = v1e1v2 . . . vp−1ep−1vp is PC in an edge-colored graph if the colors of ei
and ei+1 are different for every i ∈ [p − 2]. Note that we do not require that
colors of ep−1 and e1 are different even if v1 = vp. Thus, a PC FEV walk
might not be a PC walk if v1 = vp.
Let e = xy be an edge in an edge-colored multigraph G. The operation
of double subdivision of e replaces e with an (x, y)-path Pe with three edges
such that the weight of Pe equals that of edge e.
It is easy to see that in our study of PC walks, we may restrict ourselves
to graphs rather than multigraphs. Indeed, it suffices to double subdivide every
parallel edge e and assign the original color of e to the first and third edges of
Pe and a new color to the middle edge.
Finding PC FEV walks. Let R+ denote the set of non-negative real num-
bers. To show a polynomial-time algorithm for CPP-ECG, we will use the
following:
Lemma 1. Let G = (V, E) be a k-edge-colored graph and ω : E → R+ a
weight function. Let vertices u, v ∈ V and edge colors c1, c2 be given, where
we may have u = v. In polynomial time we can find a minimum-weight PC
FEV walk from u to v in G whose first edge has color c1 and whose last edge
has color c2, or conclude that there is no such PC FEV walk in G.
Proof. Define an auxiliary digraph H as follows. Let the vertex set of H be
{(u, 0)} ∪ {(x, i) : x ∈ V, i ∈ [k]}. For every edge xy ∈ E, of color i, we add to
H all arcs (x, j)(y, i) and (y, j)(x, i) where j ∈ [k], j 6= i. We also add an arc
from (u, 0) to (z, c1) for every edge uz ∈ E of color c1. Every arc in H retains
the weight of the corresponding edge in G. We claim that the minimum-weight
PC FEV walk we seek in G corresponds to a minimum-weight directed path
from (u, 0) to (v, c2) in H, which can be found in polynomial time, e.g., using
Dijkstra's algorithm.
On the one hand,
let (x1, d1)(x2, d2) . . . (xℓ, dℓ) be a directed path in
H such that (x1, d1) = (u, 0) and (xℓ, dℓ) = (v, c2). Then by construction,
x1e2x2 . . . eℓxℓ, where ei is an edge between xi−1 and xi of color di, is a PC
FEV walk in G with required properties. On the other hand, consider a
minimum-weight PC FEV walk W in G with the properties requested. Ori-
ent the edges of the walk away from u. We may assume that no vertex has
two in-coming directed edges in the walk of the same color, as the walk could
otherwise be shortened. As above it is not hard to verify that the walk corre-
sponds to a directed path P in H from (u, 0) to (v, c2). By construction, the
weight of P equals that of W . It remains to observe that P is a minimum-
weight directed path from (u, 0) to (v, c2), as otherwise there is a PC FEV
4
u1
u4
u2
u3
color 3
color 2
color 1
Figure 1: 3-edge-colored graph G
walk between u and v with required edge colors of weight smaller than W , a
contradiction.
Finally, we observe that the construction works without modification if
u = v.
3 Main Result
We are now ready to prove the main result.
Theorem 2. We can solve CPP-ECG in polynomial time.
Proof. Let G be an input of CPP-ECG. That is, G is a connected k-edge-
colored graph with at least one edge. We may assume that no vertex of G
is incident with edges of a single color only (G cannot have PC closed walks
through such vertices). We may also assume that k ≥ 2 is odd (if not, we
double subdivide an edge e of G and assign a new color to the middle edge of
Pe and the original color of e to the other two edges).
For a vertex u ∈ V (G) and color i ∈ [k], let di(u) be the number of edges
incident with u of color i. Let d(u) (= Pk
i=1 di(u)) be the degree of u in G. We
say that color i is dominant for u in G if 2di(u) > d(u); note that a vertex has
at most one dominant color, and is balanced if and only if it has no dominant
color.
We now show how to construct, in polynomial time, an undirected graph
H such that H has a perfect matching if and only if G has a PC closed walk
traversing all edges of G, and the minimum weight of a perfect matching in
H is equal to the minimum weight of such a walk in G minus the weight of G.
As computing a minimum-weight perfect matching can be done in polynomial
time, the claim follows. An example is shown in Figures 1 and 2.
We will build the undirected graph H as follows. Define θi(u) as follows:
θi(u) = max{0, d(u) − 2di(u)}.
Let Xi(u) be a set of independent vertices of size θi(u) and let X(u) =
i=1 Xi(u). We now consider the cases when u is balanced and when u is not
Sk
balanced, separately.
5
X3(u1)
X(u2)
artificial edge
non-artificial edge
Y (u1)
X2(u1)
X2(u3)
X(u4)
X3(u3)
Y (u3)
Figure 2: Constructed graph H from the graph of Figure 1. Some edges,
including artificial edges within X(u2) and X(u4), are omitted for clarity.
Case 1: u is not balanced. Let Y (u) be a set of (k − 2)d(u) independent
vertices and add all possible edges between Y (u) and X(u) and all possible
edges within Y (u). Let the weight of all these edges be zero and let us call
them artificial edges. Let Z(u) = X(u) ∪ Y (u).
Case 2: u is balanced. Add all possible edges within X(u). Let the weight
of all these edges be zero and let call them artificial edges. Let Z(u) = X(u).
For every pair a ∈ Xi(u) and b ∈ Xj(v) of distinct vertices such that
ab is not an artificial edge, i, j ∈ [k] and u, v ∈ V (G) (we may have i = j
and/or u = v), add an edge between a and b with the weight equal to that
of a minimum-weight PC FEV walk from u to v in G, starting in color i and
ending in color j if one exists, otherwise add no edge ab; this can be computed
in polynomial time by Lemma 1. This completes the description of H.
Assume that H has a perfect matching and M is a minimum-weight perfect
matching in H. We will show that the weight of M plus the weight of all edges
in G is the weight of an optimal solution to the CPP-ECG instance.
We begin with an observation about the structure of M .
Claim 1. We say that a vertex a ∈ X(u) for some u ∈ V (G) is affected by
M if a is incident with a non-artificial edge in M . Then M has the following
properties.
6
1. If u is unbalanced with dominant color i ∈ [k], then at least 2di(u) − d(u)
vertices of X(u) are affected by M . Furthermore Xi(u) = ∅.
2. If d(u) is odd, then an odd number of vertices of X(u) are affected by
M .
3. If d(u) is even, then an even number of vertices of X(u) are affected by
M .
Furthermore, for any matching M0 in H with no artificial edges that has the
above properties, M0 can be completed to a perfect matching by adding artificial
edges.
Proof. 1. Assume that u is unbalanced with dominant color i, i.e., 2di(u) >
d(u), i ∈ [k]. Then necessarily, 2dj (u) < d(u) for every other color j ∈ [k],
hence θj(u) = d(u) − 2dj (u) if i 6= j, whereas θi(u) = 0. Thus
X(u) = X
j6=i
(d(u) − 2dj (u)) = (k − 2)d(u) + (2di(u) − d(u)),
where the last equality uses Pj6=i d(u) = (k − 1)d(u) and Pj6=i dj(u) = d(u) −
di(u). Artificial edges on X(u) can only match vertices of X(u) against Y (u).
Since Y (u) = (k − 2)d(u) < X(u), this leaves at least X(u) − Y (u) =
2di(u)−d(u) vertices in X(u) which must be affected by M . Finally, Xi(u) = ∅
since θi(u) = 0.
2 and 3. Assume first that u is unbalanced, so there is a set of vertices
Y (u). Since k is odd, the parity of Y (u) matches the parity of d(u), hence
an odd (resp. even) number of vertices of X(u) are matched against Y (u)
if and only if d(u) is odd (resp. even). Now, as calculated in the previous
paragraph, X(u) = (k − 3)d(u) + 2di(u), which is always even. Furthermore,
the affected vertices of X(u) are exactly those not matched against Y (u). The
claim follows.
Finally, if u is balanced, then X(u) = Pj(d(u) − 2dj(u)) = (k − 2)d(u),
which again has the same parity as d(u). Since every artificial edge matches
two vertices of X(u), and the remaining vertices are exactly the affected ver-
tices in X(u), the claim follows.
Completing a non-perfect matching. Let M0 be a matching in H such that
for every vertex u, (1) if u has a dominant color i, then at least 2di(u) − d(u)
vertices of X(u) are affected by M0, and (2) an odd number of vertices of
X(u) are affected by M0 if and only if d(u) is odd. By the above, the second
point here implies that the number of unmatched vertices of Z(u) is even for
every vertex u. If u is balanced, then Z(u) = X(u) is a clique and we can
add artificial edges from the clique. If u is unbalanced, then Y (u) is entirely
unmatched, and by the first point here, the number of unmatched vertices in
X(u) is at most Y (u). Hence the completion is possible.
7
Let u and v be vertices of G and let i and j be colors. Let e = ab be an
arbitrary non-artificial edge in H, with a ∈ Xi(u) and b ∈ Xj(v), where we
may have u = v and/or i = j. An e-walk is a PC FEV walk in G, starting at
u with an edge of color i, and ending at v with an edge of color j.
We will show the theorem in two parts. First, we will show that for any
perfect matching M ′ of H, if we add an e-walk to G for every non-artificial
edge e ∈ M ′, then the resulting multigraph G′
is PC Euler. Here, to add an
e-walk to G means to duplicate every edge along the walk, duplicating an edge
multiple times if it occurs in the walk multiple times. Second, we show that
for any PC Euler graph G′ = (V, E ∪ W ), obtained by duplicating edges in
E, there exists a perfect matching M ′ of H such that W can be decomposed
into a set F of e-walks, where there is an e-walk in F if and only if e is a
non-artificial edge in M ′. This will settle the result.
We need an observation about the effect of adding an e-walk to a graph.
Claim 2. Let e ∈ H be a non-artificial edge. Adding an e-walk F to G has
the following effects.
1. For any vertex u, the parity of d(u) changes if and only if F is open and
u is either its first or last vertex.
2. If u ∈ V (G) is neither the first nor the last vertex of the walk, then for
every i ∈ [k], the value of d(u) − 2di(u) is non-decreasing in the process.
3. If F is closed, let u be its end-vertex, and let i (j, respectively) be the
colors of its first (last, respectively) edges. Then for any c ∈ [k], the value
of d(u) − 2dc(u) increases by at least 2 if c /∈ {i, j}; is non-increasing if
c ∈ {i, j} and i 6= j; and decreases by at most 2 if i = j = c.
4. If F is open, let u be an end-vertex of F , without loss of generality, the
first one. Let i be the color of the first edge. Then for any j ∈ [k], the
value of d(u) − 2dj(u) decreases by at most one if j = i, and increases
by at least one, otherwise.
Proof. The first item is easy. For the second item, we just observe that a
single transition through u increases d(u) by 2 and di(u) by at most 1. Since
the graph has no loops, the local effect on u of duplicating F decomposes into
transitions, hence the second item holds. By the same argument, if u is an
endpoint of F , then all visits to u except the first and/or last one decompose
into transitions. This leaves only the first and last edges of F , and their effects
on the end-vertices of F , to consider. The claims in items 3 and 4 follow by
considering all possibilities for these two edges.
We are now ready to show the first part of the theorem, as announced
above. Let M ′ be an arbitrary perfect matching in H, and let G′ be the result
8
of adding an arbitrary e-walk, which has the same weight as edge e, to G
for every non-artificial edge e ∈ M ′. We will show that G′ is PC Euler. By
Theorem 1, we need to show three conditions: G′ is connected, every vertex
in G′ is even, and every vertex in G′ is balanced. The first condition follows
since G is connected; the second condition follows from Claim 1(2–3) and
Claim 2(1). It remains to show that every vertex is balanced in G′, i.e., for
every u ∈ V (G) and every i ∈ [k], it holds in G′ that d(u) ≥ 2di(u). We break
this down into two cases.
Case 1: i is the dominant color for u in G. In this case, d(u) − 2di(u) < 0
in G, and we need to show that this value is nonnegative in G′. By Claim 1(1),
at least 2di(u)− d(u) vertices of X(u) are affected by M ′, and since Xi(u) = ∅,
Claim 2 gives that the value of d(u) − 2di(u) increases by at least 1 for every
such vertex, and never decreases. Thus d(u) ≥ 2di(u) in G′.
Case 2: i is not a dominant color for u in G.
In this case, Xi(u) =
d(u) − 2di(u) ≥ 0, and we need to show that this value is nonnegative in
G′. By Claim 2, the value of d(u) − 2di(u) decreases by at most as much
as the number of vertices in Xi(u) affected by M ′, in the sense of the term
used in Claim 1. Since there are only d(u) − 2di(u) such vertices, we see that
d(u) ≥ 2di(u) also in G′.
Hence we conclude that d(u) ≥ 2di(u) in G′ for every u ∈ V (G) and every
i ∈ [k], hence every vertex is balanced. This concludes the proof that G′ has a
PC Euler trail. Clearly, the weight of this trail is equal to the total weight of
E(G) plus the sum of the weight of the added e-walks, where the latter part
is exactly the weight of M ′.
Now assume that CPP-ECG on G has a solution, a PC closed walk Q in
G, and let G′ be the graph obtained from G by replacing every edge e = xy
by qe parallel edges with vertices x and y, where qe is the number of times Q
traverses e. Let W = E(G′) \ E(G), i.e., W are the edges that are added to G
in order to get the PC Euler multigraph G′. We will find a perfect matching
in H with total weight at most the sum of the weights of edges in W . This
will complete the proof.
We initially define a set W ′ of walks as the set of one-edge walks xey, where
e = xy ∈ W . We will merge walks in W ′ until we can map the remaining
walks W ′ to a matching M0 in H meeting the requirements of Claim 1, at
which point we will be done. Here to merge two walks is to replace the walks
u1e1u2 . . . eℓ−1uℓ and v1f1v2 . . . fh−1vh, where uℓ = v1 = u, with the walk
u1e1u2 . . . eℓ−1uf1v2 . . . fh−1vh. For u ∈ V (G) and i ∈ [k], let wi(u) denote the
number of times that u is an end-vertex of a walk in W ′ and that walk ends
in u with color i. Here we do not assume a fixed "first" and "last" vertex,
and thus the walks u1e1u2 . . . el−1ul and ulel−1ul−1 . . . e1u1 are the same. Note
that we will allow walks in W ′ to start and end in the same vertex, in which
9
case one walk may contribute to wi(u) twice.
By Claim 1, we need to ensure that
1. wi(u) ≤ θi(u) = Xi(u) for every u ∈ V (G) and i ∈ [k] (so that W ′
corresponds to a matching);
2. Pj∈[k]\{i} wj(u) ≥ 2di(u)−d(u) for every vertex u with a dominant color
i in G; and
3. the parity condition is met for every vertex u.
Because G′ has a PC closed walk traversing all edges, and by Theorem 1,
we have that initially W ′ satisfies the following:
4. Pj∈[k] wj(u) + d(u) ≥ 2wi(u) + 2di(u) for every vertex u ∈ V (G) and
integer i (since G′ is balanced);
5. Pj∈[k] wj(u) is even if and only if d(u) is even i.e. the parity condition
is met (since G′ is even, and hence d(u) + Pj∈[k] wj(u) is even).
We note that Condition 4 implies Condition 2 and Condition 5 implies Con-
dition 3. As long as Condition 1 is not satisfied, we will modify W ′ by merging
walks in such a way that Condition 4 and Condition 5 are still satisfied. As
each merging reduces the number of walks in W ′, we must eventually stop
with a set of walks W ′ satisfying Condition 1, Condition 2 and Condition 3.
So now assume that Condition 1 is not satisfied and let u ∈ V (G) be a
vertex such that wi(u) > θi(u) for some i ∈ [k]. If wi(u) > d(u) − 2di(u), then
we must have that wc(u) > 0 for some c 6= i, as otherwise 2wi(u) + 2di(u) >
wi(u)+d(u) = Pj∈[k] wj(u)+d(u), a contradiction to Condition 4. Thus there
are at least two colors c with wc(u) > 0.
We will choose two colors h, j, with wh(u) > 0, wj (u) > 0, h 6= j (i is not
necessarily in {j, h}), and merge a walk ending at u with color h with a walk
ending at u with color j.
(If this makes us merge both endvertices of the
same walk, we may simply remove the walk). It is clear that the new walk is
still PC, and this operation reduces the number of walks in W ′. As we have
reduced wh(u) and wj(u) by 1, and the other values are unaffected, it is clear
that Condition 5 is still satisfied. We now show how to choose h, j in such a
way that Condition 4 is still satisfied.
Let us call a color c at risk if Pj∈[k] wj(u) + d(u) ≤ 2wc(u) + 2dc(u) + 1
(informally, a color is "at risk" if removing two edges of other colors would
lead that color to dominate u). As Pj∈[k] wj(u) + d(u) is necessarily even and
Pj∈[k] wj(u) + d(u) ≥ 2wc(u) + 2dc(u) , we have that in fact Pj∈[k] wj(u) +
d(u) = 2wc(u) + 2dc(u) for any at risk color c. Furthermore, we note that
10
at most two colors in [k] can be at risk. Indeed, suppose that distinct colors
c1, c2, c3 ∈ [k] are at risk. Then 2wc1(u)+2dc1 (u)+2wc2 (u)+2dc2 (u)+2wc3 (u)+
2dc3(u) = 3(Pj∈[k] wj(u) + d(u)) > 2(Pj∈[k] wj(u) + d(u)) ≥ 2(wc1 (u) +
wc2(u) + wc3(u) + dc1(u) + dc2(u) + dc3(u)), a contradiction.
Next, suppose for a contradiction that wc(u) = 0 for an at risk color
c. Then 2dc(u) = d(u) + Pj∈[k] wj(u) and so 2di(u) ≤ 2d(u) − 2dc(u) =
d(u) − Pj∈[k] wj(u). Then wi(u) > θi(u) ≥ d(u) − 2di(u) ≥ d(u) − d(u) +
Pk∈[k] wj(u) ≥ wi(u), a contradiction. Thus, wc(u) > 0 for any at risk color
c.
We now know that there at least two colors c with wc(u) > 0, there are at
most 2 at risk colors, and if color c is at risk then wc(u) > 0. We can therefore
select two distinct colors h, j with wh(u) > 0, wj(u) > 0, such that any at
risk color is contained in {h, j}. We now merge a walk ending with color h
at u and a walk ending with color j at u, as described above. This has the
effect of reducing each of wh(u) and wj(u) by 1, and leaving wc(u) unchanged
for c ∈ [k] \ {h, j}. We now show that we still have Pj∈[k] wj(u) + d(u) ≥
2wc(u)+2dc(u) for any c ∈ [k]. If c ∈ {h, j}, then both Pj∈[k] wj(u)+d(u) and
2wc(u)+2dc(u) are reduced by 2, so the condition still holds. If c /∈ {h, j}, then
as c was not at risk we originally had Pj∈[k] wj(u)+d(u) ≥ 2wc(u)+2dc(u)+2.
As Pj∈[k] wj(u)+ d(u) is reduced by 2, we will still have Pj∈[k] wj(u)+ d(u) ≥
2wc(u) + 2dc(u), as required.
We continue the above process until there is no u ∈ V (G), i ∈ [k] for which
Condition 1 fails. We therefore have that Conditions 1, 4 and 3 hold, which
in turn implies Conditions 2 and 3 hold. Convert W ′ to a matching M0 in H
by adding for every walk F an edge e to M0 such that F is an e-walk. This is
possible since wi(u) ≤ θi(u) = Xi(u) for every i ∈ [k], u ∈ V (G). The weight
of M0 is at most the weight of W , since every edge e added to M0 this way
has a weight corresponding to a minimum-weight e-walk where F is just one
possible e-walk. By Claim 1 we can complete M0 to a perfect matching M ′ by
adding artificial edges, which does not increase the weight. Hence H admits
a perfect matching whose weight is at most the weight of W .
So in all cases we can find a perfect matching in H with weight exactly
the weight of all the (non-closed) walks in W ′. As we have already shown
that a perfect matching in H gives rise to a solution to CPP-ECG on G where
duplicated edges add the same weight as the weight of the matching, we are
done.
4 Conclusion
We considered the Chinese Postman Problem on edge-colored graphs (CPP-
ECG). This problem generalizes the Chinese Postman Problem on undirected
11
and directed graphs and the properly colored Euler trail problem on edge-
colored graphs, all of which can be solved in polynomial time. We proved that
CPP-ECG is still polynomial time solvable.
It is well-known that the number of Euler trails on digraphs can be cal-
culated in polynomial time using the so-called BEST theorem [23, 1], named
after de Bruijn, van Aardenne-Ehrenfest, Smith and Tutte. However, the
problem is much harder on undirected graphs as it is #P-complete [24]. Our
simple transformation from directed walks to PC walks described in Section 1,
shows that the problem of counting PC Euler trails on 2-edge-colored graphs
generalizes that of counting the number of Euler trails on digraphs. Assigning
each edge of an undirected graph a distinct color, shows that the problem
of counting PC Euler trails on k-edge-colored graphs is #P-complete when
k is unbounded. So it would be interesting to determine the complexity of
the problem of counting PC Euler trails on a k-edge-colored graphs when k is
bounded, in particular when k = 2.
Acknowledgments. We are thankful to the referees for careful reading of the
paper and several useful suggestions. Research of GG was partially supported
by Royal Society Wolfson Research Merit Award. Research of BS was partially
supported by China Scholarship Council.
References
[1] T. van Aardenne-Ehrenfest and N. G. de Bruijn, Circuits and trees in
oriented linear graphs, Simon Stevin 28 (1951) 203–217.
[2] A. Abouelaoualim, K. Ch. Das, W. Fernandez de la Vega, M. Karpinski,
Y. Manoussakis, C. A. Martinhon, and R. Saad, Cycles and paths in edge-
colored graphs with given degrees, J. Graph Theory 64 (1) (2010) 63–86.
[3] S.K. Ahuja, Algorithms for Routing and Channel Assignment in Wireless
Infrastructure Networks, PhD thesis, Univ. Arizona, 2010.
[4] J. Bang-Jensen and G. Gutin, Digraphs: Theory, Algorithms and Applica-
tions, 2nd Ed., Springer-Verlag, London, 2009.
[5] R. van Bevern, R. Niedermeier, M. Sorge and M. Weller. Complexity of Arc
Routing Problems. In Arc Routing: Problems, Methods and Applications
(A. Corber´an and G. Laporte, eds.), SIAM, 2014.
[6] A. Benkouar, Y. Manoussakis, V. Paschos and R. Saad, On the complexity
of finding alternating hamiltonian and Eulerian cycles in edge-coloured
graphs. Lect. Notes Comp. Sci. 557 (1991) 190–198.
12
[7] H. Fleischner, G. Sabidussi, and E. Wenger, Transforming Eulerian trails.
Discrete Math. 109(1) (1992)103–116.
[8] S. Fujita and C. Magnant, Properly colored paths and cycles. Discrete
Appl. Math. 159 (14) (2011) 1391–1397.
[9] G. Gutin, M. Jones and B. Sheng, Parameterized Complexity of the k-Arc
Chinese Postman Problem. Lect. Notes Comput. Sci. 8737 (2014) 530–541.
[10] G. Gutin, M. Jones and M. Wahlstrom, Structural Parameterizations
of the Mixed Chinese Postman Problem. Lect. Notes Comput. Sci. 9294
(2015) 668–679.
[11] B. Korte and J. Vygen, Combinatorial Optimization: Theory and Algo-
rithms. 3rd Ed., Springer, 2005.
[12] A. Kotzig, Moves without forbidden transitions in a graph, Math. Fyz.
Cazopis 18 (1968) 76–80.
[13] A. Lo, Properly coloured Hamiltonian cycles in edge-coloured complete
graphs. Combinatorica, doi: 10.1007/s00493-015-3067-1 (online).
[14] A. Lo, A Dirac type condition for properly coloured paths and cycles, J.
Graph Theory 76 (1) (2014) 60–87.
[15] A. Lo, An edge-coloured version of Dirac's theorem, SIAM J. Discrete
Math. 28 (1) (2014) 18–36.
[16] C. H. Papadimitriou. On the complexity of edge traversing. J. ACM 23
(1976) 544–554.
[17] P.A. Pevzner, DNA physical mapping and alternating Eulerian cycles in
colored graphs. Algorithmica 13 (1995) 77–105.
[18] P.A. Pevzner, Computational Molecular Biology: an algorithmic ap-
proach. MIT Press (2000).
[19] S Sankararaman, A Efrat,
S Ramasubramanian, On channel-
discontinuity-constraint routing in wireless networks, Ad Hoc Networks
134 (2014) 153–169.
[20] M. Szachniuk, M. Popenda, R.W. Adamiak and J. Blazewicz, An assign-
ment walk through 3D NMR spectrum. in: Proc. 2009 IEEE Symposium
on Computational Intelligence in Bioinformatics and Computational Biol-
ogy (2009) 215–219.
13
[21] M. Szachniuk, M.C. De Cola and G. Felici, The orderly colored longest
path problem – a survey of applications and new algorithms. RAIRO-Oper.
Res. 48 (2014) 25–51.
[22] I.L. Tseng, H.W. Chen and C.I. Lee, Obstacle-aware longest path routing
with parallel MILP solvers. in: Proc WCECS-ICCS vol. 2 (2010) 827–831.
[23] W. T. Tutte and C. A. B. Smith, On unicursal paths in a network of
degree 4, American Mathematical Monthly 48 (1994) 233–237.
[24] G. R. Brightwell and P. Winkler, Note on Counting Eulerian Circuits.
CDAM Research Report LSE-CDAM-2004-12, 2004.
14
|
1508.05143 | 2 | 1508 | 2016-04-05T21:00:07 | A Discrete and Bounded Envy-free Cake Cutting Protocol for Four Agents | [
"cs.DS",
"cs.GT"
] | We consider the well-studied cake cutting problem in which the goal is to identify a fair allocation based on a minimal number of queries from the agents. The problem has attracted considerable attention within various branches of computer science, mathematics, and economics. Although, the elegant Selfridge-Conway envy-free protocol for three agents has been known since 1960, it has been a major open problem for the last fifty years to obtain a bounded envy-free protocol for more than three agents. We propose a discrete and bounded envy-free protocol for four agents. | cs.DS | cs | A Discrete and Bounded Envy-Free Cake Cutting Protocol
for Four Agents
Haris Aziz
Simon Mackenzie
Data61 and UNSW
Sydney, Australia
{haris.aziz, simon.mackenzie}@data61.csiro.au
6
1
0
2
r
p
A
5
]
S
D
.
s
c
[
2
v
3
4
1
5
0
.
8
0
5
1
:
v
i
X
r
a
ABSTRACT
We consider the well-studied cake cutting problem in which
the goal is to identify an envy-free allocation based on a min-
imal number of queries from the agents. The problem has
attracted considerable attention within various branches of
computer science, mathematics, and economics. Although,
the elegant Selfridge-Conway envy-free protocol for three
agents has been known since 1960, it has been a major open
problem to obtain a bounded envy-free protocol for more
than three agents. The problem has been termed the cen-
tral open problem in cake cutting. We solve this problem
by proposing a discrete and bounded envy-free protocol for
four agents.
Categories and Subject Descriptors
F.2 [Theory of Computation]: Analysis of Algorithms
and Problem Complexity; I.2.11 [Distributed Artificial
Intelligence]: Multiagent Systems; J.4 [Computer Ap-
plications]: Social and Behavioral Sciences - Economics
General Terms
Algorithms, Theory, Economics
Keywords
Fair Division, Elicitation Protocols, Multiagent Resource Al-
location, Cake cutting
1.
INTRODUCTION
Cake cutting is a metaphor for the allocation of a het-
erogeneous divisible good among multiple agents with pos-
sibly different preferences over different parts of the cake.
Its main application is fair scheduling, resource allocation,
and conflict resolution [13] and hence it has been exten-
sively studied within computer science [27] and the social
sciences [36]. Since various important divisible resources
such as time and land can be captured by cake cutting, the
problem of fairly dividing the cake is a fundamental one
within the area of fair division and multiagent resource al-
location [6, 17, 26, 29, 33, 35, 36].
Formally speaking, a cake is represented by an interval
[0, 1] and each of the n agents has a value function over
pieces of the cake that specifies how much that agent val-
ues a particular subinterval. The main aim is to divide the
cake fairly. In particular, an allocation should be envy-free
so that no agent prefers to take another agent's allocation
instead of his own allocation. Although an envy-free alloca-
tion is guaranteed to exist even with n − 1 cuts [35]1, finding
an envy-free allocation is a challenging problem which has
been termed "one of the most important open problems in
20th century mathematics" by Garfunkel [16].
Motivation and Contribution.
Unlike allocation of indivisible items [12], the number of
possible allocations in cake cutting is infinite. Since the valu-
ations of agents over the subsets of the cake can be complex,
it is not practiceable to elicit each agent's complete valua-
tions function over the cake. A natural approach in cake
cutting protocols to query agents about their valuations of
different portions of the cake and based on these queries,
propose an allocation. A cake cutting protocol is envy-free
if each agent is guaranteed an envy-free piece if he reports
his real valuations.
For the case of two agents, the problem has a well-known
solution in the form of the Divide and Choose protocol: one
agent is asked to cut the cake into equally preferred pieces
and the other agent is asked to choose the preferred piece.
The protocol even features in the Book of Genesis (Chap-
ter 13) where Abraham divides the land of Canaan and Lot
chooses first. In modern times, the protocol has been en-
shrined in the Convention of the Law of the Sea (Page 10,
[6]). For the case of three agents, an elegant and bounded
protocol was independently discovered by John L. Selfridge
and John H. Conway around 1960 (Page 116, [6]). Since
then, an efficient envy-free protocol for four or more agents
has eluded mathematicians, economists, and computer sci-
entists.
In 1995, Brams and Taylor [5] made a breakthrough
by presenting an envy-free protocol
for any number of
agents [18]. Although the protocol is guaranteed to ter-
minate in finite time, there is one critical drawback of the
protocol: the running time or number of queries and even
the number of cuts required is unbounded even for four
1The existence of an envy-free cake allocation can be shown
via an interesting connection with Sperner's Lemma [35].
agents. In other words, the number of queries required to
identify an envy-free allocation can be arbitrarily large for
certain valuations functions. If a protocol is not bounded,
then its practicality is compromised [21]. Procaccia [27]
terms unboundedness as a "serious flaw". Brams and Taylor
were cognizant of their protocol's drawback and explicitly
mentioned the problem of proposing a bounded envy-free
protocol even for n = 4. Lindner and Rothe [21] write
that "even for n = 4, the development of finite bounded
envy-free cake-cutting protocols still appears to be out of
reach, and a big challenge for future research." The prob-
lem has remained open and has been highlighted in several
works [2, 5, 6, 10, 14, 31, 19, 26, 27, 22, 29, 30]. Saberi and
Wang [30] term the problem as "one of the most important
open problems in the field" and Lindner and Rothe [22] men-
tion the case for n = 4 as "the central open problem in the
field of cake-cutting". In this paper, we present a discrete
envy-free protocol for four agents that requires a bounded
number of queries as well as cuts of the cake. The maximum
number of cuts required is 203.2 Some of the techniques we
use may be useful for cake cutting protocols with other prop-
erties or for more agents. In particular, we propose a new
technique (called permutation) in which by suitably reallo-
cating portions of a partial allocation that is envy-free, we
ensure that some agent will not be envious of another agent
even if the unallocated cake is given to the latter agent.
Related Work.
Cake cutting problems originated in the 1940's when fa-
mous mathematicians such as Banach, Knaster, and Stein-
haus initiated serious mathematical work on the topic of fair
division.3 Since then, the theory of cake cutting algorithms
has become a full-fledged field with at least three books writ-
ten on the topic [3, 6, 29]. The central problem within cake
cutting is finding an envy-free allocation [15, 33].
Since the earliest works, mathematicians have been inter-
ested in the complexity of cake cutting. Steinhaus [32] wrote
that "Interesting mathematical problems arise if we are to
determine the minimal number of cuts necessary for fair di-
vision." When formulating efficient cake cutting protocols,
a typical goal is to minimize the number of cuts while ig-
noring the number of valuations queried from the agents. In
principle, the actual complexity of a problem or a protocol
depends on the number of queries. When considering how
efficient a protocol is, it is useful to have a formal query
model for cake-cutting protocols. Robertson and Webb [29]
formalized a simple query model in which there are two kinds
of queries: Evaluate and Cut.
In an Evaluate query,
an agent is asked how much he values a subinterval. In a
Cut query, an agent is asked to identify an interval, with
a fixed left endpoint, of a particular value. Although, the
query model of Robertson and Webb is very simple, it is
general enough to capture all known protocols in the liter-
ature. Note that if the number of queries is bounded, it
implies that the number of cuts is bounded in the Robert-
son and Webb model. The protocol that we present in this
2Most of the 203 cuts are technically trims but in the
Robertson and Webb model, any marking/trim on the cake
is also treated as a proper cut.
3Hugo Steinhaus presented the cake cutting problems to the
mathematical and social science communities on Sep. 17,
1947, at a meeting of the Econometric Society in Washing-
ton, D.C. [29, 32].
paper uses a bounded number of queries in the Robertson
and Webb model. Cake cutting protocols also provide an in-
teresting connection between the literature on fair division
and the field of communication complexity [20].
There is not too much known about the existence of a
bounded envy-free protocol for n ≥ 4 except that any envy-
free cake-cutting algorithm requires Ω(n2) queries in the
Robertson-Webb model [25, 27]. Also, for n ≥ 3, there ex-
ists no finite envy-free cake-cutting algorithm that outputs
contiguous allocations [34]. Brams et al. [7] and Barbanel
and Brams [4] presented envy-free protocols for four agents
that require 13 and 5 cuts respectively. However, the proto-
cols are not only unbounded but not even finite since they
are continuous protocols that require the notion of a mov-
ing knife. An alternative approach is to consider known
bounded protocols and see how well they perform in terms
of envy-freeness [21]. Apart from the unbounded Brams and
Taylor envy-free protocol for n agents, there are other gen-
eral envy-free protocols by Robertson and Webb [28] and
Pikhurko [24] that are also unbounded.
There are positive algorithmic results concerning envy-
free cake cutting when agents have restricted valuations
functions [13, 8] or when some part of the cake is left unallo-
cated [30]. There has also been work on strategyproof cake
cutting protocols for restricted valuation functions [1, 11, 23]
as well as strategic aspects of protocols [9].
Structure of the Paper.
In Section 2, the formal model is presented. In Section 3,
we present an envy-free protocol for three agents that serves
as a warm-up for the case of four agents.
In Section 4,
we presents the main protocol. The section is divided into
subsections in which three different protocols (Post Dou-
ble Domination Protocol, Core Protocol, and Permutation
Protocol ) are described. These three protocols are used as
building blocks to formulate the overall protocol.
2. PRELIMINARIES
Model.
We consider a cake which is represented by the interval
[0, 1]. A piece of cake is a finite union of disjoint subsets
of [0, 1]. We will make the standard assumptions in cake
cutting. Each agent in the set of agents N = {1, . . . , n} has
his own valuation function over subsets of interval [0, 1]. The
valuations are (i) defined on all finite unions of the intervals;
(ii) non-negative: Vi(X) ≥ 0 for all X ⊆ [0, 1]; (iii) additive:
for all disjoint X, X ′ ⊆ [0, 1], Vi(X ∪ X ′) = Vi(X) + Vi(X ′);
(iv) divisible i.e., for every X ⊆ [0, 1] and 0 ≤ λ ≤ 1, there
exists X ′ ⊆ X with Vi(X ′) = λVi(X).
We will call an allocation partial if there is some cake that
is unallocated. A partial envy-free allocation is a partial allo-
cation that is envy-free. In order to ascertain the complexity
of a protocol, Robertson and Webb presented a computa-
tional framework in which agents are allowed to make two
kinds of queries: (1) for given x ∈ [0, 1] and r ∈ R+, Cut
query asks an agent to return a point y ∈ [0, 1] such that
Vi([x, y]) = r (2) for given x, y ∈ [0, 1], Evaluate query
ask agent to return a value r ∈ R+ such that Vi([x, y]) = r.
A cake-cutting protocol specifies how agents interact with
queries and cuts. A protocol is envy-free if no agent is en-
vious if he follows the protocol truthfully. All well-known
1
2
3
1
2
γ
β
2
1
3
Initially allocated cake
Figure 1: Example of a trim. Agents 1 and 2 trim
their most preferred piece (the left most piece) to
the value equal to that of their second most pre-
ferred piece. In this instance, agent 2 trims more
than agent 1, hence his trim is to the right of agent
1's trim. Let us assume agent 1 and 3 each get a
complete piece with 1 getting his second most pre-
ferred piece. Agent 2 is not envious of other agents
if he gets the right side of the trimmed piece up till
his trim.
If agent 2 gets the part to the right of
agent 1's trim instead of his own trim, then he is
even happier. We will refer to this extra bit γ as the
'bonus' for agent 2. Agent 1 is still not envious of
agent 2 if 2 gets the bonus.
cake cutting protocols can be analyzed in terms of number
of queries required to return a fair allocation. A cake cut-
ting protocol is bounded if the number of queries required to
return a solution is bounded by a function of n irrespective
of the valuations of the agents.
Terms and Conventions.
We now define some terms and conventions that we will
use in the paper. Given a partial envy-free allocation of
the cake and an unallocated residue β, we say that agent j
dominates agent i if j does not become envious of i even if
all of β were to be allocated to i. This concept has been
referred to as i's irrevocable advantage in the cake cutting
literature [6].
In the cake cutting protocols, we will describe, an agent
may be asked to trim a piece of cake so that its value equals
the value of a less valuable piece. Agents will be asked to
trim various pieces of the cake so their remaining value is
equal to the value of the third (or in some cases their second)
most preferred complete piece. In Figure 1, we outline the
idea of trimming a piece to equal the value of some other
piece. When an agent trims a piece of cake, he will trim
it from the left side: the main piece (albeit trimmed) will
be on the the right side. The piece minus the trim will
be called the partial main piece. The remainder will be
referred to as the residue.
If an agent trims a piece, we
say he is competing for the piece. When we say an agent is
guaranteed to get his second/third/etc most favoured piece,
this guarantee is based on the ordinal preferences of the
agents over the pieces. By ordinal, we mean that agents
simply give a weak ordering over the pieces but do not tell
the exact cardinal utility difference between two pieces. If
an agent is indifferent between the top three pieces, then we
will still say that the agent is guaranteed to get his third
most valued piece.
We introduce notation to represent which agents have
trimmed which pieces. So for example 123123 represents
the scenario where one piece has three trims (by agents 1,2,3)
in any possible order and the other three pieces have one trim
each by one of the agents 1, 2, 3. We may enrich this no-
tation further as follows: 112131123 which represents that
all agents think that the piece with the three trim marks is
their most preferred or equivalently highest valued piece.
3. PROTOCOL FOR THREE AGENTS
We warm-up by presenting a protocol (Algorithm 1) for
Algorithm 1 Envy-free Protocol for 3 Agents.
1: Agent 3 divides the cake into 3 equally preferred pieces.
2: if 1 and 2 can each be given a different complete most
preferred piece then
3:
give that complete piece to the agent who prefers it
the most and give remaining piece to 3 and return.
4: else
5:
1 and 2 trim their highest valued piece from the left
side to make the right side of the trim equal to the value
of second most preferred piece (they simultaneously put
trim marks).
6: end if
7: if 1 and 2 trim the same piece then
8:
consider β1, the remainder from the left extreme of
the piece to the leftmost trim. The partial piece P 1
1
which is the most preferred piece except β1 is given to
the agent i ∈ {1, 2} who trimmed the piece more (let the
other agent be −i). Let γ1 be the part between the two
trims of agent 1 and 2. Agent −i gets his second highest
valued complete piece P 1
2 . Agent 3 gets the remaining
complete piece. The unallocated cake is β1.
′
9: end if
10: 3 cuts β1 into 3 equally preferred pieces.
11: 1 and 2 trim their highest valued piece from the left side
to make it equal to the value of second best.
12: if 1 and 2 can be each be given a different complete
most preferred piece then
13:
give that complete piece to the agent who prefers it
the most and give remaining piece to 3 and return.
14: else if 1 and 2 trim the same piece but −i trims at least
as much as i then,
15:
give −i the most preferred piece up till the leftmost
trim, give i a complete second most preferred piece, and
give 3 the remaining complete piece.
16: else if 1 and 2 trim the same piece but i trims more
again then,
17:
18:
′
let β2 be the remainder from the left hand side to the
first trim in β1. The partial piece P 2
which is the most
preferred piece except β2 is given to the agent i. Let
γ2 be the part between the two trims. Agent −i gets
his second highest valued complete piece P 2
2 . Agent 3
gets the remaining complete piece. The only unallocated
cake left if any is β2.
1
Since i again got a partial piece, i is asked to identify
to
the lesser preferred γj ∈ {γ1, γ2}. Then i gives P j
agent −i and gets P j
2 in return.
1
′
19: end if
20: If some cake is still unallocated, Agent 1 and 2 perform
Divide and Choose to allocate it.
β1
β2
2
1
3
3
2
1
3
Figure 2: Protocol for 3 agent: In case agent 2 gets
a trimmed piece two times, then we need to perform
a permutation so that 1 gets a trimmed piece.
n = 3 which we will extend to n = 4.
Although the protocol requires more cuts than the
Selfridge-Conway protocol, it bears similarities with it. It
also depends on some ideas that we will exploit for our main
protocol for four agents. The main idea of the protocol is
that in each step, the cutter (agent 3) cuts the unallocated
cake into 3 equally preferred pieces and gets one of the com-
plete pieces. In each step, a partial envy-free allocation is
maintained and then the remainder is again allocated which
results in the remainder being fully allocated or a smaller re-
mainder left. Note that when i 6= 3 is given the partial cake
piece, agent 3 dominates i. When the remainder β1 in the
first step is divided among the agents, if now 3 dominates
−i, then 3 does not care how β1 is divided among i and −i
since he dominates both. So 3 is in this sense 'eliminated'
from the protocol and we can perform Divide and Choose
for 1 and 2 on the unallocated cake. If 3 again dominates
i based on how β1 is allocated, then we enforce a permu-
tation or reallocation of some pieces of i and −i, so that 3
dominates −i (see Figure 2). Both the ideas of domination
and permutation will feature prominently in our protocol for
four agents.
4. PROTOCOL FOR FOUR AGENTS
We now extend the ideas for the case of n = 3 to n = 4.
The general flavour of the protocol is similar to that of the
protocol for 3 agents. A designated cutter is asked to cut
into equally preferred pieces. Based on some finer steps,
we are able to achieve an envy-free allocation with possibly
some cake still unallocated. We repeat the process on the
remaining unallocated piece with the goal that the cutter
dominates other agents just as we managed in the protocol
for 3 agents. A few additional complications are introduced
when dealing with 4 agents. Eliminating an agent would
require being able to ensure that we can make an agent
dominate all 3 others. Thankfully this is not necessary: we
can show that we only need a protocol that ensures a given
agent dominates 2 others. This is proved in the Double
Domination Lemma.
4.1 Post Double Domination Protocol
We now present the Post Double Domination Protocol
(Algorithm 2) that takes as input a partial envy-free alloca-
tion in which each agent dominates two agents and it returns
a complete envy-free allocation.
Algorithm 2 Post Double Domination Protocol
Agents
Input: A partial envy-free allocation and unallocated
for 4
cake such that each agent dominates 2 other agents
Output: Envy-free complete allocation.
1: There exists some agent 1 who dominates two other
agents say 3 and 4.
2: if 2 also dominates 3 and 4 then
3:
3 and 4 can divide the remainder by Divide and
Choose.
4: else if 2 does not dominate 3 and 4 then
5:
it dominates 1 and one of 3 and 4 say 4.
6:
if 3 dominates 4 then
7:
give all the remainder to 4 (since everyone domi-
nates 4).
8:
9:
else if 3 does not dominate 4 and hence dominates 1
and 2 then
then let 4 cut the residue into four equally preferred
pieces, and agents 1, 2, 3 pick their most preferred re-
maining piece in that order.
end if
10:
11: end if
12: return Envy-free complete allocation.
3
1
4
2
Figure 3: The domination graph of the final case in
the proof of the Double Domination Lemma.
Lemma 1
(Double Domination Lemma). Suppose
we have a bounded protocol which given a specified agent i
and an unallocated piece of cake returns a partial envy-free
allocation such that i dominates 2 other agents, then we
can extend this into a 4 agent envy-free bounded protocol.
Proof. If we have a bounded protocol in which one agent
can be made to dominate two other agents, then we simply
run it at most 4 times on any unallocated cake to ensure
that each agent dominates two other agents. If while doing
this, the cake is completely allocated, we are already done.
Otherwise, we can run the Post Double Domination Protocol
(Algorithm 2). We now argue for the correctness of the
protocol.
Assume 1 dominates 3 and 4. Now if 2 also dominates
3 and 4, then there exists a partial envy-free allocation in
which even if all the residue is given to 3 or 4, then agent
1 and 2 will not be envious. All the residue can be divided
among 3 and 4 using divide and choose. Agents 1 and 2
don't care because they dominate 3 and 4.
The other case is when 2 does not dominate 3 and 4.
Without loss of generality assume that 2 dominates 1 and
4. Now if 3 dominates 4, we are already done because the
whole residue can be given to 4 since everyone dominates
4. If 3 does not dominate 4 but dominates 1 and 2, then
the domination graph looks like in Figure 3, an envy-free
allocation can be found via the following method: agent
4 cuts the residue into equally preferred four pieces, and
agents 1, 2, 3 pick their most preferred remaining piece in
that order. Agent 2 dominates 1 so does not care if 1 chooses
first; agent 3 dominates 2 and 1 so that he does not care if
1 and 2 choose before him.
Since we have shown that making an agent dominate two
other agents is helpful, we will now explain how to achieve
it. The overall protocol will first achieve double domination
for each agent and if some cake is still unallocated, it will use
the Post Double Domination Protocol. In order to get an
agent to dominate other agents, we will repeat a core proto-
col multiple times which gives a partial envy-free allocation.
The core protocol is explained in the next section.
4.2 Core Protocol
In the core protocol, a specified agent is asked to cut the
cake into equally preferred pieces. The cutter gets a com-
plete piece whereas the other agents may get partial pieces.
The core protocol is recursively applied to the unallocated
cake. After a bounded number of calls of the core proto-
col, we are in a position to do some reallocation so as to
ensure that the specified cutter dominates two agents. We
can then repeat this for another specified agent until each
agent dominates two other agents.
We now give a high level description of the core protocol.
Let us say that agent 4 divides the unallocated cake into 4
pieces. Agents 1, 2, and 3 are asked to trim the left hand
side of their most preferred two pieces to make the right side
of their trim equally valuable as their third most preferred
cake piece. Each agent in set {1, 2, 3} trims at most two
pieces. In case an agent is indifferent between two or more
pieces, we will still assume that the agent trims one piece to
make it equal to the other piece. The trim in this extreme
case is a trivial trim that coincides with the left edge of the
cake. Hence the four pieces have a total of six trims. If an
agent who trims a piece most is given that piece (an agent
who trims most two pieces can choose which piece to get),
up to his trim point, then the agent is envy-free.
In fact
the agent is not envious even if each other agent who gets
a piece is given the piece up till the second rightmost trim.
This approach is useful to get a partial envy-free allocation
with some cake unallocated.
In the core protocol, we do
some extra work so that apart from the cutter, at least one
more agent is given a complete piece. In order to do this, in
a couple of cases, we may ask one or two carefully identified
agents to additionally trim their most preferred piece to the
value of their second most preferred piece in which case the
previous trims of these agents are ignored. This is helpful is
ensuring that at least two agents get complete pieces. It may
be the case that some cake is left unallocated in which case
the core protocol may be implemented on the unallocated
cake again. The main thing we will prove is that we do not
need to implement the core protocol on the unallocated cake
unbounded number of times. For this, we first show in the
Core Protocol Lemma that the core protocol returns a par-
tial envy-free allocation with the additional useful property
that the cutter and one other agent get complete pieces.
Lemma 2
(Core Protocol Lemma). For n = 4,
there exists a discrete and bounded protocol which returns
partial envy-free allocation in which one agent cuts the cake
into four equally preferred pieces and the cutter as well as
at least one other agent gets one of these four pieces.
Algorithm 3 Core Protocol for 4 agents that returns a
partial envy-free allocation
Input:
Output: Partial envy-free allocation.
1: Agent 4 is asked to cut the cake into 4 equal value pieces.
2: Agents 1, 2, 3 are asked to give their values for the 4
Specified cutter agent (say agent 4)
pieces.
3: if each agent in {1, 2, 3} can be given a most preferred
piece then
4: Allocate each agent in {1, 2, 3} a most preferred (com-
plete) piece and the remaining complete piece to the
cutter (agent 4).
return the envy-free allocation.
5:
6: end if
7: Agents 1, 2, 3 are asked to trim the left hand side of
their most and second most preferred pieces to make
them (the right side of the trim) equally valuable as
their third most preferred cake piece.
8: if no piece has exactly one trim then
9:
if we are in a case ijjkik where {i, j, k} = {1, 2, 3}.
then
10:
11:
12:
Since we have already covered the case in which
each agent can be given a complete most preferred piece,
we end up in situation i1j1j2ki2k. Without loss of
generality, the case is i1j1j2k1i2k2. In this case i and
k are asked to trim their most preferred piece to their
second most preferred piece whereas agent j is asked
to trim his two most preferred pieces to equal his third
most preferred piece. Agent i and k's trims up to their
third most preferred piece are ignored. The effective
trims look as follows now: ijjk.
if j does not have the rightmost trim in both pieces
then the right side of each piece with two trims is given
to the agent who trimmed it the most. The piece is given
up till the second rightmost trim. The remaining agent
picks his most preferred complete unallocated piece and
then 4 get the remaining unallocated piece.
else if
j trimmed both the pieces the most then
j can choose which piece (up till the second rightmost
trim) to get. The other piece with the trims is given
to the agent with the second rightmost trim up till the
second rightmost trim. The third non-cutter i or k gets
his second most preferred piece completely. The last
unallocated complete piece is given to agent 4.
end if
13:
14:
15: else if we are in a case ijkijk where {i, j, k} =
end if
{1, 2, 3}. then
16:
Ask each agent in {1, 2, 3} to trim this first and second
most preferred piece from the left side to make the right
hand side to the value of his third most preferred piece.
Each agent is given the piece he trims the most up till
the second rightmost trim.
If the same agent has the
rightmost trims for both pieces, he chooses which piece
to get. The agent with second rightmost trim gets the
other piece up till the second rightmost trim. The third
non-cutter gets his third most piece completely. The last
unallocated complete piece is given to agent 4.
17: end if {Continued on next page...}
Proof. We argue that when agent 4 cuts the cake into
equally preferred piece and then these pieces are partially
allocated to the agents, then (1) the partial allocation is
envy-free, and (2) the cutter and one other agent get com-
plete pieces. If each agent in {1, 2, 3} can be given a most
preferred piece, then both conditions are trivially met. Oth-
erwise, the algorithm distinguishes between the following
cases: (1) no piece has exactly one trim; (2) exactly one
piece has exactly one trim; (3) exactly two pieces have ex-
actly one trim; (4) exactly three pieces have exactly one
trim. In each of the cases, one non-cutter gets a complete
piece and the cutter is also given an unallocated complete
piece.
It remains to be shown that the partial allocation is envy-
free. When an agent i gets a (possibly partial) piece a, he
was the one who trimmed that piece the most. For each
other piece b that is allocated, some other agent j trimmed
b at least as much as agent i, i.e., j's trim in b was not left
of i's trim. Hence i is not envious of j if j gets b up till j's
trim from the right hand side. The reason is that i thinks
that j's piece has more value than i's allocation only if j
gets the right side of b beyond the trim of agent i. Thus, if
i has the second rightmost trim in b, then i is not envious
of j even if j gets the right side of b up till i's trim (e.g.,
see Figure 4). We note that in each of the four cases, when
an agent i get a piece, he is not envious of another agent
j because of the reason above. Moreover, if in piece a, the
second rightmost trim is strictly to the left of i's trim in a or
if there is no other trim in a, then i not only gets the right
hand side of a up to i's trim but an additional bonus up till
the second rightmost trim or the edge of the cake (whichever
comes first). Hence, no agent i is envious of another agent
j.
Remark 1. As soon as the agents' ordinal ranking of the
four pieces cut by the cutter are known, it can be ascertained
whether in the core protocol, an agent is guaranteed to get
a piece of value equal to his third or second most preferred
piece. Each agent gets a piece that is of same value as his
third most preferred piece. An agent is guaranteed to get a
second most preferred piece during the core protocol, if he is
asked in the worst case to trim his most preferred piece to
his second most preferred piece.
a
b
i
i
j
Figure 4: Example of a scenario where i has the
rightmost trim for piece a and second rightmost trim
of piece b whereas agent j has the rightmost trim for
piece b. If i gets the right hand side of a up till his
trim, and j gets the right hand side of b till i's trim
in b, then i is not envious of j's allocation.
We make another observation about the outcome of the
core protocol.
18: if exactly one piece has exactly one trim then {The
trims look like iijkjk}
19:
if the agent who trimmed it views it as his most pre-
ferred then
20:
21:
22:
23:
24:
Give the complete piece to him.
else if the agent who trimmed it find it his second
most preferred then
ask him to trim his first most preferred to equal his
second most preferred piece.
end if
The pieces are allocated up to the second rightmost
trim to the agents who trimmed them most. If 2 pieces
were trimmed most by the same agent, he decides which
to get and the other is given to the agent who trimmed
that piece second most.
25: Give the last unallocated piece completely to the cut-
ter (agent 4).
26: end if
27: if exactly 2 pieces have exactly one trim then {The
trims look like jkikij}
28: Give those pieces with exactly one trim completely to
the agents in {1, 2, 3} who trimmed them if it is their
most preferred.
29:
30:
31:
Agents who trimmed a piece with a single trim but
do not value that piece most are asked to re-trim their
most preferred piece up to their second most preferred
piece. Their trims to make them equal to their third
most preferred ignored from now on.
The right hand side of the two pieces with the two
trims are given to the agents with the rightmost trims.
The pieces are allocated up till the second rightmost
trim. If the 2 pieces were trimmed most by the same
agent (agent k), he choses which to get and the other
piece is given to whoever trimmed it second most.
If one non-cutter has not been allocated a piece, he
gets the most preferred piece among the two unallocated
complete pieces.
32: Give the last unallocated piece completely to the cut-
ter (agent 4)
33: end if.
34: if exactly 3 pieces have exactly one trim then
35:
36:
The trims look like 123123.
If an agent most prefers the piece where he made a
single trim, give him that complete piece.
37:
38:
The ones who most prefer the piece with the three
trims (call it a) compete for it by trimming up to their
second most preferred piece. Their trims to make them
equal to their most preferred ignored from now on.
Cut a at the second rightmost trim, and then allocate
a to whichever agent trimmed it most. The other agents
get their second most preferred complete piece.
39: Give the cutter (agent 4) the remaining complete
piece.
40: end if
41: return envy-free allocation and any cake that is still
not allocated.
Lemma 3. During the core protocol, when an agent i
makes trims to equal his second or third most preferred piece
respectively, i either gets such a piece completely or some
other agent gets such a piece completely that i values as much
as second or third most preferred piece respectively.
Proof. Assume agent i is not the cutter in the core pro-
tocol and he was asked to trim his first and second most
preferred pieces to equal his third most preferred piece. If
agent i gets the complete piece that is his third most pre-
ferred piece, we are already done. Let us say that he got a
partial piece. Then, at most one other agent got a partial
piece. This means that some other agent j got a complete
piece that is either agent i first, second or third most pre-
ferred piece.
Assume agent i is not the cutter in the core protocol and
he was asked to trim his most preferred pieces to equal his
second most preferred piece. This means that i got either
his most preferred piece up to the value of the second most
preferred piece or he got the second most preferred piece
completely. If he got the second most preferred piece com-
pletely, we are done. If i got the most preferred piece up
to the value of the second most preferred piece, then some
other agent got a possibly partial piece from i's most pre-
ferred piece. But note that since i was asked to trim up to
his second most piece in the protocol, only i was competing
for his second most preferred piece. Hence, some other agent
j got i's second most preferred piece completely.
The core protocol for partial envy-free allocation can be
extended to obtain protocol for partial envy-free allocation
with a single domination. If the core protocol is run again on
the unallocated cake, we show in Lemma 4 that the cutter
dominates at least one agent. When the core protocol is
implemented, then the piece from which the highest valued
(from the perspective of the cutter) residue is trimmed is
called a significant piece. We will show that the cutter can
be made to dominate an agent who got the significant piece.
If residues from both pieces that are partially allocated are
of the same value to the cutter, then we say that both non-
cutters who got partial pieces were given significant pieces.
In this case, a second run of the core protocol is enough for
the cutter to dominate two agents.
Lemma 4
(Single-Domination Protocol Lemma).
For n = 4, there exists a discrete and bounded protocol
which returns an envy-free partial allocation in which one
agent dominates another agent.
Proof. We run the core protocol a first time. This guar-
antees that the cutter (say agent 4) gets a complete piece
(of value 1/4 of the whole cake) and that the residue is com-
posed from the trims of at most 2 pieces. Since from the
cutter's perspective all pieces were equal, the residue can-
not sum up to more than 1/2 of the cake for him. Recall that
the piece which agent 4 thinks was trimmed is the significant
piece. The total residue is composed of the residue from the
significant piece (call the residue β1) and the residue from
the other trimmed piece (let us call this residue β2). The
cutter thinks that he got V4(β1) more value than the agent
who got the significant piece. Since V4(β1) ≥ V4(β2), the
value of the total residue from the cutter's perspective is at
most 2V4(β1). If we run the core protocol again, at most two
pieces are partial and hence the residue's value for the cutter
is at most 2 × 2V4(β1)/4 = V4(β1). This implies that even if
the agent who got the significant piece gets all the residue
which is of value V4(β1) to the cutter, the cutter would still
not envy him. This implies the cutter dominates the agent
who got the significant piece.
4.3 Permutation Protocol
If we run the core protocol repeatedly on the remaining
unallocated cake, it may be that the cutter keeps dominat-
ing the same agent. We show that we only need to run the
core protocol 5 times in total to achieve double domination.
It may be that each time, the core protocol is run, the same
agent gets the significant piece and hence the cutter domi-
nates the same agent. If a different agent gets a significant
part of the residue in any of the iterations, then the double
domination is already achieved with one more iteration since
from the cutter's perspective we have 2 agents who may be
given all that is left of the cake without him being envious
of them. If not, then we have one agent who ends up with
the piece from which a significant trim was obtained for all
iterations. The permutation lemma tells us that it is possi-
ble to give one of the 4 significant pieces to another agent
while still preserving envy-freeness. This ensures that agent
1 ends up dominating two agents.
Algorithm 4 Permutation Protocol for 4 Agents
Input: An outcome of a core protocol in which one agent
(say agent 4) is the cutter and another specified agent (say
agent 1) gets the significant piece.
Output: An allocation in which each agent gets a piece
equal to the value he trimmed to in the core protocol and
in which agent 2 or 3 gets the significant piece.
1: if agent 2 was competing with someone for the piece and
therefore had a trimmed piece then
1. If the agent who made the second rightmost trim
on agent 2's piece is agent 1, then we can simply
permute agents 1 and 2 i.e., exchange their pieces.
2. If the agent who made the second rightmost trim
on agent 2's piece is agent 3, then we can move 3 to
agent 2's piece. Agent 2 can be given 1's (trimmed)
piece. Agent 1 can be given one of the complete
pieces (which was given to 3 or 4). Agent 4 can be
given the remaining complete piece.
2: else if agent 2 was in possession of a complete piece for
which he was not competing with another agent then
1. If 2's piece is the piece such that agent 1 trimmed
up to that value in the core protocol, then simply
permute 1 and 2.
2. If 4 is holding the piece such that agent 1 trimmed
up to that value in the core protocol then we simply
move 4 to 2's piece since it is a complete piece and
agent 1 gets 4's piece. Agent 2 is given 1's piece.
3. If 3 has a complete piece such that agent 1 trimmed
up to that value in the core protocol and 3 is indif-
ferent between two pieces among his top 3 pieces,
then 3 can be given another complete piece (such
that he trimmed up to the value of that piece) of ei-
ther 2 and 4. Agent 1 can be given 3's piece. Agent
2 gets 1's piece and 4 gets the remaining complete
piece.
3: end if
Although the core protocol can be easily used to enable
the cutter to dominate one agent, dominating two agents
is more challenging.
In the next section, we show how to
overcome this challenge.
We will use this idea in the argument of the Permuta-
tion Lemma. Before presenting the Permutation Lemma,
we present another lemma that is useful for the proof of the
Permutation Lemma.
Lemma 5. Consider m rows each with m − 1 entries of
positive reals. Then there exists at least one row such that
for each entry in the row, the sum of other entries in the
column corresponding to that entry is greater than or equal
to the entry in the row.
Proof. It is sufficient to find a row in which each entry is
not the unique maximum entry for that column. We go row
by row and eliminate a row if it has at least one entry that
is a maximal value among all entries in the corresponding
column. Even if m − 1 rows are eliminated, we are left with
one row in which each entry is not the unique maximum
entry for that column.
We will rely on Lemma 5 while reasoning about when
reallocating pieces does not cause any envy. In particular
let us say that an agent gets slightly more than the value
he wanted to guarantee. He may not want to let go of this
extra value lest it leads to him being envious. However,
let us say he gets similar extra values again, then we may
ask the agent to choose which one of the extra values he
rates least and give this extra value to some other agent.
The other extra values, make up for this loss. Intuitively,
Lemma 5 will help identify that if we have enough subcases
(rows) then there will be one row on which an agent will be
happy to compromise.
We are now in a position to present the Permutation
Lemma.
Lemma 6
(Permutation Lemma). There exists a dis-
crete and bounded protocol for 4 agents that returns a partial
envy-free allocation in which one agent dominates two other
agents.
Proof. Assume that agents 1, 2, 3, 4 get pieces p1, p2, p3,
and p4 respectively with 4 getting complete pieces. When 4
cut the pieces, p1 is a piece P1 without the part left of the
second rightmost trim. Now, assume that in four iterations
of the 1 gets the significant piece and we want to reallocate
so that some other agent among 2 and 3 gets the signif-
icant piece. We need to show that when the reallocation
is done then barring the bonus part that agents get in the
core protocol, the agents get at least as preferred a piece. If
another agent aside from agent 1 gets the significant piece
then we are done. If agent 1 is repeatedly getting it, then
we zoom in to a case to permute it i.e., reallocate some of
the pieces so as to make sure that an agent other than 1 is
dominated. If agent 1 is repeatedly getting the significant
piece and there is no reallocation in which instead of agent 1
some other agent gets the significant piece, this means that
either (a) agent 1 likes the bonus from his significant piece
so much that he is not willing to take some other piece or
(b) for some other agent j, the bonus corresponding to the
significant piece is not enough for j to be attracted towards
the significant piece.
Claim 1. For a partial allocation as a result of the core
protocol in which agents make trims, a reallocation can be
done in which some agent other than 1 gets the significant
piece and each agent gets a piece of value corresponding to
his original trims but in which he may lose out on the addi-
tional bonus due to the second rightmost trims.
Proof. The algorithm to perform the reallocation is
stated as the Permutation Protocol (Algorithm 4).
Imag-
ine that agent 1 is holding the significant piece. For the
piece to be significant, another agent must have been com-
peting with agent 1 for it. If no other agent was competing
for the piece, then agent 1 would have got the whole piece
and hence the piece would not be significant.
Let us say that the agent who competes for the same piece
is agent 2. Since the piece was cut up to the second rightmost
trim, agent 2 is not envious of any other agent if agent 2 gets
the piece (while not getting his trim).
We now distinguish between 2 possibilities. We show how
both cases can be handled.
1. Agent 2 was competing with someone for the
piece and therefore had a trimmed piece. We
distinguish between two subcases.
(a) If the agent who made the second rightmost trim
on agent 2's piece is agent 1, then we can simply
permute agents 1 and 2 i.e., exchange their pieces.
(b) If the agent who made the second rightmost trim
on agent 2's piece is agent 3, then we can move 3
to agent 2's piece. Notice that 3 must have been
allocated a complete piece since only 2 pieces may
be trimmed. This means that 1 did not compete
with 3 for 3's piece which implies that if 1 can get
4 or 3's piece he will get the piece he was guaran-
teed before the order of the trim was determined
(either second most preferred or third most pre-
ferred). Agent 2 can be given 1's (trimmed) piece.
Agent 1 can be given one of the complete pieces
(which was given to 3 or 4). Agent 4 can be given
the remaining complete piece.
2. Agent 2 was in possession of a complete piece
for which he was not competing with another
agent. Let us focus on the piece agent 1 desires and
was guaranteed to get a piece of that value (either his
second most preferred or third most preferred). We
will use the fact that by Lemma 3, some other agent
got a complete piece that was of at least as much value
to agent 1. We distinguish between three subcases.
(a) If 2's piece is the piece agent 1 desires, then simply
permute 1 and 2.
(b) If 4 is holding the piece to be freed then we simply
move 4 to 2's piece since it is a complete piece and
agent 1 gets 4's piece. Agent 2 is given 1's piece.
(c) Let us assume that 3 is holding the piece to be
freed and agent 1 is guaranteed to get a piece of
a value as much as this complete piece (which we
refer to as a).
If 3 is indifferent among any two of his
most preferred three pieces, then either he is
indifferent among the top 2 or among the second
and third. For the former, if he got a top 2 piece,
he will be willing to get another top 2 piece and
if he got a third piece, he would certainly be
happy to get one of the complete top 2 pieces.
For the latter, if he got one of the second or third
preferred pieces, he would be happy to get the
other equally preferred one. If he had got the top
piece, then this means that only 3 had a proper
trim on a. But this means that 3 is willing to
move to another complete piece with as much
value barring the extra bonus he got in piece
a. Hence in all the cases above, 3 can be given
another complete piece, agent 1 can be given
3's complete piece, and 2 can get 1's partial piece.
We now assume that 3 is not indifferent
among any of his 3 most preferred pieces.
We show that in this scenario, we would already
be in one of the previous cases in which 1 is okay
with taking 2 or 4's piece.
We first argue that agent 3's most preferred piece
is the significant piece. Piece a cannot be agent
3's most preferred piece since otherwise 1 would
not be guaranteed to get it:
if 1 got it, 3 would
be envious. Neither can the piece held by 4 or 2
be 3's most preferred piece since envy-freeness of
the partial allocation will be violated. Since the
pieces allocated to 2, 3 and 4 are not 3's most
valued pieces, it implies that agent 3's most pre-
ferred piece is the significant piece.
Next, we argue that 3 was allocated his second
most preferred piece in which had put a proper
trim. First we assume that 3 trimmed up to his
third most preferred piece. If a is 3's third most
preferred piece, then 3 would have been envious of
either 2 or 4 whoever got 3's most preferred piece.
3 would have competed for such as piece and such
a piece would not have been allocated completely.
This means that 3 got his second most preferred
piece. Second, we assume that 3 trimmed unto
his second most preferred piece, then this means
3 was allocated his second most preferred piece
since he was guaranteed it. In both cases 3 put a
proper trim on a.
We now argue that a (3's piece) is also agent 1's
second most preferred piece.
If the piece allo-
cated to 2 or 4 was agent 1's second most pre-
ferred piece, we would be in a different case.
Therefore a is agent 1's and 3's second most pre-
ferred piece. Since 1 is guaranteed to get a piece
of value of a and since a is agent 1's second most
preferred piece, it means that only 1 has a proper
trim on a. But we have already shown that 3
also put a proper trim on a. But if two agents
put a proper trim on a piece, the piece cannot be
completely allocated to an agent hence a contra-
diction.
This complete the proof of the claim.
We have shown that reallocation can be done in which
each agent gets a piece of value corresponding the trims the
agents made.
Note that in each iteration of protocol, each agent gets
some (possibly non-zero) bonus. Since agent 1 gets a signif-
icant piece, he gets a non-zero bonus each time. Although 1
may object to any reallocation for a particular iteration of
core protocol, we show that he will not envious if one par-
ticular iteration of the core protocol is identified in which
corresponds to the row of entries in Table 1 in which for
each entry in the row, the sum of other entries in the col-
umn corresponding to that entry is greater than or equal to
the entry in the row. By Lemma 5, such a row exists. This
means that even if each agent loses his bonus because of the
permutation, he gets enough bonus in the other iterations
to make for this loss so that there is no envy.
We have shown so far that in four iterations of the core
protocol, the partial envy-free allocation is such that two
different agents got a significant piece in one of the calls
of the core protocol.
If the same agents get a significant
piece in all the first four iterations of the core protocol, then
we have shown above that the permutation protocol can
implemented to give the significant piece to another agent
and still not maintain envy-free across the four iterations
of the core protocol. The fifth iteration of the core protocol
ensures that the cutter dominates two agents and the partial
allocation is envy-free.
1
2
3
Bonus in 1st iteration
Bonus in 2nd iteration
Bonus in 3rd iteration
Bonus in 4th iteration
1
1
b1
b2
b3
b4
1
1
2
2
b1
b2
b3
b4
2
2
3
3
b1
b2
b3
b4
3
3
4
0
0
0
0
Table 1: Bonus of each agent in the first 4 calls of
the core protocol with agent 4 as the cutter.
4.4 Overall Protocol
In the previous sections, we built the building blocks for
our overall protocol: Post Double Domination Protocol,
Core Protocol, and Permutation Protocol. We are now in
a position to formalize the protocol to compute a complete
envy-free allocation and presents the main result. The pro-
tocol is formalized as Algorithm 5.
Algorithm 5 Discrete and Bounded Envy-free Protocol for
4 Agents.
1: while some agent i does not dominate two other agents
and there is still some unallocated cake do
2:
3:
4:
5:
6:
7:
Run the Core Protocol (Algorithm 3) 4 times on the
unallocated cake with i as the cutter. Return at any
point if there is no cake left unallocated. {After two
iterations, i already dominates one agent}
if the same agent (say agent j) gets the significant
piece in each of the 4 calls of the core protocol then
Identify in which call of the protocol, the non-cutter
agents get less bonus than the sum of bonuses in the
other calls of the core protocol. {see Table 1}
Implement reallocation via the Permutation Proto-
col (Algorithm 4) for pieces allocated by this particular
call of the core protocol where i is the cutter and j gets
the significant piece.
end if
Run core protocol in the unallocated cake if some cake
is still unallocated.
8: end while
9: if there is some unallocated cake then
10:
Run Post Double Domination Protocol (Algorithm 2)
on the remaining cake.
11: end if
12: return envy-free complete allocation.
Theorem 2. For four agents, there exists a discrete and
bounded envy-free protocol that requires constant number of
queries in the Robertson and Webb model.
Proof. The protocol is formalized as Algorithm 5. The
theorem follows from Lemmas 1 and 6. The protocol first en-
sures that there exists a partial envy-free allocation in which
each agent dominates two other agents. The protocol then
used the Double Domination Protocol to obtain a complete
envy-free allocation.
We now argue in more detail that the protocol is bounded
not only in terms of number of cuts but also in terms of
Robertson-Webb queries.
In each run of the core protocol, the cutter makes three
cuts and then each non-cutter makes at most two trims. Two
agents may be asked to replace their cuts and make a cut to
make their most preferred piece equal to their second most
preferred piece or in another case agents perform Divide and
Choose which requires 1 cut. Hence, in the core protocol,
there are at most 3+6+1=11 cuts. Since we run 5 iterations
of the core protocol to get one double domination, we need
50 cuts to get one double domination. Since we need each
agent to double dominate, we need 50 × 4 = 200 cuts. After
we achieve double dominations for all agents, we need at
most three more cuts. So we need at most 203 cuts.
We now count the maximum number of Robertson-Webb
query operations required during the protocol. The cutter is
asked the value of the whole cake (1 query), and then made
to cut to equal 1/4 of the value (3 queries). The non-cutter
agents are asked to give values of the pieces (3 × 4 = 12
queries.) They are then asked to trim to equal the value of
the third most preferred piece (3 × 2 queries). Two agents
may be asked to make one additional cut to make their most
preferred piece equal to their second most preferred piece
(2 × 1 queries) or in another case, agents perform Divide
and Choose which requires 4 queries. Hence the core proto-
col requires in total 1 + 3 + 12 + 6 + 4 = 26 queries. The
core protocol is run 5 times so that takes 26 × 5 queries.
In addition to running the core protocol 5 times, we also
need to run permutation protocol four times to check which
permutation to implement.
In each call of the permuta-
tion protocol, agents are queried about the amount of bonus
which they get in the core protocol which takes additional
3 queries so in total of 12 queries for the permutation pro-
tocol.
In order to get all double dominations, we require
((26 × 5) + 12) × 4 = 568 queries. After we achieve double
dominations, we require maximum 16 more queries. So the
total number of queries is 584.
5. DISCUSSION
In this paper, we proposed the first bounded and envy-
free protocol for four agents. Some of our insights such as
the one of exploiting the bonus cake given to the agents as
well as analyzing the domination graph may be useful in at-
tacking the problem for any number of agents. Our protocol
is based on three main ingredients: core protocol, permuta-
tional protocol, and post double domination protocol. The
higher level ideas of our overall protocol could be applied
to general problem for more number of agents. Some of the
proof ideas may be generalizable to more than four agents.
For example, a suitable generalization of the core protocol
is feasible. On the other hand, the Permutation Protocol
appears to be challenging to extend to arbitrary number of
agents and will require more interesting insights and tech-
niques. There are also some other interesting directions for
future research such as existence and properties of equilibria
under our new protocol.
6. ACKNOWLEDGMENTS
Data61 is funded by the Australian Government through
the Department of Communications and the Australian Re-
search Council through the ICT Centre of Excellence Pro-
gram. Part of the research was carried out when the authors
were visiting LAMSADE, University Paris Dauphine in May
2015. The authors thank Rediet Abebe, Steven Brams,
Simina Branzei, Ioannis Caragiannis, Katar´ına Cechl´arov´a,
Serge Gaspers, David Kurokawa, and Ariel Procaccia for
useful comments. They also thank the reviewers of STOC
2016 for suggestions to improve the presentation. Haris Aziz
thanks Ulle Endriss for introducing the subject to him at
the COST-ADT Doctoral School on Computational Social
Choice, Estoril, 2010.
7. REFERENCES
[1] H. Aziz and C. Ye. Cake cutting algorithms for piece-
wise constant and piecewise uniform valuations. In Pro-
ceedings of the 10th International Workshop on Internet
and Network Economics (WINE), pages 1–14, 2014.
[2] B. Barbanel and A. D. Taylor. Preference relations
and measures in the context of fair division. Ameri-
can Mathematical Monthly, 123(7):2061–2070, 1995.
[3] J. B. Barbanel. The Geometry of Efficient Fair Divi-
sion. Cambridge University Press, 2005.
[4] J. B. Barbanel and S. J. Brams. Cake division with
minimal cuts: envy-free procedures for three persons,
four persons, and beyond. Mathematical Social Sci-
ences, 48(3):251–269, 2004.
[5] S. J. Brams and A. D. Taylor. An envy-free cake di-
vision protocol. The American Mathematical Monthly,
102(1):9–18, 1995.
[6] S. J. Brams and A. D. Taylor. Fair Division: From
Cake-Cutting to Dispute Resolution. Cambridge Uni-
versity Press, 1996.
[7] S. J. Brams, A. D. Taylor, and W. S. Zwicker. A
moving-knife solution to the four-person envy-free cake
division. Proceedings of the American Mathematical So-
ciety, 125(2):547–554, 1997.
[8] S. Branzei. A note on envy-free cake cutting with
polynomial valuations. Information Processing Letters,
115(2):93–95, 2015.
[9] S. Branzei and P. B. Miltersen. A dictatorship theo-
rem for cake cutting.
In Proceedings of the 23rd In-
ternational Joint Conference on Artificial Intelligence
(IJCAI), pages 482–488. AAAI Press, 2015.
[10] C. Busch, M. S. Krishnamoorthy, and M. Magdon-
Ismail. Hardness results for cake cutting. Bulletin of
the EATCS, 86:85–106, 2005.
[11] Y. Chen, J. K. Lai, D. C. Parkes, and A. D. Procaccia.
Truth, justice, and cake cutting. Games and Economic
Behavior, 77(1):284–297, 2013.
[12] R. Cole and V. Gkatzelis. Approximating the nash so-
cial welfare with indivisible items. In Proceedings of the
47th Annual ACM Symposium on Theory of Computing
(STOC), pages 371–380. ACM Press, 2015.
[13] X. Deng, Q. Qi, and A. Saberi. Algorithmic solutions
for envy-free cake cutting. Mathematics of Operations
Research, 60(6):1461–1476, 2012.
[14] J. Edmonds and K. Pruhs. Cake cutting really is
not a piece of cake.
In Proceedings of the 17th An-
nual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 271–278, 2006.
[25] A. D. Procaccia. Thou shalt covet thy neighbor's cake.
In Proceedings of the 21st International Joint Confer-
ence on Artificial Intelligence (IJCAI), pages 239–244.
AAAI Press, 2009.
[15] G. Gamow and M. Stern. Puzzle-math. Macmillan,
[26] A. D. Procaccia. Cake cutting: Not just child's play.
1958.
[16] S. Garfunkel. For all practical purposes social choice.
COMAP, 1988.
[17] R. Guo. Cross-Border Management: Theory, Method
and Application. Springer, 2015.
[18] W. Hively. Dividing the spoils. Discover Magazine,
March, 1995.
[19] D. Kurokawa, J. Lai, and A. D. Procaccia. How to
cut a cake before the party ends.
In Proceedings of
the 27th AAAI Conference on Artificial Intelligence
(AAAI), pages 555–561. AAAI Press, 2013.
[20] E. Kushilevitz and N. Nisan. Communication Complex-
ity. Cambridge University Press, 2006.
[21] C. Lindner and J. Rothe. Degrees of guaranteed envy-
In
freeness in finite bounded cake-cutting protocols.
Proceedings of the 5th International Workshop on In-
ternet and Network Economics (WINE), volume 5929
of Lecture Notes in Computer Science (LNCS), pages
149–159. Springer-Verlag, 2009.
[22] C. Lindner and J. Rothe. Cake-cutting: Fair division
of divisible goods. In J. Rothe, editor, Economics and
Computation: An Introduction to Algorithmic Game
Theory, Computational Social Choice, and Fair Divi-
sion, chapter 7. Springer-Verlag, 2015.
[23] A. Maya and N. Nisan.
Incentive compatible two
player cake cutting. In Proceedings of the 8th Interna-
tional Workshop on Internet and Network Economics
(WINE), volume 7695 of Lecture Notes in Computer
Science (LNCS), pages 170–183. Springer-Verlag, 2012.
[24] O. Pikhurko. On envy-free cake division. The American
Mathematical Monthly, 107(8):736–738, 2000.
Communications of the ACM, 56(7):78–87, 2013.
[27] A. D. Procaccia. Cake cutting algorithms. In F. Brandt,
V. Conitzer, U. Endriss, J. Lang, and A. D. Procac-
cia, editors, Handbook of Computational Social Choice,
chapter 13. Cambridge University Press, 2016.
[28] J. M. Robertson and W. Webb. Near exact and envy-
free cake division. Ars Combinatorica, 45:97–108, 1997.
[29] J. M. Robertson and W. A. Webb. Cake Cutting Algo-
rithms: Be Fair If You Can. A. K. Peters, 1998.
[30] A. Saberi and Y. Wang. Cutting a cake for five people.
In Proceedings of the 5th International Conference on
Algorithmic Aspects in Information and Management
(AAIM), Lecture Notes in Computer Science (LNCS),
pages 292–300. Springer-Verlag, 2009.
[31] E. Segal-Halevi, A. Hassidim, and Y. Aumann. Waste
makes haste: Bounded time protocols for envy-free cake
cutting with free disposal. In Proceedings of the 14th
International Conference on Autonomous Agents and
Multi-Agent Systems (AAMAS), pages 901–908. IFAA-
MAS, 2015.
[32] H. Steinhaus. The problem of fair division. Economet-
rica, 16:101–104, 1948.
[33] I. Stewart. Division without envy. Scientific American,
1999.
[34] W. Stromquist. Envy-free cake divisions cannot be
found by finite protocols. The Electronic Journal of
Combinatorics, 15(R11), 2008.
[35] F. E. Su. Rental harmony: Sperner's lemma in fair
division. American Mathematical Monthly, 10:930–942.,
1999.
[36] W. Thomson. Children crying at birthday parties.
Why? Economic Theory, 31:501–521, 2007.
|
1301.3402 | 2 | 1301 | 2013-03-13T08:14:34 | Constraint Expressions and Workflow Satisfiability | [
"cs.DS",
"cs.CR"
] | A workflow specification defines a set of steps and the order in which those steps must be executed. Security requirements and business rules may impose constraints on which users are permitted to perform those steps. A workflow specification is said to be satisfiable if there exists an assignment of authorized users to workflow steps that satisfies all the constraints. An algorithm for determining whether such an assignment exists is important, both as a static analysis tool for workflow specifications, and for the construction of run-time reference monitors for workflow management systems. We develop new methods for determining workflow satisfiability based on the concept of constraint expressions, which were introduced recently by Khan and Fong. These methods are surprising versatile, enabling us to develop algorithms for, and determine the complexity of, a number of different problems related to workflow satisfiability. | cs.DS | cs |
Constraint Expressions and Workflow Satisfiability
Jason Crampton
Gregory Gutin
June 28, 2021
Abstract
A workflow specification defines a set of steps and the order in which those steps must be
executed. Security requirements and business rules may impose constraints on which users
are permitted to perform those steps. A workflow specification is said to be satisfiable if there
exists an assignment of authorized users to workflow steps that satisfies all the constraints.
An algorithm for determining whether such an assignment exists is important, both as a
static analysis tool for workflow specifications, and for the construction of run-time reference
monitors for workflow management systems. We develop new methods for determining
workflow satisfiability based on the concept of constraint expressions, which were introduced
recently by Khan and Fong. These methods are surprising versatile, enabling us to develop
algorithms for, and determine the complexity of, a number of different problems related to
workflow satisfiability.
1
Introduction
It is increasingly common for organizations to computerize their business and management pro-
cesses. The co-ordination of the tasks or steps that comprise a computerized business process is
managed by a workflow management system (or business process management system). Typi-
cally, the execution of these steps will be triggered by a human user, or a software agent acting
under the control of a human user, and the execution of each step will be restricted to some set
of authorized users.
A workflow is defined by the steps that comprise a business process and the order in which
those steps should be performed. Moreover, it is often the case that some form of access control,
often role-based, should be applied to limit the execution of steps to authorized users. In addition,
many workflows require controls on the users that perform groups of steps. The concept of a
Chinese wall, for example, limits the set of steps that any one user can perform [9], as does
separation-of-duty, which is a central part of the role-based access control model [1]. Hence,
it is important that workflow management systems implement security controls that enforce
authorization rules and business rules, in order to comply with statutory requirements or best
practice [6]. It is these "security-aware" workflows that will be the focus of the remainder of this
paper.
A simple, illustrative example for purchase order processing [10] is shown in Figure 1. In the
first step of the workflow, the purchase order is created and approved (and then dispatched to
the supplier). The supplier will submit an invoice for the goods ordered, which is processed by
the create payment step. When the supplier delivers the goods, a goods received note (GRN)
must be signed and countersigned. Only then may the payment be approved and sent to the
supplier. Note that a workflow specification need not be linear: the processing of the GRN and
of the invoice can occur in parallel, for example.
1
In addition to defining the order in which steps must be performed, the workflow specification
includes rules to prevent fraudulent use of the purchase order processing system. In our example,
these rules restrict the users that can perform pairs of steps in the workflow: the same user may
not sign and countersign the GRN, for example.
s3
s5
s2
6=
s1
s2
s4
s6
s5
6=
s3
s1
=
s4
6=
6=
s6
(a) Ordering on steps
(b) Constraints
s1
s2
s3
s4
s5
s6
6=
=
create purchase order
approve purchase order
sign GRN
create payment
countersign GRN
approve payment
different users must perform steps
same user must perform steps
(c) Legend
Figure 1: A simple constrained workflow for purchase order processing
It is apparent that it may be impossible to find an assignment of authorized users to workflow
steps such that all constraints are satisfied. In this case, we say that the workflow specification
is unsatisfiable. The Workflow Satisfiability Problem (WSP) is known to be NP-hard,
even when the set of constraints only includes constraints that have a relatively simple structure
(and that would arise regularly in practice).1
The rules described above can be encoded using constraints [10], the rules being enforced
if and only if the constraints are satisfied. More complex constraints, in which restrictions are
placed on the users who execute sets of steps can also be defined [3, 12, 21], can encode more
complex business requirements. (We describe these constraints in more detail in Section 2.1.) A
considerable body of work now exists on the satisfiability of workflow specifications that include
such constraints [6, 12, 21].
In this paper, we use constraint expressions to solve WSP. Constraint expressions were
introduced by Khan and Fong in their work on workflow feasibility [16]. However, the potential
of constraint expressions was not fully realized. In this paper, we show how constraint expressions
can be used to solve WSP and a number of related problems.
We also introduce a set of operators for combining workflows. This allows us to model
workflows in which the execution of steps is determined at execution time, which we will call
conditional workflows. Our model enables us to formulate the satisfiability problem for condi-
tional workflows, which we solve using constraint expressions. To our knowledge, these are the
first results on conditional workflows.
The main contributions of this paper are:
1In particular, the Graph k-Colorability problem can be reduced to a special case of WSP in which the
workflow specification only includes separation-of-duty constraints [21].
2
• to generalize the results of Wang and Li on the fixed parameter tractability of WSP
(Section 3);
• to introduce a language for workflow composition (Section 4);
• to establish new results on the satisfiability of conditional workflows (Section 4);
• to demonstrate how a problem studied by Armando et al. [3] and a problem introduced by
Crampton [10] can be solved using constraint expressions (Section 5).
In the next section we provide relevant background material. In Section 3 -- 5, we describe our
results. The proofs of our results can be found in the appendix, with the exception of the proof
of Theorem 5. We conclude with a summary of our contributions, a discussion of related work,
and our plans for future work.
2 Background
In this section, we introduce our notation and definitions, derived from earlier work [10, 21], and
then define the workflow satisfiability problem. In order to make the paper self-contained, we
also provide a short overview of parameterized complexity and summarize a number of useful
results from the literature.
2.1 The Workflow Satisfiability Problem
A directed acyclic graph G = (V, E) is defined by a set of nodes V and a set of edges E ⊆ V × V .
The reflexive, transitive closure of a directed acyclic graph defines a partial order, where v 6 w
if and only if there is a path from v to w in G. If (V, 6) is a partially ordered set, then we write
v k w if v and w are incomparable; that is, v 66 w and w 66 v. We may write v > w whenever
w 6 v. We may also write v < w whenever v 6 w and v 6= w. Finally, we will write [n] to denote
{1, . . . , n}.
Definition 1. A workflow specification is defined by a directed, acyclic graph G = (S, E), where
S is a set of steps and E ⊆ S × S. Given a workflow specification (S, E) and a set of users
U , an authorization policy for a workflow specification is a relation A ⊆ S × U . A workflow
authorization schema is a tuple (G, U, A), where G = (S, E) is a workflow specification and A is
an authorization policy.
We will use the representations of a workflow specification as a partial order and a DAG
interchangeably. The workflow specification describes a sequence of steps and the order in which
they must be performed when the workflow is executed, each such execution being called a
workflow instance. If s < s′ then s must be performed before s′ in every instance of the workflow;
if s k s′ then s and s′ may be performed in either order. User u is authorized to perform step s
only if (s, u) ∈ A.2 We assume that for every step s ∈ S there exists some user u ∈ U such that
(s, u) ∈ A.
2In practice, the set of authorized step-user pairs, A, will not be defined explicitly. Instead, A will be inferred
In particular, R2BAC -- the role-and-relation-based access control
from other access control data structures.
model of Wang and Li [21] -- introduces a set of roles R, a user-role relation UR ⊆ U × R and a role-step relation
SA ⊆ R × S from which it is possible to derive the steps for which users are authorized. For all common access
control policies (including R2BAC), it is straightforward to derive A. We prefer to use A in order to simplify the
exposition.
3
Definition 2. Let ((S, E), U, A) be a workflow authorization schema. A plan is a function
π : S → U . A plan π is authorized for ((S, E), U, A) if (s, π(s)) ∈ A for all s ∈ S.
Definition 3. A workflow constraint has the form (ρ, S1, S2), where S1, S2 ⊆ S and ρ ⊆ U × U .
A constrained workflow authorization schema is a tuple ((S, E), U, A, C), where C is a set of
workflow constraints.
Definition 4. A plan π : S → U satisfies a workflow constraint (ρ, S1, S2) if there exist s1 ∈ S1
and s2 ∈ S2 such that (π(s1), π(s2)) ∈ ρ. Given a constrained workflow authorization schema
((S, E), U, A, C), a plan π is valid if it is authorized and it satisfies all constraints in C.
We write ∆ ⊆ U × U to denote the diagonal relation {(u, u) : u ∈ U } and ∆c to denote its
complement {(u, u) : (u, u) 6∈ ∆}. Thus, the constraint on steps s1 and s2 in Figure 1 would be
written as (∆c, {s1} , {s2}).
We may now define the workflow satisfiability problem, as defined by Wang and Li [21].
Workflow Satisfiability Problem (WSP)
Input: A constrained workflow authorization schema ((S, E), U, A, C)
Output: A valid plan π : S → U or an answer that there exists no valid plan
We now discuss constraints in more detail, including the type of business rules we can encode
using our constraints and compare them to constraints in the literature. Our definition of work-
flow constraint is more general than similar definitions used when studying WSP. Crampton
defined constraints in which S1 and S2 are singleton sets: we will refer to constraints of this form
as Type 1 constraints; for brevity we will write (ρ, s1, s2) for the Type 1 constraint (ρ, {s1} , {s2}).
Wang and Li defined constraints in which at least one of S1 and S2 is a singleton set: we will
refer to constraints of this form as Type 2 constraints and we will write (ρ, s1, S2) in preference to
(ρ, {s1} , S2). Constraints in which S1 and S2 are arbitrary sets will be called Type 3 constraints.
We say that two constraints γ and γ ′ are equivalent if a plan π satisfies γ if and only if it
satisfies γ ′. The Type 2 constraint (ρ, s1, S2) is equivalent to (ρ, S2, s1) if ρ is symmetric, in
which case we will write (ρ, s1, S2) in preference to (ρ, S2, s1).
It is worth pointing out that Type 1 constraints can express requirements of the form described
in Section 1, where we wish to restrict the combinations of users that perform pairs of steps. The
plan π satisfies constraint (∆, s, s′), for example, if the same user is assigned to both steps by π,
and satisfies constraint (∆c, s, s′) if different users are assigned to s and s′. In other words, these
represent, respectively, binding-of-duty and separation-of-duty constraints. Abusing notation in
the interests of readability, we will replace ∆ and ∆c by = and 6=, respectively.
Type 2 constraints provide greater flexibility, although Wang and Li, who introduced these
constraints, do not provide a use case for which such a constraint would be needed. However,
there are forms of separation-of-duty requirements that are most naturally encoded using Type
3 constraints. Consider, for example, the requirement that a set of steps S ′ ⊆ S must not all
be performed by the same user [2]. We may encode this as the constraint (6=, S ′, S ′), which is
satisfied by a plan π only if there exists two steps in S ′ that are allocated to different users by π.
Henceforth, we will write WSP(ρ1, . . . , ρt) to denote a special case of WSP in which all
constraints have the form (ρi, S ′, S ′′) for some ρi ∈ {ρ1, . . . , ρt} and for some S ′, S ′′ ⊆ S. We
will write WSPi(ρ1, . . . , ρt) to denote a special case of WSP(ρ1, . . . , ρt), in which there are no
constraints of Type j for j > i. Thus, WSP1(=, 6=), for example, indicates an instance of WSP
in which all constraints have the form (=, s1, s2) or (6=, s1, s2) for some s1, s2 ∈ S.
We will write c, n and k to denote the number of constraints, users and steps, respectively,
in an instance of WSP. We will analyze the complexity of the workflow satisfiability problem in
terms of these parameters.
4
Note that definition of WSP given above does not make any reference to the ordering on the
set of steps. The original definition, as formulated by Crampton [10], included constraints that
were sensitive to the order in which steps were executed. If s k s′, we may define two different
constraints (ρ, s, s′) and (ρ′, s′, s), the first of which must be satisfied if s is performed before s′,
while the second must be satisfied if s′ is performed before s. To facilitate direct comparison with
the work of Wang and Li on WSP, we defer the analysis of Crampton's version of the problem
until Section 5.
2.2 Applications of WSP
In some systems, a
There are a number of different execution models for workflow systems.
tasklist is created when a workflow is instantiated. The tasklist is simply a valid plan for the
worfklow instance, allocating users to specific steps in the workflow instance. In other systems,
the workflow system maintains a pool of ready steps for each worfklow instance. We say a
step is ready in a workflow instance if all its immediate predecessor steps have been executed.
The workflow system may allocate ready steps to users; alternatively users may select steps to
perform from the pool. In both cases, the system must ensure both that the user is authorized
and that allowing the user to perform the step does not prevent the remaining steps in the
workflow instance from completing.
For systems that create tasklists, it is sufficient to know that the workflow specification
is satisfiable. Thus, an algorithm for deciding WSP is an important static analysis tool for
such systems. However, such an algorithm will only need to be executed when the workflow
specification is created or when it changes. The fact that the problem is NP-hard means that it
is important to find as efficient an algorithm as possible.
For other systems, however, the algorithm will need to be run repeatedly: every time a
user is allocated to a step. Note that the decision whether to allow a user to execute a step
in a partially completed workflow instance can be determined by solving an instance of WSP.
Specifically, suppose W = ((S, E), U, A, C) is a workflow specification, some subset S ′ of steps
have been performed in some instance of W , and the system needs to decide whether to allow
u′ to perform s′. Thus we have a partial plan π : S ′ → U . We then construct a new workflow
instance W ′ = ((S, E), U, A′, C), where (s, u) ∈ A′ if and only if one of the following conditions
holds: (i) s ∈ S ′ and u = π(s) (ii) s = s′ and u = u′ (iii) s 6∈ S ′ ∪ {s′} and (u, s) ∈ A. Clearly, the
workflow instance is satisfiable (when u′ performs s′) if and only if W ′ is satisfiable. Assuming
that these checks should incur as little delay as possible, particularly in the case when users
select steps in real time [17], it becomes even more important to find an algorithm that can
decide WSP as efficiently as possible.
The definition of workflow satisfiability given above assumes that the set of users and the
authorization relation are given. This notion of satisfiability is appropriate when the workflow
schema is designed "in-house". A number of large information technology companies develop
business process systems which are then configured by the end users of those systems. Part of
that configuration includes the assignment of users to steps in workflow schemas. The developer
of such a schema may wish to be assured that the schema is satisfiable for some set of users
and some authorization relation, since the schema is of no practical use if no such user set and
authorization relation exist. The desired assurance can be provided by solving an instance of
WSP in which there are k users, each of which is authorized for all steps. The developer may
also determine the minimum number of users required for a workflow schema to be satisfiable.
The minimum number must be between 1 and k and, using a binary search, can be determined
by examining ⌈log2 k⌉ instances of WSP.
5
2.3 Parameterized Complexity
A naıve approach to solving WSP would consider every possible assignment of users to steps in
the workflow. There are nk such assignments if there are n users and k steps, so an algorithm of
this form would have complexity O(cnk), where c is the number of constraints. Moreover, Wang
and Li showed that WSP is NP-hard, by reducing Graph k-Colorability to WSP(6=) [21,
Lemma 3]. In short, WSP is hard to solve in general. The importance of finding an efficient
algorithm for solving WSP led Wang and Li to look at the problem from the perspective of
parameterized complexity [21].
Suppose we have an algorithm that solves an NP-hard problem in time O(f (k)nd), where n
denotes the size of the input to the problem, k is some (small) parameter of the problem, f is some
function in k only, and d is some constant (independent of k and n). Then we say the algorithm is
a fixed-parameter tractable (FPT) algorithm. If a problem can be solved using an FPT algorithm
then we say that it is an FPT problem and that it belongs to the class FPT [13, 18].
Wang and Li showed, using an elementary argument, that WSP2(6=) is FPT and can be
solved in time O(kk+1N ), where N is the size of the entire input to the problem [21, Lemma
8]. They also showed that WSP2(6=, =) is FPT [21, Theorem 9], using a rather more complex
approach: specifically, they constructed an algorithm that runs in time O(kk+1(k − 1)k2k−1
N );
it follows that WSP2(=, 6=) is FPT. One of the contributions of this paper is to describe a new
method for solving WSP3(=, 6=) (that can also be used to solve WSP2(=, 6=)), thus generalizing
Wang and Li's result.
3 Solving WSP Using Constraint Expressions
In this section, we show how to extend the elementary methods used by Wang and Li to obtain
results for WSP2(=, 6=) and WSP3(=, 6=). Informally, our results make use of two observations:
• A construction used by Crampton et al. [11] can be used to transform an instance of
WSP1(=, 6=) into an equivalent instance of WSP1(6=) in time polynomial in the numbers
of constraints, steps and users.
• We can transform an instance of WSPi(=, 6=) into multiple instances of WSP1(=, 6=), the
number of instances being dependent only on the number of steps.
We use constraint expressions [16] to represent workflow constraints and to reason about multiple
constraints and the relationships between different types of constraints.
3.1 Reducing WSP1(=, 6=) to WSP1(6=)
The basic idea is to merge all steps that are related by constraints of the form (=, s1, s2)
for s1, s2 ∈ S. More formally, consider an instance I of WSP1(=, 6=), given by a workflow
((S, E), U, A, C).
(1) Construct a graph H with vertices S, in which s′, s′′ ∈ S are adjacent if C includes a
constraint (=, s′, s′′).
(2) If there is a connected component of H that contains both s′ and s′′ and C contains a
constraint (6=, s′, s′′) then I is unsatisfiable, so we may assume there is no such connected
component.
(3) For each connected component T of H,
6
(a) replace all steps of T in S by a "superstep" t;
(b) for each superstep t, authorize user u for t if and only if u was authorized (by A) for all
steps in t
(c) for each such superstep t, merge all constraints for steps in t.
Clearly, we now have an instance of WSP1(6=), perhaps with fewer steps and a modified autho-
rization relation, that is satisfiable if and only if I is satisfiable. For ease of reference, we will
refer to the procedure described above as the WSP1 constraint reduction method. The reduction
can be performed in time O(kc + kn), where c is the number of constraints: step (1) takes time
O(k + c); step (3) performs at most k merges; each merge takes O(k + c + n) time (since we
need to merge vertices, and update constraints and the authorization relation for the new vertex
set);3 finally, if k 6 c we have O(k(k + c + n) = O(k(c + n)), and if c 6 k then we perform no
more than c merges in time O(c(k + c + n)) = O(ck + cn) = O(ck + kn).
3.2 Constraint Expressions
the
our
behind
intuition
approach,
understand
To
workflow
W = ((S, E), U, A, {(ρ, S ′, S ′′)}), which defines an instance of WSP3(ρ). By definition, a plan π
satisfies the constraint (ρ, S ′, S ′′) if there exist s′ ∈ S ′ and s′′ ∈ S ′′ such that (π(s′), π(s′′)) ∈ ρ.
In other words, we could decide the satisfiability of W by considering the satisfiability of multiple
instances of WSP1: specifically, for each pair (s′, s′′) ∈ S ′ × S ′′, we consider the satisfiability of
the workflow ((S, E), U, A, {(ρ, s′, s′′)}); if any one of these instances is satisfiable, then so is W .
On the other hand, a plan satisfies a workflow W = ((S, E), U, A, {γ1, γ2}), for constraints γ1
and γ2, if and only π satisfies workflows ((S, E), U, A, {γ1}) and ((S, E), U, A, {γ2}).
consider
a
More formally, given a set of steps S, we define a constraint expression recursively:
• (ρ, s1, s2) is a (primitive) constraint expression;
• if γ and γ ′ are constraint expressions, then γ ∧ γ ′ and γ ∨ γ ′ are constraint expressions.
A plan π satisfies constraint expression:
• γ ∧ γ ′ if and only if π satisfies γ and γ ′; and
• γ ∨ γ ′ if and only if π satisfies γ or γ ′.
3.3 Reducing WSP(ρ1, . . . ρt) to WSP1(ρ1, . . . , ρt)
We now express workflow specifications using constraint expressions, rather than sets of con-
straints. A constraint (ρ, S ′, S ′′), ρ ∈ {ρ1, . . . , ρt},
is equivalent to a constraint expression
Ws′ ∈S ′,s′′ ∈S ′′ (ρ, s′, s′′), so every constraint can be written as the disjunction of primitive con-
straints. Moreover, the set of constraints {Γ1, . . . , Γc}, where each Γi is a disjunction of primitive
constraint expressions, is equivalent to the constraint expression Γ1 ∧ · · · ∧ Γc.
In other words, we can reduce the problem of determining the satisfiability of ((S, E), U, A, C)
to the problem of determining the satisfiability of a workflow of the form
((S, E), U, A, Γ1 ∧ · · · ∧ Γc),
where c = C; each clause Γi = (ρ, S ′
and each literal γi,j has the form (ρ, s′, s′′) for some s′ ∈ S ′
i, S ′′
i ) has the form γi,1 ∨ · · · ∨ γi,m(i), with m(i) = S ′
i ;
In other words,
i · S ′′
i and s′′ ∈ S ′′
i .
3We can check step (2) when we merge constraints in step 3(c).
7
we can represent any instance of WSP3(=, 6=) as a workflow containing a constraint expression
in "conjunctive normal form" in which each of the "literals" is a primitive constraint (which
corresponds to a single Type 1 constraint). Moreover, each literal is positive.
3.4 Solving WSP(=, 6=)
Given a constraint expression Γ1 ∧ · · · ∧ Γc, it is easy to see that if we can find a plan π for
some constraint expression of the form γ1 ∧ · · · ∧ γc, with γi ∈ Γi, then π is a plan for C. This
is because such a plan satisfies at least one literal in each clause Γi, thereby causing each Γi to
be satisfied; and C is satisfied if each clause is satisfied. Conversely, if π is a plan for C then it
is a plan for Γ1 ∧ · · · ∧ Γc and there exists a workflow expression of the form γ1 ∧ · · · ∧ γc for
which π is a plan. In other words, π is a plan for C if and only if it is a plan for γ1 ∧ · · · ∧ γc
for some γi ∈ Γi, where γi is a Type 1 constraint and γ1 ∧ · · · ∧ γc represents the constraint
set {γ1, . . . , γc}. We call γ1 ∧ · · · ∧ γc a simple constraint expression. That is, we have reduced
the satisfiability of an instance of WSP3(=, 6=) to determining the satisfiability of one or more
i=1 Γi, where Γi denotes the
number of literals (primitive constraint expressions) in Γi. Our strategy for solving an instance
of WSP3(=, 6=), therefore, is to try to determine the satisfiability of these related instances of
WSP1(=, 6=).
instances of WSP1(=, 6=). The number of instances is equal to Qc
Theorem 5. WSP2(=, 6=) and WSP3(=, 6=) can be decided in time
O((k − 1)c(c(k − 1)k + kn)) and O(cid:16)k2c(c(k − 1)k + kn)(cid:17),
respectively, where c is the number of constraints in the workflow instance.
Proof. We first consider an instance of WSP1(=, 6=), to which we apply the WSP1 constraint
reduction to obtain an instance of WSP1(6=). As any step with at least k authorized users can
be assigned a user that has not been assigned to any other step, we may focus on the allocation
of users to steps having fewer than k authorized users.
We consider each possible plan in turn and for each plan we check whether every constraint
is satisfied. There are no more than (k − 1)k plans to check -- since each of the steps has at
most k − 1 authorized users and there are no more than k steps -- and each constraint contains
two steps, so the time taken to solve WSP1(6=) is O(c(k − 1)k) and the time taken to solve
WSP1(6=, =) is O(c(k − 1)k + kn).
Now suppose we are given an instance of WSP(=, 6=). Then we can determine its satisfiability
by considering the satisfiability of multiple instances of WSP1(=, 6=), each instance containing
c constraints. We now determine the number of instances of WSP1(=, 6=) that need to be
considered in the worst case.
For a Type 2 constraint (ρ, s, S ′), we may assume that s 6∈ S ′: for (=, s, S ′), if s ∈ S ′, then
the constraint is satisfied by every plan and the constraint is redundant; for (6=, s, S ′), if s ∈ S ′,
then the constraint is equivalent to (6=, s, S ′ \ {s}). Hence, each Type 2 constraint (ρ, s, S ′) gives
rise to S ′ literals in a clause with S ′ < k. So we have c clauses, each of which contains no
more than k − 1 literals.
Type 1 constraints are equivalent to clauses with a single literal. Hence, for an instance of
WSP2(=, 6=) there are no more than (k − 1)c simple constraint expressions and so there are no
more than (k −1)c instances of WSP1(=, 6=) to check, which can be done in time O((k −1)c(c(k −
1)k + kn + kc)) = O((k − 1)c(c(k − 1)k + kn)).
Each Type 3 constraint (ρ, S ′, S ′′) yields a clause containing fewer than S ′ · S ′′ 6 k2
literals (which is greater than the number of clauses that can be obtained from a Type 1 or
8
Type 2 constraint). Hence, there are no more than O(k2c) simple constraint expressions and
WSP3(6=, =) can be decided in time O(k2c(c(k − 1)k + kn + kc)) = (k2c(c(k − 1)k + kn)).
Corollary 6. WSP2(=, 6=) and WSP3(=, 6=) are FPT.
3.5 Kernelization of WSP
Formally, a parameterized problem P can be represented as a relation P ⊆ Σ∗ × N over a finite
alphabet Σ. The second component is call the parameter of the problem. In particular, WSP
is a parameterized problem with parameter k, the number of steps. We denote the size of a
problem instance (I, k) by I + k.
Definition 7. Given a parameterized problem P , a kernelization of P is an algorithm that maps
an instance (I, k) to an instance (I ′, k′) in time polynomial in I + k such that (i) (I, k) ∈ P if
and only if (I ′, k′) ∈ P , and (ii) k′ + I ′ 6 g(k) for some function g; (I ′, k′) is the kernel and
g is the size of the kernel. If g(k) = kO(1), then we say (I ′, k′) is a polynomial-size kernel.
A kernelization provides a form of preprocessing aimed at compressing the given instance of
the problem. Polynomial-size kernels are particularly useful in practice as they often allow us to
reduce the size of the input of the problem under consideration to an equivalent problem with
an input of significantly smaller size.
Crampton et al. recently established that WSP1(=, 6=) has a polynomial-size kernel [11, §6].
In the case of WSP1(=, 6=), we can reduce the problem to one containing at most k users [11,
Theorem 6.5]. Crampton el al. also showed that WSP2(=, 6=) (and hence WSP3(=, 6=)) does
not have a polynomial-size kernel, so there is no efficient preprocessing step for such instances of
WSP. However, our results in this paper show we can reduce an instance of WSP(=, 6=) to at
most k2c instances of WSP1(6=) and then solve each instance by first computing a (polynomial-
size) kernel. The proof of Corollary 6 asserts that c 6 4k, although we would expect c to be
linear or quadratic in the number of steps in practice. This approach is similar to those that use
so-called Turing kernels (see [19], for example).
3.6 Negative Constraint Expressions
We could extend the syntax for constraint expressions to include negation. In other words, if γ
is a constraint expression, then ¬γ is a constraint expression. A plan π satisfies ¬γ if and only
if π violates γ. A plan π satisfies the constraint ¬(=, S1, S2), for example, if and only for all
si ∈ Si, (π(s1), π(s2)) 6∈ ∆; that is, if and only if for all si ∈ Si, π(s1) 6= π(s2).4 Thus, we can
encode any instance of WSP(=, 6=) using only constraints of the form (=, s1, s2) if we allow the
use of negation. Note, however, this means that the method for solving WSP(=, 6=) described
in Section 3.4 no longer works, because we may have negative literals in our conjunctive normal
form expressions.
However, we can determine the satisfiability of the constraint expression using any SAT solver.
A satisfying assignment returned by the SAT solver provides a "template" for a valid plan:
if
the variable (=, s1, s2) is set to true, then our plan must assign the same user to s1 and s2. This
induces a partition of the set of steps into blocks, each of which must be executed by a different
user. Hence, each satisfying assignment of the constraint expression gives rise to an instance
of WSP1(6=) in which each "step" is a block of steps in the original problem instance. We can
4This constraint is similar to the separation of duty constraints described by Basin et al. [7] and the universal
constraints described by Wang and Li [21]. Of course, we can represent this constraint as the set of Type 1
constraints {(6=, s1, s2) : s1 ∈ S1, s2 ∈ S2}.
9
solve this instance in time O((cid:0)k
2(cid:1)(b − 1)b), where b 6 k is the number of blocks, since there are
at most (cid:0)k
2(cid:1) constraints of the form (=, s1, s2). If b is small relative to k, then this may prove to
be a very efficient way of solving the original instance of WSP(=, 6=). However, we may need to
consider 2(k
2) satisfying assignments. In future work we hope to explore whether the additional
expressive power of negative constraint expressions allows us to encode business rules of practical
relevance. Further experimental work, investigating which strategies for solving WSP work best
in practice, is required.
4 Conditional Workflows
In some situations, we may wish to have conditional branching in a workflow specification, some-
times known as OR-forks [20] or exclusive gateways [22]. In our workflow system for purchase
order processing, for example, we may require that only orders with a value exceeding some
threshold amount need to be signed for twice. Informally, we can represent this extended specifi-
cation by the diagram shown in Figure 2, where s′
3 represents a step for signing a goods received
note on low-valued items. The nodes containing k and ⊕ are "orchestration" steps (or "gate-
ways") at which no processing is performed: ⊕ indicates that exactly one of the two branches is
executed, while k denotes that both branches must be executed.
⊕
s3
s′
3
s4
s5
s6
s1
s2
k
Figure 2: A workflow specification with conditional step execution
4.1 Workflow Composition
We now introduce a simple language for defining workflows. This language enables us to extend
the definition of WSP to workflows containing OR-forks, but not to arbitrary workflow patterns.
We assume every workflow specification includes a start step and a finish step, which we will
denote by α and ω, respectively, with subscripts where appropriate. These steps are orchestration
steps: no processing is performed by these steps and no constraints are applied to their execution;
they are used by the workflow management system solely to manage the initiation and completion
of workflow instances. Given two workflow specifications W1 = (S1, E1) and W2 = (S2, E2), we
may construct new workflow specifications using serial, parallel and xor composition, denoted
by W1 ; W2, W1 k W2 and W1 ⊕ W2, respectively. We assume throughout that S1 ∩ S2 = ∅. (If
this were not the case with s ∈ S1 ∩ S2, we could simply introduce subscripts or new labels to
distinguish the two copies of s.)
For serial composition, all the steps in W1 must be completed before the steps in W2. Hence,
the graph of W1 ; W2 is formed by taking the union of S1 and S2, the union of E1 and E2, and
the addition of a single edge between ω1 and α2.
10
For parallel composition, the execution of the steps in W1 and W2 may be interleaved. Hence,
the graph of W1 k W2 is formed by taking the union of S1 and S2, the union of E1 and E2, the
addition of new start and finish steps αpar and ωpar, and the addition of edges from αpar to
α1 and α2 and from ω1 and ω2 to ωpar. This form of composition is sometimes known as an
AND-fork [20] or a parallel gateway [22].5
In both serial and parallel composition, all steps in W1 and W2 are executed. In xor com-
position, either the steps in W1 are executed or the steps in W2, but not both. In other words,
xor composition represents non-deterministic choice in a workflow specification. The graph of
W1 ⊕ W2 is formed by taking the union of S1 and S2, the union of E1 and E2, the addition of
new start and finish steps αxor and ωxor, and the addition of edges from αxor to α1 and α2 and
from ω1 and ω2 to ωxor.
Henceforth, we will assume that ω1 followed by α2 will be merged to form a single (orchestra-
tion) node ǫ. Similarly, we will assume that (i) αpar followed by α1 and α2 in serial composition
will be merged to form a single node αpar; (ii) ωpar followed by ω1 and ω2 will be merged to form
a single node ωpar; (iii) αxor followed by α1 and α2 will be merged to form a single node αxor;
(iv) ωxor followed by ω1 and ω2 will be merged to form a single node ωxor.
Serial and parallel composition are illustrated in Figure 3. The structure of xor composition
is identical to that for parallel composition so it is not shown.
α1
W1
ǫ
W2
ω2
(a) Serial
Figure 3: Workflow composition
W1
αpar
ωpar
W2
(b) Parallel
4.2 Execution Sets
When we have conditional branching in a workflow, there exists more than one set of steps that
could comprise a complete workflow instance. Formally, an execution set is defined recursively:
• for a workflow specification comprising a single step s, there is a single execution set {s};
• if W1 and W2 are workflow specifications and S1 and S2 are execution sets for W1 and W2,
respectively, then
-- S1 ∪ S2 is an execution set for W1 ; W2,
-- S1 ∪ S2 is an execution set for W1 k W2,
-- S1 and S2 are execution sets for W1 ⊕ W2.
5The workflows that arise from serial and parallel composition have a lot in common with series-parallel graphs;
see [5], for example, for further details.
11
In our running example, both {s1, s2, s3, s4, s5, s6} and {s1, s2, s′
3, s4, s6} represent possible ex-
ecution sets, with the second set representing a workflow instance in which the value of goods
ordered is lower than the threshold requiring the GRN to be countersigned.6
4.3 Workflow Formulas and Trees
Clearly, each workflow step represents a workflow specification, in fact the simplest possible
specification. Hence, we may represent the example workflow specification in Figure 2 as the
workflow formula
(s1 ; s2) ; (((s3 ; s5) ⊕ s′
3) k s4) ; s6.
Thus, we may also represent the workflow specification as a workflow tree, as illustrated in
Figure 4.
;
k
s6
⊕
s4
s′
3
;
;
;
s1
s2
s3
s5
Figure 4: A workflow tree
The number of different possible execution sets is determined by the structure of the workflow
formula. Specifically, let ♯(W ) denote the number of possible execution sets for workflow W . For
a workflow W comprising a single step, we have ♯(W ) = 1. In general, we have
♯(W1 ; W2) = ♯(W1 k W2) = ♯(W1) · ♯(W2)
♯(W1 ⊕ W2) = ♯(W1) + ♯(W2),
where · denotes multiplication.
Using a post-order traversal of the workflow tree, we can compute the number of possible
execution sets: we assign the value 1 to each leaf node; we compute the number of possible
execution sets for each non-leaf node using the values assigned to its children and the appropriate
formula for the operation associated with the node. The root node in the tree depicted in Figure 4
is assigned the value 2, for example.
We write ♭(W ) to denote the maximum number of steps in any possible execution set for a
workflow specification W . Then
♭(W1 ; W2) = ♭(W1 k W2) = ♭(W1) + ♭(W2)
♭(W1 ⊕ W2) = max {♭(W1), ♭(W2)} .
6The concept of an execution set is related to, but simpler than, the concept of an execution history [7]: for
any execution set {s1, . . . , sm}, an execution history is a set {(s1, u1), . . . , (sm, um)} for some users u1, . . . , um.
An execution history also has some similarity to our concept of a plan.
12
Clearly, we can compute ♭(W ) from the workflow tree associated with W using a similar algorithm
to the one described above for calculating ♯(W ).
4.4 Constraints in Conditional Workflows
Let W1 and W2 be two workflow specifications with constraints C1 and C2, respectively. When
we form W1 ; W2 or W1 k W2, we include all constraints in C1 and C2. In addition, we may create
new constraints, governing the execution of some steps in S1 and some steps in S2. However,
we prohibit the addition of constraints in which all the steps are contained in either S1 or S2
(the assumption being that they would have been created earlier, if required). In other words,
any constraint that is added when we form W1 ; W2 (or W1 k W2) has the form (ρ, S ′, S ′′), where
S ′ ∪ S ′′ 6⊆ S1 and S ′ ∪ S ′′ 6⊆ S2.
In contrast, since xor composition requires that we either perform the steps in S1 or those in
S2, any constraint that includes steps from both S1 and S2 serves no purpose. Hence, we assume
that we add no constraints when we form W1 ⊕ W2.
4.5 Derived Deterministic Workflows
We say a workflow specification is deterministic if it has a single execution set (and non-
deterministic otherwise). Each possible execution set in a non-deterministic workflow speci-
fication gives rise to a different, deterministic workflow specification.
In particular, given a
workflow specification W = (S, E) with execution sets {S1, . . . , Sm}, we define Wi = (Si, Ei),
where
Ei
def= (Si × Si) ∩ E.
Then Wi is a (derived) deterministic workflow specification.
For a constrained workflow specification W = ((S, E), A, C) with possible execution sets
{S1, . . . , Sm}, we define Wi = (Si, Ei, Ai, Ci), where
Ai
def= (Si × U ) ∩ A,
and, for each γ = (ρ, S1, S2) ∈ C such that S1 ∩ Si 6= ∅ and S2 ∩ Si 6= ∅,
γi
def
= (ρ, S1 ∩ Si, S2 ∩ Si) ∈ Ci.
Each Wi is a deterministic, constrained workflow specification. Notice that when we form γi, S1 ∩
Si 6= ∅ and S2∩Si 6= ∅: this follows by a simple induction on the structure of the workflow formula
and the assumptions we make about the addition of constraints when we compose workflows (as
described in Section 4.4).
Hence, we may model any non-deterministic workflow specification as a collection of deter-
ministic workflow specifications. We may define the notion of weakly satisfiable and strongly
satisfiable for a non-deterministic specification: the former holds if there exists a derived, de-
terministic workflow specification that is satisfiable; the latter holds if all derived, deterministic
workflow specifications are satisfiable. In practice, it is likely that a workflow specification should
be strongly satisfiable (otherwise there exist execution paths that can never complete).
Proposition 8. Let W be an instance of WSP1(=, 6=). Then we can determine whether W is
weakly or strongly satisfiable in time O(♯(W )(♭(W ) − 1)♭(W )).
Note that we can extend this result to WSP3(=, 6=) as described in the proof of Theorem 5
(that is, using the reduction to multiple instances of WSP1(=, 6=), where the number of instances
13
is O(♭(W )2)). The above result asserts that the complexity of checking whether a workflow is
strongly satisfiable is determined by ♭(W ) and ♯(W ). Crude upper bounds for these parameters
are k and 2k, both functions of k only. Thus, determining whether a conditional workflow is
strongly satisfiable is FPT.
Of course, these bounds can be improved: the upper bound for ♭(W ) is only attained if no
xor composition is used, in which case ♯(W ) = 1; conversely, introducing xor composition may
reduce the maximum length, and using only xor composition reduces the number of derived
specifications to k. The question is: What deployment of k − 1 composition operators for k steps
yields the worst-case complexity? We have the following result.
Theorem 9. Given k workflow steps, a workflow has no more than:
• 3k′
execution sets if k = 3k′;
• 4 · 3k′
−1 execution sets if k = 3k′ + 1; and
• 2 · 3k′
execution sets if k = 3k′ + 2.
Remark 10. The proof of the above result (see appendix) is constructive, in the sense that it
tells us how to maximize the number of execution sets for a fixed set of k steps. Given k steps,
we obtain a workflow with the greatest possible number of execution sets by taking the serial (or
parallel) composition of sub-workflows ⊕2 and ⊕3, where ⊕i denotes the xor composition of i
steps. More specifically, if k = 3a, we take the serial composition of a copies of ⊕3; if k = 3a + 1,
we take the serial composition of a − 1 copies of ⊕3 and two copies of ⊕2; and if k = 3a + 2, we
take the serial composition of a copies of ⊕3 and one copy of ⊕2. We may conclude that ♭(W )
for such a workflow is no greater than ⌈k/3⌉.
Remark 11. Note that using xor composition reduces ♭(W ). And note that the exponential term
in the complexity of solving WSP1(=, 6=) is determined by the number of steps in the workflow,
for which an upper bound is ♭(W ) in the case of non-deterministic workflow specifications. For
a fixed k, it follows from Theorem 9 that the worst-case complexity for WSP1(=, 6=) occurs for a
workflow specification with a single execution set (of k steps).
5 Further Applications
In this section, we study two problems from the literature and establish that they are fixed-
parameter tractable. In both cases, we represent the problem as a workflow satisfiability problem
using constraint expressions.
5.1 Ordered WSP
We note that the version of WSP considered so far in this paper makes no use of the order
relation on the set of steps. This is a simplification introduced by Wang and Li [21]. In fact, the
definition of workflow constraints by Crampton [10] prohibited constraints of the form (ρ, s, s′)
for s > s′. Moreover, a plan was required to specify an execution order for the steps in the
workflow (in addition to the assignment of steps to users). This, in turn, means that Crampton's
definition of constraint satisfaction (and hence of the workflow satisfiability problem) is more
complex. More formally, we have the following definitions.
Definition 12. Let W = ((S, E), U, A, C) be a workflow comprising k steps. A tuple (s1, . . . , sk)
is an execution schedule for W if {s1, . . . , sk} = S and, for all 1 6 i < j 6 k, si 6> sj.7 We say
si precedes sj in an execution schedule if i < j.
7In other words, an execution schedule is a linear extension or topological sort of (S, 6).
14
For the workflow depicted in Figure 1, (s2, s1, . . . ) is not an execution schedule, for example,
but (s1, s2, s3, s5, s4, s6) and (s1, s2, s3, s4, s5, s6) are.
Definition 13. The (Type 1) constraint (ρ, s, s′) is satisfied by execution schedule σ and plan π
if one of the following holds: (i) s precedes s′ in σ and (π(s), π(s′)) ∈ ρ; (ii) s′ precedes s in σ.
The intuition here is that a constraint (ρ, s, s′) is well-formed only if s could precede s′ in
the execution of some instance of the workflow (that is, either s < s′ or s k s′). Moreover, if
s does occur before s′, then the execution of s′ is constrained by ρ and the identity of the user
that performed s. A modified version of WSP, based on the above definitions, is defined in the
following way.
Ordered WSP (OWSP)
Input: A constrained workflow authorization schema ((S, E), U, A, C).
Output: True if there exists an execution schedule σ and a plan π that satisfy all
constraints in C, and False otherwise.
Note that it may not be possible to find a valid plan π for a particular execution schedule
σ. Conversely, there may be a plan π for which there exist schedules σ and σ′ such that (σ, π)
satisfies all constraints but (σ′, π) does not. Consider, for example, a plan π that is valid for our
purchase order workflow such that π(s3) = π(s4). If we add the constraint (6=, s3, s4), then π is
valid for any execution schedule in which s4 precedes s3 and invalid otherwise.
The above example also shows there exist workflows for which a plan π is not a solution
to WSP, but for which (σ, π) is a solution to OWSP for certain choices of σ. Crampton
introduced the notion of a well-formed workflow, which has the following property: for all si k sj,
(ρ, si, sj) ∈ C if and only if (ρ, sj, si) ∈ C, where ρ is defined to be {(u, u′) ∈ U × U : (u′, u) ∈ ρ}.
To ensure that the workflow in the above example is well-formed, we would add the constraint
(6=, s4, s3) to C. It is easy to see that OWSP for well-formed workflows and WSP are essentially
equivalent, since a valid plan for one execution schedule will be a valid plan for any execution
schedule [10, Lemma 9].
Nevertheless, there will be business processes that cannot be represented using a well-formed
workflow schema. In the purchase order example illustrated in Figure 1, for example, it would
be quite reasonable to impose constraints on s3 and s4 that would mean the resulting workflow
schema was not well-formed. Suppose, for example, that ∼ is an equivalence relation on U , where
u ∼ u′ if and only if u and u′ belong to the same department. Then the constraints (≁, s3, s4)
and (6=, s4, s3) require that if s3 (the sign GRN step) is performed before s4 (the create payment
step), then the user that performs s4 must be in a different department from the user that
performs s3; whereas if the steps are performed in the reverse order, we only require the users to
be different (since the more commercially sensitive step has been performed first in this case).
Note that OWSP is only defined for Type 1 constraints (see Definition 13). Wang and
Li showed that WSP is W[1]-hard [13] for arbitrary constraint relations (even if only Type 1
constraints) are used [21]. Moreover, any instance of WSP defines an instance of OWSP. Thus,
OWSP is W[1]-hard. However, there is a strong connection between WSP and OWSP.
Proposition 14. OWSP1(ρ1, . . . , ρt) is FPT if WSP1(ρ1, . . . , ρt) is.
A stronger notion of satisfiability for OWSP would require that there exists a plan for every
execution schedule (as for conditional workflows). In this case, we simply require that every one
of the O(k!) derived instances of WSP is satisfiable. The worst-case complexity of determining
"weak" and "strong" satisfiability for OWSP is, therefore, the same. Note that an instance of
WSP is satisfiable if the corresponding instance of OWSP is strongly satisfiable.
15
5.2
Identifying Constraint Violation
Consider the following problem: Given a constrained workflow specification ((S, E), U, A, C),
does there exist a plan α such that (s, α(s)) ∈ A for all s ∈ S and at least one constraint C that
is not satisfied? This question is of interest because if we know that no such plan exists, then we
do not need a reference monitor: any allocation of (authorized) users to steps will satisfy all the
constraints. This question has been studied by Armando and colleagues [2, 4] and solutions have
been computed using model checkers. We answer this question by examining the satisfiability of
the "negation" of the problem, rewritten using the language of constraint expressions.
Theorem 15. Determining whether there exists a plan that violates a workflow specification
((S, E), U, A, C), where all constraints have the form (=, S1, S2) or (6=, S1, S2), is FPT.
The approach described above can also be used to "prune" a workflow specification. Given
a workflow specification ((S, E), U, A, C), we can identify, with the same (worst-case) time com-
plexity, all constraints in C that can be violated. This enables us to remove any constraints that
cannot be violated, leaving a workflow specification ((S, E, U, A, C ′), with C ′ ⊆ C. In Section 2.2,
we identified situations in which we may be required to solve WSP for a workflow specification
multiple times. Thus, reducing the set of constraints will reduce the complexity of subsequent
attempts to determine the satisfiability of the workflow specification.
6 Concluding Remarks
In this paper, we have explored the use of constraint expressions as a means of translating
different versions of the workflow satisfiability problem into one or more instances of WSP1(6=).
Constraint expressions provide a uniform way of representing the workflow satisfiability problem
and related problems, such as WSP for conditional workflows (Section 4), ordered WSP and
the identification of constraints that can be violated (Section 5). This, in turn, enables us
to establish the complexity of solving these problems. We also believe our characterization of
workflow composition, the representation of workflows as trees, and execution sets may be useful
modeling tools for future research on authorization in workflow systems.
6.1 Related Work
Work on computing plans for workflows that must simultaneously satisfy authorization policies
and constraints goes back to the seminal paper of Bertino et al. [8]. This work considered
linear workflows and noted the existence of an exponential algorithm for computing valid plans.
Crampton extended the model for workflows to partially ordered sets (equivalently, directed
acyclic graphs) and to directed acyclic graphs with loops [10]. Wang and Li further extended
this model to include Type 2 constraints [21].
Wang and Li first investigated the computational complexity and, significantly, the existence
of fixed-parameter tractable algorithms for the workflow satisfiability problem [21]. One or their
main results [21, Theorem 9] is very similar to the result we prove for WSP2(=, 6=) (Theorem 5),
although our approach is more direct and generalizes to WSP3(=, 6=). Crampton et al. introduced
a new method for solving the problem [12], which yields significantly better complexity bounds
for WSP3(=, 6=). However, their methods only apply for certain kinds of constraints; indeed, it
is not clear whether their approach extends to relations other than ∆, ∆c and constraints using
equivalence relations defined on the user set.
The use of constraint expressions to represent and reason about the complexity of the work-
flow satisfiability problem appears, therefore, to have some significant advantages, one specific
16
advantage being its versatility, over existing approaches. Khan and Fong introduced the notion
of a constraint expression to reason about the problem of workflow feasibility [16], which asks:
Given a set of constraints and restrictions on admissible authorization policies, does there exist
an authorization policy from which we can construct a valid plan? Their work was undertaken
in the context of the relationship-based access control model [14], in which the "shape" of au-
thorization policies is restricted, and does not explore fully the possibility of using constraints
expressions to solve the "classical" workflow satisfiability problem.
It is widely accepted that it is useful to have conditional branching in workflow specifica-
tions [20, 22]. However, there is very little prior work on the workflow satisfiability problem,
or its complexity, for conditional workflows. Khan's master's thesis includes work on existential
satisfiability (what we have called weak satisfiability) and universal (strong) satisfiability [15,
Chapter 8] but does not consider fixed parameter tractability.
6.2 Future Work
There are a number of opportunities for future work. Crampton et al. studied the workflow
satisfiability problem in the presence of constraints specified using an equivalence relation ∼
defined on U [12]. The relation ∆ may be viewed as an equivalence relation, in which each
equivalence class is a single user. We would like to investigate whether our methods can be
extended to solve WSP(=, 6=, ∼, ≁), where ∼ is not equal to ∆. This is a non-trivial problem
as we cannot use our trick of considering only those steps for which there are fewer than k
authorized users. A second problem we would like to consider is the optimal workflow-aware
authorization administration problem, which determines whether it is possible to modify the
authorization relation, subject to some bound on the "cost" of the changes, when the workflow
is unsatisfiable [7]. Finally, we would like (a) to remove the restriction that (S, E) is an ayclic
graph, so that we can model sub-workflows that can be repeated, and (b) to include inclusive
gateways [22], allowing for one or more sub-workflows to be executed. Both of these extensions
can be readily modeled using execution sets (or multisets).
If, for example, S1 and S2 are
execution sets for W1 and W2, respectively, then S1, S2 and S1 ∪ S2 are execution sets for
W1 + W2, where + indicates inclusive-or composition.
References
[1] American National Standards Institute. ANSI INCITS 359-2004 for Role Based Access
Control, 2004.
[2] A. Armando, E. Giunchiglia, and S. E. Ponta. Formal specification and automatic analysis of
business processes under authorization constraints: An action-based approach. In S. Fischer-
Hubner, C. Lambrinoudakis, and G. Pernul, editors, TrustBus, volume 5695 of Lecture Notes
in Computer Science, pages 63 -- 72. Springer, 2009.
[3] A. Armando and S. Ponta. Model checking of security-sensitive business processes.
In
P. Degano and J. D. Guttman, editors, Formal Aspects in Security and Trust, volume 5983
of Lecture Notes in Computer Science, pages 66 -- 80. Springer, 2009.
[4] A. Armando and S. Ranise. Automated analysis of infinite state workflows with access
control policies. In C. Meadows and M. C. F. Gago, editors, STM, volume 7170 of Lecture
Notes in Computer Science, pages 157 -- 174. Springer, 2011.
[5] J. Bang-Jensen and G. Gutin. Digraphs: Theory, Algorithms and Applications. Springer,
2nd edition, 2009.
17
[6] D. A. Basin, S. J. Burri, and G. Karjoth. Obstruction-free authorization enforcement:
Aligning security with business objectives. In CSF, pages 99 -- 113. IEEE Computer Society,
2011.
[7] D. A. Basin, S. J. Burri, and G. Karjoth. Optimal workflow-aware authorizations.
In
V. Atluri, J. Vaidya, A. Kern, and M. Kantarcioglu, editors, SACMAT, pages 93 -- 102.
ACM, 2012.
[8] E. Bertino, E. Ferrari, and V. Atluri. The specification and enforcement of authorization
constraints in workflow management systems. ACM Trans. Inf. Syst. Secur., 2(1):65 -- 104,
1999.
[9] D. F. C. Brewer and M. J. Nash. The Chinese Wall security policy. In IEEE Symposium
on Security and Privacy, pages 206 -- 214. IEEE Computer Society, 1989.
[10] J. Crampton. A reference monitor for workflow systems with constrained task execution.
In E. Ferrari and G.-J. Ahn, editors, SACMAT, pages 38 -- 47. ACM, 2005.
[11] J. Crampton, G. Gutin, and A. Yeo.
On the parameterized complexity and
CoRR, abs/1205.0852, 2012.
kernelization of
http://arxiv.org/abs/1205.0852.
the workflow satisfiability problem.
[12] J. Crampton, G. Gutin, and A. Yeo. On the parameterized complexity of the workflow
satisfiability problem. In T. Yu, G. Danezis, and V. D. Gligor, editors, ACM Conference on
Computer and Communications Security, pages 857 -- 868. ACM, 2012.
[13] R. G. Downey and M. R. Fellows. Parameterized Complexity. Springer Verlag, 1999.
[14] P. W. L. Fong. Relationship-based access control: protection model and policy language.
In R. S. Sandhu and E. Bertino, editors, CODASPY, pages 191 -- 202. ACM, 2011.
[15] A. A. Khan. Satisfiability and feasibility in a relationship-based workflow authorization
model. Master's thesis, University of Calgary, 2012.
[16] A. A. Khan and P. W. L. Fong. Satisfiability and feasibility in a relationship-based workflow
authorization model. In S. Foresti, M. Yung, and F. Martinelli, editors, ESORICS, volume
7459 of Lecture Notes in Computer Science, pages 109 -- 126. Springer, 2012.
[17] M. Kohler and A. Schaad. Proactive access control for business process-driven environments.
In ACSAC, pages 153 -- 162. IEEE Computer Society, 2008.
[18] R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford University Press, 2006.
[19] A. Schafer, C. Komusiewicz, H. Moser, and R. Niedermeier. Parameterized computational
complexity of finding small-diameter subgraphs. Optimization Letters, 6(5):883 -- 891, 2012.
[20] W. M. P. van der Aalst, A. H. M. ter Hofstede, B. Kiepuszewski, and A. P. Barros. Workflow
patterns. Distributed and Parallel Databases, 14(1):5 -- 51, 2003.
[21] Q. Wang and N. Li. Satisfiability and resiliency in workflow authorization systems. ACM
Trans. Inf. Syst. Secur., 13(4):40, 2010.
[22] S. A. White and D. Miers. BPMN Modeling and Reference Guide. Future Strategies,
Incorporated, 2008.
18
A Proofs
Proof of Corollary 6. For WSP2(=, 6=), in the worst case, each constraint has the form (ρ, s, S ′),
with s 6∈ S ′. Hence, the number of Type 2 constraints can be no greater than k2k−1. It now
follows from Theorem 5 that WSP2(=, 6=) is FPT. For WSP3(=, 6=), in the worst case, each
constraint has the form (ρ, S ′, S ′′). Thus, noting that (ρ, S ′, S ′′) is equivalent for WSP3(=, 6=),
the number of Type 3 constraints can be no greater than 2k · 2k = 22k, from which it follows that
WSP3(=, 6=) is FPT.
Proof of Proposition 8. The result follows by noting that determining strong satisfiability re-
quires us to check whether all ♯(W ) derived instances of W are satisfiable, while determining
weak satisfiability requires us to check whether at least one derived instance is satisfiable. The
complexity, in the worst case, is the same. The complexity of checking a single instance is
(k′ − 1)k′
, where k′ is the number of steps in the derived instance. The result now follows.
Proof of Theorem 9. First observe that may disregard the k operator in computing an upper
bound on ♯(W ). To see this, note that the parallel operator requires, like the serial operator,
that all steps in the sub-workflows are performed. In particular, an execution set for workflow
W1 k W2 has the form S1 ∪ S2, where Si is an execution set for workflow Wi.
Recall ⊕i represents the xor composition of i steps and ♯(⊕i ; ⊕j) = ♯(⊕j ; ⊕i) = ij.
We proceed by induction on k. For k = 2, we may construct
⊕1 ; ⊕1
and ⊕2,
thus the result holds for k = 2. For k = 3, we may construct three different workflows:
thus the result holds for k = 3. Finally, for k = 4, we may construct
⊕1 ; ⊕1 ; ⊕1, ⊕1 ; ⊕2,
and ⊕3,
⊕1 ; ⊕1 ; ⊕1 ; ⊕1, ⊕1 ; ⊕1 ; ⊕2, ⊕1 ; ⊕3, ⊕2 ; ⊕2
and ⊕4,
thus the result holds for k = 4.
Now consider k > 4 steps and suppose the result holds for all workflows constructed from
k − 1 or fewer steps. Then for any split of k into workflows W1 and W2 comprising k1 and k2
steps, respectively, such that k1 + k2 = k, we may form W1 ; W2 or W1 ⊕ W2. Clearly, for k > 4,
♯(W1 ; W2) > ♯(W1 ⊕ W2). Moreover, ♯(W1 ; W2) = ♯(W1) · ♯(W2).
First consider the case k = 3a and let ki = 3ai + bi, i = 1, 2, with bi ∈ {0, 1, 2}. We assume
If b1 = b2, then k1 and k2 are divisible by 3 and
(without loss of generality) that b1 6 b2.
♯(Wi) 6 3ai by the inductive hypothesis, whence
♯(W ) = ♯(W1) · ♯(W2) 6 3a1 · 3a2 = 3a.
If b1 = 1 and b2 = 2, then we have a1 + a2 = a − 1 and
♯(W ) = ♯(W1) · ♯(W2) 6 4 · 3a1−1 · 2 · 3a2 = 8 · 3a−2 < 3a
and the result holds.
Now consider the case k = 3a + 1. If b1 = 0, then b2 = 1 and a1 + a2 = a. Hence, by the
inductive hypothesis, we have
♯(W ) 6 3a1 · 4 · 3a2−1 = 4 · 3a−1,
19
as required. If b1 = b2, then we have bi = 2 and a1 + a2 = a − 1. Hence, by the inductive
hypothesis, we have
♯(W ) 6 2 · 3a1 · 2 · 3a2 = 4 · 3a−1,
as required.
Finally, consider the case k = 3a + 2. If b1 = 0, then b2 = 2 and a1 + a2 = a. Hence, by the
inductive hypothesis, we have
♯(W ) 6 3a1 · 2 · 3a2 = 2 · 3a,
as required. If b1 = b2, then bi = 1 and a1 + a2 = a. Hence, we have
♯(W ) 6 16 · 3a1+a2−2 =
16
9
· 3a < 2 · 3a.
as required.
Proof of Proposition 14. An instance of OWSP1 contains a set of constraints C and we may as-
sume that C contains at least two constraints of the form (ρi, s, s′) and (ρj, s′, s) with ρi 6= eρj. (If
no such constraints exist then OWSP1(ρ1, . . . , ρt) is identical to an instance of WSP1(ρ1, . . . , ρt).)
Observe that the number of linear extensions of (S, 6) (and hence possible execution schedules)
is determined only by k. Specifically, the number of linear extensions is no greater than k!. Note
also that in any execution of the workflow, either s precedes s′ or vice versa. Hence each linear
extension allows us to discard either (ρi, s, s′) or (ρj , s′, s) (since exactly one of them will be
irrelevant to the schedule defined by the linear extension), thus defining an instance of WSP
that contains fewer constraints than the original problem. In other words, we may consider our
instance of OWSP1 to be the disjunction of k! instances of WSP1. If each instance of WSP1 is
FPT, we can solve each of these instances, thus solving the original instance of OWSP1.
Proof of Theorem 15. A Type 1 constraint (ρ, s, s′) is satisfied by a plan α if (α(s), α(s′)) ∈ ρ and
is not satisfied ("violated") otherwise. In other words, (ρ, s, s′) is violated by α if (α(s), α(s′)) 6∈ ρ.
Equivalently, a constraint (ρ, s, s′) is violated iff (ρ, s, s′) is satisfied, where
ρ def= {(u, u′) ∈ U × U : (u, u′) 6∈ ρ} .
A Type 2 constraint (ρ, s, S ′), S ′ ⊆ S is violated if (ρ, s, s′) is violated for all s′ ∈ S ′. In other
words, (ρ, s, S ′) is violated iff the constraint expression
^
s′∈S ′
(ρ, s, s′)
is satisfied. Similarly, a Type 3 constraint (ρ, S ′, S ′′) is violated iff the constraint expression
^
s′ ∈S ′,s′′ ∈S ′′
(ρ, s′, s′′)
is satisfied. Finally a set of constraints {c1, . . . , ct} is violated if at least one ci is violated. In
other words, we can determine whether there exists a plan that violates a set of constraints by de-
termining if there exists a plan α that satisfies a constraint expression in disjunctive normal form,
where each clause is a conjunction of Type 1 constraints. We make the following observations.
• There are no more than c disjuncts, where c is the number of constraints in the original
workflow specification.
20
• A Type 2 constraint, when rewritten in the above way, gives rise to a conjunction of no
more than k − 1 Type 1 constraints, while a Type 3 constraint gives rise to no more than
k2 Type 1 constraints.
• There can be no more than k2k Type 2 constraints in a workflow specification and no more
than 4k Type 3 constraints.
• ∆ is ∆c and ∆c is ∆.
• By Theorem 5, the time taken to solve WSP1(∆, ∆c) (that is, WSP1(=, 6=)) is O(c(k −
1)k + kn), where c is the number of constraints.
Therefore, there exists an FPT algorithm to determine whether there exists a plan π in which
each user is authorized and a constraint that π does not satisfy, since we need only find a single
disjunct that is true, and each disjunct represents a workflow specification containing only Type
1 constraints. The time taken to solve this new problem is O(k2k−1((k − 1)k+1 + kn)) for Type
2 constraints and O(4k(k(k − 1)k+1 + kn)) for Type 3 constraints.
21
|
1806.01119 | 1 | 1806 | 2018-06-04T13:51:29 | Covering with Clubs: Complexity and Approximability | [
"cs.DS"
] | Finding cohesive subgraphs in a network is a well-known problem in graph theory. Several alternative formulations of cohesive subgraph have been proposed, a notable example being $s$-club, which is a subgraph where each vertex is at distance at most $s$ to the others. Here we consider the problem of covering a given graph with the minimum number of $s$-clubs. We study the computational and approximation complexity of this problem, when $s$ is equal to 2 or 3. First, we show that deciding if there exists a cover of a graph with three $2$-clubs is NP-complete, and that deciding if there exists a cover of a graph with two $3$-clubs is NP-complete. Then, we consider the approximation complexity of covering a graph with the minimum number of $2$-clubs and $3$-clubs. We show that, given a graph $G=(V,E)$ to be covered, covering $G$ with the minimum number of $2$-clubs is not approximable within factor $O(|V|^{1/2 -\varepsilon})$, for any $\varepsilon>0$, and covering $G$ with the minimum number of $3$-clubs is not approximable within factor $O(|V|^{1 -\varepsilon})$, for any $\varepsilon>0$. On the positive side, we give an approximation algorithm of factor $2|V|^{1/2}\log^{3/2} |V|$ for covering a graph with the minimum number of $2$-clubs. | cs.DS | cs |
Covering with Clubs: Complexity and Approximability
Riccardo Dondi1, Giancarlo Mauri2, Florian Sikora3, and Italo Zoppis2
1 Universit`a degli Studi di Bergamo, Bergamo, Italy [email protected]
2 Universit`a degli Studi di Milano-Bicocca, Milano - Italy {mauri,zoppis}@disco.unimib.it
3 Universit´e Paris-Dauphine, PSL Research University, CNRS UMR 7243, LAMSADE, 75016 Paris,
France [email protected]
Abstract. Finding cohesive subgraphs in a network is a well-known problem in graph
theory. Several alternative formulations of cohesive subgraph have been proposed, a notable
example being s-club, which is a subgraph where each vertex is at distance at most s to the
others. Here we consider the problem of covering a given graph with the minimum number of
s-clubs. We study the computational and approximation complexity of this problem, when
s is equal to 2 or 3. First, we show that deciding if there exists a cover of a graph with three
2-clubs is NP-complete, and that deciding if there exists a cover of a graph with two 3-clubs
is NP-complete. Then, we consider the approximation complexity of covering a graph with
the minimum number of 2-clubs and 3-clubs. We show that, given a graph G = (V, E) to
be covered, covering G with the minimum number of 2-clubs is not approximable within
factor O(V 1/2−ε), for any ε > 0, and covering G with the minimum number of 3-clubs is
not approximable within factor O(V 1−ε), for any ε > 0. On the positive side, we give an
approximation algorithm of factor 2V 1/2 log3/2 V for covering a graph with the minimum
number of 2-clubs.
1 Introduction
The quest for modules inside a network is a well-known and deeply studied problem in network
analysis, with several application in different fields, like computational biology or social network
analysis. A highly investigated problem is that of finding cohesive subgroups inside a network
which in graph theory translates in highly connected subgraphs. A common approach is to look for
cliques (i.e. complete graphs), and several combinatorial problems have been considered, notable
examples being the Maximum Clique problem ([11, GT19]), the Minimum Clique Cover problem ([11,
GT17]), and the Minimum Clique Partition problem ([11, GT15]). This last is a classical problem in
theoretical computer science, whose goal is to partition the vertices of a graph into the minimum
number of cliques. The Minimum Clique Partition problem has been deeply studied since the seminal
paper of Karp [15], studying its complexity in several graph classes [5,6,21,9].
In some cases, asking for a complete subgraph is too restrictive, as interesting highly connected
graphs may have some missing edges due to noise in the data considered or because some pair
may not be directly connected by an edge in the subgraph of interest. To overcome this limitation
of the clique approach, alternative definitions of highly connected graphs have been proposed,
leading to the concept of relaxed clique [16]. A relaxed clique is a graph G = (V, E) whose vertices
satisfy a property which is a relaxation of the clique property. Indeed, a clique is a subgraph whose
vertices are all at distance one from each other and have the same degree (the size of the clique
minus one). Different definitions of relaxed clique are obtained by modifying one of the properties
of clique, thus leading to distance-based relaxed cliques, degree-based relaxed cliques, and so on
(see for example [16]).
In this paper, we focus on a distance-based relaxation. In a clique all the vertices are required
to be at distance at most one from each other. Here this constraint is relaxed, so that the ver-
tices have to be at distance at most s, for an integer s (cid:62) 1. A subgraph whose vertices are all
distance at most s is called an s-club (notice that, when s = 1, an s-club is exactly a clique).
The identification of s-clubs inside a network has been applied to social networks [19,1,18,20,23],
and biological networks [3]. Interesting recent studies have shown the relevance of finding s-clubs
2
in a network [18,20], in particular focusing on finding 2-clubs in real networks like DBLP or a
European corporate network.
Contributions to the study of s-clubs mainly focus on the Maximum s-Club problem, that is
the problem of finding an s-club of maximum size. Maximum s-Club is known to be NP-hard, for
each s (cid:62) 1 [4]. Even deciding whether there exists an s-club larger than a given size in a graph
of diameter s + 1 is NP-complete, for each s (cid:62) 1 [3]. The Maximum s-Club problem has been
studied also in the approximability and parameterized complexity framework. A polynomial-time
approximation algorithm with factor V 1/2 for every s (cid:62) 2 on an input graph G = (V, E) has
been designed [2]. This is optimal, since the problem is not approximable within factor V 1/2−ε,
on an input graph G = (V, E), for each ε > 0 and s (cid:62) 2 [2]. As for the parameterized complexity
framework, the problem is known to be fixed-parameter tractable, when parameterized by the
size of an s-club [22,17,7]. The Maximum s-Club problem has been investigated also for structural
parameters and specific graph classes [13,12].
In this paper, we consider a different combinatorial problem, where we aim at covering the
vertices of a network with a set of subgraphs. Similar to Minimum Clique Partition, we consider the
problem of covering a graph with the minimum number of s-clubs such that each vertex belongs
to an s-club. We denote this problem by Min s-Club Cover, and we focus in particular on the cases
s = 2 and s = 3. We show some analogies and differences between Min s-Club Cover and Minimum
Clique Partition. We start in Section 3 by considering the computational complexity of the problem
of covering a graph with two or three s-clubs. This is motivated by the fact that Clique Partition
is known to be in P when we ask whether there exists a partition of the graph consisting of two
cliques, while it is NP-hard to decide whether there exists a partition of the graph consisting of
three cliques [10]. As for Clique Partition, we show that it is NP-complete to decide whether there
exist three 2-clubs that cover a graph. On the other hand, we show that, unlike Clique Partition,
it is NP-complete to decide whether there exist two 3-clubs that cover a graph. These two results
imply also that Min 2-Club Cover and Min 3-Club Cover do not belong to the class XP for the
parameter "number of clubs" in a cover.
Then, we consider the approximation complexity of Min 2-Club Cover and Min 3-Club Cover.
We recall that, given an input graph G = (V, E), Minimum Clique Partition is not approximable
within factor O(V 1−ε), for any ε > 0, unless P = N P [24]. Here we show that Min 2-Club Cover
has a slightly different behavior, while Min 3-Club Cover is similar to Clique Partition. Indeed, in
Section 4 we prove that Min 2-Club Cover is not approximable within factor O(V 1/2−ε), for any
ε > 0, unless P = N P , while Min 3-Club Cover is not approximable within factor O(V 1−ε), for
any ε > 0, unless P = N P . In Section 5, we present a greedy approximation algorithm that has
factor 2V 1/2 log3/2 V for Min 2-Club Cover, which almost match the inapproximability result for
the problem. We start the paper by giving in Section 2 some definitions and by formally defining
the problem we are interested in.
2 Preliminaries
Given a graph G = (V, E) and a subset V (cid:48) ⊆ V , we denote by G[V (cid:48)] the subgraph of G induced
by V (cid:48). Given two vertices u, v ∈ V , the distance between u and v in G, denoted by dG(u, v), is the
length of a shortest path from u to v. The diameter of a graph G = (V, E) is the maximum distance
between two vertices of V . Given a graph G = (V, E) and a vertex v ∈ V , we denote by NG(v) the
set of neighbors of v, that is NG(v) = {u : {v, u} ∈ E}. We denote by NG[v] the close neighborhood
of V , that is NG[v] = NG(v) ∪ {v}. Define N l
G(v) = {u : u has distance at most l from v}, with
1 (cid:54) l (cid:54) 2. Given a set of vertices X ⊆ V and l, with 1 (cid:54) l (cid:54) 2, define N l
G(u). We
may omit the subscript G when it is clear from the context. Now, we give the definition of s-club,
which is fundamental for the paper.
Definition 1. Given a graph G = (V, E), and a subset V (cid:48) ⊆ V , G[V (cid:48)] is an s-club if it has
diameter at most s.
G(X) =(cid:83)
u∈X N l
Notice that an s-club must be a connected graph. We present now the formal definition of the
Minimum s-Club Cover problem we are interested in.
3
Minimum s-Club Cover (Min s-Club Cover)
Input: a graph G = (V, E) and an integer s (cid:62) 2.
Output: a minimum cardinality collection S = {V1, . . . , Vh} such that, for each i with 1 (cid:54) i (cid:54) h,
Vi ⊆ V , G[Vi] is an s-club, and, for each vertex v ∈ V , there exists a set Vj, with 1 (cid:54) j (cid:54) h, such
that v ∈ Vj.
We denote by s-Club Cover(h), with 1 (cid:54) h (cid:54) V , the decision version of Min s-Club Cover that
asks whether there exists a cover of G consisting of at most h s-clubs.
Notice that while in Minimum Clique Partition we can assume that the cliques that cover a graph
G = (V, E) partition V , hence the cliques are vertex disjoint, we cannot make this assumption
for Min s-Club Cover. Indeed, in a solution of Min s-Club Cover, a vertex may be covered by more
than one s-club, in order to have a cover consisting of the minimum number of s-clubs. Consider
the example of Fig. 1. The two 2-clubs induced by {v1, v2, v3, v4, v5} and {v1, v6, v7, v8, v9} cover
G, and both these 2-clubs contain vertex v1. However, if we ask for a partition of G, we need at
least three 2-clubs. This difference between Minimum Clique Partition and Min s-Club Cover is due
to the fact that, while being a clique is a hereditary property, this is not the case for being an
s-club. If a graph G is an s-club, then a subgraph of G may not be an s-club (for example a star
is a 2-club, but the subgraph obtained by removing its center is not anymore a 2-club).
G
v3
v2
v6
v7
v1
v4
v5
v9
v8
Fig. 1. A graph G and a cover consisting of two 2-clubs (induced by the vertices in the ovals). Notice that
the 2-clubs of this cover must both contain vertex v1.
3 Computational Complexity
In this section we investigate the computational complexity of 2-Club Cover and 3-Club Cover and
we show that 2-Club Cover(3), that is deciding whether there exists a cover of a graph G with
three 2-clubs, and 3-Club Cover(2), that is deciding whether there exists a cover of a graph G with
two 3-clubs, are NP-complete.
3.1 2-Club Cover(3) is NP-complete
In this section we show that 2-Club Cover(3) is NP-complete by giving a reduction from the
3-Clique Partition problem, that is the problem of computing whether there exists a partition of a
graph Gp = (V p, Ep) in three cliques. Consider an instance Gp = (V p, Ep) of 3-Clique Partition,
we construct an instance G = (V, E) of 2-Club Cover(3) (see Fig. 2). The vertex set V is defined
as follows:
V = {wi : vi ∈ V p} ∪ {wi,j : {vi, vj} ∈ Ep ∧ i < j}}
The set E of edges is defined as follows:
E = {{wi, wi,j},{wi, wh,i} : vi ∈ V p, wi, wi,j, wh,i ∈ V }∪
{{wi,j, wi,l},{wi,j, wh,i},{wh,i, wz,i} : wi,j, wi,l, wh,i, wz,i ∈ V }
Before giving the main results of this section, we prove a property of G.
4
Lemma 2. Let Gp = (V p, Ep) be an instance of 3-Clique Partition and let G = (V, E) be the cor-
responding instance of 2-Club Cover(3). Then, given two vertices vi, vj ∈ V p and the corresponding
vertices wi, wj ∈ V :
– if {vi, vj} ∈ Ep, then dG(wi, wj) = 2
– if {vi, vj} /∈ Ep, then dG(wi, wj) (cid:62) 3
Proof. Notice that NG(wi) = {wi,z : {vi, vz} ∈ Ep ∧ i < z} ∪ {wh,i : {vi, vh} ∈ Ep ∧ h < i}. It
follows that wj ∈ N 2
G(wi) if and only if there exists a vertex wi,j (or wj,i), which is adjacent to
both wi and wj. But then, by construction, wj ∈ N 2
(cid:117)(cid:116)
G(wi) if and only if {vi, vj} ∈ Ep.
We are now able to prove the main properties of the reduction.
Lemma 3. Let Gp = (V p, Ep) be a graph input of 3-Clique Partition and let G = (V, E) be the
corresponding instance of 2-Club Cover(3). Then, given a solution of 3-Clique Partition on Gp =
(V p, Ep), we can compute in polynomial time a solution of 2-Club Cover(3) on G = (V, E).
3 ⊆ V p be
Proof. Consider a solution of 3-Clique Partition on Gp = (V p, Ep), and let V p
the sets of vertices of Gp that partition V p. We define a solution of 2-Club Cover(3) on G = (V, E)
as follows. For each d, with 1 (cid:54) d (cid:54) 3, define
1 , V p
2 , V p
Vd = {wj ∈ V : vj ∈ V p
d } ∪ {wi,j : vi ∈ V p
d }
We show that each G[Vd], with 1 (cid:54) d (cid:54) 3, is a 2-club. Consider two vertices wi, wj ∈ Vd, with
1 (cid:54) i < j (cid:54) V . Since they correspond to two vertices vi, vj ∈ V p that belong to a clique of Gp,
it follows that {vi, vj} ∈ Ep and wi,j ∈ Vd. Thus dG[Vd](wi, wj) = 2. Now, consider the vertices
wi ∈ Vd, with 1 (cid:54) i (cid:54) V , and wh,z ∈ Vd, with 1 (cid:54) h < z (cid:54) V . If i = h or i = z, assume
w.l.o.g. i = h, then by construction dG[Vd](wi, wi,z) = 1. Assume that i (cid:54)= h and i (cid:54)= z (assume
w.l.o.g. that i < h < z), since wh,z ∈ Vd, it follows that wh ∈ Vd. Since wi, wh ∈ Vd, it follows
that wi,h ∈ Vd. By construction, there exist edges {wi,h, wh,z}, {wi, wi,h} in Ep, thus implying
that dG[Vd](wi, wh,z) = 2. Finally, consider two vertices wi,j, wh,z ∈ Vd, with 1 (cid:54) i < j (cid:54) V and
1 (cid:54) h < z (cid:54) V . Then, by construction, wi ∈ Vd and wh ∈ Vd. But then, wi,h belongs to Vd, and,
by construction, {wi,j, wi,h} ∈ E and {wh,z, wi,h} ∈ E. It follows that dG[Vd](wi,j, wh,z) = 2.
that V = V1 ∪ V2 ∪ V3, thus G[V1], G[V2], G[V3] covers G.
We conclude the proof observing that, by construction, since V p
3 partition V p, it holds
(cid:117)(cid:116)
1 , V p
2 , V p
Gp
v4
v5
v1
v2
G
w4
w1,4
w1
w4,5
w5
w1,2
w2,5
w2
v3
w1,3
w2,3
w3
Fig. 2. An example of a graph Gp input of 3-Clique Partition and the corresponding graph G input of
2-Club Cover(3).
Based on Lemma 2, we can prove the following result.
Lemma 4. Let Gp = (V p, Ep) be a graph input of 3-Clique Partition and let G = (V, E) be the
corresponding instance of 2-Club Cover(3). Then, given a solution of 2-Club Cover(3) on G =
(V, E), we can compute in polynomial time a solution of 3-Clique Partition on Gp = (V p, Ep).
5
Proof. Consider a solution of 2-Club Cover(3) on G = (V, E) consisting of three 2-clubs G[V1],
G[v2], G[V3]. Consider a 2-club G[Vd], with 1 (cid:54) d (cid:54) 3. By Lemma 2, it follows that, for each
wi, wj ∈ Vd, {vi, vj} ∈ E. As a consequence, we can define three cliques Gp[V p
2 ], Gp[V p
3 ]
in Gp as follows. For each d, with 1 (cid:54) d (cid:54) 3, V p
1 ], Gp[V p
d is defined as:
d = {vi : wi ∈ Vd}
V p
Next, we show that G[V p
it holds {vi, vj} ∈ E, thus by construction {vi, vj} ∈ Ep and G[V p
since V1 ∪ V2 ∪ V3 = V , then V p
3 = V p. Notice that V p
but, starting from V p
cliques.
d ], with 1 (cid:54) d (cid:54) 3, is indeed a clique. By Lemma 2 if wi, wj ∈ Vd then
d ] is a clique in Gp. Moreover,
1 , V p
3 may not be disjoint,
3 , it is easy to compute in polynomial time a partition of Gp in three
(cid:117)(cid:116)
1 ∪ V p
2 ∪ V p
1 , V p
2 , V p
2 , V p
Now, we can prove the main result of this section.
Theorem 5. 2-Club Cover(3) is NP-complete.
Proof. By Lemma 3 and Lemma 4 and from the NP-hardness of 3-Clique Partition [15], it follows
that 2-Club Cover(3) is NP-hard. The membership to NP follows easily from the fact that, given
three 2-clubs of G, it can be checked in polynomial time whether they are 2-clubs and cover all
(cid:117)(cid:116)
vertices of G.
3.2 3-Club Cover(2) is NP-complete
In this section we show that 3-Club Cover(2) is NP-complete by giving a reduction from a variant
of Sat called 5-Double-Sat. Recall that a literal is positive if it is a non-negated variable, while it
is negative if it is a negated variable.
Given a collection of clauses C = {C1, . . . , Cp} over the set of variables X = {x1, . . . , xq}, where
each Ci ∈ C, with 1 (cid:54) i (cid:54) p, contains exactly five literals and does not contain both a variable and
its negation, 5-Double-Sat asks for a truth assignment to the variables in X such that each clause
Ci, with 1 (cid:54) i (cid:54) p, is double-satisfied. A clause Ci is double-satisfied by a truth assignment f to the
variables X if there exist a positive literal and a negative literal in Ci that are both satisfied by f .
Notice that we assume that there exist at least one positive literal and at least one negative literal
in each clause Ci, with 1 (cid:54) i (cid:54) p, otherwise Ci cannot be doubled-satisfied. Moreover, we assume
that each variable in an instance of 5-Double-Sat appears both as a positive literal and a negative
literal in the instance. Notice that if this is not the case, for example a variable appears only as
a positive literal, we can assign a true value to the variable, as defining an assignment to false
does not contribute to double-satisfy any clause. First, we show that 5-Double-Sat is NP-complete,
which may be of independent interest.
Theorem 6. 5-Double-Sat is NP-complete.
Proof. We reduce from 3-Sat, where given a set X3 of variables and a set C3 of clauses, which are
a disjunction of 3 literals (a variable or the negation of a variable), we want to find an assignment
to the variables such that all clauses are satisfied. Moreover, we assume that each clause in C3
does not contain a positive variable x and its negation x, since such a clause is obviously satisfied
by any assignment. The same property holds also for the instance of 5-Double-Sat we construct.
Consider an instance (X3,C3) of 3-Sat, we construct an instance (X,C) of 5-Double-Sat as
follows. Define X = X3 ∪ XN , where X3 ∩ XN = ∅ and XN is defined as follows:
The set C of clauses is defined as follows:
XN = {xC,i,1, xC,i,2 : Ci ∈ C3}
C = {Ci,1, Ci,2 : Ci ∈ C3}
where Ci,1, Ci,2 are defined as follows. Consider Ci ∈ C3 = (li,1 ∨ li,2 ∨ li,3), where li,p, with
1 (cid:54) p (cid:54) 3 is a literal, that is a variable (a positive literal) or a negated variable (a negative literal),
the two clauses Ci,1 and Ci,2 are defined as follows:
6
– Ci,1 = li,1 ∨ li,2 ∨ li,3 ∨ xC,i,1 ∨ xC,i,2
– Ci,2 = li,1 ∨ li,2 ∨ li,3 ∨ xC,i,1 ∨ xC,i,2
We claim that (X3,C3) is satisfiable if and only if (X,C) is double-satisfiable.
Assume that (X3,C3) is satisfiable and let f be an assignment to the variables on X that
satisfies C3. Consider a clause Ci in C3, with 1 (cid:54) i (cid:54) C3. Since it is satisfied by f , it follows that
there exists a literal li,p of Ci, with 1 (cid:54) p (cid:54) 3, that is satisfied by f . Define an assignment f(cid:48) on
X that is identical to f on X3 and, if li,p is positive, then assigns value false to both xC,i,1 and
xC,i,2, if li,p is negative, then assigns value true to both xC,i,1 and xC,i,2. It follows that both Ci,1
and Ci,2 are double-satisfied by f(cid:48).
Assume that (X,C) is double-satisfied by an assignment f(cid:48). Consider two clauses Ci,1 and
Ci,2, with 1 (cid:54) i (cid:54) C, that are double-satisfied by f(cid:48), we claim that there exists at least one
literal of Ci,1 and Ci,2 not in XN which is satisfied. Assume this is not the case, then, if Ci,1
is double-satisfied, it follows that xC,i,1 is true and xC,i,2 is false, thus implying that Ci,2 is not
double-satisfied. Then, an assignment f that is identical to f(cid:48) restricted to X3 satisfies each clause
in C.
Now, since 3-Sat is NP-complete [15], it follows that 5-Double-Sat is NP-hard. The membership
to NP follows from the observation that, given an assignment to the variables on X, we can check
in polynomial-time whether each clause in C is double-satisfied or not.
(cid:117)(cid:116)
Let us now give the construction of the reduction from 5-Double-Sat to 3-Club Cover(2). Con-
sider an instance of 5-Double-Sat consisting of a set C of clauses C1, . . . , Cp over set X = {x1, . . . , xq}
of variables. We assume that it is not possible to double-satisfy all the clauses by setting at most
two variables to true or to false (this can be easily checked in polynomial-time).
Before giving the details, we present an overview of the reduction. Given an instance (X,C)
of 5-Double-Sat, for each positive literal xi, with 1 (cid:54) i (cid:54) q, we define vertices xT
i,2 and for
i . Moreover, for each clause Cj ∈ C,
each negative literal xi, with 1 (cid:54) i (cid:54) q, we define a vertex xF
with 1 (cid:54) j (cid:54) p, we define a vertex vC,j. We define other vertices to ensure that some vertices
have distance not greater than three and to force the membership to one of the two 3-clubs of
the solution (see Lemma 7). The construction implies that for each i with 1 (cid:54) i (cid:54) q, xT
i,1 and xF
i
belong to different 3-clubs (see Lemma 8); this corresponds to a truth assignment to the variables
in X. Then, we are able to show that each vertex vC,j belongs to the same 3-club of a vertex
i,1, with 1 (cid:54) i (cid:54) q, and of a vertex xF
h , with 1 (cid:54) h (cid:54) q, adjacent to vC,j (see Lemma 10); these
xT
vertices correspond to a positive literal xi and a negative literal xh, respectively, that are satisfied
by a truth assignment, hence Cj is double-satisfied.
Now, we give the details of the reduction. Let (X,C) be an instance of 5-Double-Sat, we con-
struct an instance G = (V, E) of 3-Club Cover(2) as follows (see Fig. 3). The vertex set V is defined
as follows:
i,1, xT
V = {r, r(cid:48), rT , r(cid:48)
T , r∗
T , rF , r(cid:48)
F} ∪ {xT
i,1, xT
i,2, xF
i
: xi ∈ X} ∪ {vC,j : Cj ∈ C} ∪ {y1, y2, y}
The edge set E is defined as follows:
E = {{r, r(cid:48)},{{r(cid:48), rT},{r(cid:48), r∗
∪{{rF , xF
i } : xi ∈ X} ∪ {{r(cid:48)
T , xT
{{xT
i,1, xT
i,2} : xi ∈ X} ∪ {{r∗
i,1} : xi ∈ X} ∪ {{r(cid:48)
T}{r(cid:48), rF}} ∪ {{rT , xT
F , xF
i,2},{y1, xT
T , xT
i,1} : xi ∈ X}
i } : xi ∈ X}∪
i,2} : xi ∈ X}∪
i , vC,j} : xi ∈ Cj}∪
F}}
T},{y1, r(cid:48)
{{xT
i,2, xF
j } : xi, xj ∈ X, i (cid:54)= j} ∪ {{xT
i,1, vC,j} : xi ∈ Cj} ∪ {{xF
{{vC,j, y} : Cj ∈ C} ∪ {{y, y2},{y1, y2},{y1, r(cid:48)
We start by proving some properties of the graph G.
Lemma 7. Consider an instance (C, X) of 5-Double-Sat and let G = (V, E) be the corresponding
instance of 3-Club Cover(2). Then, (1) dG(r(cid:48), y) > 3, (2) dG(r, y) > 3, (3) dG(r, vC,j) > 3, for
each j with 1 (cid:54) j (cid:54) p, and (4) dG(r, r(cid:48)
F ) > 3, dG(r, r(cid:48)
T ) > 3.
7
i,1, xT
T or rF is adjacent to vertices xT
Proof. We start by proving (1). Notice that any path from r(cid:48) to y must pass through rT , r∗
T or rF .
i , with 1 (cid:54) i (cid:54) q (in addition to r(cid:48)),
Each of rT , r∗
and none of these vertices is adjacent to y, thus concluding that dG(r(cid:48), y) > 3. Moreover, observe
that for each vertex vC,j, with 1 (cid:54) j (cid:54) p, there exists a vertex xT
h , with
1 (cid:54) h (cid:54) q, that is adjacent to vC,j, with 1 (cid:54) j (cid:54) p, thus dG(r(cid:48), vCj ) = 3, for each j with 1 (cid:54) j (cid:54) p.
As a consequence of (1), it follows that (2) holds, that is dG(r, y) > 3. Since dG(r(cid:48), vCj ) = 3, for
each j with 1 (cid:54) j (cid:54) p, it holds (3) dG(r, vC,j) > 3.
T , rT , rF} and that none of the vertices in
Finally, we prove (4). Notice that N 2
(cid:117)(cid:116)
G(r) is adjacent to r(cid:48)
N 2
i,1, with 1 (cid:54) i (cid:54) q, or xF
T , thus dG(r, r(cid:48)
G(r) = {r(cid:48), r∗
F and r(cid:48)
i,2 and xF
F ) > 3.
rT
xT
i,1
r(cid:48)
T
r
r(cid:48)
r∗
T
xT
i,2
y1
vC,j
y
y2
rF
xF
i
r(cid:48)
F
Fig. 3. Schematic construction for the reduction from 5-Double-Sat to 3-Club Cover(2).
Consider two sets V1 ⊆ V and V2 ⊆ V , such that G[V1] and G[V2] are two 3-clubs of G that
cover G. As a consequence of Lemma 7, it follows that r and r(cid:48) are in exactly one of G[V1], G[V2],
F , y and vC,j, for each j with 1 (cid:54) j (cid:54) p, belong to G[V2] and not to
w.l.o.g. G[V1], while r(cid:48)
G[V1].
T , r(cid:48)
Next, we show a crucial property of the graph G built by the reduction.
Lemma 8. Given an instance (C, X) of 5-Double-Sat, let G = (V, E) be the corresponding instance
of 3-Club Cover(2). Then, for each i with 1 (cid:54) i (cid:54) q, dG(xT
Proof. Consider a path π of minimum length that connects xT
notice that, by construction, the path π after xT
i,2 or vC,j, with 1 (cid:54) j (cid:54) p.
xT
We consider the first case, that is the path π after xT
in π is either r(cid:48) or xT
that in this case the path π has length greater than three.
We consider the second case, that is the path π after xT
i , with 1 (cid:54) i (cid:54) q. First,
i,1 must pass through one of these vertices: rT , r(cid:48)
T ,
i,1 passes through rT . Now, the next vertex
i , it follows
h,1, with 1 (cid:54) h (cid:54) q. Since both r(cid:48) and xT
h,1 are not adjacent to xF
i,1 passes through r(cid:48)
T . Now, after r(cid:48)
i,1 and xF
i ) > 3.
i,1, xF
passes through either y1 or xT
it follows that in this case the path π has length greater than three.
h,1, with 1 (cid:54) h (cid:54) q. Since both y1 and xT
T , π
h,1 are not adjacent to xF
i ,
8
T or y1 or xF
i,1 passes through xT
We consider the third case, that is the path after xT
We consider the last case, that is the path after xT
h , with 1 (cid:54) h (cid:54) q and h (cid:54)= i. Since r∗
of π is either r∗
xF
i , it follows that in this case the path π has length greater than three.
i,2. Now, the next vertex
h are not adjacent to
i,1 passes through vC,j, with 1 (cid:54) j (cid:54) p. We
have assumed that xi and xi do not belong to the same clause, thus by construction xF
is not
i
incident in vC,j. It follows that after vC,j, the path π must pass through either y or xT
h,1, with
1 (cid:54) h (cid:54) q, or xF
z are not adjacent to
xF
i , it follows that also in this case the path π has length greater than three, thus concluding the
(cid:117)(cid:116)
proof.
z , 1 (cid:54) z (cid:54) q and z (cid:54)= i. Once again, since y, xT
T , y1 and xF
h,1 and xF
Now, we are able to prove the main results of this section.
Lemma 9. Given an instance (C, X) of 5-Double-Sat, let G = (V, E) be the corresponding instance
of 3-Club Cover(2). Then, given a truth assignment that double-satisfies C, we can compute in
polynomial-time two 3-clubs that cover G.
Proof. Consider a truth assignment f on the set X of variables that double-satisfies C. In the
following we construct two 3-clubs G[V1] and G[V2] that cover G. The two sets V1, V2 are defined
as follows:
V1 = {r, r(cid:48), rT , r∗
T , rF} ∪ {xT
i,1, xT
i,2 : f (xi) = f alse} ∪ {xF
i , : f (xi) = true}
V2 = {r(cid:48)
T , r(cid:48)
F , y, y1, y2} ∪ {xT
i,2 : f (xi) = true} ∪ {xF
i,1, xT
{vC,j : 1 (cid:54) j (cid:54) p}
i
: f (xi) = f alse∪}
i,1, xT
h,2 and xF
i,1, from each vertex xT
i,2, and from each vertex xF
i,1) = 2 and dG[V1](r(cid:48), xT
i,1, with 1 (cid:54) i (cid:54) q, xT
i,1, rT , xT
Let us first consider G[V1]. By construction, dG[V1](r, xT
h,2, with 1 (cid:54) h (cid:54) q and xF
h,1 and xT
h,2, vertices xT
i,1) = 3 and dG[V1](r, xT
i,2) = 2, for each i with 1 (cid:54) i (cid:54) q, and dG[V1](r(cid:48), xF
Next, we show that G[V1] and G[V2] are indeed two 3-clubs that cover G. First, notice that
V1 ∪ V2 = V , hence G[V1] and G[V2] cover G. Next, we show that both G[V1] and G[V2] are indeed
3-clubs.
i,2) = 3, for
i ) = 3, for each i with 1 (cid:54) i (cid:54) i (cid:54) q. Moreover,
each i with 1 (cid:54) i (cid:54) i (cid:54) q, and dG[V1](r, xF
dG[V1](r(cid:48), xT
i ) = 2, for
each i with 1 (cid:54) i (cid:54) i (cid:54) q. As a consequence, it holds that rT , r(cid:48)
T and rF have distance at most
three in G[V1] from each vertex xT
i . Since r, rT ,
T and rF are in N (r(cid:48)), it follows that r, r(cid:48), rT , r∗
r∗
T and rF are at distance at most 2 in G[V1].
j , with 1 (cid:54) j (cid:54) q. Since
Hence, we focus on vertices xT
i,1, xT
there exists a path that passes trough xT
h,1 are at distance
at most two in G[V1], while xT
h,2 are at distance at most three in G[V1] (if i = h they are at
j are at distance one in G[V1], since h (cid:54)= j and {xT
j } ∈ E
distance one). Vertices xT
by construction. Finally, xT
j are at distance two in G[V1], since there exists a path that
passes trough xT
i,2 ∈ V2, then
We now consider G[V2]. We recall that, for each i with 1 (cid:54) i (cid:54) q, if xT
i ∈ V1. Furthermore, we recall that we assume that each xi appears as a positive and a negative
xF
i,1, with 1 (cid:54) i (cid:54) q, and each vertex xF
literal in the instance of 5-Double-Sat, thus each vertex xT
h ,
with 1 (cid:54) h (cid:54) q, are connected to some VC,j, with 1 (cid:54) j (cid:54) p.
First, notice that vertex y is at distance at most three in G[V2] from each vertex of V2, since
it has distance one in G[V2] from each vertex vC,j, with 1 (cid:54) j (cid:54) p, thus distance two from xT
i,1,
with 1 (cid:54) i (cid:54) q, and xF
F . Since y
is adjacent to y2, it has distance one from y2 and two from y1.
Now, consider a vertex vC,j, with 1 (cid:54) j (cid:54) p. Since f double-satisfies C, it follows that there
z , with 1 (cid:54) z (cid:54) q, which are connected to
T and from r(cid:48)
F , and at most 3 from each
z ∈ V2, with 1 (cid:54) z (cid:54) q. Furthermore, notice that,
h,2 ∈ V2, with 1 (cid:54) h (cid:54) q and h (cid:54)= z, then
exist two vertices in V2, xT
vC,j. It follows that vC,j has distance 2 in G[V2] from r(cid:48)
h,1 ∈ V2, with 1 (cid:54) h (cid:54) q, and from each xF
xT
since vC,j is adjacent to xF
in G[V1], as i (cid:54)= j. It follows that G[V1] is a 3-club.
i,1, xT
h , with 1 (cid:54) h (cid:54) q, and three from xT
i,1, with 1 (cid:54) i (cid:54) q, and xF
i,2, with 1 (cid:54) i (cid:54) q, r(cid:48)
z and xF
z is adjacent to each xT
i,1, xT
i,2 and xF
j
T and r(cid:48)
i,1 and xF
h,2, xF
9
Consider a vertex xT
h,2 of G[V2]. Moreover, it has distance two from y1 and three from y2 and r(cid:48)
vC,j has distance at most two in G[V2] from each xT
has distance two and three respectively, from y2 and y1, in G[V2].
most three in G[V2] from any vC,j, with 1 (cid:54) j (cid:54) p, and two from y. Since xT
has distance at most two from each other vertex xT
vertex xT
is adjacent to every vertex xF
most two from every vertex xF
h,2 ∈ V2. Finally, since vC,j is adjacent to y, it
i,1 ∈ V2, with 1 (cid:54) i (cid:54) q. We have already shown that it has distance at
T , it
h,1, with 1 (cid:54) h (cid:54) q, and three from each other
F . Since xT
i,2
h,1 has distance at
i,2 ∈ V2, with 1 (cid:54) i (cid:54) q. We have already shown that it has distance at
i,1, it has distance three from y and
z ∈ V2, with 1 (cid:54) z (cid:54) q,
i,2 has distance two from each vertex
i,2 is
most two from each vC,j in G[V2]. Since it is connected to xT
two from r(cid:48)
i,2 has distance at most two from r(cid:48)
xT
h,2 in G[V2], with 1 (cid:54) i (cid:54) q, since by construction they are both adjacent to y1. Since xT
xT
adjacent to y1, thus it has distance at most two from y2 in G[V2].
z ∈ V2, with 1 (cid:54) z (cid:54) q, as z (cid:54)= i, it follows that xT
z ∈ V2.
i,2 is adjacent to every vertex xF
T in G[V2]. By construction xT
i,1 is adjacent to r(cid:48)
F in G[V2]. Moreover, xT
Consider a vertex xT
Consider a vertex xF
h , with 1 (cid:54) h (cid:54) q. It has distance one from r(cid:48)
two from y1 and three from y2 in G[V2]. Moreover, xF
thus it has distance two from each xT
there exists at least one vC,j, with 1 (cid:54) j (cid:54) p, adjacent to xF
and three from each vC,z in G[V2].
Finally, we consider vertices r(cid:48)
i,1 and distance three from r(cid:48)
T , r(cid:48)
h is adjacent to each xT
F in G[V2], and thus distance
i,2 ∈ V2, with 1 (cid:54) i (cid:54) q,
T in G[V2]. Since by construction
h has distance two from y
h , thus xF
T , r(cid:48)
T , r(cid:48)
F , y1 and y2. Notice that it suffices to show that these vertices
have pairwise distance at most three in G[V2], since we have previously shown that any other vertex
F , y2 ∈ N (y1), they are
of V2 has distance at most three from these vertices in G[V2]. Since r(cid:48)
(cid:117)(cid:116)
all at distance at most two. It follows that G[V2] is a 3-club, thus concluding the proof.
Lemma 10. Given an instance (C, X) of 5-Double-Sat, let G = (V, E) be the corresponding in-
stance of 3-Club Cover(2). Then, given two 3-clubs that cover G, we can compute in polynomial
time a truth assignment that double-satisfies C.
Proof. Consider two 3-clubs G[V1], G[V2], with V1, V2 ⊆ V , that cover G. First, notice that by
F ∈ V2 \ V1 and vC,j ∈ V2 \ V1, for each j
Lemma 7 we assume that r, r(cid:48) ∈ V1 \ V2, while y, r(cid:48)
with 1 (cid:54) j (cid:54) p. Moreover, by Lemma 8 it follows that for each i with 1 (cid:54) i (cid:54) q, xT
i do
not belong to the same 3-club, that is exactly one belongs to V1 and exactly one belongs to V2.
By construction, each path of length at most three from a vertex vC,j, with 1 (cid:54) j (cid:54) p, to r(cid:48)
h , with 1 (cid:54) h (cid:54) q. Similarly, each path of length at most three from a
must pass through some xF
vertex vC,j, with 1 (cid:54) j (cid:54) p, to r(cid:48)
i,1. Assume that vC,j, with 1 (cid:54) j (cid:54) p,
h ∈ V2, with 1 (cid:54) h (cid:54) p respectively). It
i,1 ∈ V2, with 1 (cid:54) i (cid:54) q (xF
is not adjacent to a vertex xT
u,1, with 1 (cid:54) u (cid:54) q,
w, with 1 (cid:54) w (cid:54) q (xT
follows that vC,j is only adjacent to y and to vertices xF
respectively) in G[V2]. In the first case, notice that y is adjacent only to vC,z, with 1 (cid:54) z (cid:54) p, and
F , respectively), thus implying that this path from vC,j to r(cid:48)
y2, none of which is adjacent to r(cid:48)
T
(to r(cid:48)
w (xT
F , respectively) has length at least 4. In the second case, xF
u,1, respectively) is adjacent
T (r(cid:48)
to r(cid:48)
T , rT , vC,j, xT
F , rF , vC,j and xT
F ,
respectively), implying that also in this case the path from vC,j to r(cid:48)
T (to r(cid:48)
F , respectively) has
F , vC,j ∈ V2, it follows that, for each vC,j, the set V2 contains a vertex
length at least 4. Since r(cid:48)
h , with 1 (cid:54) h (cid:54) q, connected to vC,j.
i,1, with 1 (cid:54) i (cid:54) q, and a vertex xF
xT
i,1, xF
By Lemma 8 exactly one of xT
i belongs to V2, thus we can construct a truth assignment f
i ∈ V2. The assignment f double-satisfies
i,1, for some i with 1 (cid:54) i (cid:54) q, and a
(cid:117)(cid:116)
as follows: f (xi) := true, if xT
each clause of C, since each vC,j is connected to a vertex xT
vertex xF
u,2, respectively), none of which is adjacent to r(cid:48)
h , for some h with 1 (cid:54) h (cid:54) q.
i,1 ∈ V2, f (xi) := false, if xF
T must pass through some xT
i,2 (r(cid:48)
T , r(cid:48)
i,1 and xF
T (r(cid:48)
F
Based on Lemma 9 and Lemma 10, and on the NP-completeness of 5-Double-Sat (see Theorem
6), we can conclude that 3-Club Cover(2) is NP-complete.
10
Theorem 11. 3-Club Cover(2) is NP-complete.
Proof. By Lemma 9 and Lemma 10, and from the NP-hardness of 5-Double-Sat (see Theorem 6),
it follows that 3-Club Cover(2) is NP-hard. The membership in NP follows easily from the fact
that, given two 3-clubs, it can be checked in polynomial time whether are 3-clubs and cover all
(cid:117)(cid:116)
vertices of G.
4 Hardness of Approximation
consider
section we
the approximation complexity of Min 2-Club Cover
In this
and
Min 3-Club Cover and we prove that Min 2-Club Cover
is not approximable within factor
O(V 1/2−ε), for each ε > 0, and that Min 3-Club Cover is not approximable within factor
O(V 1−ε), for each ε > 0. The proof for Min 2-Club Cover is obtained with a reduction very
similar to that of Section 3.1, except from the fact that we reduce Minimum Clique Partition to
Min 2-Club Cover.
Corollary 12. Unless P = N P , Min 2-Club Cover is not approximable within factor O(V 1/2−ε),
for each ε > 0.
present
a
Proof. We
to
Min 2-Club Cover. Let Gp = (V p, Ep) be a graph input of Minimum Clique Partition, we
compute in polynomial time a corresponding instance G = (V, E) of Min 2-Club Cover as in
Section 3.1. In what follows we prove the following results that are useful for the reduction.
from Minimum Clique Partition
preserving-factor
reduction
Lemma 13. Let Gp = (V p, Ep) be a graph input of Minimum Clique Partition and let
G = (V, E) be the corresponding instance of Min 2-Club Cover. Then, given a solution of
Minimum Clique Partition on Gp = (V p, Ep) consisting of k cliques, we can compute in polyno-
mial time a solution of Min 2-Club Cover on G = (V, E) consisting of k 2-clubs.
Proof. Consider a solution of Minimum Clique Partition on Gp = (V p, Ep) where {V p
k }
2 , . . . , V p
is the set of k cliques that partition V P . We define a solution of Min 2-Club Cover on G = (V, E)
consisting of k 2-clubs as follows. For each d, 1 (cid:54) d (cid:54) k, let
1 , V p
Vd = {wj ∈ V : vj ∈ V p
d } ∪ {wi,j : vi ∈ V p
d ∧ i < j}
2 . . . V p
k .
it follows that for each d, G[Vd]
As for the proof of Lemma 9,
is a 2-club. Furthermore,
G[V1], . . . , G[Vk] cover each vertex of V , as each vi ∈ V p is covered by one of the cliques
(cid:117)(cid:116)
V p
1 , V p
Lemma 14. Let Gp = (V p, Ep) be a graph input of Minimum Clique Partition and let G = (V, E)
be the corresponding instance of Min 2-Club Cover. Then, given a solution of Min 2-Club Cover
on G = (V, E) consisting of k 2-clubs, we can compute in polynomial time a solution of
Minimum Clique Partition on Gp = (V p, Ep) with k cliques.
Proof. Consider the 2-clubs G[V1], . . . , G[Vk] that cover G. As for the proof of Lemma 10, the
result follows from the fact that by Lemma 2, given wi, wj ∈ Vd, for each d with 1 (cid:54) d (cid:54) k, it
holds that {vi, vj} ∈ E. As a consequence, we can define a solution of Minimum Clique Partition
on Gp = (V p, Ep) consisting of k cliques as follows, for each d, 1 (cid:54) d (cid:54) k:
d = {vi : wi ∈ Vd}
V p
(cid:117)(cid:116)
The inapproximability of Min 2-Club Cover follows from Lemma 13 and Lemma 14, and from
the inapproximability of Minimum Clique Partition, which is known to be inapproximable within
factor O(V p1−ε(cid:48)
) [24] (where Gp = (V p, Ep) is an instance of Hence Min 2-Club Cover is not
approximable within factor O(V p1−ε(cid:48)
), for each ε(cid:48) > 0, unless P = N P , hence Min 2-Club Cover
is not approximable within factor O(V p(1−ε(cid:48))). By the definition of G = (V, E), it holds V =
V p + Ep (cid:54) V p2 hence, for each ε > 0, Min 2-Club Cover is not approximable within factor
(cid:117)(cid:116)
O(V 1/2−ε), unless P = N P .
11
Next, we show that Min 3-Club Cover is not approximable within factor O(V 1−ε), for each
ε > 0, unless P = N P , by giving a preserving-factor reduction from Minimum Clique Partition.
Consider an instance Gp = (V p, Ep) of Minimum Clique Partition, we construct an instance
G = (V, E) of Min 3-Club Cover by adding a pendant vertex connected to each vertex of V p.
Formally, V = {ui, wi : vi ∈ V p}, E = {{ui, wi} : 1 (cid:54) i (cid:54) V p} ∪ {{ui, uj} : {vi, vj} ∈ Ep}}.
We prove now the main properties of the reduction.
Lemma 15. Let Gp = (V p, Ep) be an instance of Minimum Clique Partition and let G = (V, E) be
the corresponding instance of Min 3-Club Cover. Then, given a solution of Minimum Clique Partition
on Gp = (V p, Ep) consisting of k cliques, we can compute in polynomial time a solution of
Min 3-Club Cover on G = (V, E) consisting of k 3-clubs.
Proof. Consider a solution of Minimum Clique Partition on Gp = (V p, Ep), consisting of the cliques
{Gp[Vc,1], Gp[Vc,2], . . . , Gp[Vc,k]}. Then, for each i, with 1 (cid:54) h (cid:54) k, define the following subset
Vh ⊆ V :
Vh = {uj, wj ∈ V : vj ∈ V p
h }
1 , V p
2 . . . V p
Since V p
k partition V p, it follows that V1, V2 . . . Vk partition (hence cover) G. Now, we
show that each G[Vh], with 1 (cid:54) h (cid:54) k, is a 3-club. First, notice that since G[V p
h ], is a clique,
then the set {uj : uj ∈ Vh} induces a clique in G. Then, it follows that, for each ui, wj, wz ∈ Vh,
(cid:117)(cid:116)
dG[Vh](ui, wj) (cid:54) 2 and dG[Vh](wj, wz) (cid:54) 3, thus concluding the proof.
Lemma 16. Let Gp = (V p, Ep) be a graph input of Minimum Clique Partition and let G = (V, E)
be the corresponding instance of Min 3-Club Cover. Then, given a solution of Min 3-Club Cover
on G = (V, E) consisting of k 3-clubs, we can compute in polynomial time a solution of
Minimum Clique Partition on Gp = (V p, Ep) consisting of k cliques.
Proof. Consider the k 3-clubs G[V1], . . . , G[Vk] that cover G. First, we show that for each Vh, 1 (cid:54)
h, (cid:54) k, ∀wi, wj ∈ Vh, with 1 (cid:54) i, j (cid:54) V p, it holds that ui, uj ∈ Vh. Indeed, notice that N (wi) =
{ui} and N (wj) = {uj}, and by the definition of a 3-club we must have dG[vh](wi, wj) (cid:54) 3, it
follows that ui, uj ∈ Vh. Hence, we can define a set of cliques of Gp. For each Vh, with 1 (cid:54) h (cid:54) k,
define a set V p
h :
h = {vi : wi ∈ Vh}
V p
Notice that each V p
wi, wj ∈ Vh, and this implies {vi, vj} ∈ Ep. Notice that the cliques V p
but starting from V p
consisting of at most k cliques.
h , 1 (cid:54) h (cid:54) k, induces a clique in Gp, as by construction if vi, vj ∈ V p
1 , . . . , V p
h , then
k may overlap,
k , we can easily compute in polynomial time a clique partition of Gp
(cid:117)(cid:116)
1 , . . . , V p
Lemma 15 and Lemma 16 imply the following result.
Theorem 17. Min 3-Club Cover is not approximable within factor O(V 1−ε), for each ε > 0,
unless P = N P .
Proof. The result follows from Lemma 15 and Lemma 16, as these results imply that we have
defined a factor-preserving reduction, and from the inapproximability of Minimum Clique Partition,
which is known to be inapproximable within factor O(V p1−ε), for each ε > 0, unless P = N P [24]
(where Gp = (V p, Ep) is an instance of Minimum Clique Partition). Thus, Min 3-Club Cover is not
approximable within factor O(V p1−ε), for each ε > 0, unless P = N P , and since it holds
V = 2V p, Min 3-Club Cover is not approximable within factor O(V 1−ε), unless P = N P .
(cid:117)(cid:116)
5 An Approximation Algorithm for Min 2-Club Cover
In this section, we present an approximation algorithm for Min 2-Club Cover that achieves an
approximation factor of 2V 1/2 log3/2 V . Notice that, due to the result in Section 4, the approxi-
12
mation factor is almost tight. We start by describing the approximation algorithm, then we present
the analysis of the approximation factor.
Algorithm 1: Club-Cover-Approx
Data: a graph G
Result: a cover S of G
1 V (cid:48) := V ; /* V (cid:48) is the set of uncovered vertices of G, initialized to V */
2 S := ∅;
3 while V (cid:48) (cid:54)= ∅ do
Let v be a vertex of V such that N [v] ∩ V (cid:48) is maximum;
Add N [v] to S;
V (cid:48) := V (cid:48) \ N [v];
Club-Cover-Approx
similar
the
is
6
4
5
to
algorithm for
Minimum Dominating Set and Minimum Set Cover. While there exists an uncovered vertex
of G, the Club-Cover-Approx algorithm greedily defines a 2-club induced by the set N [v] of
vertices, with v ∈ V , such that N [v] covers the maximum number of uncovered vertices (notice
that some of the vertices of N [v] may already be covered). While for Minimum Dominating Set the
choice of each iteration is optimal, here the choice is suboptimal. Notice that indeed computing a
maximum 2-club is NP-hard.
approximation
greedy
Clearly the algorithm returns a feasible solution for Min 2-Club Cover, as each set N [v] picked
by the algorithm is a 2-club and, by construction, each vertex of V is covered. Next, we show the
approximation factor yielded by the Club-Cover-Approx algorithm for Min 2-Club Cover.
First, consider the set VD of vertices v ∈ V picked by the Club-Cover-Approx algorithm, so
that N [v] is added to S. Notice that VD = S and that VD is a dominating set of G, since, at
each step, the vertex v picked by the algorithm dominates each vertex in N [v], and each vertex in
V is covered by the algorithm, so it belongs to some N [v], with v ∈ VD.
Let D be a minimum dominating set of the input graph G. By the property of the greedy
approximation algorithm for Minimum Dominating Set, the set VD has the following property [14]:
VD (cid:54) D log V
(1)
The size of a minimum dominating set in graphs of diameter bounded by 2 (hence 2-clubs) has
been considered in [8], where the following result is proven.
Lemma 18 ([8]). Let H = (VH , EH ) be a 2-club, then H has a dominating set of size at most
1 +(cid:112)VH + ln(VH).
The approximation factor 2V 1/2 log3/2 V for Club-Cover-Approx is obtained by combining
Lemma 18 and Equation 1.
Theorem 19. Let OP T be an optimal solution of Min 2-Club Cover, then Club-Cover-Approx
returns a solution having at most 2V 1/2 log3/2 V OP T 2-clubs.
Proof. Let D be a minimum dominating set of G and let OP T be an optimal solution of
Min 2-Club Cover. We start by proving that D (cid:54) 2OP TV 1/2 log1/2 V . For each 2-club G[C],
with C ⊆ V , that belongs to OP T , by Lemma 18 there exists a dominating set DC of size
at most 1 +(cid:112)C + ln(C) (cid:54) 2(cid:112)C + ln(C). Since C (cid:54) V , it follows that each 2-club
G[C] that belongs to OP T has a dominating set of size at most 2(cid:112)V + ln(V ). Consider
D(cid:48) = (cid:83)
ers G. Since D(cid:48) contains OP T sets DC and DC (cid:54) 2(cid:112)V + ln(V ), for each G[C] ∈ OP T ,
it follows that D(cid:48) (cid:54) 2OP T(cid:112)V + ln(V ). Since D is a minimum dominating set, it follows
that D (cid:54) D(cid:48) (cid:54) 2OP T((cid:112)V + ln(V )). By Equation 1, it holds VD (cid:54) 2D log V thus
C∈OP T DC. It follows that D(cid:48) is a dominating set of G, since the 2-clubs in OP T cov-
VD (cid:54) 2V 1/2 ln1/2 V log V OP T (cid:54) 2V 1/2 log3/2 V OP T.
(cid:117)(cid:116)
6 Conclusion
There are some interesting direction for the problem of covering a graph with s-clubs. From the
computational complexity point of view, the main open problem is whether 2-Club Cover(2) is NP-
complete or is in P. Moreover, it would be interesting to study the computational/parameterized
complexity of the problem in specific graph classes, as done for Minimum Clique Partition [5,6,21,9].
13
References
1. Alba, R.D.: A graph-theoretic definition of a sociometric clique. The Journal of Mathematical Sociol-
ogy 3, 113–126 (1973)
2. Asahiro, Y., Doi, Y., Miyano, E., Samizo, K., Shimizu, H.: Optimal Approximation Algorithms for
Maximum Distance-Bounded Subgraph Problems. Algorithmica (Jul 2017)
3. Balasundaram, B., Butenko, S., Trukhanov, S.: Novel Approaches for Analyzing Biological Networks.
J. Comb. Optim. 10(1), 23–39 (2005)
4. Bourjolly, J., Laporte, G., Pesant, G.: An exact algorithm for the maximum k-club problem in an
undirected graph. European Journal of Operational Research 138(1), 21–28 (2002)
5. Cerioli, M.R., Faria, L., Ferreira, T.O., Martinhon, C.A.J., Protti, F., Reed, B.A.: Partition into
cliques for cubic graphs: Planar case, complexity and approximation. Discrete Applied Mathematics
156(12), 2270–2278 (2008)
6. Cerioli, M.R., Faria, L., Ferreira, T.O., Protti, F.: A note on maximum independent sets and minimum
clique partitions in unit disk graphs and penny graphs: complexity and approximation. RAIRO -
Theor. Inf. and Applic. 45(3), 331–346 (2011)
7. Chang, M., Hung, L., Lin, C., Su, P.: Finding large k-clubs in undirected graphs. Computing 95(9),
739–758 (2013)
8. Desormeaux, W.J., Haynes, T.W., Henning, M.A., Yeo, A.: Total domination in graphs with diameter
2. Journal of Graph Theory 75(1), 91–103 (2014)
9. Dumitrescu, A., Pach, J.: Minimum Clique Partition in Unit Disk Graphs. Graphs and Combinatorics
27(3), 399–411 (2011)
10. Garey, M.R., Johnson, D.S., Stockmeyer, L.J.: Some Simplified NP-Complete Graph Problems. Theor.
Comput. Sci. 1(3), 237–267 (1976)
11. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory of NP-Completeness
(1979)
12. Golovach, P.A., Heggernes, P., Kratsch, D., Rafiey, A.: Finding clubs in graph classes. Discrete Applied
Mathematics 174, 57–65 (2014)
13. Hartung, S., Komusiewicz, C., Nichterlein, A.: Parameterized Algorithmics and Computational Ex-
periments for Finding 2-Clubs. J. Graph Algorithms Appl. 19(1), 155–190 (2015)
14. Johnson, D.S.: Approximation Algorithms for Combinatorial Problems. J. Comput. Syst. Sci. 9(3),
256–278 (1974)
15. Karp, R.M.: Reducibility among combinatorial problems. In: Miller, R.E., Thatcher, J.W. (eds.) Pro-
ceedings of a symposium on the Complexity of Computer Computations, held March 20-22, 1972, at
the IBM Thomas J. Watson Research Center, Yorktown Heights, New York. pp. 85–103. The IBM
Research Symposia Series, Plenum Press, New York (1972)
16. Komusiewicz, C.: Multivariate Algorithmics for Finding Cohesive Subnetworks. Algorithms 9(1), 21
(2016)
17. Komusiewicz, C., Sorge, M.: An algorithmic framework for fixed-cardinality optimization in sparse
graphs applied to dense subgraph problems. Discrete Applied Mathematics 193, 145–161 (2015)
18. Laan, S., Marx, M., Mokken, R.J.: Close communities in social networks: boroughs and 2-clubs. Social
Netw. Analys. Mining 6(1), 20:1–20:16 (2016)
19. Mokken, R.: Cliques, clubs and clans. Quality & Quantity: International Journal of Methodology
13(2), 161–173 (1979)
20. Mokken, R.J., Heemskerk, E.M., Laan, S.: Close communication and 2-clubs in corporate networks:
Europe 2010. Social Netw. Analys. Mining 6(1), 40:1–40:19 (2016)
21. Pirwani, I.A., Salavatipour, M.R.: A weakly robust PTAS for minimum clique partition in unit disk
graphs. Algorithmica 62(3-4), 1050–1072 (2012)
22. Schafer, A., Komusiewicz, C., Moser, H., Niedermeier, R.: Parameterized computational complexity
of finding small-diameter subgraphs. Optimization Letters 6(5), 883–891 (2012)
14
23. Zoppis, I., Dondi, R., Santoro, E., Castelnuovo, G., Sicurello, F., Mauri, G.: Optimizing social inter-
action - A computational approach to support patient engagement. In: Zwiggelaar, R., Gamboa, H.,
Fred, A.L.N., i Badia, S.B. (eds.) Proceedings of the 11th International Joint Conference on Biomed-
ical Engineering Systems and Technologies (BIOSTEC 2018) - Volume 5: HEALTHINF, Funchal,
Madeira, Portugal, January 19-21, 2018. pp. 651–657. SciTePress (2018)
24. Zuckerman, D.: Linear Degree Extractors and the Inapproximability of Max Clique and Chromatic
Number. Theory of Computing 3(1), 103–128 (2007)
|
1506.07568 | 1 | 1506 | 2015-06-24T21:25:49 | Towards Resistance Sparsifiers | [
"cs.DS"
] | We study resistance sparsification of graphs, in which the goal is to find a sparse subgraph (with reweighted edges) that approximately preserves the effective resistances between every pair of nodes. We show that every dense regular expander admits a $(1+\epsilon)$-resistance sparsifier of size $\tilde O(n/\epsilon)$, and conjecture this bound holds for all graphs on $n$ nodes. In comparison, spectral sparsification is a strictly stronger notion and requires $\Omega(n/\epsilon^2)$ edges even on the complete graph.
Our approach leads to the following structural question on graphs: Does every dense regular expander contain a sparse regular expander as a subgraph? Our main technical contribution, which may of independent interest, is a positive answer to this question in a certain setting of parameters. Combining this with a recent result of von Luxburg, Radl, and Hein~(JMLR, 2014) leads to the aforementioned resistance sparsifiers. | cs.DS | cs |
Towards Resistance Sparsifiers
Michael Dinitz∗
Robert Krauthgamer†
Tal Wagner‡
May 29, 2021
Abstract
We study resistance sparsification of graphs, in which the goal is to find a sparse subgraph
(with reweighted edges) that approximately preserves the effective resistances between every pair
of nodes. We show that every dense regular expander admits a (1 + ε)-resistance sparsifier of
size O(n/ε), and conjecture this bound holds for all graphs on n nodes. In comparison, spectral
sparsification is a strictly stronger notion and requires Ω(n/ε2) edges even on the complete
graph.
Our approach leads to the following structural question on graphs: Does every dense regular
expander contain a sparse regular expander as a subgraph? Our main technical contribution,
which may of independent interest, is a positive answer to this question in a certain setting of
parameters. Combining this with a recent result of von Luxburg, Radl, and Hein (JMLR, 2014)
leads to the aforementioned resistance sparsifiers.
1
Introduction
Compact representations of discrete structures are of fundamental importance, both from an appli-
cations point of view and from a purely mathematical perspective. Graph sparsification is perhaps
one of the simplest examples: given a graph G(V, E), is there a subgraph that represents G truth-
fully, say up to a small approximation? This notion has had different names in different contexts,
depending on the property that is being preserved: preserving distances is known as a graph span-
ner [PS89], preserving the size of cuts is known as a cut sparsifier [BK96], while preserving spectral
properties is known as a spectral sparsifier [ST04]. These concepts are known to be related, for
example, every spectral sparsifier is clearly also a cut sparsifier, and spectral sparsifiers can be
constructed by an appropriate sample of spanners [KP12].
Our work is concerned with sparsification that preserves effective resistances. We define this
in Section 1.1, but informally the effective resistance between two nodes u and v is the voltage
differential between them when we regard the graph as an electrical network of resistors with one
unit of current injected at u and extracted at v. Effective resistances are very useful in many
applications that seek to cluster nodes in a network (see [vLRH14] and references therein for
a comprehensive list), and are also of fundamental mathematical interest. For example, they
have deep connections to random walks on graphs (see [Lov96] for an excellent overview of this
connection). Most famously, the commute time between two nodes u and v (the expected time for
∗Johns Hopkins University, Baltimore, MD. Work supported in part by NSF award #1464239.
Email:
[email protected]
†Weizmann Institute of Science, Rehovot, Israel. Work supported in part by a US-Israel BSF grant #2010418 and
‡Massachussetts Institute of Technology, Cambridge, MA. Email: [email protected]
an Israel Science Foundation grant #897/13. Email: [email protected]
1
a random walk starting at u to hit v plus the expected time for a random walk starting at v to hit
u) is exactly 2m times the effective resistance between u and v, where throughout n := V and
m := E. Hence, we are concerned with sparsification which preserves commute times.
We ask whether graphs admit a good resistance sparsifier : a reweighted subgraph G(cid:48)(V, E(cid:48), w(cid:48))
in which the effective resistances are equal, up to a (1 + ε)-factor, to those in the original graph.
The short answer is yes, because every (1+ε)-spectral sparsifier is also a (1+ε)-resistance sparsifier.
Using the spectral-sparsifiers of [BSS12], we immediately conclude that every graph admits a (1+ε)-
resistance sparsifier with O(n/ε2) edges.
Interestingly, the same 1/ε2 factor loss appears even when we interpret "sparsification" far
more broadly. For example, a natural approach to compressing the effective resistances is to use
a metric embedding (instead of looking for a subgraph): map the nodes into some metric, and
use the metric's distances as our resistance estimates. This approach is particularly attractive
since it is well-known that effective resistances form a metric space which embeds isometrically
into (cid:96)2-squared (i.e., the metric is of negative type, see e.g. [DL97]). Hence, using the Johnson-
Lindenstrauss dimension reduction lemma, we can represent effective resistances up to a distortion
of (1 + ε) using vectors of dimension O(ε−2 log n), i.e., using total space O(n/ε2). In fact, this very
approach was used by [SS11] to quickly compute effective resistance estimates, which were then
used to construct a spectral sparsifier.
Since a 1/ε2 term appears in both of these natural ways to compactly represent effective resis-
tances, an obvious question is whether this is necessary. For the stronger requirement of spectral
sparsification, we know the answer is yes -- every spectral sparsifier of the complete graph requires
Ω(n/ε2) edges [BSS12, Section 4] (see also [AKW14]). However, it is currently unknown whether
such a bound holds also for resistance sparsifiers, and the starting point of our work is the ob-
servation (based on [vLRH14]) that for the complete graph, every O(1/ε)-regular expander is a
(1 + ε)-resistance sparsifier, despite not being a (1 + ε)-spectral sparsifier! We thus put forward the
following conjecture.
Conjecture 1.1. Every graph admits a (1 + ε)-resistance sparsifier with O(n/ε) edges.
We make the first step in this direction by proving the special case of dense regular expanders
(which directly generalize the complete graph). Even this very special case turns out to be non-
trivial, and in fact leads us to another beautiful problem which is interesting in its own right.
Question 1.2. Does every dense regular expander contain a sparse regular expander as a subgraph?
Our positive answer to this question (for a certain definition of expanders) forms the bulk of
our technical work (Sections 2 and 3), and is then used to find good resistance sparsifiers for dense
regular expanders (Section 4).
1.1 Results and Techniques
Throughout, we consider undirected graphs, and they are unweighted unless stated otherwise. In
a weighted graph, i.e., when edges have nonnegative weights, the weighted degree of a vertex is
the sum of weights on incident edges, and the graph is considered regular if all of its weighted
degrees are equal. Typically, a sparsifying subgraph must be weighted even when the host graph
is unweighted, in order to exhibit comparable parameters with far fewer edges.
Before we can state our results we first need to recall some basic definitions from spectral graph
theory. Given a weighted graph G, let D be the diagonal n × n matrix of weighted degrees, and
let A be the weighted adjacency matrix. The Laplacian of G is defined as L := D − A, and the
normalized Laplacian is the matrix L := D−1/2LD−1/2.
2
Definition 1.3 (Effective Resistance). Let G(V, E, w) be a weighted graph, and let P the Moore-
Penrose pseudo-inverse of its Laplacian matrix. The effective resistance (also called resistance
distance) between two nodes u, v ∈ V is
RG(u, v) := (eu − ev)T P (eu − ev),
where eu and ev denote the standard basis vectors in RV that correspond to u and v respectively.
When the graph G is clear from context we will omit it and write R(u, v). We can now define
the main objects that we study.
Definition 1.4 (Resistance Sparsifier). Let G(V, E, w) be a weighted graph, and let ε ∈ (0, 1).
A (1 + ε)-resistance sparsifier for G is a subgraph H(V, E(cid:48), w(cid:48)) with reweighted edges such that
(1 − ε)RH (u, v) ≤ RG(u, v) ≤ (1 + ε)RH (u, v), for all u, v ∈ V .
It will turn out that in order to understand resistance sparsifiers, we need to use expansion
properties.
Definition 1.5 (Graph Expansion). The edge-expansion (also known as the Cheeger constant) of
a weighted graph G(V, E, w) is
(cid:27)
(cid:26) w(S, ¯S)
S
φ(G) := min
: S ⊂ V, 0 < S ≤ V /2
,
where w(S, ¯S) denotes the total weight of edges with exactly one endpoint in S ⊂ V . The spec-
tral expansion of G, denoted λ2(G), is the second-smallest eigenvalue of the graph's normalized
Laplacian.
Our main result is the following. Throughout this paper, "efficiently" means in randomized
polynomial time.
Theorem 1.6. Fix β, γ > 0, let n be sufficiently large, and 1/n0.99 < ε < 1. Every D-regular graph
G on n nodes with D ≥ βn and φ(G) ≥ γD contains (as a subgraph) a (1 + ε)-resistance sparsifier
with at most ε−1n(log n)O(1/βγ2) edges, and it can be found efficiently.
While dense regular expanders may seem like a simple case, even this special case requires
significant technical work. The most obvious idea, of sparsifying through random sampling, does
not work -- selecting each edge of G uniformly at random with probability O(1/(Dε)) (the right
probability for achieving a subgraph with O(n/ε) edges) need not yield a (1 + ε)-resistance spar-
sifier. Intuitively, this is because the variance of independent random sampling is too large (see
Theorem 4.1 for the precise effect), and the easiest setting to see this is the case of sparsifying
the complete graph. If we sparsify through independent random sampling, then to get a (1 + ε)-
resistance sparsifier requires picking each edge independently with probability at least 1/(ε2n), and
we end up with n/ε2 edges. To beat this, we need to use correlated sampling. More specifically,
it turns out that a random O(1/ε)-regular graph is a (1 + ε)-resistance sparsifier of the complete
graph, despite not being a (1 + ε)-spectral sparsifier. So instead of sampling edges independently
(the natural approach, and in fact the approach used to construct spectral sparsifiers by Spielman
and Srivastava [SS11]), we need to sample a random regular graph.
In order to prove Theorem 1.6, we actually need to generalize this approach beyond the complete
graph. But what is the natural generalization of a random regular graph when the graph we start
with is not the complete graph? It turns out that what we need is an expander, which is sparse
but maintains regularity of its degrees. This motivates our main structural result, that every dense
regular expander contains a sparse regular expander (as a subgraph). This can be seen as a type
of sparsification result that retains regularity.
3
Theorem 1.7. Fix β, γ > 0 and let n be sufficiently large. Every D-regular graph G on n nodes
with D ≥ βn and φ(G) ≥ γD contains a weighted d-regular subgraph H with d = (log n)O(1/βγ2)
and φ(H) ≥ 1
3 . All edge weights in H are in {1, 2}, and H can be found efficiently.
To prove this theorem, we analyze a modified version of the cut-matching game of Khandekar,
Rao, and Vazirani [KRV09]. This game has been used in the past to construct expander graphs,
but in order to use it for Theorem 1.7 we need to generalize beyond matchings, and also show how
to turn the graphs it creates (which are not necessarily subgraphs of G) into subgraphs of G.
The expansion requirement for G in Theorem 1.7 is equivalent to λ2(G) = Ω(1), when β and
γ are viewed as absolute constants. We note that H is a much weaker expander, satisfying only
λ2(H) = Ω(1/polylog(n)), but this is nonetheless sufficient for Theorem 1.6. Also, H is regular in
weighted degrees. For completeness we give a variant of Theorem 1.7 that achieves an unweighted
H by requiring stronger expansion from G, but this is not necessary for our application to resistance
sparsifiers, which anyway involves reweighting the edges.
Theorem 1.8. For every β > 0 there is 0 < γ < 1 such the following holds for sufficiently large
n. Every D-regular graph G on n nodes with D ≥ βn and φ(G) ≥ γD contains an (unweighted)
d-regular subgraph H with d = (log n)O(1/βγ) and φ(H) ≥ 1
3 , and it can be found efficiently.
The algorithm underlying Theorems 1.6, 1.7 and 1.8 turns out to be quite straightforward:
decompose the host graph into disjoint perfect matchings or Hamiltonian cycles (which are "atomic"
regular components), and subsample a random subset of them of size d to form the target subgraph.
However, since the decomposition leads to large dependencies between inclusion of different edges
in the subgraph, it is unclear how to approach this algorithm with direct probabilistic analysis.
Instead, our analysis uses the adaptive framework of [KRV09] to quantify the effect of gradually
adding random matching/cycles from the decomposition to the subgraph.
1.2 Related Work
The line of work most directly related to resistance sparsifiers is the construction of spectral spar-
sifiers. This was initiated by Spielman and Teng [ST04], and was later pushed to its limits by
Spielman and Teng [ST11], Spielman and Srivastava [SS11], and Batson, Spielman, and Srivas-
tava [BSS12], who finally proved that every graph has a (1 + ε)-spectral sparsifier with O(n/ε2)
edges and that this bound is tight (see also [AKW14]).
The approach by Spielman and Srivastava [SS11] is particularly closely related to our work.
They construct almost-optimal spectral sparsifiers (a logarithmic factor worse than [BSS12]) by
sampling each edge independently with probability proportional to the effective resistance between
the endpoints. This method naturally leads us to try the same thing for resistance sparsification,
but as discussed, independent random sampling (even based on the effective resistances) cannot give
improved resistance sparsifiers. Interestingly, in order to make their algorithm extremely efficient
they needed a way to estimate effective resistances very quickly, so along the way they showed
how to create a sketch of size O(n log n/ε2) from which every resistance distance can be read
off in O(log n) time (essentially through an (cid:96)2-squared embedding and a Johnson-Lindenstrauss
dimension reduction).
2 Sparse Regular Expanding Subgraphs
In this section we prove Theorem 1.7, building towards it in stages. Our starting point is the
Cut-Matching game of Khandekar, Rao and Vazirani (KRV) [KRV09], which is a framework to
4
constructing sparse expanders by iteratively adding perfect matchings across adaptively chosen
bisections of the vertex set. The resulting graph H is regular, as it is the union of perfect matchings,
and if the matchings are contained in the input graph G then H is furthermore a subgraph of G, as
desired. In Section 2.1, we employ this approach to prove Theorem 1.7 in the case D/n = 3
4 + Ω(1).
To handle smaller D, we observe that the perfect matchings in the KRV game can be replaced
with a more general structure that we call a weave, defined as a set of edges where for every vertex
at least one incident edge crosses the given bisection. To ensure that H is regular (all vertices
have the same degree), we would like the weaves to be regular. We thus decompose the input
graph to disjoint regular elements -- either perfect matchings or Hamiltonian cycles -- and use them
as building blocks to construct regular weaves. Leveraging the fact that for some bisections, G
contains no perfect matching but does contain a weave, we use this extension in Section 2.2 to
handle the case D/n = 1
2 + Ω(1).
Finally, for the general case D/n = Ω(1), we need to handle a graph G that contains no weave
on some bisections. The main portion of our proof constructs a weave that is not contained in G,
but rather embeds in G with small (polylogarithmic) congestion. Repeating this step sufficiently
many times as required by the KRV game, yields a subgraph H as desired.
Notation and terminology. For a regular graph G, we denote deg(G) the degree of each vertex.
We say that a graph H is an edge-expander if φ(H) > 1
3 . A bisection of a vertex set of size n is a
partition (S, ¯S) with equal sizes 1
2 n if n is even, or with sizes (cid:98) 1
2 n(cid:99) and (cid:100) 1
2 n(cid:101) if n is odd.
2.1 The Cut-Matching Game
Khandekar, Rao and Vazirani [KRV09] described the following game between two players. Start
with an empty graph (no edges) H on a vertex set of even size n. In each round, the cut player
chooses a bisection, and the matching player answers with a perfect matching across the bisection.
The game ends when H is an edge-expander. Informally, the goal of the cut player is to reach this
as soon as possible, and that of the matching player is to delay the game's ending.
Theorem 2.1 ([KRV09, KKOV07]). The cut player has an efficiently computable strategy that
wins (i.e., is guaranteed to end the game) within O(log2 n) rounds, and a non-efficient strategy that
wins within O(log n) rounds.
The following result illustrates the use of the KRV framework in our setting.
Theorem 2.2. Let δ > 0 and let n be even and sufficiently large (n ≥ n0(δ)). Then every n-
vertex graph G(V, E) with minimum degree D ≥ ( 3
4 + δ)n contains an edge-expander H that is
d-regular for d = O(log n), and also an efficiently computable edge-expander H(cid:48) that is a d(cid:48)-regular
for d(cid:48) = O(log2 n).
Proof. Apply the Cut-Matching game on V with the following player strategies. For the cut player,
execute the efficient strategy from Theorem 2.1 that wins within O(log2 n) rounds. For the matching
player, given a bisection (S, ¯S), consider the bipartite subgraph G[S, ¯S] of G induced by (S, ¯S). Each
2 n − 1 of them are in S, and the rest
vertex in S has in G at least D ≥ 3
must be in ¯S, which implies that G[S, ¯S] has minimum degree ≥ 1
4 n. Hence, as a simple consequence
of Hall's theorem (see Proposition A.2), it contains a perfect matching that can be efficiently found.
The matching player returns this matching as his answer. We then remove this matching from G
before proceeding to the next round, to ensure that different iterations find disjoint matchings. The
slackness parameter δ (and n being sufficiently large) ensure that the minimum degree of G does
not fall below 3
4 n during the O(log2 n) iterations, so the above argument holds in all rounds.
4 n neighbors, but at most 1
5
The game ends with an edge-expander H(cid:48) which is a disjoint union of d(cid:48) = O(log2 n) perfect
matchings contained in G, and hence is a d(cid:48)-regular subgraph of G, as required. To obtain the
graph H, apply the same reasoning but using the non-efficient strategy from Theorem 2.1 that
wins within O(log n) rounds.
2.2 The Cut-Weave Game
For values of D below 3
4 n, we can no longer guarantee that every bisection in G admits a perfect
matching. However, we observe that one can allow the matching player a wider range of strategies
while retaining the ability of the cut player to win within a small number of rounds.
Definition 2.3 (weave). Given a bisection (S, ¯S) of a vertex set V , a weave on (S, ¯S) is a subgraph
in which every node has an incident edge crossing (S, ¯S).
Definition 2.4 (Cut-Weave Game). The Cut-Weave game with parameter r is the following game
of two players. Start with a graph H on a vertex set of size n and no edges. In each round, the cut
player chooses a bisection of the vertex set, and the weave player answers with an r-regular weave
on the bisection. The edges of the weave are added to H.
Note that the r = 1 case is the original Cut-Matching game (when n is even). The following
theorem is an extension of Theorem 2.1. For clarity of presentation, its proof is deferred to Section 3.
Theorem 2.5. In the Cut-Weave game with parameter r, the cut player has an efficient strategy
that wins within O(r log2 n) rounds, and furthermore ensures φ(H) ≥ 1
2 r.
In order to construct regular weaves, we employ a decomposition of G into disjoint Hamiltonian
cycles. The following theorem was proven by Perkovic and Reed [PR97], and recently extended by
Csaba, Kuhn, Lo, Osthus and Treglown [CKL+14].
Theorem 2.6. Let δ > 0. Every D-regular graph G on n nodes with D ≥ ( 1
decomposition of its edges into (cid:98) 1
is odd). Furthermore, the decomposition can be found efficiently.
2 + δ)n, admits a
2 D(cid:99) Hamiltonian cycles and possibly one perfect matching (if D
Now we can use the Cut-Weave framework to make another step towards Theorem 1.7.
Theorem 2.7. Let δ > 0 and let n be sufficiently large. Then every n-vertex graph G(V, E) with
minimum degree D ≥ ( 1
2 + δ)n contains a d-regular edge-expander H with d = O(log3 n), which
furthermore can be efficiently found.
Proof. We simulate the Cut-Weave game with r = 16δ−1 log n. The proof is the same as Theo-
rem 2.2, only instead of a perfect matching we need to construct an r-regular weave across a given
bisection (S, ¯S). We apply Theorem 2.6 to obtain a Hamiltonian decomposition of G. For simplic-
ity, if D is odd we discard the one perfect matching from Theorem 2.6. Let C be the collection of
Hamiltonian cycles in the decomposition.
2 n neighbors in S, and hence
at least δn incident edges crossing to ¯S. We set up a Set-Cover instance of the cycles C against the
nodes in S, where a node v is considered covered by a cycle C is v has an incident edge crossing to
¯S, that belongs to C. This is a dense instance: since each cycle visits v only twice, v can be covered
2 δn cycles. Therefore, 4δ−1 log n randomly chosen cycles form a cover with high probability (see
by 1
Proposition A.3 for details). We then repeat the same procedure to cover the nodes on side ¯S. The
result is a collection of 8δ−1 log n = 1
2 r disjoint Hamiltonian cycles, whose union forms an r-regular
weave on (S, ¯S), which we return as the answer of the weave player. Applying Theorem 2.5 with
r = O(log n) concludes the proof of Theorem 2.7.
2 n(cid:101). Every v ∈ S has at most S − 1 ≤ 1
Suppose w.l.o.g. S = (cid:100) 1
6
Observe that in the proof of Theorem 2.7, the weave player is in fact oblivious to the queries
of the cut player: all she does is sample random cycles from C, and the output subgraph H is
the union of those cycles. Therefore, in order to construct H, it is sufficient to decompose G into
disjoint Hamiltonian cycles, and choose a random subset of size O(log3 n) of them. There is no
need to actually simulate the cut player, and in particular, the proof does not require her strategy
(from Theorem 2.5) to be efficient.
2.3 Reduction to Double Cover
We now begin to address the full range of parameters stated in Theorem 1.7. In this range there
is no Hamiltonian decomposition theorem (or a result of similar flavor) that we are aware of, so
we replace it with a basic argument which incurs edge weights w : V × V → {0, 1, 2} in the target
subgraph H, as well as a loss in its degree.
Given the input graph G(V, E), we construct its double cover, which is the bipartite graph
G(cid:48)(cid:48)(V (cid:48)(cid:48), E(cid:48)(cid:48)) defined by V (cid:48)(cid:48) = V × {0, 1} and E(cid:48)(cid:48) = {((v, 0)(u, 1)) : vu ∈ E}. It is easily seen that if
G is D-regular then so is G(cid:48)(cid:48), and since V (cid:48)(cid:48) = 2V we have D ≥ 1
2 βV (cid:48)(cid:48). It also well known that
λ2(G) = λ2(G(cid:48)(cid:48)), and therefore by the discrete Cheeger inequalities,
φ(G(cid:48)(cid:48)) ≥ 1
2 λ2(G(cid:48)(cid:48))D = 1
2 λ2(G)D ≥ 1
2 γ2(G)D.
2 β and γ(cid:48)(cid:48) = 1
G(cid:48)(cid:48) satisfies the requirements of Theorem 1.7 with β(cid:48)(cid:48) = 1
2 γ2. Suppose we find in G(cid:48)(cid:48)
a d-regular edge-expander H(cid:48)(cid:48) with d = (log n)O(1/β(cid:48)(cid:48)γ(cid:48)(cid:48)) = (log n)O(1/βγ2). We carry it over to a sub-
graph H of G, by including each edge uv ∈ E in H with weight {(v, 0)(u, 1), (u, 0)(v, 1)} ∩ E(H(cid:48)(cid:48)),
where E(H(cid:48)(cid:48)) denotes the edge set of H(cid:48)(cid:48). Each edge then appears in H with weight either 1 or 2
(or 0, which means it is not present in H). It can be easily checked that H is d-regular in weighted
degrees, and φ(H) ≥ 1
2 φ(H(cid:48)(cid:48)). Therefore H is a suitable target subgraph for Theorem 1.7.
The above reduction allows us to restrict our attention to regular bipartite graphs G, but on
the other hand we are forced to look for a subgraph H which is unweighted and d-regular with
d = (log n)O(1/βγ) (which is tighter than stated in Theorem 1.7). We take this approach in the
remainder of the proof. The gain is that such G admits a decomposition into disjoint perfect
matchings, which can be efficiently found, as a direct consequence of Hall's theorem. We will use
this fact where we have previously used Theorem 2.6.
2.4 Constructing an Embedded Weave
We now get to the main technical part of the proof. Given a bisection (S, ¯S) queried by the
cut player, we need to construct an r-regular weave on the bisection, where this time we choose
r = (log n)O(1/βγ). Unlike the proof of Theorem 2.7, we cannot hope to find a weave which is a
subgraph of G, since if D < 1
2 n, any bisection in which one side contains some vertex and all its
neighbors would not admit a weave in G. Instead, we aim for a weave which embeds into G with
polylogarithmic congestion.
We will use two types of graph operations: The union of two graphs on the same vertex set V
is obtained by simply taking the set union of their edge sets, whereas the sum of the two graphs
is given by keeping parallel edges if they appear in both graphs. We now construct the weave in 4
steps.
Step 1. Fix µ = βγ2
following process:
4 . We partition the entire vertex set V into subsets S0, S1, . . . , St by the
7
1. Set S0 ← ¯S and T ← S.
2. While T (cid:54)= ∅, take Si ⊆ T to be the subset of nodes with at least µD neighbors in Si−1, and
set T → T \ Si.
βγ iterations.
βγ that ends with T (cid:54)= ∅. Denote ¯T = V \ T = ∪i
Lemma 2.8. The process terminates after t ≤ 2
Proof. Consider an iteration i ≤ 2
j=0Sj. By the
hypothesis φ(G) ≥ γD we have at least γDT edges crossing from T to ¯T , so by averaging over
the nodes in T , there is v ∈ T with γD neighbors in ¯T . For every j < i, v must have less than µD
neighbors in Sj, or it would already belong to Sj+1 ⊆ ¯T . Summing over j = 0, . . . , i − 1, we see
that v has less than iµD ≤ 1
2 γD neighbors in Si. This implies
Si ≥ 1
βγ iterations either terminates the process or
removes 1
2 γD. We have shown that each of the first 2
2 γD ≥ 1
2 γD neighbors in ¯T \ Si, so at least 1
2 γβn nodes from T , so after 2
βγ iterations we must have T = ∅.
Step 2. By Section 2.3 we have a decomposition of all the edges in G into a collection M of
D disjoint perfect matchings. For every i = 1, . . . , t, we now cover the nodes in Si with perfect
matchings, similar to the proof of Theorem 2.7. A node v ∈ Si is considered covered by a matching
if v has an incident edge with the other endpoint in Si−1, and that edge lies on the matching. Since
v has µD incident edges crossing to Si−1, and each matching touches v with at most one edge,
we have µD matchings that can cover v. Therefore k = 1
µ log n randomly chosen matchings from
M form a cover of Si (see Proposition A.3), which we denote as Ki. Thus, for each i we have a
subgraph Ki which is k-regular, such that each node in Si has an incident edge in Ki with the other
endpoint in Si−1. Denote henceforth
K = ∪t
i=1Ki.
Note that K is a regular subgraph of G, since it is a union of disjoint perfect matchings from M,
and deg(K) ≤ kt.
In this step we construct a graph K∗ from the subgraph K. As discussed, K∗ will not
Step 3.
be a subgraph of G but will embed into it with reasonable congestion. Let us formally define the
notion of graph embedding that we will be using.
Definition 2.9 (Graph embedding with congestion). Let G(V, E) and G(cid:48)(V, E(cid:48)) be graphs on the
same vertex set. Denote by PG the set of simple paths in G. An embedding of G(cid:48) into G is a map
f : E(cid:48) → PG such that every edge in G(cid:48) is mapped to a path in G with the same endpoints.
The congestion of f on an edge e ∈ E is cngf (e) := e(cid:48) ∈ E(cid:48) : e ∈ f (e(cid:48)). The congestion of
f is cng(f ) := maxe∈E cngf (e). We say that G(cid:48) embeds into G with congestion c if there is an
embedding f with cng(f ) = c.
The following claim is a simple observation and we omit its proof.
Claim 2.10. If G(cid:48) embeds into G with congestion c, then φ(G) ≥ 1
c φ(G(cid:48)).
We generate K∗ with the following inductive construction.
Lemma 2.11. Let ρ0 = c0 = 0. We can efficiently construct subgraphs K∗
have parallel edges and self-loops), such that for every i = 1, . . . , t,
1 , . . . , K∗
t (which may
1. K∗
2. K∗
i is ρi-regular, where ρi = k(1 + ρi−1).
i embeds into K with congestion ci, where ci = 1 + kci−1.
8
3. Every v ∈ Si has an incident edge in K∗
i with the other endpoint in S0.
Proof. We go by induction on i. For the base case i = 1 we simply set K∗
as we recall that
1 = K1. The claim holds
1. K1 is k-regular.
2. K1 is a subgraph of K, hence it embeds into K with congestion 1 = 1 + kc0.
3. By Step 2, every v ∈ S1 has an incident edge in K1 crossing to S0.
We turn to the inductive step i > 1. Start with a graph K(cid:48) which is a fresh copy of K∗
i−1, with
each edge duplicated into k parallel edges. By induction, K(cid:48) is (kρi−1)-regular. Now sum Ki into
K(cid:48); recall this means keeping parallel edges instead of unifying them. Since Ki is k-regular, K(cid:48) is
ρi-regular.
Let v ∈ Si. By Step 2, there is an edge vw ∈ Ki such that w ∈ Si−1. By induction, there is an
edge wu ∈ K∗
i−1 such that u ∈ S0. Note that both edges vw and wu are present in K(cid:48). Perform
the following crossing operation on K(cid:48): Remove the edges vw and wu, and add an edge vu and a
self-loop on w.
Perform this on every v ∈ Si. The resulting graph is K∗
i . We need to show that it is well defined
in the following sense: we might be using the same edge wu for several v's, and we need to make
sure each wu appears sufficiently many times, to be removed in all the crossing operations in which
it is needed. Indeed, we recall that Ki is the union of k disjoint perfect matchings, and therefore
each w ∈ Si−1 has at most k edges in Ki incoming from Si. Since K(cid:48) contains k copies of each edge
wu, we have enough copies to be removed in all necessary crossing operations.
Lastly we show that K∗
1. Since K(cid:48) was ρi-regular, and the switching operations do not effect vertex degrees, we see
i satisfies all the required properties.
that K∗
is ρi-regular.
2. Each edge vu in K∗
i
i which is not original from K(cid:48), corresponds to a path (of length 2) in
K(cid:48) that was removed upon adding that edge; hence K∗
i embeds into K(cid:48) with congestion 1.
K(cid:48) is the sum of Ki, which is a subgraph of K, and k copies of K∗
i−1, which by induction
embeds into K with congestion ci−1. Hence K(cid:48) embeds into K with congestion 1 + kci−1 = ci.
Therefore, K∗
i embeds into K with congestion ci.
3. For every v ∈ Si, we added to K∗
i an edge vu such that u ∈ S0.
Step 4.
i=1 K∗
the latter point holds, recall that we put ¯S = S0.)
i . By Lemma 2.11, K∗ is ((cid:80)t
i=1 ρi)-regular, embeds into K with
i=1 ci, and every v ∈ S has an incident edge vu ∈ K∗ such that u ∈ ¯S. (To see why
We now take K∗ =(cid:80)t
congestion(cid:80)t
This results in a subgraph ¯K of G which is kt-regular, and a graph ¯K∗ which is ((cid:80)t
embeds into ¯K with congestion(cid:80)t
embeds into K∪ ¯K (and hence into G, which contains K∪ ¯K) with congestion c, where r = 2(cid:80)t
and c = 2(cid:80)t
In this final step we repeat Steps 1 -- 3, only with the roles of S and ¯S interchanged.
i=1 ρi)-regular,
i=1 ci, and every v ∈ ¯S has an incident edge vu ∈ ¯K∗ such that
Our final weave is K∗ + ¯K∗. By the above it is clearly a weave, and moreover it is r-regular and
i=1 ρi
i=1 ci. By inspecting the recurrence formulas from Lemma 2.11, in which ρi and ci
were defined, we can bound ρi, ci ≤ (2k)i ≤ (2k)t for every i, and hence r, c ≤ 2t(2k)t. Recalling
that t ≤ 2
µ log n = O(log n), we find r, c ≤ (log n)O(1/βγ).
βγ + 1 and k = 1
u ∈ S.
9
2.5 Completing the Proof of Theorem 1.7
(cid:96)=1W(cid:96) and H∗ =(cid:80)L
We play the Cut-Weave game for L rounds, where L = O(r log2 n) is the number of rounds required
by the efficient strategy in Theorem 2.5. For each round (cid:96) = 1, . . . , L, we constructed above an
(cid:96) = K∗ + ¯K∗, that embeds into a subgraph W(cid:96) = K ∪ ¯K of G with congestion c.
r-regular weave W ∗
(cid:96) . Then H is a union of disjoint perfect matchings from M,
Let H = ∪L
and hence regular. Moreover deg(H) ≤ 2ktL, since H is the union of L subgraphs {W(cid:96)}L
(cid:96)=1, where
each W(cid:96) is a union W(cid:96) of two kt-regular graphs K, ¯K.
(cid:96) embeds into W(cid:96) with congestion c, we see that H∗ embeds
into H with congestion (at most) cL. By Theorem 2.5 we have φ(H∗) ≥ 1
2 r, and this now implies
φ(H) ≥ r
2cL .
Now consider H∗. Since each W ∗
1=(cid:96) W ∗
Recalling the parameters:
t = O(1) ; k = O(log n) ; r, c = O(logO(1/βγ) n) ; L = O(r log2 n),
we see that H is a d-regular subgraph of d = (log n)O(1/βγ) and φ(H) ≥ 1/(log n)O(1/βγ). We can
now repeat this Cut-Weave game (log n)O(1/βγ) disjoint times, because if each time we remove the
graph H we have found, we decrease the degree D = βn of each node by only polylog (n). By
repeating the game this many times and taking the union of the disjoint resulting subgraphs, we
find a regular subgraph H of G with deg(H) = (log n)O(1/βγ) and φ(H) ≥ 1. Lastly recall that
unfolding the reduction from Section 2.3 puts on H edge weight in {1, 2}, and weakens the degree
bound to deg(H) = (log n)O(1/βγ2). This completes the proof of Theorem 1.7.
Regarding the algorithm to construct H, the observation made after Theorem 2.7 applies here
as well. The weave player's strategy is oblivious to the queries of the cut player, since she just
samples random matchings from M to form H. The cut player strategy does not actually need
to be simulated, nor the graphs K∗ need to actually be constructed. The algorithm to construct
H then amounts to the following: Construct the double cover graph G" of G; decompose G" into
disjoint perfect matchings; choose a random subset of (log n)O(1/βγ2) of them to form a subgraph
H" of G"; and unfold the double cover construction to obtain the final subgraph H from H".
2.6 Proof of Theorem 1.8
The theorem follows from replacing the reduction to the double cover in Section 2.3 by a Hamil-
tonian decomposition result that holds for this stronger expansion requirement, due to Kuhn and
Osthus [KO14, Theorem 1.11]. The trade-off between β and γ is inherited from their theorem (in
which it is unspecified). Circumventing Section 2.3 also improves the dependence of d on γ. The
proof of Theorem 1.8 is otherwise identical to the proof of Theorem 1.7.
3 Proof of the Cut-Weave Theorem
Recall the setting of the Cut-Weave game with parameter r: The game starts with a graph G0 on
n vertices and without edges. In each round t = 1, 2, . . ., the weave player queries a bisection of the
vertex set, and the weave player answers with an r-regular weave Ht on that bisection. The weave
is then unified into the graph, putting Gt = Gt−1 ∪ Ht.
We now prove Theorem 2.5 by an adaptation of the analysis from [KRV09]. The main change
is in Lemma 3.6.
For each step t, let Mt be the matrix describing one step of the natural lazy random walk on
2r move to a neighbor. The cut player
2 stay in the current vertex, and with probability 1
Ht: W.p. 1
strategy is as follows:
10
• Choose a random unit vector z ⊥ 1 in Rn.
• Compute u = MtMt−1 . . . M1z.
• Output the bisection (S, . . . S) where S is the (cid:98)n/2(cid:99) vertices with smallest values in u.
Let us analyze the game with this strategy. In the graph Gt (which equals ∪t
t(cid:48)=1Ht(cid:48)), we consider
the following t-steps random walk: Take one (lazy) step on H1, then on H2, and so on until Ht. In
other words, the walk is given by applying sequentially M1, then M2, and so on.
Let Pij(t) denote the probability to go from node j to node i within t steps. Let Pi denote the
vector (Pi1, Pi2, . . . , Pji). We use the following potential function:
n(cid:88)
Lemma 3.1. For every t and every i ∈ V , we have(cid:80)
(Pij − 1/n)2 =
(cid:88)
Ψ(t) =
i,j∈V
i=1
(cid:107)Pi − 1/n(cid:107)2
2.
j∈V Pij(t) = 1.
Proof. By induction on t: It holds initially, and in each step t, vertex i trades exactly half of its
total present probability with its neighbors in Ht. (Note that this relies on the fact that Ht is
regular.)
2 r.
Lemma 3.2. If Ψ(t) < 1/4n2 then G = Gt has edge-expansion at least 1
2n for all i, j ∈ V . Hence the graph Kt on V , in which each
Proof. If Ψ(t) < 1/4n2 then Pji(t) ≥ 1
edge ij has weight Pji(t) + Pij(t), has edge-expansion 1
2 . We finish by showing that Kt embeds
into Gt with congestion 1/r. Proof by induction: Consider the transition from Gt−1 to Gt, which
is unifying Ht into Gt−1. Let i, j ∈ V be connected with an edge in Ht, and let k be any vertex. In
the transition from Kt−1 to Kt, we need to ship 1
2r Pik) to
j, and similarly, ship 1
2r Pjk probability from j to i. (The "type-k" probabiility is probability mass
that was originally located in k.) In total, we need to ship 1
2r from i to j and a
2r
similar amount from j to i. In total the edge ij in Ht needs to support 1
r flow (of probability) in
the transition, so the claim follows.
2r of the type-k probability in i (namely 1
k∈V Pik = 1
(cid:80)
We turn to analyzing the change in potential in a single fixed round t. To simplify notation we
let
Moreover recall we have a vector u generated by the cut player in the current round:
Pji = Pji(t)
; Qji = Pji(t + 1).
u = MtMt−1 . . . M1z.
Denote its entries by u1, . . . , un. We are now adding the graph Ht+1 to Gt to produce Gt+1.
Lemma 3.3. For every i, ui is the projection of Pi on r, i.e. ui = P T
i z.
Proof. Fix i. Abbreviate M = MtMt−1 . . . M1( 1
P T
i φ is the probability that the random walk lands in vertex i after t steps, meaning
n 1). If φ is any distribution on the vertices then
(M φ)i = P T
i φ.
(1)
n(cid:107)z(cid:107)∞ z. Applying Equation (1) with φ = z(cid:48) + 1
Let z(cid:48) = 1
Applying Equation (1) again with φ = 1
i z(cid:48), which implies ui = (M z)i = P T
P T
i z.
n 1 gives (M 1
n 1 gives (M (z(cid:48) + 1
n 1).
n 1) and together we get (M z(cid:48))i =
n 1))i = P T
i (z(cid:48) + 1
n 1)i = P T
i ( 1
11
Lemma 3.4. With probability 1 − 1/nΩ(1) over the choice of z, for all pairs i, j ∈ V ,
(cid:107)Pi − Pj(cid:107)2
2 ≥ n − 1
C log n
ui − uj2.
Proof. Similar to [KRV09, Lemma 3.4].
Lemma 3.5. Let E(S, ¯S) denote the set of edges in Ht+1 that cross the bisection (S, ¯S) produced
by the cut player (from the vector u). Then,
(cid:88)
ij∈E(S, ¯S)
(n − 1)E
ui − uj2
≥ Ψ(t).
Proof. Denote by deg(S, ¯S)(i) the number of edges in E(S, ¯S) incident to vertex i. Note that
deg(S, ¯S)(i) ≥ 1 for every i ∈ V , since Ht+1 is a weave on (S, ¯S). Recall that S contains the
vertices with smallest entries in u. Hence there is a number η ∈ R such that i ≤ η ≤ j for each
edge ij ∈ E(S, ¯S). Hence,(cid:88)
((ui − η)2 + (η − uj)2)
ij∈E(S, ¯S)
ij∈E(S, ¯S)
=
ui − uj2 ≥ (cid:88)
(cid:88)
≥(cid:88)
(cid:88)
≥(cid:88)
i∈V
i∈V
i∈V
=
u2
i ,
deg(S, ¯S)(i)(ui − η)2
(ui − η)2
i − 2η
u2
(cid:88)
ui + nη2
i∈V
where the last equality is by noting that z ⊥ 1, hence u ⊥ 1, hence(cid:80)
i∈V
Next, since ui = P T
Pi − 1/n on z. By properties of random projections we have E[u2
in [KRV09]), hence
i z and z ⊥ 1 we have ui = (Pi − 1/n)T z. Hence ui is the projection of
2 (see details
i ] = 1
i ui = 0.
n−1(cid:107)Pi − 1/n(cid:107)2
(cid:35)
(cid:34)(cid:88)
i∈V
E
(cid:88)
i∈V
u2
i
=
1
n − 1
(cid:107)Pi − 1/n(cid:107)2
2 =
1
n − 1
Ψ(t),
and the lemma follows from combining this with the above.
Lemma 3.6. Let Et+1 denote the edge set of Ht+1. The potential reduction is
Ψ(t) − Ψ(t + 1) =
1
r
(cid:107)Pi − Pj(cid:107)2
2.
(cid:88)
ij∈Et+1
Proof. We construct from G a graph G(cid:48) by splitting each vertex i into r copies i1, . . . , ir, assigning
arbitrarily one edge from the r edges incident to i in Et+1 to the copies, and distributing the type-j
probability in i, for each j, evenly among the copies. We denote by Pjik the amount of type-j
probability on ik before adding Et+1 to G(cid:48), and by Qjik the type-j probability in i after adding
r Pji for all i, j ∈ V and k ∈ [r], but for the Qjik 's all we
Et+1. Note that we have defined Pjik = 1
12
k=1 Qjik = Qji, so Qji may be distributed arbitrarily among the Qjik 's. As usual
Pik denotes the vector with entries Pjik , and Qik is defined similarly.
Define the potential of G(cid:48) as:
(cid:88)
r(cid:88)
i∈V
k=1
Ψ(cid:48)(t) =
r(cid:88)
(cid:88)
i∈V
k=1
(cid:107)Pik − 1/nr(cid:107)2
2.
r(cid:88)
r 1. Since we have(cid:80)r
(cid:107)Pik − 1/nr(cid:107)2
(cid:88)
2 = r
i∈V
(cid:107)Pi − 1/n(cid:107)2
2 = r
Pi − 1/nr(cid:107)2
(cid:107) 1
r
2 = rΨ(cid:48)(t).
to min(cid:107)x − c1(cid:107) s.t. x ∈ Rr,(cid:80)
To relate Ψ(t + 1) to Ψ(cid:48)(t + 1), we use the general fact that for any constants c and X, the solution
k=1 Qjik = Qji for
i xi = X is attained on x = X
k=1
know is that(cid:80)r
We thus have
Ψ(t) =
(cid:88)
i∈V
all i, j, we infer
Ψ(t + 1) =
(cid:107)Qi − 1/n(cid:107)2
2
=
i∈V
i,j∈V
(cid:88)
(cid:88)
(cid:88)
≤ (cid:88)
(cid:88)
i,j∈V
i,j∈V
=
= r
= rΨ(cid:48)(t + 1).
i∈V
k=1
(Qji − 1/n)2
r
(
k=1
1
r
Qji − 1/nr)2
r(cid:88)
r(cid:88)
(Qjik − 1/nr)2
r(cid:88)
(cid:107)Qik − 1/nr(cid:107)2
k=1
2
r
We have thus proven,
Ψ(t) − Ψ(t + 1) ≥ r(Ψ(cid:48)(t) − Ψ(cid:48)(t + 1)).
Now observe that Et+1 is, by construction, a perfect matching on G(cid:48). Therefore by [KRV09, Lemma
3.3] (which the current lemma generalizes),
Ψ(cid:48)(t) − Ψ(cid:48)(t + 1) ≥ (cid:88)
(cid:88)
(cid:88)
=
ik,jk(cid:48)∈Et+1
ik,jk(cid:48)∈Et+1
1
r2
i,j∈Et+1
=
(cid:107)Pik − Pjk(cid:48)(cid:107)2
2
Pj(cid:107)2
2
(cid:107) 1
Pi − 1
r
r
(cid:107)Pi − Pj(cid:107)2
2,
and the lemma follows.
Proof of Theorem 2.5. The initial potential is Ψ(0) = n − 1, and by Lemma 3.2 we need to get it
below 1/4n2. Putting Lemmas 3.4 to 3.6 together, we see that in each step we have in expectation
Ψ(t + 1) ≤ (1− 1
Cr log n )Ψ(t). Hence, in expectation, it is enough to play for O(r log2 n) rounds.
13
4 Resistance Sparsification
We prove Theorem 1.6 by combining Theorem 1.7 with the following known result.
Theorem 4.1 (von Luxburg, Radl and Hein [vLRH14]). Let G be a non-bipartite weighted graph
with maximum edge weight wmax and minimum weighted degree dmin. Let u, v be nodes in G with
weighted degrees du, dv respectively. Then
(cid:12)(cid:12)(cid:12)(cid:12)RG(u, v) −
(cid:18) 1
+
1
dv
du
(cid:18) 1
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2
+ 2
λ2(G)
(cid:19) wmax
.
d2
min
Qualitatively, the theorem asserts that in a sufficiently regular expander, the resistance distance
is essentially determined by vertex degrees. Therefore an expanding subgraph H of G with the same
weighted degrees can serve as a resistance sparsifier. In particular, in order to resistance-sparsify a
regular expander, all we need is a regular expanding subgraph, as we have by Theorem 1.7. Since
Theorem 4.1 does not apply to bipartite graphs, we will use the following variant that holds also
for bipartite graphs as long as they are regular. Its proof appears in Appendix A.1.
Theorem 4.2. Let G be a weighted graph which is d-regular in weighted degrees, with maximum
edge weight wmax. Let u, v be nodes in G. Then
(cid:12)(cid:12)(cid:12)(cid:12)RG(u, v) − 2
d
(cid:12)(cid:12)(cid:12)(cid:12) ≤ 12
(cid:18) 1
+ 2
λ2(G)
(cid:19) wmax
.
d2
Proof of Theorem 1.6. Using Theorem 1.7 we obtain a d-regular subgraph H of G with φ(H) > 1
3 .
By removing the obtained subgraph H from G and iterating, we can apply the theorem 3d/ε
times and obtain disjoint subgraphs H. Since d = (log n)O(1) and D = Ω(n), the degree of G
does not significantly change in the process, and the requirements of Theorem 1.7 continue to hold
throughout the iterations (with a loss only in constants). Taking the union of the disjoint subgraphs
produced in this process, we obtain a subgraph H of G which is (3d2/ε)-regular with φ(H) ≥ d/ε.
By the discrete Cheeger inequality,
(cid:18) φ(H)
deg(H)
(cid:19)2 ≥ 1
18d2 .
λ2(H) ≥ 1
2
Recall that H has edge weights in {1, 2}. We now multiply each weight by εD/(3d2), rendering it
D-regular in weighted degrees. This does not affect λ2(H) since it is an eigenvalue of the normalized
Laplacian.
Let u, v ∈ V . Apply Theorem 4.2 on both G and H. As G is D-regular with wmax = 1 and
d2 )
(cid:1). And as H is D-regular with wmax = O( εD
(cid:1). Putting these together, we get
D ± O(cid:0) ε
(cid:1) = 1 ± O (ε) , where the last equality holds for sufficiently large n since
D ± O(cid:0) 1
D
λ2(G) = Ω(1), we know that RG(u, v) = 2
and λ2(H) = Ω(1/d2), we know that RH (u, v) = 2
RH (u,v)
RG(u,v) = 1 ± O(cid:0)ε + 1
D2
D
D = Ω(n). Scaling ε down by the constant hidden in the last O(ε) notation yields the theorem.
Acknowledgements
We thank Uriel Feige for useful comments on this work.
14
References
[AKW14] A. Andoni, R. Krauthgamer, and D. P. Woodruff. The sketching complexity of graph
cuts. CoRR, abs/1403.7058, 2014. arXiv:1403.7058.
[BK96]
A. A. Bencz´ur and D. R. Karger. Approximating s-t minimum cuts in O(n2) time.
In 28th Annual ACM Symposium on Theory of Computing, pages 47 -- 55. ACM, 1996.
doi:10.1145/237814.237827.
[BSS12]
J. D. Batson, D. A. Spielman, and N. Srivastava. Twice-ramanujan sparsifiers. SIAM
J. Comput., 41(6):1704 -- 1721, 2012. doi:10.1137/090772873.
[CKL+14] B. Csaba, D. Kuhn, A. Lo, D. Osthus, and A. Treglown. Proof of the 1-factorization
and Hamilton decomposition conjectures. ArXiv e-prints, abs/1401.4159, 2014. arXiv:
1401.4159.
[DL97]
M. M. Deza and M. Laurent. Geometry of cuts and metrics. Springer-Verlag, Berlin,
1997.
[KKOV07] R. Khandekar, S. A. Khot, L. Orecchia, and N. K. Vishnoi. On a cut-matching game for
the sparsest cut problem. Technical Report UCB/EECS-2007-177, EECS Department,
University of California, Berkeley, 2007. Available from: http://www.eecs.berkeley.
edu/Pubs/TechRpts/2007/EECS-2007-177.html.
[KO14]
[KP12]
D. Kuhn and D. Osthus. Decompositions of complete uniform hypergraphs into hamil-
ton berge cycles. J. Comb. Theory, Ser. A, 126:128 -- 135, 2014. doi:10.1016/j.jcta.
2014.04.010.
M. Kapralov and R. Panigrahy. Spectral sparsification via random spanners. In 3rd
Innovations in Theoretical Computer Science Conference, pages 393 -- 398. ACM, 2012.
doi:10.1145/2090236.2090267.
[KRV09] R. Khandekar, S. Rao, and U. V. Vazirani. Graph partitioning using single commodity
flows. J. ACM, 56(4), 2009. doi:10.1145/1538902.1538903.
[Lov96]
[PR97]
[PS89]
[SS11]
[ST04]
L. Lov´asz. Random walks on graphs: a survey. In Combinatorics, Paul Erdos is eighty,
Vol. 2 (Keszthely, 1993), volume 2 of Bolyai Soc. Math. Stud., pages 353 -- 397. J´anos
Bolyai Math. Soc., Budapest, 1996.
L. Perkovic and B. Reed. Edge coloring regular graphs of high degree.
Math.,165/166, pages 567 -- 578, 1997.
In Discrete
D. Peleg and A. A. Schaffer. Graph spanners. J. Graph Theory, 13(1):99 -- 116, 1989.
doi:10.1002/jgt.3190130114.
D. A. Spielman and N. Srivastava. Graph sparsification by effective resistances. SIAM
J. Comput., 40(6):1913 -- 1926, December 2011. doi:10.1137/080734029.
D. A. Spielman and S.-H. Teng. Nearly-linear time algorithms for graph partitioning,
graph sparsification, and solving linear systems. In 36th Annual ACM Symposium on
Theory of Computing, pages 81 -- 90. ACM, 2004. doi:10.1145/1007352.1007372.
15
[ST11]
D. A. Spielman and S.-H. Teng. Spectral sparsification of graphs. SIAM J. Comput.,
40(4):981 -- 1025, July 2011. doi:10.1137/08074489X.
[vLRH14] U. von Luxburg, A. Radl, and M. Hein. Hitting and commute times in large random
neighborhood graphs. Journal of Machine Learning Research, 15(1):1751 -- 1798, 2014.
Available from: http://jmlr.org/papers/v15/vonluxburg14a.html.
A Appendix: Omitted Proofs
A.1 Proof of Theorem 4.2
In the non-bipartite case, Theorem 4.2 follows from Theorem 4.1. We henceforth assume that
G = (V, E, w) is bipartite with bipartition V = V1 ∪ V2. Note that since it is regular, we must have
V1 = V2 = 1
2V . Furthermore, as a weighted regular bipartite graph, G is a convex combination
of perfect matchings and hence is regular also in unweighed degrees. Let d(cid:48) denote the unweighted
degree of each vertex in G. If d(cid:48) ≤ 2 then it is easy to verify that the theorem holds (due to poor
expansion), so we henceforth assume d(cid:48) ≥ 3.
For brevity we denote the error term in Theorem 4.1 as
(cid:18) 1
(cid:19) wmax
.
d2
err := 2
+ 2
λ2(G)
We will use the notion of hitting time: For a pair of vertices u, v, the hitting time HG(u, v) is
defined as the expected time it takes a random walk in G that starts at u, to hit v. Define the
normalized hitting time hG(u, v) = 1
2W HG(u, v), where W is the sum of all edge weights in G. We
then have,
RG(u, v) = hG(u, v) + hG(v, u).
(2)
We will use the following bound on the normalized hitting time, which is given in the same theorem
by von Luxburg, Radl and Hein [vLRH14].
Theorem A.1. In the same setting of Theorem 4.1,
∀u (cid:54)= v ∈ V,
hG(u, v) =
± err.
1
dv
(Like Theorem 4.1, this theorem does not apply to bipartite graphs, and this is the obstacle we
are now trying to circumvent.)
We begin by handling pairs of vertices contained within the same partition side, say V1. We
construct from G a weighted graph G1 on the vertex set V1, with weights w1, by putting
∀i (cid:54)= j ∈ V1, w1(i, j) =
1
d
w(i, k)w(j, k).
(cid:88)
k∈V2
We argue that HG1(u, v) = 1
2 HG(u, v). This follows by observing that we set the weights w1 such
that for any i, j ∈ V1, the probability to walk in one step from i to j in G1 equals the probability to
walk in two steps from i to j in G via an intermediate node in V2. Furthermore, we have normalized
2V , we have
the weights w1 such that G1 is d-regular in weighted degrees. Recalling that V1 = 1
hG1(u, v) =
1
dV1 HG1(u, v) =
2
dV · 1
2
HG(u, v) = hG(u, v).
16
Recalling that the unweighted degree in G is d(cid:48) ≥ 3, we see that by construction, G1 contains a
triangle and hence is non-bipartite. Hence we can apply to it Theorem A.1 and obtain hG1(u, v) =
d ± err1, where err1 is the error term of G1. Note that for every i (cid:54)= j ∈ V1 we have w1(i, j) ≤
1
w(i, k) = wmax, so the maximum edge weight in G1 is bounded by wmax, and λ2(G1) ≥
λ2(G) (easy to verify by construction), so err1 ≤ err, and we have hG1(u, v) = 1
d ± err. Hence,
(cid:80)
k∈V2
wmax
d
Recalling that RG(u, v) = hG(u, v) + hG(v, u), we have established that
hG(u, v) =
± err.
1
d
RG(u, v) =
± 2err
2
d
for every pair u, v ∈ V1. The same arguments hold for every pair u, v ∈ V2 as well. We are left to
handle the case u ∈ V1, v ∈ V2. Recalling the definition of hitting time, we have
HG(u, v) = 1 +
w(u, v)
d
· 0 +
w(u, x)
d
HG(x, v)
(factoring out the first step)
= 1 +
w(u, v)
· 0 +
w(u, x)
· 2W · hG(x, v)
(cid:88)
(cid:88)
x∈V2\{v}
x∈V2\{v}
w(u, x)
(cid:18) 1
(cid:19)(cid:18) 1
d
d
± err
(cid:19)
(cid:19)
.
± err
d
(cid:18)
d
(cid:88)
= 1 + 2W
= 1 + 2W
(cid:18)
d
x∈V2\{v}
1 − w(u, v)
d
hG(u, v) =
1
2W
+
1 − w(u, v)
d
hG ≤ 1
2W
± err
1
d
(cid:19)
(cid:16)
+
(cid:17)(cid:18) 1
(cid:88)
d
j∈V
d = deg(i) =
w(i, j) ≤ nwmax,
Therefore
which implies
and
(since v, x ∈ V2)
(cid:19)(cid:18) 1
d
(cid:19)
± err
,
hG(u, v) ≥ 1
2W
1 − wmax
d
± 2err.
2W ± 2err. Now, since for an arbitrary vertex i we have
± err
1
2W
d + 1
1
d
+
+
=
Together, hG(u, v) = 1
we see that
1
2W = 1
nd ≤ wmax
d2 ≤ err and hence
hG(u, v) =
± 3err.
1
d
Plugging this into RG(u, v) = hG(u, v) + hG(v, u), we find
± 6err,
RG(u, v) =
2
d
which completes the proof of Theorem 4.2.
17
4 n. Then G contains a perfect matching.
4 n neighbors in U , we have N (S) ≥ N ({v}) ≥ 1
A.2 Further Omitted Proofs
Proposition A.2. Let G(V, U ; E) be a bipartite graph on n nodes with V = U = 1
minimum degree ≥ 1
Proof. Let S ⊂ V be non-empty, and denote N (S) ⊂ U the set of nodes with a neighbor in S. If
S ≤ 1
4 n then since any v ∈ S has 1
4 n ≥ S. If
S > 1
4 n then by the minimum degree condition on side U , every u ∈ U must have a neighbor in
S, and hence N (S) = U = V ≥ S. The same arguments apply for S ⊂ U , so the condition of
Hall's Marriage Theorem is verified, and it implies that G contains a perfect matching.
Proposition A.3. Consider an instance of Set Cover with a set S of n elements, and a family M
of subsets of S. Suppose each x ∈ S belongs to at least a µ-fraction of the subsets in M. Then for
sufficiently large n, we can efficiently find a cover M ⊂ M with M ≤ 1.1
Proof. Pick q uniformly random sets (with replacement) from M to form M . The probability that
a given element in S is not covered by M is upper-bounded by (1− µ)q. Taking a union bound over
the element, we need to ensure that n(1− µ)q < 1 in order to ensure that with constant probability,
M is a solution to the given Set Cover instance. This can be achieved by q ≤ 1.1
2 n, and
µ log n.
µ log n.
18
|
1209.5821 | 3 | 1209 | 2013-11-16T15:08:23 | Faster spectral sparsification and numerical algorithms for SDD matrices | [
"cs.DS"
] | We study algorithms for spectral graph sparsification. The input is a graph $G$ with $n$ vertices and $m$ edges, and the output is a sparse graph $\tilde{G}$ that approximates $G$ in an algebraic sense. Concretely, for all vectors $x$ and any $\epsilon>0$, $\tilde{G}$ satisfies $$ (1-\epsilon) x^T L_G x \leq x^T L_{\tilde{G}} x \leq (1+\epsilon) x^T L_G x, $$ where $L_G$ and $L_{\tilde{G}}$ are the Laplacians of $G$ and $\tilde{G}$ respectively. We show that the fastest known algorithm for computing a sparsifier with $O(n\log n/\epsilon^2)$ edges can actually run in $\tilde{O}(m\log^2 n)$ time, an $O(\log n)$ factor faster than before. We also present faster sparsification algorithms for slightly dense graphs. Specifically, we give an algorithm that runs in $\tilde{O}(m\log n)$ time and generates a sparsifier with $\tilde{O}(n\log^3{n}/\epsilon^2)$ edges. This implies that a sparsifier with $O(n\log n/\epsilon^2)$ edges can be computed in $\tilde{O}(m\log n)$ time for graphs with more than $O(n\log^4 n)$ edges. We also give an $\tilde{O}(m)$ time algorithm for graphs with more than $n\log^5 n (\log \log n)^3$ edges of polynomially bounded weights, and an $O(m)$ algorithm for unweighted graphs with more than $n\log^8 n (\log \log n)^3 $ edges and $n\log^{10} n (\log \log n)^5$ edges in the weighted case. The improved sparsification algorithms are employed to accelerate linear system solvers and algorithms for computing fundamental eigenvectors of slightly dense SDD matrices. | cs.DS | cs |
Faster spectral sparsification
and numerical algorithms for SDD matrices
Ioannis Koutis
CSD-UPRRP
[email protected]
Alex Levin
MATH-MIT
[email protected]
Richard Peng
CSD-CMU
[email protected]
May 2, 2014
Abstract
We study algorithms for spectral graph sparsification. The input is a graph G with n vertices and
m edges, and the output is a sparse graph G that approximates G in an algebraic sense. Concretely, for
all vectors x and any ǫ > 0, the graph G satisfies
(1 − ǫ)xT LGx ≤ xT L Gx ≤ (1 + ǫ)xT LGx,
where LG and L G are the Laplacians of G and G respectively.
We show that the fastest known algorithm for computing a sparsifier with O(n log n/ǫ2) edges can
actually run in O(sf m log2 n) time1, an O(log n) factor faster than before. We also present faster sparsifi-
cation algorithms for slightly dense graphs. Specifically, we give an algorithm that runs in O(tf m log n)
time and generates a sparsifier with O(sf n log3 n/ǫ2) edges. We also give an O(m log log n) time al-
f n log5 n log log n edges and an O(m) algorithm for graphs with
gorithm for graphs with more than s2
f n log8 n edges. The improved
more than t2
sparsification algorithms are employed to accelerate linear system solvers and algorithms for computing
fundamental eigenvectors of slightly dense SDD matrices.
f n log10 n edges and unweighted graphs with more than s3
f s3
1
Introduction
The efficient transformation of dense instances of graph problems to nearly equivalent sparse instances is
a very powerful tool in algorithm design. The idea, widely known as graph sparsification, was originally
introduced by Bencz´ur and Karger [3] in the context of cut problems. Spielman and Teng [15] generalized
the cut-preserving sparsifiers of Bencz´ur and Karger to the more powerful spectral sparsifiers, which preserve
in an algebraic sense the Laplacian matrix of the dense graph. The main motivation of spectral sparsifiers
was the design of nearly-linear time algorithms for the solution of symmetric diagonally dominant (SDD)
linear systems. A matrix A is SDD if it is symmetric and for all i, Aii ≥ Pj6=i Aij.
Bencz´ur and Karger proved that, for arbitrary ǫ, cuts can be preserved within a factor of 1 ± ǫ by
a graph with O(n log n/ǫ2) edges. This graph can be computed by a randomized algorithm that runs in
O(m log3 n) time2, where m is the number of edges in the dense graph. Spielman and Teng gave the first
construction of spectral sparsifiers, but the edge count of these objects was several log factors bigger than
1We denote by sf and tf the O(log log n) factors that appear in the stretch and time guarantees of the best currently
known algorithm for computing low-stretch trees [1]. It is conjectured that sf = O(1) and tf = O(1) is possible.
2All sparsification algorithms in this paper are randomized with a probability failure inversely proportional to n. They
consist of a preprocessing phase followed by the generation of the sparsifier which in general can be performed in time
proportional to the number of edges in it (e.g. O(n log n/ǫ2)). For the sake of conciseness our running time statements will
include only the time for preprocessing and will omit the failure probability.
1
that of Bencz´ur and Karger’s cut-preserving sparsifiers. However, recent progress that we review below
allows for the construction of spectral sparsifiers with O(n log n/ǫ2) edges in O(sf m log3 n) time.
Sparsification can be employed to immediately accelerate algorithms for numerous problems. In several
cases and depending on the density of the instance, the sparsification routine dominates the running time of
the sparsifier-enhanced algorithm. This is a strong incentive for speeding up the construction of sparsifiers
even further.
This problem was undertaken in the context of cut-preserving sparsifiers by Fung et al. [5]. Improving
upon the work of Bencz´ur and Karger, they proved that there is an O(m log2 n) time algorithm that
computes a sparsifier with O(n log n/ǫ2) edges. This stands as the fastest known algorithm with this
sparsity guarantee for general graphs. However, Fung et al. also showed that we can do even better on
slightly more dense graphs. More concretely, they proved that there is an O(m) time algorithm that
computes a sparsifier with O(n log2 n/ǫ2) edges. Note that by transitivity, a combination of the two
algorithms can produce a graph with O(n log n/ǫ2) edges in O(m + n log4 n) time. In other words, there is
a linear time sparsification algorithm for graphs with more than n log4 n edges.
This leads us to the main question we address in this paper: Is something analogous possible for spectral
sparsification? We answer the question in the affirmative. We first show that a slight modification of the
Spielman-Srivastava algorithm [14] can improve the run time to O(sf m log2 n). This nearly matches the
general case algorithm of [5]. We present three additional sparsification algorithms. The first is a variation
of the Spielman-Srivastava algorithm that generates a sparsifier with O(sf n log3 n/ǫ2) edges in O(tf m log n)
time. The second produces a sparsifier with O(n log n/ǫ2) edges in O(m log log n) time, assuming the
f n log5 n log log n edges. The third produces a sparsifier with O(n log n/ǫ2) edges in
input has more than s2
f n log8 n edges, or if it has more than
O(m) time assuming the input is unweighted and has more than s3
t2
f s3
Applications in numerical algorithms
The (1 ± ǫ)-sparsifiers we obtain can be employed in a standard way as preconditioners for SDD lin-
ear systems, giving us faster solvers for slightly dense graphs: (i) an O(m log log n) time solver for sys-
f n log5 n log log n non-zero entries and (ii) an O(m) time solver for Laplacians of
tems with more than s2
f n log10 n non-zero entries. The best previously known algorithm [13] runs in
f s3
graphs with more than t2
O(sf m log n log(1/δ)) time.
f n log10 n in the weighted case.
In addition, our sparsification algorithms accelerate the computation of an approximate Fiedler eigen-
vector of a graph Laplacian LG. An (1+ǫ)-approximate eigenvector is a unit norm vector x such that xT LGx
is within a factor 1 + ǫ from the eigenvalue λ2 of LG. The algorithm consists of two steps: (i) computing
a spectral sparsifier G that (1 ± ǫ/2)-approximates the input graph G and (ii) computing a (1 + ǫ/3)-
approximate eigenvector of G; this will automatically be an (1 ± ǫ)-approximate eigenvector of the (more)
dense input graph because the spectral sparsification step preserves the eigenvalues of G within 1 ± ǫ/2.
Hence combining our sparsification algorithms with the inverse power method [16] (which consists of solving
f n log5 n log(1/ǫ)/ǫ2) time.
O(log n log(1/ǫ)) systems in L G) gives an approximate eigenvector in O(m + s2
The fastest previously known algorithm runs in time O(sf m log2 n log(1/ǫ)). The same result applies to
the computation of the Fiedler eigenvector of a normalized Laplacian D−1/2LGD−1/2; applying the inverse
power method on D−1/2L GD−1/2 gives the required eigenvector.
We note here that one practical application of eigenvectors is in partitioning algorithms; the analysis of
Cheeger’s inequality [4] tells us how to turn an approximate Fiedler vector into a partition. Hence, we give
an improvement to the running time of a fundamental graph partitioning algorithm. Finally we note that
the computation of additional eigenvectors can be performed in the same amount of time (per vector) by
restricting the action of the matrix to the complement of the subspace spanned by the previously computed
eigenvectors.
2
2 Overview of our techniques
2.1 Brief background on spectral sparsification
The first algorithm for edge-efficient spectral sparsifiers was given by Spielman and Srivastava [14]. Their
algorithm produces a sparsifier with O(n log n/ǫ2) edges in a very elegant way:
it samples edges with
replacement. The probability of sampling an edge is proportional to its weight multiplied by its effective
resistance in the resistive electrical network associated with the given graph.
Computing the effective resistance of a given edge requires—almost by definition—the solution of a
linear system on the graph Laplacian.3 However, Spielman and Srivastava also provided a way of estimating
all m effective resistances, via solving O(log n) SDD linear systems. This holds under the assumption that
the SDD solver is direct, i.e. it outputs an exact solution. The use of a nearly-linear time iterative solver that
computes approximate solutions introduces an additional source of imprecision; Spielman and Srivastava
showed that solving the systems up to an inverse polynomial precision is sufficient for sparsification. This
brings the running time of their algorithm to O(sf m logc+2 n), where c is the constant appearing in the
running time of the SDD solver.
2.2 The O(sf m log2 n) time algorithm
While the work of Spielman and Srivastava did not improve the running time of the SDD solver, it proved
to be a decisive step towards the fast SDD solver of Koutis, Miller, and Peng [12, 13], which runs in time
O(sf m log n log(1/δ)), where δ is the desired precision. Using this solver in the Spielman and Srivastava
sparsification sampling scheme immediately yields an O(sf m log3 n/ǫ2) time algorithm. This brings us
to the first contribution of this paper, a tighter analysis of the Spielman and Srivastava algorithm.
In
Section 5 we show that solving the systems up to fixed precision is actually sufficient for sparsification.
This decreases the running time to O(sf m log2 n/ǫ2).
2.3 Faster algorithms: The main idea
To get our two faster algorithms, we will trade accuracy in the computation of effective resistances for
speed. The idea is to transform the input graph G into another graph H where effective resistances can be
computed faster while still providing good bounds for the true effective resistances in G. These approximate
effective resistances can still be used for sparsification at the expense of additional sampling [12] that yields
slightly more dense sparsifiers. These sparsifiers can be re-sparsified to O(n log n/ǫ2) edges by applying
the fast general-case algorithm.
2.4 The O(tf m log n) time algorithm
The O(tf m log n) time algorithm is based on the observation that the Spielman-Srivastava scheme can be
implemented to run in O(tf m log n) time on a spine-heavy approximation H of G. The spine-heavy graph
H is derived in O(tf m log n) time from G by computing a low-stretch tree of G and scaling it up by a
O(sf log2 n) factor. In [13] it was shown that linear systems involving the Laplacian of H can be solved
in O(m) time, enabling the faster implementation of the Spielman-Srivastava scheme on H. At the same
time the effective resistances in H are at most a O(sf log2 n) factor smaller than those in G. Sampling with
respect to these estimates, allows us to get a sparsifier G′ with O(sf n log3 n/ǫ2) edges. Re-sparsifying G′
gives a sparsifier G with O(n log n/ǫ2) edges in O(s2
f n log5 n/ǫ2) time. The details are given in Section 5.
2.5 The O(m log log n) and O(m) time algorithms
There are two major bottlenecks in the O(tf m log n) time algorithm. In order to work around them, we
introduce several ideas of independent interest.
3Laplacian matrices are SDD.
3
(a) The first bottleneck is in the computation of the low-stretch tree; all known algorithms for comput-
ing a low-stretch tree run in time at least O(m log n). The solution to this problem involves two steps. We
first observe that if we can settle for a weaker stretch guarantee, it is enough to find a low-stretch tree of a
subgraph H of the input graph G which preserves all cuts of G within a polylogarithmic factor. This allows
the existing low-stretch tree computation to run faster, assuming the subgraph is sparser by an O(tf log n)
factor. The subgraph H can be viewed as an approximate cut sparsifier of G, a notion very similar to
the incremental spectral sparsifiers from [12]; these are computed by first finding a low-stretch tree and
sampling off-tree edges with probability proportional to their stretch over the tree. Inspired by this idea,
we give an even simpler algorithm for computing the graph H: we find a maximum weight spanning tree
and then we sample uniformly the off-tree edges. The proof that a variant of this simple procedure returns
the desired incremental cut sparsifier relies on Karger’s earlier work on cut sparsification [8].
(b) Having removed the low-stretch tree computation obstacle we can now attempt to mimic the steps
of the O(tf m log n) time algorithm. In fact it is possible to take care of the system-solving part of the
Spielman-Srivastava scheme in O(m) time by merely scaling-up the low-stretch tree by a larger factor.
However this is not enough; there is still a bottleneck that lies in the computations after the solution of
the linear systems in the heavier-spine graphs. These are m simple manipulations of vectors of dimension
O(log n). In an earlier version of this work, we attempted to work around this problem by reducing the
dimension of these vectors, but at a significant loss of sparsity [11].
Here we take a different route. We first use the low-stretch tree to construct an approximate sparsifier
i.e. a sparse approximation for the input graph, but of moderate quality. The significant departure from the
Spielman-Srivastava scheme comes in the next step which computes estimates of the effective resistances
using a combinatorial rather than algebraic approach. Concretely, we observe that with some additional
work we can ‘leverage’ the approximate sparsifier to compute a sparsifier for G as well.
Indeed,
let H be a κ-approximation of G (see Definition 3.1) and H be a sparsifier of H with
O(n log n) edges; H is the approximate sparsifier. Then we generate a low-stretch spanning tree T of
H in O(sf n log2 n) time and approximate the effective resistances of G over T in O(m) time. We will be
able to claim that these approximate values are enough to generate a sparsifier G′ for G with O(sf nκ log3 n)
edges. Finally from G′ we can compute a sparsifier G with O(n log n/ǫ2) edges in O(s2
f κn log5 n) time using
our first algorithm.
We will derive our O(m) algorithm via a single application of the above ‘leveraging’ idea for κ =
f log5 n). To improve over the O(sf m log n) algorithm for even sparser graphs we will progressively
O(s3
sparsify a sequence of t = O(log log n) graphs H = H0, H1, . . . , Ht = G, such that Hi is a 2-approximation
of Hi+1: given the sparsifier for Hi we can construct the sparsifier for Hi+1 via the leveraging idea. The
details are given in Section 6.
3 Background on spectral graph theory and sparsification
3.1 The graph Laplacian and its pseudoinverse
Let G = (V, E, w) be an undirected weighted graph on n vertices, which we identify with the integers
{1, 2, . . . , n}, and m edges, where the weight of edge e is given by we. Without loss of generality we will
assume that minimum weight is 1. We will also assume that matrices discussed below are represented as
adjacency lists.
The Laplacian of G is denoted by LG. It is a symmetric n × n matrix with zero row and column sums,
where the (i, j) off-diagonal entry is given by −w(i,j) if (i, j) is an edge of G and 0 otherwise. The ith
diagonal entry is given by the weighted degree of vertex i.
If G is a connected graph, then LG is a matrix of rank n − 1, with its kernel spanned by 1 (the vector
G denote the Moore-Penrose pseudoinverse of LG; this is a matrix that acts as the
G = In−1, where In−1 is the projection onto the
of all 1’s). We let L+
inverse of LG on (ker LG)⊥, and satisfies L+
GLG = LGL+
4
(n − 1)-dimensional image of LG.
notation to graphs, with the obvious meaning.
Given the one-to-one correspondence of graphs and their Laplacians we will often apply algebraic
3.2 Spectral approximation and sparsification
In this paper we concentrate on symmetric diagonally dominant matrices. For two matrices A and B of
the same dimension, we write A (cid:22) B if xT Ax ≤ xT Bx for all vectors x. For two graphs G and H, we write
G (cid:22) H if the Laplacians satisfy LG (cid:22) LH.
Definition 3.1 We say that a graph H is a κ-approximation of a graph G if G (cid:22) H (cid:22) κG.
It is not hard to show that if H is a graph that κ-approximates a graph G then we have
1
κ
G (cid:22) L+
L+
H (cid:22) L+
G
(3.1)
Definition 3.2 Given a graph G, we say that a (sparser) graph H is a 1 ± ǫ spectral sparsifier of G if
(3.2)
(1 − ǫ)G (cid:22) H (cid:22) (1 + ǫ)G.
It is easy to see that if H is a 1± ǫ spectral sparsifier of G then 1
1−ǫ -approximates
G. By the definition, it is also easy to verify transitivity. If G1 is a 1 ± ǫ1 sparsifier of G and G2 is a
1 ± ǫ2 of G1 then G2 is a (1 ± ǫ1)(1 ± ǫ2) sparsifier of G.
3.3 Graphs as resistive electrical networks
1−ǫ H is a graph that 1+ǫ
We can consider our graph G as an electrical network of nodes (vertices) and wires (edges), where edge e
has resistivity of w−1
e Ohms.
In this context it is very useful to give another definition of the Laplacian LG, in terms of its incidence
matrix BG. To define BG, fix an arbitrary orientation for each edge in G. For a vertex i let χi be its
(n × 1) characteristic vector, with a 1 at the ith entry and 0’s everywhere else. Let e = (i, j) be an edge
and define be = χi − χj. Then BG is the m × n matrix whose eth row is the vector be. Let WG be the
m × m diagonal matrix whose eth diagonal entry is we. With these definitions, it is easy to verify that
LG = BT
GWGBG = X
e∈G
webebT
e .
For notational convenience, we will drop the subscripts on LG, BG, and WG when the graph we are
dealing with is clear from context.
Going back to the electrical analogy, the effective resistance between vertices i and j, denoted by
RG(i, j) or RG(e) when (i, j) is an edge e, is the voltage difference that has to be applied between i and j
in order to drive one unit of external current between the two vertices. Algebraically it is given by
RG(i, j) = (χi − χj)T L+
The above equation allows us to apply (3.1) and see that
G(χi − χj)
G (cid:22) H (cid:22) κG ⇒ (1/κ)RG(e) ≤ RH(e) ≤ RG(e).
(3.3)
(3.4)
The definition of the effective resistance for (i, j) in (3.3) shows directly that it can be computed by
solving the system LGx = (χi − χj). In light of this, (3.4) will be of central importance in our proofs.
Informally, it states that if H is a κ-approximation of G, then the effective resistance of any edge in G
5
can be approximated by the effective resistance of the same edge in H, which can be done by solving the
system LHx = (χi − χj). This will allow us to construct special approximations H for which solving with
LH is easier than with LG.
3.4 Low-stretch subgraphs, spine-heavy graphs, and SDD solvers
Let S be a graph on the same vertex set as a graph G. Let e = (i, j) be an edge of G. If p is a path
e1, e2, . . . , eν between i and j in S we say that the stretch of e over p is stretchp(e) := we Pν
ei , i.e.
the weight of e multiplied by the sum of inverse weights of tree edges on the path from i to j. If P(e) is
the set of all paths between i and j in S we define
i=1 w−1
stretchS(e) = min
p∈P(e)
stretchp(e).
We will use the term stretch of e over S for stretchS(e). The definition is simpler when S is a tree. In this
case there is a unique path between the endpoints of e. We denote by stretchS(G) the sum of stretches in
S of all edges of G, i.e.
stretchS(G) = X
e∈G
stretchS(e).
It is known that every graph G has a spanning tree T with stretchT (G) = O(m log n log log n), known
as a low-stretch tree. The tree can computed in O(m log n log log n) time [1]. Because these guarantees
are still open to improvement we will state our results with respect to two parameters: We will denote
by tf the factor in excess of O(m log n) in the time required for computing a low-stretch tree on a graph
with m edges, via the algorithm in [1]. Similarly, we will denote by sf the factor in excess of O(m log n)
in stretchT (G) provided by the same algorithm. That is, as noted above, the best current guarantees are
sf = O(log log n) and tf = O(log log n).
We call a graph spine-heavy if it has a spanning tree with stretchT (G) = O(m/ log n). Given a graph
G we can compute a spine-heavy graph H that O(sf log2 n)-approximates it by computing a low-stretch
tree and then scaling up the weights of tree edges in G by the O(sf log2 n) factor. This is summarized in
the following lemma.
Lemma 3.3 For every graph G with n vertices there is a spine-heavy graph H that O(sf log2 n)-approximates
G. The graph H can be constructed in time dominated by the computation of a low-stretch tree for G.
Finally we state a lemma that summarizes the recent work on fast SDD solvers [13].
Lemma 3.4 Let A be an SDD matrix. There is a symmetric operator Aδ such that
and that for any vector b, the vector A+
Moreover, if A is the Laplacian of a spine-heavy graph and its low-stretch tree is given, then A+
evaluated in O(m log(1/δ)) time.
(1 − δ)A (cid:22) Aδ (cid:22) (1 + δ)A
δ b can be evaluated in O(tf m log n + sf m log n log(1/δ)) time.
δ b can be
3.5 Sampling for sparsification
In a remarkable work, Spielman and Srivastava [14] analyzed a spectral sparsification algorithm based on
a simple sampling procedure. The procedure will be central in our algorithms and we review it here. It
takes as input a weighted graph G and frequencies p′
e for each edge e. These frequencies are normalized to
probabilities pe summing to 1. It then picks in q rounds exactly q samples, which are weighted copies of the
edges. The probability that given edge e is picked in a given round is pe. The weight of the corresponding
6
Sample
e
Input: Graph G = (V, E, w), p′ : E → R+.
Output: Graph G′ = (V,L, w′).
1: t := Pe p′
2: q := Cst log t/ǫ2
3: pe := p′
4: G′ := (V,L, w′) with L = ∅
5: for q times do
6:
e/t
(* CS is an explicitly known constant *)
Sample one e ∈ E with probability of picking e being pe
Add l, a sample of e, to L with weight w′
l = we/pe
7:
8: end for
9: For all l ∈ L, let w′
10: return G′
l := w′
l/q
sample is set so that the expected weight of the edge e after sampling is equal to its actual weight in the
input graph. The details are given in the following pseudocode.
Spielman and Srivastava analyzed the case when p′
e = weRG(e), where RG(e) is the effective resistance
of e in G. The following generalization characterizes the quality of G′ as a spectral sparsifier for G. It is
shown in [9] and it was originally proved with a weaker success guarantee in [12].
Theorem 3.5 (Oversampling) Let G = (V, E, w) be a graph. Assuming that p′
edge e ∈ E the graph G′ = Sample(G, p′) is a (1 ± ǫ) sparsifier of G with probability at least 1 − 1/n2.
3.6
Incremental spectral sparsifiers
e ≥ weRG(e) for each
In [12], Koutis, Miller, and Peng asked whether they could get anything useful out of the Spielman-
Srivastava construction, without having to use effective resistances whose computation requires the solution
of linear system solvers. The oversampling Theorem 3.5 offers a possibility as long as we are able to
compute, more efficiently, upper bounds to the quantities weRG(e).
The idea in [12] was to use spanning trees and use as an upper bound the stretch of an edge over the
tree. The idea generalizes readily to spanning subgraphs. The reason is that for any subgraph S, we have
stretchS(e) ≥ weRG(e) by Rayleigh’s theorem.
In the case S is a tree, we can compute the stretches of all the edges in O(m) time (using an offline
lowest common ancestor algorithm [18, 6]). For one of our results we will also take S to be a so-called
O(log n)-spanner of the graph. In this case, by definition of spanner, we have stretchS(e) = O(log n) for
all e; we will use directly these estimates without further computations.
In order to get a useful approximation via oversampling, we need the number of edges in the resulting
object to be significantly smaller than m. The intuition is that the lower we get stretchS(G) the fewer sam-
ples we need to take, since the sum of the overestimates of the probabilities defines the number of samples.
In [12] this intuition is coupled with a scaling-up technique, where instead of G we apply oversampling on
a graph H that is the same as G except that the weights of the edges of S are scaled up by a suitably
chosen factor κ. By doing that, we lose a factor of κ in the approximation guarantee but at the same time
we in fact lower the total stretch by a κ factor, and, for suitable κ, we will have have a small enough edge
count in the graph I we output. We call this graph the incremental spectral sparsifier.
We summarize the result in Theorem 3.6 below, which is an adaptation of a corresponding Theorem in
[13], coupled with the stronger probabilistic guarantee of success of Theorem 3.5, and slightly generalized
to handle general subgraphs rather than only spanning trees.
7
Theorem 3.6 Let G be a graph and S be a spanning subgraph of G. Assume that stretchS(e) is known
for each edge e of G. Then, there is an algorithm for constructing a graph I such that, given κ:
• H (cid:22) I (cid:22) 2H, where H = G + κS.
• I has n − 1 + O(stretchS(G) log n/κ) edges
assuming stretchS(G) is polynomially bounded. The algorithm succeeds with high probability and runs in
in O(m + stretchS(G) log n/κ) time.
4 The general case: An O(sf m log2 n) time algorithm
4.1 Estimating effective resistances
e = weRG(e). For
As we discussed above, Spielman and Srivastava [14] use the Sample algorithm with p′
the efficient implementation of their algorithm they first obtain a different expression for the effective
resistance, via a simple algebraic manipulation:
RG(i, j) = (χi − χj)T L+(χi − χj)
= (χi − χj)T L+LL+(χi − χj)
= (χi − χj)T L+BT W 1/2W 1/2BL+(χi − χj)
= kW 1/2BL+(χi − χj)k2
The advantage of this definition is that it expresses the effective resistance as the squared Euclidean distance
between two points, given by the ith and jth column of the matrix W 1/2BL+. This new expression still
involves the solution of a linear system with L. The natural idea is to replace L with an approximation
Lδ satisfying the properties described in Lemma 3.4. So instead of RG(i, j) we compute the quantities
RG(i, j) = kW 1/2B L+
Of course, there are still m systems to be solved. To work around this hurdle, Spielman and Srivastava
observe that projecting the vectors to an O(log n)-dimensional space preserves the Euclidean distances
within a factor of 1±ǫ/8, by the Johnson- Lindenstrauss theorem. Algebraically this amounts to computing
the quantities kQW 1/2B L+
δ (χi − χj)k2, where Q is a properly defined random matrix of dimension k × m
for k = O(log n). The authors invoke the result of Achlioptas [2], which states that one can use a matrix
Q each of whose entries is randomly chosen in {±1/√k}.
δ (χi − χj)k2.
The construction of the sparsifiers can can thus be broken up into three steps.
1. Compute QW 1/2B. This takes time O(km), since B has only two non-zero entries per row.
2. Apply the linear operator L+
δ to the k columns of the matrix (QW 1/2B)T , using Lemma 3.4. This
gives the matrix Z = QW 1/2B L+
δ .
3. Compute all the (approximate) effective resistances (time O(km)) via the square norm of the differ-
ences between columns of the matrix Z. Then sample the edges.
4.2 The O(sf m log2 n) time algorithm
Spielman and Srivastava prove that the approximations RG(i, j) can be used to obtain the sparsifier if they
satisfy
(1 − ǫ/4)RG(i, j) ≤ RG(i, j) ≤ (1 + ǫ/4)RG(i, j).
Then they show that this can be satisfied if δ, the accuracy guarantee of the linear system solver, is taken
to be an inverse polynomial in n. Thus their algorithm is dominated by the second step (the applications
of L+
δ ) and takes time O(tf m log n + sf m log3 n log(1/ǫ)).
8
The following lemma shows that in fact it is enough to take δ to be a constant. Furthermore, our proof
significantly simplifies the corresponding analysis of [14].
Lemma 4.1 For a given ǫ, if L satisfies (1 − δ)L (cid:22) L (cid:22) (1 + δ)L where δ = ǫ/8, then the approximate
effective resistance values RG(u, v) = kW 1/2B L+(χu − χv)k2 satisfy:
(1 − ǫ)RG(u, v) ≤ RG(u, v) ≤ (1 + ǫ)RG(u, v).
Proof. We only show the first half of the inequality, as the other half follows similarly. Since L and
L have the same null space, by (3.1) the given condition is equivalent to:
Since
1
1+δ L+ (cid:22) L+, we have
1
1 + δ
L+ (cid:22) L+ (cid:22)
1
1 − δ
L.
RG(u, v) = (χu − χv)T L+(χu − χv)
≤ (1 + δ)(χu − χv)T L+(χu − χv)
= (1 + δ)(χu − χv)T L+ L L+(χu − χv).
Applying the fact that L (cid:22) (1 + δ)L to the vector L+(χu − χv) in turn gives:
RG(u, v) ≤ (1 + δ)2(χu − χv)T L+L L+(χu − χv)
= (1 + δ)2kW 1/2B L+(χu − χv)k2 = (1 + δ)2 RG(u, v)
The rest of the proof follows from 1
(1+δ)2 ≤ 1 − ǫ/4 by choice of δ.
(cid:3)
This proves our first theorem.
Theorem 4.2 There is a 1 ± ǫ sparsification algorithm that runs in O(tf m log n + sf m log2 n log(1/ǫ))
time.
5 The O(tf m log n) time algorithm
Informally, the oversampling Theorem 3.5 states that if we use estimates to the effective resistances, rather
than the true values, the Spielman-Srivastava scheme still works; but in order to produce the sparsifier we
have to compensate by taking more samples. We exploit this in our second Theorem.
Theorem 5.1 There is a (1± ǫ)-sparsification algorithm that runs in O(tf m log n + m log n log(1/ǫ)) time
and returns a sparsifier with O(sf n log3 n/ǫ2) edges. As a result, we can compute an (1± ǫ)-sparsifier with
O(n log n/ǫ2) edges in O(tf m log n + m log n log(1/ǫ) + s2
f n log5 n/ǫ2) time.
Proof. Given the input graph G we construct a spine-heavy graph H that O(sf log2 n)-approximates G.
The construction can be done in O(tf m log n) time, by Lemma 3.3. We then run the Spielman-Srivastava
scheme (Section 4) on H to approximate the effective resistances RH(i, j) within a factor of 1 ± ǫ. Step 2
of the Spielman-Srivastava scheme runs in O(m log n log(1/ǫ)) time on H, by Lemma 3.4. We adjust the
approximate effective resistances in H down by a factor of 1 + ǫ to accommodate for the upper side of the
error in Lemma 4.1. Then, by (3.2) the calculated approximate effective resistances satisfy
1
O(sf log2 n)
RG(i, j) ≤ RH(i, j) ≤ RG(i, j).
9
So Theorem 3.5 applies if we take p′
e = O(sf log2 n)we RH(i, j). We have
X
e
e = O(sf log2 n) X
p′
e
we RH(i, j) ≤ O(sf log2 n) X
e
weRG(i, j) = O(sf n log2 n).
The last equality follows from the fact that Pe weRG(i, j) = n−1 for any graph G (e.g. see [14]). Hence the
total number of samples we need to take in order to produce an (1 ± ǫ)-sparsifier is O(sf n log3 n/ǫ2). The
second sparsifier is computed by re-sparsifying with the general case algorithm (and appropriate settings
for ǫ).
(cid:3)
6 The fastest algorithms
6.1 A near-linear stretch tree in O(m) time
The first problem in trying to accelerate the algorithm of the previous Section lies in the computation of the
low-stretch tree; known algorithms for the task take time at least O(m log n). To work around this problem
we will trade-off stretch for time. In the remainder of this subsection we show that we can in O(m) time
f m log3 n).
produce a spanning tree with a slightly weaker stretch guarantee, namely stretchT (G) = O(s2
The construction goes through the computation of incremental cut sparsifiers, which may be of independent
interest.
A cut-based characterization of stretch. We start by giving a simple alternative characterization of
the stretch of a graph over a tree; for this we will need some notation. Let G = (V, EG) be a graph and
T = (V, ET ) be a spanning tree of G. Every edge e ∈ ET defines in the obvious way a partition of V into
two sets Ve and V − Ve. Indeed, removing the edge disconnects the tree, and the partition is formed by
the vertices in the two connected components; we arbitrarily let Ve be the vertices in one of them. Let
capG(Ve, V − Ve) be the total weight of the edges in G with endpoints in Ve and V − Ve. We have
stretchT (G) = X
e∈ET
w−1
e capG(Ve, V − Ve).
(6.5)
To see why, let us go back to the definition of stretchT (G) given in Section 3. We have
stretchT (G) = X
e′∈EG
we′ X
e∈pT (e)
w−1
e ,
where pT (e) is the unique path in T between the two endpoints of e. It is then clear that stretchT (G) is a
sum of terms of the form we′w−1
for e′ ∈ EG and e ∈ ET . Instead of grouping the terms with respect to
e′ ∈ EG as is customary in the above definition, we group them with respect to e ∈ ET . It can be seen that
for any fixed e ∈ ET , the term w−1
appears as a factor multiplying we′ for each edge e′ ∈ EG that has its
two endpoints in Ve and V − Ve, precisely when pT (e′) has to use e. This directly gives us the alternative
characterization in Equation 6.5.
e
e
Incremental cut sparsifiers. Inspired by the incremental spectral sparsifiers of [12], we give an analogous
construction of incremental cut sparsifiers, i.e. sparsifiers that approximately preserve cuts but are only
mildly sparser relative to the input graph. We will say that a graph H is a τ -cut approximation of G if for
all S ⊆ V we have
τ · capH(S, V − S) > capG(S, V − S).
We claim the following Lemma.
Lemma 6.1 There is an O(t2
The graph H can be computed in O(m) time with high probability.
f log3 n)-cut approximation H of a graph G with O(m/(tf log n) + n) edges.
10
The algorithm and its proof is based on Karger’s earlier work on cut sparsification [8]. Before we proceed
with it, we review necessary definitions and a Lemma from [8]. In the following context graphs are allowed
to have multiple edges.
A graph is k-connected if the value of each cut in G is at least k. A k-strong component is a k-connected
vertex induced subgraph of a graph. The strong connectivity of an edge e, denoted ke, is the maximum
value of k such that a k-strong component contains e. Finally, a graph G is said to be c-smooth if for every
edge e, ke ≥ cwe.
Now let G(p) be the subgraph of G resulting by keeping each edge of G with probability p.
Lemma 6.2 [8] Let G be a c-smooth graph. Let p = ρǫ/c where ρǫ = O(log n/ǫ2). Then with high
probability every cut in G(p) has a value in the range (1 ± ǫ) times its expectation (which is p times its
original value).
We now proceed with the algorithm and its proof.
Proof.
(of Lemma 6.1) We can assume without loss of generality that the edge weights in G are
integers. We first find a max-weight spanning tree T of G; since the weights are integer this can be done
in O(m) time. We then form an intermediate graph G′ by multiplying the weight of every edge in T by
⌈tf log2 n⌉; let T ′ be T with the scaled up weights.
Consider now an edge e in G′ which is not in T ′. Let pT ′(e) be the path connecting the endpoints of e
in T ′. Because T is a maximum-weight spanning tree in G every edge along pT (e) has weight at least we
in G. Therefore the subgraph of G′ induced by the vertices in pT ′(e) is (we⌈tf log2 n⌉)-strong. Therefore
the connectivity ke satisfies ke ≥ (⌈tf log2 n⌉)we.
In order to be able to apply Lemma 6.2 we will modify G′ to be a multigraph, by viewing each e ∈ T ′ as
⌈tf log2 n⌉ parallel edges of weight we. Under this definition, the connectivity of each such parallel edge e′
trivially satisfies ke′ ≥ (⌈tf log2 n⌉)we. The same holds for all other edges of G′ as shown above. It follows
that the multi-graph G′ is O(tf log2 n)-soft.
We can now apply Lemma 6.2, setting ǫ = 1/2 to form G(p) for p = O(1/(tf log n)). We get that
(2/3)G(p) is an O(tf log n)-cut approximation for G′. By an easy transitivity argument, the graph H =
f log3 n)-cut approximation of G. The claim about the number of edges
2G(p)/(3⌈tf log2 n⌉) is then an O(t2
of H follows by application of Chernoff’s inequality.
(cid:3)
Computing a low-stretch tree faster. We conclude this subsection with the main Lemma.
Lemma 6.3 Given a graph G, a spanning tree T such that stretchT (G) = O(t2
in O(m) time.
f m log3 n) can be computed
Proof. We first produce in O(m) time the graph H of Lemma 6.1. We then apply the low-stretch
algorithm of [1] on graph H to get a spanning tree T ′ of H; given the number of edges in H this step takes
O(m) time as well, while T ′ satisfies stretchT ′(H) = O(m). Notice now that the edge weights in H are by
definition smaller than those in G. Therefore, the corresponding tree T in G (i.e. the tree with the original
weights) satisfies
stretchT (H) = O(m).
Given the number of edges in H the low-stretch algorithm runs in O(m) time. Then using the definition
11
of cut approximation, and equation 6.5 we get that
stretchT (G) = X
e∈ET
w−1
e capG(Ve, V − Ve).
≤ O(t2
= O(t2
= O(t2
f log3 n) X
e∈ET
w−1
e capH (Ve, V − Ve)
f log3 n) · stretchT (H)
f m log3 n).
(cid:3)
6.2 Leveraging an approximate sparsifier
The purpose of this section is to show that if we are given an approximate sparsifier for a graph G we can
efficiently produce a sparsifier for G. So far we have been using only low-stretch subgraphs of G to get
approximations to the effective resistances in G. A key to our fastest algorithm is the realization that we
can find low-stretch trees that are not necessarily subgraphs of the given graph G; the total stretch will
actually be only a near-linear function of n. This is based on the following Lemma.
Lemma 6.4 Let H ′ (cid:22) H. Then for any tree T
stretchT (H ′) ≤ stretchT (H).
Proof. Let A+ denote the Moore-Penrose pseudoinverse of a matrix A and λi(A) denote the ith largest
eigenvalue of A. By Spielman and Woo [17] we know that for any graph G
stretchT (G) = trace(LGL+
T ) = X
i
λi(LGL+
T )
Because LG and LT have the same null space (the constant vector), we can write
λi(LGL+
T ) = λi(L+/2
T LGL+/2
T
).
Notice now that we have
H ′ (cid:22) H ⇒ (L+/2
T LH ′L+/2
T
) (cid:22) (L+/2
T LHL+/2
T
).
This follows easily by definition. It is also easy to prove (see for example [10], Ch. 6.1) that
A (cid:22) B ⇒ λi(A) ≤ λi(B).
Hence
(cid:3)
λi(L+/2
trace(L+/2
) ≤ λi(L+/2
T LH ′L+/2
) ≤ trace(L+/2
T LH ′L+/2
stretchT (H ′) ≤ stretchT (H).
T LHL+/2
) ⇒
T LHL+/2
T
T
T
T
) ⇒
We are now ready to prove the main Lemma in this subsection. To avoid confusion we will use mH to
denote the number of edges of a graph H.
12
Lemma 6.5 (Leveraging) Let H ′ be a κ-approximation of H. Suppose we are given H ′, a 4-approximation
of H ′ with O(n log n) edges. Then we can construct an (1 ± ǫ)-approximation of H with O(sf κn log3 n/ǫ2)
edges in O(mH + tf n log2 n) time. We can also construct am (1 ± ǫ)-sparsifier of H with O(n log n/ǫ2)
edges in O(mH + s2
f κn log5 n/ǫ2) time.
Proof. We compute a low-stretch spanning tree T of H ′
in O(tf n log2 n) time. Because H ′ has
O(n log n) edges we have
stretchT ( H ′) = O(sf n log2 n).
We then compute in O(mH) time the effective resistance RT (e) of each edge e of H over T . This can
be done in O(mH ) time using off-line LCA algorithms by Gabow and Tarjan [18, 6]. Since H ′ (cid:22) 4H ′ and
H ′ (cid:22) H we can apply (3.2) twice and get
By Rayleigh’s monotonicity theorem (e.g. see [12]) we get that:
RH(e) ≤ RH ′
(e) ≤ 4R
H ′
(e).
H ′
R
(e) ≤ RT (e).
e = 4weRT (e) in Theorem 3.5 allows us to sparsify H. To get a (1 ± ǫ)-approximation of
Hence setting p′
H, the number of samples we need to take is O(q log q/ǫ2) where q = Pe 4weRT (e) = 4stretchT (H). Since
H ′ (cid:22) H ′ and H (cid:22) κH ′, we apply Lemma 6.4 twice and we get that
stretchT (H) ≤ κ · stretchT (H ′) ≤ κ · stretchT ( H ′) = O(sf κn log2 n).
This proves the first claim. The second claim follows via picking the appropriate ǫ and re-sparsifying the
first sparsifier with the general case sparsification algorithm.
(cid:3)
6.3 The O(m) and O(m log log n) time algorithms
We are now ready to prove our main claims.
Theorem 6.6 Given a graph G we can compute an (1± ǫ)-spectral sparsifier of G with O(n log n/ǫ2) edges
in O(m + t2
f n log10 n/ǫ2) time.
f s3
Proof. Let T be the spanning tree of G provided by Lemma 6.3. Let H and I be the two graphs from
f sf log5 n). We use our algorithm from
the application of Theorem 3.6 on G with respect to T and κ = O(t2
Theorem 5.1 to sparsify I and by transitivity we get a 4-approximation H ′ of H with O(n log n) edges in
f sf log5 n)-approximation of G we can ‘leverage’ its sparsifier
O(m + s2
f n log10 n/ǫ2)
H ′ to compute a (1± ǫ)-sparsifier of G via Lemma 6.5 (with κ = O(t2
time.
f n log5 n) time. Because H is an O(t2
f sf log5 n)) in O(m + t2
f s3
(cid:3)
Theorem 6.7 Given an unweighted graph G we can compute an (1 ± ǫ)-spectral sparsifier of G with
O(n log n/ǫ2) edges in O(m + s3
f n log8 n/ǫ2) time.
Proof. Let S be an O(log n)-spanner of G, as discussed in Section 3.6. Such an S with O(n) edges can
be computed in O(m) time [7]. Let H and I be the two graphs from the application of Theorem 3.6 on G
with respect to S and κ = O(sf log3 n). We then use our algorithm from Theorem 5.1 to sparsify I and by
f n log5 n) time. Because
transitivity we get a 4-approximation H ′ of H with O(n log n) edges in O(m + s3
H is an O(sf log3 n)-approximation of G we can ‘leverage’ its sparsifier H ′ to compute a (1 ± ǫ)-sparsifier
of G via Lemma 6.5 (with κ = O(sf log3 n)) in O(m + s3
f n log8 n/ǫ2) time.
(cid:3)
13
Theorem 6.8 Given a graph G we can compute an (1± ǫ)-spectral sparsifier of G with O(n log n/ǫ2) edges
in O((m + s2
f n log5 n/ǫ2) log log n) time.
Proof. Let T be the spanning tree of G provided by Lemma 6.3. Let H = G + O(t2
f sf log5 n)T . We
construct a sequence of graphs,
H = H0, H1, . . . , Ht = G
f sf log5 n)T /2i, for i = 1 . . . , t − 1 with some appropriate t = O(log log n). Notice
where Hi = G + O(t2
that all graphs Hi have m edges, so this takes O(m log log n) time. For i = 0, . . . , t − 1, let Hi denote
It is clear that H0 can be computed in O(m) time.
a 4-approximation of Hi with O(n log n) edges.
Provided now that we have Hj we can apply Lemma 6.5 (with κ = 2, ǫ = 1/2 and a proper scaling of
the (1 ± ǫ)-sparsifier) to get a 4-approximation Hj+1 in O(m + s2
f n log5 n) time. From Ht−1 we produce
a (1 ± ǫ)-sparsifier of Ht = G again by Lemma 6.5, using the desired value for ǫ. Because we apply the
algorithm of Lemma 6.5 O(log log n) times, we get the claimed running time.
(cid:3)
7 Final Remarks
The original algorithm of Spielman and Teng remains the only known combinatorial sparsification algorithm
that does not rely on solving systems. Designing a spectral sparsification algorithm that does not depend
on a linear system solver and that outputs a very sparse graph with O(n log n) or O(n log2 n) edges is a
challenging open problem. Since achieving this may prove to be difficult, an alternate approach could be
algorithms that compute very sparse κ-approximations for small values of κ. Such algorithms could play
a significant role in the development of more practical SDD solvers.
8 Acknowledgments
We would like to thank Jonathan Kelner and Gary Miller for useful discussions and the anonymous referees
for carefully reading an earlier version of this paper. Ioannis Koutis is supported by NSF CAREER award
CCF-1149048. Alex Levin is supported by a National Science Foundation graduate fellowship. Richard
Peng was at Microsoft Research New England for part of this work and is supported by a Microsoft
Research Fellowship. He is also partially supported by the National Science Foundation under grant
number CCF-1018463.
References
[1] Ittai Abraham and Ofer Neiman. Using petal-decompositions to build a low stretch spanning tree. In
Howard J. Karloff and Toniann Pitassi, editors, STOC, pages 395–406. ACM, 2012.
[2] Dimitris Achlioptas. Database-friendly random projections. In PODS ’01: Proceedings of the Twenti-
eth ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, pages 274–281,
2001.
[3] Andr´as A. Bencz´ur and David R. Karger. Approximating s-t minimum cuts in O(n2) time. In STOC
’96: Proceedings of the Twenty-Eighth Annual ACM symposium on Theory of Computing, pages 47–55,
1996.
[4] Fan Chung. Random walks and local cuts in graphs. Linear Algebra and its applications, 423(1):22–32,
2007.
[5] Wai Shing Fung, Ramesh Hariharan, Nicholas J. A. Harvey, and Debmalya Panigrahi. A general
In STOC ’11: Proceedings of the 43rd ACM Symposium on
framework for graph sparsification.
Theory of Computing, pages 71–80, 2011.
14
[6] Harold N. Gabow and Robert Endre Tarjan. A linear-time algorithm for a special case of disjoint set
union. In Proceedings of the fifteenth annual ACM symposium on Theory of computing, STOC ’83,
pages 246–251, New York, NY, USA, 1983. ACM.
[7] Eran Halperin and Uri Zwick. Linear time deterministic algorithm for computing spanners for un-
weighted graphs. 1996. manuscript.
[8] David R. Karger. Better random sampling algorithms for flows in undirected graphs. In Proceed-
ings of the ninth annual ACM-SIAM symposium on Discrete algorithms, SODA ’98, pages 490–499,
Philadelphia, PA, USA, 1998. Society for Industrial and Applied Mathematics.
[9] JonathanA. Kelner and Alex Levin. Spectral sparsification in the semi-streaming setting. Theory of
Computing Systems, 53(2):243–262, 2013.
[10] Ioannis Koutis. Combinatorial and algebraic tools for optimal multilevel algorithms. PhD thesis,
Carnegie Mellon University, Pittsburgh, May 2007. CMU CS Tech Report CMU-CS-07-131.
[11] Ioannis Koutis, Alex Levin, and Richard Peng. Improved spectral sparsification and numerical algo-
rithms for sdd matrices. In 29th International Symposium on Theoretical Aspects of Computer Science,
STACS, pages 266–277, 2012.
[12] Ioannis Koutis, Gary L. Miller, and Richard Peng. Approaching optimality for solving SDD systems.
In FOCS ’10: Proceedings of the 51st Annual IEEE Symposium on Foundations of Computer Science,
2010.
[13] Ioannis Koutis, Gary L. Miller, and Richard Peng. A nearly-m log n solver for SDD linear systems. In
FOCS ’11: Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer Science,
2011.
[14] Daniel A. Spielman and Nikhil Srivastava. Graph sparsification by effective resistances. In STOC ’08:
Proceedings of the 40th Annual ACM symposium on Theory of Computing, pages 563–568, 2008.
[15] Daniel A. Spielman and Shang-Hua Teng. Nearly-linear time algorithms for graph partitioning, graph
sparsification, and solving linear systems. In STOC ’04: Proceedings of the 36th Annual ACM Sym-
posium on Theory of Computing, pages 81–90, 2004.
[16] Daniel A. Spielman and Shang-Hua Teng. Nearly-linear time algorithms for preconditioning and
solving symmetric, diagonally dominant linear systems. CoRR, abs/cs/0607105, 2006.
[17] Daniel A. Spielman and Jaeoh Woo. A note on preconditioning by low-stretch spanning trees. CoRR,
abs/0903.2816, 2009.
[18] Robert Endre Tarjan. Applications of path compression on balanced trees. J. ACM, 26(4):690–715,
1979.
15
|
1801.06216 | 1 | 1801 | 2018-01-18T19:57:39 | Degree-constrained 2-partitions of graphs | [
"cs.DS",
"math.CO"
] | A $(\delta\geq k_1,\delta\geq k_2)$-partition of a graph $G$ is a vertex-partition $(V_1,V_2)$ of $G$ satisfying that $\delta(G[V_i])\geq k_i$ for $i=1,2$. We determine, for all positive integers $k_1,k_2$, the complexity of deciding whether a given graph has a $(\delta\geq k_1,\delta\geq k_2)$-partition.
We also address the problem of finding a function $g(k_1,k_2)$ such that the $(\delta\geq k_1,\delta\geq k_2)$-partition problem is ${\cal
NP}$-complete for the class of graphs of minimum degree less than $g(k_1,k_2)$ and polynomial for all graphs with minimum degree at least $g(k_1,k_2)$. We prove that $g(1,k)=k$ for $k\ge 3$, that $g(2,2)=3$ and that $g(2,3)$, if it exists, has value 4 or 5. | cs.DS | cs |
Degree-constrained 2-partitions of graphs
J. Bang-Jensen∗
St´ephane Bessy†
July 27, 2021
Abstract
A (δ ≥ k1, δ ≥ k2)-partition of a graph G is a vertex-partition (V1, V2) of G satisfying that
δ(G[Vi]) ≥ ki for i = 1, 2. We determine, for all positive integers k1, k2, the complexity of deciding
whether a given graph has a (δ ≥ k1, δ ≥ k2)-partition.
We also address the problem of finding a function g(k1, k2) such that the (δ ≥ k1, δ ≥ k2)-partition
problem is NP-complete for the class of graphs of minimum degree less than g(k1, k2) and poly-
nomial for all graphs with minimum degree at least g(k1, k2). We prove that g(1, k) = k for k ≥ 3,
that g(2, 2) = 3 and that g(2, 3), if it exists, has value 4 or 5.
Keywords: NP-complete, polynomial, 2-partition, minimum degree.
Introduction
1
A 2-partition of a graph G is a partition of V (G) into two disjoint sets. Let P1, P2 be two graph
properties, then a (P1, P2)-partition of a graph G is a 2-partition (V1, V2) where V1 induces a graph
with property P1 and V2 a graph with property P2. For example a (δ ≥ k1, δ ≥ k2)-partition of a
graph G is a 2-partition (V1, V2) where δ(G[Vi]) ≥ ki, for i = 1, 2
There are many papers dealing with vertex-partition problems on (di)graphs. Examples (from a
long list) are [1, 2, 4, 6, 7, 8, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 24, 26, 27, 28, 29, 30, 31, 33, 34].
Examples of 2-partition problems are recognizing bipartite graphs (those having has a 2-partition into
two independent sets) and split graphs (those having a 2-partition into a clique and an independent
set) [15]. It is well known and easy to show that there are linear algorithms for deciding whether a
graph is bipartite, respectively, a split graph. It is an easy exercise to show that every graph G has a
2-partition (V1, V2) such that the degree of each vertex in G[Vi], i ∈ [2] is at most half of its original
degree. Furthermore such a partition can be found efficiently by a greedy algorithm.
In [16, 17]
and several other papers the opposite condition for a 2-partition was studied. Here we require the
that each vertex has at least half of its neighbours inside the set it belongs to in the partition. This
problem, known as the satisfactory partition problem, is NP-complete for general graphs [5].
A partition problem that has received particular attention is that of finding sufficient conditions
for a graph to possess a (δ ≥ k1, δ ≥ k2)-partition. Thomassen [31] proved the existence of a function
f (k1, k2) so that every graph of minimum degree at least f (k1, k2) has a (δ ≥ k1, δ ≥ k2)-partition.
He proved that f (k1, k2) ≤ 12 · max{k1, k2}. This was later improved by Hajnal [19] and Haggkvist
(see [31]). Thomassen [31, 32] asked whether it would hold that f (k1, k2) = k1 + k2 + 1 which would
be best possible because of the complete graph Kk1+k2+1. Stiebitz [29] proved that indeed we have
f (k1, k2) = k1 + k2 + 1. Since this result was published, several groups of researchers have tried to find
extra conditions on the graph that would allow for a smaller minimum degree requirement. Among
others the following results were obtained.
Theorem 1.1 [20] For all integers k1, k2 ≥ 1 every triangle-free graph G with δ(G) ≥ k1 + k2 has a
(δ ≥ k1, δ ≥ k2)-partition.
∗Department of Mathematics and Computer Science, University of Southern Denmark, Odense DK-5230, Denmark
(email:[email protected]). This work was done while the first author was visiting LIRMM, Universit´e de Montpellier,
France. Hospitality is gratefully acknowledged. The research of Bang-Jensen was supported by the Danish research
council under grant number 7014-00037B.
†LIRMM, Universit´e de Montpellier, France (email: [email protected]).
1
Theorem 1.2 [26] For all integers k1, k2 ≥ 2 every graph G with no 4-cycle and with δ(G) ≥ k1+k2−1
has a (δ ≥ k1, δ ≥ k2)-partition.
Theorem 1.3 [23]
• For all integers k1, k2 ≥ 1, except for K3, every graph G with no K4 − e and with δ(G) ≥ k1 + k2
has a (δ ≥ k1, δ ≥ k2)-partition.
• For all integers k1, k2 ≥ 2 every triangle-free graph G in which no two 4-cycles share an edge
and with δ(G) ≥ k1 + k2 − 1 has a (δ ≥ k1, δ ≥ k2)-partition.
The original proof that f (k1, k2) = k1 + k2 + 1 in [29] is not constructive and neither are those
of Theorems 1.2 and 1.3. In [7] Bazgan et al. gave a polynomial algorithm for constructing a (δ ≥
k1, δ ≥ k2)-partition of a graph with minimum degree at least k1 + k2 + 1 or at least k1 + k2 when the
input is triangle-free.
The main result of this paper is a full characterization of the complexity of the (δ ≥ k1, δ ≥ k2)-
partition problem.
Theorem 1.4 Let k1, k2 ≥ 1 with k1 ≤ k2 be integers. When k1 + k2 ≤ 3 it is polynomial to decide
whether a graph has a 2-partition (V1, V2) such that δ(G[Vi]) ≥ ki for i = 1, 2. For all other values of
k1, k2 it is NP-complete to decide the existence of such a partition.
A result in [13] implies that (δ ≥ 3, δ ≥ 3)-partition is NP-complete already for 4-regular graphs
and there are other results about the complexity of finding partitions with lower and/or upper bounds
on the degrees inside each partition, such as [5, 6, 16, 17, 34], but we did not find anything which
implies Theorem 1.4.
The result of Stiebitz [29] insures that if the minimum degree of the input graph is large enough,
at least k1 + k2 + 1, then the (δ ≥ k1, δ ≥ k2)-partition always exists. We conjecture that if this
minimum degree is large but less than k1 + k2 + 1 then the (δ ≥ k1, δ ≥ k2)-partition is not always
trivial but can be solved in polynomial time.
Conjecture 1.5 There exists a function g(k1, k2) so that for all 1 ≤ k1 ≤ k2 with k1 + k2 ≥ 3 the
(δ ≥ k1, δ ≥ k2)-partition problem is NP-complete for the class of graphs of minimum degree less than
g(k1, k2) and polynomial for all graphs with minimum degree at least g(k1, k2).
In the next section we introduce notions and tools that will be used later. In Section 3 we give
the proof of Theorem 1.4 and in Section 4 we provide some partial results concerning Conjecture 1.5.
In particular we prove that g(1, k) = k for k ≥ 3, that g(2, 2) = 3 and that g(2, 3), if it exists, has
value 4 or 5. Finally in Section 5 we address some other partition problems mainly dealing with
(edge-)connectivity in each part of the partition.
Notice that regarding the results that we establish in this paper the first open case of Conjecture 1.5
is the following problem.
Problem 1.6 What is the complexity of the (δ ≥ 2, δ ≥ 3)-partition problem for graphs of minimum
degree 4?
2 Notation, definitions and preliminary results
Notation is standard and follows [3, 12]. In this paper graphs have no parallel edges and no loops.
We use the shorthand notation [k] for the set {1, 2, . . . , k}.
2.1 Special Graphs
We first define some graphs that will be used frequently in our proofs to ensure that certain vertices
have a sufficiently high degree.
For all k ≥ 3 we let Xk,2 be the graph we obtain from Kk+1 by subdividing one edge by a vertex
x. Let X3,1 be obtained from X3,2 by adding a vertex x(cid:48) adjacent to the degree 2 vertex x of X3,2.
2
Let Y4,1 be the graph on 7 vertices which we obtain from a 5-wheel by adding a new edge e linking 2
non adjacent vertices of the outer cycle of the 5-wheel, a new vertex joined to the 3 vertices of this
outer cycle not incident to e and to another new vertex y. For k = 3 let Zk be the graph X3,1 that we
defined above and let z = x(cid:48). For k ≥ 4 let Zk be the graph that we obtain from Kk−2,k−1 by adding
a cycle on the k − 1 vertices of degree k − 2 and then adding a new vertex z adjacent to all the k − 2
vertices of degree k − 1. And finally let Wk be the graph we obtain from Kk+1 by deleting one edge
u(cid:48)v(cid:48) and the adding two new vertices u, v and the edges uu(cid:48), vv(cid:48).
All these graphs are depicted in Figure 1.
y
z
Y4,1
X3,2
x
x
k − 2 vertices
x(cid:48)
X3,1
u
u(cid:48)
v
v(cid:48)
Zk
Kk+1
Xk,2
Kk+1
Wk
k − 1 vertices
Figure 1: The graphs Y4,1, X3,2, X3,1, Zk, Xk,2 and Wk.
2.2 Connected instances of 3-SAT
For a given instance F of 3-SAT with clauses C1, C2, . . . , Cm and variables x1, . . . , xn, where each
variable xi occurs at least once as the literal xi and at least once as ¯xi, we define the bipartite graph
B(F) as the graph with vertex set {v1, ¯v1, v2, ¯v2, . . . , vn, ¯vn} ∪ {c1, c2, . . . , cm}, where the first set
corresponds to the literals and the second one to the clauses, and edge set containing an edge between
the vertex cj and each of the 3 vertices corresponding to the literals of Cj for every j ∈ [m]. We say
that F is a connected instance if B(F) is connected.
Lemma 2.1 3-SAT is NP-complete for instances F where B(F) is connected
Proof: Suppose B(F) has connected components X1, X2, . . . , Xk, where k ≥ 2. Fix a literal
vertex (cid:96)i ∈ Xi for each i ∈ [k], add a new variable y and k − 1 new clauses C(cid:48)
k−1 where
j = ((cid:96)j−1 ∨ y ∨ (cid:96)j), j ∈ [k − 1]. Let F(cid:48) be the new formula obtained by adding the variable y and the
C(cid:48)
(cid:5)
clauses C(cid:48)
By adding a few extra variables, if necessary, we can also obtain an equivalent connected instance
k−1. It is easy to check that F(cid:48) is equivalent to F and that B(F(cid:48)) is connected.
1, . . . , C(cid:48)
1, . . . , C(cid:48)
in which each literal occurs at least twice. We leave the easy details to the interested reader.
2.3 Ring graphs and 3-SAT
We first introduce an important class of graphs that will play a central role in our proofs. The
directed analogue of these graphs was used in [2, 4]. A ring graph is the graph that one obtains
by taking two or more copies of the complete bipartite graph on 4 vertices {a1, a2, b1, b2} and edges
{a1b1, a1b2, a2b1, a2b2} and joining these in a circular manner by adding a path Pi,1 from the vertex
3
bi,1 to ai+1,1 and a path Pi,2 from bi,2 to ai+1,2 where bi,1 is the ith copy of b1 etc and indices are
'modulo' n (bn+1,j = b1,j for j ∈ [2] etc). Our proofs are all reductions from NP-complete variants of
the 3-SAT problems. We call the copies of {a1b1, a1b2, a2b1, a2b2} switch vertices.
We start by showing how we can associate a ring graph to a given 3-SAT formula. Let F =
C1 ∧ C2 ∧ . . . ∧ Cm be an instance of 3-SAT consisting on m clauses C1, . . . , Cm over the same set
of n boolean variables x1, . . . , xn. Each clause Ci is of the form Ci = ((cid:96)i,1 ∨ (cid:96)i,2 ∨ (cid:96)i,3) where each
(cid:96)i,j belongs to {x1, x2, . . . , xn, ¯x1, ¯x2, . . . , ¯xn} and ¯xi is the negation of variable xi. By adding extra
clauses to obtain an equivalent formula, if necessary, we can ensure that every literal occurs at least
twice in F. We shall use this fact in one of our proofs.
For each variable xi the ordering of the clauses above induces an ordering of the occurrences of
xi, resp ¯xi, in the clauses. Let qi (resp. pi) denote the number of times xi (resp. ¯xi) occurs in the
clauses Let R(F) = (V, E) be the ring graph defined as follows. Its vertex set is
V = {a1,1, . . . an,1, a1,2 . . . , an,2} ∪ {b1,1, . . . , bn,1, b1,2, . . . , bn,2} ∪ n(cid:91)
• (cid:83)n
i=1{ai,1bi,1, ai,1bi,2, ai,2bi,1, ai,2bi,2}
Its edge set E consists of the following edges:
i=1
{vi,1, . . . , vi,qi, v(cid:48)
i,1, . . . , v(cid:48)
i,pi
}
• the edges of the paths Pi,1, Pi,2, i ∈ [n] where Pi,1 = bi,1vi,1 . . . vi,qiai+1,1 and
Pi,2 = bi,2v(cid:48)
i,1 . . . v(cid:48)
ai+1,2.
i,pi
For 1 ≤ j ≤ m, we associate the clause Cj = ((cid:96)j,1 ∨ (cid:96)j,2 ∨ (cid:96)j,3) with the set Oj consisting of three
vertices of R(F) representing the occurrences of the literals of Cj in F: if (cid:96)j,1 = xi for some i ∈ [n]
and this is the r'th occurrence of xi in the clauses, then Oj contains the vertex vi,r. If (cid:96)j,1 = ¯xi for
some i ∈ [n] and this is the r'th occurrence of ¯xi in the clauses, then Oj contains the vertex v(cid:48)
i,r.
The other two vertices of Oj are defined similarly. In our proofs below we will often add a vertex ci
adjacent to all the vertices of Oi for i ∈ [n]. An example is depicted in Figure 2.
b1,1
a1,1
v1,1
v4,2
v1,2
v(cid:48)
1,1
b1,2
a1,2
v(cid:48)
4,1
v4,1
a2,1
b2,1
a2,2
b2,2
v2,1
v(cid:48)
2,1
c1
c2
c4
c3
a3,2
b3,2
a3,1
b3,1
v(cid:48)
3,1
v(cid:48)
3,2
v(cid:48)
3,3
v3,1
b4,2
a4,2
b4,1
a4,1
Figure 2: The ring graph R(F) corresponding to the formula F = (x1 ∨ x4 ∨ ¯x3) ∧ (x1 ∨ ¯x2 ∨ ¯x3) ∧
(x2 ∨ ¯x3 ∨ x4) ∧ (¯x1 ∨ x3 ∨ ¯x4). The grey boxes contain the switch vertices, the white vertices are the
variable vertices and we added the clauses vertices c1, c2, c3 and c4 (these are not part of the ring
graph R(F)).
4
The following observation which forms the base of many of our proofs is easy to prove (for a proof
of result a very similar to this see [4]).
Theorem 2.2 Let F be a 3-SAT formula and let R(F) be the corresponding ring graph. Then R =
R(F) contains a cycle C which intersects all the sets O1, . . . , Om so that R − C is a cycle C(cid:48) if and
only if F is a 'Yes'-instance of 3-SAT.
3 Proof of Theorem 1.4
3.1 The case k1 + k2 ≤ 3
We start with a trivial observation.
Proposition 3.1 Every graph G with δ(G) ≥ 1 and at least 4 vertices has a (δ ≥ 1, δ ≥ 1)-partition
except if G is a star.
Proposition 3.2 There is a polynomial algorithm for testing whether a graph has a (δ ≥ 1, δ ≥ 2)-
partition
Proof: We try for every choice of adjacent vertices u, v whether there is a solution with u, v ∈ V1.
Clearly G is a 'yes'-instance if and only if at least one of these O(n2) attempts will succeed. Hence
by starting with V1 = {u, v} and then moving vertices with at most one neighbour in V − V1 to V1 we
either end with a good partition or V1 = V in which case no partition exists for that choice {u, v}. (cid:5)
3.2 The case k1 = 1 and k2 ≥ 3
The following variant of satisfiability, which we call ≤ 3-SAT(3), is known to be NP-complete: Given
a boolean CNF formula F consisting of clauses C1, C2, . . . , Cm over variables x1, x2, . . . , xn such that
each clause has 2 or 3 literals, no variable occurs in more than 3 clauses and no literal appears more
than twice; decide whether F can be satisfied.
As we could not find a proper reference for a proof that ≤ 3-SAT(3) is NP-complete, we give one
here as it is presented on pages 281-283 in a set of course notes1 by Prof. Yuh-Dauh Lyuu, National
Taiwan University:
Assume that F is an instance of 3-SAT in which the variable x occurs a total of r ≥ 4 times in
the formula (as x or ¯x) in clauses Ci1, . . . , Cir . Introduce new variables x1, . . . , xr and replace the
first occurrence of x (in Ci1) by x1 is x is not negated in Ci1 and otherwise replace it by ¯x1 in Ci1 .
Similarly we replace the occurrence of x in Cij , j ≥ 2 by xj or ¯xj. Finally we add the new clauses
(¯x1 ∨ x2) ∧ (¯x2 ∨ x3) ∧ . . . ∧ (¯xr ∨ x1). These clauses (which have size 2) will force all the variables
x1, . . . , xr to take the same value under any satisfying truth assignment. Repeating this replacement
for all variables of the original formula F we obtain an equivalent instance F(cid:48) of ≤ 3-SAT(3).
Below we will need another variant which we call ≤ 3-SAT(5) where clauses still have size 2 or 3
and each variable is allowed to occur at most 5 times and at most 3 times as the same literal. By
following the scheme above and for each original variable occurring at least 4 times adding r extra
clauses (x1 ∨ ¯x2) ∧ (x2 ∨ ¯x3) ∧ . . . ∧ (xr ∨ ¯x1) we obtain an equivalent instance F(cid:48)(cid:48) and because the
2r new clauses will form a cycle in the bipartite graph B(F(cid:48)(cid:48)) of F(cid:48)(cid:48) it is easy to see that F(cid:48)(cid:48) is
a connected instance of ≤ 3-SAT(5) if F is a connected instance of 3-SAT. Hence, by Lemma 2.1,
connected ≤ 3-SAT(5) is NP-complete.
Theorem 3.3 For all k ≥ 3 it is NP-complete to decide whether a graph has a (δ ≥ 1, δ ≥ k)-
partition.
Proof: We show how to reduce an instance of connected ≤ 3-SAT(5) to (δ ≥ 1, δ ≥ k)-partition
where k ∈ {3, 4} in polynomial time and then show how to extend the construction to higher values
of k. We start the construction for k = 3.
Below we will use several disjoint copies of the graphs X3,1 and X3,2 to achieve our construction.
Let F be a connected instance of ≤ 3-SAT(5) with clauses C1, . . . , Cm and variables x1, . . . xn. We
1https://www.csie.ntu.edu.tw/~lyuu/complexity/2008a/20080403.pdf
5
may assume that each of the 2n literals occur at least once in F (this follows from the fact that we
may assume this for any instance of normal 3-SAT and the reduction above to ≤ 3-SAT(5) preserves
this property). We will construct G = G(F) as follows:
• For each variable xi, i ∈ [n] we introduce three new vertices yi, vi, ¯vi and two the edges yivi, yi¯vi.
• For each i ∈ [n]: if the literal xi (¯xi) occurs precisely once in F, then we identify vi (¯vi) with
the vertex x in a private copy of X3,2. If xi (¯xi) occurs precisely twice, then we identify vi (¯vi)
with the vertex x(cid:48) in a private copy of X3,1.
• Now we add new vertices c1, . . . , cm, where ci corresponds to the clause Ci, i ∈ [m], and join
each cj by an edge to those (2 or 3) vertices from {v1, . . . , vn, ¯v1, . . . , ¯vn} which correspond to
its literals. If cj gets only two edges this way, we identify it with the vertex x(cid:48) in a private copy
of X3,1.
• Add 2m new vertices z1, z2, . . . , z2m and the edges of the 2m-cycle z1z2, z2z3, . . . , z2m−1z2m, z2mz1.
• Finally we add, for each j ∈ [m] the edges cjz2j−1, cjz2j.
We claim that G(F) has a (δ ≥ 1, δ ≥ 3)-partition if and only if F can be satisfied. Suppose first
that t is a satisfying truth assignment. Then it is easy to check that (V1, V2) is a good 2-partition
if we take V1 to be the union of {y1, . . . , yn} and the n vertices from {v1, . . . , vn, ¯v1, . . . , ¯vn} which
corresponds to the false literals. Note that if a vertex vi (¯vi) is in V2 then it will have degree 3 via
its private copy of one of the graphs X1, X2 or because the corresponding literal occurred 3 times in F.
Conversely assume that (V1, V2) is a (δ ≥ 1, δ ≥ 3)-partition. Then we claim that we must have
all the vertices z1, z2, . . . , z2m in V2: If one of these is in V1, then they all are as they have degree
exactly 3. Clearly we also have {y1, . . . , yn} ⊂ V1. However, by construction, the literal and clause
vertices all have degree 3 and they induce a connected graph (here we use that the instance F has a
connected bipartite graph B(F)). Thus all of these vertices must be in V2, but then each vertex yi
is isolated, contradiction. Hence all the vertices z1, z2, . . . , z2m are in V2 and this implies that all of
c1, . . . , cm are also in V2. The vertices y1, . . . , yn are in V1 and hence, for each i ∈ [n], at least one of
the vertices vi, ¯vi is also in V1. Now define a truth assignment a follows. For each i ∈ [n]: If both vi
and ¯vi are in V1, or vi is in V2 we put xi true; otherwise we put xi false. Since each cj must have a
neighbour from {v1, . . . , vn, ¯v1, . . . , ¯vn} in V2 this is a satisfying truth assignment.
To obtain the construction for k = 4 we replace each copy of X3,1 above by a copy of Y4,1, each
copy of X3,2 by a copy of X4,2 and identify each of the literal and clause vertices with the vertex y
in an extra private copy of Y4,1. Finally we identify each vertex zt, t ∈ [2m] with the vertex y in a
private copy of Y4,1. Now it is easy to see that we can complete the proof as we did for the case k = 3.
For all k ∈ {3 + 2a, 4 + 2aa ≥ 1} we can increase the degree of all literal, clause and zj vertices by
2a by identifying these with the x vertices of a private copies of Xk,2 and repeat the proof above. (cid:5)
Corollary 3.4 The (δ ≥ 1, δ ≥ k)-partition problem is NP-complete for graphs of minimum degree
k − 1.
Proof: Recall that in our proof above the vertices corresponding to clauses must always belong
to V2 in any good partition (V1, V2) hence if we connect each vertex yi, i ∈ [n] to c1, . . . , ck−3 by edges
we obtain a graph of minimum degree k − 1 which has a good partition if and only if F is satisfiable
(The vertices yi, i ∈ [n] must belong to V1 as they have degree k − 1).
(cid:5)
3.3 (δ ≥ k1, δ ≥ k2)-partition when 2 ≤ k1 ≤ k2
Theorem 3.5 For every choice of natural numbers 2 ≤ k1 ≤ k2 it is NP-complete to decide whether
a graph has a (δ ≥ k1, δ ≥ k2)-partition.
Proof: We show how to reduce 3-SAT to (δ ≥ k1, δ ≥ k2)-partition. Given an instance F of
3-SAT with clauses C1, C2, . . . , Cm and variables x1, . . . , xn we proceed as follows. Start from a copy
of the ring graph R = R(F) and then add the following:
6
• If k2 ≥ 3 we identify each vertex of R with the vertex z in a private copy of Zk2.
• For each i ∈ [n] add a new vertex ui and join it by edges to the vertices ai,1, ai,2. If k1 ≥ 3 we
identify ui with the vertex z in a private copy of Zk1.
• For each i ∈ [n] add a new vertex u(cid:48)
i and join it by edges to the vertices bi,1, bi,2. If k1 ≥ 3 we
identify u(cid:48)
i with the vertex z in a private copy of Zk1.
• For each j ∈ [m] we add a new vertex cj, identify cj with the vertex z in a private copy of Zk1
if k1 ≥ 3 and add three edges from cj to the three vertices of R which correspond to the literals
of Cm.
• Finally add a new vertex r and join this to all of the vertices in {u1, . . . , un, u(cid:48)
n, c1, . . . , cm}
via private copies of Wk1 by identifying the vertex u with r and v with the chosen vertex from
{u1, . . . , un, u(cid:48)
nc1, . . . , cm}.
1, . . . , u(cid:48)
1, . . . , u(cid:48)
We claim that the final graph G has a (δ ≥ k1, δ ≥ k2)-partition if and only if F is satisfiable.
First we make some observations about G:
• Every vertex v of R which is not a switch vertex has degree exactly k2 + 1 as it has degree 2 in
R(F), is adjacent to exactly one cj, j ∈ [m] and if k2 ≥ 3 then v has been identified with one
vertex z of a private copy of Zk2.
1, . . . , u(cid:48)
• Switch vertices all have degree exactly k2 + 2.
• The vertices u1, . . . , un, u(cid:48)
n have degree exactly k1 + 1.
• All vertices in copies of Za have degree exactly a when a ≥ 3.
• All vertices in copies of Wk1 have degree exactly k1
• All vertices cj, j ∈ [m] have degree k1 + 2.
• The vertex r has degree m + 2n which we may clearly assume is at least k1.
For convenience in writing, below we define Z2 to be the empty graph so that we can talk about Za's
without having to condition this on a being at least 3. Suppose first that F is satisfiable. By Theorem
2.2 this means that R(F) has a cycle C which intersects the neighbourhood of each cj, j ∈ [m] and so
that R− C is another cycle C(cid:48). Now we let V1 consist of the vertices of C, their corresponding private
copies of Zk2, all the vertices {u1, . . . , un, u(cid:48)
n, c1, . . . , cm} along with their private copies of Zk1
and finally the vertex r and the vertices of all copies of Wk1 that we used. Let V2 = V (G) − V1, that
is V2 contains the vertices of C(cid:48) and their private copies of Zk2 . It is easy to check that δ(G[V1]) ≥ k1
and that δ(G[V2]) ≥ k2 so (V1, V2) is a good partition. Now assume that G has a good 2-partition
n, c1, . . . , cm} via copies of
(V1, V2). The way we connected r to the vertices in {u1, . . . , un, u(cid:48)
Wk1 implies that these must belong to the same set Vi as r. If k1 < k2 this must be V1 and otherwise
we can rename the sets so that i = 1. Since each cj, j ∈ [m] has degree k1 + 2 at least one of the
vertices corresponding to a literal of Cj must belong to V1. Suppose that some vertex corresponding
a literal (cid:96) is in V1, then all the vertices of the path in R corresponding to that literal, including the
two end vertices which are switch vertices, must belong to V1. This follows from the fact that all
n, c1, . . . , cm} ⊂ V1.
these vertices have degree k2 + 1 and have a neighbour in {u1, . . . , un, u(cid:48)
Moreover if ai,j (resp. bi,j) belongs to V2 then, as ai,j (resp. bi,j) has degree k2 + 2 and ui (resp. u(cid:48)
i)
belongs to V1, at least one of the vertices of {bi,1, bi,2} (resp. {ai,1, ai,2}) belongs to V2. And as ui
i) belongs to V1, one of {ai,1, ai,2} (resp. {bi,1, bi,2}) must lie in V1. So since V2 is not empty
(resp. u(cid:48)
this implies that the restriction of V2 to R is a cycle consisting of paths Q1, . . . , Qn where Qi is either
the path Pi,1 or the path Pi,2. Hence R[V (R) ∩ V1] is a cycle intersecting each of the neighbourhoods
(cid:5)
of the vertices cj, j ∈ [m] and hence F is satisfiable by Theorem 2.2.
1, . . . , u(cid:48)
1, . . . , u(cid:48)
1, . . . , u(cid:48)
Combining the results of this section concludes the proof of Theorem 1.4.
7
4 Higher degrees
We study the borderline between polynomial and NP-complete instances of the partition problems.
That is, we try to see how close we can get to the bound k1 + k2 + 1 on the minimum degree and still
have an NP-complete instance. For k1 = 1 we can give the precise answer by combining Corollary
3.4 and the result below.
Proposition 4.1 There is a polynomial algorithm for checking whether a graph G of minimum degree
at least k has a (δ ≥ 1, δ ≥ k)-partition.
Proof: It suffices to see that we can test for a given edge uv of G whether there is a (δ ≥ 1, δ ≥ k)-
partition (V1, V2) with u, v ∈ V1. This is done by starting with V1 = {u, v} and then moving vertices
from V − V1 to V1 when these vertices do not have at least k neighbours in V − V1. Note that this
process preserves the invariant that δ(G[V1]) ≥ 1. Hence if the process terminates before V1 = V we
have found the desired partition and otherwise we proceed to the next choice for an edge to start
(cid:5)
from.
For the (δ ≥ 2, δ ≥ 2)-partition problem we can also give the precise borderline between polynomial
and NP-complete instances.
Proposition 4.2 There exists a polynomial algorithm for checking whether a given graph of minimum
degree at least 3 has a (δ ≥ 2, δ ≥ 2)-partition.
Proof: First test whether G has two disjoint cycles C1, C2. This can be done in polynomial time
[9, 25]. If no such pair exists G is a 'no'-instance, so assume that we found a pair of disjoint cycles
C1, C2. Now put the vertices of C1 in V1 and continue to move vertices of V − V1 − V (C2) to V1
if they have at least two neighbours in the current V1. When this process stops the remaining set
V2 = V − V1 induces a graph of minimum degree at least 2, since the vertices we did not move have
(cid:5)
at most one neighbour in V1.
We now proceed to partitions where 2 ≤ k1 ≤ k2 and try to raise the minimum degree above k1
to see whether we can still prove NP-completeness.
Theorem 4.3 For every a ≥ 3 it is NP-complete to decide whether a graph of minimum degree a + 1
has a (δ ≥ a, δ ≥ a)-partition.
Proof: We give the proof for a = 3 and then explain how to extend it to larger a. Let F be an
instance of 3-SAT with n variable and m clauses C1, . . . , Cm. Let R(cid:48) = R(cid:48)(F) be obtained from R(F)
by adding, for all i ∈ [n], an edge between all vertices at distance 2 in one of the paths Pi,1, Pi,2, i ∈ [n]
(that is, we replace each of these paths by their square). Now we construct the graph H = H(F)
starting from R(cid:48) as follows:
• For each j ∈ [m]: add two vertices cj,1, cj,2 and join them to the vertices in R(cid:48) which correspond
to the literals of Cj.
• add the vertices of a 2m -cycle y1y2 . . . y2my1.
• For each j ∈ [m] add the two edges y2j−1cj,1, y2j−1cj,2 and the two edges y2jcj,2, y2j, cj+1,1,
where cm+1,1 = c1,1.
The resulting graph H has minimum degree 4 and we claim that H has a (δ ≥ 3, δ ≥ 3)-partition
if and only if F is satisfiable, which we know, by Theorem 2.2 and the previous proofs, where we
used the same approach, is if and only if the vertex set of R (which is the same as that of R(cid:48)) can be
partitioned into two cycles C, C(cid:48) so that C contains a neighbour of each of the vertices cj,1, cj,2, j ∈ [m].
Again the proof is easy when F is satisfiable: Let C, C(cid:48) be as above and let V2 = V (C(cid:48)) and
V1 = V (H) − V2. It is easy to check that δ(H[Vi]) ≥ 3 for i = 1, 2, because C contains a neighbour
of each cj,1, cj,2, j ∈ [m] (and we assume that each literal appears at least twice in F to insure that
δ(H[V2]) ≥ 3). Suppose now that H has a (δ ≥ 3, δ ≥ 3)-partition (V1, V2). Since adjacent vertices in
{y1, y2, . . . , y2m} have degree 4 and share a neighbour they must all belong to the same set Vi, i ∈ [2]
8
and this set must also contain all the vertices cj,1, cj,2, j ∈ [m]. Without loss of generality we have
i = 1. Thus V2 is a subset of V (R(cid:48)). The vertices of R have degree at most 4 in R(cid:48) and the initial and
terminal vertex of each path Pi,1 or Pi,2 has degree 3. Using this is not difficult to see that if some
vertex of a path Pi,1 or Pi,2 is in V2 then all the vertices of that path and the two adjacent switch
vertices are in V2. If there is some i ∈ [n] so that both of the vertices ai,1, ai,2 or both of the vertices
bi,1, bi,2 are in in V2, then, using the observation we just made, all vertices of R(cid:48) would be in V2 which
is impossible. Similarly we can show that V1 cannot contain both of the vertices ai,1, ai,2 or both of
the vertices bi,1, bi,2 for some i ∈ [n]. Hence for each switch {ai,1, ai,2, bi,1, bi,2}, exactly one of the
vertices ai,1, ai,2 and exactly one of the vertices bi,1, bi,2 is in V2. Now we see that the vertices in V1
and V2 both induces a cycle in R and as the vertices cj,1, cj,2 have degree 2 outside R, the cycle in R
which is in V1 must contain a neighbour of each of cj,1, cj,2, j ∈ [m]. Hence, by Theorem 2.2, F is
satisfiable.
We obtain the result for higher values of a by induction where we just proved the base case a = 3
above. Assume we have already constructed Ha = Ha(F) with δ(Ha) ≥ a + 1 such that Ha has a
(δ ≥ a, δ ≥ a)-partition if and only if F is satisfiable. Construct Ha+1 from two copies of Ha by joining
copies of the same vertex by an edge. It is easy to check that Ha+1 has a (δ ≥ a+1, δ ≥ a+1)-partition
if and only if Ha has a (δ ≥ a, δ ≥ a)-partition.
(cid:5)
Theorem 4.4 Deciding whether a graph of minimum degree 3 has a (δ ≥ 2, δ ≥ 3)-partition is NP-
complete.
Proof: Let X = X(F) be the graph we obtain by starting from the ring graph R(F) and then
adding the following:
• Add the vertices of a tree T whose internal vertices have all degree 3 and which has 2m + 4n
leaves denoted by
u1,1, u1,2 . . . , un,1, un,2, u(cid:48)
ui,1, ui,2 and u(cid:48)
i,1, u(cid:48)
j ∈ [m]. Here the vertices cj,1, cj,2 correspond to the clause Cj
n,2, c1,1, c1,2, c2,1, c2,2, . . . , cm,1, cm,2, where the pairs
i,2 have the same parent in T for i ∈ [n] and so do each of the pairs cj,1, cj,2,
1,2, . . . , u(cid:48)
n,1, u(cid:48)
1,1, u(cid:48)
• Join each vertex cj,1, cj,2, j ∈ [m] to the 3 vertices in R which correspond to the literals which
correspond to Cj and add the edge cj,1cj,2.
• Add 4n new vertices wi,1, wi,2, w(cid:48)
• For each i ∈ [n] add the edges wi,1ai,1, wi,2ai,2, w(cid:48)
• For each i ∈ [n] join each of ui,1 and ui,2 by edges to the vertices wi,1, wi,2 and add the edge
i,2, i ∈ [n]. Add the edges wi,1wi,2, w(cid:48)
i,2, i ∈ [n].
i,1bi,1, w(cid:48)
i,1, w(cid:48)
i,1w(cid:48)
i,2bi,2.
ui,1ui,2.
i,2 and add the edge
• For each i ∈ [n] join each of u(cid:48)
i,1 and u(cid:48)
i,2 by edges to the vertices w(cid:48)
i,1, w(cid:48)
u(cid:48)
i,1u(cid:48)
i,2.
We first prove that in any (δ ≥ 2, δ ≥ 3)-partition (V1, V2) of X all the vertices of T must be in
the same set Vi. Note that we can not have cj,1 and cj,2 in different sets of the partition because then
one of their 3 neighbours in V (R) must be in both sets. By a similar argument, for each i ∈ [n] the
i,1 and u(cid:48)
vertices ui,1 and ui,2 must belong to the same set in the partition and the vertices u(cid:48)
i,2 must
belong to the same set in the partition. It is easy to check that this implies our claim for T .
If F is satisfiable, then by Theorem 2.2, we can find vertex disjoint cycles C, C(cid:48) in R such that C
contains a vertex corresponding to a literal of Cj for each j and V (R) = V (C) ∪ V (C(cid:48)). Now we let
V1 = V (C(cid:48)) and V2 = V (X) − V1. It is easy to check that this is a (δ ≥ 2, δ ≥ 3)-partition because
C must contain exactly one of the vertices ai,1, ai,2 and exactly one of the vertices bi,1, bi,2 for each
i ∈ [n].
Suppose now that (V1, V2) is a good partition of X. By the argument above we have V (T ) ⊂ Vi for
i = 1 or i = 2. This must be i = 2 since in the graph X − V (T ) all vertices except the switch vertices
have degree 2. Thus we have V (T ) ⊂ V2 and each of the vertices cj,1, cj,2, j ∈ [m] have a neighbour in
9
V (R) which is also in V2. As in earlier proofs it is easy to check that if some vertex of one of the paths
Pi,1, Pi,2 is in V2 then all vertices of that path are in V2. As in the proof of the previous theorem, we
now conclude that for each switch {ai,1, ai,2, bi,1, bi,2}, exactly one of the vertices ai,1, ai,2 and exactly
one of the vertices bi,1, bi,2 is in V2. Now we see that the vertices in V1 and V2 both induce a cycle in
R and as the vertices cj,1, cj,2 have degree 2 outside R, the cycle in R which is in V2 must contain a
neighbour of each of cj,1, cj,2, j ∈ [m]. Hence, by Theorem 2.2, F is satisfiable.
(cid:5)
Corollary 4.5 For every k ≥ 2 it is NP-complete to decide if a given graph with minimum degree at
least k + 1 has a (δ ≥ k, δ ≥ k + 1)-partition.
Proof: This follows by induction on k with Theorem 4.4 as the base case in the same way as we
(cid:5)
proved the last part of Theorem 4.3.
Proposition 4.6 There is a polynomial algorithm for deciding whether a given graph of minimum
degree 5 has a (δ ≥ 2, δ ≥ 3)-partition.
Proof: Let G have δ(G) ≥ 5. By Theorem 1.1 and the algorithmic version of this result from [7]
we may assume that G has a 3-cycle C. Denote its vertex set by {a, b, c}. If δ(G[V − V (C)] ≥ 3)
we are done as we can take V1 = V (C), so assume there is a vertex d which adjacent to all vertices
of C. Then {a, b, c, d} induce a K4.
If δ(G[V − {a, b, c, d}) ≥ 2 we can take V2 = {a, b, c, d}, so
we can assume that there is a vertex e which is adjacent to all 4 vertices in {a, b, c, d} and now
{a, b, c, d, e} induce a K5. Now if G[V − {a, b, c, d, e}] contains a cycle C(cid:48) we can conclude by starting
with V2 = {a, b, c, d, e} and adding vertices of V − V (C(cid:48)) − {a, b, c, d, e} to V2 as long as there is one
with at least 3 neighbours in V2. When the process stops (V − V2, V2) is a good partition. Hence we
can assume that G[V − {a, b, c, d, e}] is acyclic. If one connected component of G[V − {a, b, c, d, e}] is
non trivial with a spanning tree T , then two leaves u, v of T will share a neighbour in {a, b, c, d, e}.
Without loss of generality this is e and now the K4 induced by a, b, c, d and the cycle formed by e, u, v
and the path between u and v in T are disjoint and we can find a good partition as we did above.
Hence if we have not found the partition yet we must have that G[V −{a, b, c, d, e}] is an independent
set I, all of whose vertices are joined to all vertices in {a, b, c, d, e}. If I ≥ 2 it is easy to find a good
partition consisting of a 3-cycle on a, b and one vertex from I as V1 and the remaining vertices as V2.
(cid:5)
Finally if I = 1 there is no solution.
5 Further 2-partition problems
In [31] Thomassen proved that every graph G of connectivity at least k1 + k2 − 1 and minimum degree
at least 4k1 + 4k2 + 1 has a 2-partition (V1, V2) so that G[Vi] is ki-connected for i = 1, 2.
with prescribed lower bounds on the (edge-)connectivity of G[Vi], i ∈ [2].
It is natural to ask about the complexity of deciding whether a graph has a 2-partition (V1, V2)
We start with a simple observation.
Proposition 5.1 There exits a polynomial algorithm for deciding whether a given graph has a 2-
partition (V1, V2) such that G[V1] is connected and G[V2] is 2-edge-connected.
Proof: Suppose first that G is not 2-edge-connected. If G has more than two connected compo-
nents it is a 'no'-instance. If it has two components, it is a 'yes'-instance if and only if one of these
is 2-edge-connected. Hence we can assume that G is connected but not 2-edge-connected. Now it is
easy to see that there is a good partition if and only if the block-cutvertex tree of G has a nontrivial
block which is a leaf in the block-cutvertex tree. Thus assume below that G is 2-edge-connected. Now
consider an ear-decomposition (sometimes called a handle-decomposition) of G where we start from
an arbitrary cycle C. Let P be the last non-trivial ear that we add and let u, v be the end vertices of
P . Then V1 = V (P ) − {u, v} and V2 = V − V1 is a good partition.
(cid:5)
Perhaps a bit surprisingly, if we require just a bit more for the connected part, the problem becomes
Both NP-completeness proofs below use reductions from a given 3-SAT formula F so we only
NP-complete.
describe the necessary modifications of R(F).
10
Theorem 5.2 It is NP-complete to decide whether an undirected graph G = (V, E) has a vertex
partition (V1, V2) so that G[V1] is 2-edge-connected and G[V2] is connected and non-acyclic.
Proof: We add vertices and edges to R = R(F) as follows:
• For each clause Cj, j ∈ [m] we add a vertex cj and join it by three edges to the three literal
vertices of R corresponding to Cj (as we did in several proofs above).
1, c(cid:48)
j, j ∈ [m].
i, i ∈ [n]
i, i ∈ [n].
m and edges cjc(cid:48)
1, . . . , α(cid:48)
1, . . . , β(cid:48)
n and the edges αiai,1, αiai,2, αiα(cid:48)
n and the edges βibi,1, βibi,2, βiβ(cid:48)
• Add new vertices c(cid:48)
2, . . . , c(cid:48)
• Add new vertices α1, . . . , αn, α(cid:48)
• Add new vertices β1, . . . , βn, β(cid:48)
We claim that the resulting graph G has a vertex partition (V1, V2) such that G[V1] is 2-edge-
connected and G[V2] is connected and non-acyclic if and only if F is satisfiable. Note that, by
construction, for every good partition every vertex of G which is not in R must belong to V2.
In
particular if a path Pi,j contains a vertex of V2 then all the vertices of Pi,j are in V2. Since we want
G[V2] to be connected, the edges between αi, βi, i ∈ [n] and R imply that for every i ∈ [n] at most
one of the vertices ai,1, ai,2 and a most one of the vertices bi,1, bi,2 can belong to V1. This implies that
exactly one of ai,1, ai,2 and exactly one of the vertices bi,1, bi,2 belong to V1 as otherwise V1 would
be empty. Now it is easy to check that the desired partition exists if and only if R contains a cycle
C(cid:48) which uses precisely one of the paths Pi,1, Pi,2 for i ∈ [n] and avoids at least one literal vertex for
every clause of F. Thus, by Theorem 2.2, F is satisfiable if and only if G has a good partition.
(cid:5)
Since we can decide whether a graph has two vertex disjoint cycles in polynomial time [9, 25] the
following result, whose easy proof we leave to the interested reader, implies that it is polynomial to
decide whether a graph has a 2-partition into two connected and non-acyclic graphs.
Proposition 5.3 A graph G has a 2-partition (V1, V2) such that G[Vi] is connected and has a cycle
for i = 1, 2 if and only if G has a pair of disjoint cycles and either G is connected or it has exactly
two connected components, each of which contain a cycle.
Theorem 5.4 It is NP-complete to decide whether a graph G = (V, E) has a vertex partition (V1, V2)
so that each of G[V1] and G[V2] are 2-edge-connected.
Proof: Let F be a 3-SAT formula and let G = G(F) be the graph we constructed in the proof above.
Let G1 be the graph obtained by adding the following vertices and edges to G:
mqm
1 q1c(cid:48)(cid:48)
2 q2 . . . c(cid:48)(cid:48)
j , j ∈ [m], qj, j ∈ {0} ∪ [m] and γ
• add new vertices c(cid:48)(cid:48)
• add the edges c(cid:48)(cid:48)
j, j ∈ [m]
j c(cid:48)
• add the edges of the path q0c(cid:48)(cid:48)
• complete this path into a cycle W by adding the edges γq0, γqm
• add an edge from γ to all vertices in {α(cid:48)
We claim that G(cid:48) has a vertex-partition into two 2-edge-connected graphs if and only if F is
satisfiable. First observe that in any good partition (V1, V2) we must have all vertices of W inside V1
or V2. This follows from the fact that each qi has degree 2 so it needs both its neighbours in the same
set. Without loss of generality, W is a cycle in V2. After deleting the vertices of W we have exactly
the graph G in the proof of Theorem 5.2 above and it is easy to see that all vertices not in V (R) must
belong to V2 in any good partition. This implies that G(cid:48) has the desired vertex-partition if and only
2 ) so that G[V1] is 2-edge-connected and G[V (cid:48)
if the graph G has a partition (V1, V (cid:48)
2 ] is connected and
non-acyclic. This problem is NP-complete by Theorem 5.2 so the proof is complete.
(cid:5)
1, . . . , α(cid:48)
1, . . . , β(cid:48)
n, β(cid:48)
n}.
By inspecting the proof above it is not difficult to see that the following holds.
Theorem 5.5 It NP-complete to decide whether a graph has a 2-partition (V1, V2) such that each of
the graphs G[Vi], i = 1, 2 are 2-connected.
It may be worth while to try and extend the results of this section to higher (edge)-connectivities.
11
References
[1] N. Alon. Splitting digraphs. Combin. Prob. Comput., 15:933 -- 937, 2006.
[2] J. Bang-Jensen, N. Cohen, and F. Havet. Finding good 2-partitions of digraphs II. Enumerable
properties. Theor. Comput. Sci., 640:1 -- 19, 2016.
[3] J. Bang-Jensen and G. Gutin. Digraphs: Theory, Algorithms and Applications. Springer-Verlag,
London, 2nd edition, 2009.
[4] J. Bang-Jensen and F. Havet. Finding good 2-partitions of digraphs I. Hereditary properties.
Theor. Comput. Sci., 636:85 -- 94, 2016.
[5] C. Bazgan, Z. Tuza, and D.Vanderpooten. The satisfactory partition problem. Discrete Appl.
Math., 154(8):1236 -- 1245, 2006.
[6] C. Bazgan, Z. Tuza, and D. Vanderpooten. Degree-constrained decompositions of graphs:
Bounded treewidth and planarity. Theor. Comput. Sci., 355(3):389 -- 395, 2006.
[7] C. Bazgan, Z. Tuza, and D. Vanderpooten. Efficient algorithms for decomposing graphs under
degree constraints. Discrete Applied Mathematics, 155(8):979 -- 988, 2007.
[8] J. Bensmail. On the complexity of partitioning a graph into a few connected subgraphs. J.
Combin. Optim., 30:174 -- 187, 2015.
[9] B. Bollob´as. Extremal Graph Theory. Academic Press, London, 1978.
[10] M. Bonamy, K.K. Dabrowski, C. Feghali, M. Johnson, and D. Paulusma. Recognizing graphs
In 42nd International Symposium on Mathematical Foundations of
close to bipartite graphs.
Computer Science, MFCS 2017, August 21-25, 2017 - Aalborg, Denmark, pages 70:1 -- 70:14, 2017.
[11] M. Bonamy, K.K. Dabrowski, C. Feghali, M. Johnson, and D. Paulusma. Independent feedback
vertex sets for graphs of bounded diameter. Information Proc. Letters., 131:26 -- 32, 2018.
[12] J.A. Bondy and U.S.R. Murty. Graph Theory, volume 244 of Graduate Texts in Mathematics.
Springer-Verlag, Berlin, 2008.
[13] V. Chv´atal. Recognizing decomposable graphs. J. Graph Theory, 8(1):51 -- 53, 1984.
[14] M.E. Dyer and A.M. Frieze. On the complexity of partitioning graphs into connected subgraphs.
Discrete. Appl. Math., 10:139 -- 153, 1985.
[15] S. Foldes and P. Hammer. Split graphs. Congress. Numer., 19:311 -- 315, 1977.
[16] M. Gerber and D. Kobler. Algorithmic approach to the satisfactory graph partitioning problem.
European J. of Operation Research, 125:283 -- 291, 2000.
[17] M.U. Gerber and D. Kobler. Classes of graphs that can be partitioned to satisfy all their vertices.
Australasian J. of Combin., 29:201 -- 214, 2004.
[18] A. Grigoriev and R. Sitters. Connected feedback vertex set in planar graphs.
In WG 2009:
Graph Theoretical concepts in Computer Science, volume 5911 of Lect. Notes Comp. Sci., pages
143 -- 153. Springer Verlag, Berlin, 2009.
[19] P. Hajnal. Partition of graphs with condition on the connectivity and minimum degree. Combi-
natorica, 3:95 -- 99, 1983.
[20] A. Kaneko. On decomposition of triangle-free graphs under degree constraints. J. Graph Theory,
27(1):7 -- 9, 1998.
[21] D. Kuhn and D. Osthus. Partitions of graphs with high minimum degree or connectivity. J.
Combin. Theory Ser. B, 88:29 -- 43, 2003.
12
[22] H.-O. Le, V.B. Le, and H. Muller. Splitting a graph into disjoint induced paths or cycles. Discrete.
Appl. Math., 131:199 -- 212, 2003.
[23] M. Liu and B. Xu. On partitions of graphs under degree constraints. Disc. Appl. Math., 226:87 -- 93,
2017.
[24] M.H. Liu and B.G. Xu. Bipartition of graph under degree constraints. Science China Mathemat-
ics, 58:869 -- 874, 2015.
[25] L. Lovas´z. On graphs not containing independent circuits (in Hungarian). MatLapok, 16:289299,
1965.
[26] J. Ma and T. Yang. Decomposing C4-free graphs under degree constraints. arXiv:1706.07292,
2017.
[27] N. Misra, G. Philip, V. Raman, S. Saurabh, and S. Sikdar. FPT algorithms for connected feedback
vertex set. J. Combin. Optim., 24:131 -- 146, 2012.
[28] M. Stiebitz. Decomposition of graphs and digraphs. In KAM Series in Discrete Mathematics-
Combinatorics-Operations Research-Optimization 95-309. Charles University Prague, 1995.
[29] M. Stiebitz. Decomposing graphs under degree constraints. J. Graph Theory, 23:321 -- 324, 1996.
[30] H. Suzuki, N. Takahashi, and T. Nishizeki. A linear algorithm for bipartion of biconnected graphs.
Inform. Process. Lett., 33:227 -- 231, 1990.
[31] C. Thomassen. Graph decomposition with constraints on the connectivity and minimum degree.
J. Graph Theory, 7:165 -- 167, 1983.
[32] C. Thomassen. Paths, circuits and subdivisions. In Selected topics in graph theory Vol. 3, pages
97 -- 131. Academic Press, 1988.
[33] P. van't Hof, D. Paulusma, and G.J. Woeginger. Partitioning graphs into connected parts. Theor.
Comput. Sci., 410:4834 -- 4843, 2009.
[34] M. Xiao and H. Nagamochi. Complexity and kernels for bipartition into degree-bounded induced
graphs. Theor. Comput. Sci., 659:72 -- 82, 2017.
13
|
1906.10615 | 1 | 1906 | 2019-06-25T15:45:56 | Krivine diffusions attain the Goemans--Williamson approximation ratio | [
"cs.DS"
] | Answering a question of Abbasi-Zadeh, Bansal, Guruganesh, Nikolov, Schwartz and Singh (2018), we prove the existence of a slowed-down sticky Brownian motion whose induced rounding for MAXCUT attains the Goemans--Williamson approximation ratio. This is an especially simple particular case of the general rounding framework of Krivine diffusions that we investigate elsewhere. | cs.DS | cs |
KRIVINE DIFFUSIONS ATTAIN THE GOEMANS -- WILLIAMSON APPROXIMATION RATIO
RONEN ELDAN AND ASSAF NAOR
ABSTRACT. Answering a question of Abbasi-Zadeh, Bansal, Guruganesh, Nikolov, Schwartz and Singh (2018),
we prove the existence of a slowed-down sticky Brownian motion whose induced rounding for MAXCUT
attains the Goemans -- Williamson approximation ratio. This is an especially simple particular case of the
general rounding framework of Krivine diffusions that we investigate elsewhere.
Bibliographical comment. This text amounts to an extraction of a subsection from our preprint [EN14],
with a modicum of background details added so as to make it self-contained. [EN14] is an outgrowth
of a project that we have been pusuing over the past 5 years to understand a framework, which we call
Krivine diffusions, to consistently "round" high-dimensional vectors to signs, i.e., to elements of {−1, 1}.
We decided to post the present note separately, prior to the publication of the more extensive manu-
script [EN14] (at which time [EN14] will subsume the present note), because recently a special case of the
approach of [EN14] arose in [AZBG+18], and we realized that even this (much simpler) setting of [EN14]
resolves one of the main open questions of [AZBG+18]. The present extract from [EN14] establishes this
fact in a self-contained manner that can be read as a standalone article without waiting for the appear-
ance of [EN14], though [EN14] situates the ensuing reasoning within proper context and derives more
delicate statements in other settings of interest (notably for the Grothendieck constant, which entails
treating cancellations that do not arise in the present setting).
1. INTRODUCTION
Throughout what follows, we fix d ∈ N and let 〈·,·〉 : Rd → R be the standard scalar product on Rd , with
the associated unit sphere denoted by Sd−1 = {x ∈ Rd : 〈x, x〉 = 1}.
Suppose that n ∈ N and x1, . . . , xn ∈ Sd−1. It is a ubiquitous problem in several areas to meaningfully
associate to these vectors signs σ1, . . . , σn ∈ {−1, 1}. The interpretation of the term "meaningfully" here is
problem-specific, and it encompasses issues that may include the efficiency of determining these signs.
This is of central importance to approximation algorithms, when x1, . . . , xn are an output of a continuous
relaxation of a combinatorial optimization problem. Grothendieck famously broached [Gro53] this mat-
ter for (formidable) analytic applications (see also the work of Lindenstrauss and Pełczynski [LP68] and
the survey [Pis12]), and it occurs in quantum information theory; see e.g. the work of Tsirelson [Tsi85].
The above generic "rounding" problem has the following seemingly naïve solution: Simply output
the "first bit" (namely, the sign of the first entry) of each of these vectors, perhaps after first apply-
ing a suitable transformation, such as a uniformly random rotation as in the work of Goemans and
Williamson [GW95] or a more sophisticated "preprocessing" as in the work of Krivine [Kri77]. Despite
its simplistic appearance, this procedure is of major importance to algorithm design, as demonstrated
decisively in the seminal work [GW95] and in numerous subsequent investigations.
Superior performance of the above rounding task can sometimes be achieved by associating a sign
to a vector in a manner that is more sophisticated than considering its first bit after preprocessing. This
may be somewhat nonintuitive at times, e.g. Krivine [Kri77] stated that this is not so for the Grothendieck
R. E. is supported by a European Reseach Council Starting Grant (ERC StG) and by the Israel Science Foundation.
A. N. was supported by the Packard Foundation and the Simons Foundation.
The research presented here was conducted under the auspices of the Simons Algorithms and Geometry (A&G) Think Tank.
1
problem (to be described below), but [BMMN13] established that Krivine's conjecture has a negative an-
swer (note that despite the fact that for historical reasons this is formally described as a "negative answer"
to a conjecture, it entails the design of a better rounding method, so it is in fact a positive result). Such
higher-dimensional "Krivine rounding schemes" are shown to be optimal for the Grothendieck problem
in [NR14], and this is the reason for the nomenclature that we choose below to describe a new rounding
method. There is a remarkable generic procedure [Rag08, RS09a, RS09b, Aus10] to devise for certain con-
straint satisfaction problems (CSPs) a higher-dimensional rounding method that has the best possible
polynomial-time performance unless the Unique Games Conjecture [Kho02] has a negative answer.
Despite the overall usefulness of the above "first bit rounding" (a.k.a. hyperplane rounding, after pre-
processing), as well as its optimality in important settings such as MAXCUT [FS02, KKMO07], combined
with the fact that sharp rounding methods have been proven to exist for a large family of CSPs, there
is great need to devise further rounding methods (and ways to analyse them) because there are im-
portant continuous relaxations of combinatorial optimization problems for which the aforementioned
results do not shed light on the best-possible value of the associated integrality gap; understanding the
Grothendieck constant is a prime example of a longstanding mystery of this nature. Of course, there is no
reason to expect that such constants could be expressed in terms of elementary functions, but one might
hope to find enlightening characterizations of them, better estimates, or to "merely" obtain a versatile
rounding method that could be used heuristically in practice. More ways to reason about the rounding
problem are needed also because the aforementioned generic rounding methods for CSPs are sharp (as-
suming the Unique Games Conjecture) only when the integrality gap is O(1) as n → ∞, while there are
important questions (e.g. Sparset Cut) where the integrality gap is unbounded as a function of n, so the
pertinent issue becomes to devise methods that could be analysed so as to yield sharp asymptotics.
Our preprint [EN14] considers the following general rounding method that we call Krivine diffusions,
in reference to the power of Krivine rounding schemes that was established in [BMMN13, NR14]. Note,
however, that while the works [BMMN13, NR14] build on a key idea that Krivine introduced in [Kri77],
their success relies on the usefulness of degrees of freedom that Krivine expected [Kri77] not to be useful.
Nonetheless, Krivine's approach is the fundamental inspiration for these ideas, thus justifying naming
the method in homage to his contribution. The ensuing framework allows one to bring in tools from
stochastic calculus to analyse the rounding question, and in [EN14] this is shown to relate integrality gaps
to certain variational questions in partial differential equations. The present text, however, demonstrates
how in the special setting of MAXCUT, an explicit sharp closed-form solution could be obtained.
1.1. Krivine diffusions. Consider the following method to associate to u ∈ Rd a random sign σu ∈ {−1, 1}.
We call any such procedure a discrete-time Krivine diffusion. We will pass later to a natural continuous-
time limit in order to use analytic tools to understand the behavior of the rounding procedure.
• Datum. Each instance of the method is determined by an integer T ∈ N, a function f : R → R and
for every t ∈ {1, . . . ,T } a function Ft : Rt → R.
• Randomness. Let γ1, . . . γT be independent standard Gaussian random variables in Rd , i.e., their
density at (x1, . . . , xd ) ∈ Rd is proportional to exp(−(x2
• Evolution. For every u ∈ Rd consider the random numbers Xu(0), . . . , Xu(T ) ∈ R that are defined
by setting Xu(0) = 0, and proceeding recursively by setting
1 + . . .+ x2
d )/2).
(1)
def
Xu(t + 1)
= Xu(t )+ Ft¡Xu(1), . . . , Xu(t )¢f ¡〈γt ,u〉¢.
• Output. For each u ∈ Rd , if there is t ∈ {1, . . . ,T } with Xu(t ) Ê 1, then let Tu be the first time that
this occurs, i.e., Tu = min{t ∈ {1, . . . ,T } : Xu(t ) Ê 1}, and output σu = sign(Xu(Tu)). Otherwise,
namely if Xu(t ) < 1 for all t ∈ {1, . . . ,T }, then let σu be uniformly distributed over {−1, 1}.
In other words, the datum of a Krivine diffusion consists of the recursion coefficients in (1) and the
number T of recursion steps. With this given, one chooses T independent Gaussians and evolves the
random numbers {Xu(t )}T
t=0 according to the update rule (1) in which the increment Xu(t + 1)− Xu(t ) is
2
proportional to a function of the random dot-product 〈γt ,u〉, where the proportion factor may depend
on the values Xu(1), . . . , Xu(t ) that were determined thus far. The output σu ∈ {−1, 1} is the sign of this
process at the first time that it exists the interval (−1, 1) if such a time exists, and otherwise it is a uni-
formly random element of {−1, 1}. Note that the same Gaussians γ1, . . . , γT are used here for every u ∈ Rd ,
thus inducing an important and subtle correlation between the signs {σu}u∈Rd .
There are degrees of freedom that the above construction did not exploit. One obvious variant is to
allow the recursion coefficient f (〈γt ,u〉) in (1) to have a further dependence on t of the form ft (〈γt ,u〉)
for some ft : R → R. We dropped this flexibility because thus far we did not find any use for it.
A more substantial variant is to consider distributions over families of Krivine diffusions. This will not
be needed in the present note, but it is worthwhile to mention in passing that distributions over pairs of
Krivine diffusions are used in the following theorem from [EN14].
Theorem 1 (Krivine diffusions attain the Grothendieck constant). For any ε > 0, there exists T ∈ N and
a distribution µ over pairs of Krivine diffusions {(Xu(t ),Yu(t ))}T
t=0 such that for every m,n ∈ N, every
u1, . . . ,um, v1, . . . , vn ∈ Sd−1 and every m by n matrix A = (ai j ) ∈ Mm×n(R),
n
Xi=1
n
Xj=1
ai j〈ui , v j〉 É (KG + ε)
n
Xi=1
n
Xj=1
ai j E£sign¡Xu(T )¢sign¡Yu(T )¢¤+ ε
n
Xi=1
n
Xj=1ai j,
(2)
where KG is the Grothendieck constant and the expectation in (2) is with respect to both µ and the under-
lying randomness of the Krivine diffusions, namely over the Gaussians γ1, . . . , γT .
A much more well-understood rounding problem concerns the special case of (2) in which the coeffi-
cients {ai j : (i , j ) ∈ {1, . . . ,n}2} are non-positive. An especially important instance of this is the MAXCUT
problem; the lack of cancellations between the summands in the setting of MAXCUT allows for the suc-
cess of understanding it fully assuming the Unique Games Conjecture [GW95, KKMO07].
Abbasi-Zadeh, Bansal, Guruganesh, Nikolov, Schwartz and Singh studied in [AZBG+18] a special case
of Krivine diffusions (which they discovered independently of [EN14]), called sticky Brownian motion
(and a "slowed down" version thereof ). It is shown in [AZBG+18] that the integrality gap for MAXCUT
that can be attained by sticky Brownian motion is quite close to the Goemans -- Williamson approximation
ratio [GW95], relying on analytic reasoning and computer-assisted numerical computations. One of the
main open questions that were posed in [AZBG+18] is whether the Goemans -- Williamson approximation
ratio can be achieved exactly using this approach. Here we prove that this is indeed the case.
1.2. The continuous setting. We begin by recalling the slowed-down sticky Brownian motion rounding
scheme of [AZBG+18]. Its input is a sequence of unit vectors u1, . . . ,un ∈ Rd and its goal is to produce
signs σ1, . . . , σn ∈ {−1, 1}. To do so, we will use below basic facts from stochastic calculus, all of which are
standard and appear in introductory texts such as [Ok03, KS91]. Let {B (t )}tÊ0 be a Brownian motion in
Rd adapted to the filtration {Ft }tÊ0. Sticky Brownian motion refers to the associated random signs
∀i ∈ {1, . . . ,n},
σi
def
= 〈ui ,B (Ti )〉,
where
Ti
def
= min{t Ê 0 : 〈ui ,B (t )〉= 1}.
Thus, the sign σi is extracted when the stochastic process {〈ui ,B (t )〉}tÊ0 first reaches the set {−1, 1}; hence
the terminology "sticky." Next, consider the following modification of this procedure. Fix a continuous
function ϕ : [−1, 1] → [0,∞) that satisfies
lim
s→1−
ϕ(s) = lim
s→−1+
ϕ(s) = 0
and
∀ s ∈ (−1, 1),
ϕ(s) > 0.
Let {W ϕ
u (t )}(u,t )∈Rd×[0,∞) be a stochastic process that satisfies the stochastic differential equation
∀(u, t ) ∈ Rd × [0,∞),
W ϕ
u (0) = 0,
and
3
dW ϕ
u (t ) = ϕ¡W ϕ
u (t )¢〈u, dB (t )〉.
(3)
(4)
The existence and uniqueness (in distribution) of {W ϕ
u (t )}(u,t )∈Rd×[0,∞) follows from the rudimentary the-
ory of stochastic differential equations; thorough treatments of this appear in e.g. [KS91, Ok03]. In par-
ticular, {W ϕ
u (t )}tÊ0 is a martingale (with respect to the Brownian filtration {Ft }tÊ0) for every u ∈ Rd , and
by the martingale convergence theorem the following limit exists almost-surely.
ϕ
u
σ
def
= lim
t→∞
W ϕ
u (t ).
ϕ
u ∈ {−1, 1} almost-surely. Since the speed ϕ satisfies ϕ(s) → 0 as
The assumptions (3) on ϕ ensure that σ
s → {±1}, this process is called "slowed-down" sticky Brownian motion.
Remark 2. For each fixed u ∈ Rd the equation (4) that governs the evolution of the process {W ϕ
u (t )}τÊ0 is
a one-dimensional stochastic differential equation. One can solve (4) separately for each u ∈ Rd , but it is
motion that is used to drive this family of stochastic differential equations is the same for all u ∈ Rd , thus
inducing important correlations which are exploited crucially both below and in [EN14, AZBG+18].
important to stress that the resulting rounding procedure relies on the fact that the underlying Brownian
A mechanical approximation argument verifies that the above processes can be obtained as continu-
ous limits of Krivine diffusions; we will not include the straightforward details here (they are meaningful
for understanding the context, and also a discretization is needed for efficient implementation of the
rounding method, but such issues are not needed for the ensuing proof ).
The work [AZBG+18] makes the choice ϕ(s) = (1− s2)α for α > 0. Here we pinpoint the following dif-
ferent choice which recovers the Goemans -- Williamson approximation ratio and thus answers positively
the question raised in [AZBG+18]. Define ξ : [−1, 1] → [0,∞) by setting
2 ¢2
Φ−1¡ 1−s
∀ s ∈ [−1, 1],
p2
pπ
e− 1
ξ(s)
(5)
=
def
,
2
where Φ : R → R is the standard Gaussian cumulative distribution function, i.e.,
∀ x ∈ R,
Φ(x)
def
=
e− s2
2 ds,
thus
Φ′′(x) = −xΦ′(x).
(6)
1
p2π x
∞
With the following theorem at hand, one concludes identically to [GW95] that the slowed-down sticky
Brownian motion with speed ξ attains the Goemans -- Williamson approximation ratio for MAXCUT.
Theorem 3. For u ∈ Rd and t Ê 0 write W ξ
∀u, v ∈ Sd−1,
ξ
u = σu ∈ {−1, 1}, where ξ is given in (5). Then,
u (t ) = Wu(t ) and σ
2
arcsin(〈u, v〉).
E£σu σv¤ =
(7)
π
2. PROOF OF THEOREM 3
Let {B (t )}tÊ0 be a Brownian motion in Rd adapted to the filtration {Ft }tÊ0. Consider the process
∀(x, t ) ∈ Rd × [0,∞),
Zx (t )
def
= t
0
e− s
2 〈x, dB (s)〉.
0 e−sd s implies that the following limit exists almost-surely.
The convergence of the integral´ ∞
Zx (∞)
Zw (t ) = ∞
Fix u, v ∈ Sd−1. Because u and v are unit vectors and´ ∞
and Zv (∞) are standard Gaussian random variables, and their covariance equals
2〈x, dB (s)〉.
= lim
t→∞
e− s
def
0
0 e−s ds = 1, it follows from (9) that both Zu(∞)
E£Zu(∞)Zv (∞)¤ (9)
= ∞
0
e−s d[〈u,B (s)〉〈v,B (s)〉]s = ∞
0
e−s〈u, v〉 ds = 〈u, v〉.
(8)
(9)
4
Hence, by the original computation of Goemans and Wiliamson [GW95] (using elementary planar ge-
ometry), we have the following equality (Grothendieck's identity [Gro53, LP68]) .
E£sign¡Zu(∞)¢sign¡Zv (∞)¢¤ =
2
π
arcsin(〈u, v〉).
(10)
Due to (10), Theorem 3 will be proven if we show that (σu , σv ) and (sign(Zu(∞)), sign(Zv (∞))) have the
same distribution. More generally, by working with continuous modifications of {Wx (t )}(x,t )∈Sd−1×[0,∞)
and {Zx (t )}(x,t )∈Sd−1×[0,∞) (as we may, see e.g. [Kun84]), the ensuing argument establishes that the sign-
valued processes {σx }x∈Sd−1 and {sign(Zx (∞))}x∈Sd−1 are equal almost-surely.
For every u ∈ Rd , consider the martingale that is defined by
By the martingale convergence theorem, we almost-surely have
∀ t Ê 0,
Mu(t )
def
= E£sign¡Zu(∞)¢¯¯
Ft¤ .
(11)
Mu(t )
lim
t→∞
= E£sign¡Zu(∞)¢¯¯
F∞¤ = sign¡Zu(∞)¢.
We will show below that the stochastic processes {Mu(t )}u∈Sd−1×[0,∞) and {Wu(t )}u∈Sd−1×[0,∞) are equal.
Since, by definition, σu = limt→∞ Wu(t ) almost-surely, together with (12) this would imply the desired
equi-distribution of (σu, σv ) and (sign(Zu(∞)), sign(Zv (∞))).
Note that´ ∞t e−s/2〈u, dB (s)〉 is independent of Zu(t ) and has the same distribution as e−t /2
γ, where γ
is a standard Gaussian random variable that is independent of {Zu(t )}(u,t )∈Rd×[0,∞]. Hence,
2 Zu(t )¢.
Ft¤− 1 = 2P£Zu(t )+ e− t
2 γ > 0¤− 1 = 1− 2Φ¡− e
Mu(t )
(13)
(11)
t
Itô's formula yields,
(11)
(12)
(14)
= 2P£Zu(∞) Ê 0¯¯
2 Zu(t )¢e
t
dMu(t ) = 2Φ′¡− e
(6)∧(13)
t
2 dZu(t )− Φ′′¡− e
t
= 2Φ′µΦ−1³ 1− Mu(t )
2
´¶ e
t
2 dZu(t )
t
2 Zu(t )¢ dt + e
2 Zu(t )Φ′¡− e
(5)∧(8)
= ξ¡Mu(t )¢〈u, dB (t )〉.
t
2 Zu(t )¢ dt
Recalling that {Wu (t )}u∈Sd−1×[0,∞) satisfies the stochastic differential equation (4) with ϕ = ξ, we see
from (14) that the processes {Mu(t )}u∈Sd−1×[0,∞) and {Wu (t )}u∈Sd−1×[0,∞) satisfy the same stochastic dif-
ferential equation, so by Itô's uniqueness theorem Wu(t ) = Mu(t ) for all t ∈ [0,∞) and all u ∈ Sd−1.
(cid:3)
REFERENCES
[Aus10]
[AZBG+18] S. Abbasi-Zadeh, N. Bansal, G. Guruganesh, A. Nikolov, R. Schwartz, and M. Singh. Sticky brownian rounding and
P. Austrin. Towards sharp inapproximability for any 2-CSP. SIAM J. Comput., 39(6):2430 -- 2463, 2010.
its applications to constraint satisfaction problems, 2018. ArXiv:1812.07769.
[BMMN13] M. Braverman, K. Makarychev, Y. Makarychev, and A. Naor. The Grothendieck constant is strictly smaller than
[EN14]
[FS02]
[Gro53]
[GW95]
[Kho02]
Krivine's bound. Forum Math. Pi, 1:e4, 42, 2013.
R. Eldan and A. Naor. Krivine diffusions, 2014. Preprint.
U. Feige and G. Schechtman. On the optimality of the random hyperplane rounding technique for MAX CUT. Ran-
dom Structures Algorithms, 20(3):403 -- 440, 2002. Probabilistic methods in combinatorial optimization.
A. Grothendieck. Résumé de la théorie métrique des produits tensoriels topologiques. Bol. Soc. Mat. São Paulo,
8:1 -- 79, 1953.
M. X. Goemans and D. P. Williamson. Improved approximation algorithms for maximum cut and satisfiability prob-
lems using semidefinite programming. J. Assoc. Comput. Mach., 42(6):1115 -- 1145, 1995.
S. Khot. On the power of unique 2-prover 1-round games. In Proceedings of the Thirty-Fourth Annual ACM Sympo-
sium on Theory of Computing, pages 767 -- 775. ACM, New York, 2002. doi:10.1145/509907.510017.
[KKMO07] S. Khot, G. Kindler, E. Mossel, and R. O'Donnell. Optimal inapproximability results for MAX-CUT and other 2-
[Kri77]
[KS91]
variable CSPs? SIAM J. Comput., 37(1):319 -- 357, 2007.
J.-L. Krivine. Sur la constante de Grothendieck. C. R. Acad. Sci. Paris Sér. A-B, 284(8):A445 -- A446, 1977.
I. Karatzas and S. E. Shreve. Brownian motion and stochastic calculus, volume 113 of Graduate Texts in Mathematics.
Springer-Verlag, New York, second edition, 1991. ISBN 0-387-97655-8. doi:10.1007/978-1-4612-0949-2.
5
[Kun84]
[LP68]
[NR14]
[Ok03]
[Pis12]
[Rag08]
[RS09a]
[RS09b]
[Tsi85]
H. Kunita. Stochastic differential equations and stochastic flows of diffeomorphisms. In École d'été de probabilités
de Saint-Flour, XII -- 1982, volume 1097 of Lecture Notes in Math., pages 143 -- 303. Springer, Berlin, 1984. doi:10.
1007/BFb0099433.
J. Lindenstrauss and A. Pełczy ´nski. Absolutely summing operators in Lp -spaces and their applications. Studia
Math., 29:275 -- 326, 1968.
A. Naor and O. Regev. Krivine schemes are optimal. Proc. Amer. Math. Soc., 142(12):4315 -- 4320, 2014.
B. Ø ksendal. Stochastic differential equations. Universitext. Springer-Verlag, Berlin, sixth edition, 2003. ISBN 3-540-
04758-1. doi:10.1007/978-3-642-14394-6. An introduction with applications.
G. Pisier. Grothendieck's theorem, past and present. Bull. Amer. Math. Soc. (N.S.), 49(2):237 -- 323, 2012.
P. Raghavendra. Optimal algorithms and inapproximability results for every CSP? [extended abstract]. In STOC'08,
pages 245 -- 254. ACM, New York, 2008. doi:10.1145/1374376.1374414.
P. Raghavendra and D. Steurer. How to round any csp. In In Proc. 50th IEEE Symp. on Foundations of Comp. Sci.
2009.
P. Raghavendra and D. Steurer. Towards computing the Grothendieck constant. In Proceedings of the Twentieth
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 525 -- 534. SIAM, Philadelphia, PA, 2009.
B. S. Tsirelson. Quantum analogues of Bell's inequalities. The case of two spatially divided domains. Zap. Nauchn.
Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI), 142:174 -- 194, 200, 1985. Problems of the theory of probability
distributions, IX.
6
|
1109.6178 | 2 | 1109 | 2011-11-30T04:27:23 | Space-efficient Local Computation Algorithms | [
"cs.DS"
] | Recently Rubinfeld et al. (ICS 2011, pp. 223--238) proposed a new model of sublinear algorithms called \emph{local computation algorithms}. In this model, a computation problem $F$ may have more than one legal solution and each of them consists of many bits. The local computation algorithm for $F$ should answer in an online fashion, for any index $i$, the $i^{\mathrm{th}}$ bit of some legal solution of $F$. Further, all the answers given by the algorithm should be consistent with at least one solution of $F$.
In this work, we continue the study of local computation algorithms. In particular, we develop a technique which under certain conditions can be applied to construct local computation algorithms that run not only in polylogarithmic time but also in polylogarithmic \emph{space}. Moreover, these local computation algorithms are easily parallelizable and can answer all parallel queries consistently. Our main technical tools are pseudorandom numbers with bounded independence and the theory of branching processes. | cs.DS | cs | Space-efficient Local Computation Algorithms
Noga Alon∗
Ronitt Rubinfeld†
Shai Vardi ‡
Ning Xie §
Abstract
Recently Rubinfeld et al. (ICS 2011, pp. 223 -- 238) proposed a new model of sublinear algorithms
called local computation algorithms. In this model, a computation problem F may have more than one
legal solution and each of them consists of many bits. The local computation algorithm for F should
answer in an online fashion, for any index i, the ith bit of some legal solution of F . Further, all the
answers given by the algorithm should be consistent with at least one solution of F .
In this work, we continue the study of local computation algorithms.
In particular, we develop
a technique which under certain conditions can be applied to construct local computation algorithms
that run not only in polylogarithmic time but also in polylogarithmic space. Moreover, these local
computation algorithms are easily parallelizable and can answer all parallel queries consistently. Our
main technical tools are pseudorandom numbers with bounded independence and the theory of branching
processes.
1
1
0
2
v
o
N
0
3
]
S
D
.
s
c
[
2
v
8
7
1
6
.
9
0
1
1
:
v
i
X
r
a
∗Sackler School of Mathematics and Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel and
Institute for Advanced Study, Princeton, New Jersey 08540, USA. E-mail: [email protected]. Research supported in part by
an ERC Advanced grant, by a USA-Israeli BSF grant and by NSF grant No. DMS-0835373.
†CSAIL, MIT, Cambridge, MA 02139, USA and School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel. E-
mail: [email protected]. Supported by NSF grants CCF-0728645 and CCF-1065125, Marie Curie Reintegration grant
PIRG03-GA-2008-231077 and the Israel Science Foundation grant nos. 1147/09 and 1675/09.
‡School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel. E-mail: [email protected]. Supported by
Israel Science Foundation grant no. 1147/09.
§CSAIL, MIT, Cambridge MA 02139, USA. E-mail: [email protected]. Research supported by NSF grants CCF-
0728645, CCF-0729011 and CCF-1065125.
0
1 Introduction
The classical view of algorithmic analysis, in which the algorithm reads the entire input, performs a com-
putation and then writes out the entire output, is less applicable in the context of computations on massive
data sets. To address this difficulty, several alternative models of computation have been adopted, including
distributed computation as well as various sub-linear time and space models.
Local computation algorithms (LCAs) were proposed in [24] to model the scenario in which inputs to
and outputs from the algorithms are large, such that writing out the entire output requires an amount of time
that is unacceptable. On the other hand, only small portions of the output are required at any point in time
by any specific user. LCAs support queries to the output by the user, such that after each query to a specified
location i, the LCA outputs the value of the output at location i. LCAs were inspired by and intended as a
generalization of several models that appear in the literature, including local algorithms, locally decodable
codes and local reconstruction algorithms. LCAs whose time complexity is efficient in terms of the amount
of solution requested by the user have been given for various combinatorial and coding theoretic problems.
One difficulty is that for many computations, more than one output is considered to be valid, yet the
values returned by the LCA over time must be consistent. Often, the straightforward solutions ask that the
LCA store intermediate values of the computations in order to maintain consistency for later computations.
Though standard techniques can be useful for recomputing the values of random coin tosses in a straightfor-
ward manner, some algorithms (e.g., many greedy algorithms) choose very different solutions based on the
order of input queries. Thus, though the time requirements of the LCA may be efficient for each query, it
is not always clear how to bound the storage requirements of the LCA by a function that is sublinear in the
size of the query history. It is this issue that we focus on in this paper.
1.1 Our main results
Before stating our main results, we mention two additional desirable properties of LCAs. Both of these
properties are achieved in our constructions of LCAs with small storage requirements. The first is that an
LCA should be query oblivious, that is the outputs of A should not depend on the order of the queries but
only on the input and the random bits generated on the random tape of A. The second is that the LCA should
be parallelizable, i.e., that it is able to answer multiple queries simultaneously in a consistent manner.
All the LCAs given in [25] suffer from one or more of the following drawbacks: the worst case space
complexity is linear, the LCA is not query oblivious, and the LCA is not parallelizable. We give new
techniques to construct LCAs for the problems studied in [25] which run in polylogarithmic time as well as
polylogarithmic space. Moreover, all of the LCAs are query oblivious and easily parallelizable.
Theorem 1.1 (Main Theorem 1 (informal)). There is an LCA for Hypergraph Coloring that runs in poly-
logarithmic time and space. Moreover, the LCA is query oblivious and parallelizable.
Theorem 1.2 (Main Theorem 2 (informal)). There is an LCA for Maximal Independent Set that runs in
polylogarithmic time and space. Moreover, the LCA is query oblivious and parallelizable.
We remark that following [25], analogous techniques can be applied to construct LCAs with all of the
desirable properties for the radio network problem and k-CNF problems.
1.2 Techniques
There are two main technical obstacle in making the LCAs constructed in [25] space efficient, query obliv-
ious and parallelizable. The first is that LCAs need to remember all the random bits used in computing
1
previous queries. The second issue is more subtle -- [25] give LCAs based on algorithms which use very
little additional time resources per query as they simulate greedy algorithms. These LCAs output results
that depend directly on the orders in which queries are fed into the algorithms.
We address the randomness issue first. The space inefficient LCAs constructed in [25] for the problems
of concern to us are probabilistic by nature. Consistency among answers to the queries seems to demand
that the algorithm keeps track of all random bits used so far, which would incur linear space complexity.
A simple but very useful observation is that all the computations are local and thus involve a very small
number of random bits. Therefore we may replace the truly random bits with random variables of limited
independence. The construction of small sample space k-wise independent random variables of Alon et
al. [3] allows us to reduce the space complexity from linear to polylogarithmic. This allows us to prove
our main theorem on the LCA for the maximal independent set problem. It is also an important ingredient
in constructing our LCA for Hypergraph Coloring. We believe such a technique will be a standard tool in
future development of LCAs.
For Hypergraph Coloring, we need to also address the second issue raised above. The original LCA for
Hypergraph Coloring in [25] emulates Alon's algorithm [2]. Alon's algorithm runs in three phases. During
the first phase, it colors all vertices in an arbitrary order. Such an algorithm looks "global" in nature and
it is therefore non-trivial to turn it into an LCA. In [25], they use the order of vertices being queried as the
order of coloring in Alon's algorithm, hence the algorithm needs to store all answers to previous queries and
requires linear space in computation.
We take a different approach to overcome this difficulty. Observe that there is some "local" dependency
among the colors of vertices -- namely, the color of any vertex depends only on the colors of at most a
constant number, say D, other vertices. The colors of these vertices in turn depend on the colors of their
neighboring vertices, and so on. We can model the hypergraph coloring process by a query tree: Suppose
the color of vertex x is now queried. Then the root node of the query tree is x, the nodes on the first level are
the vertices whose colors the color of x depends on. In general, the colors of nodes on level i depends on1
the colors of nodes on level i + 1. Note that the query tree has degree bound D and moreover, the size of
the query tree clearly depends on the order in which vertices are colored, since the color of a vertex depends
only on vertices that are colored before it. In particular, if x is the kth vertex to be colored, then the query
tree contains at most k vertices.
An important fact to note is that Alon's algorithm works for any order, in particular, it works for a
random order. Therefore we can apply the random order method of Nguyen and Onak [20]: generate a
random number r ∈ [0, 1], called the rank, and use these ranks to prune the original query tree into a
random query tree T . Specifically, T is defined recursively: the root of T is still x. A node z is in T if its
parent node y in the original query tree is in T and r(z) < r(y). Intuitively, a random query tree is small
and indeed it is surprisingly small [20]: the expected size of T is eD−1
D , a constant!
Therefore, if we color the vertices in the hypergraph in a random order, the expected number of vertices
we need to color is only a constant. However, such an "average case" result is insufficient for our LCA
purpose: what we need is a "worst case" result which tells almost surely how large a random query tree will
be. In other words, we need a concentration result on the sizes of the random query trees. The previous
techniques in [20, 28] do not seem to work in this setting.
Consider the worst case in which the rank of the root node x is 1. A key observation is, although there
are D child nodes of x, only the nodes whose ranks are close to 1 are important, as the child nodes with
smaller ranks will die out quickly. But in expectation there will be very few important nodes! This inspires
1In fact, they may depend on the colors of some nodes on levels lower than i. However, as we care only about query complexity,
we will focus on the worst case that the query relations form a tree.
2
us to partition the random query tree into D + 1 levels based on the ranks of the nodes, and analyze the
sizes of trees on each level using the theory of branching processes. In particular, we apply a quantitative
bound on the total number of off-springs of a Galton-Watson process [22] to show that, for any m > 0, with
probability at least 1 − 1/m2 the size of a random query tree has at most C(D) logD+1 m vertices, where
C(D) is some constant depending only on D. We conjecture that the upper bound can be further reduced to
C(D) log m.
However, the random order approach raise another issue: how do we store the ranks of all vertices? Ob-
serve that in constructing a random query tree, the actual values of the ranks are never used -- only the relative
orders between vertices matter. This fact together with the fact that all computations are local enables us to
replace the rank function with some pseudorandom ordering among the vertices, see Section 4 for formal
definition and construction. The space complexity of the pseudorandom ordering is only polylogarithmic,
thus making the total space complexity of the LCA also polylogarithmic.
1.3 Other related work
Locally decodable codes [9] which given an encoding of a message, provide quick access to the requested
bits of the original message, can be viewed as LCAs. Known constructions of LDCs are efficient and use
small space [27]. LCAs generalize the reconstruction models described in [1, 6, 26, 7]. These models
describe scenarios where an input string that has a certain property, such as monotonicity, is assumed to be
corrupted at a relatively small number of locations. The reconstruction algorithm gives fast query access to
an uncorrupted version of the string that is close to the original input. Most of the works mentioned are also
efficient in terms of space.
In [24], it is noted that the model of LCAs is related to local algorithms, studied in the context of
distributed computing [19, 17, 12, 13, 14, 11, 10]. This is due to a reduction given by Parnas Ron [23]
which allows one to construct (sequential) LCAs based on constant round distributed algorithms. Note that
this relationship does not immediately yield space-efficient local algorithms, nor does it yield sub-linear
time LCAs when used with parallel or distributed algorithms whose round complexity is O(log n).
Recent exciting developments in sublinear time algorithms for sparse graph and combinatorial optimiza-
tion problems have led to new constant time algorithms for approximating the size of a minimum vertex
cover, maximal matching, maximum matching, minimum dominating set, minimum set cover, packing and
covering problems (cf. [23, 16, 20, 28]). For example, for Maximal Independent Set, these algorithms con-
struct a constant-time oracle which for most, but not all, vertices outputs whether or not the vertex is part of
the independent set. For the above approximation algorithms, it is not necessary to get the correct answer
for each vertex, but for LCAs, which must work for any sequence of online inputs, the requirements are
more stringent, thus the techniques are not applicable without modification.
1.4 Organization
The rest of the paper is organized as follows. Some preliminaries and notations that we use throughout the
paper appear in Section 2. We then prove our main technical result, namely the bound on the sizes of random
query trees in Section 3. In Section 4 we construct pseudorandom orderings with small space. Finally we
apply the techniques developed in Section 3 and Section 4 to construct LCAs for the hypergraph coloring
problem and the maximal independent set problem in Section 5 and Section 6, respectively.
3
2 Preliminaries
Unless stated otherwise, all logarithms in this paper are to the base 2. Let n ≥ 1 be a natural number. We
use [n] to denote the set {1, . . . , n}.
All graphs in this paper are undirected graphs. Let G = (V, E) be a graph. The distance between two
vertices u and v in V (G), denoted by dG(u, v), is the length of a shortest path between the two vertices.
We write NG(v) = {u ∈ V (G) : (u, v) ∈ E(G)} to denote the neighboring vertices of v. Furthermore, let
N +
G (v) = N (v) ∪ {v}. Let dG(v) denote the degree of a vertex v.
2.1 Local computation algorithms
We present our model of local computation algorithms: Let F be a computational problem and x be an input
to F . Let F (x) = {y y is a valid solution for input x}. The search problem is to find any y ∈ F (x).
Definition 2.1 ((t, s, δ)-local algorithms [25]). Let x and F (x) be defined as above. A (t(n), s(n), δ(n))-
local computation algorithm A is a (randomized) algorithm which implements query access to an arbitrary
y ∈ F (x) and satisfies the following: A gets a sequence of queries i1, . . . , iq for any q > 0 and after
each query ij it must produce an output yij satisfying that the outputs yi1, . . . , yiq are substrings of some
y ∈ F (x). The probability of success over all q queries must be at least 1 − δ(n). A has access to a
random tape and local computation memory on which it can perform current computations as well as store
and retrieve information from previous computations. We assume that the input x, the local computation
tape and any random bits used are all presented in the RAM word model, i.e., A is given the ability to access
a word of any of these in one step. The running time of A on any query is at most t(n), which is sublinear in
n, and the size of the local computation memory of A is at most s(n). Unless stated otherwise, we always
assume that the error parameter δ(n) is at most some constant, say, 1/3. We say that A is a strongly local
computation algorithm if both t(n) and s(n) are upper bounded by logc n for some constant c.
Two important properties of LCAs are as follows:
Definition 2.2 (Query oblivious[25]). We say an LCA A is query order oblivious (query oblivious for short)
if the outputs of A do not depend on the order of the queries but depend only on the input and the random
bits generated on the random tape of A.
Definition 2.3 (Parallelizable[25]). We say an LCA A is parallelizable if A supports parallel queries, that
is the LCA is able to answer multiple queries simultaneously so that all the answers are consistent.
2.2 k-wise independent random variables
Let 1 ≤ k ≤ n be an integer. A distribution D : {0, 1}n → R≥0 is k-wise independent if restricting D
to any index subset S ⊂ [n] of size at most k gives rise to a uniform distribution. A random variable is
said to be k-wise independent if its distribution function is k-wise independent. Recall that the support of
a distribution D, denoted supp(D), is the set of points at which D(x) > 0. We say a discrete distribution
D is symmetric if D(x) = 1/supp(D) for every x ∈ supp(D). If a distribution D : {0, 1}n → R≥0 is
symmetric with supp(D) ≤ 2m for some m ≤ n, then we may index the elements in the support of D
by {0, 1}m and call m the seed length of the random variable whose distribution is D. We will need the
following construction of k-wise independent random variables over {0, 1}n with small symmetric sample
space.
4
Theorem 2.4 ([3]). For every 1 ≤ k ≤ n, there exists a symmetric distribution D : {0, 1}n → R≥0 of
support size at most n⌊ k
2 ⌋ and is k-wise independent. That is, there is a k-wise independent random variable
x = (x1, . . . , xn) whose seed length is at most O(k log n). Moreover, for any 1 ≤ i ≤ n, xi can be
computed in space O(k log n).
3 Bounding the size of a random query tree
3.1 The problem and our main result
Consider the following scenario which was first studied by [20] in the context of constant-time approxi-
mation algorithms for maximal matching and some other problems. We are given a graph G = (V, E) of
bounded degree D. A real number r(v) ∈ [0, 1] is assigned independently and uniformly at random to every
vertex v in the graph. We call this random number the rank of v. Each vertex in the graph G holds an input
x(v) ∈ R, where the range R is some finite set. A randomized Boolean function F is defined inductively on
the vertices in the graph such that F (v) is a function of the input x(v) at v as well as the values of F at the
neighbors w of v for which r(w) < r(v). The main question is, in order to compute F (v0) for any vertex
v0 in G, how many queries to the inputs of the vertices in the graph are needed?
Here, for the purpose of upper bounding the query complexity, we may assume for simplicity that the
graph G is D-regular and furthermore, G is an infinite D-regular tree rooted at v0. It is easy to see that
making such modifications to G can never decrease the query complexity of computing F (v0).
Consider the following question. We are given an infinite D-regular tree T rooted at v0. Each node w in
T is assigned independently and uniformly at random a real number r(w) ∈ [0, 1]. For every node w other
than v0 in T , let parent(w) denote the parent node of w. We grow a (possibly infinite) subtree T of T rooted
at v as follows: a node w is in the subtree T if and only if parent(w) is in T and r(w) < r(parent(w)) (for
simplicity we assume all the ranks are distinct real numbers). That is, we start from the root v, add all the
children of v whose ranks are smaller than that of v to T . We keep growing T in this manner where a node
w′ ∈ T is a leaf node in T if the ranks of its D children are all larger than r(w′). We call the random tree T
constructed in this way a query tree and we denote by T the random variable that corresponds to the size
of T . We would like to know what are the typical values of T .
Following [20, 21], we have that, for any node w that is at distance t from the root v0, Pr[w ∈
T ]=1/(t+1)! as such an event happens if and only if the ranks of the t + 1 nodes along the shortest path
from v0 to w is in monotone decreasing order. It follows from linearity of expectation that the expected
value of T is given by the elegant formula [21]
E[T ] =
∞
Xt=0
Dt
(t + 1)!
=
eD − 1
D
,
which is a constant depending only on the degree bound D.
Our main result in this section can be regarded as showing that in fact T is highly concentrated around
its mean:
Theorem 3.1. For any degree bound D ≥ 2, there is a constant C(D) which depends on D only such that
for all large enough N ,
Pr[T > C(D) logD+1 N ] < 1/N 2.
5
3.2 Breaking the query tree into levels
A key idea in the proof is to break the query tree into levels and then upper bound the sizes of the subtrees
D+1 , 1 − i−1
on each level separately. First partition the interval [0, 1] into D + 1 sub-intervals: Ii := (1 − i
D+1 ]
1
for i = 1, 2, . . . , D and ID+1 = [0,
D+1 ]. We then decompose the query tree T into D + 1 levels such that
a node v ∈ T is said to be on level i if r(v) ∈ Ii. For ease of exposition, in the following we consider the
worst case that r(v0) ∈ I1. Then the vertices of T on level 1 form a tree which we call T1 = T (1)
rooted at
v0. The vertices of T on level 2 will in general form a set of trees {T (1)
}, where the total number
of such trees m2 is at most D times the number of nodes in T1 (we have only inequality here because some
of the child nodes in T of the nodes in T1 may fall into levels 2, 3, etc). Finally the nodes on level D + 1
form a forest {T (1)
i } are generated by the same stochastic
process, as the ranks of all nodes in T are i.i.d. random variables. The next lemma shows that each of the
subtrees on any level is of size O(log N ) with probability at least 1 − 1/N 3,
}. Note that all these trees {T (j)
, . . . , T (m2)
D+1, . . . , T (mD+1)
D+1
2
2
1
Lemma 3.2. For any 1 ≤ i ≤ D + 1 and any 1 ≤ j ≤ mi, with probability at least 1 − 1/N 3, T (j)
O(log N ).
i
=
One can see that Theorem 3.1 follows directly from Lemma 3.2: Once again we consider the worst case
that r(v0) ∈ I1. By Lemma 3.2, the size of T1 is at most O(log N ) with probability at least 1 − 1/N 3. In
what follows, we always condition our argument upon that this event happens. Notice that the root of any
tree on level 2 must have some node in T1 as its parent node; it follows that m2, the number of trees on level
2, is at most D times the size of T1, hence m2 = O(log N ). Now applying Lemma 3.2 to each of the m2
trees on level 2 and assume that the high probability event claimed in Lemma 3.2 happens in each of the
subtree cases, we get that the total number of nodes at level 2 is at most O(log2 N ). Once again, any tree
on level 3 must have some node in either level 1 or level 2 as its parent node, so the total number of trees
on level 3 is also at most D(O(log N ) + O(log2 N )) = O(log2 N ). Applying this argument inductively,
we get that mi = O(logi−1 N ) for i = 2, 3, . . . , D + 1. Consequently, the total number of nodes at all
D + 1 levels is at most O(log N ) + O(log2 N ) + · · · + O(logD+1 N ) = O(logD+1 N ), assuming the high
probability event in Lemma 3.2 holds for all the subtrees in all the levels. By the union bound, this happens
with probability at least 1 − O(logD+1 N )/N 3 > 1 − 1/N 2, thus proving Theorem 3.1.
The proof of Lemma 3.2 requires results in branching processes, in particular the Galton-Watson pro-
cesses.
3.3 Galton-Watson processes
Consider a Galton-Watson process defined by the probability function p := {pk; k = 0, 1, 2, . . .}, with
pk ≥ 0 and Pk pk = 1. Let f (s) = P∞
k=0 pksk be the generating function of p. For i = 0, 1, . . . , let Zi
be the number of off-springs in the ith generation. Clearly Z0 = 1 and {Zi : i = 0, 1, . . .} form a Markov
chain. Let m := E[Z1] = Pk kpk be the expected number of children of any individual. The classical result
of the Galton-Watson processes is that the survival probability (namely limn→∞ Pr[Zn > 0]) is zero if and
only if m ≤ 1. Let Z = Z0 + Z1 + · · · be the sum of all off-springs in all generations of the Galton-Watson
process. The following result of Otter is useful in bounding the probability that Z is large.
Theorem 3.3 ([22]). Suppose p0 > 0 and that there is a point a > 0 within the circle of convergence of
f for which af ′(a) = f (a). Let α = a/f (a). Let t = gcd{r : pr > 0}, where gcd stands for greatest
common divisor. Then
6
Pr[Z = n] =
a
2παf ′′(a)(cid:17)1/2
t(cid:16)
0,
α−nn−3/2 + O(α−nn−5/2),
if n ≡ 1 (mod t);
if n 6≡ 1 (mod t).
(1)
In particular, if the process is non-arithmetic, i.e. gcd{r : pr > 0} = 1, and
a
αf ′′(a) is finite, then
Pr[Z = n] = O(α−nn−3/2),
and consequently Pr[Z ≥ n] = O(α−n).
3.4 Proof of Lemma 3.2
To simplify exposition, we prove Lemma 3.2 for the case of tree T1. Recall that T1 is constructed recursively
as follows: for every child node v of v0 in T , we add v to T1 if r(v) < r(v0) and r(v) ∈ I1. Then for every
child node v of v0 in T1, we add the child node w of v in T to T1 if r(w) < r(v) and r(w) ∈ I1. We repeat
this process until there is no node that can be added to T1.
Once again, we work with the worst case that r(v0) = 1. To upper bound the size of T1, we consider a
1. The process that
1 and w is a child node of v
1 as long as r(w) ∈ I1, but give up the requirement that r(w) < r(v). Clearly, we
related random process which also grows a subtree of T rooted at v0, and denote it by T ′
grows T ′
in T , then we add w to T ′
always have T1 ⊆ T ′
1 is the same as that of T1 except for the following difference: if v ∈ T ′
1 and hence T ′
1 ≥ T1.
Note that the random process that generates T ′
1 is in fact a Galton-Watson process, as the rank of each
node in T is independently and uniformly distributed in [0, 1]. Since I1 = 1/(D + 1), the probability
function is
p = {(1 − q)D,(cid:18)D
1(cid:19)q(1 − q)D−1,(cid:18)D
2(cid:19)q2(1 − q)D−2, . . . , qD},
where q := 1/(D + 1) is the probability that a child node in T appears in T ′
Note that the expected number of children of a node in T ′
tree with probability one.
1 is Dq = D/(D + 1) < 1, so the tree T ′
1 when its parent node is in T ′
1.
1 is a finite
The generating function of p is
f (s) = (1 − q + qs)D,
as the probability function {pk} obeys the binomial distribution pk = b(k, D, q). In addition, the conver-
gence radius of f is ρ = ∞ since {pk} has only a finite number of non-zero terms.
Solving the equation af ′(a) = f (a) yields a = 1−q
q(D−1) = D
D−1. It follows that (since D ≥ 2)
D − 1(cid:19)D−2
1 − q
> 0,
f ′′(a) = q2D(D − 1)(cid:18)1 − q +
hence the coefficient in (1) is non-singular.
7
Let α(D) := a/f (a) = 1/f ′(a), then
1/α(D) = f ′(a) =
D
D + 1
(
)D−1
D2
D2 − 1
1
)(D2−1)/(D+1) D
D + 1
= (1 +
D2 − 1
< e1/(D+1) D
< (cid:18)(1 +
1
D
D + 1
)D+1(cid:19)1/(D+1) D
D + 1
where in the third and the fourth steps we use the inequality (see e.g. [18]) that (1 + 1
for any positive integer t. This shows that α(D) is a constant greater than 1.
t )t < e < (1 + 1
t )t+1
= 1,
case) gives that, for all large enough n, Pr[T ′
Now applying Theorem 3.3 to the Galton-Watson process which generates T ′
1 (note that t = 1 in our
1 ≥
i=n 2−ci ≤ 2−Ω(n) for all large enough n. Hence for all large enough N, with probability at least
1 = n] ≤ 2−cn for some constant c. It follows that Pr[T ′
1 − 1/N 3, T1 ≤ T ′
1 = O(log N ). This completes the proof of Lemma 3.2.
n] ≤ P∞
4 Construction of almost k-wise independent random orderings
An important observation that enables us to make some of our local algorithms run in polylogarithmic space
is the following. In the construction of a random query tree T , we do not need to generate a random real
number r(v) ∈ [0, 1] independently for each vertex v ∈ T ; instead only the relative orderings among the
vertices in T matter. Indeed, when generating a random query tree, we only compare the ranks between
a child node w and its parent node v to see if r(w) < r(v); the absolute values of r(w) and r(v) are
irrelevant and are used only to facilitate our analysis in Section 3. Moreover, since (almost surely) all our
computations in the local algorithms involve only a very small number of, say at most k, vertices, so instead
of requiring a random source that generates total independent random ordering among all nodes in the graph,
any pseudorandom generator that produces k-wise independent random ordering suffices for our purpose.
We now give the formal definition of such orderings.
Let m ≥ 1 be an integer. Let D be any set with m elements. For simplicity and without loss of
generality, we may assume that D = [m]. Let R be a totally ordered set. An ordering of [m] is an injective
function r : [m] → R. Note that we can project r to an element in the symmetric permutation group Sm in
a natural way: arrange the elements {r(1), . . . , r(m)} in R in the monotone increasing order and call the
permutation of [m] corresponding to this ordering the projection of r onto Sm and denote it by PSmr. In
general the projection PSm is not injective. Let r = {ri}i∈I be any family of orderings indexed by I. The
random ordering Dr of [m] is a distribution over a family of orderings r. For any integer 2 ≤ k ≤ m, we
say a random ordering Dr is k-wise independent if for any subset S ⊆ [m] of size k, the restriction of the
projection onto Sm of Dr over S is uniform over all the k! possible orderings among the k elements in S. A
random ordering Dr is said to ǫ-almost k-wise independent if the statistical distance between Dr is at most
ǫ from some k-wise independent random ordering. Note that our definitions of k-wise independent random
ordering and almost k-wise independent random ordering are different from that of k-wise independent
permutation and almost k-wise independent permutation (see e.g. [8]), where the latter requires that the
function to be a permutation (i.e., the domain and the range of the function are the same set).
In this
8
1
m2 -almost k-wise independent random ordering whose seed length is
section we give a construction of
O(k log2 m). In our later applications k = polylogm so the seed length of the almost k-wise independent
random ordering is also polylogarithmic.
Theorem 4.1. Let m ≥ 2 be an integer and let 2 ≤ k ≤ m. Then there is a construction of 1
k-wise independent random ordering over [m] whose seed length is O(k log2 m).
m2 -almost
Proof. For simplicity we assume that m is a power of 2. Let s = 4 log m. We generate s independent copies
of k-wise independent random variables Z1, . . . , Zs with each Zℓ, 1 ≤ ℓ ≤ s, in {0, 1}m. By Theorem 2.4,
the seed length of each random variable Zℓ is O(k log m) and therefore the total space needed to store these
random seeds is O(k log2 m). Let these k-wise independent m-bit random variables be
Z1 = z1,1, . . . , z1,m;
Z2 = z2,1, . . . , z2,m;
. . . . . .
Zs = zs,1, . . . , zs,m.
Now for every 1 ≤ i ≤ m, we view each r(i)def= z1,iz2,i · · · zs,i as an integer in {0, 1, . . . , 2s − 1} written
in the s-bit binary representation and use r : [m] → {0, 1, . . . , 2s − 1} as the ranking function to order the
m elements in the set. We next show that, with probability at least 1 − 1/m2, r(1), . . . , r(m) are distinct m
integers.
Let 1 ≤ i < j ≤ m be any two distinct indices. For every 1 ≤ ℓ ≤ s, since zℓ,1, . . . , zℓ,m are k-wise
independent and thus also pair-wise independent, it follows that Pr[zℓ,i = zℓ,j] = 1/2. Moreover, as all
Z1, . . . , Zs are independent, we therefore have
Pr[r(i) = r(j)] = Pr[zℓ,i = zℓ,j for every 1 ≤ ℓ ≤ s]
s
Pr[zℓ,i = zℓ,j]
=
Yℓ=1
= (1/2)s
= 1/m4.
1 − 1/m2, all these m numbers are distinct.
Applying a union bound argument over all (cid:0)m
2(cid:1) distinct pairs of indices gives that with probability at least
Since each Zℓ, 1 ≤ ℓ ≤ s, is a k-wise independent random variable in {0, 1}m, therefore for any subset
{i1, . . . , ik} of k indices, (r(i1), . . . , r(ik)) is distributed uniformly over all 2ks tuples. By symmetry,
conditioned on that r(i1), . . . , r(ik) are all distinct, the restriction of the ordering induced by the ranking
function r to {i1, . . . , ik} is completely independent. Finally, since the probability that r(1), . . . , r(m)
1
are not distinct is at most 1/m2, it follows that the random ordering induced by r is
m2 -almost k-wise
independent.
5 LCA for Hypergraph Coloring
We now apply the technical tools developed in Section 3 and Section 4 to the design and analysis of LCAs.
Recall that a hypergraph H is a pair H = (V, E) where V is a finite set whose elements are called
nodes or vertices, and E is a family of non-empty subsets of V , called hyperedges. A hypergraph is called
9
k-uniform if each of its hyperedges contains precisely k vertices. A two-coloring of a hypergraph H is a
mapping c : V → {red, blue} such that no hyperedge in E is monochromatic. If such a coloring exists,
then we say H is two-colorable. In this paper we assume that each hyperedge in H intersects at most d
other hyperedges. Let N be the number of hyperedges in H. Here and after we think of k and d as fixed
constants and all asymptotic forms are with respect to N. By the Lovász Local Lemma (see, e.g. [4]) when
e(d + 1) ≤ 2k−1, the hypergraph H is two-colorable.
Following [25], we let m be the total number of vertices in H. Note that m ≤ kN, so m = O(N ).
For any vertex x ∈ V , we use E(x) to denote the set of hyperedges x belongs to. For any hypergraph
H = (V, E), we define a vertex-hyperedge incidence matrix M ∈ {0, 1}m×N so that, for every vertex x
and every hyperedge e, Mx,e = 1 if and only if e ∈ E(x). Because we assume both k and d are constants,
the incidence matrix M is necessarily very sparse. Therefore, we further assume that the matrix M is
implemented via linked lists for each row (that is, vertex x) and each column (that is, hyperedge e).
Let G be the dependency graph of the hyperedges in H. That is, the vertices of the undirected graph G
are the N hyperedges of H and a hyperedge Ei is connected to another hyperedge Ej in G if Ei ∩ Ej 6= ∅.
It is easy to see that if the input hypergraph is given in the above described representation, then we can find
all the neighbors of any hyperedge Ei in the dependency graph G (there are at most d of them) in O(log N )
time.
5.1 Overview of Alon's algorithm
We now give a sketch of Alon's algorithm [2]; for a detailed description of the algorithm in the context of
LCA see [25].
The algorithm runs in three phases. In the first phase, we go over all the vertices in the hypergraph in
any order and color them in {red, blue} uniformly at random. During this process, if any hyperedge has too
many vertices (above some threshold) in it are colored in one color and no vertex is colored in the other color,
then this hyperedge is said to become dangerous. All the uncolored vertices in the dangerous hyperedges are
then frozen and will be skipped during Phase 1 coloring. A hyperedge is called survived if it does not have
vertices in both colors at the end of Phase 1. The basic lemma, based on the breakthrough result of Beck [5],
claims that after Phase 1, almost surely all connected components of the dependency graph H of survived
hyperedges are of sizes at most O(log N ). We then proceed to the second phase of the algorithm which
repeats the same coloring process (with some different threshold parameter) for each connected component
and gives rise to connected components of size O(log log N ). Finally in the third phase we perform a brute-
force search for a valid coloring whose existence is guaranteed by the Lovász local lemma. As each of the
connected components in Phase 3 has at most O(log log N ) vertices, the running time of each brute force
search is thus bounded by polylogN.
To turn Alon's algorithm into an LCA, Rubinfeld et al. [25] note that one may take the order that vertices
are queried as the order to color the vertices and then in Phase 2 and Phase 3 focus only on the connected
components in which the queried vertex lie. This leads to an LCA with polylogarithmic running time but
the space complexity can be linear in the worst case (as the algorithm needs to remember the colors of
all previously queried or colored vertices).
In addition, the LCA is not query oblivious and not easily
parallelizable.
5.2 New LCA for Hypergraph Coloring
To remedy these, we add several new ingredients to the LCA in [25] and achieve an LCA with both time and
space complexity are polylogarithmic. In addition, the LCA is query oblivious and easily parallelizable.
10
1st ingredient: bounded-degree dependency. We first make use of the following simple fact: the color of
any fixed vertex in the hypergraph depends only on the colors of a very small number of vertices. Specif-
ically, if vertex x lies in hyperedges E1, . . . , Ed′, then the color of x depends only on the colors of all the
vertices in E1, . . . , Ed′. As every hyperedge is k-uniform and each hyperedge intersects at most d other
hyperedges, the color of any vertex depends on at most the colors of D = k(d + 1) other vertices.
2nd ingredient: random permutation. Note that in the first phase of Alon's coloring algorithm, any order
of the vertices will work. Therefore, we may apply the idea of random ordering in [20]. Specifically, suppose
we are given a random number generator r : [m] → [0, 1] which assign a random number uniformly and
independently to every vertex in the hypergraph. Suppose the queried vertex is x. Then we build a (random)
query tree T rooted at x using BFS as follows: there are at most D other vertices such that the color of x
depends on the colors of these vertices. Let y be any of such vertex. If r(y) < r(x), i.e. the random number
assigned to y is smaller than that of x, then we add y as a child node of x in T . We build the query tree this
way recursively until there is no child node can be added to T . By Theorem 3.1, with probability at least
1 − 1/m2, the total number of nodes in T is at most polylogm and is thus also at most polylogN. This
implies that, if we color the vertices in T in the order from bottom to top (that is, we color the leaf nodes
first, then the parent nodes of the leaf nodes and so on, and color the root node x last), then for any vertex
x, with probability at least 1 − 1/m2 we can follow Alon's algorithm and color at most polylogN vertices
(and ignore all other vertices in the hypergraph) before coloring x. Therefore the running time of the first
phase of our new LCA is (almost surely) at most polylogN.
3rd ingredient: k-wise independent random ordering. The random permutation method requires linear
space to store all the random numbers that have been revealed in previous queries in order to make the
answers consistent. However, two useful observations enable us to reduce the space complexity of random
ordering from linear to polylogarithmic. First, only the relative orderings among vertices matter: in building
the query tree T we only check if r(y) < r(x) but the absolute value of r(x) and r(y) are irrelevant. There-
fore we can replace the random number generator r with an equivalent random ordering function r ∈ Sm,
where Sm is the symmetric group on m elements. Second, as the query tree size is at most polylogarith-
mic almost surely, the random ordering function r need not be totally random but a polylogarithmic-wise
independent permutation suffices2. Therefore we can use the construction in Theorem 4.1 of
1
m2 -almost
k-wise independent random ordering of all the vertices in the hypergraph with k = polylogN. The space
complexity of such a random ordering, or the seed length, is O(k log2 m) = polylogN.
4th ingredient: k-wise independent random coloring. Finally, the space complexity for storing all the
random colors assigned to vertices is also linear in worst case. Once again we exploit the fact that all
computations in LCAs are local to reduce the space complexity. Specifically, the proof of the basic lemma of
Alon's algorithm (see e.g. [4, Claim 5.7.2]) is valid as long as the random coloring of the vertices is c log N-
wise independent, where c is some absolute constant. Therefore we can replace the truly random numbers
in {0, 1}m used for coloring with a c log N-wise independent random numbers in {0, 1}m constructed in
Theorem 2.4 thus reducing the space complexity of storing random colors to O(log2 N ).
5.3 Pseudocode of the LCA and main result
To put everything together, we have the following LCA for Hypergraph Coloring as illustrated in Fig. 1,
Fig. 2 and Fig. 3. In the preprocessing stage, the algorithm generates O( log N
log log N ) copies of pseudo-random
2 Since the full query tree has degree bound D, so the total number of nodes queried in building the random query tree T is at
most DT , which is also at most polylogarithmic.
11
LCA for Hypergraph Coloring
Preprocessing:
1. generate O( log N
2. generate a 1
log log N ) copies of c log N -wise independent random variables in {0, 1}m
m2 -almost polylogN -wise independent random ordering over [m]
Input: a vertex x ∈ V
Output: a color in {red, blue}
1. Use BFS to grow a random query tree T rooted at x
2. Color the vertices in T bottom up
3.
If x is colored red or blue, return the color
Else run Phase 2 Coloring(x)
Figure 1: Local computation algorithm for Hypergraph Coloring
colors for every vertex in the hypergraph and a pseudorandom ordering of all the vertices. To answer each
query, the LCA runs in three phases. Suppose the color of vertex x is queried. During the first phase,
the algorithm uses BFS to build a random query tree rooted at x and then follows Alon's algorithm to
color all the vertices in the query tree.
If x gets colored in Phase 1, the algorithm simply returns that
color; if x is frozen in Phase 1, then Phase 2 coloring is invoked. In the second phase, the algorithm first
explores the connected components around x of survived hyperedges. Then Alon's algorithm is performed
again, but this time only on the vertices in the connected component. For some technical reason, the random
log log N ) times3, until a good coloring is found which makes all the surviving
coloring process is repeated O( log N
connected components after Phase 2 very small. If x gets colored in the good coloring, then that color is
returned; otherwise the algorithm runs the last phase, in which a brute-force search is performed to find the
color of x.
The time and space complexity as well as the error bound of the LCA are easy to analyze and we have
the following main result of LCA for Hypergraph Coloring:
Theorem 5.1. Let d and k be such that there exist three positive integers k1, k2 and k3 such that the follow-
ings hold:
k1 + k2 + k3 = k,
16d(d − 1)3(d + 1) < 2k1,
16d(d − 1)3(d + 1) < 2k2,
2e(d + 1) < 2k3.
Then there exists a (polylogN, polylogN, 1/N )-local computation algorithm which, given a hypergraph H
and any sequence of queries to the colors of vertices (x1, x2, . . . , xs), returns a consistent coloring for all
xi's which agrees with some 2-coloring of H.
6 LCA for Maximal Independent Set
Recall that an independent set (IS) of a graph G is a subset of vertices such that no two vertices in the set
are adjacent. An independent set is called a maximal independent set (MIS) if it is not properly contained in
any other IS.
3This is why the algorithm generates many copies of independent pseudorandom colorings at the beginning of the LCA.
12
Phase 2 Coloring(x)
Input: a vertex x ∈ V
Output: a color in {red, blue} or FAIL
1. Start from E(x) to explore G in order to find the connected
component C1(x) of survived hyperedges around x
2.
If the size of the component is larger than c2 log N
Abort and return FAIL
3. Repeat the following O( log N
log log N ) times and stop if a good coloring is founda
(a) Color all the vertices in C1(x) uniformly at random
(b) Explore the dependency graph of GS1(x)
(c) Check if the coloring is good
4.
If x is colored in the good coloring, return that color
Else run Phase 3 Coloring(x)
a Following [25], let S1(x) be the set of surviving hyperedges in C1(x) after all vertices in C1(x)
are either colored or are frozen. Now we explore the dependency graph of S1(x) to find out all the
connected components. We say a Phase 2 coloring is good if all connected components in GS1 (x)
have sizes at most c3 log log N, where c3 is some absolute constant.
Figure 2: Local computation algorithm for Hypergraph Coloring: Phase 2
Phase 3 Coloring(x)
Input: a vertex x ∈ V
Output: a color in {red, blue}
1. Start from E(x) to explore G in order to find the connected
component of all the survived hyperedges around x
2. Go over all possible colorings of the connected component
and color it using a feasible coloring.
3. Return the color c of x in this coloring.
Figure 3: Local computation algorithm for Hypergraph Coloring: Phase 3
13
In [24, 25], a two-phase LCA is presented for MIS. For completeness, we present the pseudocode of the
LCA in Appendix A. Let G be a graph with maximum degree d and suppose the queried vertex is v. In
the first phase, the LCA simulates Luby's algorithm for MIS [15]. However, instead of running the parallel
algorithm for O(log n) rounds as the original Luby's algorithm, the LCA simulates the parallel algorithm
for only O(d log d) rounds. Following an argument of Parnas and Ron [23], the sequential running time for
simulating the parallel algorithm to determine whether a given node is in the MIS is dO(log d). If v or any of
v's neighbors is put into the independent set during the first phase, then the algorithm return "Yes" or "No",
respectively. If, on the other hand, v lies in some connected component of "surviving" vertices after running
the first phase, then the algorithm proceeds to the second phase algorithm, in which a simple linear-time
greedy search for an MIS of the component is performed. A key result proved in [24, 25] is that, after the
first phase of the algorithm, almost surely all connected components have sizes at most O(poly(d) log n).
Therefore the running time4 of the second phase is dO(log d) log n.
To implement such a two-phase LCA and ensure that all answers are consistent, we need to maintain a
random tape that keeps a record of all the generated random bits during previous runs, which implies the
space complexity of the LCA is linear in the worst case. To see this, suppose two vertices u and v are
connected in G and u is queried first. Suppose further that the LCA runs on u and finds out during the first
phase that u is in the IS. If vertex v is queried at some time later, we need to ensure that, when simulating
Luby's algorithm u is put in the IS in some round (hence v is deleted in the round after that). This in turn
requires that we retrieve the random bits used during the run of LCA on u.
A simple but crucial observation which enables us to reduce the space complexity of the LCA for MIS
is, since all the computations are "local", we may replace the truly random bits used in the algorithm with
random bits of limited independence constructed in Theorem 2.4.
First we round the degree bound of G to d = 2⌈log d⌉. Note that d ≤ d < 2d. Now we can generate the
probability 1/2 d used in Luby's algorithm (c.f. Figure 4) by tossing log d = ⌈log d⌉ independent fair coins.
Since the second phase of the LCA is deterministic, we can therefore focus on the first phase only. The
running time of the first phase is shown to be dO(log d) [25]. Following the notation in [25], for a vertex v
in G, let Av be the event that v is a surviving vertex at the end of Phase 1 and let Bv be the event that v is
in state "⊥" after running MISB for O(d log d) rounds, where MISB is a variant of MIS, a subroutine of
the first phase algorithm. It was shown in [25] that Av ⊆ Bv (Claim 4.2) and for any subset of vertices W ,
Pr[all vertices in W are surviving vertices]
= Pr[∩v∈W Av]
≤ Pr[∩v∈W Bv].
Following the proof of Lemma 4.6 in [25], a graph H on the vertices V (G) is called a dependency graph
for {Bv}v∈V (G) if for all v the event Bv is mutually independent of all Bu such that (u, v) /∈ H. Let H 3
denote the "distance-3" graph of H, that is, vertices u and v are connected in H 3 if their distance in H is
exactly 3. Let W be a subset of vertices in H 3. Then, since all vertices in W are at least 3-apart, all the
events {Bv}v∈W are mutually independent, it follows that the probability that all vertices in W are surviving
vertices satisfies
Pr[∩v∈W Bv] = Yv∈W
Pr[Bv].
Finally in the proof of Lemma 4.6 in [25], the size of W is taken to be c1 log n for some constant c1 to
4Note that we need to run a BFS starting from v to explore the connected component in which v lies. Each step of the BFS
incurs a run on the explored node of the first phase LCA.
14
show that, almost surely all connected components of surviving vertices after Phase 1 are of sizes at most
poly(d) log n.
Now we try to replace the true random bits used in the LCA in [25] with pseudorandom bits of limited
independence. Firstly, since the running time of the first phase is dO(log d), hence this is also the running time
of the algorithm if the subroutine MIS is replaced with MISB. It follows that each event Bv depends on at
most dO(log d) · log d = dO(log d) random bits. Secondly, the argument we sketched in the last paragraph is
still valid as long as the events {Bv}v∈H 3 are c1 log n-wise independent. Such a condition is satisfied if the
random bits used in the algorithm are k-wise independent, where k = dO(log d) · c1 log n = dO(log d) log n.
Note that the total number of random bits used during the first phase for all vertices is m = dO(log d) · n.
Therefore all we need is a k-wise independent random variable in {0, 1}m. By Theorem 2.4, such random
variables can be constructed with seed length O(k log m) = dO(d log d) log2 n and each random bit can be
computed in time O(k log m) = dO(d log d) log2 n.
To put everything together, we proved the following theorem regarding the LCA for MIS5:
Theorem 6.1. Let G be an undirected graph with n vertices and maximum degree d. Then there is a
dO(d log d) log3 n, dO(log d) log2 n, 1/n)-local computation algorithm which, on input a vertex v, decides if v
is in a maximal independent set. Moreover, the algorithm will give a consistent MIS for every vertex in G.
Acknowledgments
We would like to thank Tali Kaufman and Krzysztof Onak for enlightening discussions.
References
[1] N. Ailon, B. Chazelle, S. Comandur, and D. Liu. Property-preserving data reconstruction. Algorith-
mica, 51(2):160 -- 182, 2008.
[2] N. Alon. A parallel algorithmic version of the Local Lemma. Random Structures and Algorithms,
2:367 -- 378, 1991.
[3] N. Alon, L. Babai, and A. Itai. A fast and simple randomized algorithm for the maximal independent
set problem. Journal of Algorithms, 7:567 -- 583, 1986.
[4] N. Alon and J. Spencer. The Probabilistic Method. John Wiley and Sons, second edition, 2000.
[5] J. Beck. An algorithmic approach to the Lovász Local Lemma. Random Structures and Algorithms,
2:343 -- 365, 1991.
[6] B. Chazelle and C. Seshadhri. Online geometric reconstruction. In SoCG, pages 386 -- 394, 2006.
[7] M. Jha and S. Raskhodnikova. Testing and reconstruction of Lipschitz functions with applications to
data privacy. In Proc. 52nd Annual IEEE Symposium on Foundations of Computer Science, 2011.
[8] E. Kaplan, M. Naor, and O. Reingold. Derandomized constructions of k-wise (almost) independent
permutations. Algorithmica, 55(1):113 -- 133, 2009.
5Note that the space complexity of storing the pseudorandom bits dominates the space complexity of local computation for each
query.
15
[9] J. Katz and L. Trevisan. On the efficiency of local decoding procedures for error-correcting codes. In
Proc. 32nd Annual ACM Symposium on the Theory of Computing, pages 80 -- 86, 2000.
[10] F. Kuhn. Local multicoloring algorithms: Computing a nearly-optimal tdma schedule in constant time.
In STACS, pages 613 -- 624, 2009.
[11] F. Kuhn and T. Moscibroda. Distributed approximation of capacitated dominating sets. In SPAA, pages
161 -- 170, 2007.
[12] F. Kuhn, T. Moscibroda, T. Nieberg, and R. Wattenhofer. Fast deterministic distributed maximal inde-
pendent set computation on growth-bounded graphs. In DISC, pages 273 -- 287, 2005.
[13] F. Kuhn, T. Moscibroda, and R. Wattenhofer. The price of being near-sighted. In Proc. 17th ACM-SIAM
Symposium on Discrete Algorithms, pages 980 -- 989, 2006.
[14] F. Kuhn and R. Wattenhofer. On the complexity of distributed graph coloring. In Proc. 25th ACM
Symposium on Principles of Distributed Computing, pages 7 -- 15, 2006.
[15] M. Luby. A simple parallel algorithm for the maximal independent set problem. SIAM Journal on
Computing, 15(4):1036 -- 1053, 1986. Earlier version in STOC'85.
[16] S. Marko and D. Ron. Distance approximation in bounded-degree and general sparse graphs.
In
APPROX-RANDOM'06, pages 475 -- 486, 2006.
[17] A. Mayer, S. Naor, and L. Stockmeyer. Local computations on static and dynamic graphs. In Proceed-
ings of the 3rd Israel Symposium on Theory and Computing Systems (ISTCS), 1995.
[18] D. S. Mitrinovi´c. Analytic inequalities. Springer-Verlag, 1970.
[19] M. Naor and L. Stockmeyer. What can be computed locally?
SIAM Journal on Computing,
24(6):1259 -- 1277, 1995.
[20] H. N. Nguyen and K. Onak. Constant-time approximation algorithms via local improvements. In Proc.
49th Annual IEEE Symposium on Foundations of Computer Science, pages 327 -- 336, 2008.
[21] K. Onak. New Sublinear Methods in the Struggle Against Classical Problems. PhD thesis, MIT, 2010.
[22] R. Otter. The multiplicative process. Annals of mathematical statistics, 20(2):206 -- 224, 1949.
[23] M. Parnas and D. Ron. Approximating the minimum vertex cover in sublinear time and a connection
to distributed algorithms. Theoretical Computer Science, 381(1 -- 3):183 -- 196, 2007.
[24] R. Rubinfeld, G. Tamir, S. Vardi, and N. Xie. Fast local computation algorithms. In Proc. 2nd Sympo-
sium on Innovations in Computer Science, pages 223 -- 238, 2011.
[25] R. Rubinfeld, G. Tamir, S. Vardi, and N. Xie. Fast local computation algorithms. Technical report,
April 2011. http://arxiv.org/abs/1104.1377.
[26] M. E. Saks and C. Seshadhri. Local monotonicity reconstruction. SIAM Journal on Computing,
39(7):2897 -- 2926, 2010.
16
MAXIMAL INDEPENDENT SET: PHASE 1
Input: a graph G and a vertex v ∈ V
Output: {"true", "false", "⊥"}
For i from 1 to r = 20d log d
(a) If MIS(v, i) = "selected"
return "true"
(b) Else if MIS(v, i) = "deleted"
return "false"
(c) Else
return "⊥"
MIS(v, i)
Input: a vertex v ∈ V and a round number i
Output: {"selected", "deleted", "⊥"}
1.
If v is marked "selected" or "deleted"
return "selected" or "deleted", respectively
2. For every u in N (v)
If MIS(u, i − 1) = "selected"
mark v as " deleted" and return "deleted"
3. v chooses itself independently with probability 1
2d
If v chooses itself
(i) For every u in N (v)
If u is marked "⊥", u chooses itself independently with probability 1
2d
(ii) If v has a chosen neighbor
return "⊥"
(iii) Else
mark v as "selected" and return "selected"
Else
return "⊥"
Figure 4: Local computation algorithm for MIS: Phase 1
[27] S. Yekhanin. Locally decodable codes. In 6th International Computer Science Symposium in Russia,
pages 289 -- 290, 2011.
[28] Y. Yoshida, Y. Yamamoto, and H. Ito. An improved constant-time approximation algorithm for maxi-
mum matchings. In Proc. 41st Annual ACM Symposium on the Theory of Computing, pages 225 -- 234,
2009.
A Pseudocode of the LCA for Maximal Independent Set
In this section we present the pseudocode of the LCA for Maximal Independent Set. This is taken from [25]
with slight modifications and we also refer interested readers to [25] for detailed description and analysis of
the algorithm.
17
MISB(v, i)
Input: a vertex v ∈ V and a round number i
Output: {"picked", "⊥"}
1.
If v is marked "picked"
return "picked"
2. v chooses itself independently with probability 1
2d
If v chooses itself
(i) For every u in N (v)
u chooses itself independently with probability 1
2d
(ii) If v has a chosen neighbor
return "⊥"
(iii) Else
mark v as "picked" and return "picked"
Else
return "⊥"
Figure 5: Algorithm MISB
MAXIMAL INDEPENDENT SET: PHASE 2
Input: a graph G and a vertex v ∈ V
Output: {"true", "false"}
1. Run BFS starting from v to grow a connected component of surviving vertices
(If a vertex u is in the BFS tree and w ∈ N (u) in G, then w is in the BFS tree
if and only if running the first phase LCA on w returns "⊥")
2.
(Run the greedy search algorithm on the connected component for an MIS)
Set S = ∅
Scan all the vertices in the connected component in order
If a vertex u is not deleted
add u to S
delete all the neighbors of u
3.
If v ∈ S
return "true"
else "false"
Figure 6: Local computation algorithm for MIS: Phase 2
18
|
1905.04941 | 1 | 1905 | 2019-05-13T10:08:07 | An improved algorithm for the submodular secretary problem with a cardinality constraint | [
"cs.DS"
] | We study the submodular secretary problem with a cardinality constraint. In this problem, $n$ candidates for secretaries appear sequentially in random order. At the arrival of each candidate, a decision maker must irrevocably decide whether to hire him. The decision maker aims to hire at most $k$ candidates that maximize a non-negative submodular set function. We propose an $(\mathrm{e} - 1)^2 / (\mathrm{e}^2 (1 + \mathrm{e}))$-competitive algorithm for this problem, which improves the best one known so far. | cs.DS | cs |
An improved algorithm for the submodular secretary
problem with a cardinality constraint
Kaito Fujii
University of Tokyo
[email protected]
May 14, 2019
Abstract
We study the submodular secretary problem with a cardinality constraint. In this problem, n can-
didates for secretaries appear sequentially in random order. At the arrival of each candidate, a decision
maker must irrevocably decide whether to hire him. The decision maker aims to hire at most k candidates
that maximize a non-negative submodular set function. We propose an (e − 1)2/(e2(1 + e))-competitive
algorithm for this problem, which improves the best one known so far.
1
Introduction
In the classical secretary problem, a decision maker aims to hire the best one out of n candidates. Each
candidate has a real value that expresses his skill. The decision maker knows only the number n of candidates
in the beginning. The candidates appear in random order one by one, and at the arrival of each candidate,
his value is revealed. Just after observing the value, the decision maker must decide whether to hire this
candidate or not. This decision is irrevocable, and the decision maker can hire only one candidate. The
goal of this problem is to hire the candidate with the largest value with probability as high as possible. An
asymptotically optimal strategy for this problem was stated by Dynkin [6]. This strategy ignores the first
⌊n/e⌋ candidates without hiring and hires the first candidate better than these first ⌊n/e⌋ candidates. This
algorithm hires the best one with probability at least 1/e [6].
A multiple-choice variant was proposed by Kleinberg [15]. In contrast to the classical secretary problem,
k candidates can be selected in this variant, where k ∈ Z>0 is the maximum number of hired candidates. The
decision maker knows n and k in advance. Let V be the set of all candidates and w : V → R≥0 a non-negative
weight of each candidate. At each arrival of candidates, the decision maker must decide whether to hire the
candidate or not according to the revealed value of the candidate. The objective is to maximize the sum of
values of the hired candidates, that is, Pv∈S w(v), where S ⊆ V is the set of the hired candidates.
As a further extension of the multiple-choice secretary problem, Bateni et al. [3] and Gupta et al. [13]
proposed submodular secretary problems, in which the objective function f : 2V → R≥0 is not the sum
of values of the hired candidates, but a submodular set function. A set function f : 2V → R≥0 is called
submodular if it satisfies f (S ∪ {v}) − f (S) ≥ f (T ∪ {v}) − f (T ) for all S ⊆ T ⊆ V and v ∈ V \ T . We
assume that for any subset S of candidates that already appeared, the value of f (S) can be computed in O(1)
time. Among various constraints for submodular secretary problems, in this study we focus on cardinality
constraints, under which the decision maker can hire at most k candidates out of n candidates.
Submodular secretary problems have many applications. A key application is online auctions [2], in
which the seller decides who gets products in an online fashion out of buyers declaring their bids one by one.
One of the goals is to design a mechanism that maximizes the utility of the agents. Submodularity of utility
functions is often assumed due to its equivalence to the property of diminishing returns. Also, submodular
1
secretary problems are applied to machine learning tasks such as stream-based active learning [12] and the
interpretation of neural networks [7].
The quality of an algorithm for submodular secretary problems is evaluated by its competitive ratio. Let
A(f, σ) ⊆ V be the (possibly randomized) output of algorithm A, where f : 2V → R≥0 is the objective
function and σ ∈ Σn is a permutation of n candidates. The competitive ratio α(A) of an algorithm A is
defined as
α(A) = inf
f
E[f (A(f, σ))]
maxS ∗⊆V f (S∗)
,
where the expectation is taken over the random permutation and random factors in the algorithm.
1.1 Related work
For the multiple-choice secretary problem, that is, the case where f is a linear function, (1 − O(1/√k))-
competitive [15] and (1/e)-competitive [1] algorithms were proposed.
For the submodular secretary problem with a cardinality constraint, Gupta et al. [13] proposed a 1/1417-
competitive algorithm and Bateni et al. [3] proposed a 1/(8e2)-competitive algorithm. Several studies of
submodular secretary problems assume the monotonicity of the objective function, which is defined as the
condition that f (S) ≤ f (T ) for all S ⊆ T ⊆ V . Under the monotonicity assumption, Bateni et al. [3] proposed
a 0.090-competitive algorithm and Feldman et al. [11] showed this algorithm achieves 0.170-approximation.
The best one for the monotone case is (1 − O(1/√k))/e-approximation with exponential running time by
Kesselheim and Tonnis [14].
Submodular secretary problems under more general constraints have been considered in several studies
such as Ma et al. [16] and Feldman and Zenklusen [10]. Secretary problems with a set function lacking
submodularity also have been studied, such as monotone subadditive functions by Rubinstein and Singla
[18] and monotone functions with bounded supermodular degree by Feldman and Izsak [9].
Submodular maximization with a cardinality constraint in the offline setting has been studied extensively.
In the offline setting, the ground set is given in advance and the objective value of any subset can be
obtained from the beginning. The best algorithm known so far achieves 0.385-approximation by Buchbinder
and Feldman [4]. Under the monotonicity assumption, the greedy algorithm is known to achieve (1 − 1/e)-
approximation [17].
1.2 Our result
In this study, we propose an algorithm with an improved competitive ratio for the submodular secretary
problem with a cardinality constraint. The objective function is assumed to be non-negative and submodular,
but not necessarily monotone. Our algorithm is based on the one proposed for the monotone case by Bateni
et al. [3]. We slightly modify their algorithm so that it works for the non-monotone case. In the analysis
of the competitive ratio, we use a lemma proved by Buchbinder et al. [5], which was originally used for the
offline setting. While Buchbinder et al. [5] designed a randomized algorithm and applied this lemma to the
analysis of its approximation ratio, we utilize a random factor of the ordering of candidates. The resulting
algorithm is (e − 1)2/(e2(1 + e)) ≈ 0.107-competitive. To the best of our knowledge, this competitive ratio
improves the best one known so far, which is 1/(8e2) ≈ 0.0169 [3].
2 Preliminaries
Let V be a finite set of size n. The elements of V arrive one by one in random order, i.e., the order of
V is chosen uniformly at random out of n! permutations. The non-negative submodular objective function
f : 2V → R≥0 can be accessed through an value oracle, which returns the value of f (S) in O(1) time for any
subset S of those who already arrived. At the arrival of element v ∈ V , the algorithm must decide whether
to add it to the solution or reject it irrevocably. The algorithm can select at most k elements. We define the
marginal gain of adding an element v ∈ V to the current solution S ⊆ V as f (vS) := f (S ∪ {v}) − f (S).
2
Algorithm 1 The modified classical secretary algorithm
Require: A randomly ordered elements v1,··· , vn.
1: Generate n numbers from the uniform distribution on [0, 1] and sort them in ascending order. Let (ti)n
i=1
be these sorted numbers.
2: Regard ti as the arrival time of the ith element vi.
3: Ignore all elements that arrive before time 1/e.
4: if there is no element that arrives before time 1/e then
5:
Select the first element with probability 1/(et1). Otherwise, no element is selected.
6: else
7:
8:
Let θ be the largest value of the elements that arrive before time 1/e.
Select the first element with value at least θ.
A continuous-time model [11] is a problem setting equivalent to the random order model. In this model,
each element is assigned to an arrival time that is generated from the uniform distribution on [0, 1], and the
decision maker can observe each element at its arrival time. We can transform a problem instance in the
random order model to one in the continuous-time model. Suppose we generate n random numbers from
the uniform distribution on [0, 1] and sort them in ascending order in advance. By assigning these values to
each element in order of arrival, we can transform the random order model to the continuous-time model.
3 The modified classical secretary algorithm
In this section, we describe a modified version of the classical secretary algorithm, which is used as a
subroutine of our proposed method. The modified classical secretary algorithm is based on the continuous-
time version of the classical secretary problem [11], which ignores all elements that arrive before time 1/e
and selects the first element with value higher than all ignored elements. They proved that their method
selects the best one with probability at least 1/e. However, we need another property: The probability that
the algorithm selects each element must be upper-bounded. Their method does not satisfy this property if n
is small. Thus we make a small modification to this algorithm, and obtain the one satisfying this property.
This modification is applied only when there exists no element that arrives before time 1/e. In this case,
the modified algorithm selects the first element with probability 1/(et1), where t1 is the time when the first
element arrives, while the original one always selects it. This modified algorithm is described in Algorithm 1.
Lemma 1. Algorithm 1 is 1/e-competitive and for each element v ∈ V , the probability that Algorithm 1
selects v is at most 1/e.
Proof. Let v∗ ∈ V be an element with the largest value. Suppose v∗ appears after time 1/e and its arrival
time is t∗ ∈ [1/e, 1]. Below, we fix the arrival time t∗ and consider the event that v∗ is selected by the
algorithm. This event happens if one of the following two conditions holds.
• v∗ is the first element and the algorithm decides to select v∗.
• v∗ is not the first element, and v appears before time 1/e, where v is the best element among those
who appear before time t∗.
The probability that v∗ is the first element is equal to the probability that all other elements arrive after
time t∗, then it is (1− t∗)n−1. Since the algorithm selects the first element with probability 1/(et1), the first
condition holds with probability (1 − t∗)n−1/(et∗).
To consider the second condition, we fix the set of elements S ⊆ V that arrive before time t∗. Then the
largest element among them, v ∈ argmaxv∈Sw(v), is also determined. Under this condition, the probability
that v arrives before time 1/e is 1/(et∗) since the arrival time of v conforms to the uniform distribution on
3
Algorithm 2 Our proposed method
Input: A randomly ordered elements v1,··· , vn.
1: Generate n numbers from the uniform distribution on [0, 1] and sort them in ascending order. Let (ti)n
i=1
be these sorted numbers. Regard ti as the arrival time of the ith element vi.
that arrive from time (l − 1)/k to time l/k for each l ∈ [k].
2: Partition V into k segments V1,··· , Vk, where Vl = {vi ∈ V ti ∈ [(l − 1)/k, l/k]} is the set of elements
3: Let S0 := ∅.
4: for each l = 1,··· , k do
5: Apply Algorithm 1 to Vl with weight f (vSl−1) for each v ∈ Vl.
6:
7:
8:
if Algorithm 1 selects an element sl ∈ Vl and f (slSl−1) ≥ 0 then
else
Sl ← Sl−1 ∪ {sl}.
Sl ← Sl−1.
9:
10: return Sk.
[0, t∗]. Since this holds for any S, the second condition holds with probability {1 − (1 − t∗)n−1}/(et∗). In
total, the probability that the algorithm selects v∗ is
(1 − t∗)n−1
et∗
+
1 − (1 − t∗)n−1
et∗
dt∗
1/e
Z 1
=Z 1
1/e
1
e
=
1
et∗ dt∗
Then this algorithm is 1/e-competitive.
Let v ∈ V be any element and t be its arrival time. If v is selected, one of the following two conditions
holds.
• v is the first element and the algorithm decides to select v.
• v is not the first element, v appears before time 1/e, and the value of v is less than that of v, where v
is the best element among those who appear before time t.
Since the second condition is stronger than the second condition for the optimal element v∗, the probability
that one of these conditions is satisfied is less than that of v∗. Therefore, the probability that any element
is selected is at most 1/e.
4 Algorithm and analysis
In this section, we illustrate our proposed method and provide an analysis of its competitive ratio.
Our proposed method utilizes the idea of the continuous-time model. To transform the random order
model to the continuous-time model, we generate n random numbers from the uniform distribution on [0, 1]
and sort them in ascending order in advance. Let t1 ≤ ··· ≤ tn be these numbers. By regarding ti as the
arrival time of the ith element, we can obtain an instance of the continuous-time model. We partition the
sequence of the elements into k segments V1,··· , Vk by separating [0, 1] into equal-length time windows.
Then we apply Algorithm 1 to each segment with regarding the marginal gain f (vSl−1) as the weight for
element v ∈ Vl, where Sl−1 is the solution just before segment Vl. If Algorithm 1 selects an element sl ∈ Vl
and its marginal gain is non-negative, we add it to the solution. The detailed description of the algorithm is
provided in Algorithm 2.
In the proof, we utilize the following lemmas.
4
Lemma 2 (Lemma 2.2 of [8]). Let f : 2V → R be submodular. Denote by A(p) a random subset of A where
each element appears with probability p (not necessarily independently). Then, E[f (A(p))] ≥ (1 − p)f (∅) +
p · f (A).
Lemma 3 (Lemma 2.2 of [5]). Let f : 2V → R be submodular. Denote by A(p) a random subset of A
where each element appears with probability at most p (not necessarily independently). Then, E[f (A(p))] ≥
(1 − p)f (∅).
Theorem 4. Algorithm 2 is (e − 1)2/(e2(1 + e))-competitive for any non-negative submodular objective
function and cardinality constraint.
Proof. Let S∗ ∈ argmaxS : S≤kf (S) be an optimal solution. Let U be a set with the largest objective value
that has at most one element in each partition, i.e., U ∈ argmax{f (U ) U ⊆ S∗, ∀l ∈ [k] : U ∩ Vl ≤ 1}. Let
U be a random subset of S∗ obtained by selecting an element uniformly at random from each non-empty
S∗ ∩ Vl where l ∈ [k]. From the definition, we have E[f (U )] ≥ E[f ( U )].
with probability 1/k for each l ∈ [k]. Then we have
From the independence between arrival times of elements, each element of S∗ is included in partition Vl
Pr(Vl ∩ S∗ 6= ∅) = 1 −(cid:18)1 −
1
k(cid:19)S ∗
for each l ∈ [k]. From the linearity of expectation, we have
k
=
E[ U] = E" k
1Vl∩S ∗6=∅#
Xl=1
Xl=1
Pr(Vl ∩ S∗ 6= ∅)
k(cid:19)S ∗!
= k 1 −(cid:18)1 −
e(cid:19) .
≥ S∗(cid:18)1 −
1
1
By considering the randomness of Vl for all l ∈ [k], we can see that each element in S∗ is included in U
with the same probability, then we can define p := Pr(v ∈ U ) for all v ∈ S∗. Since E[ U] = pS∗, we have
p ≥ 1 − 1/e. By applying Lemma 2, we obtain
E[f (U )] ≥ E[f ( U )]
≥ (1 − p)f (∅) + pf (S∗)
≥ (1 − 1/e)f (S∗).
(1)
Fix partitions V1,··· , Vk and U . We consider the randomness of the ordering of each partition. Since
Algorithm 1 selects each element in Vl with probability at most 1/e for each l ∈ [k] from Lemma 1, we have
Pr(v ∈ Sk V1,··· , Vk, U ) ≤ 1/e for all v ∈ V . Let g(Sk) = f (Sk ∪ U ) − f (U ). By applying Lemma 3 to g,
we have
E[f (Sk ∪ U )V1,··· , Vk, U ] = E[g(Sk)V1,··· , Vk, U ]
≥ (1 − 1/e)g(∅)
= (1 − 1/e)f (U ).
(2)
We consider the marginal gain of the algorithm obtained in the lth segment by fixing Sl−1. If U ∩ Vl 6= ∅, it is
a singleton, and let ul be an element in U ∩ Vl. If U ∩ Vl = ∅, let ul be a dummy element that does not affect
5
the objective value. Define sl similarly as an element in Sk ∩ Vl or a dummy element. Since Algorithm 1 is
1/e-competitive and elements with negative marginal gain are not selected, we have
E[f (slSl−1)V1,··· , Vk, U, Sl−1] ≥
1
e
max{f (ulSl−1), 0}.
(3)
j=1 Vj(cid:17). Combining the above inequalities, we obtain
Let Ul = U ∩(cid:16)Sl
E[f (Sk) V1,··· , Vk, U ] =
E[f (slSl−1) V1,··· , Vk, U ]
k
k
=
Xl=1
Xl=1
Xl=1
Xl=1
≥
= E(cid:20) 1
≥
k
k
e
V1,··· , Vk, U(cid:21)
E[E[f (slSl−1) V1,··· , Vk, U, Sl−1] V1,··· , Vk, U ]
max{f (ulSl−1), 0}(cid:12)(cid:12)(cid:12)(cid:12)
E(cid:20) 1
f (ulSk ∪ Ul−1)(cid:12)(cid:12)(cid:12)(cid:12)
E(cid:20) 1
e {f (Sk ∪ U ) − f (Sk)}(cid:12)(cid:12)(cid:12)(cid:12)
E[f (Sk) V1,··· , Vk, U ] = E(cid:20) 1
V1,··· , Vk, U(cid:21)
V1,··· , Vk, U(cid:21) .
V1,··· , Vk, U(cid:21)
1 + e
e
f (Sk ∪ U )(cid:12)(cid:12)(cid:12)(cid:12)
e(cid:19) f (U ).
1
≥
1
1 + e(cid:18)1 −
(due to (3))
(due to the submodularity)
By a simple calculation, we have
(due to (2))
By taking the expectation about V1,··· , Vk and U and substituting (1), we obtain
E[f (Sk)] ≥
1
1 + e(cid:18)1 −
1
e(cid:19)2
f (S∗).
Acknowledgement
This study is supported by JSPS KAKENHI Grant Number JP 18J12405.
References
[1] Moshe Babaioff, Nicole Immorlica, David Kempe, and Robert Kleinberg. A knapsack secretary problem
with applications. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and
Techniques (APPROX), pages 16 -- 28, 2007.
[2] MohammadHossein Bateni. Secretary problems and online auctions. In Encyclopedia of Algorithms,
pages 1910 -- 1913. 2016.
[3] MohammadHossein Bateni, Mohammad Taghi Hajiaghayi, and Morteza Zadimoghaddam. Submodular
secretary problem and extensions. ACM Trans. Algorithms, 9(4):32:1 -- 32:23, 2013.
6
[4] Niv Buchbinder and Moran Feldman. Constrained submodular maximization via a non-symmetric
technique. CoRR, abs/1611.03253, 2016.
[5] Niv Buchbinder, Moran Feldman, Joseph Naor, and Roy Schwartz. Submodular maximization with car-
dinality constraints. In Proceedings of the 25th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 1433 -- 1452, 2014.
[6] E. B. Dynkin. The optimum choice of the instant for stopping a markov process. Soviet Math. Dokl.,
(4):627 -- 629, 1963.
[7] Ethan Elenberg, Alexandros G Dimakis, Moran Feldman, and Amin Karbasi. Streaming weak submod-
ularity: Interpreting neural networks on the fly. In Advances in Neural Information Processing Systems
(NIPS) 30, pages 4047 -- 4057. 2017.
[8] Uriel Feige, Vahab S. Mirrokni, and Jan Vondr´ak. Maximizing non-monotone submodular functions.
SIAM J. Comput., 40(4):1133 -- 1153, 2011.
[9] Moran Feldman and Rani Izsak. Building a good team: Secretary problems and the supermodular
degree. In Proceedings of the 28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 1651 -- 1670, 2017.
[10] Moran Feldman and Rico Zenklusen. The submodular secretary problem goes linear. SIAM J. Comput.,
47(2):330 -- 366, 2018.
[11] Moran Feldman, Joseph Naor, and Roy Schwartz. Improved competitive ratios for submodular secretary
In Approximation, Randomization, and Combinatorial Optimization.
problems (extended abstract).
Algorithms and Techniques (APPROX), pages 218 -- 229, 2011.
[12] Kaito Fujii and Hisashi Kashima. Budgeted stream-based active learning via adaptive submodular
maximization. In Advances in Neural Information Processing Systems (NIPS) 29, pages 514 -- 522, 2016.
[13] Anupam Gupta, Aaron Roth, Grant Schoenebeck, and Kunal Talwar. Constrained non-monotone
submodular maximization: Offline and secretary algorithms. In Proceedings of the 6th International
Workshop on Internet and Network Economics (WINE), pages 246 -- 257, 2010.
[14] Thomas Kesselheim and Andreas Tonnis. Submodular secretary problems: Cardinality, matching, and
linear constraints. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and
Techniques (APPROX), pages 16:1 -- 16:22, 2017.
[15] Robert D. Kleinberg. A multiple-choice secretary algorithm with applications to online auctions. In
Proceedings of the 16th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 630 -- 631,
2005.
[16] Tengyu Ma, Bo Tang, and Yajun Wang. The simulated greedy algorithm for several submodular matroid
secretary problems. Theory Comput. Syst., 58(4):681 -- 706, 2016.
[17] George L. Nemhauser, Laurence A. Wolsey, and Marshall L. Fisher. An analysis of approximations for
maximizing submodular set functions - I. Math. Program., 14(1):265 -- 294, 1978.
[18] Aviad Rubinstein and Sahil Singla. Combinatorial prophet inequalities.
In Proceedings of the 28th
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1671 -- 1687, 2017.
7
|
1408.4490 | 2 | 1408 | 2016-10-05T21:07:41 | Tractable Pathfinding for the Stochastic On-Time Arrival Problem | [
"cs.DS"
] | We present a new and more efficient technique for computing the route that maximizes the probability of on-time arrival in stochastic networks, also known as the path-based stochastic on-time arrival (SOTA) problem. Our primary contribution is a pathfinding algorithm that uses the solution to the policy-based SOTA problem---which is of pseudo-polynomial-time complexity in the time budget of the journey---as a search heuristic for the optimal path. In particular, we show that this heuristic can be exceptionally efficient in practice, effectively making it possible to solve the path-based SOTA problem as quickly as the policy-based SOTA problem. Our secondary contribution is the extension of policy-based preprocessing to path-based preprocessing for the SOTA problem. In the process, we also introduce Arc-Potentials, a more efficient generalization of Stochastic Arc-Flags that can be used for both policy- and path-based SOTA. After developing the pathfinding and preprocessing algorithms, we evaluate their performance on two different real-world networks. To the best of our knowledge, these techniques provide the most efficient computation strategy for the path-based SOTA problem for general probability distributions, both with and without preprocessing. | cs.DS | cs |
Tractable Pathfinding for the
Stochastic On-Time Arrival Problem
Mehrdad Niknami1 and Samitha Samaranayake2
1 Electrical Engineering and Computer Science, UC Berkeley
2 School of Civil and Environmental Engineering, Cornell University
Abstract. We present a new and more efficient technique for computing
the route that maximizes the probability of on-time arrival in stochastic
networks, also known as the path-based stochastic on-time arrival (SOTA)
problem. Our primary contribution is a pathfinding algorithm that uses
the solution to the policy-based SOTA problem-which is of pseudo-
polynomial-time complexity in the time budget of the journey-as a
search heuristic for the optimal path. In particular, we show that this
heuristic can be exceptionally efficient in practice, effectively making it
possible to solve the path-based SOTA problem as quickly as the policy-
based SOTA problem. Our secondary contribution is the extension of
policy-based preprocessing to path-based preprocessing for the SOTA
problem. In the process, we also introduce Arc-Potentials, a more efficient
generalization of Stochastic Arc-Flags that can be used for both policy- and
path-based SOTA. After developing the pathfinding and preprocessing
algorithms, we evaluate their performance on two different real-world
networks. To the best of our knowledge, these techniques provide the
most efficient computation strategy for the path-based SOTA problem for
general probability distributions, both with and without preprocessing.
1
Introduction
Modern advances in graph theory and empirical computational power have
essentially rendered deterministic point-to-point routing a solved problem. While
the ubiquity of routing and navigation tools in our everyday lives is a testament
to the success and usefulness of deterministic routing technology, inaccurate
predictions remain a fact of life, resulting in missed flights, late arrivals to
meetings, and failure to meet delivery deadlines. Recent research in transportation
engineering, therefore, has focused on the collection of traffic data and the
incorporation of uncertainty into traffic models, allowing for the optimization of
relevant reliability metrics desirable for the user.
The point-to-point stochastic on-time arrival problem [1], or SOTA for short,
concerns itself with this reliability aspect of routing. In the SOTA problem,
the network is assumed to have uncertainty in the travel time across each
link, represented by a strictly positive random variable. The objective is then to
maximize the probability of on-time arrival when traveling between a given origin-
destination pair with a fixed time budget.3 It has been shown that the SOTA
solution can appreciably increase the probability of on-time arrival compared to
the classical least expected travel time (LET) path [3], motivating the search for
efficient solutions to this problem.
1.1 Variants
There exist two primary variants of the SOTA problem. The path-based SOTA
problem, which is also referred to as the shortest-path problem with on-time
arrival reliability (SPOTAR) [4], consists of finding the a-priori most reliable
path to the destination. The policy-based SOTA problem, on the other hand,
consists of computing a routing policy-rather than a fixed path-such that,
at every intersection, the choice of the next direction depends on the current
state (i.e., the remaining time budget).4 While a policy-based approach provides
better reliability when online navigation is an option, in some situations it can
be necessary to determine the entire path prior to departure.
The policy-based SOTA problem, which is generally solved in discrete-time,
can be solved via a successive-approximation algorithm, as shown by Fan and Nie
[5]. This approach was subsequently improved by Samaranayake et al. [3] to a
pseudo-polynomial-time label-setting algorithm based on dynamic-programming
with a sequence of speedup techniques and the use of zero-delay convolution [6, 7].
It was then demonstrated in Sabran et al. [8] that graph preprocessing techniques
such as Reach [9] and Arc-Flags [10] can be used to further reduce query times
for this problem.
In contrast with the policy-based problem, however, no polynomial-time solu-
tion is known for the general path-based SOTA problem [4]. In the special case
of normally-distributed travel times, Nikolova et al. [11] present an 𝑂(𝑛𝑂(𝑙𝑜𝑔 𝑛))-
algorithm for computing exact solutions, while Lim et al. [12] present a poly-
logarithmic-time algorithm for approximation solutions. To allow for more general
probability distributions, Nie and Wu [4] develop a label-correcting algorithm
that solves the problem by utilizing the first-order stochastic dominance prop-
erty of paths. While providing a solution method for general distributions, the
performance of this algorithm is still insufficient to be of practical use in many
real-world scenarios; for example, while the stochastic dominance approach pro-
vides a reasonable computation time (on the order of half a minute per instance)
for networks of a few hundred to a thousand vertices, it fails to perform well on
metropolitan road networks, which easily exceed tens of thousands of vertices.
In contrast, our algorithm easily handles networks of tens of thousands of edges
in approximately the same amount of time without any kind of preprocessing.5
3 The target objective can in fact be generalized to utility functions other than the
probability of on-time arrival [2] with little effect on our algorithms, but for our
purposes, we limit our discussion to this scenario.
4 In this article, we only consider time-invariant travel-time distributions. The problem
can be extended to incorporate time-varying distributions as discussed in [3].
5 Parmentier and Meunier [13] have concurrently also developed a similar approach
concerning stochastic shortest paths with risk measures.
With preprocessing, our techniques further reduce the running time to less than
half a second, making the problem tractable for larger networks.6
1.2 Contributions
Our primary contribution in this article is a practically efficient technique for
solving the path-based SOTA problem, based on the observation that the solution
to the policy-based SOTA problem is in practice itself an extremely efficient
heuristic for solving the path-based problem.
Our secondary contribution is to demonstrate how graph preprocessing can
be used to speed up the computation of the policy heuristic, and thus the
optimal path, while maintaining correctness.7 Toward this goal, we present Arc-
Potentials, a more efficient generalization of the existing preprocessing technique
known as Stochastic Arc-Flags that can be used for both policy- and path-based
preprocessing.
After presenting these techniques, we evaluate the performance of our al-
gorithms on two real-world networks while comparing the trade-off between
their scalability (in terms of memory and computation time) and the speedups
achieved. Our techniques, to the best of our knowledge, provide the most efficient
computation strategy for the path-based SOTA problem with general probability
distributions, both with and without preprocessing.
2 Preliminaries
We are given a stochastic network in the form of a directed graph 𝐺 = (𝑉, 𝐸)
where each edge (𝑖, 𝑗) ∈ 𝐸 has an associated probability distribution 𝑝𝑖𝑗(·)
representing the travel time across that edge.8 The source is denoted by 𝑠 ∈ 𝑉 ,
the destination by 𝑑 ∈ 𝑉 , and the travel time budget by 𝑇 ∈ R+.
For notational simplicity, we present the SOTA problem in continuous-time
throughout this article, with the understanding that the algorithms are applied
after discretization with a time interval of 𝛥𝑡.
Definition 1 (SOTA Policy). Let 𝑢𝑖𝑗(𝑡) be the probability of arriving at the
destination 𝑑 with time budget 𝑡 when first traversing edge (𝑖, 𝑗) ∈ 𝐸 and sub-
sequently following the optimal policy. Let 𝛿𝑖𝑗 > 0 be the minimum travel time
along edge (𝑖, 𝑗), i.e. min{𝜏 : 𝑝𝑖𝑗(𝜏 ) > 0}. Then, the on-time arrival probability
𝑢𝑖(𝑡) and the policy (optimal subsequent node) 𝑤𝑖(𝑡) at node 𝑖, can be defined
via the dynamic programming equations below [1]. Note that 𝛥𝑡 must satisfy
𝛥𝑡 ≤ 𝛿𝑖𝑗 ∀(𝑖, 𝑗) ∈ 𝐸.
6 It should be noted that the largest network we consider only has approximately
71,000 edges and is still much smaller than networks used to benchmark deterministic
shortest path queries, which can have millions of edges [14].
7 As explained later, there is a potential pitfall that must be avoided when the
preprocessed policy is to be used as a heuristic for the path.
8 We assume that at most one edge exists between any pair of nodes in each direction.
∫︁ 𝑡
𝑢𝑖𝑗(𝑡) =
𝑢𝑑(·) = 1
𝛿𝑖𝑗
𝑢𝑗(𝑡 − 𝜏 )𝑝𝑖𝑗(𝜏 ) 𝑑𝜏
𝑢𝑖(𝑡) = max
𝑗: (𝑖,𝑗)∈𝐸
𝑤𝑖(𝑡) = arg max
𝑗: (𝑖,𝑗)∈𝐸
𝑢𝑖𝑗(𝑡)
𝑢𝑖𝑗(𝑡)
The solution to the policy-based SOTA problem can be obtained by solving this
system of equations using dynamic programming as detailed in [3]. This requires
evaluating a set of convolution integrals. The computation of the policy 𝑢𝑠(·) for
a source 𝑠 and time budget 𝑇 using direct convolution takes 𝑂(𝐸𝑇 2) time, as
computing 𝑢𝑠(𝑇 ) could in the worst case require convolutions of length 𝑂(𝑇 ) for
each edge in the graph. However, by using an online convolution technique known
as zero-delay convolution (ZDC) [15, 6], the time complexity can be reduced
to 𝑂(𝐸𝑇 log2 𝑇 ). Justifications for the results and time complexity, including
details on how to apply ZDC to the SOTA problem, can be found in [3, 7].
Assumptions. Our work, as with other approaches to both the policy-based
and path-based SOTA problems, makes a number of assumptions about the
nature of the travel time distributions. The three major assumptions are that the
travel time distributions are (1) time invariant, (2) exogenous (not impacted by
individual routing choices), and (3) independent. The time-invariance assumption-
which prevents accounting for traffic variations throughout the day-can be
relaxed under certain conditions as described in [3]. Furthermore, the exogeneity
assumption is made even in the case of deterministic shortest path problems.
This leaves the independence assumption as a major concern for this problem.
It might, in fact, be possible to partially relax this assumption [3] to allow for
conditional distributions at the cost of increasing the computation time by a factor
linear in the number of states to be conditioned on. (If we assume the Markov
property for road networks, the number of conditioning states becomes the in-
degree of each vertex, a small enough constant that may make generalizations
in this direction practical.) Nevertheless, we will only focus on the independent
setting and make no claim to have solved the path-based SOTA problem in full
generality, as the problem already lacks efficient solution methods even in this
simplified setting. Our techniques should, however, provide a foundation that
allows for relaxing these assumptions in the future.
3 Path-Based SOTA
In the deterministic setting, efficient solution strategies (from Dijkstra's algorithm
to state-of-the-art solutions) generally exploit the sub-path optimality property:
namely, the fact that any optimal route to a destination node 𝑑 that includes
some intermediate node 𝑖 necessarily includes the optimal path from 𝑖 to 𝑑.
Unfortunately, this does not hold in the stochastic setting. Furthermore, blind
enumeration of all possible paths in the graph is absolutely intractable for all but
the most trivial networks, as the number of simple paths grows exponentially
with the number of nodes in the graph. Naturally, this leads us to seek a heuristic
Algorithm 1 Algorithm for computing the optimal SOTA path
Notation: * is the convolution operator and ‖ is the concatenation operator
for all 𝑖 ∈ 𝑉 , 0 ≤ 𝑡 ≤ 𝑇 do
Compute the optimal policy's reliability10 𝑢𝑖(𝑡)
𝑄 ← PriorityQueue()
𝑄.Push(𝑢𝑠(𝑇 ), ([1.0], [𝑠]))
while 𝑄 is not empty do
(𝑟, (𝑞, 𝑃 )) ← 𝑄.PopMax()
𝑖 ← 𝑃 [𝑃 − 1]
if 𝑖 = 𝑑 then
return 𝑃
for all 𝑗 ∈ 𝐸.Neighbors(𝑖) do
𝑄.Push((𝑞 * 𝑢𝑗)[𝑇 ], (𝑞 * 𝐸.Cost(𝑖, 𝑗), 𝑃 ‖ [𝑗]))
return nil
◁ Push (reliability, (cost dist., initial path))
◁ Extract most reliable path so far
◁ Get the last node in the path
◁ Append new edge.
◁ No path found
to guide us toward the optimal path efficiently, while not compromising its
optimality.
3.1 Algorithm
Consider a fixed path 𝑃 from the source 𝑠 to node 𝑖. Let 𝑞𝑃
𝑠𝑖(𝑡) be the travel
time distribution along 𝑃 from node 𝑠 to node 𝑖, i.e., the convolution of the
travel time distributions of every edge in 𝑃 . Upon arriving at node 𝑖 at time 𝑡,
let the user follow the optimal policy toward 𝑑, therefore reaching 𝑑 from 𝑠 with
𝑠𝑖(𝑡)𝑢𝑖(𝑇 − 𝑡). The reliability of following path 𝑃 to node 𝑖
probability density 𝑞𝑃
and subsequently following the optimal policy toward 𝑑 is9:
∫︁ 𝑇
𝑟𝑃
𝑠𝑖(𝑇 ) =
𝑠𝑖(𝑡)𝑢𝑖(𝑇 − 𝑡) 𝑑𝑡
𝑞𝑃
0
Note that the route from 𝑠 → 𝑖 is a fixed path while that from 𝑖 → 𝑑 is a policy.
The optimal path is found via the procedure in Algorithm 1. Briefly, starting
at the source 𝑠, we add the hybrid (path + policy) solution 𝑟𝑃
𝑠𝑖(𝑇 ) for each
neighbor 𝑖 of 𝑠 to a priority queue. Each of these solutions gives an upper bound
on the solution (success probability). We then dequeue the solution with the
highest upper bound, repeating this process until a path to the destination is
found.
Essentially, algorithm 1 performs an 𝐴* search for the destination, using
the policy as a heuristic. While it is obvious that the algorithm would find the
optimal path eventually if the search were allowed to continue indefinitely, it is
less obvious that the first path found will be optimal. We show this by showing
9 The bounds of this integral can be slightly tightened through inclusion of the minimum
travel times, but this has been omitted for simplicity.
10 Can be limited to those 𝑖 and 𝑡 reachable from 𝑠 in time 𝑇 , and can be further sped
up through existing policy preprocessing techniques such as Arc-Flags.
that the policy is an admissible heuristic for the path, and consequently, by the
optimality of 𝐴* [16], the first returned path must be optimal.
Proposition 1 (Admissibility). The solution to policy-based SOTA problem is
an admissible heuristic for the optimal solution to the path-based SOTA problem
using Algorithm 1.
Proof. When finding a minimum cost path, an admissible heuristic is a heuristic
that never overestimates the actual cost [17]. In our context, since the goal is
to maximize the reliability (success probability), this corresponds to a heuristic
that never underestimates the reliability of a routing strategy. The reliability of
an optimal SOTA policy clearly provides an upper bound on the reliability of
any fixed path with the same source, destination, and travel budget. (Otherwise,
a better policy would be to simply follow the fixed path irrespective of the time
remaining, contradicting the assumption that the policy is optimal.) Therefore,
the SOTA policy is an admissible heuristic for the optimal SOTA path.
3.2 Analysis
The single dominant factor in this algorithm's (in)efficiency is the length of the
priority queue (i.e., the number of paths considered by the algorithm), which
in turn depends on the travel time distribution along each road. As long as the
number of paths considered is approximately linear in length of the optimal path,
the path computation time is easily dominated by the policy computation time,
and the algorithm finds the optimal path very quickly. In the worst-case scenario
for the algorithm, the optimal path at a node corresponds to the direction for the
worst policy at that node. Such a scenario, or even one in which the optimal policy
frequently chooses a suboptimal path, could result in a large (even exponential)
running time as well as space usage. However, it is difficult to imagine this
happening in practice. As shown later, experimentally, we came across very few
cases in which the path computation time dominated the policy computation
time, and even in those cases, they were still quite reasonable and extremely far
from such a worst-case scenario. We conjecture that such situations are extremely
unlikely to occur in real-world road networks.
An interesting open problem is to define characteristics (network structure,
shape of distributions, etc.) that guarantee pseudo-polynomial running times in
stochastic networks, similar in nature to the Highway Dimension property [18] in
deterministic networks, which guarantees logarithmic query times when networks
have a low Highway Dimension.
4 Preprocessing
In deterministic pathfinding, preprocessing techniques such as Arc-Flags [10],
reach-based routing [9, 19], contraction hierarchies [20], and transit node routing
[21] have been very successfully used to decrease query times by many orders of
magnitude by exploiting structural properties of road networks. Some of these
approaches allow for pruning the search space based solely on the destination node,
while others also take the source node into account, allowing for better pruning
at the cost of additional preprocessing. The structure of the SOTA problem,
however, makes it more challenging to apply such techniques to it. Previously,
Arc-Flags and Reach have been successfully adapted to the policy-based problem
in [8], resulting in Stochastic Arc-Flags and Directed Reach. While at first glance
one may be tempted to directly apply these algorithms to the computation of
the policy heuristic for the path-based problem, a naive application of source-
dependent pruning (such as Directed Reach or source-based Arc-Flags) can result
in an incorrect solution, as the policy needs to be recomputed for source nodes that
correspond to different source regions. This effectively limits any preprocessing
of the policy heuristic to destination-based (i.e., source-independent) techniques
such as Stochastic Arc-Flags, precluding the use of source-based approaches such
as Directed Reach for the policy computation.
With sufficient preprocessing resources (as explained in section 5.2), however,
one can improve on this through the direct use of path-based preprocessing-that
is, pruning the graph to include only those edges which may be part of the most
reliable path. This method allows us to simultaneously account for both source
and destination regions, and generally results in a substantial reduction of the
search space on which the policy needs to be computed. However, as path-based
approaches require computing paths between all ≈ 𝑉 2 pairs of vertices in the
graph, this approach may become computationally prohibitive for medium- to
large-scale networks. In such cases, we would then need to either find alternate
approaches (e.g. approximation techniques), or otherwise fall back to the less
aggressive policy-based pruning techniques, which only require computing 𝑉
separate policies (one per destination).
4.1 Efficient Path-based Preprocessing
Path-based preprocessing requires finding the optimal paths for each time budget
up to the desired time budget 𝑇 for all source-destination pairs. Naively, this
can be done by placing Algorithm 1 in a loop, executing it for all time budgets
from 1 to 𝑇 . This requires 𝑇 times the work of finding the path for a single time
budget, which is clearly prohibitive for any reasonable value of 𝑇 . However, we
can do far better by observing that many of the computations in the algorithm
are independent of the time budget and can be factored out when the path does
not change with 𝑇 .
To improve the efficiency of the naive approach in this manner, we make two
observations. First, we observe that, in Algorithm 1, only the computation of
the path's reliability (priority) in the priority queue ever requires knowledge of
the time budget. Crucially, the convolution 𝑞 * 𝐸.Cost(𝑖, 𝑗) only depends on the
maximum time budget 𝑇 for truncation purposes, which is a fixed value. This
means that the travel time distribution of any path under consideration can be
computed once for the maximum time budget, and re-used for all lower time
budgets thereafter. Second, we observe that when a new edge is appended, the
priority of the new path is the inner product of the vector 𝑞 and (the reverse of)
the vector 𝑢𝑗, shifted by 𝑇 . As noted in the algorithm itself, this quantity in fact
the convolution of the two aforementioned vectors evaluated at 𝑇 . Thus, when a
new edge is appended, instead of recomputing the inner product, we can simply
convolve the two vectors once, and thereafter look up the results instantly for
other time budgets.
Together, these two observations allow us to compute the optimal paths
for all budgets far faster than would seem naively possible, making path-based
preprocessing a practical option.
4.2 Arc-Potentials
As noted earlier, Arc-Flags, a popular method for graph preprocessing, has been
adapted to the SOTA problem as Stochastic Arc-Flags [8]. Instead of applying
it directly, however, we present Arc-Potentials, a more natural generalization of
Arc-Flags to SOTA that can still be directly applied to the policy- and path-based
SOTA problems alike, while allowing for more efficient preprocessing.
Consider partitioning the graph 𝐺 into 𝑅 regions (we choose 𝑅 = 𝑂(log 𝐸),
described below), where 𝑅 is tuned to trade off storage space for pruning accuracy.
In the deterministic setting, Arc-Flags allow us to preprocess and prune the
search space as follows. For every arc (edge) (𝑖, 𝑗) ∈ 𝐸, Arc-Flags defines a
bit-vector of length 𝑅 that denotes whether or not this arc belongs to an optimal
path ending at some node in region 𝑅. We then pre-compute these Arc-Flags,
and store them for pruning the graph at query time. (This approach has been
extended to the dynamic setting [22] in which the flags are updated with low
recomputation cost after the road network is changed.)
Sabran et al. [8] apply Arc-Flags to the policy-based SOTA problem as
follows: each bit vector is defined to represent whether or not its associated arc is
realizable, meaning that it belongs to an optimal policy to some destination in the
target region associated with each bit. The problem with this approach, however,
is that it requires computing arc-flags for all target budgets (or more, practically,
some ranges of budgets), each of which takes a considerable amount of space.
Instead, we propose a more efficient alternative 2, which we call Arc-Potentials.
Definition 2 (Arc-Potentials). For a given destination region 𝐷, we define
the arc activation potential 𝜑𝑖𝑗 of the edge from node 𝑖 to node 𝑗 to be the
minimum time budget at which the arc becomes part of an optimal policy to some
destination 𝑑 ∈ 𝐷.
The Arc-Potentials pruning algorithm only stores the "activation" potential
of every edge. As expected, this implies that for large time budgets, every edge is
potentially active. We could have further generalized the algorithm to allow for
asymptotically exact pruning at relatively low cost by simply storing the actual
potential intervals during which the arc is active, rather than merely their first
activation potential. However, in our experiments this was deemed unnecessary as
Arc-Potentials were already sufficient for significant pruning in the time budgets
of interest in our networks.
The computation of the set of realizable edges (and nodes) under a given
policy is essentially equivalent to the computation of the policy itself, except that
updates are performed in the reverse order (from the source to the destination).
The activation potentials 𝜑 can then be obtained from this set. As with Arc-Flags,
we limit the space complexity to 𝑂(𝐸𝑅) = 𝑂(𝐸 log 𝐸) by choosing 𝑅 to be
proportional log 𝐸, tuning it as desired to increase the pruning accuracy. In our
𝑅. Note, however,
experiments, we simply used a rectangular grid of size
that the preprocessing time does not depend on 𝑅, as the paths between all
≈ 𝑉 2 pairs of nodes must be eventually computed.
𝑅 × √
√
5 Experimental Results
We evaluated the performance of our algorithms on two real-world test networks:
a small San Francisco network with 2643 nodes and 6588 edges for which real-
world travel-time data was available as a Gaussian mixture model [23], and a
second (relatively larger) Luxembourg network with 30647 nodes and 71655 edges
for which travel-time distributions were synthesized from road speed limits, as
real-world data was unavailable. The algorithms were implemented in C++ (2003)
and executed on a cluster of 1.9 GHz AMD OpteronTM 6168 CPUs. The SOTA
queries were executed on a single CPU and the preprocessing was performed in
parallel as explained below.
The SOTA policies were computed as explained in [3, 7] using zero-delay
convolution with a discretization interval of 𝛥𝑡 = 1 second.11 To generate random
problem instances, we independently picked a source and a destination node
uniformly at random from the graph and computed the least expected travel-time
(LET) path between them. We then evaluated our pathfinding algorithm for
budgets chosen uniformly at random from the 5th to 95th percentile of LET path
travel times (those of practical interest) on 10, 000 San Francisco and 1000
Luxembourg problems instances.
First, we discuss the speed of our pathfinding algorithm, and afterward, we
evaluate the effectiveness and scalability of our preprocessing strategies.
5.1 Evaluation
We first evaluate the performance of our path-based SOTA algorithm without
any graph preprocessing. Experimental results, as can be seen in Figure 1, show
that the run time of our solution is dominated by the time taken to obtain the
solution to the policy-based SOTA problem, which functions as a search heuristic
for the optimal path.
The stochastic-dominance (SD) approach [4], which to our knowledge is
the fastest published solution for the path-based SOTA problem with general
probability distributions, takes, on average, between 7 and 18 seconds (depending
on the variance of the link travel time distributions) to compute the optimal path
11 Recall that we must have 𝛥𝑡 ≤ min(𝑖,𝑗)∈𝐸 𝛿𝑖𝑗, which is ≈ 1 sec for our networks.
Fig. 1: Running time of the pathfinding algorithm as a function of the travel
time budget for random unpruned (i.e., non-preprocessed) instantiations of each
network. We can see that the path computation time is dominated by the policy
computation time, effectively reducing the path-based SOTA problem to the
policy-based SOTA problem in terms of computation time.
for 100 time-step budgets. For comparison, our algorithm solves for paths on the
San Francisco network with budgets of up to 1400 seconds (= 1400 time-steps)
in ≈ 7 seconds, even achieving query times below 1 second for budgets less than
550 seconds without any preprocessing at all. Furthermore, it also handles most
queries on the 71655-edge Luxembourg network in ≈ 10 seconds (almost all of
them in 20 seconds), where the network and time budgets are more than an order
of magnitude larger than the 2950-edge network handled by the SD approach in
the same amount of time.
Of course, this speedup-which increases more dramatically with the problem
size-is hardly surprising or coincidental; indeed, it is quite fundamental to the
nature of the algorithm: by drawing on the optimal policy as an upper bound
(and quite often an accurate one) for the reliability of the final path, it has a
very clear and fundamental informational advantage over any search algorithm
that lacks any knowledge of the final solution. This allows the algorithm to direct
itself toward the final path in an intelligent manner.
It is, however, less clear and more difficult to see how one might compare
the performance of our generic discrete-time approach with Gaussian-restricted,
continuous-time approaches [12, 24]. Such approaches operate under drastically
different assumptions and, in the case of [12], use approximation techniques,
which we have yet to employ for additional performance improvements. When the
true travel times cannot be assumed to follow Gaussian distributions, however,
our method, to the best of our knowledge, presents the most efficient means for
solving the path-based SOTA problem.
As we show next, combining our algorithm with preprocessing techniques
allows us to achieve even further reductions in query time, making it more
tractable for industrial applications on appropriately sized networks.
Preprocessing. Figure 2 demonstrates policy-based and path-based preprocessing
using Arc-Potentials for two random San Francisco and Luxembourg problem
8001000120014001600TimeBudget(s)0246810RunningTime(s)SanFranciscoPolicyPath15002000250030003500TimeBudget(s)0510152025LuxembourgPolicyPath instances. As can be seen in the figure, path-based preprocessing is in general
much more effective than policy-based preprocessing.
0 regions (unpruned)
132 regions (policy-pruned) 102 regions (path-pruned)
0 regions (unpruned)
232 grid (policy-pruned)
172 regions (path-pruned)
Fig. 2: Policy- vs. path-based pruning for random instances of San Francisco
(𝑇 = 837 seconds, source at top) and Luxembourg (𝑇 = 3165 seconds, source at
bottom). Light-gray edges are pruned from the graph and blue edges belong to
the optimal path, whereas red edges belong to (sub-optimal) paths that were on
the queue at the termination of the algorithm.
Figure 3, summarized in table 1, shows how the computation times scale with
the preprocessing parameters. As expected, path-based preprocessing performs
much better than purely policy-based preprocessing, and both become faster as
we use more fine-grained regions. Nevertheless, we see that the majority of the
speedup is achieved via a small number of regions, implying that preprocessing
can be very effective even with low amounts of storage. (For example, for a
17 × 17 grid in Luxembourg, this amounts to 71655 × 172 ≈ 21M floats.)
Fig. 3: Running time of pathfinding algorithm as a function of the time budget
for each network. Red dots represent the computation time of the policy, and
blue crosses represent the computation of the path using that policy.
10−210−1100101RunningTime(s)SanFrancisco0regions(nopruning)10−210−1100RunningTime(s)10²regions(policypruning)10−210−1100RunningTime(s)26²regions(policypruning)10−310−210−1100RunningTime(s)10²regions(pathpruning)8001000120014001600TimeBudget(s)10−310−210−1100RunningTime(s)26²regions(pathpruning)10−210−1100101Luxembourg0regions(nopruning)10−210−110010117²regions(policypruning)10−210−110010134²regions(policypruning)10−210−110017²regions(pathpruning)15002000250030003500TimeBudget(s)10−210−134²regions(pathpruning) Table 1: The average query time with both policy-based and path-based pruning
at various grid sizes and time budgets on the San Francisco network (left) and
the Luxembourg network (right). We can see that in both cases, most of the
speedup occurs at low granularity (and thus low space requirements).
Grid/
Pruning
Unpruned
Time budget (seconds)
800 1000 1200 1400 1600
1.81 3.00 4.10 5.11 5.23
10 × 10, policy 0.30 0.69 1.11 1.66 1.72
26 × 26, policy 0.17 0.40 0.63 0.93 0.97
10 × 10, path 0.11 0.38 0.63 0.87 0.90
26 × 26, path 0.02 0.04 0.06 0.07 0.08
Grid/
Time budget (seconds)
Pruning
Unpruned
1500 2000 2500 3000 3500
0.54 1.29 3.53 6.27 9.95
17 × 17, policy 0.25 0.83 2.31 4.57 7.97
34 × 34, policy 0.21 0.71 2.09 3.79 7.07
17 × 17, path 0.03 0.06 0.09 0.13 0.18
34 × 34, path 0.02 0.04 0.06 0.08 0.12
5.2 Scalability
Path-based preprocessing requires routing between all ≈ 𝑉 2 pairs of vertices,
which is quadratic in the size of the network and intractable for moderate size
networks. In practice, this meant that we had to preprocess every region lazily
(i.e. on-demand), which on our CPUs took 9000 CPU-hours. It is therefore
obvious that this becomes intractable for large networks, leaving policy-based
preprocessing as the only option. One possible approach for large-scale path-
based preprocessing might be to consider the boundary of each region rather
than its interior [8]. While currently uninvestigated, such techniques may prove
to be extremely useful in practice, and are potentially fruitful topics for future
exploration.
6 Conclusion and Future Work
We have presented an algorithm for solving the path-based SOTA problem by
first solving the easier policy-based SOTA problem and then using its solution
as a search heuristic. We have also presented two approaches for preprocessing
the underlying network to speed up computation of the search heuristic and
path query, including a generalization of the Arc-Flags preprocessing algorithm
that we call Arc-Potentials. We have furthermore applied and implemented these
algorithms on moderate-sized transportation networks and demonstrated their
potential for high efficiency in real-world networks.
While unobserved in practice, there remains the possibility that our algorithm
may perform poorly on stochastic networks in which the optimal policy is a
poor heuristic for the path reliability. Proofs in this direction have remained
elusive, and determining whether such scenarios can occur in realistic networks
remains an important step for future research. In the absence of theoretical
advances, however, our algorithm provides a more tractable alternative to the
state-of-the-art techniques for solving the path-based SOTA problem.
While our approach is tractable for larger networks than were possible with
previous solutions, it does not scale well enough to be used with regional or
continental sized networks that modern deterministic shortest path algorithms
can handle with ease. In the future, we hope to investigate how our policy-
based approach might be combined with other techniques such as the first-order
stochastic dominance [4] and approximation methods such approximate Arc-
Flags [8] for further speedup, and to also look into algorithms that allow for at
least a partial relaxation of the independence assumption. We hope that our
techniques will provide a strong basis for even better algorithms to tackle this
problem for large-scale networks in the foreseeable future.
Bibliography
[1] Yueyue Fan, Robert Kalaba, and JE Moore II. Arriving on time. Journal of
Optimization Theory and Applications, 127(3):497–513, 2005.
[2] Arthur Flajolet, S´ebastien Blandin, and Patrick Jaillet. Robust adaptive
routing under uncertainty. arXiv:1408.3374, 2014.
[3] Samitha Samaranayake, Sebastien Blandin, and Alexandre Bayen. A
tractable class of algorithms for reliable routing in stochastic networks.
Transportation Research Part C, 20(1):199–217, 2012.
[4] Yu Marco Nie and Xing Wu. Shortest path problem considering on-time
arrival probability. Transportation Research Part B: Methodological, 43(6):
597–613, 2009.
[5] Yueyue Fan and Yu Nie. Optimal routing for maximizing travel time
reliability. Networks and Spatial Economics, 6(3-4):333–344, 2006.
[6] Brian C Dean. Speeding up stochastic dynamic programming with zero-delay
convolution. Algorithmic Operations Research, 5(2):Pages–96, 2010.
[7] Samitha Samaranayake, Sebastien Blandin, and Alexandre Bayen. Speedup
techniques for the stochastic on-time arrival problem. In ATMOS, pages
83–96, 2012.
[8] Guillaume Sabran, Samitha Samaranayake, and Alexandre Bayen. Precom-
putation techniques for the stochastic on-time arrival problem. In ALENEX,
pages 138–146. SIAM, 2014.
[9] Ronald Gutman. Reach-based routing: A new approach to shortest path
algorithms optimized for road networks. In ALENEX/ANALC, pages 100–
111, 2004.
[10] Moritz Hilger, Ekkehard Kohler, Rolf Mohring, and Heiko Schilling. Fast
point-to-point shortest path computations with arc-flags. Ninth DIMACS
Implementation Challenge, 74:41–72, 2009.
[11] Evdokia Nikolova, Jonathan Kelner, Matthew Brand, and Michael Mitzen-
In
macher. Stochastic shortest paths via quasi-convex maximization.
Algorithms–ESA 2006. Springer, 2006.
[12] Sejoon Lim, Christian Sommer, Evdokia Nikolova, and Daniela Rus. Practical
route planning under delay uncertainty: Stochastic shortest path queries.
Robotics, page 249, 2013.
[13] Axel Parmentier and Fr´ed´eric Meunier. Stochastic shortest paths and risk
measures. arXiv:1408.0272, 2014.
[14] Daniel Delling, Peter Sanders, Dominik Schultes, and Dorothea Wagner.
Engineering route planning algorithms. Algorithmics of Large and Complex
Networks, 2:117–139, 2009.
[15] William G Gardner. Efficient convolution without input/output delay. In
Audio Engineering Society Convention 97. Audio Engineering Society, 1994.
[16] Rina Dechter and Judea Pearl. Generalized best-first search strategies and
the optimality of 𝐴*. Journal of the ACM (JACM), 32(3):505–536, 1985.
[17] Stuart J. Russell and Peter Norvig. Artificial Intelligence: A Modern Ap-
proach. Prentice-Hall, Inc., 1995. ISBN 0-13-103805-2.
[18] Ittai Abraham, Amos Fiat, Andrew Goldberg, and Renato Werneck. Highway
dimension, shortest paths, and provably efficient algorithms. In Proceedings
of the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms,
pages 782–793. Society for Industrial and Applied Mathematics, 2010.
[19] Andrew Goldberg, Haim Kaplan, and Renato Werneck. Reach for 𝐴*:
Efficient point-to-point shortest path algorithms. In ALENEX, volume 6,
pages 129–143. SIAM, 2006.
[20] Robert Geisberger, Peter Sanders, Dominik Schultes, and Daniel Delling.
Contraction hierarchies: Faster and simpler hierarchical routing in road
networks. In Experimental Algorithms, pages 319–333. Springer, 2008.
[21] Holger Bast, Stefan Funke, and Domagoj Matijevic. Transit: ultrafast
shortest-path queries with linear-time preprocessing. 9th DIMACS Imple-
mentation Challenge [1], 2006.
[22] Gianlorenzo DAngelo, Daniele Frigioni, and Camillo Vitale. Dynamic arc-
flags in road networks. In Experimental Algorithms, pages 88–99. Springer,
2011.
[23] Timothy Hunter, Pieter Abbeel, and Alexandre Bayen. The path inference
filter: model-based low-latency map matching of probe vehicle data. In
Algorithmic Foundations of Robotics X, pages 591–607. Springer, 2013.
[24] Sejoon Lim, Hari Balakrishnan, David Gifford, Samuel Madden, and Daniela
Rus. Stochastic motion planning and applications to traffic. The Interna-
tional Journal of Robotics Research, 2010.
|
1205.0192 | 2 | 1205 | 2012-05-11T11:22:55 | Large-scale compression of genomic sequence databases with the Burrows-Wheeler transform | [
"cs.DS",
"q-bio.GN"
] | Motivation
The Burrows-Wheeler transform (BWT) is the foundation of many algorithms for compression and indexing of text data, but the cost of computing the BWT of very large string collections has prevented these techniques from being widely applied to the large sets of sequences often encountered as the outcome of DNA sequencing experiments. In previous work, we presented a novel algorithm that allows the BWT of human genome scale data to be computed on very moderate hardware, thus enabling us to investigate the BWT as a tool for the compression of such datasets.
Results
We first used simulated reads to explore the relationship between the level of compression and the error rate, the length of the reads and the level of sampling of the underlying genome and compare choices of second-stage compression algorithm.
We demonstrate that compression may be greatly improved by a particular reordering of the sequences in the collection and give a novel `implicit sorting' strategy that enables these benefits to be realised without the overhead of sorting the reads. With these techniques, a 45x coverage of real human genome sequence data compresses losslessly to under 0.5 bits per base, allowing the 135.3Gbp of sequence to fit into only 8.2Gbytes of space (trimming a small proportion of low-quality bases from the reads improves the compression still further).
This is more than 4 times smaller than the size achieved by a standard BWT-based compressor (bzip2) on the untrimmed reads, but an important further advantage of our approach is that it facilitates the building of compressed full text indexes such as the FM-index on large-scale DNA sequence collections. | cs.DS | cs |
Doc-StartBIOINFORMATICS
Large-scale compression of genomic sequence
databases with the Burrows-Wheeler transform
Anthony J. Cox,1, ∗ Markus J. Bauer, 1 Tobias Jakobi 2 and Giovanna Rosone 3
1Computational Biology Group, Illumina Cambridge Ltd., Chesterford Research Park, Little
Chesterford, Essex CB10 1XL, United Kingdom
2Computational Genomics, CeBiTec, Bielefeld University, Bielefeld, Germany
3University of Palermo, Dipar timento di Matematica e Informatica,Via Archirafi 34, 90123 Palermo,
Italy
Received on XXXXX; revised on XXXXX; accepted on XXXXX
Vol. 00 no. 00 2012
Pages 1–6
Associate Editor : XXXXXXX
ABSTRACT
Motivation:
The Burrows-Wheeler transform (BWT) is the foundation of many
algorithms for compression and indexing of text data, but the cost
of computing the BWT of very large string collections has prevented
these techniques from being widely applied to the large sets of
sequences often encountered as the outcome of DNA sequencing
experiments. In previous work (Bauer et al. (2011)), we presented a
novel algorithm that allows the BWT of human genome scale data
to be computed on very moderate hardware,
thus enabling us to
investigate the BWT as a tool for the compression of such datasets.
Results:
We first used simulated reads to explore the relationship between
the level of compression and the error rate, the length of the reads and
the level of sampling of the underlying genome and compare choices
of second-stage compression algorithm.
We demonstrate that compression may be greatly improved by
a par ticular reordering of the sequences in the collection and give
a novel
‘implicit sor ting’ strategy that enables these benefits to
be realised without the overhead of sor ting the reads. With these
techniques, a 45× coverage of real human genome sequence data
compresses losslessly to under 0.5 bits per base, allowing the
135.3Gbp of sequence to fit into only 8.2Gbytes of space (trimming
a small propor tion of low-quality bases from the reads improves the
compression still fur ther).
This is more than 4 times smaller than the size achieved by a
standard BWT-based compressor (bzip2) on the untrimmed reads,
but an impor tant fur ther advantage of our approach is that it facilitates
the building of compressed full
text
indexes such as the FM-
index (Ferragina and Manzini (2000)) on large-scale DNA sequence
collections.
Availability:
Code to construct the BWT and SAP-array on large genomic data
sets is par t of the BEETL library, available as a github respository at
[email protected]:BEETL/BEETL.git.
Contact: [email protected]
∗ to whom correspondence should be addressed
c(cid:13) Oxford University Press 2012.
1 INTRODUCTION
In this paper we present strategies for the lossless compression of
the large number of short DNA sequences that comprise the raw
data of a typical sequencing experiment.
Much of the early work on the compression of DNA sequences
was motivated by the notion that the compressibility of a DNA
sequence could serve as a measure of its information content and
hence as a tool for sequence analysis. This concept was applied
to topics such as feature detection in genomes (Grumbach and
Tahi (1994); Rivals et al. (1996); Milosavljevic and Jurka (1993))
and alignment-free methods of sequence comparison (Chen et al.
(2002)) - a comprehensive review of the field up to 2009 is given by
Giancarlo et al. (2009) . However, Grumbach and Tahi in 1994 have
been echoed by many subsequent authors in citing the exponential
growth in the size of nucleotide sequence databases as a reason to
be interested in compression for its own sake. The recent and rapid
evolution of DNA sequencing technology has given the topic more
practical relevance than ever.
The outcome of a sequencing experiment typically comprises
a large number of short sequences - often called ‘reads’ - plus
metadata associated with each read and a ‘quality score’ that
estimates the confidence of each base. Tembe et al. (2010)
and Deorowicz and Grabowski (2011) both describe methods for
compressing the FASTQ file format in which such data is often
stored. The metadata is usually highly redundant, whereas the
quality scores can be hard to compress, and these two factors
combine to make it hard to estimate the degree of compression
achieved for the sequences themselves. However, both schemes
employ judicious combinations of standard text compression
methods such as Huffman and Lempel-Ziv, with which it is hard
to improve substantially upon the naive method of using a different
2-bit code for each of the 4 nucleotide bases. For example,
GenCompress (Chen et al. (2002)) obtains 1.92 bits per base
(henceforth bpb) compression on the E.coli genome.
An experimenter wishing to sequence a diploid genome such
as a human might aim for 20-fold average coverage or more,
with the intention of ensuring a high probability of capturing both
alleles of any heterozygous variation. This oversampling creates an
opportunity for compression that is additional to any redundancy
1
Cox et al.
inherent in the sample being sequenced. However,
in a whole-
genome shotgun experiment, the multiple copies of each locus are
randomly dispersed among the many millions of reads in the dataset,
making this redundancy inaccessible to any compression method
that relies on comparison with a small buffer of recently-seen data.
This can be addressed by ‘reference-based’ compression
(Kozanitis et al. (2010); Fritz et al. (2011)), which saves space
by sorting aligned reads by the position they align to on a
reference sequence and expressing their sequences as compact
encodings of the differences between the reads and the reference.
However this is fundamentally a lossy strategy that achieves
best compression by retaining only reads that closely match
the reference,
limiting the scope for future reanalyses such as
realignment to a refined reference (containing, perhaps, ethnicity
specific haplotypes (Dewey et al. (2011))) or any sort of de novo
discovery on reads that did not align well initially. Moreover,
as Yanovsky (2011) points out, a reference-based approach is
inapplicable to experiments for which a reference sequence is not
clearly defined (metagenomics) or entirely absent (de novo).
Yanovsky (2011) describes a lossless compression method
ReCoil for sets of reads that works in external memory (i.e.
via sequential access to files held on disk) and is therefore not
constrained in scale by available RAM. A graph of similarities
between the reads is first constructed and then each read is expressed
as a traversal on that graph, encodings of these traversals then being
passed to a general-purpose compressor (ReCoil uses 7-Zip1 ).
The two-stage nature of this procedure is shared by the family of
compression algorithms based on the Burrows-Wheeler transform
(BWT). The BWT is simply a permutation of the letters of the
text and so is not a compression method per se. Its usefulness
for compression stems from the facts that the BWT tends to be
more compressible than its originating text (since it tends to group
symbols into ‘runs’ of like letters, which are easy to compress) and,
remarkably, that the originating text can be reconstructed from it
(thus allowing the BWT to represent the originating text without
loss of information). Once generated,
the BWT is compressed
by standard techniques: a typical scheme would follow an initial
move-to-front encoding with run length encoding and then Huffman
encoding.
The widely-used BWT-based compressor bzip22 divides a text
into blocks of (at most, and by default) 900 kbytes and compresses
each separately, so is only able to take advantage of local similarities
in the data. Mantaci et al. (2005) gave the first extension of the
BWT to a collection of sequences and used it as a preprocessing
step for the simultaneous compression of the sequences of the
collection. They show that this method is more effective than the
technique used by a classic BWT-based compressor, because one
could potentially access redundancy arising from these long-range
correlations in the data. Until recently, computing the BWT of a
large collection of sequences was prevented from being feasible
on very large scales by the need either to store the suffix array
of the set of reads in RAM (requiring 400Gbytes of RAM for
a dataset of 1 billion 100-mer reads) or to resort to “divide-and-
conquer then merge” strategies at considerable cost in CPU time.
in Bauer et al. (2011),
three of the present authors
However,
1 www.7-zip.org, Igor Pavlov
2 www.bzip.org, Julian Seward
2
described fast and RAM-efficient methods capable of computing
the BWT of sequence collections of the size encountered in human
whole genome sequencing experiments, computing the BWT of
collections as large as 1 billion 100-mers.
Unlike the transformation in Mantaci et al. (2005), the algorithms
in Bauer et al. (2011) require ordered and distinct ‘end-marker’
characters to be appended to the sequences in the collection, making
the collection of sequences an ordered multiset, i.e. the order of
the sequences in the collection is determined by the lexicographical
order of these end-markers. It is easy to see that the ordering of
sequences in the collection can affect the compression, since the
same or similar sequences might be distant in the collection. We
will outline different ordering strategies for the input sequences and
will show their effect on simulated and real data.
We have created an open-source C++ library BEETL that makes
it practical to compute the BWT of large collections of DNA
sequences and and thus enables the redundancy present in large-
scale genomic sequence datasets to be fully exploited by generic
second-stage compressors such as bzip2 and 7-Zip. We use
simulated read collections of up to 60× redundancy to investigate
the effect of genome coverage, sequencing error and read length on
the level of compression. Furthermore, we show the effect of read
trimming and choices of second-stage compressors on the level of
compression. Finally, we describe an extension of our method that
implicitly reorders sequences in the collection so as to drastically
improve compression.
2 METHODS
Consider a string s comprising k symbols from some finite ordered
alphabet Σ = {c1 , c2 , . . . , cσ } of size σ . We mark the end of
s by appending an additional symbol $ that we consider to be
lexicographically smaller than any symbol in Σ. We can build k + 1
distinct suffixes from s by starting at different symbols of the string
and continuing rightwards until we reach $. If we imagine placing
then the Burrows-Wheeler
these suffixes in alphabetical order,
transform of s (Burrows and Wheeler (1994)) can be defined such
that the i-th element of the BWT is the symbol in s that precedes the
first symbol of the i-th member of this ordered list of suffixes. Each
symbol in the BWT therefore has an associated suffix in the string.
We recommend Adjeroh et al. (2008) for further reading.
One way to generalize the notion of the BWT to a collection of
m strings S = {s1 , . . . , sm } (see also Mantaci et al. (2005)) is to
imagine that each member si of the collection is terminated by a
distinct end marker $i such that $1 < · · · < $m . Moreover, we
assume that the end markers are lexicographically smaller than any
symbol in Σ. In Bauer et al. (2011), we give two related methods
for computing the BWT of large collections of DNA sequences by
making use of sequential reading and writing of files from disk. The
first variant BCR requires 14 bytes of RAM for each sequence in the
collection to be processed (and is hence capable of processing over
a billion reads in 16Gbytes of RAM), whereas the second variant
BCRext uses negligible RAM at the expense of a larger amount of
disk I/O.
To understand how a BWT string might be compressed, we can
think of it as the concatenation of a set of ‘runs’ of like letters,
each of which can be described by its constituent symbol c plus
an integer i denoting the number of times c is repeated. We assume
all runs are maximal, i.e. they do not abut another run of the same
character. Intuitively, for two strings of the same length, we expect
the string that consists of fewer (and hence, on average, longer) runs
to compress to a smaller size.
Given S , our strategy is to search for permutations S → S (cid:48) of
the sequences in the collection such that the BWT of the permuted
collection S (cid:48) can be more readily compressed than the BWT of S .
For the BWT of S , we define a bit array called the SAP-array (for
‘same-as-previous’) whose elements are set to 1 if and only if the
suffixes associated with their corresponding characters in the BWT
are identical (their end markers excluded) to those associated with
the characters that precede them. Thus each 0 value in the SAP-
array denotes the start of a new SAP-interval in the BWT, within
which all characters share an identical associated suffix.
The BWTs of S and S (cid:48) can only differ within SAP-intervals
that contains more than one distinct symbol. Within such an SAP-
interval, the ordering of the characters is entirely determined by the
ordering of the reads they are contained in, so best compression
is achieved if we permute the characters so they are grouped into
as few runs as possible. In doing this we are implicitly permuting
the sequences in the collection. As illustrated by Figure 1, if the
collection is sorted such that the reverses of the sequences are in
lexicographic order (we call this reverse lexicographic order, or
RLO for short), then the symbols in an SAP-interval of length l
must group into at most φ runs, where φ ≤ σ is the number of
distinct characters encountered within the interval. This compares
with an upper bound up to l runs if no sorting is applied. We would
therefore expect the RLO-sorting of a collection to compress better
than the original and our experiments in the next section show this
to be the case in practice.
However, we wish our approach to scale to many millions of
reads and so we would prefer to avoid sorting the collection as
a preprocessing step. If we know the SAP-array of a BWT, it
is a simple single-pass procedure to read both the BWT and its
SAP-array from disk in tandem, identify SAP-intervals, sort the
characters within them and output a modified BWT. We note there
is scope to experiment with different heuristics - e.g. in Figure
1, placing the T prior to the two Gs in the ACCT-interval would
eliminate another run by extending the run of Ts begun in the
preceding ACCACT-interval.
It remains to compute the SAP-array. In order to do this, we
show that the method of BWT construction introduced by three
of the present authors in Bauer et al. (2011) can be modified to
compute the SAP-array alongside the BWT with minimal additional
overhead. We do not attempt to describe this previous work in full
detail, but the BWT construction algorithm proceeds in k stages, k
being the length of the reads in the collection.3 At stage j , the j -
suffixes (0 ≤ j ≤ k) of the reads (that is, the suffixes of length
j , the 0-suffix being defined as the suffix that contains only the end
marker $) are processed in lexicographic order and the characters
that precede them are merged into a ‘partial BWT’.
The partial BWT at step j can be thought of as the concatenation
of σ + 1 segments Bj (0), Bj (1), . . . , Bj (σ), where each Bj (h)
corresponds to the symbols in the partial BWT segments that
precede all suffixes of S that are of length smaller or equal to
Large-scale BWT compression
j starting with c0 = $4 , for h = 0, and with ch ∈ Σ, for
h = 1, . . . , σ .
At each step j , with 0 < j ≤ k the main idea is to compute
the positions of all characters that precede the j -suffixes in the old
partial BWT in order to obtain an updated partial BWT that also
contains the symbols associated with the suffixes of length j . The
two variants BCR and BCRext of the Bauer et al. (2011) algorithm
contrive in slightly different ways (with different trade-offs between
RAM use and I/O) to arrange matters so that the symbols we have to
insert into the partial BWT are processed by lexicographic order of
their associated j -suffixes, which allows each segment of the partial
BWT to be held on disk and accessed sequentially. Crucially, this
also means that the characters in an SAP-interval will be processed
consecutively. When we write a BWT character, we can also write
its SAP status at the same time, since it is not going to change in
subsequent iterations.
We consider two suffixes u and v of length j of two different
reads sr and st satisfying 1 ≤ r, t ≤ m and r (cid:54)= t. Both sr and st
belong to S and assume that u is lexicographically less than v . Then,
they are identical (up to the end-markers) and so the SAP status of
v must be set to 1 if and only if the following conditions hold: the
symbols associated with suffixes u(cid:48) and v (cid:48) of length (j − 1) of sr
and st are equal which guarantees that u and v begin with the same
symbol, and they are stored in the same BWT segment, implying
that u(cid:48) and v (cid:48) begin with the same symbol. The SAP status of all
suffixes between u(cid:48) and v (cid:48) must be set to 1 which ensures that all
suffixes between u(cid:48) and v (cid:48) in iteration j − 1 coincide and have the
length j − 1. Note that the symbols in the BWT segment associated
with suffixes between u(cid:48) and v (cid:48) could be different from the symbol
preceding u(cid:48) (and v (cid:48) ).
Actually, we compute the SAP status of the suffixes u and v at the
step j − 1 when we insert the corresponding suffixes u(cid:48) and v (cid:48) . In
particular, when we copy the values from the old to the new partial
BWT and insert the new m values, we can read the SAP values and
verify the condictions stated above at the same time. It means that
at the end of the iteration j − 1, we obtain the partial BWT and the
SAP values of the next suffixes to insert and, at iteration j , we can
directly set the SAP status of the j -suffixes and compute the SAP
status for the suffixes of the next iteration.
i.e. we can
We can compute the SAP-interval by induction,
compute the SAP status of the j -suffixes at iteration j − 1 by using
the symbols associated with the (j − 1)-suffixes and their SAP
values. At iteration j − 1, we assume that we know the SAP values
of the (j − 1)-suffixes and we have to compute the SAP values of
the j -suffixes need for the next iteration. For simplicity, we focus on
the computation of the SAP values of a fixed BWT segment, i.e. we
only consider the insertion of the (j − 1)-suffixes starting with the
same symbol. In our implementation, we use a counter A for each
symbol of the alphabet and a generic counter Z .
The element A[h], for each h = 1, . . . , σ and ch ∈ Σ (we can
ignore the end-marker because the reads have the same length and
hence it does not appear in the partial BWT), contains the number of
SAP intervals between the first position and the position of the last
inserted symbol associated with a read sq (for some 1 ≤ q ≤ m)
equal to ch in the considered BWT segment. The counter Z contains
3 Our implementation and, for simplicity, our description here assume all
reads in S to have the same length, but this restriction is not in fact intrinsic
to the algorithm.
4 In our implementation, we assume that all end-markers are equal, and
sr [k] < st [k], if r < t.
3
Cox et al.
the number of the SAP intervals between the first position and the
position where we have to insert cp . The symbol cp is associated
with the new suffix of length j − 1 of read st , with 1 ≤ t ≤ m,
that we want to insert. If the value A[p] is equal to Z , then the SAP
status of the j -suffix of st (obtained by concatenating cp (j − 1)-
suffix of st ) must be set to 1, otherwise it is set to 0. We observe
that if A[p] = Z holds true, then this implies that j -suffixes of sr
and st are in the same SAP interval.
B (0) SAP (0)
0
T
1
T
1
T
1
T
B (1) SAP (1)
0
T
0
G
1
T
G
1
0
C
0
T
1
G
G
0
B (2) SAP (2)
0
C
0
A
1
A
1
A
1
A
C
0
1
C
1
A
C
1
B (3) SAP (3)
0
A
A
1
0
$4
$2
0
B (4) SAP (4)
0
C
C
1
1
C
1
C
0
$3
0
A
$1
0
suffixes
$1
$2
$3
$4
suffixes
ACCAC T $2
ACCT$1
ACCT$2
ACCT$4
AC T $3
AGACCT$1
AGACCT$4
AT ACC T $2
suffixes
CAC T $3
CCAC T $3
CCT$1
CCT$2
CCT$4
CT$1
CT$2
CT$3
CT$4
suffixes
GACCT$1
GACCT$4
GAGACC T $4
GAT ACC T $2
suffixes
T$1
T$2
T$3
T$4
T ACCAC T $3
T ACC T $2
T AGACC T $1
⇒
BSAP (0)
T
T
T
T
BSAP (1)
T
G
G
T
C
G
T
G
BSAP (2)
C
A
A
A
A
A
C
C
C
BSAP (3)
A
A
$4
$2
BSAP (4)
C
C
C
C
$3
A
$1
Fig. 1. Columnwise from left, we have the BWT of the collection
{T AGACC T , GAT ACC T , T ACCAC T , GAGACC T },
=
S
the
SAP
bit
and
the
suffix
associated with
each
symbol.
collection
The
the
of
BWT
the
shows
side
right-hand
{T ACCAC T , T AGACC T , GAGACC T , GAT ACC T }
obtained
by sorting the elements of S into reverse lexicographic order (RLO). This
permutes the symbols within SAP-intervals so as to minimise the number of
runs.
various coverage levels for compression both on the original reads
and the BWT transform.
We found the PPMd mode (-m0=PPMd) of 7-Zip to be a good
choice of second-stage compressor for the BWT strings (referred to
as PPMd (default) in the following). RLO-sorting the datasets led to
a BWT that was slightly more compressible than the SAP-permuted
BWT, but the difference was small (0.36bpb versus 0.38bpb at 60×,
both over 30% less than the 0.55bpb taken up by the compressed
BWT of the unsorted reads).
In contrast, when gzip5 , bzip2 and default PPMd were applied
to the original reads, each gave a compression level that was
consistent across all levels of coverage and none was able to
compress below 2 bits per base. However, a sweep of the PPMd
parameter space yielded a combination -mo=16 -mmem=2048m
that attained 0.50bpb on the 60× dataset (in the following we will
refer to this parameter setting as PPMd (large)). This is because the
E.coli genome is small enough to permit several-fold redundancy
of the genome to be captured in the 2Gbytes of working space
that this combination specifies. For a much larger genome such as
human, this advantage disappears. Figure 3 summarizes results from
a set of 192 million human reads6 previously analyzed by Yanovsky
(2011): PPMd(large) compresses the BWT of the reads less well
than PPMd(default), as well as being several times slower.
We also investigated the effects of sequencing errors on the
compression ratios by simulating 40× data sets of 100bp reads with
different rates of uniformly distributed substitution error, finding
that an error rate of 1.2% approximately doubled the size of the
compressed BWT (0.90bpb, compared with 0.47bpb for error-free
data at the same coverage).
We were interested in the behaviour of BWT-based compression
techniques as a function of the read length. To this end, we fixed
a coverage of 40× and simulated error-free E.coli reads of varying
lengths. As the read length increased from 50bp to 800bp, the size
of the compressed BWTs shrank from 0.54bpb to 0.32bpb. This
is not surprising since longer reads allow repetitive sequences to be
grouped together, which could otherwise potentially be disrupted by
suffixes of homologous sequences.
Finally, we assessed the performance of the compressors on
a typical human genome resequencing experiment7 , containing
135.3Gbp of 100-base reads, or about 45× coverage of the human
genome. In addition to this set of reads, we created a second dataset
by trimming the reads based on their associated quality scores
according to the scheme described in bwa (Li and Durbin (2009)).
Setting a quality score of 15 as the threshold removes 1.3% of
the input bases. We again constructed the corresponding data sets
in RLO and SAP order. Table 4 shows the improvement in terms
of compression after the trimming. Eliminating 1.3% of the bases
improves the compression ratio by about 4.5%, or compressing the
entire 133.6Gbp down to 7.7Gbytes.
3 RESULTS
Reads simulated from the E.coli genome (K12 strain) allowed us to
assess separately the effects of coverage, read length and sequencing
error on the level of compression achieved. First, a 60× coverage of
error-free 100 base reads was subsampled into datasets as small as
10×. Figure 2 shows a summary plot of the compression ratios at
5 www.gzip.org, Jean-loup Gailly and Mark Adler
6 http://www.ebi.ac.uk/ena/data/view/SRX001540
7 Available
at
http://www.ebi.ac.uk/ena/data/view/
ERA015743
4
Large-scale BWT compression
Input size BWT BWT-RLO BWT-SAP
untrimmed 135.3Gbp
0.484
0.528
0.746
trimmed 133.6Gbp
0.721
0.504
0.462
Fig. 4. Different combinations of first-stage (BWT, SAP- and RLO-
permuted BWT) and PPMd as the second-stage compression method
compared on a 45× human data set. Values state the compression achieved
in bits per input base. We trimmed the reads by following the strategy
described for bwa which removed 1.3% of the bases. This leads to an
improvement in terms of compression ratio of 4.5%.
4 DISCUSSION
Analysing the SRX001540 dataset allowed us to compare our results
with ReCoil, but the reads are noisier and shorter than more recent
datasets and, at under 3×, the oversampling of the genome is too
low to manifest the kind of coverage-induced compression that
Figure 2 illustrates. Nevertheless, we were able to compress the
data to 1.21bpb in just over an hour, compared to 1.34bpb achieved
by ReCoil in 14 hours (albeit on a slower processor). On the
more recent ERA015743 data we obtained 0.48bpb, allowing the
135.3Gbp of data to fit in 8.2Gbytes of space.
In both cases, reordering the strings in the collection - implicitly
or explicitly - is necessary for best compression. One situation
where this might be problematic would be if the reads are paired
- typically, sequenced from different ends of a library of DNA
templates. An obvious way to capture this relationship would be to
store the two reads in a pair at adjacent positions in S - i.e. the first
two reads in the collection are reads 1 and 2 of the first read-pair,
and so forth. If we wanted to retain this information after a change
in ordering, we would incur the overhead of storing an additional
pointer to associate each read with its mate.
In Bauer et al. (2012), we give efficient algorithms for inverting
the BWT and recovering the original reads. However, when
augmented with relatively small additional data structures,
the
compressed BWT forms the core of the FM-index (Ferragina and
Manzini (2000, 2005); Ferragina et al. (2007)) that allows count
(“how many times does a k-mer occur in the dataset?”) and locate
(“at what positions in the collection does the k-mer occur?”)
queries. Several variants of this algorithmic scheme exist which
make different tradeoffs between the compression achieved and the
time needed for the count and locate operations. For instance,
Ferragina et al. (2007) describes a variant of the FM-index that
indexes a string T of length n within nHk (T )+ o(n) bits of storage,
where Hk (T ) is the k-th order empirical entropy of T . This index
counts the occurrences in T of a string of length p in O(p) time and
it locates each occurrence of the query string in O(log1+n) time,
for any constant 0 < < 1.
Most sequencing technologies associate a quality score with each
sequenced base and Kozanitis et al. (2010) found that significant
lossless compression of these scores can be hard to achieve, but also
showed that a lossy reduction in resolution of the scoring scheme
could obtain better compression while having limited impact on
the accuracy of variant calls derived from the sequences. Future
work we are pursuing would complement such approaches by using
queries to a BWT-based index for de novo identification of bases that
are not likely to be important in downstream variant calls, whose
quality scores can then be discarded entirely.
5
Fig. 2. We simulated 60× coverage of error-free 100 base reads from the
E.coli genome and subsampled this into datasets as small as 10×. We
compared different compression schemes, both on the raw input sequences
and the BWT. RLO denotes a reverse lexicographical ordering of the reads,
SAP is the data set where all the reads are ordered according to the same-as-
previous array. The x-axis gives the coverage level whereas the y -axis shows
the number of bits used per input symbol. Gzip, Bzip2, PPMd (default)
and PPMd (large) show compression achieved on the raw sequence data.
BWT, BWT-SAP and BWT-RLO give compression results on the BWT
using PPMd (default) as second-stage compressor.
Stage 1
Reads
BWT
Method
Stage 2
Bzip2
PPMd (default)
PPMd (large)
-mx9
Bzip2
PPMd (default)
PPMd (large)
-mx9
Bzip2
PPMd (default)
PPMd (large)
-mx9
BWT-SAP
3520
-
Time
Stage 1 Stage 2
905
324
5155
17974
818
353
4953
16709
601
347
3116
11204
3520
Compression
2.25
2.04
2.00
1.98
2.09
1.93
2.05
2.09
1.40
1.21
1.28
1.34
Fig. 3. Different combinations of first-stage (BWT, SAP-permuted
BWT) and second-stage (bzip2 with default parameters, PPMd mode of
7-Zip with default parameters, PPMd mode of 7-Zip with -mo=16
-mmem=2048m, deflate mode of 7-Zip with -mx9) compression
compared on 192 million human reads previous analyzed by Yanovsky
(2011). Time is in CPU seconds, as measured on a single core of an Intel
Xeon X5450 (Quad-core) 3GHz processor. Compression is in bits per base.
1020304050600.00.51.01.52.02.5coveragebits per baselllllllGzipBzip2PPMd (default)BWTPPMd (large)BWT−SAPBWT−RLO Cox et al.
More broadly, many computationally intensive tasks in sequence
analysis might be facilitated by having a set of reads in indexed
form. For example, Simpson and Durbin (2010, 2011) show how
the overlap of a set of reads can be computed from its FM-index
and apply this insight to build a practical tool for de novo assembly.
By making it feasible to index sets of reads on the scale necessary
for whole human genome sequencing, our work raises the intriguing
possibility of a new generation of algorithms that operate directly on
sets of reads held in compressed index form.
ACKNOWLEDGEMENT
The authors would like to thank Dirk Evers for his support
throughout this project.
Funding: M.J.B. and A.J.C. are employees of Illumina Inc., a public
company that develops and markets systems for genetic analysis,
and receive shares as part of their compensation. Part of T.J.’s
contribution was made while on a paid internship at Illumina’s
offices in Cambridge, UK.
REFERENCES
Adjeroh, D., Bell, T., and Mukherjee, A. (2008). The Burrows-Wheeler Transform:
Data Compression, Suffix Arrays, and Pattern Matching. Springer Publishing
Company, Incorporated, 1st edition.
Bauer, M. J., Cox, A. J., and Rosone, G. (2011). Lightweight BWT construction for
very large string collections. In CPM 2011, volume 6661 of LNCS, pages 219–231.
Springer.
Bauer, M. J., Cox, A. J., and Rosone, G. (2012). Lightweight algorithms for
constructing and inverting the BWT of string collections. Theoretical Computer
Science. Available online 10 February 2012.
Burrows, M. and Wheeler, D. J. (1994). A block sorting data compression algorithm.
Technical report, DIGITAL System Research Center.
Chen, X., Li, M., Ma, B., and Tromp, J. (2002). DNACompress: fast and effective
DNA sequence compression. Bioinformatics, 18(12), 1696–1698.
Deorowicz, S. and Grabowski, S. (2011). Compression of genomic sequences in
FASTQ format. Bioinformatics, 27(6), 860–862.
Dewey, F. E., Chen, R., Cordero, S. P., Ormond, K. E., Caleshu, C., Karczewski, K. J.,
Whirl-Carrillo, M., Wheeler, M. T., Dudley, J. T., Byrnes, J. K., Cornejo, O. E.,
Knowles, J. W., Woon, M., Sangkuhl, K., Gong, L., Thorn, C. F., Hebert, J. M.,
Capriotti, E., David, S. P., Pavlovic, A., West, A., Thakuria, J. V., Ball, M. P.,
Zaranek, A. W., Rehm, H. L., Church, G. M., West, J. S., Bustamante, C. D.,
Snyder, M., Altman, R. B., Klein, T. E., Butte, A. J., and Ashley, E. A. (2011).
Phased whole-genome genetic risk in a family quartet using a major allele reference
sequence. PLoS Genet, 7(9), e1002280.
Ferragina, P. and Manzini, G. (2000). Opportunistic data structures with applications.
In Proceedings of the 41st Annual Symposium on Foundations of Computer Science,
pages 390–398, Washington, DC, USA. IEEE Computer Society.
Ferragina, P. and Manzini, G. (2005). Indexing compressed text. J. ACM , 52, 552–581.
Ferragina, P., Manzini, G., M akinen, V., and Navarro, G. (2007). Compressed
representations of sequences and full-text indexes. ACM Trans. Algorithms, 3.
Fritz, M. H., Leinonen, R., Cochrane, G., and Birney, E. (2011). Efficient storage of
high throughput DNA sequencing data using reference-based compression. Genome
Research, 21(5), 734–740.
Giancarlo, R., Scaturro, D., and Utro, F. (2009). Textual data compression in
computational biology: a synopsis. Bioinformatics, 25(13), 1575–1586.
Grumbach, S. and Tahi, F. (1994). A new challenge for compression algorithms:
Genetic sequences. In Information Processing and Management, volume 30, pages
875–886.
Kozanitis, C., Saunders, C., Kruglyak, S., Bafna, V., and Varghese, G. (2010).
Compressing genomic sequence fragments using SlimGene. In RECOMB, volume
6044 of LNCS, pages 310–324. Springer.
Li, H. and Durbin, R. (2009). Fast and accurate short read alignment with Burrows-
Wheeler transform. Bioinformatics, 25(14), 1754–1760.
Mantaci, S., Restivo, A., Rosone, G., and Sciortino, M. (2005). An Extension of the
Burrows Wheeler Transform and Applications to Sequence Comparison and Data
Compression. In CPM 2005, volume 3537 of LNCS, pages 178—189.
Milosavljevic, A. and Jurka, J. (1993). Discovering simple DNA sequences by
the algorithmic significance method. Computer applications in the biosciences :
CABIOS, 9(4), 407–411.
Rivals, E., Dauchet, M., Delahaye, J., and Delgrange, O. (1996). Compression and
genetic sequence analysis. Biochimie, 78(5), 315–22.
Simpson, J. T. and Durbin, R. (2010). Efficient construction of an assembly string graph
using the FM-index. Bioinformatics, 26(12), i367–i373.
Simpson, J. T. and Durbin, R. (2011). Efficient de novo assembly of large genomes
using compressed data structures. Genome Research.
Published in Advance
December 7, 2011.
Tembe, W., Lowey, J., and Suh, E. (2010). G-SQZ: compact encoding of genomic
sequence and quality data. Bioinformatics, 26(17), 2192–2194.
Yanovsky, V. (2011). ReCoil - an algorithm for compression of extremely large datasets
of DNA data. Algorithms for Molecular Biology, 6(1), 23.
6
|
1007.2021 | 4 | 1007 | 2011-10-06T00:46:31 | Parameterizing by the Number of Numbers | [
"cs.DS",
"cs.DM"
] | The usefulness of parameterized algorithmics has often depended on what Niedermeier has called, "the art of problem parameterization". In this paper we introduce and explore a novel but general form of parameterization: the number of numbers. Several classic numerical problems, such as Subset Sum, Partition, 3-Partition, Numerical 3-Dimensional Matching, and Numerical Matching with Target Sums, have multisets of integers as input. We initiate the study of parameterizing these problems by the number of distinct integers in the input. We rely on an FPT result for ILPF to show that all the above-mentioned problems are fixed-parameter tractable when parameterized in this way. In various applied settings, problem inputs often consist in part of multisets of integers or multisets of weighted objects (such as edges in a graph, or jobs to be scheduled). Such number-of-numbers parameterized problems often reduce to subproblems about transition systems of various kinds, parameterized by the size of the system description. We consider several core problems of this kind relevant to number-of-numbers parameterization. Our main hardness result considers the problem: given a non-deterministic Mealy machine M (a finite state automaton outputting a letter on each transition), an input word x, and a census requirement c for the output word specifying how many times each letter of the output alphabet should be written, decide whether there exists a computation of M reading x that outputs a word y that meets the requirement c. We show that this problem is hard for W[1]. If the question is whether there exists an input word x such that a computation of M on x outputs a word that meets c, the problem becomes fixed-parameter tractable. | cs.DS | cs | Parameterizing by the Number of Numbers(cid:63)
Michael R. Fellows1, Serge Gaspers2, and Frances A. Rosamond1
1 School of Engineering and IT, Charles Darwin University, NT 0909, Australia.
{michael.fellows,frances.rosamond}@cdu.edu.au
2 Institute of Information Systems, Vienna University of Technology, Vienna, Austria.
[email protected]
Abstract. The usefulness of parameterized algorithmics has often depended on what
Niedermeier has called "the art of problem parameterization". In this paper we intro-
duce and explore a novel but general form of parameterization: the number of numbers.
Several classic numerical problems, such as Subset Sum, Partition, 3-Partition,
Numerical 3-Dimensional Matching, and Numerical Matching with Target
Sums, have multisets of integers as input. We initiate the study of parameterizing
these problems by the number of distinct integers in the input. We rely on an FPT
result for Integer Linear Programming Feasibility to show that all the above-
mentioned problems are fixed-parameter tractable when parameterized in this way. In
various applied settings, problem inputs often consist in part of multisets of integers or
multisets of weighted objects (such as edges in a graph, or jobs to be scheduled). Such
number-of-numbers parameterized problems often reduce to subproblems about tran-
sition systems of various kinds, parameterized by the size of the system description.
We consider several core problems of this kind relevant to number-of-numbers param-
eterization. Our main hardness result considers the problem: given a non-deterministic
Mealy machine M (a finite state automaton outputting a letter on each transition), an
input word x, and a census requirement c for the output word specifying how many
times each letter of the output alphabet should be written, decide whether there exists
a computation of M reading x that outputs a word y that meets the requirement c.
We show that this problem is hard for W [1]. If the question is whether there exists an
input word x such that a computation of M on x outputs a word that meets c, the
problem becomes fixed-parameter tractable.
1 Introduction
Parameterized complexity and algorithmics has been developing for more than twenty
years. Some important progress of the field has depended on what Niedermeier has
called "the art of problem parameterization" (see Chapter 5 of his monograph [23]).
For example, it was Cristina Bazgan who first suggested that the parameter might
be k = 1/ in the study of the complexity of approximation, leading eventually to
the study of EPTASs [3].
Here we explore, for the first time (to our knowledge), a parameterization that
seems widely relevant: the number of numbers. Many problems take as input informa-
tion that consists (in part) of multisets of integers or multisets of weighted objects,
such as weighted edges in a weighted graph, the time-requirements of jobs to be
(cid:63) A preliminary version of this paper appeared in the proceedings of IPEC 2010 [7]. M.R.F. and
F.A.R. acknowledge support from the Australian Research Council. S.G. acknowledges partial
support from the European Research Council (COMPLEX REASON, 239962), from Conicyt
Chile (Basal-CMM), and from the Australian Research Council.
1
1
0
2
t
c
O
6
]
S
D
.
s
c
[
4
v
1
2
0
2
.
7
0
0
1
:
v
i
X
r
a
2
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
scheduled, or the sequence of molecular weights of a spectrographic dataset. Our in-
vestigations are of importance for problem input distributions where the number of
distinct numerical values is small compared to the overall input size, and when the
modeling of the problem allows rounding as a way to get to fewer distinct values.
In classical complexity, this "parameterization" has been explored in distribution-
sensitive algorithmics [29]. For example, while Ω(n log n) is a lower bound on sorting
n values in the comparison model [18], a multiset of cardinality n and h distinct
values can be sorted using O(n log h) comparisons [22].
It is perhaps surprising that this parameterization in the sense of Niedermeier's
"art of problem parameterization" [23, 24] has not been considered before in parame-
terized complexity, as it seems entirely well-motivated. While weighted combinatorial
optimization problems have generally strong claims to model realism, it is often the
case that, e.g., the jobs to be scheduled may be of certain standard sizes arising in
a limited number of ways, or that the costs of the nodes in a network problem may
depend on the model and vendor of the device, of which there are a limited num-
ber of possibilities. Many similar scenarios easily come to mind. A bounded number
of numbers may also arise naturally and implicitly in parameterized problems when
numbers are associated to other parameterized aspects of a problem, such as alphabet
size.
As an initial foray, we first show that a number of classic NP-hard problems
about multisets of integers, when parameterized in this way, become fixed-parameter
tractable. The proofs are easy, and the knowledgeable reader might anticipate them
almost as exercises today -- they use the relatively deep result that Integer Linear
Programming, parameterized by the number of variables, is FPT. Until recently, as
noted in the 2006 monograph by Niedermeier [23], there were not so many interesting
applications of this fundamental result (see [1, 11, 12, 16] for some exceptions).
At a deeper level of engagement with this parameterization, we describe some
examples of how number-of-numbers parameterized problems reduce to numerical
problems about Mealy machines, parameterized by the size of the description of the
machine. We show that one basic problem about Mealy machines, parameterized in
this way, is FPT, and that another is W [1]-hard.
2 Preliminaries
Integer Linear Programming In the Integer Linear Programming Feasibility
problem (ILPF), the input is an m × n matrix A of integers and an m-vector b of
integers, the parameter is n, and the question is whether there exists an n-vector x of
integers satisfying the m inequalities Ax ≤ b. ILPF, parameterized by the number of
variables, was shown to be fixed-parameter tractable by Lenstra [20] and the running
time has been improved by Kannan [17] and by Frank and Tardos [14].
Multisets Let A be a multiset. The cardinality of A, denoted A, is the total number
of elements in A, including repeated memberships. The variety of A, denoted A, is
the number of distinct elements in A. Element a has multiplicity m in A if it occurs
m times in A. We denote the set of integers from 1 to n by [n] = {1, . . . , n}.
Parameterizing by the Number of Numbers
3
Graphs Let G = (V, E) be a graph, v ∈ V be a vertex of G, and S ⊆ V be a subset
of vertices of G. The subgraph of G induced on S is the graph G[S] = (S, E ∩ {uv :
u, v ∈ S}). The set S is a clique of G if G[S] is complete, i.e. there is an edge between
every two distinct vertices of G[S]. The set S is an independent set of G if G[S] is
empty, i.e. G[S] has no edge. The neighborhood of v is the set of vertices incident to v
and denoted N (v). The degree of v is d(v) = N (v). We also define NS(v) = N (v)∩ S
and dS(v) = NS(v).
Words Let Σ be an alphabet. The elements of Σ are called letters, and a word x of
length n = x is a sequence of n letters. The symbol denotes the empty letter. We
denote the concatenation of two words x1, x2 ∈ Σ∗ by x1x2. The ith power of a word
x is denoted xi or (x)i and represents the word xx . . . x
.
i times
(cid:124) (cid:123)(cid:122) (cid:125)
Parameterized Complexity We define the basic notions of Parameterized Complex-
ity and refer to other sources [6, 13, 23] for an in-depth treatment. A parameterized
problem is a set of pairs (I, k), the instances, where I is the main part and k is
the parameter. A parameterized problem is fixed-parameter tractable if there exist a
computable function f and an algorithm that solves any instance (I, k) of size n in
time f (k)nO(1). FPT denotes the class of all fixed-parameter tractable parameterized
decision problems.
Parameterized complexity offers a completeness theory that allows the accumula-
tion of strong theoretical evidence that some parameterized problems are not fixed-
parameter tractable. This theory is based on a hierarchy of complexity classes
FPT ⊆ W[1] ⊆ W[2] ⊆ W[3] ⊆ ··· ⊆ XP.
where all inclusions are believed to be strict. Each class W[i] contains all parameter-
ized decision problems that can be reduced to a canonical parameterized satisfiability
problem Pi under parameterized reductions. These are many-to-one reductions where
the parameter for one problem maps into the parameter for the other. More specifi-
cally, a parameterized problem L reduces to a parameterized problem L(cid:48) if there is a
mapping R from instances of L to instances of L(cid:48) such that
1. (I, k) is a Yes-instance of L if and only if (I(cid:48), k(cid:48)) = R(I, k) is a Yes-instance
2. there is a computable function g such that k(cid:48) ≤ g(k), and
3. there is a computable function f such that R can be computed in time f (k)·nO(1),
of L(cid:48),
where n denotes the size of (I, k).
A parameterized problem L is then in W[i], for i ∈ N, if it has a parameterized
reduction to the problem of deciding whether a Boolean decision circuit (a decision
circuit is a circuit with exactly one output), with AND, OR, and NOT gates, of
constant depth such that on each path from an input to the output, all but i gates
have a constant number of inputs, parameterized by the number of ones in a satisfying
assignment to the inputs of the circuit [6].
A parameterized problem is in XP if there exist computable functions f and g
and an algorithm that solves any instance (I, k) of size n in time f (k)ng(k).
4
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
3 Subset Sum and Partition
We start with two classic problems on multisets and show that they are fixed-para-
meter tractable, parameterized by the number of numbers.
variety-Subset Sum (var -SubSum)
Input:
Parameter: k = A, the number of distinct integers in A.
Question:
Is there a multiset X ⊆ A such that(cid:80)
A multiset A of integers and an integer s.
a∈X a = s?
A multiset A of integers.
variety-Partition (var -Part)
Input:
Parameter: k = A.
Question:
The parameterizon of Subset Sum by X is W [1]-hard [9]. This hardness also holds
for the parameterization of Partition by X as an easy reduction from Subset
a∈A a)/2,
a∈A a) − s
a∈X a =(cid:80)
Sum adds the integer ((cid:80)
a∈A a)/2, and if s > ((cid:80)
the reduction looks instead for the complement set A \ X that sums to ((cid:80)
Is there a multiset X ⊆ A such that(cid:80)
a∈A a) − 2s to A if s ≤ ((cid:80)
b∈A\X b?
and uses the previous construction.
Our FPT results use a deep result of Lenstra, stating that Integer Linear
Programming Feasibility (ILPF), parameterized by the number of variables, is
FPT. They are obtained by very natural formulations of the respective problems as
integer programs.
Theorem 1. var-SubSum is fixed-parameter tractable.
Proof. Given an instance (A, s) for var -SubSum, with A = k, we create an equiv-
alent instance of ILPF whose number of variables is upper bounded by a function of
k. Let a1, . . . , ak denote the distinct elements of A and let m1, . . . , mk denote their
respective multiplicities in A. The ILPF instance has the integer variables x1, . . . , xk
and the following inequalities and equalities.
xi ≤ mi
xi ≥ 0
xi · ai = s.
k(cid:88)
i=1
∀i ∈ [k]
∀i ∈ [k]
For each i ∈ [k], the variable xi represents the number of times ai occurs in X, the
set summing to s in a valid solution. Using standard techniques in mathematical
programming, these constraints can be transformed into the form Ax ≤ b.
(cid:117)(cid:116)
A very similar proof shows that var -Part is fixed-parameter tractable.
Theorem 2. var-Part is fixed-parameter tractable.
Parameterizing by the Number of Numbers
5
Proof. Given an instance A for var -Part, with A = k, we create an equivalent
instance of ILPF whose number n of variables is upper bounded by a function of k.
Let a1, . . . , ak denote the distinct elements of A and let m1, . . . , mk denote their
respective multiplicities in A. The ILPF instance has the integer variables x1, . . . , xk
and the following inequalities and equalities.
xi ≤ mi
xi ≥ 0
(cid:88)
xi · ai =
k(cid:88)
b∈A\X b =(cid:80)
a∈A
i=1
that(cid:80)
a∈X a =(cid:80)
∀i ∈ [k]
∀i ∈ [k]
a/2.
For each i ∈ [k], the variable xi represents the number of times ai occurs in X, such
Using standard techniques in mathematical programming, these constraints can be
transformed such that they respect the form Ax ≤ b.
(cid:117)(cid:116)
a∈A a/2 in a valid solution.
4 Other Classic Numerical Problems
Using the ILPF machinery, we show in this section that several other problems,
which are often used in NP-hardness proofs, become fixed-parameter tractable when
parameterized by the number of numbers.
Three multisets A, B, C of n integers each and an integer s.
variety-Numerical 3-Dimensional Matching (var -Num3-DM)
Input:
Parameter: k = A ∪ B ∪ C.
Question: Are there n triples S1, . . . , Sn, each containing one element from
each of A, B, and C such that for every i ∈ [n],(cid:80)
a = s?
a∈Si
Theorem 3. var-Num3-DM is fixed-parameter tractable.
Proof. Let (A, B, C, s) be an instance for var -Num3-DM, with k1 = A, k2 = B,
k3 = C, and k = A ∪ B ∪ C. Let a1, . . . , ak1 denote the distinct elements of
A, b1, . . . , bk2 denote the distinct elements of B, and c1, . . . , ck3 denote the distinct
elements of C. Also, let m1,a, . . . , mk1,a, m1,b, . . . , mk2,b, m1,c, . . . , mk3,c denote their
respective multiplicities in A, B, and C. We create an instance of ILPF with at most
k3 integer variables xi,j,(cid:96), for i ∈ [k1], j ∈ [k2], (cid:96) ∈ [k3]:
(cid:88)
(cid:88)
(cid:88)
xi,j,(cid:96) = 0
xi,j,(cid:96) = mi,a
xi,j,(cid:96) = mj,b
xi,j,(cid:96) = m(cid:96),c
for each (i, j, (cid:96)) ∈ [k1] × [k2] × [k3]
such that ai + bj + c(cid:96) (cid:54)= s
∀i ∈ [k1]
∀j ∈ [k2]
∀(cid:96) ∈ [k3]
(j,(cid:96))∈([k2],[k3])
(i,(cid:96))∈([k1],[k3])
(i,j)∈([k1],[k2])
6
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
A variable xi,j,(cid:96) represents the number of times the elements ai ∈ A, bj ∈ B and
c(cid:96) ∈ C are used together to form a triple summing to s. The first constraint makes
sure that such a triple is formed only if it sums to s. The remaining equalities make
sure that each element of A ∪ B ∪ C appears in a triple. Thus n such triples are
(cid:117)(cid:116)
formed, all summing to s if the integer program is feasible.
Note that the problem is also fixed-parameter tractable if parameterized by A∪ B
only: we face a No-instance if C > {a + b : a ∈ A, b ∈ B}. A closely related, well
known numerical problem, is the following.
Three multisets A, B, S of n integers each.
variety-Numerical Matching with Target Sums (var -NMTS)
Input:
Parameter: k = A ∪ B ∪ S.
Question: Are there n triples C1, . . . , Cn ∈ A × B × S, such that the A-
element and the B-element from each Ci sum to its S-element?
Corollary 1. var-NMTS is fixed-parameter tractable.
By the previous discussion, the natural parameterization by A ∪ B is also fixed-
parameter tractable. A straightforward adaptation of the proof of Theorem 3 shows
that variety-3-Partition is fixed-parameter tractable.
variety-3-Partition (var -3-Part)
Input:
Parameter: k = A.
Question: Are there n triples S1, . . . , Sn ⊆ A, all summing to the same
A multiset A of 3n integers.
number?
s =(cid:80)
Theorem 4. var-3-Part is fixed-parameter tractable.
Proof. Let A be an instance for var -3-Part, with A = k and A = 3n. Let
a∈A a/n. Let a1, . . . , ak denote the distinct elements of A and let m1, . . . , mk
denote their multiplicities in A. We create an instance of ILPF with at most k3
integer variables xi,j,(cid:96), for i, j, (cid:96) ∈ [k]:
xi,j,(cid:96) = 0
for each i, j, (cid:96) ∈ [k]
such that ai + aj + a(cid:96) (cid:54)= s
(xi,j,(cid:96) + xj,i,(cid:96) + xj,(cid:96),i)
(cid:88)
+2 ·(cid:88)
j,(cid:96)∈[k]
j,(cid:96)(cid:54)=i
j∈[k]
j(cid:54)=i
(xi,i,j + xi,j,i + xj,i,i)
+3 · xi,i,i = mi
∀i ∈ [k]
A variable xi,j,(cid:96) represents the number of times the elements ai, aj and a(cid:96) are used
together to form a triple summing to s. The first constraint makes sure that such a
Parameterizing by the Number of Numbers
7
triple is formed only if it sums to s. The second set of equalities make sure that each
element of A appears in a triple. Thus n such triples are formed, all summing to s if
(cid:117)(cid:116)
the integer program is feasible.
5 Mealy Machines
In this section, we explore how far we can generalize the rather simple FPT results of
the previous two sections. To this end, we investigate the parameterized complexity of
two problems about Mealy Machines. Both problems can be viewed as parameterized
problems implicitly parameterized by the number of numbers, because in each case
the size of the alphabet is part of the parameterization, and each letter of the alphabet
is associated with a census requirement. The richer structure of these problems means
that a simple appeal to integer linear programming no longer suffices: one turns out
to be FPT, and the other W[1]-hard. In Section 6, we show that other problems
parameterized by the number of numbers reduce to these two seemingly general
problems of this kind.
Mealy machines [21] are finite-state transducers, generating an output based on
their current state and input. They have important applications in cryptanalysis [2],
computational linguistics [27], and control and system theory [30]. A deterministic
Mealy machine is a dual-alphabet state transition system given by a 5-tuple M =
(S, s0, Γ, Σ, T ):
-- a finite set of states S,
-- a start state s0 ∈ S,
-- a finite set Γ , called the input alphabet,
-- a finite set Σ, called the output alphabet, and
-- a transition function T : S × Γ → S × Σ mapping pairs of a state and an input
letter to the corresponding next state and output letter.
The alphabets Γ and Σ may contain the empty letter , as in [28]. This eases some
of the description, but all our results also hold if we restrict /∈ Γ ∪ Σ.
In a non-deterministic Mealy machine, the only difference is that the transition
function is defined T : S × Γ → P(S × Σ) as for a given state and input letter, there
may be more than one possibility for the next state and output letter. (Here P(X)
denotes the powerset of a set X.)
A census requirement c : Σ \ {} → N is a function assigning a non-negative
integer to each letter of the output alphabet (except ). It is used to constrain how
many times each letter should appear in the output of a Mealy machine. A word
x ∈ Σ∗ meets the census requirement if every letter b ∈ Σ \ {} appears exactly c(b)
times in x.
The notion of census requirement is related to Parikh images [25]. Let Σ \ {} =
{b1, . . . , bσ}. For x ∈ Σ∗, the Parikh image is Ψ (x) = (c(b1), . . . , c(bσ)), where c is
the census requirement such that x meets c. The Parikh image of a language L is
Ψ (L) = {Ψ (x) : x ∈ L}. Parikh's theorem [25] states that the Parikh image of a
context-free language is semilinear, i.e., that for every context-free language there is
a regular language with the same Parikh image.
8
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
Our first problem about Mealy machines asks whether there exists an input word
and a computation of the Mealy machine such that the output word meets the census
requirement.
variety-Exists Word Mealy Machine (var -EWMM)
Input:
Parameter: S + Γ + Σ.
Question: Does there exist a word x ∈ Γ ∗ for which a computation of M
A non-deterministic Mealy machine M = (S, s0, Γ, Σ, T ), and a
census requirement c : Σ \ {} → N.
on input x generates an output y that meets c?
Our proof that var -EWMM is fixed-parameter tractable is inspired by the proof from
[10] showing that Bandwidth is fixed-parameter tractable when parameterized by
the maximum number of leaves in a spanning tree of the input graph. We need the
following definition and lemma from [10].
In a digraph D, two directed walks ∆ and ∆(cid:48) from a vertex s to a vertex t are
arc-equivalent, if for every arc a of D, ∆ and ∆(cid:48) pass through a the same number of
times.
Lemma 1 ([10]). Any directed walk ∆ through a finite digraph D on n vertices from
a vertex s to a vertex t of D is arc-equivalent to a directed walk ∆(cid:48) from s to t, where
∆(cid:48) has the form:
(1) ∆(cid:48) consists of an underlying directed walk ρ from s to t of length at most n2,
(2) together with some number of short loops, where each such short loop l begins
and ends at a vertex of ρ, and has length at most n.
The algorithm will first subdivide state transitions in order to make the underlying
directed graph simple. As suggested by Lemma 1, the algorithm goes over all possible
choices for selecting an underlying directed walk ρ starting from s0. For every short
loop starting and ending at a vertex from ρ, the algorithm associates an integer
variable representing the number of times this short loop is executed while moving
along ρ. Again by Integer Linear Programming Feasibility, it can be checked
whether there is a set of integers, representing the number of executions of the short
loops, such that the number of times each output letter is written is compatible with
the census requirement.
Theorem 5. var-EWMM is fixed-parameter tractable.
Proof. Let (M(cid:48) = (S(cid:48), s(cid:48)
0, Γ (cid:48), Σ(cid:48), T (cid:48)), c) be an instance for var -EWMM with k =
S(cid:48) + Γ (cid:48) + Σ(cid:48). As M(cid:48) might have multiple transitions from one state to another,
we first subdivide each transition in order to obtain a simple digraph underlying
the Mealy machine (so we can use Lemma 1): create a new non-deterministic Mealy
0, Γ = Γ (cid:48) ∪ {}, and
machine M = (S, s0, Γ, Σ, T ) such that, initially, S = S(cid:48), s0 = s(cid:48)
Σ = Σ(cid:48) ∪ {}; for each transition t of T (cid:48) from a couple (si,(cid:104)i(cid:105)) to a couple (so,(cid:104)o(cid:105)),
add a new state st to S and add the transition from (si,(cid:104)i(cid:105)) to (st,(cid:104)o(cid:105)) and the
transition from (st, ) to (so, ) to T . Clearly, there is at most one transition between
every two states in M .
Parameterizing by the Number of Numbers
9
Our algorithm goes over all transition walks in M of length at most S2 that start
from s0. There are at most S(S2) such transition walks and each such transition
walk has at most SS short loops, as they have length at most S by Lemma 1. Let
P = (s0, s1, . . . , sP) be such a transition walk and L = ((cid:96)1, (cid:96)2, . . . , (cid:96)L) be its short
loops. It remains to check whether there exists a set of integers X = {x1, x2, . . . , xL}
such that a word output by a computation of M moving from s0 to sP along the
walk P , and executing xi times each short loop (cid:96)i, 1 ≤ i ≤ L, meets the census
requirement. Note that if one such word meets the census requirement, then all such
words meet the census requirement, as it does not matter in which order the short
loops are executed. We verify whether such a set X exists by ILPF.
Let Σ \ {} = {(cid:104)(cid:96), 1(cid:105),(cid:104)(cid:96), 2(cid:105), . . . ,(cid:104)(cid:96), σ(cid:105)}. Define m(i, j), for 1 ≤ i ≤ L, 1 ≤ j ≤ σ,
to denote the number of times that M writes the letter (cid:104)(cid:96), j(cid:105) when it executes the
loop (cid:96)i once. Define m(j), for 1 ≤ j ≤ σ, to be the number of times that M writes
the letter (cid:104)(cid:96), j(cid:105) when it transitions from s0 to sP along the walk P . Then, we only
need to verify that there exist integers x1, x2, . . . , xL such that
L(cid:88)
m(j) +
xi · m(i, j) = c((cid:104)(cid:96), j(cid:105)),
∀j ∈ [σ].
i=0
By construction, S ≤ S(cid:48) + T (cid:48) ≤ S(cid:48) + S(cid:48)2 · Γ (cid:48) · Σ(cid:48) ≤ k + k4. As the number of
integer variables of this program is at most L ≤ SS ≤ (k+k4)k+k4, and the number
of transition walks that the algorithm considers is at most S(S2) ≤ (k+k4)k2+2k5+k8,
(cid:117)(cid:116)
var -EWMM is fixed-parameter tractable.
We note that the proof in [10] concerned a special case of a deterministic Mealy
machine where the input and output alphabet are the same, and all transitions that
read a letter (cid:104)(cid:96)(cid:105) also write (cid:104)(cid:96)(cid:105).
In our second Mealy machine problem, the question is whether, for a given input
word, there is a computation of the Mealy machine which outputs a word that meets
the census requirement.
variety-Given Word Mealy Machine (var -GWMM)
Input:
Parameter: S + Γ + Σ
Question:
A non-deterministic Mealy machine M = (S, s0, Γ, Σ, T ), a word
x ∈ Γ ∗, and a census requirement c : Σ \ {} → N.
Is there a computation of M on input x generating an output y
that meets c?
By dynamic programming we show that two restrictions of this problem are in XP.
In the first one, the census requirement is encoded in unary. This restriction of the
problem seems lenient, especially when one is actually interested in finding the output
word, as the census function acts then as a placeholder for the produced word.
Theorem 6. var-GWMM is in XP if c is encoded in unary.
Proof. Let Σ \ {} = σ and Σ \ {} = {b1, . . . , bσ}. Our dynamic programming
algorithm computes the entries of a boolean table A. The table A has an entry A[s, c1,
10
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
. . . , cσ, i, p] for each state s ∈ S, each cj ∈ {0, . . . , c(bj)}, j ∈ [σ], each index i ∈
{0, . . . ,x}, and each integer p ∈ P = {0, . . . ,S − 1}. The entry A[s, c1, . . . , cσ, i, p]
is set to true if there exists a computation of M reading the first i letters of x,
outputting a word y in which the letter bj occurs cj times, for each j ∈ [σ], followed
by p transitions that read and write , and ending up in state s, and to false
otherwise.
Set A[s, c1, . . . , cσ, 0, 0] to true if s = s0 and c1 = . . . = cσ = 0, and to false
i=1 ci, index i,
otherwise. Compute the values of the table by increasing values of(cid:80)σ
propagation integer p, and state number s:
A[s, c1, . . . , cσ, i, p] =
A[s(cid:48), c1, . . . , cj−1, cj − 1, cj+1, . . . , cσ, i − 1, p(cid:48)]
(cid:95)
(cid:95)
(cid:95)
(cid:95)
∨
∨
∨
s(cid:48)∈S,bj∈Σ\{},p(cid:48)∈P :
(s,bj )∈T (s(cid:48),x[i])
s(cid:48)∈S,p(cid:48)∈P :
(s,)∈T (s(cid:48),x[i])
s(cid:48)∈S,bj∈Σ\{},p(cid:48)∈P :
(s,bj )∈T (s(cid:48),)
s(cid:48)∈S:(s,)∈T (s(cid:48),)
A[s(cid:48), c1, . . . , cσ, i − 1, p(cid:48)]
A[s(cid:48), c1, . . . , cj−1, cj − 1, cj+1, . . . , cσ, i, p(cid:48)]
A[s(cid:48), c1, . . . , cσ, i, p − 1]
census requirement if and only if(cid:87)
Finally, there exists an x-computation of M generating a word y that meets the
s∈S,p∈P A[s, c(b1), . . . , c(bσ),x, p] is true. Denote
by n is the length of the description of an input instance. The table has S·x·P·
j=1c(bj) ≤ S2 · nσ+1 entries, and each entry can be computed in time O(S2 · σ).
Π σ
The running time of the algorithm is thus upper bounded by O(nσ+1 · k5), where k
(cid:117)(cid:116)
is the parameter.
For the version where c is encoded in binary, a restriction on the input alphabet gives
an XP algorithm as well.
Corollary 2. var-GWMM is in XP if /∈ Γ .
Proof. If(cid:80)
b∈Σ\{} c(b) > x, then return false, as M cannot output more than x
letters. Otherwise, run the algorithm described in the proof of Theorem 6. Its running
(cid:117)(cid:116)
time is O(k5 · x · Π σ
j=1c(bj)) = O(nσ+1 · k5).
Note that the XP-results also hold if the parameter is only Σ.
To show that var -GWMM is W [1]-hard, we reduce from the Multicolored
Clique problem, which is W [1]-hard [26, 8].
Parameterizing by the Number of Numbers
11
Multicolored Clique (MCC)
Input:
An integer k and a connected undirected graph G = (V (1) ∪
V (2) . . . ∪ V (k), E) such that for every i ∈ [k], the vertices of
V (i) induce an independent set in G.
Parameter: k.
Question:
Is there a clique of size k in G?
Clearly, a solution to this problem has one vertex from each color.
Our parameterized reduction encodes G in the input word x of the Mealy machine
M , and the description of M depends only on k. The Mealy machine is divided into
k parts, one for each color class V (i), with 1 ≤ i ≤ k. Its ith part is responsible for
selecting a vertex vi from V (i) and edges vivj for every vj ∈ V (j), with 1 ≤ j (cid:54)= i ≤ k.
All consistency issues and communication is done via the census requirement. Within
part i, we need to make sure that the selected edges are all incident to the selected
vertex vi. This is achieved by making M output p times each letter (cid:104)l, i, j(cid:105), with
1 ≤ j (cid:54)= i ≤ k, if it selects the pth vertex in V (i). The census requirement for (cid:104)l, i, j(cid:105)
is V (i) + 1, meaning that (cid:104)l, i, j(cid:105) needs to be output V (i) + 1 − p times later. To
select an edge vivj, the machine M will be constrained to select this edge among the
edges incident to vi. To achieve this, edges from V (i) to V (j) appear in x grouped by
the vertex from V (i) on which they are incident. After each group of edges incident
on one vertex from V (i), there is a special state where (cid:104)l, i, j(cid:105) is output if and only
if the edge towards V (j) has already been selected. As (cid:104)l, i, j(cid:105) needs to be output
exactly V (i) + 1 − p times, we force in this way that an edge is selected which is
incident on vi. This enforces that all edges selected in the ith part are incident on the
same vertex. It remains to make sure that distinct parts i and j select the same edge
between V (i) and V (j). This is again achieved by a census requirement where a part
of the census of letter (cid:104)l, ¯e, i, j(cid:105) is output in the ith part and the remaining part in
the jth part of M .
Theorem 7. var-GWMM is W [1]-hard.
Proof. Let (k, G = (V (1) ∪ V (2) . . . ∪ V (k), E)) be an instance of MCC. Suppose
V (i) = {vi,1, vi,2, . . . , vi,V (i)} is the vertex set of color i, for each color class i ∈ [k],
E = {e1, e2, . . . , eE}, and E(i, j) = {e(i, j, 1), e(i, j, 2), . . . , e(i, j,E(i, j))} is the
subset of edges with one vertex in color class i and the other in color class j, for i, j ∈
[k]. Moreover, suppose E(i, j) follows the same order as E, that is if ep = e(i, j, p(cid:48)),
eq = e(i, j, q(cid:48)), and p ≤ q, then p(cid:48) ≤ q(cid:48). For a vertex vi,p and two integers j ∈ [k] \ {i}
and q ∈ [dV (j)(vi,p) + 1], we define gap(vi,p, j, q) = t− s, where e(i, j, t) is the qth edge
in E(i, j) incident to vi,p (respectively, t = E(i, j) if q = dV (j)(vi,p) + 1) and e(i, j, s)
is the (q − 1)th edge in E(i, j) incident to vi,p (respectively, s = 0 if q = 1).
We construct an instance (M = (S, s0, Γ, Σ, T ), x, c) for var -GWMM as follows.
M 's input alphabet, Γ , is {(cid:104)i(cid:105),(cid:104)i, j(cid:105),(cid:104)¯e, i, j(cid:105),(cid:104)e, i, j(cid:105) : i, j ∈ [k], i (cid:54)= j}. M 's output
alphabet, Σ, is {} ∪ {(cid:104)l, i, j(cid:105),(cid:104)l, ¯e, i, j(cid:105) : i, j ∈ [k], i (cid:54)= j}. The word x is defined
12
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
x := x1x2 . . . xk
xi := xi,0xi,1 . . . xi,i−1xi,i+1xi,i+2 . . . xi,k(cid:104)i(cid:105)
xi,0 := ((cid:104)i, 1(cid:105)(cid:104)i, 2(cid:105) . . .(cid:104)i, i − 1(cid:105)(cid:104)i, i + 1(cid:105)(cid:104)i, i + 2(cid:105) . . .(cid:104)i, k(cid:105))V (i)
xi,j := (cid:104)i, j(cid:105)xi,j,1(cid:104)i, j(cid:105)xi,j,2 . . .(cid:104)i, j(cid:105)xi,j,V (i)(cid:104)i, j(cid:105)
xi,j,p := (cid:104)¯e, i, j(cid:105)gap(vi,p,j,1)(cid:104)e, i, j(cid:105)(cid:104)¯e, i, j(cid:105)gap(vi,p,j,2)(cid:104)e, i, j(cid:105)
. . .(cid:104)¯e, i, j(cid:105)gap(vi,p,j,dV (j)(vi,p))(cid:104)e, i, j(cid:105)(cid:104)¯e, i, j(cid:105)gap(vi,p,j,dV (j)(vi,p)+1).
∀i ∈ [k]
∀i ∈ [k]
∀i, j ∈ [k], i (cid:54)= j
The census requirement c is, for every i, j ∈ [k], i (cid:54)= j,
c((cid:104)l, i, j(cid:105)) := V (i) + 1
c((cid:104)l, ¯e, i, j(cid:105)) := E(i, j).
On reading a subword xi, the Mealy machine will select a vertex vi,p in V (i) and one
edge incident to vi,p for each color class j ∈ [k] \ {i}. The vertex vi,p is selected in
the subword xi,0 of xi. Next, for each j ∈ [k] \ {i}, a vertex in V (i) and a vertex in
V (j) are selected in the subword xi,j. The census requirement for (cid:104)l, i, j(cid:105) makes sure
that the vertex from V (i) is vi,p. The subword xi,j,p ensures that vi,p and the vertex
that is selected from V (j) are joined by an edge. Finally, the census requirement for
(cid:104)l, ¯e, i, j(cid:105) is responsible for the inter-partition communication and makes sure that
the edge selected in xi,j is equal to the edge selected in xj,i.
The Mealy machine M consists of k parts. The ith part of M is depicted in Fig. 1.
Its initial state is sv,1. There is a transition from the last state of each part, s(4)
e,i,k, to
the first state of the following part, sv,i+1 (from the kth part, there is a transition to a
final state): it reads the letter (cid:104)i(cid:105) and writes the letter . We set (cid:104)l(cid:48), ¯e, i, j(cid:105) = (cid:104)l, ¯e, j, i(cid:105)
for all i (cid:54)= j ∈ [k]. Note that, in the description of M , the letter (cid:104)l, i, j(cid:105) can only
be output on reading (cid:104)i, j(cid:105), and (cid:104)l, ¯e, i, j(cid:105) can only be output on reading (cid:104)¯e, i, j(cid:105) or
(cid:104)¯e, j, i(cid:105).
Let us first verify that the parameter for var -GWMM is a function of k, and
that there exists a function f such that the size of the instance for var -GWMM is
f (k) · nO(1), where n is the number of vertices of G. We have Γ = k + 3 · k · (k − 1),
Σ = 1 + 2· k· (k− 1), and S = 1 + k· (2 + 4· (k− 1)). The parameter of var -GWMM
is thus bounded by a function of k. The length of x is O(k2 · n3). Now, we show that
(M, x, c) is a Yes-instance for var -GWMM if and only if (G, k) is a Yes-instance
for MCC.
First, suppose (M = (S, s0, Γ, Σ, T ), x, c) is a Yes-instance for var -GWMM.
We say that M selects a vertex vi,p if it makes a transition from state sv,i to state
v,i reading (cid:104)i, k(cid:105) (respectively (cid:104)i, k − 1(cid:105) if i = k) for the pth time. In other words,
s(cid:48)
in the ith part of M , it reads p · (k − 1) − 1 letters of xi,0, staying in state sv,i and
outputs the letter (cid:104)l, i, r(cid:105) for each letter (cid:104)i, r(cid:105) it reads; then it transitions to state s(cid:48)
on reading (cid:104)i, k(cid:105) (respectively (cid:104)k, k − 1(cid:105) if i = k) and outputs (cid:104)l, i, k(cid:105) (respectively
(cid:104)l, k, k − 1(cid:105)); in the state s(cid:48)
v,i it outputs the empty letter for each letter (cid:104)i, r(cid:105) it reads.
v,i
Parameterizing by the Number of Numbers
13
Fig. 1. The ith part of the Mealy machine M . It does not have the states s(1)
there is instead a transition from s(4)
transition from s(4)
at the last state. (Drawing all this would have cluttered the figure too much.)
e,i,i, and s(4)
e,i,i;
e,i,i+1 reading (cid:104)i − 1(cid:105) and writing , and there is a
e,k,k−1 to the last state reading (cid:104)k(cid:105) and writing . There are no transitions starting
e,i,i−1 to s(1)
e,i,i, s(2)
e,i,i, s(3)
We say that M selects an edge e(i, j, q) if it makes a transition from state s(2)
e,i,j
e,i,j after having read the letter (cid:104)¯e, i, j(cid:105) of xi,j,p exactly q times, where vi,p
to state s(3)
is the vertex of color i that e(i, j, q) is incident on. In other words, in the ith part of
M , it transitions from the state s(1)
e,i,j on reading the first letter of
xi,j,p (if it did this transition any later, the census requirement of (cid:104)l, ¯e, i, j(cid:105) could not
be met, as shown in the proof of Claim 2 below); then it stays in the state s(2)
e,i,j until
it has read q times the letter (cid:104)¯e, i, j(cid:105) of xi,j,p; then it transitions to the state s(3)
e,i,j on
reading (cid:104)e, i, j(cid:105); it stays in this state and outputs (cid:104)l(cid:48), ¯e, i, j(cid:105) for each letter (cid:104)¯e, i, j(cid:105) it
reads until transitioning to the state s(4)
e,i,j to the state s(2)
e,i,j on reading the letter following xi,j,p.
sv,is0v,is(1)e,i,1s(2)e,i,1s(3)e,i,1s(4)e,i,1s(1)e,i,2s(2)e,i,2s(3)e,i,2s(4)e,i,2s(1)e,i,ks(2)e,i,ks(3)e,i,ks(4)e,i,khi−1i,hi,1i,hl,i,1ihi,2i,hl,i,2i...hi,ki,hl,i,kihi,ki,hl,i,kihi,1i,hi,2i,...hi,ki,hi,1i,hi,1i,he,i,1i,h¯e,i,1i,h¯e,i,1i,hl,¯e,i,1ihe,i,1i,h¯e,i,1i,hl,¯e,i,1ihe,i,1i,he,i,1i,h¯e,i,1i,hl0,¯e,i,1ihi,1i,hl,i,1ihi,1i,hl,i,1ihe,i,1i,h¯e,i,1i,hi,2i,hi,2i,he,i,2i,h¯e,i,2i,h¯e,i,2i,hl,¯e,i,2ihe,i,2i,h¯e,i,2i,hl,¯e,i,2ihe,i,2i,he,i,2i,h¯e,i,2i,hl0,¯e,i,2ihi,2i,hl,i,2ihi,2i,hl,i,2ihe,i,2i,h¯e,i,2i,hi,ki,he,i,ki,h¯e,i,ki,h¯e,i,ki,hl,¯e,i,kihe,i,ki,h¯e,i,ki,hl,¯e,i,kihe,i,ki,he,i,ki,h¯e,i,ki,hl0,¯e,i,kihi,ki,hl,i,kihi,ki,hl,i,kihe,i,ki,h¯e,i,ki,hii, 14
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
The following claims ensure that the edge-selection and the vertex-selection are
compatible, i.e., that exactly one edge is selected from color i to color j, and that
this edge is incident on the selected vertex of color i.
Claim 1. Let i be a color and let vi,p be the vertex selected in the ith part of M . In
its ith part, M selects one edge incident to vi,p and to a vertex of color j, for each
j ∈ [k] \ {i}.
Proof. After M has selected vi,p, it has output p times each of the letters (cid:104)l, i, 1(cid:105),
(cid:104)l, i, 2(cid:105), . . . ,(cid:104)l, i, i − 1(cid:105),(cid:104)l, i, i + 1(cid:105),(cid:104)l, i, i + 2(cid:105), . . . ,(cid:104)l, i, k(cid:105). For each j ∈ [k] \ {i}, the
only other transitions that output (cid:104)l, i, j(cid:105) are the transition from s(3)
e,i,j and a
e,i,j. To meet the census requirement of V (i)+1 for (cid:104)l, i, j(cid:105),
transition that loops on s(4)
M selects an edge while reading xi,j,p. This edge is incident on vi,p by construction.
(cid:117)(cid:116)
e,i,j to s(4)
The following claim makes sure that the edge selected from color i to color j is the
same as the edge selected from color j to color i.
Claim 2. Suppose M selects the edge e(i, j, q) in its ith part. Then, M selects the
edge e(j, i, q) in its jth part.
Proof. Before M selects e(i, j, q), it has output q(cid:48) ≤ q times the letter (cid:104)l, ¯e, i, j(cid:105). On
selecting e(i, j, q) it transitions to the state s(3)
e,i,j, and after the selection it outputs
(cid:104)l(cid:48), ¯e, i, j(cid:105) for every letter (cid:104)¯e, i, j(cid:105) of xi,j,p it reads. As it reads
dV (j)(vi,p)+1(cid:88)
(
gap(vi,p, j, r)) − q = E(i, j) − q
r=1
times the letter (cid:104)¯e, i, j(cid:105) of xi,j,p after it has selected e(i, j, q), the Mealy machine
outputs E(i, j) − q times the letter (cid:104)l(cid:48), ¯e, i, j(cid:105) in its ith part.
The only other transition where it outputs (cid:104)l, ¯e, i, j(cid:105) = (cid:104)l(cid:48), ¯e, j, i(cid:105) is the transition
e,j,i that reads (cid:104)¯e, j, i(cid:105) and outputs (cid:104)l(cid:48), ¯e, j, i(cid:105). To meet
in the jth part of M looping on s(3)
the census requirement for (cid:104)l, ¯e, i, j(cid:105), this transition must be used exactly E(i, j)− q(cid:48)
times.
The only other transitions where it outputs (cid:104)l(cid:48), ¯e, i, j(cid:105) = (cid:104)l, ¯e, j, i(cid:105) are two transi-
tions in the jth part of M : the transition from s(1)
e,j,i and the transition looping
e,j,i, both reading (cid:104)¯e, j, i(cid:105) and writing (cid:104)l, ¯e, j, i(cid:105). These transitions can be used at
on s(2)
most q(cid:48) times as the transition of the previous paragraph is used E(i, j) − q(cid:48) times.
These transitions have to be used at least q times to meet the census requirement for
(cid:104)l(cid:48), ¯e, i, j(cid:105). Thus, these transitions are used exactly q times and q = q(cid:48).
e,j,i happens after having read q times the
letter (cid:104)¯e, j, i(cid:105) of some vertex xj,i,p(cid:48), p(cid:48) ∈ [V (j)], which means that M selects the edge
(cid:117)(cid:116)
e(j, i, q) in its jth part.
Finally, the transition from s(2)
e,j,i to s(3)
e,j,i to s(2)
By Claims 1 and 2, the k vertices that are selected by M form a multicolored clique.
Thus, (k, G = (V (1) ∪ V (2) . . . ∪ V (k), E)) is a Yes-instance for MCC.
Parameterizing by the Number of Numbers
15
Now, suppose that (k, G = (V (1) ∪ V (2) . . . ∪ V (k), E)) is a Yes-instance for MCC.
Let {v1,p1, v2,p2, . . . , vk,pk} be a multicolored clique in G. We will construct a word
y meeting c such that a computation of M on input x generates y. For two adjacent
vertices vi,pi and vj,pj , define edge(vi,pi, vj,pj ) = t such that e(i, j, t) = vi,pivj,pj . The
word y is y1y2 . . . yk, where yi, for i ∈ [k], is
((cid:104)l, i, 1(cid:105)(cid:104)l, i, 2(cid:105) . . .(cid:104)l, i, i − 1(cid:105)(cid:104)l, i, i + 1(cid:105)(cid:104)l, i, i + 2(cid:105) . . .(cid:104)l, i, k(cid:105))pi
yi,1yi,2 . . . yi,i−1yi,i+1yi,i+2 . . . yi,k
and yi,j, for i (cid:54)= j ∈ [k], is
(cid:104)l, ¯e, i, j(cid:105)edge(vi,pi ,vj,pj
)(cid:104)l(cid:48), ¯e, i, j(cid:105)E(i,j)−edge(vi,pi ,vj,pj
)(cid:104)l, i, j(cid:105)V (i)−pi+1.
We note that y meets the census requirement c. Moreover, the computation of M on
input x, which selects (as defined in the first part of the proof) exactly the vertices
and edges of the multicolored clique {v1,p1, v2,p2, . . . , vk,pk}, outputs y. Thus (M =
(cid:117)(cid:116)
(S, s0, Γ, Σ, T ), x, c) is a Yes-instance for var -GWMM.
The theorem holds if we restrict /∈ Γ ∪ Σ. Indeed, /∈ Γ in the target instance,
and one can add a new letter e to Σ, which replaces and has census requirement
i,j∈[k],i(cid:54)=j(c((cid:104)l, i, j(cid:105)) + c((cid:104)l, ¯e, i, j(cid:105)). This instance is equivalent since the
c(e) = x −(cid:80)
modified M outputs one letter for each letter in x.
6 Applications
In this section we sketch two examples that illustrate how number-of-numbers pa-
rameterized problems may reduce to census problems about Mealy machines, param-
eterized by the size of the machine. For another application, see [10].
Example 1: Heat-Sensitive Scheduling. In a recent paper Chrobak et al. [5]
introduced a model for the issue of temperature-aware task scheduling for micropro-
cessor systems. The motivation is that different jobs with the same time requirements
may generate different heat loads, and it may be important to schedule the jobs so
that some temperature threshold is not breached.
In the model, the input consists of a set of jobs that are all assumed to be of
unit length, with each job assigned a numerical heat level. If at time t the processor
temperature is Tt, and if the next job that is scheduled has heat level H, then the
processor temperature at time t + 1 is
Tt+1 = (Tt + H)/2
It is also allowed that perhaps no job is scheduled for time t + 1 (that is, idle time is
scheduled), in which case H = 0 in the above calculation of the updated temperature.
The relevant decision problem is whether all of the jobs can be scheduled, meet-
ing a specified deadline, in such a way that a given temperature threshold is never
exceeded. This problem has been shown to be NP-hard [5] by a reduction from 3-
Dimensional Matching. An image instance of the reduction, however, involves
16
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
arbitrarily many distinct heat levels asymptotically close to H = 2, for a tempera-
ture threshold of 1.
In the spirit of the "deconstruction of hardness proofs" advocated by Komusiewicz
et al. [19] (see also [4, 24]), one might regard this problem as ripe for parameterization
by the number of numbers, for example (scaling appropriately), a model based on 2k
equally-spaced heat levels and a temperature threshold of k. Furthermore, if the heat
levels of the jobs are only roughly classified in this way, it also makes sense to treat
the temperature transition model similarly, as:
Tt+1 = (cid:100)(Tt + H)/2(cid:101)
The input to the problem can now be viewed equivalently as a census of how
many jobs there are for each of the 2k + 1 heat levels, with the available potential
units of idle time allowed to meet the deadline treated as "jobs" for which H = 0.
Because of the ceiling function modeling the temperature transition, the problem now
immediately reduces to var -EWMM, for a machine on k + 1 states (that represent
the temperature of the processor) and an alphabet of size at most 2k +1. By Theorem
5, the problem is fixed-parameter tractable.
Example 2: A Problem in Computational Chemistry. The parameterized
problem of Weighted Splits Reconstruction for Paths that arises in com-
putational chemistry [15] reduces to a special case of var -GWMM. The input to
the problem is obtained from time-series spectrographic data concerning molecular
weights. The problem as defined in [15] is equivalent to the following two-processor
scheduling problem. The input consists of
-- a sequence x of positive integer time gaps taken from a set of positive integers Γ ,
and
-- a census requirement c on a set of positive integers Σ of job lengths.
The question is whether there is a "winning play" for the following one-person two-
processor scheduling game. At each step, first, Nature plays the next positive integer
"gap" of the sequence of time gaps x -- this establishes the next immediate deadline.
Second, the Player responds by scheduling on one of the two processors, a job that
begins at the last stop-time on that processor, and ends at the immediate deadline.
The Player wins if there is a sequence of plays (against x) that meets the census
requirement c on job lengths. Fig. 2 illustrates such a game.
Processor 1
Processor 2
x =
4
4
5
1
3
2
3
1
1
1
3
1
5
4
Fig. 2. A winning game for the census: 1 (1), 3 (3), 4 (1), 5 (2)
This problem easily reduces to a special case of var -GWMM. Whether this special
case is also W [1]-hard remains open.
Parameterizing by the Number of Numbers
17
7 Concluding Remarks
The practical world of computing is full of computational problems where inputs
are "weighted" in a realistic model -- weighted graphs provide a simple example
relevant to many applications. Here we have begun to explore parameterizing on the
numbers of numbers as a way of mitigating computational complexity for problems
that are numerically structured. One might view some of the impulse here as moving
approximation issues into the modeling, as illustrated by Example 1 in Section 6. We
believe this line of attack may be widely applicable.
To date, there has been little attention to parameterized complexity issues in
the context of cryptography, control theory, and other numerically structered areas
of application. Number of numbers parameterization may provide some inroads into
these underdeveloped areas.
Our main FPT result, Theorem 5, has a poor worst-case running-time guarantee.
Can this be improved -- at least in important special cases?
Acknowledgment. We thank Iyad Kanj for stimulating conversations about this
work.
References
1. Noga Alon, Yossi Azar, Gerhard J. Woeginger, and Tal Yadid. Approximation schemes for
scheduling on parallel machines. Journal of Scheduling, 1:55 -- 66, 1998.
2. Gregory V. Bard. Algebraic cryptanalysis. Springer, 2009.
3. Cristina Bazgan. Sch´emas d'approximation et complexit´e param´etr´ee. Master's thesis, Universit´e
Paris Sud, 1995.
4. Nadja Betzler, Michael R. Fellows, Jiong Guo, Rolf Niedermeier, and Frances A. Rosamond.
Fixed-parameter algorithms for kemeny rankings. Theoretical Computer Science, 410(45):4554 --
4570, 2009.
5. Marek Chrobak, Christoph Durr, Mathilde Hurand, and Julien Robert. Algorithms for
In Rudolf Fleischer and Jin-
temperature-aware task scheduling in microprocessor systems.
hui Xu, editors, AAIM 2008, volume 5034 of Lecture Notes in Computer Science, pages 120 -- 130.
Springer, 2008.
6. Rodney G. Downey and Michael R. Fellows. Parameterized complexity. Springer-Verlag, New
York, 1999.
7. Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond. Parameterizing by the number of
numbers. In Venkatesh Raman and Saket Saurabh, editors, IPEC 2010, volume 6478 of Lecture
Notes in Computer Science, pages 123 -- 134. Springer, 2010.
8. Michael R. Fellows, Danny Hermelin, Frances Rosamond, and St´ephane Vialette. On the parame-
terized complexity of multiple-interval graph problems. Theoretical Computer Science, 410(1):53 --
61, 2009.
9. Michael R. Fellows and Neal Koblitz. Fixed-parameter complexity and cryptography.
In
G´erard D. Cohen, Teo Mora, and Oscar Moreno, editors, AAECC 1993, volume 673 of Lec-
ture Notes in Computer Science, pages 121 -- 131. Springer, 1993.
10. Michael R. Fellows, Daniel Lokshtanov, Neeldhara Misra, Matthias Mnich, Frances A. Rosamond,
and Saket Saurabh. The complexity ecology of parameters: An illustration using bounded max
leaf number. Theory of Computing Systems, 45(4):822 -- 848, 2009.
11. Michael R. Fellows, Daniel Lokshtanov, Neeldhara Misra, Frances A. Rosamond, and Saket
Saurabh. Graph layout problems parameterized by vertex cover. In Seok-Hee Hong, Hiroshi
Nagamochi, and Takuro Fukunaga, editors, ISAAC 2008, volume 5369 of Lecture Notes in Com-
puter Science, pages 294 -- 305. Springer, 2008.
12. Jir´ı Fiala, Petr A. Golovach, and Jan Kratochv´ıl. Parameterized complexity of coloring problems:
Treewidth versus vertex cover. Theoretical Computer Science, 412(23):2513 -- 2523, 2011.
18
Michael R. Fellows, Serge Gaspers, and Frances A. Rosamond
13. Jorg Flum and Martin Grohe. Parameterized Complexity Theory, volume XIV of Texts in The-
oretical Computer Science. An EATCS Series. Springer, Berlin, 2006.
14. Andr´as Frank and ´Eva Tardos. An application of simultaneous diophantine approximation in
combinatorial optimization. Combinatorica, 7(1):49 -- 65, 1987.
15. Serge Gaspers, Mathieu Liedloff, Maya J. Stein, and Karol Suchan. Complexity of splits recon-
struction for low-degree trees. In Jan Kratochv´ıl, editor, WG 2011, volume 6986 of Lecture Notes
in Computer Science. Springer, 2011. To appear.
16. Jens Gramm, Rolf Niedermeier, and Peter Rossmanith. Fixed-parameter algorithms for closest
string and related problems. Algorithmica, 37(1):25 -- 42, 2003.
17. Ravi Kannan. Minkowski's convex body theorem and integer programming. Mathematics of
Operations Research, 12(3):415 -- 440, 1987.
18. Donald E. Knuth. The Art of Computer Programming, Volume III: Sorting and Searching.
Addison-Wesley, 1973.
19. Christian Komusiewicz, Rolf Niedermeier, and Johannes Uhlmann. Deconstructing intractabil-
ity - a multivariate complexity analysis of interval constrained coloring. Journal of Discrete
Algorithms, 9(1):137 -- 151, 2011.
20. Hendrik W. Lenstra. Integer programming with a fixed number of variables. Mathematics of
Operations Research, 8(4):538 -- 548, 1983.
21. George H. Mealy. A method for synthesizing sequential circuits. Bell System Technical Journal,
34(5):1045 -- 1079, 1955.
22. Ian Munro and Philip M. Spira. Sorting and searching in multisets. SIAM Journal on Computing,
5(1):1 -- 8, 1976.
23. Rolf Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford Lecture Series in Mathe-
matics and Its Applications. Oxford University Press, Oxford, 2006.
24. Rolf Niedermeier. Reflections on multivariate algorithmics and problem parameterization. In
Jean-Yves Marion and Thomas Schwentick, editors, STACS 2010, volume 5 of LIPIcs, pages
17 -- 32. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2010.
25. Rohit J. Parikh. On context-free languages. Journal of the ACM, 13(4):570 -- 581, 1966.
26. Krzysztof Pietrzak. On the parameterized complexity of the fixed alphabet shortest common
supersequence and longest common subsequence problems. Journal of Computer and System
Sciences, 67(4):757 -- 771, 2003.
27. Emmanuel Roche and Yves Schabes. Finite-state language processing. The MIT Press, 1997.
28. John E. Savage. Models of computation - exploring the power of computing. Addison-Wesley,
1998.
29. Sandeep Sen and Neelima Gupta. Distribution-sensitive algorithms. Nordic Journal of Comput-
ing, 6:194 -- 211, 1999.
30. Eduardo Sontag. Mathematical Control Theory: Deterministic Finite Dimensional Systems,
Second Edition. Springer, 1998.
|
1205.1312 | 1 | 1205 | 2012-05-07T08:02:48 | Converting online algorithms to local computation algorithms | [
"cs.DS"
] | We propose a general method for converting online algorithms to local computation algorithms by selecting a random permutation of the input, and simulating running the online algorithm. We bound the number of steps of the algorithm using a query tree, which models the dependencies between queries. We improve previous analyses of query trees on graphs of bounded degree, and extend the analysis to the cases where the degrees are distributed binomially, and to a special case of bipartite graphs.
Using this method, we give a local computation algorithm for maximal matching in graphs of bounded degree, which runs in time and space O(log^3 n).
We also show how to convert a large family of load balancing algorithms (related to balls and bins problems) to local computation algorithms. This gives several local load balancing algorithms which achieve the same approximation ratios as the online algorithms, but run in O(log n) time and space.
Finally, we modify existing local computation algorithms for hypergraph 2-coloring and k-CNF and use our improved analysis to obtain better time and space bounds, of O(log^4 n), removing the dependency on the maximal degree of the graph from the exponent. | cs.DS | cs |
Converting Online Algorithms to Local Computation
Algorithms
Yishay Mansour1 ⋆, Aviad Rubinstein1 ⋆⋆, Shai Vardi1 ⋆ ⋆ ⋆, and Ning Xie2 †
1 School of Computer Science, Tel Aviv University, Israel
2 CSAIL, MIT, Cambridge MA 02139, USA
Abstract. We propose a general method for converting online algorithms to local
computation algorithms,3 by selecting a random permutation of the input, and
simulating running the online algorithm. We bound the number of steps of the
algorithm using a query tree, which models the dependencies between queries.
We improve previous analyses of query trees on graphs of bounded degree, and
extend the analysis to the cases where the degrees are distributed binomially, and
to a special case of bipartite graphs.
Using this method, we give a local computation algorithm for maximal matching
in graphs of bounded degree, which runs in time and space O(log3 n).
We also show how to convert a large family of load balancing algorithms (related
to balls and bins problems) to local computation algorithms. This gives several
local load balancing algorithms which achieve the same approximation ratios as
the online algorithms, but run in O(log n) time and space.
Finally, we modify existing local computation algorithms for hypergraph 2-coloring
and k-CNF and use our improved analysis to obtain better time and space bounds,
of O(log4 n), removing the dependency on the maximal degree of the graph from
the exponent.
1 Introduction
1.1 Background
The classical computation model has a single processor which has access to a given
input, and using an internal memory, computes the output. This is essentially the von
Newmann architecture, which has been the driving force since the early days of com-
putation. The class of polynomial time algorithms is widely accepted as the definition
⋆ [email protected]. Supported in part by the Google Inter-university center for
Electronic Markets and Auctions, by a grant from the Israel Science Foundation, by a grant
from United States-Israel Binational Science Foundation (BSF), by a grant from Israeli Centers
of Research Excellence (ICORE), and by a grant from the Israeli Ministry of Science (MoS).
⋆⋆ [email protected].
⋆ ⋆ ⋆ [email protected]. Supported in part by the Google Inter-university center for
Electronic Markets and Auctions
† [email protected]. Supported by NSF grants CCF-0728645, CCF-0729011 and
CCF-1065125.
3 For a given input x, local computation algorithms support queries by a user to values of spec-
ified locations yi in a legal output y ∈ F (x).
2
of efficiently computable problems. Over the years many interesting variations of this
basic model have been studied, focusing on different issues.
Online algorithms (see, e.g., [7]) introduce limitations in the time domain. An online
algorithm needs to select actions based only on the history it observed, without access to
future inputs that might influence its performance. Sublinear algorithms (e.g. [11, 15])
limit the space domain, by limiting the ability of an algorithm to observe the entire
input, and still strive to derive global properties of it.
Local computation algorithms (LCAs) [16] are a variant of sublinear algorithms.
The LCA model considers a computation problem which might have multiple admis-
sible solutions, each consisting of multiple bits. The LCA can return queries regarding
parts of the output, in a consistent way, and in poly-logarithmic time. For example, the
input for an LCA for a job scheduling problem consists of the description of n jobs
and m machines. The admissible solutions might be the allocations of jobs to machines
such that the makespan is at most twice the optimal makespan. On any query of a job,
the LCA answers quickly the job's machine. The correctness property of the LCA guar-
antees that different query replies will be consistent with some admissible solution.
1.2 Our results
Following [2], we use an abstract tree structure - query trees to bound the number of
queries performed by certain algorithms. We use these bounds to improve the upper
bound the time and space requirements of several algorithms introduced in [2]. We also
give a generic method of transforming online algorithms to LCAs, and apply it to obtain
LCAs to maximal matching and several load balancing problems.
1.2.1 Bounds on query trees Suppose that we have an online algorithm where the
reply to a query depends on the replies to a small number of previous queries. The reply
to each of those previous queries depends on the replies to a small number of other
queries and so on. These dependencies can be used to model certain problems using
query trees – trees which model the dependency of the replies to a given query on the
replies to other queries.
Bounding the size of a query tree is central to the analyses of our algorithms. We
show that the size of the query tree is O(log n) w.h.p., where n is the number of vertices.
d, the degree bound of the dependency graph, appears in the constant. 4 This answers in
the affirmative the conjecture of [2]. Previously, Alon et al. [2] show that the expected
size of the query tree is constant, and O(logd+1 n) w.h.p.5 Our improvement is signif-
icant in removing the dependence on d from the exponent of the logarithm. We also
show that when the degrees of the graph are distributed binomially, we can achieve the
same bound on the size of the query tree. In addition, we show a trivial lower bound of
Ω(log n/ log log n).
4 Note that, however, the hidden constant is exponentially dependent on d. Whether or not this
bound can be improved to have a polynomial dependency on d is an interesting open question.
5 Notice that bounding the expected size of the query tree is not enough for our applications,
since in LCAs we need to bound the probability that any query fails.
3
We use these results on query trees to obtain LCAs for several online problems –
maximal matching in graphs of bounded degree and several load balancing problems.
We also use the results to improve the previous algorithms for hypergraph 2-coloring
and k-CNF.
1.2.2 Hypergraph 2-coloring We modify the algorithm of [2] for an LCA for hyper-
graph 2-coloring, and coupled with our improved analysis of query tree size, obtain an
LCA which runs in time and space O(log4 n), improving the previous result, an LCA
which runs O(logd+1 n) time and space.
1.2.3 k-CNF Building on the similarity between hypergraph 2-coloring and k-CNF,
we apply our results on hypergraph 2-coloring to give an an LCA for k-CNF which runs
in time and space O(log4 n).
We use the query tree to transform online algorithms to LCAs. We simulate online
algorithms as follows: first a random permutation of the items is generated on the fly.
Then, for each query, we simulate the online algorithm on a stream of input items ar-
riving according to the order of the random permutation. Fortunately, because of the
nature of our graphs (the fact that the degree is bounded or distributed binomially), we
show that in expectation, we will only need to query a constant number of nodes, and
only O(log n) nodes w.h.p. We now state our results:
1.2.4 Maximal matching We simulate the greedy online algorithm for maximal
matching, to derive an LCA for maximal matching which runs in time and space O(log3 n).
1.2.5 Load Balancing We give several LCAs to load balancing problems which run
in O(log n) time and space. Our techniques include extending the analysis of the query
tree size to the case where the degrees are selected from a binomial distribution with
expectation d, and further extending it to bipartite graphs which exhibit the characteris-
tics of many balls and bins problems, specifically ones where each ball chooses d bins
at random. We show how to convert a large class of the "power of d choices" online
algorithms (see, e.g., [3, 6, 18]) to efficient LCAs.
1.3 Related work
Nguyen and Onak [13] focus on transforming classical approximation algorithms into
constant-time algorithms that approximate the size of the optimal solution of problems
such as vertex cover and maximum matching. They generate a random number r ∈
[0, 1], called the rank, for each node. These ranks are used to bound the query tree size.
Rubinfeld et al. [16] show how to construct polylogarithmic time local computa-
tion algorithms to maximal independent set computations, scheduling radio network
broadcasts, hypergraph coloring and satisfying k-SAT formulas. Their proof technique
uses Beck's analysis in his algorithmic approach to the Lovász Local Lemma [4], and
a reduction from distributed algorithms. Alon et al. [2], building on the technique of
4
[13], show how to extend several of the algorithms of [16] to perform in polylogarith-
mic space as well as time. They further observe that we do not actually need to assign
each query a rank, we only need a random permutation of the queries. Furthermore,
assuming the query tree is bounded by some k, the query to any node depends on at
most k queries to other nodes, and so a k-wise independent random ordering suffices.
They show how to construct a 1/n2-almost k-wise independent random ordering6 from
a seed of length O(k log2 n).
Recent developments in sublinear time algorithms for sparse graph and combinato-
rial optimization problems have led to new constant time algorithms for approximating
the size of a minimum vertex cover, maximal matching, maximum matching, minimum
dominating set, and other problems (cf. [15, 11, 13, 20]), by randomly querying a con-
stant number of vertices. A major difference between these algorithms and LCAs is
that LCAs require that w.h.p., the output will be correct on any input, while optimiza-
tion problems usually require a correct output only on most inputs. More importantly,
LCAs reuire a consistent output for each query, rather than only approximating a given
global property.
There is a vast literature on the topic of balls and bins and the power of d choices.
(e.g. [3, 6, 9, 18]). For a survey on the power of d choices, we refer the reader to [12].
1.4 Organization of our paper
The rest of the paper is organized as follows: Some preliminaries and notations that we
use throughout the paper appear in Section 2. In Section 3 we prove the upper bound of
O(log n) on the size of the query tree in the case of bounded and binomially distributed
degrees. In section 4, we use this analysis to give improved algorithms for hypergraph
2-coloring and k-CNF. In Section 5 we give an LCA for finding a maximal matching
in graphs of bounded degree. Section 6 expands our query tree result to a special case
of bipartite graphs; we use this bound for bipartite graph to convert online algorithms
for balls and bins into LCAs for the same problems. The appendices provide in-depth
discussions of the hypergraph 2-coloring and analogous k-CNF LCAs, and a lower
bound to the query tree size.
2 Preliminaries
Let G = (V, E) be an undirected graph. We denote by NG(v) = {u ∈ V (G) : (u, v) ∈
E(G)} the neighbors of vertex v, and by degG(v) we denote the degree of v. When it
is clear from the context, we omit the G in the subscript. Unless stated otherwise, all
logarithms in this paper are to the base 2. We use [n] to denote the set {1, . . . , n}, where
n ≥ 1 is a natural number.
We present our model of local computation algorithms (LCAs): Let F be a com-
putational problem and x be an input to F . Let F (x) = {y y is a valid solution
for input x}. The search problem for F is to find any y ∈ F (x).
6 A random ordering Dr is said to be ǫ-almost k-wise independent if the statistical distance
between Dr and some k-wise independent random ordering by at most ǫ.
5
A (t(n), s(n), δ(n))-local computation algorithm A is a (randomized) algorithm
which solves a search problem for F for an input x of size n. However, the LCA A
does not output a solution y ∈ F (x), but rather implements query access to y ∈ F (x).
A receives a sequence of queries i1, . . . , iq and for any q > 0 satisfies the following: (1)
after each query ij it produces an output yij , (2) With probability at least 1 − δ(n) A is
consistent, that is, the outputs yi1 , . . . , yiq are substrings of some y ∈ F (x). (3) A has
access to a random tape and local computation memory on which it can perform current
computations as well as store and retrieve information from previous computations.
We assume that the input x, the local computation tape and any random bits used
are all presented in the RAM word model, i.e., A is given the ability to access a word
of any of these in one step. The running time of A on any query is at most t(n), which
is sublinear in n, and the size of the local computation memory of A is at most s(n).
Unless stated otherwise, we always assume that the error parameter δ(n) is at most
some constant, say, 1/3. We say that A is a strongly local computation algorithm if
both t(n) and s(n) are upper bounded by O(logc n) for some constant c.
Two important properties of LCAs are as follows. We say an LCA A is query order
oblivious (query oblivious for short) if the outputs of A do not depend on the order of
the queries but depend only on the input and the random bits generated on the random
tape of A. We say an LCA A is parallelizable if A supports parallel queries, that is A
is able to answer multiple queries simultaneously so that all the answers are consistent.
3 Bounding the size of a random query tree
3.1 The problem and our main results
In online algorithms, queries arrive in some unknown order, and the reply to each query
depends only on previous queries (but not on any future events). The simplest way to
transform online algorithms to LCAs is to process the queries in the order in which they
arrive. This, however, means that we have to store the replies to all previous queries,
so that even if the time to compute each query is polylogarithmic, the overall space is
linear in the number of queries. Furthermore, this means that the resulting LCA is not
query-oblivious. The following solution can be applied to this problem ([13] and [2]):
Each query v is assigned a random number, r(v) ∈ [0, 1], called its rank, and the queries
are performed in ascending order of rank. Then, for each query x, a query tree can be
constructed, to represent the queries on which x depends. If we can show that the query
tree is small, we can conclude that each query does not depend on many other queries,
and therefore a small number of queries need to be processed in order to reply to query
x. We formalize this as follows:
Let G = (V, E) be an undirected graph. The vertices of the graph represent queries,
and the edges represent the dependencies between the queries. A real number r(v) ∈
[0, 1] is assigned independently and uniformly at random to every vertex v ∈ V ; we call
r(v) the rank of v. This models the random permutation of the vertices. Each vertex
v ∈ V holds an input x(v) ∈ R, where the range R is some finite set. The input is the
content of the query associated with v. A randomized function F is defined inductively
on the vertices of G such that F (v) is a (deterministic) function of x(v) as well as the
6
values of F at the neighbors w of v for which r(w) < r(v). F models the output of
the online algorithm. We would like to upper bound the number of queries to vertices
in the graph needed in order to compute F (v0) for any vertex v0 ∈ G, namely, the time
to simulate the output of query v0 using the online algorithm.
To upper bound the number of queries to the graph, we turn to a simpler task of
bounding the size of a certain d-regular tree, which is an upper bound on the number of
queries. Consider an infinite d-regular tree T rooted at v0. Each node w in T is assigned
independently and uniformly at random a real number r(w) ∈ [0, 1]. For every node w
other than v0 in T , let parent(w) denote the parent node of w. We grow a (possibly
infinite) subtree T of T rooted at v as follows: a node w is in the subtree T if and only
if parent(w) is in T and r(w) < r(parent(w)) (for simplicity we assume all the ranks
are distinct real numbers). That is, we start from the root v0, add all the children of v0
whose ranks are smaller than that of v0 to T . We keep growing T in this manner where
a node w′ ∈ T is a leaf node in T if the ranks of its d children are all larger than r(w′).
We call the random tree T constructed in this way a query tree and we denote by T
the random variable that corresponds to the size of T . Note that T is an upper bound
on the number of queries since each node in T has at least as many neighbors as that in
G and if a node is connected to some previously queried nodes, this can only decrease
the number of queries. Therefore the number of queries is bounded by the size of T .
Our goal is to find an upper bound on T which holds with high probability.
We improve the upper bound on the query tree of O(logd+1 N ) given in [2] for
the case when the degrees are bounded by a constant d and extend our new bound to
the case that the degrees of G are binomially distributed, independently and identically
with expectation d, i.e., deg(v) ∼ B(n, d/n).
Our main result in this section is bounding, with high probability, the size of the
query tree T as follows.
Lemma 1. Let G be a graph whose vertex degrees are bounded by d or distributed
independently and identically from the binomial distribution: deg(v) ∼ B(n, d/n).
Then there exists a constant C(d) which depends only on d, such that
Pr[T > C(d) log n] < 1/n2,
where the probability is taken over all the possible permutations π ∈ Π of the vertices
of G, and T is a random query tree in G under π.
3.2 Overview of the proof
Our proof of Lemma 1 consists of two parts. Following [2], we partition the query tree
into levels. The first part of the proof is an upper bound on the size of a single (sub)tree
on any level. For the bounded degree case, this was already proved in [2] (the result is
restated as Proposition 1).
We extend the proof of [2] to the binomially distributed degrees case. In both cases
the bound is that with high probability each subtree is of size at most logarithmic in the
size of the input.
The second part, which is a new ingredient of our proof, inductively upper bounds
the number of vertices on each level, as the levels increase. For this to hold, it crucially
7
depends on the fact that all subtrees are generated independently and that the probability
of any subtree being large is exponentially small. The main idea is to show that although
each subtree, in isolation, can reach a logarithmic size, their combination is not likely
to be much larger. We use the distribution of the sizes of the subtrees, in order to bound
the aggregate of multiple subtrees.
3.3 Bounding the subtree size
i
1
L+1 ], for i = 1, 2, · · · , L and IL+1 = [0,
As in [2], we partition the query tree into levels and then upper bound the probability
that a subtree is larger than a given threshold. Let L > 1 be a function of d to be
determined later. First, we partition the interval [0,1] into L sub-intervals: Ii = (1 −
L+1 , 1 − i−1
L+1 ]. We refer to interval Ii as
level i. A vertex v ∈ T is said to be on level i if r(v) ∈ Ii. We consider the worst case,
in which r(v0) ∈ I1. In this case, the vertices on level 1 form a tree T1 rooted at v0.
Denote the number of (sub)trees on level i by ti. The vertices on level 2 will form a
set of trees {T (1)
}, where the total number of subtrees is at most the sum
of the children of all the vertices in T1 (we only have inequality because some of the
children of the vertices of T1 may be assigned to levels 3 and above.) The vertices on
}. Note that all these subtrees {T (j)
level i > 1 form a set of subtrees {T (1)
i }
are generated independently by the same stochastic process, as the ranks of all nodes in
T are i.i.d. random variables. In the following analysis, we will set L = 3d.
2 , · · · , T (t2)
2
, · · · T (ti)
i
i
For the bounded degree case, bounding the size of the subtree follows from [2]:
Proposition 1 ([2]). 7 Let L ≥ d+1 be a fixed integer and let T be the d-regular infinite
query tree. Then for any 1 ≤ i ≤ L and 1 ≤ j ≤ ti, Pr[T (j)
i=n 2−ci ≤
2−Ω(n), for all n ≥ β, where β is some constant. In particular, there is an absolute
constant c0 depending on d only such that for all n ≥ 1,
≥ n] ≤ P∞
i
Pr[T (j)
i
≥ n] ≤ e−c0n.
3.3.1 The binomially distributed degrees case We are interested in bounding the
subtree size also in the case that the degrees are not a constant d, but rather selected
independently and identically from a binomial distribution with mean d.
Proposition 2. Let T be a tree with vertex degree distributed i.i.d. binomially with
deg(v) ∼ B(n, d/n). For any 1 ≤ i ≤ L and any 1 ≤ j ≤ ti, Pr[T (j)
≥ n] ≤
P∞
i=n 2−ci ≤ 2−Ω(n), for n ≥ β, for some constant β > 0.
Proof. The proof of Proposition 2 is similar to the proof of Proposition 1 in [2]; we em-
ploy the theory of Galton-Watson processes. For a good introduction to Galton-Watson
branching processes see e.g. [10].
i
0, 1, 2, . . .}, with pk ≥ 0 and Pk pk = 1. Let f (s) = P∞
Consider a Galton-Watson process defined by the probability function p := {pk; k =
k=0 pksk be the gen-
erating function of p. For i = 0, 1, . . . , let Zi be the number of offsprings in the
7 In [2], this lemma is proved for the case of L = d+1. This immediately establishes Proposition
1, since the worse case is L = d + 1.
8
ith generation. Clearly Z0 = 1 and {Zi : i = 0, 1, . . .} form a Markov chain. Let
m := E[Z1] = Pk kpk be the expected number of children of any individual. Let
Z = Z0 + Z1 + · · · be the sum of all offsprings in all generations of the Galton-Watson
process. The following result of Otter is useful in bounding the probability that Z is
large.
Theorem 1 ([14]). Suppose p0 > 0 and that there is a point a > 0 within the circle of
convergence of f for which af ′(a) = f (a). Let α = a/f (a). Let t = gcd{r : pr > 0},
where gcd stands for greatest common divisor. Then
Pr[Z = n] =
a
2παf ′′(a)(cid:17)1/2
t(cid:16)
0,
α−nn−3/2 + O(α−nn−5/2),
if n ≡ 1 (mod t);
if n 6≡ 1 (mod t).
In particular, if the process is non-arithmetic, i.e. gcd{r : pr > 0} = 1, and
finite, then
a
αf ′′(a) is
Pr[Z = n] = O(α−nn−3/2),
and consequently Pr[Z ≥ n] = O(α−n).
i
We prove Proposition 2 for the case of tree T1 – the proof actually applies to all
subtrees T (j)
. Recall that T1 is constructed inductively as follows: for v ∈ N (v0) in
T , we add v to T1 if r(v) < r(v0) and r(v) ∈ I1. Then for each v in T1, we add the
neighbors w ∈ N (v) in T to T1 if r(w) < r(v) and r(w) ∈ I1. We repeat this process
until there is no vertex that can be added to T1.
Once again, we work with the worst case that r(v0) = 1. To upper bound the size
of T1, we consider a related random process which also grows a subtree of T rooted at
v0, and denote it by T ′
1 is the same as that of T1 except for
1 and w is a child vertex of v in T , then we add w to
the following difference: if v ∈ T ′
1 as long as r(w) ∈ I1. In other words, we give up the requirement that r(w) < r(v).
T ′
Clearly, we always have T1 ⊆ T ′
1. The process that grows T ′
1 and hence T ′
1 ≥ T1.
Note that the random process that generates T ′
1 is in fact a Galton-Watson process,
as the rank of each vertex in T is independently and uniformly distributed in [0, 1]. We
take vertex v to be the parent node. Since I1 = 1/L, then for any vertex u ∈ V (G),
u 6= v, the probability that u is a child node of v in T ′
1 is
d/n · 1/L = d/nL,
as the random process that connects w to v and the random process that generates the
rank of w are independent (each edge is chosen with probability d/n, and the probability
that r(w) is in Iv = 1/L). It follows that we have a binomial distribution for the number
of child nodes of v in T ′
1:
p = {(1 − q)n,(cid:18)n
1(cid:19)q(1 − q)n−1,(cid:18)n
2(cid:19)q2(1 − q)n−2, . . . , qn},
9
1 when its
where q := d/nL is the probability that a child vertex in T appears in T ′
1 is
parent vertex is in T ′
nq = d/L < 1, so from the classical result on the extinction probability of Galton-
Watson processes (see e.g. [10]), the tree T ′
1. Note that the expected number of children of a vertex in T ′
1 is finite with probability one.
The generating function of p is
f (s) = (1 − q + qs)n,
as the probability function {pk} obeys the binomial distribution pk = Pr[X = k] where
X ∼ B(n, q). In addition, the convergence radius of f is ρ = ∞ since {pk} has only a
finite number of non-zero terms.
f ′(s) = nq(1 − q + qs)n−1
Solving the equation af ′(a) = f (a) yields anq(1 − q + qa)n−1 = (1 − q + qa)n
and hence anq = 1 − q + qa. Consequently, solving for a gives
a =
1 − q
q(n − 1)
=
=
=
1 − d/nL
d(n − 1)/nL
nL − d
d(n − 1)
3n − 1
n − 1
.
We can lower bound α as
α =
a
f (a)
=
1
f ′(a)
L
=
=
≥
d(1 − q + q 3n−1
n−1 )n−1
3
3(n−1)(cid:17)n−1
(cid:16)1 + 2
3
e2/3 > 1.
Finally we calculate f ′′(a):
f ′′(a) = q2n(n − 1)(1 − q + qa)n−2
=
=
d2n(n − 1)
n2L2
n − 1
9n (cid:18)1 +
= Θ(1),
+
3n − 1
n − 1 (cid:19)n−2
·
d
nL
d
nL
(cid:18)1 −
3(n − 1)(cid:19)n−2
2
10
therefore
a
αf ′′(a) is a bounded constant.
Now applying Theorem 1 to the Galton-Watson process which generates T ′
1 (note
that t = 1 in our case) gives that, there exists a constant n0 such that for n > n0,
Pr[T ′
1 ≥ n] ≤
P∞
i=n 2−ci ≤ 2−Ω(n) for all n > n0. Hence for all large enough n, with probabil-
⊓⊔
1 = n] ≤ 2−cn for some constant c > 0. It follows that Pr[T ′
ity at least 1 − 1/n3, T1 ≤ T ′
1 = O(log n).
The following corollary stems directly from Propositions 1 and 2:
Corollary 1. Let T be any infinite d-regular query tree or tree with vertex degree dis-
tributed i.i.d. binomially with deg(v) ∼ B(n, d/n). For any 1 ≤ i ≤ L and any
1 ≤ j ≤ ti, with probability at least 1 − 1/n3, T (j)
= O(log n).
i
3.4 Bounding the increase in subtree size as we go up levels
From Corollary 1 we know that the size of any subtree, in particular T1, is bounded
by O(log n) with probability at least 1 − 1/n3 in both the degree d and the binomial
degree cases. Our next step in proving Lemma 1 is to show that, as we increase the
levels, the size of the tree does not increase by more than a constant factor for each
level. That is, there exists an absolute constant η depending on d only such that if the
number of vertices on level k is at most Tk, then the number of vertices on level k + 1,
i=1 Ti + O(log n) ≤ 2ηTk + O(log n). Since there
are L levels in total, this implies that the number of vertices on all L levels is at most
O((2η)L log n) = O(log n).
Tk+1 satisfies Tk+1 ≤ ηPk
The following Proposition establishes our inductive step.
Proposition 3. For any infinite query tree T with constant bounded degree d (or de-
grees i.i.d. ∼ B(n, d/n)), for any 1 ≤ i < L, there exist constants η1 > 0 and
i+1 ≥ η1η2 log n] < 1/n2
j=1 T (j)
j=1 T (j)
i
η2 > 0 s.t. if Pti
for all n > β, for some β > 0.
≤ η1 log n then Pr[Pti+1
i=1 zi.
i=1 Zi. Assume
that each vertex i on level ≤ k is the root of a tree of size zi on level k + 1. Notice that
Proof. Denote the number of vertices on level k by Zk and let Yk = Pk
Zk+1 = PYk
By Proposition 1 (or Proposition 2), there are absolute constants c0 and β depending
on d only such that for any subtree T (i)
k = n] ≤
e−c0n. Therefore, given (z1, . . . , zYk), the probability of the forest on level k+1 consist-
i=1 e−c0(zi−β) = e−c0(Zk+1−βYk).
Notice that, given Yk (the number of nodes up to level k), there are at most(cid:0)Zk+1+Yk−1
(cid:1)
ing of exactly trees of size (z1, . . . , zYk ) is at mostQYk
< (cid:0)Zk+1+Yk
(cid:1) vectors (z1, . . . , zYk ) that can realize Zk+1.
k on level k and any n > β, Pr[T (i)
Yk−1
Yk
We want to bound the probability that Zk+1 = ηYk for some (large enough) con-
stant η > 0. We can bound this as follows:
Pr[Tk+1 = Zk+1] < (cid:18)Zk+1 + Yk
Yk
(cid:19)e−c0(Zk+1−βYk)
11
Yk
e−c0(Zk+1−βYk)
(cid:19)Yk
< (cid:18) e · (Zk+1 + Yk)
= (e(1 + η))Yk e−c0(η−β)Yk
= eYk(−c0(η−β)+ln(η+1)+1)
≤ e−c0ηYk/2,
It follows that there is some absolute constant c′ which depends on d only such
that Pr[Tk+1 ≥ ηYk] ≤ e−c′ηYk . That is, if ηYk = Ω(log n), the probability that
Tk+1 ≥ ηYk is at most 1/n3. Adding the vertices on all L levels and applying the
union bound, we conclude that with probability at most 1/n2, the size of T is at most
O(log n).
⊓⊔
4 Hypergraph 2-coloring and k-CNF
We use the bound on the size of the query tree of graphs of bounded degree to improve
the analysis of [2] for hypergraph 2-coloring. We also modify their algorithm slightly
to further improve the algorithm's complexity. As the algorithm is a more elaborate
version of the algorithm of [2] and the proof is somewhat long, we only state our main
theorem for hypergraph 2-coloring; we defer the proof to Appendix A.
Theorem 2. Let H be a k-uniform hypergraph s.t. each hyperedge intersects at most d
other hyperedges. Suppose that k ≥ 16 log d + 19.
Then there exists an (O(log4 n), O(log4 n), 1/n)-local computation algorithm which,
given H and any sequence of queries to the colors of vertices (x1, x2, . . . , xs), with
probability at least 1 − 1/n2, returns a consistent coloring for all xi's which agrees
with a 2-coloring of H. Moreover, the algorithm is query oblivious and parallelizable.
Due to the similarity between hypergraph 2-coloring and k-CNF, we also have the
following theorem; the proof is in Appendix B.
Theorem 3. Let H be a k-CNF formula with k ≥ 2. Suppose that each clause inter-
sects no more than d other clauses, and furthermore suppose that k ≥ 16 log d + 19.
Then there exists a (O(log4 n), O(log4 n), 1/n)-local computation algorithm which,
given a formula H and any sequence of queries to the truth assignments of variables
(x1, x2, . . . , xs), with probability at least 1 − 1/n2, returns a consistent truth assign-
ment for all xi's which agrees with some satisfying assignment of the k-CNF formula
H. Moreover, the algorithm is query oblivious and parallelizable.
12
5 Maximal matching
We consider the problem of maximal matching in a bounded-degree graph. We are
given a graph G = (V, E), where the maximal degree is bounded by some constant d,
and we need to find a maximal matching.A matching is a set of edges with the property
that no two edges share a common vertex. The matching is maximal if no other edge
can be added to it without violating the matching property.
Assume the online scenario in which the edges arrive in some unknown order. The
following greedy online algorithm can be used to calculate a maximal matching: When
an edge e arrives, we check whether e is already in the matching. If it is not, we check
if any of the neighboring edges are in the matching. If none of them is, we add e to the
matching. Otherwise, e is not in the matching.
We turn to the local computation variation of this problem. We would like to query,
for some edge e ∈ E, whether e is part of some maximal matching. (Recall that all
replies must be consistent with some maximal matching).
We use the technique of [2] to produce an almost O(log n)-wise independent ran-
dom ordering on the edges, using a seed length of O(log3 n).8 When an edge e is
queried, we use a BFS (on the edges) to build a DAG rooted at e. We then use the
greedy online algorithm on the edges of the DAG (examining the edges with respect to
the ordering), and see whether e can be added to the matching.
As the query tree is an upper-bound on the size of the DAG, we derive the following
theorem from Lemma 1.
Theorem 4. Let G = (V, E) be an undirected graph with n vertices and maximum
degree d. Then there is an (O(log3 n), O(log3 n), 1/n) - local computation algorithm
which, on input an edge e, decides if e is in a maximal matching. Moreover, the algo-
rithm gives a consistent maximal matching for every edge in G.
6 The bipartite case and local load balancing
We consider a general "power of d choices" online algorithm for load balancing. In this
setting there are n balls that arrive in an online manner, and m bins. Each ball selects a
random subset of d bins, and queries these bins. (Usually the query is simply the current
load of the bin.) Given this information, the ball is assigned to one of the d bins (usually
to the least loaded bin). We denote by LB such a generic algorithm (with a decision rule
which can depend in an arbitrary way on the d bins that the ball is assigned to). Our
main goal is to simulate such a generic algorithm.
The load balancing problem can be represented by a bipartite graph G = ({V, U }, E),
where the balls are represented by the vertices V and the bins by the vertices U . The
random selection of a bin u ∈ U by a ball v ∈ V is represented by an edge. By defi-
nition, each ball v ∈ V has degree d. Since there are random choices in the algorithm
LB we need to specify what we mean by a simulation. For this reason we define the
input to be the following: a graph G = ({V, U }, E), where V = n, U = m, and
8 Since the query tree is of size O(log n) w.h.p., we don't need a complete ordering on the
vertices; an almost O(log n)-wise independent ordering suffices.
13
n = cm for some constant c ≥ 1. We also allocate a rank r(u) ∈ [0, 1] to every u ∈ U .
This rank represents the ball's arrival time: if r(v) < r(u) then vertex v arrived before
vertex u. Furthermore, all vertices can have an input value x(w). (This value represents
some information about the node, e.g., the weight of a ball.) Given this input, the al-
gorithm LB is deterministic, since the arrival sequence is determined by the ranks, and
the random choices of the balls appear as edges in the graph. Therefore by a simulation
we will mean that given the above input, we generate the same allocation as LB.
We consider the following stochastic process: Every vertex v ∈ V uniformly and
independently at random chooses d vertices in U . Notice that from the point of view
of the bins, the number of balls which chose them is distributed binomially with X ∼
B(n, d/m). Let Xv and Xu be the random variables for the number of neighbors of
vertices v ∈ V and u ∈ U respectively. By definition, Xv = d, since all balls have d
neighbors, and hence each Xu is independent of all Xv's. However, there is a depen-
dence between the Xu's (the number of balls connected to different bins). Fortunately
this is a classical example where the random variables are negatively dependent (see
e.g. [9]). 9
6.1 The bipartite case
Recall that in Section 3, we assumed that the degrees of the vertices in the graph were
independent. We would like to prove an O(log n) upper bound on the query tree T
for our bipartite graph. As we cannot use the theorems of Section 3 directly, we show
that the query tree is smaller than another query tree which meets the conditions of our
theorems.
The query tree for the binomial graph is constructed as follows: a root v0 ∈ V is
selected for the tree. (v0 is the ball whose bin assignment we are interested in deter-
mining.) Label the vertices at depth j in the tree by Wj. Clearly, W0 = {v0}. At each
depth d, we add vertices one at a time to the tree, from left to right, until the depth is
"full" and then we move to the next depth. Note that at odd depths (2j + 1) we add bin
vertices and at even depths (2j) we add ball vertices.
Specifically, at odd depths (2j + 1) we add, for each v ∈ W2j its d neighbors
u ∈ N (v) as children, and mark each by u.10 At even depths (2j) we add for each
node marked by u ∈ W2j−1 all its (ball) neighbors v ∈ N (u) such that r(v) <
r(parent(u)), if they have not already been added to the tree. Namely, all the balls
that are assigned to u by time
A leaf is a node marked by a bin uℓ for whom all neighboring balls v ∈ N (uℓ) −
{parent(uℓ)} have a rank larger than its parent, i.e., r(v) > r(parent(uℓ)). Namely,
parent(uℓ) is the first ball to be assigned to bin uℓ. This construction defines a stochas-
tic process F = {Ft}, where Ft is (a random variable for) the size of T at time t. (We
start at t = 0 and t increases by 1 for every vertex we add to the tree).
We now present our main lemma for bipartite graphs.
9 We remind the reader that two random variables X1 and X2 are negatively dependent if
Pr[X1 > xX2 = a] < Pr[X1 > xX2 = b], for a > b and vice-versa.
10 A bin can appear several times in the tree. It appears as different nodes, but they are all marked
so that we know it is the same bin. Recall that we assume that all nodes are unique, as this
assumption can only increase the size of the tree.
14
Lemma 2. Let G = ({V, U }, E) be a bipartite graph, V = n and U = m and
n = cm for some constant c ≥ 1, such that for each vertex v ∈ V there are d edges
chosen independently and at random between v and U . Then there is a constant C(d)
which depends only on d such that
Pr[T < C(d) log n] > 1 − 1/n2,
where the probability is taken over all of the possible permutations π ∈ Π of the
vertices of G, and T is a random query tree in G under π.
To prove Lemma 2, we look at another stochastic process F ′, which constructs
a tree T ′: we start with a root v′
0. Label the vertices at depth j in the tree by W ′
j.
Assign every vertex y that is added to the tree a rank r(y) ∈ [0, 1] independently and
uniformly at random. Similarly to T , W ′
0}. At odd depths (2j + 1) we add to
2j, d children (from left to right). At even depths (2j) we add to each node
each v′ ∈ W ′
u′ of different nodes
u′ ∈ W ′
are i.i.d. Of the nodes added in this level, we remove all those vertices y′ for which
r(y′) > r(parent(parent(y′))).
u′ ∼ B(n, 2d/m) and the X ′
u′ children, where X ′
0 = {v′
2j−1, X ′
Importantly, the neighbor distributions of the vertices in the tree are independent of
each other. If at any point T ′ has "more than half the bins", i.e., the sum of nodes on
odd levels is at least m/2, we add n + m bin children of rank 0 to some even-level node
in the tree.
Given a tree T we define squash(T ) to be the tree T with the odd levels deleted, and
a node v in level 2j is connected to node v′ in level 2j + 2 if v = parent(parent(v′)).
Lemma 3. There is a constant C(d) which depends only on d such that for all large
enough n, Pr[squash(T ′) > C(d) log n] < 1/n2.
Because d·squash(T ′) ≥ T ′ ≥ squash(T ′), we immediately get the following
corollary:
Corollary 2. There is a constant C(d) which depends only on d such that for all large
enough n, Pr[T ′ > C(d) log n] < 1/n2.
We first make the following claim:
Claim. squash(T ′) has vertex degree distributed i.i.d. binomially with deg(v) ∼ B(dn,
2d/m).
2j has d children, each with degree distributed binomially ∼ B(n, 2d/m).
Proof. Each v ∈ W ′
For any independent r.v.'s Y1, Y2, · · · where ∀i > 0, Yi ∼ B(n, p), we know that
q
Xi=1
Yi ∼ B(qn, p). The Claim follows.
⊓⊔
We can now turn to the proof of Lemma 3:
Proof. As long as squash(T ′) < m/2, the proof of the lemma follows the proof of
Lemma 1 with slight modifications to constants and will therefore be omitted.
15
We notice that squash(T ′) < m/2 w.p. at least 1 − 1/n2: the proof of Lemma
1 is inductive - we show that at level I1, the size of the subtree is at most O(log n),
and then bound the increase in tree size as we move to the next level. By the level IL,
squash(T ′) = O(log n) w.p. at least 1−1/n2. Therefore it follows that squash(T ′) <
m/2 w.p. at least 1 − 1/n2.
⊓⊔
Before we can complete the proof of Lemma 2, we need to define the notion of first
order stochastic dominance:
Definition 1 (First order stochastic dominance). We say a random variable X first
order stochastically dominates (dominates for short) a random variable Y if Pr[X ≥
a] ≥ Pr[Y ≥ a] for all a and Pr[X ≥ a] > Pr[Y ≥ a] for some a. If X dominates Y ,
then we write X ≥ Y .
Lemma 4. For every t, F ′
t first-order stochastically dominates Ft.
Proof. Assume we add a (bin) vertex u ∈ U to T at time t, the random variable for the
number of u's neighbors is negatively dependent on all other Xw, w ∈ Tt ∩ U . We label
this variable ¯X = Xu{Xw}, w ∈ Tt.
We first show that F ′
t ≥ Ft when T has less than m/2 bins, and then show that
t ≥ Ft when T ′ has more than m/2 bins. (It is easy to see why this is enough).
F ′
Assume Tt ∩U ≤ m/2. Xu is dependent on at most m/2 other random variables, Xw.
Because the dependency is negative, Xu is maximized when ∀w, Xw = 0. Therefore,
in the worst case, Xu is dependent on m/2 bins with 0 children. If m/2 bins have
0 children, all edges in G must be distributed between the remaining bins. Therefore
¯X ≤ X ′
When T ′ has more than m/2 bins, by the construction of F ′
vertices, and so F ′
t , it has more than m + n
⊓⊔
t trivially dominates Ft.
u, where X ′
u ∼ B(n, 2d/m).
Combining Corollary 2 and Lemma 4 completes the proof of Lemma 2.
6.2 Local load balancing
The following theorem states our basic simulation result.
Theorem 5. Consider a generic online algorithm LB which requires constant time per
query, for n balls and m bins, where n = cm for some constant c > 0. There exists
an (O(log n), O(log n), 1/n)-local computation algorithm which, on query of a (ball)
vertex v ∈ V , allocates v a (bin) vertex u ∈ U , such that the resulting allocation is
identical to that of LB with probability at least 1 − 1/n.
Proof. Let K = C(d) log U for some constant C(d) depending only on d. K is the
upper bound given in Lemma 2. (In the following we make no attempt to provide the
exact values for C(d) or K.)
We now describe our (O(log n), O(log n), 1/n)-local computation algorithm for
LB. A query to the algorithm is a (ball) vertex v0 ∈ V and the algorithm will chose a
(bin) vertex from the d (bin) vertices connected to v0.
16
We first build a query tree as follows: Let v0 be the root of the tree. For every
u ∈ N (u0), add to the tree the neighbors of u, v ∈ V such that r(v) < r(v0). Continue
inductively until either K nodes have been added to the random query tree or no more
nodes can be added to it. If K nodes have been added to the query tree, this is a failure
event, and assign to v0 a random bin in N (v0). From Lemma 2, this happens with
probability at most 1/n2, and so the probability that some failure event will occur is at
most 1/n. Otherwise, perform LB on all of the vertices in the tree, in order of addition
to the tree, and output the bin to which ball v0 is assigned to by LB.
⊓⊔
A reduction from various load balancing algorithms gives us the following corollar-
ies to Theorem 5.
Corollary 3. (Using [6]) Suppose we wish to allocate m balls into n bins of uniform
capacity, m ≥ n, where each ball chooses d bins independently and uniformly at ran-
dom. There exists a (log n, log n, 1/n) LCA which allocates the balls in such a way that
the load of the most loaded bin is m/n + O(log log n/ log d) w.h.p.
Corollary 4. (Using [19]) Suppose we wish to allocate n balls into n bins of uniform
capacity, where each ball chooses d bins independently at random, one from each of d
groups of almost equal size θ( n
d ). There exists a (log n, log n, 1/n) LCA, which allo-
cates the balls in such a way that the load of the most loaded bin is ln ln n/(d−1) ln 2+
O(1) w.h.p. 11
Corollary 5. (Using [5]) Suppose we wish to allocate m balls into n ≤ m bins,
where each bin i has a capacity ci, and Pi ci = m. Each ball chooses d bins at ran-
dom with probability proportional to their capacities. There exists a (log n, log n, 1/n)
LCA which allocates the balls in such a way that the load of the most loaded bin is
2 log log n + O(1) w.h.p.
Corollary 6. (Using [5]) Suppose we wish to allocate m balls into n ≤ m bins, where
each bin i has a capacity ci, and Pi ci = m. Assume that the size of a large bin is at
least rn log n, for large enough r. Suppose we have s small bins with total capacity ms,
and that ms = O((n log n)2/3). There exists a (log n, log n, 1/n) LCA which allocates
the balls in such a way that the expected maximum load is less than 5.
Corollary 7. (Using [8]) Suppose we have n bins, each represented by one point on a
circle, and n balls are to be allocated to the bins. Assume each ball needs to choose
d ≥ 2 points on the circle, and is associated with the bins closest to these points. There
exists a (log n, log n, 1/n) LCA which allocates the balls in such a way that the load of
the most loaded bin is ln ln n/ ln d + O(1) w.h.p.
11 In fact, in this setting the tighter bound is ln ln n
d ln φd
Fibonacci sequence, i.e. φd = limk→∞
and for k ≥ 1 Fd(k) = Pd
i=1 Fd(k − i)
+ O(1), where φd is the ratio of the d-step
kpFd(k), where for k < 0, Fd(k) = 0, Fd(1) = 1,
17
6.3 Random ordering
In the above we assume that we are given a random ranking for each ball. If we are not
given such random rankings (in fact, a random permutation of the vertices in U will also
suffice), we can generate a random ordering of the balls. Specifically, since w.h.p. the
size of the random query is O(log n), an O(log n)-wise independent random ordering12
suffices for our local computation purpose. Using the construction in [2] of 1/n2-almost
O(log n)-wise independent random ordering over the vertices in U which uses space
O(log3 n), we obtain (O(log3 n), O(log3 n), 1/n)-local computation algorithms for
balls and bins.
12 See [2] for the formal definitions of k-wise independent random ordering and almost k-wise
independent random ordering.
18
References
[1] N. Alon. A parallel algorithmic version of the Local Lemma. Random Structures and
Algorithms, 2:367–378, 1991.
[2] N. Alon, R. Rubinfeld, S. Vardi, and N. Xie. Space-efficient local computation algorithms.
In Proc. 23rd ACM-SIAM Symposium on Discrete Algorithms, pages 1132–1139, 2012.
[3] Y. Azar, A. Z. Broder, A. R. Karlin, and E. Upfal. Balanced allocations. SIAM Journal on
Computing, 29(1):180–200, 1999.
[4] J. Beck. An algorithmic approach to the Lovász Local Lemma. Random Structures and
Algorithms, 2:343–365, 1991.
[5] P. Berenbrink, A. Brinkmann, T. Friedetzky, and L. Nagel. Balls into non-uniform bins.
In Proceedings of the 24th IEEE International Parallel and Distributed Processing Sympo-
sium (IPDPS), pages 1–10. IEEE, 2010.
[6] P. Berenbrink, A. Czumaj, A. Steger, and B. Vöcking. Balanced allocations: The heavily
loaded case. SIAM J. Comput., 35(6):1350–1385, 2006.
[7] A. Borodin and Ran El-Yaniv. Online Computation and Competitive Analysis. Cambridge
University Press, 1998.
[8] John W. Byers, Jeffrey Considine, and Michael Mitzenmacher. Simple load balancing for
distributed hash tables. In Proc. of Intl. Workshop on Peer-to-Peer Systems(IPTPS), pages
80–87, 2003.
[9] D. Dubhashi and D. Ranjan. Balls and bins: A study in negative dependence. Random
Structures and Algorithms, 13:99–124, 1996.
[10] T. Harris. The Theory of Branching Processes. Springer, 1963.
[11] S. Marko and D. Ron. Distance approximation in bounded-degree and general sparse
graphs. In APPROX-RANDOM'06, pages 475–486, 2006.
[12] M. Mitzenmacher, A. Richa, and R. Sitaraman. The power of two random choices: A
survey of techniques and results. In Handbook of Randomized Computing, Vol. I, edited by
P. Pardalos, S. Rajasekaran, J. Reif, and J. Rolim, pages 255–312. Norwell, MA: Kluwer
Academic Publishers, 2001.
[13] H. N. Nguyen and K. Onak. Constant-time approximation algorithms via local improve-
ments. In Proc. 49th Annual IEEE Symposium on Foundations of Computer Science, pages
327–336, 2008.
[14] R. Otter. The multiplicative process. Annals of mathematical statistics, 20(2):206–224,
1949.
[15] M. Parnas and D. Ron. Approximating the minimum vertex cover in sublinear time and a
connection to distributed algorithms. Theoretical Computer Science, 381(1–3), 2007.
[16] R. Rubinfeld, G. Tamir, S. Vardi, and N. Xie. Fast local computation algorithms. In Proc.
2nd Symposium on Innovations in Computer Science, pages 223–238, 2011.
[17] R. Rubinfeld, G. Tamir, S. Vardi, and N. Xie. Fast local computation algorithms, 2011.
[18] K. Talwar and U. Wieder. Balanced allocations: the weighted case. In Proc. 39th Annual
ACM Symposium on the Theory of Computing, pages 256–265, 2007.
[19] Berthold Vöcking. How asymmetry helps load balancing. J. ACM, 50:568–589, July 2003.
[20] Y. Yoshida, Y. Yamamoto, and H. Ito. An improved constant-time approximation algo-
rithm for maximum matchings. In Proc. 41st Annual ACM Symposium on the Theory of
Computing, pages 225–234, 2009.
19
A Hypergraph two-coloring
Recall that a hypergraph H is a pair H = (V, E) where V is a finite set whose ele-
ments are called nodes or vertices, and E is a family of non-empty subsets of V , called
hyperedges. A hypergraph is called k-uniform if each of its hyperedges contains pre-
cisely k vertices. A two-coloring of a hypergraph H is a mapping c : V → {red, blue}
such that no hyperedge in E is monochromatic. If such a coloring exists, then we say
H is two-colorable. We assume that each hyperedge in H intersects at most d other
hyperedges. Let n be the number of hyperedges in H. Here we think of k and d as fixed
constants and all asymptotic forms are with respect to n. By the Lovász Local Lemma,
when e(d + 1) ≤ 2k−1, the hypergraph H is two-colorable (e.g. [1]).
Following [17], we let m be the total number of vertices in H. Note that m ≤ kn,
so m = O(n). For any vertex x ∈ V , we use E(x) to denote the set of hyperedges x
belongs to. For any hypergraph H = (V, E), we define a vertex-hyperedge incidence
matrix M ∈ {0, 1}m×n so that, for every vertex x and every hyperedge e, Mx,e = 1
if and only if e ∈ E(x). Because we assume both k and d are constants, the incidence
matrix M is necessarily very sparse. Therefore, we further assume that the matrix M
is implemented via linked lists for each row (that is, vertex x) and each column (that is,
hyperedge e).
Let G be the dependency graph of the hyperedges in H. That is, the vertices of
the undirected graph G are the n hyperedges of H and a hyperedge Ei is connected to
another hyperedge Ej in G if Ei ∩ Ej 6= ∅. It is easy to see that if the input hypergraph
is given in the above described representation, then we can find all the neighbors of any
hyperedge Ei in the dependency graph G (there are at most d of them) in constant time
(which depends on k and d).
A natural question to ask is: Given a two-colorable hypergraph H, and a vertex v,
can we quickly compute the coloring of v? Alon et al. gave ([2]) a polylog(n)-time
and space LCA based on Alon's 3-phase parallel hypergraph coloring algorithm ([1]),
where the exponent of the logarithm depends on d. We get rid of the dependence on
d (in the exponent of the logarithm) using the improved analysis of the query tree in
section 3, together with a modified 4-phase coloring algorithm.
Our main result in this section is, given a two-colorable hypergraph H whose two-
coloring scheme is guaranteed by the Lovász Local Lemma (with slightly weaker pa-
rameters), we give a (O(log4 n), O(log4 n), 1/n) - local computation algorithm. We
restate our main theorem:
Theorem 6. Let H be a k-uniform hypergraph s.t. each hyperedge intersects at most d
other hyperedges. Suppose that k ≥ 16 log d + 19.
Then there exists an (O(log4 n), O(log4 n), 1/n)-local computation algorithm which,
given H and any sequence of queries to the colors of vertices (x1, x2, . . . , xs), with
probability at least 1 − 1/n2, returns a consistent coloring for all xi's which agrees
with a 2-coloring of H. Moreover, the algorithm is query oblivious and parallelizable.
In fact, we only need:
k ≥ 3⌈log 16d(d − 1)3(d + 1)⌉ + ⌈log 2e(d + 1)⌉
20
Throughout the following analysis, we set: k1 = k, and
ki = ki−1 − ⌈log 16d(d − 1)3(d + 1)⌉
Notice that the theorem's premise simply implies that 2k4−1 ≥ e(d + 1), as required by
the Lovász Local Lemma.
A.1 The general phase - random coloring
In each phase we begin with subsets Vi and Ei of V and E, such that each edge contains
at least ki vertices. We sequentially assign colors at random to the vertices, as long as
every monochromatic edge has at least ki+1 uncolored vertices. Once the phase is over
we do not change this assignment.
If an edge has all of its vertices besides ki+1 colored in one color, it is labeled
dangerous. All the uncolored vertices in a dangerous edge are labeled saved and we do
not color them in this phase. We proceed until all vertices in Vi are either red, blue, or
saved. Let the survived hyperedges be all the edges that do not contain both red and
blue vertices. Each survived edge contains some vertices colored in one color, and at
least ki+1 saved vertices.
Let Si be the set of survived edges after a random coloring in Phase i, and con-
sider GSi, the restriction of G to Si The probability that GSi contains a connected
component of size d3u at most Vi2−u ([1]). In particular, after repeating the random
coloring procedure ti times, there is no connected component of size greater than d3u
with probability
(cid:0)Vi2−u(cid:1)ti
If the query vertex x has been assigned a color in the i-th phase, we can simply
return this color. Otherwise, if it is a saved vertex we let Ci(x) be the connected com-
ponent containing x in GSi. Finally, since the coloring of Ci(x) is independent of all
other uncolored vertices, we can restrict ourselves to Ei+1 = Ci(x) in the next phase.
A.2 Phase 1: partial random coloring
In the first phase we begin with the whole hypergraph, i.e. V1 = V , E1 = E, and
k1 = k. Thus, we cannot even assign a random coloring to all the vertices in sublinear
complexity. Instead, similarly to the previous sections, we randomly order the vertices
of the hypergraph and use a query tree to randomly assign colors to all the vertices that
arrive before x and may influence it. Note that this means that we can randomly assign
the colors only once.
If x is a saved vertex, we must compute E2 = C1(x), the connected component
containing x in GS1. Notice that the size C1(x) is bounded w.h.p.
Pr(cid:2)C1(x) > 4d3 log n(cid:3) < n2−4 log n = n−3
In order to compute C1(x), we run BFS on GS1. Whenever we reach a new node,
we must first randomly assign colors to the vertices in its query tree, like we did for x's
query tree. Since (w.h.p.) there are at most O(log n) edges in C1(x), we query for trees
21
LCA for Hypergraph Coloring
Preprocessing:
1. Generate O(log n) independent ensembles, consisting each of O(log n)-wise inde-
pendent random variables in {0, 1}m
2. Generate π, a 1
m2 -almost log2 n-wise independent random ordering over [m]
Phase 1:
Input: a vertex x ∈ V
Out¯put¯: a color in {red, blue}
1. Use BFS to find the query tree T rooted at x, based on the ordering π
2. Randomly color the vertices in T according to the order defined by π
3. If x is colored red or blue, return the color
4. Else:
(a) Starting from E (x)a run BFS in GS1
b in order to find the connected component
E2 = C1(x) of survived hyperedges around x
(b) Let V2 be the set of uncolored vertices in E2
Run Phase 2 Coloring(x, E2, V2)
a Recall that E (x) is the set of hyperedges containing (x).
b S1 denotes the set of survived hyperedges in E.
Fig. 1. Local computation algorithm for Hypergraph Coloring
vertices.
of vertices in at most O(log n) edges. Therefore in total we color at most O(cid:0)log2(n)(cid:1)
Finally, since we are only interested in O(cid:0)log2(n)(cid:1) vertices, we may consider a
coloring which is only O(cid:0)log2(n)(cid:1)-wise independent, and a random ordering which is
only n−3-almost O(cid:0)log2(n)(cid:1)-independent. Given the construction in [2], this can be
done in space and time complexity O(cid:0)log4(n)(cid:1)
A.3 Phase 2 and 3: gradually decreasing the component size
Phase 2 and 3 are simply iterations of the general phase with parameters as described
below. With high probability we have that E2 ≤ 4d3 log n, and each edge has k2
uncolored vertices. After at most t2 = log n repetitions of the random coloring proce-
dure, we reach an assignment that leaves a size 2d3 log log n-connected component of
survived edges with probability
(cid:0)(cid:0)4d3k2 log n(cid:1) 2−2 log log n(cid:1)log n
< n−3
Similarly, in the third phase we begin with E3 < 2d3 log log n, and after t3 = log n
repetitions we reach an assignment that leaves a size log log n
-connected component of
survived edges with probability
k4
(cid:18)(cid:0)2d3k3 log log n(cid:1) 2
− log log n
d3 k4 (cid:19)log n
< n−3
22
i ∈ 2, 3
Phase i Coloring(x,Ei,Vi)
Input: a vertex x ∈ Vi and subsets Ei ⊆ E and Vi ⊆ V
Out¯put¯: a ¯color in {red, blue} or FAIL
1. Repeat the following log n times and stop if a good coloring is founda
(a) Sequentially try to color every vertex in Vi uniformly at random
(b) Explore the dependency graph of GSi
(c) Check if the coloring is good
2. If x is colored in the good coloring, return that color
Else
(a) Compute the connected connected component Ci(x) = Ei+1 and then also Vi+1
(b) Run Phase i + 1 Coloring(x, Ei+1, Vi+1)
a Following [17], let Si be the set of survived hyperedges in Ei after all vertices in Vi are
either colored or are saved. Now we explore the dependency graph of Si to find out all
the connected components.
We say a Phase 2 coloring is good if all connected components in GS2 have sizes at
most 2d3 log log n.
Similarly, we say a Phase 3 coloring is good if all connected components in GS3 have
sizes at most log log n
.
k4
Fig. 2. Local computation algorithm for Hypergraph Coloring: Phase 2 and Phase 3
Phase 4 Coloring(x, E4, V4)
Input: a vertex x ∈ V4 and subsets E4 ⊆ E and V4 ⊆ V
Out¯put¯: a ¯color in {red, blue}
1. Go over all possible colorings of the connected component V4 and color it using a
feasible coloring.
2. Return the color c of x in this coloring.
Fig. 3. Local computation algorithm for Hypergraph Coloring: Phase 4
23
A.4 Phase 4: brute force
Finally, we are left with a connected component of E4 < log log n
, and each edge has
k4 uncolored vertices. By the Lovasz Local Lemma, there must exists a coloring (see
e.g. Theorem 5.2.1 in [1]). We can easily find this coloring via brute force search in
time O(log n).
k4
B k-CNF
As another application, our hypergraph coloring algorithm can be easily modified to
compute a satisfying assignment of a k-CNF formula, provided that the latter satisfies
some specific properties.
Let H be a k-CNF formula on m Boolean variables x1, . . . , xm. Suppose H has n
clauses H = A1 ∧ · · · ∧ An and each clause consists of exactly k distinct literals.13
We say two clauses Ai and Aj intersect with each other if they share some variable (or
the negation of that variable). As in the case for hypergraph coloring, k and d are fixed
constants and all asymptotics are with respect to the number of clauses n (and hence
m, since m ≤ kn). Our main result is the following.
Theorem 7. Let H be a k-CNF formula with k ≥ 2. Suppose that each clause inter-
sects no more than d other clauses, and furthermore suppose that k ≥ 16 log d + 19.
Then there exists a (O(log4 n), O(log4 n), 1/n)-local computation algorithm which,
given a formula H and any sequence of queries to the truth assignments of variables
(x1, x2, . . . , xs), with probability at least 1 − 1/n2, returns a consistent truth assign-
ment for all xi's which agrees with some satisfying assignment of the k-CNF formula
H. Moreover, the algorithm is query oblivious and parallelizable.
Proof [Sketch]: We follow a 4-phase algorithm similar to that of hypergraph two-
coloring as presented in appendix A. In every phase, we sequentially assign random
values to a subset of the remaining variables, maintaining a threshold of ki unassigned
variables in each unsatisfied clause. Since the same (in fact, slightly stronger) bounds
that hold for the connected components in the hyperedges dependency graph also hold
for the clauses dependency graph ([17]), we can return an answer which is consistent
with a satisfying assignment with probability at least 1 − 1/n2. ⊓⊔
C Lower bound on the size of the query tree
We prove a lower bound on the size of the query tree.
Theorem 8. Let G be a random graph whose vertex degree is bounded by d ≥ 2
or distributed independently and identically from the binomial distribution: deg(v) ∼
B(n, d/n) (d ≥ 2). Then
Pr[T > log n/ log log n] > 1/n,
13 Our algorithm works for the case that each clause has at least k literals; for simplicity, we
assume that all clauses have uniform size.
24
where the probability is taken over all random permutations π ∈ Π of the vertices, and
T is the largest query tree in G (under π).
Proof. For both the bounded degree and the binomial distribution cases, there exists a
path of length at least k = log n/ log log n in the graph w.h.p. Label the vertices on the
path v1, v2, . . . , vk. There are k! possible permutations of the weights of the vertices
on the path. The probability of choosing the permutation in which w(v1) < w(v2) <
. . . < (vk) is 1/k!.
k! = (log n/ log log n)!
< (log n/ log log n)log n/ log log n
< n.
Therefore, 1/k! > 1/n and so the probability of the query tree having size
log n/ log log n is at least 1/n.
⊓⊔
|
1811.10767 | 1 | 1811 | 2018-11-27T01:52:13 | The Batched Set Cover Problem | [
"cs.DS"
] | We introduce the batched set cover problem, which is a generalization of the online set cover problem. In this problem, the elements of the ground set that need to be covered arrive in batches. Our main technical contribution is a tight $\Omega(H_{m - 2^z + 1})$ lower bound on the competitive ratio of any fractional batched algorithm given an adversary that is required to produce batches of VC-dimension at least $z$, for some $z \in \mathbb{N}^0$. This restriction on the adversary is motivated by the fact that, in some real world applications, decisions are made after collecting batches of data of non-trivial VC-dimension. In particular, ridesharing systems rely on the batch assignment of trip requests to vehicles, and some related problems such as that of optimal congregation points for passenger pickups and dropoffs can be modeled as a batched set cover problem with VC-dimension greater than or equal to two. Furthermore, we note that while any online algorithm may be used to solve the batched set cover problem by artificially sequencing the elements in a batch, this procedure may neglect the rich information encoded in the complex interactions between the elements of a batch and the sets that contain them. Therefore, we propose a minor modification to an online algorithm found in [8] to obtain an algorithm that attempts to exploit such information. Unfortunately, we are unable to improve its analysis in a way that reflects this intuition. However, we present computational experiments that provide empirical evidence of a constant factor improvement in the competitive ratio. To the best of our knowledge, we are the first to use the VC-dimension in the context of online (batched) covering problems. | cs.DS | cs |
The Batched Set Cover Problem
Juan C. Mart´ınez Mori1 and Samitha Samaranayake2
1 Systems, Cornell University
[email protected]
2 School of Civil and Environmental Engineering, Cornell University
[email protected]
Abstract. We introduce the batched set cover problem, which is a gen-
eralization of the online set cover problem. In this problem, the elements
of the ground set that need to be covered arrive in batches. Our main
technical contribution is a tight Ω(Hm−2z +1) lower bound on the com-
petitive ratio of any fractional batched algorithm given an adversary
that is required to produce batches of VC-dimension at least z, for some
z ∈ N0. This restriction on the adversary is motivated by the fact that, in
some real world applications, decisions are made after collecting batches
of data of non-trivial VC-dimension. In particular, ridesharing systems
rely on the batch assignment of trip requests to vehicles, and some re-
lated problems such as that of optimal congregation points for passenger
pickups and dropoffs can be modeled as a batched set cover problem with
VC-dimension greater than or equal to two. Furthermore, we note that
while any online algorithm may be used to solve the batched set cover
problem by artificially sequencing the elements in a batch, this procedure
may neglect the rich information encoded in the complex interactions be-
tween the elements of a batch and the sets that contain them. Therefore,
we propose a minor modification to an online algorithm found in [8] to
obtain an algorithm that attempts to exploit such information. Unfortu-
nately, we are unable to improve its analysis in a way that reflects this
intuition. However, we present computational experiments that provide
empirical evidence of a constant factor improvement in the competitive
ratio. To the best of our knowledge, we are the first to use the VC-
dimension in the context of online (batched) covering problems.
Keywords: set cover · batched · online · primal-dual · VC-dimension ·
ridesharing
1
Introduction
1.1 Background
Let X = {1,··· , n} be a ground set of n elements and S = {S1,··· , Sm} be a
collection of m subsets of X. A set cover is a sub-collection of S such that its
union is X. The set cover problem is to find a minimum cardinality set cover of
X. It is a classic NP-hard problem that is also hard to approximate to within a
factor of (1 − α) ln n in polynomial time for any α > 0 [11,9].
2
Mart´ınez Mori, J. C. and Samaranayake, S.
In the online setting [2,7,8], the members of S are identified a priori, but the
elements of the ground set that need to be covered, along with their respective
set membership, are revealed sequentially. More precisely, the online set cover
problem consists of a game between an algorithm and an oblivious adversary;
one which knows the algorithm but not the realization of any random choices3.
The adversary produces, in advance, a sequence σ = σ1, σ2,··· of elements of X,
which it reveals to the algorithm one at a time. Upon the arrival of an element,
the algorithm must either conclude that the element is already in the set cover,
or irrevocably extend the set cover with a member of S containing the element.
Alon et al. [2] gave a deterministic O(log m log n)-competitive algorithm for
the online set cover problem and a nearly matching lower bound for any online
algorithm. Buchbinder and Naor [7,8] later proposed a general scheme for the
design and analysis of online algorithms, namely the primal-dual method4, and
used it to obtain new algorithms for the online set cover problem. Their algo-
rithms generally consist of two phases: i ) a deterministic O(log m) primal-dual
subroutine for the fractional online set cover problem, which is optimal up to
constant terms, and ii ) a randomized rounding procedure whose expected cost
is O(log n) times the cost of the fractional solution, ultimately producing ran-
domized O(log m log n)-competitive algorithms. The rounding procedure can be
derandomized, producing deterministic O(log m log n)-competitive algorithms.
1.2 Contributions
Herein we introduce the batched set cover problem, which is a generalization of
the online set cover problem. However, as in [7,8], our focus is on its fractional
counterpart; this corresponds to phase i ) of the primal-dual scheme. We im-
mediately recover the integral case through the rounding procedures in phase
ii ), which we leave untouched. In essence, the batched set cover problem differs
from the online set cover problem in that the adversary produces a sequence of
batches of elements of X. Thus, the online set cover problem is a special case of
the batched set cover problem in which each element revealed by the adversary
is its own batch.
Note that the problem we consider is distinct from the capacitated online set
cover problem with set requests treated by Bhawalkar et al. [5]. They argue that
the uncapacitated problem is not meaningful because the elements in a batch can
be thought of as arriving sequentially, whereas we argue that this is not always
the case. Our main technical contribution is a tight lower bound on the compet-
itive ratio of any fractional batched algorithm given a parametrized restriction
on the adversary. Specifically, if we consider adversaries that are required to pro-
duce batches of Vapnik Chervonenkis (VC)-dimension [12] at least z, for some
z ∈ N0, any fractional batched algorithm is Ω(Hm−2z+1)-competitive. For z > 0,
this bound is more generous (to the algorithm) than the Ω(Hm) bound of the
online setting [7,8], which we recover when z = 0.
3 If the algorithm is deterministic, an oblivious adversary is equivalent to an adaptive
adversary; one which makes requests adaptively in response to the algorithm [4].
4 This scheme first appeared in the context of approximation algorithms [15].
The Batched Set Cover Problem
3
In addition, we propose a minor modification to an online algorithm found in
[8] to obtain a dedicated batched algorithm. The main idea is the simultaneous
update of the dual variables that correspond to unsatisfied primal constraints,
which is reminiscent of a primal-dual algorithm in [14] for the generalized Steiner
tree problem. Unfortunately, we are unable to analyze this algorithm in a way
that exhibits the effects of the more generous (parametrized) bound. Alterna-
tively, we provide computational results that suggest that while the greedy strat-
egy proposed in [5] is inoffensive for the worst case instance given 0 ≤ z ≤ 1,
it compromises on the competitive ratio obtained on the worst case instance
given z > 1. Our experiments suggest that proposed algorithm improves on the
competitive ratio obtained by the greedy strategy by some constant factor.
The significance of this problem stems from the fact that, in some real-world
applications, decisions are made after collecting a batch of data. Moreover, in
many of these applications, batches of data are rarely produced by an absolute
worst case adversary. The intent of our restricted adversarial model is to mimic
the worst case instances that may effectively arise in the real world. For example,
high-capacity ridesharing systems rely on the batch assignment of trip requests
to vehicles [3], and some related problems such as that of optimal congregation
points for passenger pickups and dropoffs can be modeled as a batched set cover
problem. Intuitively, sequencing a batch of travel requests defeats the purpose
of preparing the batch in the first place. Moreover, the batches that arise in
this setting tend to have a VC-dimension greater than or equal to two, as the
application revolves around exploiting the overlaps between distinct requests (see
Section 2). Our results formalize this intuition. As listed in [5], further examples
of applications of the batched set cover problem may be found in distributed
computing, facility planning, and subscription markets.
To the best of our knowledge, we are the first to use the VC-dimension in
the context of online (batched) covering problems. The VC-dimension has been
used successfully in the context of approximation algorithms for (offline) set
cover problems [6,10], as well as in the context of improved running time bounds
for unconstrained [1] and constrained [13] shortest path algorithms. In both of
these settings, the algorithms exploit the low VC-dimension of the set systems
on which they operate. Perhaps surprisingly, algorithms for the batched set cover
problem may instead exploit the high VC-dimension of the set systems on which
they operate, which we model as a restriction on the adversary. Intuitively, the
reason is that the adversary is forced to reveal complex, intertwined batches.
A dedicated algorithm attempts to exploit the richness of the information re-
vealed, while a greedy algorithm is myopic to the interactions between the set
memberships of the elements in a batch.
1.3 Organization
In Section 2 we formally introduce our problems and definitions. In Section 3 we
consider bounds for the online fractional set cover problem. We present a known
lower bound of Ω(Hm) on the competitive ratio of any online algorithm. While
the tightness of this lower bound (up to constants) follows immediately from
4
Mart´ınez Mori, J. C. and Samaranayake, S.
the existence of O(Hm)-competitive fractional algorithms [7,8], we present an
inductive proof that shows the tightness of the lower bound without the explicit
need of a competitive algorithm. This technique is used in Section 4 to show the
tightness of a Ω(Hm−2z+1) lower bound on the competitive ratio of any batched
algorithm, given our restricted adversary parametrized by z. The reason for doing
this is our argument that batching may offer a constant factor improvement in
the competitive ratio. Hence, tightness up to constants is not informative enough
for our purposes. In Section 5 we formalize the greedy strategy suggested in [5]
and present our minor modification to an online algorithm found in [8]. We also
present the results of our computational experiments.
2 Preliminaries
Let xj ∈ {0, 1} be set to 1 if Sj is brought to the set cover and to 0 otherwise.
Now, consider LP 1, which describes the linear programming relaxation of the
offline set cover problem. We refer to LP 1 as the primal covering problem. Here,
cj > 0 refers to the cost of bringing some set Sj ∈ S to the set cover, and the
objective is to minimize the total cost incurred. In the unweighted case, cj = 1
for all j = 1,··· , m. Constraints (1.1) ensure that every element i = 1,··· , n in
the ground set X is covered. Note that the set membership information of each
element is encoded in its respective constraint.
minimize
s.t.
(1.1)
(cid:80)
j:i∈Sj
xj ≥ 1, i = 1,··· , n
xj ≥ 0, j = 1,··· , m
(LP 1)
The primal covering problem has an associated dual packing problem, described
in LP 2. We refer to this primal-dual formulation throughout this work. We will
refer to the collection of sets in S that individually contain σi ∈ X by S(σi).
m(cid:80)
j=1
cjxj
n(cid:80)
i=1
yi
maximize
s.t.
(2.1)
(cid:80)
i∈Sj
yi ≤ cj, j = 1,··· , m
yi ≥ 0, i = 1,··· , n
(LP 2)
In the fractional online setting [7,8], the objective function of LP 1 is known
a priori, but constraints (1.1) are revealed one by one. This corresponds to
the algorithm identifying the costs of the sets in S a priori, but the adversary
revealing a sequence σ = σ1, σ2,··· of elements of X, along with their respective
set membership, in an online fashion. Equivalently, the right hand side of the
constraints (2.1) of LP 2 are known a priori, but the variables involved in them
and in the objective function are revealed one by one.
The Batched Set Cover Problem
5
Now, consider the following batched version of the set cover problem, which
is also a game between an algorithm and an oblivious adversary. In the batched
set cover problem, S is identified a priori, but the adversary produces a sequence
β = β1, β2,··· of batches of elements of X, which it reveals one batch at a time.
For instance, βk = {σk,1,··· , σk,βk} ⊆ X, where βk denotes the size of the kth
batch. When a batch arrives, all of its elements, along with their respective set
membership information, are revealed simultaneously. The fractional batched
setting is analogous to the fractional online setting, except constraints (1.1)
appear in tandem. Equivalently, the variables involved in the objective function
and constraints (2.1) are revealed in tandem. Note that the online setting is
trivially recovered when each batch is a singleton. We refer to the union of sets
in S that individually cover the elements in βk = {σk,1,··· , σk,βk}, namely
S(σk,1) ∪ ··· ∪ S(σk,βk), by S(βk).
We define an instance I of the online set cover problem as a collection S
together with the adversarial sequence. We introduce the following performance
measures. The batched setting for both of these measures is analogous.
Definition 1. An online algorithm ALGO is said to be c-competitive if for every
instance I of the problem it outputs a solution of cost at most c · OPT(I), where
OPT(I) is the cost of the optimal offline solution.
Definition 2. An online adversary ADVO is said to be c-advantaged if it pro-
duces an instance I such that every online algorithm ALGO outputs a solution of
cost at least c· OPT(I), where OPT(I) is the cost of the optimal offline solution.
Our analysis in Section 4 relies on imposing a minimum on the VC-dimension
of any batch βk produced by the adversary. The VC-dimension was first proposed
by Vapnik and Chernovekis [12], and it is a widely used measure of complexity
in computational learning theory. We work with the following definitions.
Definition 3 (Set System). A set system (X,S) is a ground set X together
with a collection S of subsets of X.
Definition 4 (Shattering). A subset B ⊆ X is said to be shattered by S if
{S ∩ B : S ∈ S} = P(B), where P(B) is the power set of B.
Definition 5 (VC-dimension). The VC-dimension of a set system (X,S) is
the cardinality of the largest subset B ⊆ X to be shattered by S. We denote it by
VCD(X,S).
In particular, upon the arrival of a batch βk we obtain a set system (βk,S),
where S is known a priori. Moreover, note that i ) restricting the adversary to
produce batches βk with VC-dimension VCD(βk,S) ≥ z is only meaningful when
m = S ≥ 2z; otherwise the adversary is unable to produce any batches, and ii )
by definition, any batch satisfying VCD(βk,S) ≥ z also satisfies βk ≥ z. This is
illustrated in Figure 1, which showcases how a batch βk satisfying VCD(βk,S) ≥
z can be constructed with m = 2z and βk = z. Observe that in each of the
cases, βk is shattered since each of its subsets is the intersection of βk with some
6
Mart´ınez Mori, J. C. and Samaranayake, S.
S ∈ S. Of course, given z, there may be instances for which m > 2z, or for which
the adversary produces batches satisfying βk > z, or both. We consider these
cases in our analysis in Section 4.
(a) VCD(βk,S) ≥ 0.
(b) VCD(βk,S) ≥ 1.
(c) VCD(βk,S) ≥ 2.
(d) VCD(βk,S) ≥ 3.
Fig. 1: Construction of S and βk satisfying VCD(βk,S) ≥ z, for z = 0, 1, 2, 3,
with the minimum possible cardinality requirements on m = S and βk. The
construction for z > 3 is analogous. The nodes labeled S1,··· , Sm represent the
sets in S, whereas the circles around them represent the elements of βk ⊆ X
they contain (i.e., constraints in LP 1).
In a ridesharing context, we may interpret each S ∈ S as a possible congre-
gation point (e.g., an intersection in a street network), whereas each constraint
corresponds to the set of compatible congregation points (e.g., within walking
distance) for each trip origin or destination. Note that the only reasons why
we would have VCD(βk,S) ≤ 1 are if i ) βk ≤ 1, or ii ) βk > 1 but all the
travel request form either a collection of proper subsets or a collection completely
disjoint subsets of the possible congregation points. Given sufficiently high and
heterogeneous demand, batches with VCD(βk,S) ≤ 1 are unlikely to arise; the
batches that arise look more like the those in Figure 1 (c) and (d).
S1S2S1S3S2S4S1S8S6S5S7S2S3S4S1 The Batched Set Cover Problem
7
3 Fractional Online Set Cover
Lemma 1 (Variation of Buchbinder and Naor [8]). There exists an in-
stance I∗ of the unweighted fractional online set cover problem such that any
online algorithm is Ω (Hm)-competitive on this instance.
Sj∈S
(cid:84)
Sj. Upon the arrival of σ1, ALGO must satisfy(cid:80)
Proof. Consider the following instance I∗, which is particular to the sequence σ
produced by an adversary in response to an arbitrary online algorithm ALGO.
Let σ1 ∈
j:σ1∈Sj xj ≥ 1.
Thus, it must let xj ≥ 1/m for at least one Sj ∈ S. Refer to such Sj ∈ S as S1 and
to its corresponding variable as x1. Now, let σ2 ∈
Sj∈S\S1 Sj. Upon the arrival
j:σ2∈Sj xj ≥ 1. Thus, it must let xj ≥ 1/(m − 1)
for at least one Sj ∈ S \ S1. Again, refer to such Sj ∈ S \ S1 as S2 and to its
corresponding variable as x2. In general, an adversary may continue revealing
elements σi satisfying
of σ2, ALGO must satisfy (cid:80)
(cid:84)
(cid:92)
(cid:83)
σi ∈
Sj∈S\
i(cid:48)<i Si(cid:48)
Sj,
(cid:83)
(cid:83)
(cid:48)
(cid:48)
i(cid:48)<i Si
i(cid:48)<i Si
forcing the algorithm to let xj ≥ 1/(m− i + 1) for at least one Sj ∈ S \
.
as Si and to its corresponding variable as
Refer to such Sj ∈ S \
xi. After m steps, the total cost ALGO(I∗) incurred by the algorithm, namely
x1 + ··· + xm, is at least
1
m
+ 1 = Hm.
+ ··· +
1
2
+
1
m − 1
Meanwhile, the total cost OP T (I∗) incurred by an optimal offline solution is 1,
which corresponds to simply letting xm = 1. Thus, Hm ≤ ALGO(I
∗)
OP T (I∗) .
Note that I∗ depends on ALGO only in the sense that the particular adver-
sarial sequence σ produced is a response to the particular algorithm; the lower
bound on the competitive factor, on the other hand, is independent of the al-
gorithm. Thus, we may parametrize the instance I∗ in Lemma 1 by m = S to
obtain the instance class I∗(m). In other words, I∗(m) refers to the instances
that produce a lower bound of Hm on the competitive factor of any online algo-
rithm as a result of the adversary following the strategy in the proof of Lemma 1.
If we vary m, we obtain the family of instance classes I∗ = {I∗(m) : m ∈ Z+}.
The tightness of this lower bound (up to constants) is an immediate result
of the existence of O(log m)-competitive algorithms for the fractional online set
cover problem [7,8]. In Lemma 3 we present a different approach to show that
this lower bound is tight, this time without relying on a particular algorithm.
We use Lemma 2, whose proof is in the Appendix.
Lemma 2. Hr > 1
2 (Hr−t + Ht) for any integers r ≥ t ≥ 0.
8
Mart´ınez Mori, J. C. and Samaranayake, S.
Lemma 3. There exists an online algorithm ALGO for the unweighted frac-
tional online set cover problem such that any adversary is O(Hm)-advantaged.
In particular, this bound is matched by any adversary that follows the strategy
in the family of instance classes I∗, described in the proof of Lemma 1.
Proof. By Lemma 1, there exists an adversary that is Hm-advantaged, namely
one that follows the strategy of instance class I∗(m). We need to show that this
is indeed the best an adversary can be guaranteed to achieve. We prove this by
strong induction on m and by focusing on an arbitrary adversary ADVO. We will
make use of the existence of a randomized algorithm ALGO that, in principle,
produces specific outcomes with non-zero probability.
Base Case (m = 1): When ADVO reveals any element σ1, it must be the
case that σ1 ∈ S1. Then, ALGO must let x1 = 1, achieving a competitive factor
of H1, as in I∗(1)
Inductive Step: Assume, by way of strong induction, than the statement is
true for m = 2,··· , w. We need to show that the statement is true for m =
w + 1. By Lemma 1, there exists an adversary that is Hw+1-advantaged, namely
one that follows the strategy of instance class I∗(w + 1). Now, consider the
case in which ADVO deviates from the strategy of I∗(w + 1) on some arbitrary
step i. Let σi∗ be the ith element according to the strategy of I∗(w + 1). If
ADVO reveals an element σi such that S(σi) ∩ S(σi∗ ) is empty, the cost of
the optimal solution increases by 1, which by Lemma 2 irrevocably decreases
the advantage of ADVO. Therefore, suppose that ADVO reveals an element σi
such that S(σi) ∩ S(σi∗ ) is non-empty. Then, with non-zero probability ALGO
disregards all Sj ∈ S(σi) \ S(σi∗ ), if any, making such deviation futile. Thus,
safely assume that ADVO instead reveals an element σi such that S(σi) ⊂ S(σi∗ ).
Let t = S(σi), r = S(σi∗ ), and note that r > t. Let i(cid:48) > i be the first step after
step i such that S(σi) ∩ S(σi(cid:48)) (cid:54)= ∅. As before, with non-zero probability ALGO
disregards all Sj ∈ S(σi(cid:48)) \ S(σi), if any, so safely assume that ADVO reveals
an element σi(cid:48) such that S(σi(cid:48)) ⊂ S(σi). Then, by the inductive hypothesis,
given that element σi satisfied S(σi) ⊂ S(σi∗ ), the best ADVO can do is to
recreate I∗(t) on the remainder of the steps, starting with i(cid:48). In particular, the
best ADVO can do is to reveal an element σi(cid:48) such that S(σi(cid:48)) ⊂ S(σi) and
S(σi(cid:48)) + 1 = S(σi). A symmetric argument can be made about concurrently
recreating I∗(r − t) on the remainder of the steps, which is disjoint from I∗(t)
after the ith step and hence increases the offline solution by one. However, by
Lemma 2, this achieves a strictly lower competitive advantage for ADVO.
4 Fractional Batched Set Cover
4.1 General Case
Lemma 4. There exists an instance I∗ of the unweighted fractional batched set
cover problem such that any batched algorithm is Ω (Hm)-competitive on this
instance.
The Batched Set Cover Problem
9
Lemma 5. There exists a batched algorithm ALGB for the unweighted fractional
batched set cover problem such that any adversary is O(Hm)-advantaged.
Lemma 4 follows from the fact that the fractional online set cover problem is a
special case of the fractional batched set cover problem, together with Lemma 1.
In Section 4.2 we consider the case in which the adversary is imposed a mini-
mum VC-dimension for any batch βk produced. In the proof of Lemma 7, we
mention why an adversary never benefits from producing batches with a VC-
dimension larger than the minimum required. This, together with Lemma 3,
yields Lemma 5.
4.2 Restricted Adversary
Our intent now is to characterize instance classes that distinguish the fractional
batched set cover problem from the fractional online set cover problem. In partic-
ular, we restrict the adversary in that it is forced to produce batches βk satisfying
VCD(βk,S) ≥ z, for some z ∈ N0. Given z, we assume that m = S ≥ 2z and
βk ≥ z, as described in Section 2.
Lemma 6. There exists an instance I∗z of the unweighted fractional batched set
cover problem, with an adversary satisfying VCD(βk,S) ≥ z for any batch βk,
such that any batched algorithm is Ω (Hm−2z+1)-competitive on this instance.
Proof. This proof is analogous to that of Lemma 1. Consider the following in-
stance I∗z , which is particular to the sequence β produced by an adversary sat-
isfying VCD(βk,S) ≥ z for any batch βk, in response to an arbitrary batched
algorithm ALGB. In the following, our ordering of S1,··· , Sm ∈ S is arbitrary.
Let β1 = {σ1,1,··· , σ1,z, σ1,z+1} such that S({σ1,1,··· , σ1,z}) is as in the
diagrams in Figure 1 on sets S1,··· , S2z−1, S2z with β1∩S2z = β1, in addition to
each σ1,q ∈ β1 being also contained in each Sj ∈ {S2z , S2z+1,··· , Sm}. This last
property cannot decrease the VC-dimension, so β1 is a valid batch. Then, because
of the constraint corresponding to σ1,z+1 ∈ β1, ALGB must let xj ≥
m−2z+1
for at least one Sj ∈ {S2z ,··· , Sm}. For clarity and without loss of generality,
assume such Sj is S2z .
Then, let β2 = {σ2,1,··· , σ2,z, σ2,z+1} such that S({σ2,1,··· , σ2,z}) is as in
the diagrams in Figure 1 on sets S2,··· , S2z , S2z+1 with β2 ∩ S2z+1 = β2, in
addition to each σ2,q ∈ β2 being also contained in each Sj ∈ {S2z+1,··· , Sm}.
Then, because of the constraint corresponding to σ2,z+1 ∈ β2, ALGB must let
m−2z−1+1 for at least one Sj ∈ {S2z+1,··· , Sm}. For clarity and without
xj ≥
loss of generality, assume such Sj is S2z+1.
In general, an adversary may continue revealing βk = {σk,1,··· , σk,z, σk,z+1}
such that S({σk,1,··· , σk,z}) is as in the diagrams in Figure 1 on sets Sk,··· ,
S2z+k−2, S2z+k−1 with βk ∩ S2z+k−1 = βk, in addition to each σk,q ∈ βk being
also contained in each Sj ∈ {S2z+k−1,··· , Sm}. Then, because of the constraint
corresponding to σk,z+1 ∈ βk, ALGB must let xj ≥
m−2z−(k−1)+1 for at least
one Sj ∈ {S2z+k−1,··· , Sm}. For clarity and without loss of generality, assume
such Sj is S2z+k−1.
1
1
1
10
Mart´ınez Mori, J. C. and Samaranayake, S.
After the m− 2z + 1 possible steps, the total cost ALGB(I∗z ) incurred by the
algorithm, namely x2z + ··· + xm, is at least
Hm−2z+1 =
1
m − 2z + 1
+
1
m − 2z + ··· +
1
2
+ 1.
Meanwhile, the total cost OP T (I∗z ) incurred by an optimal offline solution is 1,
∗
which corresponds to simply letting xm−2z+1 = 1. Thus, Hm−2z+1 ≤ ALGO(I
z )
z ) .
OP T (I∗
In simple terms, the adversary may capitalize on Sj ∈ {S2z ,··· , Sm} while
assuming a sunk cost on Sj ∈ {S1,··· , S2z−1}.
As in Section 3, we parametrize the instances I∗z in Lemma 6 by m. Thus, given
z, we obtain the family of instance classes I∗z = {I∗z (m) : m ∈ Z+}. We recover
the general case when z = 0. Analogous to Lemma 3, Lemma 7 shows the lower
bound given z is tight. Its proof can be found in the Appendix.
Lemma 7. There exists a batched algorithm ALGB for the unweighted fractional
online set cover problem such that any adversary satisfying VCD(βk,S) ≥ z for
any batch βk is O (Hm−2z+1)-advantaged. In particular, this bound is matched
by any adversary that follows the strategy in the family of instance classes I∗z ,
described in the proof of Lemma 6.
5 Batched Algorithms
5.1 Analysis
Note that since any batch βk could be artificially decomposed into a sequence
of βk elements that are 'revealed' one at a time, any competitive algorithm
for online set cover would produce a competitive feasible solution; we refer to
this as the 'trivial' approach. More precisely, the trivial approach consists of two
phases: i ) some (possibly randomized) subroutine that executes a mapping f :
), where σk,q(cid:48) is the q(cid:48)th element of the artificial sequence,
βk → (σk,1(cid:48),··· , σk,b(cid:48)
followed by ii ) any competitive algorithm for the online set cover problem.
k
An example of such approach is the O(log m)-competitive primal-dual sub-
routine in Algorithm 1, which we refer to as ALGB,T , followed by any O(log n)
rounding technique (i.e., the second phase of the primal-dual method). ALGB,T
is a minor modification of the O(log m)-competitive Algorithm 2 in Section 4.2
of [8] for the batched setting. Note that d = maxk,q S(σk,q) ≤ m. Its correctness
follows immediately from Theorem 4.2 of [8]. For conciseness, we only mention
that the proof relies on showing three claims together with weak duality: i )
the algorithm produces a primal feasible solution to LP 1, ii ) the algorithm
produces a dual feasible solution to LP 2, and iii ) the objective value of LP 1
is bounded above by O(log m) times the objective value of LP 2. Clearly, the
complete algorithm is O(log m log n)-competitive.
Nevertheless, unless βk = 1 for all k, the trivial approach may fail to leverage
the rich information that is possibly implicit in the fact that all elements of a
The Batched Set Cover Problem
11
Algorithm 1: ALGB,T : 'Trivial' batched algorithm.
/* Upon the arrival of batch k := {σk,1,··· , σk,βk}:
1 for σk,q ∈ f ({σk,1,··· ,βk}) do
*/
j:σk,q∈Sj
while (cid:80)
exp
xj ← 1
d
xj < 1 do
ln (1+d)
cj
Increase yk,q continuously
(cid:80)
2
3
4
− 1
,∀j : σk,q ∈ Sj
yk(cid:48),q(cid:48)
k(cid:48),q(cid:48):
σk(cid:48) ,q(cid:48)∈Sj
end
5
6 end
given batch are revealed simultaneously. We refer to an algorithm that attempts
to leverage any such information as a 'dedicated' algorithm. We obtain such an
algorithm if we replace ALGB,T with Algorithm 2, which we refer to as ALGB,D.
Algorithm 2: ALGB,D: 'Dedicated' batched algorithm.
/* Upon the arrival of batch k := {σk,1,··· , σk,βk}:
1 while ∃σk,q such that (cid:80)
ln (1+d)
Increase yk,q continuously, ∀q : (cid:80)
− 1
,∀j : ∃σk,q ∈ Sj
exp
xj ← 1
xj < 1 do
j:σk,q∈Sj
j:σk,q∈Sj
(cid:80)
xj < 1
yk(cid:48),q(cid:48)
d
cj
2
3
k(cid:48),q(cid:48):
σk(cid:48) ,q(cid:48)∈Sj
*/
4 end
Note the difference in how the dual variables are updated: sequentially in
ALGB,T and simultaneously in ALGB,D. This is reminiscent of the approach
of increasing multiple variables at once in a primal-dual algorithm by [14] for
the generalized Steiner tree problem. As expected, ALGB,D is also O(log m)-
competitive5; it is also a minor modification of the O(log m)-competitive Algo-
rithm 2 in Section 4.2 of [8], and its correctness also follows immediately from
Theorem 4.2 of [8]. Unfortunately, we are unable to improve its analysis in a way
that reflects the intuition that batching should improve the competitive factor
obtained under certain conditions. For example, we expect batching to help in
the families of instances I∗z , shown in Section 4 to produce a more generous lower
bound on the competitive factor of any algorithm due to the VC-dimension re-
quirement on the batches produced by the adversary. We leave presenting such
5 To be precise, both algorithms are 2 ln(1 + d)-competitive.
12
Mart´ınez Mori, J. C. and Samaranayake, S.
an analysis as an open problem. As an alternative, in Section 5.2 we present
the results of computational experiments that compare the performance (i.e.,
competitive factor) of ALGB,T and ALGB,D on instances of I∗z .
5.2 Computational Experiments
Figure 2 (a) and (b) present the results of computational experiments that
compare the worst-case performance (i.e., competitive factor) of ALGB,T and
ALGB,D obtained on instances of I∗z , respectively, for various values of z and
m. We discretize both algorithms with a step size of = 0.001. We justify the
use of instances of I∗z by the fact that it is not some arbitrary family of instance
classes; Lemma 6 ans Lemma 7 imply it produces tight bounds. Also, note that
because of the symmetric nature of the batches βk that arise in I∗z , the partic-
ular order produced by the mapping f : βk → (σk,1(cid:48),··· , σk,βk(cid:48)) in ALGB,T is
irrelevant for evaluating the worst-case performance of the trivial approach so
long as σk,z+1 is the last element 'revealed'. This notion does not translate to
ALGB,D, since the dual variable updates occur simultaneously. Also, recall that
for instances in Iz the optimal offline solution is one.
(a) ALGB,T .
(b) ALGB,D.
Fig. 2: Competitive ratio of (a) ALGB,T and (b) ALGB,D obtained on instances
of I∗z for z = 0, 1, 2, 3, 4. The solid line indicates to the competitive ratio obtained
by the algorithm, while the dashed line indicates the Ω(Hm−2z+1) lower bound
on the competitive ratio of any batched algorithm.
Note that for z = 0, 1, both algorithms match the theoretical lower bound.
This is expected, as in both of these cases, per the description of I∗z , all the
batches are singletons. On the other hand, for z > 1, neither of the algorithms
match said lower bound. However, it can be observed that ALGB,D is closer
than ALGB,T to the theoretical lower bound on the competitive ratio; this is not
surprising as the algorithm has only be shown to be optimal up to constants. The
0102030405060m=S1.01.52.02.53.03.54.04.55.0ALGB,T(I∗z)/OPT(I∗z)z≥0z≥1z≥2z≥3z≥4z≥50102030405060m=S1.01.52.02.53.03.54.04.55.0ALGB,D(I∗z)/OPT(I∗z)z≥0z≥1z≥2z≥3z≥4z≥5 The Batched Set Cover Problem
13
improvement is more significant as z increases, though it is only by a constant
factor; both curves remain logarithmic. These results provide empirical evidence
of the benefits of batching when the VC-dimension is high.
Acknowledgements
The first author was funded through the Center for Transportation, Environ-
ment, and Community Health (CTECH) as well as a Federal Highway Admin-
istration (FHWA) Dwight David Eisenhower Transportation Fellowship.
14
Mart´ınez Mori, J. C. and Samaranayake, S.
References
1. Abraham, I., Delling, D., Fiat, A., Goldberg, A.V., Werneck, R.F.: Vc-dimension
and shortest path algorithms. In: International Colloquium on Automata, Lan-
guages, and Programming. pp. 690 -- 699. Springer (2011)
2. Alon, N., Awerbuch, B., Azar, Y.: The online set cover problem. In: Proceedings
of the 35th annual ACM Symposium on the Theory of Computation. pp. 100 -- 105
(2003)
3. Alonso-Mora, J., Samaranayake, S., Wallar, A., Frazzoli, E., Rus, D.: On-
demand high-capacity ride-sharing via dynamic trip-vehicle assignment. Pro-
ceedings of
the National Academy of Sciences 114(3), 462 -- 467 (2017).
https://doi.org/10.1073/pnas.1611675114
4. Ben-David, S., Borodin, A., Karp, R., Tardos, G., Wigderson, A.: On the power
of randomization in on-line algorithms. Algorithmica 11(1), 2 -- 14 (1994)
5. Bhawalkar, K., Gollapudi, S., Panigrahi, D.: Online
set
In: LIPIcs-Leibniz
International Proceedings
requests.
vol.
set
ics.
https://doi.org/10.4230/LIPIcs.APPROX-RANDOM.2014.64
28. Schloss Dagstuhl-Leibniz-Zentrum fuer
cover with
in Informat-
(2014).
Informatik
6. Bronnimann, H., Goodrich, M.T.: Almost optimal set covers in finite vc-dimension.
Discrete & Computational Geometry 14(4), 463 -- 479 (1995)
7. Buchbinder, N., Naor, J.: Online primal-dual algorithms for covering and packing
problems. In: Proceedings of the 13th Annual European Symposium on Algorithms.
pp. 689 -- 701 (2005)
8. Buchbinder, N., Naor, J.: The design of competitive online algorithms via a primal-
dual approach. Foundations and Trends in Theoretical Computer Science 3(2-3),
93 -- 263 (2009). https://doi.org/10.1561/0400000024
9. Dinur, I., Steurer, D.: Analytical approach to parallel repetition. In: Proceedings of
the Forty-sixth Annual ACM Symposium on Theory of Computing. pp. 624 -- 633.
STOC '14, ACM (2014). https://doi.org/10.1145/2591796.2591884
10. Even, G., Rawitz, D., Shahar, S.M.: Hitting sets when the vc-dimension is small.
Information Processing Letters 95(2), 358 -- 362 (2005)
11. Feige, U.: A threshold of ln n for approximating set cover. Journal of the ACM
45(4), 634 -- 652 (1998). https://doi.org/10.1145/285055.285059
12. Vapnik, V.N., Chervonenkis, A.Y.: On the uniform convergence of relative frequen-
cies of events to their probabilities. In: Measures of complexity, pp. 11 -- 30. Springer
(2015)
13. Vera, A., Banerjee, S., Samaranayake, S.: Computing constrained shortest-paths
at scale (2017)
14. Williamson, D.P., Goemans, M.X., Mihail, M., Vazirani, V.V.: A primal-dual ap-
proximation algorithm for generalized steiner network problems. Combinatorica
15(3), 435 -- 454 (1995). https://doi.org/10.1007/BF01299747
15. Williamson, D.P., Shmoys, D.B.: The Design of Approximation Algorithms. Cam-
bridge University Press, New York, NY, USA, 1st edn. (2011)
The Batched Set Cover Problem
15
Appendix
5.3 Proof of Lemma 2.
Proof. Assume, without loss of generality, that r − t ≥ t. Note that
Hr =
1
r
+
1
r − 1
+ ··· +
1
t + 1
+ Ht.
Likewise, note that
1
2
(Hr−t + Ht) =
=
=
1
2
1
2
1
2
(cid:18) 1
(cid:18) 1
(cid:18) 1
r − t
r − t
r − t
+
+
+
1
r − t − 1
1
r − t − 1
1
r − t − 1
Clearly,
1
r
+
1
r − 1
+ ··· +
1
t + 1
>
1
2
so the proof is complete.
5.4 Proof of Lemma 7.
(cid:18) 1
(cid:19)
+ 1
1
t − 1
1
t − 1
1
+ ··· +
2
1
2
+ ··· +
+ 1
+ ··· +
+ 1 +
+
1
t
1
2
1
(cid:19)
(cid:19)
+ ···
+ ···
t + 1
1
t + 1
+
1
t
+
+ Ht.
(cid:19)
,
+
1
r − t − 1
+ ···
1
t + 1
r − t
Proof. This proof is analogous to that of Lemma 3. By Lemma 7, there exists
an adversary satisfying VCD(βk,S) ≥ z for any batch βk that is Hm−2z+1-
advantaged, namely one that follows the strategy of instance class I∗z (m). We
need to show that this is indeed the best an adversary can be guaranteed to
achieve. We prove this by strong induction on m and by focusing on an arbitrary
batched adversary ADVB. We will make use of the existence of a randomized
algorithm ALGB that, in principle, produces specific outcomes with non-zero
probability.
the adversary cannot produce any batches.
Base Case (1 ≤ m < 2z). If m < 2z, the statement is vacously true because
Base Case (m = 2z). When ADVB reveals any batch β1 = {σ1,1,··· , σ1,z}, it
must be the case that S(β1) is as in the diagrams in Figure 1 on sets S1,··· , Sm=2z .
Then, with non-zero probability, ALGB disregards any Sj /∈ {Sm}. In such case,
ALGB lets xm = 1, achieving a competitive factor of H1. This is in agreement
with the description of I∗z (1).
Inductive Step. Assume, by way of strong induction, than the statement
is true for m = 2z + 1,··· , w. We need to show that the statement is true
for m = w + 1. By Lemma 6, there exists an adversary that is Hw+1−2z+1-
advantaged, namely one that follows the strategy of instance class I∗z (w + 1).
Now, consider the case in which ADVB deviates from the strategy of I∗z (w + 1)
16
Mart´ınez Mori, J. C. and Samaranayake, S.
on some arbitrary step k. Let βk = {σk,1,··· , σk,l}, where l ≥ z, and let
βk∗ = {σk∗,1,··· , σk∗,z, σk∗,z+1} be the kth batch according to the strategy
of I∗z (w + 1). Further, denote S∩βk = ∩l
q=1S(σk,q), with t = S∩βk, as well as
S∩βk∗ = ∩z
q=1S(σk∗,z), with r = S∩βk∗. If S∩βk∩S∩βk∗ = ∅, then the cost of the
optimal offline solution increases by 1, which by Lemma 2 irrevocably decreases
the advantage of ADVB. Therefore, suppose that ADVB reveals a batch βk such
that S∩βk ∩ S∩βk∗ (cid:54)= ∅. Moreover, with non-zero probability ALGB disregards
all Sj ∈ S(βk) \ S∩βk , so safely assume this is the case for the rest of the proof.
Now, with non-zero probability ALGB disregards all Sj ∈ S∩βk \ S∩βk∗ , if any,
making such deviation futile. Thus, safely assume that S∩βk ⊂ S∩βk∗ , implying
that t < r. Let k(cid:48) > k be the first step after step k such that S∩βk ∩ S(βk(cid:48)) (cid:54)= ∅.
As before, with non-zero probability ALGB disregards all Sj ∈ S(βk(cid:48))\S∩βk(cid:48) , so
safely assume that this is the case for the rest of the proof. Then, with non-zero
probability ALGB also disregards any Sj ∈ S∩βk(cid:48) \S∩βk , if any, so safely assume
that S∩βk(cid:48) ⊂ S∩βk . Then, by the inductive hypothesis, given that batch βk satis-
fied S∩βk ⊂ S∩βk∗ , the best ADVB can do is to recreate I∗z (t) on the remainder
of the steps, starting with k(cid:48). In particular, this requires S∩βk(cid:48) ⊂ S∩βk and
S∩βk(cid:48) + 1 = S∩βk. A symmetric argument can be made about concurrently
recreating I∗z (r − t) on the remainder of the steps, which is disjoint from I∗z (t)
after the kth step and hence increases the offline solution by one. However, by
Lemma 2, this achieves a strictly lower competitive advantage for ADVB.
Note that ADVB cannot be guaranteed to benefit from producing batches βk
satisfying VCD(βk,S) > z because with non-zero probability ALGB disregards
any Sj ∈ S(βk)\S∩βk , making such deviation futile with non-zero probability. In
fact, a larger VC-dimension would involve more sets, possibly making t = S∩βk
(and hence the attainable competitive advantage) smaller.
|
1306.3857 | 1 | 1306 | 2013-06-17T13:48:44 | Computing Tree-depth Faster Than $2^{n}$ | [
"cs.DS",
"math.CO"
] | A connected graph has tree-depth at most $k$ if it is a subgraph of the closure of a rooted tree whose height is at most $k$. We give an algorithm which for a given $n$-vertex graph $G$, in time $\mathcal{O}(1.9602^n)$ computes the tree-depth of $G$. Our algorithm is based on combinatorial results revealing the structure of minimal rooted trees whose closures contain $G$. | cs.DS | cs |
Computing Tree-depth Faster Than 2n
Fedor V. Fomin(cid:63), Archontia C. Giannopoulou∗, and Micha(cid:32)l Pilipczuk∗
Department of Informatics, University of Bergen,
{fominarchontia.giannopouloumichal.pilipczuk}@ii.uib.no
P.O. Box 7803, N-5020 Bergen, Norway
Abstract. A connected graph has tree-depth at most k if it is a sub-
graph of the closure of a rooted tree whose height is at most k. We give
an algorithm which for a given n-vertex graph G, in time O(1.9602n)
computes the tree-depth of G. Our algorithm is based on combinatorial
results revealing the structure of minimal rooted trees whose closures
contain G.
1
Introduction
The tree-depth of a graph G, denoted td(G), is the minimum number k such that
there is a rooted forest F , not necessarily a subgraph of G, with the following
properties.
-- V (G) = V (F ),
-- Every tree in F is of height at most k, i.e. the longest path between the root
of the tree and any of its leaves contains at most k vertices,
-- G is a subgraph of the closure of F , which is the graph obtained from F by
adding all edges between every vertex of F and the vertices contained in the
path from this vertex to the root of the tree that it belongs to.
This parameter has increasingly been receiving attention since it was defined
by Nesetril and Ossona de Mendez in [13] and played a fundamental role in
the theory of classes of bounded expansion [14,15,16,17]. Tree-depth is a very
natural graph parameter, and due to different applications, was rediscovered
several times under different names as the vertex ranking number [2], the ordered
coloring [10], and the minimum height of an elimination tree of a graph [13].
From the algorithmic perspective, it has been known that the problem of
computing tree-depth is NP-hard even when restricted to bipartite graphs [2,13].
However, it also admits polynomial time algorithms for specific graph classes [6,12].
For example, when the input graph is a tree its tree-depth can be computed in
linear time [20]. Moreover, as tree-depth is closed under minors, from the results
of Robertson and Seymour [18,19], the problem is in FPT when parameterized by
the solution size. In [2], Bodlaender et al. showed that the computation of tree-
depth is also in XP when parameterized by treewidth. From the point of view
(cid:63) Supported by European Research Council (ERC) Grant "Rigorous Theory of Pre-
processing", reference 267959.
of approximation, tree-depth can be approximated in polynomial time within a
factor of O(log2 n) [4], where n is the number of vertices of the input graph.
Moreover, there is a simple approximation algorithm that, given a graph G, re-
turns a forest F such that G is contained in the closure of F and the height
of F is at most 2td(G) [17]. Finally, it is easy to see that there exists an exact
algorithm for the computation of tree-depth running in O∗(2n) time1.
We are interested in tree-depth from the perspective of exact exponential
time algorithms. Tree-depth is intimately related to another two well studied
parameters, treewidth and pathwidth. The treewidth of a graph can be defined
as the minimum taken over all possible completions into a chordal graph of the
maximum clique size minus one. Similar, path-width can be defined in terms of
completion to an interval graph. One of the equivalent definitions of tree-depth
is as the largest clique size in a completion to a trivially perfect graph. These
graph classes form the following chain
trivially perfect ⊂ interval ⊂ chordal,
corresponding to the parameters tree-depth, pathwidth, and treewidth.
However, while for the computation of treewidth and pathwidth there exist
O∗(cn), c < 2, time algorithms [7,8,11,21], to the best of our knowledge no such
algorithm for tree-depth has been known prior to our work. In this paper, we
construct the first exact algorithm which for any input graph G computes its tree-
depth in time O∗(cn), c < 2. The running time of the algorithm is O∗(1.9602n).
The approach is based on the structural characteristics of the minimal forest
that defines the tree-depth of the graph.
The rest of the paper is organized as follows. In Section 2 we give some
basic definitions and preliminary combinatorial results on the minimal trees for
tree-depth and in Section 3, based on the results from Section 2, we present the
O(1.9602n) time algorithm for tree-depth. Finally, in Section 4 we conclude with
open problems.
2 Minimal Rooted Forests for Tree-depth
2.1 Preliminaries
For a graph G = (V, E), we use V (G) to denote V and E(G) to denote E. If
S ⊆ V (G) we denote by G \ S the graph obtained from G after removing the
vertices of S. In the case where S = {u}, we abuse notation and write G \ u
instead of G \ {u}. We denote by G[S] the subgraph of G induced by set S. For
S ⊆ V (G), the open neighborhood of S in G, NG(S), is the set {u ∈ G \ S
∃v ∈ S : {u, v} ∈ E(G)}. Again, in the case where S = {v} we abuse notation
and write NG(v) instead of NG({v}). Given two vertices v and u we denote by
distG(v, u) their distance in G. We use C(G) to denote the set of connected
components of G.
1 The O∗(·) notation suppresses factors that are polynomial in the input size.
2
2.2 Tree-depth
A rooted forest is a disjoint union of rooted trees. The height of a vertex x in
a rooted forest F is the number of vertices of the path from the root (of the
tree to which x belongs) to x and is denoted by height(x, F ). The height of
F is the maximum height of the vertices of F and is denoted by height(F ).
Let x, y be vertices of F . The vertex x is an ancestor of y if x belongs to
the path linking y and the root of the tree to which y belongs. The closure
clos(F ) of a rooted forest F is the graph with vertex set V (F ) and edge set
{{x, y} x is an ancestor of y in F, x (cid:54)= y}. For every vertex y of F we denote
by Py the unique path linking y and the root of the tree to which y belongs, and
denote by p(y) the parent of y in F , i.e. the neighbor of y in Py. Vertices whose
parent is y are called the children of y. We call a vertex x of F a branching point
if x is not a root of F and degF (x) > 2 or if x is a root of F and degF (x) ≥ 2.
For a vertex v of a rooted tree T , we denote by Tv the maximal subtree of T
rooted in v. For example, if v is the root of T , then Tv = T .
Let G be a graph. The tree-depth of G, denoted td(G), is the least k ∈ N
such that there exists a rooted forest F on the same vertex set as G such that
G ⊆ clos(F ) and height(F ) = k. Note that if G is connected then F must be a
tree, and the tree-depth of a disconnected graph is the maximum of tree-depth
among its connected components. Thus, when computing tree-depth we may
focus on the case when G is connected and F is required to be a rooted tree.
With every rooted tree T of height h we associate a sequence (t1, t2, t3, . . . ),
where ti = {v height(v, T ) = i}, i ∈ N, that is, ti is the number of vertices of
the tree T of height i, i ∈ N. Note that since T is finite, this sequence contains
only finitely many non-zero values.
1, t1
2, t1
Let T1 and T2 be two rooted trees with heights h1 and h2, and corresponding
3, . . . ), respectively. We say that T1 ≺ T2 if
sequences (t1
and only if there exists an i ∈ N such that t1
j , for every j > i.
Note in particular that if h1 < h2, then taking i = h2 in this definition proves
that T1 ≺ T2.
3, . . . ) and (t2
j = t2
i < t2
i and t1
1, t2
2, t2
Definition 1. Let G be a connected graph. A rooted tree T is minimal for G if
1. V (T ) = V (G) and G ⊆ clos(T ), and
2. there is no tree T (cid:48) such that V (T (cid:48)) = V (G), G ⊆ clos(T (cid:48)), and T (cid:48) ≺ T .
The next observation follows from the definitions of ≺ and of tree-depth.
Observation 1. Let G be a connected graph and T be a rooted tree for G such
that V (T ) = V (G), G ⊆ clos(T ), and height(T ) > td(G). Then there exists
a rooted tree T (cid:48) such that V (T (cid:48)) = V (G), G ⊆ clos(T (cid:48)), and height(T (cid:48)) <
height(T ), and thus T (cid:48) ≺ T .
The following combinatorial lemmata reveal the structures of minimal trees
which will be handy in the algorithm.
3
Lemma 1. Let T 1 be a rooted tree with root r, v ∈ V (T 1), and T ∗ be a rooted
tree with root r∗ such that T ∗ ≺ T 1
v . If T 2 is the rooted tree obtained from T 1
after considering the union of T 1 \ V (T 1
v ) with T ∗ and adding an edge between
r∗ and p(v) (if p(v) exists), then T 2 ≺ T 1.
As T ∗ ≺ T 1
Proof. Notice first that the claim trivially holds for the case where v = r as then
v = T 1 and T 2 = T ∗. Thus, from now on we prove the claim assuming that
T 1
v (cid:54)= r. Let then h be the height of the vertex p(v) in T 1.
v , there exists an index i such that the number of vertices of height
i in T ∗ is strictly smaller than the number of vertices of height i in T 1
v , and for
every j > i, the number of vertices of height j is equal in both T ∗ and T 1
v . This
implies that the number of vertices of height h + i in T 2 is strictly smaller than
the number of vertices of height h + i in T 1, and for every j > h + i, the number
of vertices of height j is equal in both T 1 and T 2. Thus, we again conclude that
(cid:117)(cid:116)
T 2 ≺ T 1.
Lemma 2. Let G be a connected graph. If T is a minimal tree for G with root
r then for every v ∈ V (T ),
1. G[V (Tv)] is connected,
2. Tv is a minimal tree for G[V (Tv)], and
3. if v(cid:48) ∈ V (Tv) is a branching point such that distTv (v, v(cid:48)) is minimum then
NG(v) ∩ V (Tu) (cid:54)= ∅, for every child u of v(cid:48).
Proof. We first prove (1). Assume in contrary that there exists a vertex v ∈ V (T )
such that the graph G[V (Tv)] is not connected. Notice that we may choose v
in such a way that distT (r, v) is maximum. We first exclude the case where
v = r. Indeed, notice that if v = r, then G[V (Tr)] = G is connected by the
hypothesis. Thus, v (cid:54)= r. Notice also that if v is a leaf of T then Tv is the graph
consisting of one vertex, so it is again connected. Therefore, v is not a leaf of T .
Let v1, v2, . . . , vp be the children of v. The choice of v (maximality of distance
from r) implies that G[V (Tvi)] is a connected component of G[V (Tv)]\ v, i ∈ [p].
Moreover, from the fact that G[V (Tv)] is not connected, it follows that there
exists at least one i0 ∈ [p] such that NG(v) ∩ V (Tvi0
) = ∅. Let T (cid:48) be the tree
obtained from T by removing the edge {v, vi0} and adding the edge {p(v), vi0}.
Observe that G ⊆ clos(T (cid:48)). Moreover, notice that by construction of T (cid:48), we may
consider T (cid:48) as the tree obtained from the union of T \ V (Tp(v)) with T (cid:48)
p(v) after
p(v) ≺ Tp(v).
adding the edge {p(v), p(p(v))} (if p(v) (cid:54)= r). It is easy to see that T (cid:48)
Therefore, from Lemma 1, we end up with a contradiction to the minimality of
T .
To prove (2), we assume in contrary that there exists a vertex v ∈ V (T )
such that Tv is not a minimal tree for G[V (Tv)]. By the hypothesis that T is
a minimal tree for G, it follows that v (cid:54)= r. As Tv is not a minimal tree for
G[V (Tv)], there exists a rooted tree T (cid:48) with root r(cid:48) such that V (T (cid:48)) = V (Tv),
G[V (Tv)] ⊆ clos(T (cid:48)), and T (cid:48) ≺ Tv. Let now T ∗ be the rooted tree obtained
from the union of T \ V (Tv) with T (cid:48) after adding an edge between p(v) and r(cid:48).
4
Notice then that G ⊆ clos(T ∗). Moreover, from Lemma 1, we get that T ∗ ≺ T ,
a contradiction to the minimality of T .
We now prove (3). Let v be a vertex of T and v(cid:48) be a branching point of Tv
such that distTv (v, v(cid:48)) is minimum, that is, v(cid:48) is the highest branching point in
Tv. Assume in contrary that there exists a child u of v(cid:48) such that NG(v)∩V (Tu) =
∅. Let T (cid:48) be the tree obtained from T by switching the position of the vertices
v and v(cid:48), where T (cid:48) = T if v = v(cid:48). Notice that clos(T ) = clos(T (cid:48)) and T and
T (cid:48) are isomorphic, hence T (cid:48) is also a minimal tree for G. Moreover, children of
v in T (cid:48) are exactly children of v(cid:48) in T . Observe also that if w is a child of v(cid:48) in
w)) ⊆ {v}.
T , hence also a child of v in T (cid:48), then Tw = T (cid:48)
w and NG[V (T (cid:48)
As NG(v) ∩ V (Tu) = ∅, we obtain that G[V (T (cid:48)
v)] is not connected. However,
T (cid:48) is a minimal tree for G and therefore, from (1), G[V (T (cid:48)
v)] is connected, a
contradiction. This completes the proof of the last claim and of the lemma. (cid:117)(cid:116)
v)](V (T (cid:48)
3 Computing tree-depth
3.1 The naive DP, and pruning the space of states
To construct our algorithm, we need an equivalent recursive definition of tree-
depth.
1
Proposition 1 ([13]). The tree-depth of a connected graph G is equal to
td(G) =
1 + min
v∈V (G)
if V (G) = 1
max
H∈C(G\v)
td(H) otherwise
(1)
Proposition 1 already suggests a dynamic programming algorithm computing
tree-depth of a given graph G in O∗(2n) time. Assume without loss of general-
ity that G is connected, as otherwise we may compute the tree-depth of each
connected component of G separately. For every X ⊆ V (G) such that G[X] is
connected, we compute td(G[X]) using (1). Assuming that the tree-depth of all
the connected graphs induced by smaller subsets of vertices has been already
computed, computation of formula (1) takes polynomial time. Hence, if we em-
ploy dynamic programming starting with the smallest sets X, we can compute
td(G) in O∗(2n) time. Let us denote this algorithm by A0.
The reason why A0 runs in pessimistic O∗(2n) time is that the number of
subsets of V (G) inducing connected subgraphs can be as large as O(2n). There-
fore, if we aim at reducing the time complexity, we need to prune the space of
states significantly. Let us choose some ε, 0 < ε < 1
6 , to be determined later,
and let G be a connected graph on n vertices. We define the space of states Sε
as follows:
Sε = {S ⊆ V (G) 1 ≤ S ≤(cid:0) 1
∃X ⊆ V (G) : X ≤(cid:0) 1
2 − ε(cid:1) n and G[S] is connected, or
2 − ε(cid:1) n and G[S] ∈ C(G \ X)}.
Observe that thus all the sets belonging to Sε induce connected subgraphs of
G. The subsets S ∈ Sε considered in the first part of the definition will be called
5
(cid:19)(cid:19)
time.
n
2−ε)n
( 1
. Moreover,
Proof. For sets of the first type, there are at most n ·
n
2 − ε(cid:1) n
(cid:0) 1
n
of the first type, and the ones considered in the second part will be called of the
second type. Note that V (G) ∈ Sε since it is a subset of second type for X = ∅.
(cid:1)(cid:17)
Lemma 3. If G is a graph on n vertices, then Sε = O∗(cid:16)(cid:0)
Sε may be enumerated in O∗(cid:18)(cid:18)
(cid:18)
(cid:19)
(cid:0) 1
2 − ε(cid:1) n
2 − ε(cid:1) n. Moreover, one can enumerate them in O∗(cid:18)(cid:18)
at most (cid:0) 1
(cid:0) 1
2 − ε(cid:1) n
2 − ε(cid:1) n in O∗(cid:18)(cid:18)
(cid:19)(cid:19)
all the vertex sets X of size at most(cid:0) 1
(cid:0) 1
2 − ε(cid:1) n
time, and for each run a polynomial-time check whether it induces a connected
subgraph. For the sets of the second type, we can in the same manner enumerate
time, and
for each of them take all of the at most n connected components of G \ X. (cid:117)(cid:116)
In our algorithms we store the family Sε as a collection of binary vectors
of length n in a prefix tree (a trie). Thus when constructing Sε we can avoid
enumerating duplicates, and then test belonging to Sε in O(n) time.
We now define the pruned dynamic programming algorithm Aε that for every
X ∈ Sε computes value td∗(G[X]) defined as follows:
sets S of size
(cid:19)(cid:19)
n
n
1
td∗(G[X]) =
H∈C(G[X]\v), V (H)∈Sε
max
td∗(H)
if X = 1
otherwise
(2)
1 + min
v∈X
We use convention that td∗(G[X]) = +∞ if X /∈ Sε. The algorithm Aε can
be implemented in a similar manner as A so that its running time is O∗(Sε).
We consider sets from Sε in increasing order of cardinalities (sorting Sε with
respect to cardinalities takes O∗(Sε) time) and simply apply formula (2). Note
that computation of formula (2) takes polynomial time, since we need to consider
at most n vertices v, and for every connected component H ∈ C(G \ v) we can
test whether its vertex set belongs to Sε in O(n) time.
For a set S ∈ Sε and T being a minimal tree for G[S], we say that T is
covered by Sε if V (Tv) ∈ Sε for every v ∈ S. The following lemma expresses the
crucial property of td∗.
Lemma 4. For any connected graph G and any subset S ⊆ V (G), it holds that
td∗(G[S]) ≥ td(G[S]). Moreover, if S ∈ Sε and there exists a minimal tree T
for G[S] that is covered by Sε, then td∗(G[S]) = td(G[S]).
Proof. We first prove the first claim by induction with respect to the cardinality
of S. If td∗(G[S]) = +∞ then the claim is trivial. Therefore, we assume that S ∈
Sε, there exists some r ∈ S such that td∗(G[S]) = 1 + maxH∈C(G[S]\r) td∗(H),
and V (H) ∈ Sε for each H ∈ C(G[S] \ r). By the induction hypothesis, since
V (H) ≤ S for each H ∈ C(G[S] \ r), we infer that td∗(H) ≥ td(H) for
6
each H ∈ C(G[S] \ r). On the other hand, by (1) we have that td(G[S]) ≤
1 + maxH∈C(G[S]\r) td(H). Therefore,
td(G[S]) ≤ 1 + max
≤ 1 + max
td∗(H) = td∗(G[S]),
H∈C(G[S]\r)
td(H)
H∈C(G[S]\r)
and the induction step follows.
We now prove the second claim, again by induction with respect to the
cardinality of S. Let T be a minimal tree for G[S] that is covered by Sε. Let
r be the root of T and let v1, v2, . . . , vp be the children of r in T . By (2) of
Lemma 2, we have that Tvi is a minimal tree for G[V (Tvi )], for each i ∈ [p].
Moreover, since T was covered by Sε, then so does each Tvi. By the induction
hypothesis we infer that td∗(G[V (Tvi )]) = td(G[V (Tvi )]) for each i ∈ [p], since
V (Tvi) < S. Moreover, since T and each Tvi are minimal, we have that
td(G[S]) = height(T ) = 1 + max
i∈[p]
height(Tvi) = 1 + max
i∈[p]
td(G[V (Tvi)])
= 1 + max
i∈[p]
td∗(G[V (Tvi)]) ≥ td∗(G[S]).
The last inequality follows from the fact that, by (1) of Lemma 2, G[V (Tvi)] are
connected components of G[S] \ r and moreover that their vertex sets belong
to Sε. Hence, vertex r was considered in (2) when defining td∗(G[S]). We infer
that td(G[S]) ≥ td∗(G[S]), and td(G[S]) ≤ td∗(G[S]) by the first claim, so
(cid:117)(cid:116)
td∗(G[S]) = td(G[S]).
Lemma 4 implies that the tree-depth is already computed exactly for all
2 − ε)n, then td∗(G[S]) = td(G[S]).
connected subgraphs induced by significantly less than half of the vertices.
Corollary 1. For any connected graph G on n vertices and any S ∈ Sε, if
S ≤ ( 1
Proof. If T is a minimal tree for G[S], then for every v ∈ V (T ), G[V (Tv)] is
connected by (1) of Lemma 2, and V (Tv) ≤ ( 1
2 −ε)n. Hence, for every v ∈ V (T )
it holds that V (Tv) ∈ Sε and the corollary follows from Lemma 4.
(cid:117)(cid:116)
Finally, we observe that for any input graph G the algorithm Aε already
computes the tree-depth of G unless every minimal tree for G has a very special
structure. Let T be a minimal tree for G. A vertex v ∈ V (G) is called problematic
2 − ε)n. We say that a minimal
if (i) V (Tv) > ( 1
tree T for G is problematic if it contains some problematic vertex.
2 − ε)n, and (ii) V (Pp(v)) > ( 1
Corollary 2. For any connected graph G, if G admits a minimal tree that is
not problematic, then td∗(G) = td(G).
Proof. We prove that any minimal tree T for G that is not problematic, is in
fact covered by Sε. Then the corollary follows from Lemma 4.
7
G[V (Tv)] is connected by (1) of Lemma 2. Hence if V (Tv) ≤(cid:0) 1
V (Pp(v)) ≤(cid:0) 1
2 − ε(cid:1) n, then it
2 − ε(cid:1) n, since v is not problematic. Note then that NG(V (Tv)) ⊆
Take any v ∈ V (G); we need to prove that V (Tv) ∈ Sε. First note that
trivially holds that V (Tv) ∈ Sε by the definition of Sε. Otherwise we have that
V (Pp(v)) and so G[V (Tv)] is a connected component of V (G) \ V (Pp(v)). Conse-
(cid:117)(cid:116)
quently, V (Tv) is a subset of second type for X = V (Pp(v)).
3.2 The algorithm
Corollary 2 already restricts cases when the pruned dynamic program Aε misses
the minimal tree: this happens only when all the minimal trees for the input
graph G are problematic. Therefore, it remains to investigate the structure of
problematic minimal trees to find out, if some problematic minimal tree could
have smaller height than the one computed by Aε.
problematic, we have that Z >(cid:0) 1
Let G be the input graph on n vertices. Throughout this section we assume
that G admits some problematic minimal tree T . Let v be a problematic vertex
in T . Let moreover v(cid:48) be the highest branching point in Tv (possibly v(cid:48) = v
if v is already a branching point in T ), or v(cid:48) be the only leaf of Tv in case Tv
does not contain any branching points. Let Z = V (Pv(cid:48)); observe that since v is
of T rooted in NT (Z \{v(cid:48)}), that is, in the children of vertices of Z \{v(cid:48)} that do
not belong to Z, and let R1, R2, . . . , Rb be the subtrees of T rooted in children
of v(cid:48). Note that trees Q1, Q2, . . . , Qa, R1, R2, . . . , Rb are pairwise disjoint, and
by the definition of a minimal tree we have that NG(V (Qi)), NG(V (Rj)) ⊆ Z
for any i ∈ [a], j ∈ [b]. See Figure 1 for reference.
2 − ε(cid:1) n. Let Q1, Q2, . . . , Qa be all the subtrees
Let Q = (cid:83)a
i=1 V (Qi) and R = (cid:83)b
Z ⊇ NG(V (Rj) ∪(cid:83)a
j=1 V (Rj). For any problematic mini-
mal tree T and a problematic vertex v in it, we say that v defines the sets
Z, Q1, Q2, . . . , Qa, R1, R2, . . . , Rb, Q, R in T .
Observation 2. If b > 0, then Z = NG(V (Rj) ∪ Q) for any j ∈ [b].
Proof. As NG(V (Qi)), NG(V (Rj)) ⊆ Z for any i ∈ [a], j ∈ [b], we have that
Take any z ∈ Z, and let z(cid:48) be the highest branching point in Tz; note that z(cid:48)
is always defined since b > 0 and thus v(cid:48) is a branching point. If z(cid:48) = v(cid:48), then by
(3) of Lemma 2 we infer that z ∈ NG(V (Rj)), j ∈ [b]. Otherwise, we have that if
z(cid:48) ∈ Z \ {v(cid:48)}. Since z(cid:48) is a branching point, there exists some subtree Qi rooted
in a child of z(cid:48). We can again use (3) of Lemma 2 to infer that z ∈ NG(V (Qi)),
so also z ∈ NG(Q).
(cid:117)(cid:116)
i=1 V (Qi)). We proceed to proving the reverse inclusion.
Observe that if b = 0, then we trivially have that Z = V (G) \ Q.
Observation 3. Q < 2εn.
2 − ε(cid:1) n and V (Tv) >
Proof. Since v is problematic, we have that V (Pp(v)) >(cid:0) 1
2 − ε(cid:1) n. Observe also that V (Pp(v)) ∪ V (Tv) = Z ∪ R by the definition of Z.
(cid:0) 1
Since V (Pp(v)) ∩ V (Tv) = ∅ we have that:
Z ∪ R = V (Pp(v)) ∪ V (Tv) > (1 − 2ε)n.
8
r
Q1
Q2
Q
Z
Qa
R1
v
v(cid:48)
R2
. . .
Rb
R
Fig. 1. A problematic minimal tree T rooted at r, the problematic vertex v, the branch-
ing vertex v(cid:48) in Tv, the set Z (the set consisting of white vertices), and the sets Qi,
i ∈ [a], and Rj, j ∈ [b]. We would like to remark here that the figure only aims to
facilitate identification of the above sets and that it is possible that, for some such tree
T , the vertices in Z \ v(cid:48) have more children or there exists a vertex q in Q such that
height(q, T ) > height(r, T ) for every vertex r ∈ R.
Since Q = V (G) \ (Z ∪ R), the claim follows.
Observation 4. R <(cid:0) 1
Proof. Since R = V (G) \ (Z ∪ Q) and Z >(cid:0) 1
2 + ε(cid:1) n − Q.
2 − ε(cid:1) n, we have that
R = V (G) \ (Z ∪ Q) = n − Z − Q <
+ ε
(cid:117)(cid:116)
(cid:117)(cid:116)
(cid:18) 1
2
(cid:19)
(cid:18) 1
4
n − Q.
(cid:19)
+
ε
2
n − Q
2
.
Observation 5. If b > 0, then b ≥ 2 and min
j∈[b]
V (Rj) <
Proof. If b > 0 then v(cid:48) is a branching point and it has at least two children. It
follows that b ≥ 2. For the second claim, observe that since b ≥ 2 we have that
(cid:117)(cid:116)
minj∈[b] V (Rj) ≤ R/2 and the claim follows from Observation 4.
We can proceed to the description of our main algorithm, denoted further A.
Similarly as before, without loss of generality let us assume that G is connected.
First, the algorithm constructs the family Sε using Lemma 3, and runs the algo-
.
rithm Aε on it. Note that these steps can be performed in time O∗(cid:16)(cid:0)
(cid:1)(cid:17)
n
2−ε)n
( 1
9
(cid:18) 1
(cid:19)
(cid:18) 1
(cid:19)
ε
2
(cid:18) 1
4
(cid:19)
3ε
2
We can therefore assume that the value td∗(G[S]) is computed for every S ∈ Sε,
and in particular for S = V (G).
Now the algorithm proceeds to checking whether a better problematic mini-
mal tree T with problematic vertex v can be constructed. We adopt the notation
introduced in the previous paragraphs for a problematic minimal tree T . We aim
at identifying set Z and sets V (Q1), V (Q2), . . . , V (Qa), V (R1), V (R2), . . . , V (Rb).
Without loss of generality assume that if b > 0, then V (R1) has the smallest car-
dinality among V (R1), V (R2), . . . , V (Rb), i.e., V (R1) ≤ V (R2), . . . ,V (Rb).
Let then Y = Q ∪ R1 if b > 0, and Y = Q if b = 0.
The algorithm branches into at most (n + 1) subbranches, in each fixing the
expected cardinality y of Y . Note that by Observations 3 and 5 and the fact
that ε < 1
6 we may assume that
y < Q +
n − Q
Then the algorithm branches into (cid:0)n
ε
2
=
+
4
2
Q
2
y
4
+
+
n <
(cid:1) subbranches, in each fixing a different
n.
+
subset of vertices of size smaller than y as the set Y . Note that sets V (Q1),
V (Q2), . . . , V (Qa), V (R1) are then defined as vertex sets of connected compo-
nents of G[Y ]. The algorithm branches further into (n + 1) cases. In one case
the algorithm assumes that b = 0 and therefore concludes that Q = Y . In other
cases the algorithm assumes that b > 0 and picks one of the components of G[Y ]
assuming that its vertex set is V (R1), thus recognizing Q as Y \ V (R1), i.e., the
union of vertex sets of remaining components of G[Y ].
In the case when b = 0 the algorithm concludes that Z = V (G) \ Q. In the
cases when b > 0, the algorithm concludes that Z = NG(Y ) using Observation 2.
Having identified Z, the sets V (R1), V (R2), . . . , V (Rj) can be recognized as ver-
tex sets of connected components of V (G) \ (Z ∪ Q). Observations 2, 3, and 5
ensure that for every problematic minimal tree T for G, there will be at least one
subbranch where sets Z, V (Q1), V (Q2), . . . , V (Qa), V (R1), V (R2), . . . , V (Rb) are
fixed correctly. Observe also that in each of at most (n + 1) branches where
y has been fixed, we produced at most (n + 1) ·(cid:0)n
(cid:1) subbranches. We per-
form also sanity checks: whenever any produced branch does not satisfy any
of Observations 2, 3, 4 or 5, or the fact that V (R1) is a smallest set among
V (R1), V (R2), . . . , V (Rj), we terminate the branch immediately.
The algorithm now computes td(G[V (Qi)]) and td(G[V (Rj)]) for all i ∈ [a],
j ∈ [b]. Recall that by Corollary 1, for any set X ⊆ V (G) such that G[X] is
connected and X ≤ ( 1
2 − ε)n, we have that td(G[X]) = td∗(G[X]), and hence
the value td(G[X]) has been already computed by algorithm Aε. Since Q ≤ 2εn
6 , we infer that this is the case for every set V (Qi) for i ∈ [a], and values
and ε < 1
td(G[V (Qi)]) are already computed. The same holds for every Rj assuming that
V (Rj) ≤ ( 1
2 − ε)n, i.e.,
we have no guarantee that the algorithm Aε computed td(G[V (Rj0)]) correctly.
Note that by Observation 4 and the fact that ε < 1
6 , there can be at most one
such j0. Furthermore, if this is the case, then by Observation 5 we have that
Assume then that there exists some j0 such that V (Rj0 ) > ( 1
2 − ε)n.
y
10
b ≥ 2 and V (Rj0) cannot be the smallest among sets V (R1), V (R2), . . . , V (Rb);
hence, j0 (cid:54)= 1 and V (Rj0) ⊆ V (G)\ (Z ∪ Y ). Therefore, we must necessarily have
that
y = Y ≤ V (G) − Z − V (Rj0 ) < n −
− ε
n −
− ε
n = 2εn,
(cid:19)
(cid:18) 1
(cid:18) 1
2
2
(cid:19)
(cid:18) 1
(cid:18) 1
2
(cid:19)
(cid:19)
− ε
n − y =
n − y.
+ ε
2
and moreover
V (Rj0 ) ≤ V (G) − Z − Y < n −
y
2 +ε)n−y
2−ε)n
( 1
2
Formally, if none of these assertions holds, the branch would be terminated
by the sanity check. To compute td(G[V (Rj0 )]) we employ the naive dynamic
programming routine on G[V (Rj0 )], i.e., algorithm A0. Observe, however, that
in this application we do not need to recompute the values for subsets of size at
2 −ε)n, since the values for them were already computed by the algorithm
most ( 1
6 , the application of algorithm
Aε. Therefore, since Rj0 ≤(cid:0) 1
2 + ε(cid:1) n−y and ε < 1
A takes at most O∗((cid:0)( 1
(cid:1)) time.
Summarizing, for every choice of y (recall that y < (cid:0) 1
rithm produced at most (n + 1) ·(cid:0)n
could have used extra O∗((cid:0)( 1
2 +ε)n−y
2−ε)n
( 1
(cid:1) n), the algo-
(cid:1) branches, and in branches with y < 2εn it
(cid:1)) time for computing values td(G[V (Rj)])
4 + 3ε
whenever there was no guarantee that algorithm Aε computed them correctly.
We arrive at the situation where in each branch the algorithm already identi-
fied set Z, sets V (Q1), V (Q2), . . . , V (Qa), V (R1), V (R2), . . . , V (Rb), and values
td(G[V (Qi)]) and td(G[V (Rj)]) for i ∈ [a], j ∈ [b]. Note, however, that the
algorithm does not have yet the full knowledge of the shape of tree T , because
we have not yet determined in which order the vertices of Z appear on the path
Pv(cid:48), and thus we do not know where the trees Qi and Rj are attached to this
path. Fortunately, it turns out that finding an optimum such ordering of vertices
of Z is polynomial-time solvable.
For i ∈ [a + b] let Mi = Qi if i ≤ a and Mi = Ri−a otherwise, and let
hi = td(G[V (Mi)]). Note that since T is minimal, by (2) of Lemma 2 we have
that hi = height(Mi) for each i ∈ [a + b]. Let also Zi = NG(V (Mi)); note that
since G ⊆ clos(T ), we have that Zi ⊆ Z. Let σ be an ordering of Z, i.e., σ is a
bijective function from Z to [Z]. Finally, we define the weight of σ as follows:
(cid:18)
Z, max
i∈[a+b]
(cid:19)
µ(σ) = max
(max(σ(Zi)) + hi)
.
(3)
Lemma 5. Let G be the input graph, and let Z,{V (Mi)}i∈[a+b] be any parti-
tioning of vertices of G such that Zi = NG(V (Mi)) is a subset of Z for any
i ∈ [a + b]. Moreover, let hi = td(G[V (Mi)]) and for σ being an ordering of Z,
let µ(σ) be defined by (3). Then td(G) ≤ µ(σ) for any ordering σ of Z. However,
if G admits a problematic minimal tree T and Z,{V (Mi)}i∈[a+b] are defined by
any problematic vertex in this tree, then td(G) = minσ µ(σ).
11
Proof. We first prove that td(G) ≤ µ(σ) for any such partitioning {V (Mi)}i∈[a+b],
Z of V (G) and ordering σ of Z. Construct a rooted tree T (cid:48) as follows. First, cre-
ate a path on vertex set Z, where the vertices are ordered as in ordering σ
and σ−1(1) is the root of the tree. Then, for every i ∈ [a + b] construct a min-
imal tree Ti for G[V (Mi)], and attach its root using one edge to the vertex
σ−1(mi), where mi = max(σ(Zi)). Observe that every neighbor of V (Mi) is
before σ−1(mi) in the ordering σ, and hence it follows that G ⊆ clos(T (cid:48)). Con-
sequently, td(G) ≤ height(T (cid:48)). However, by the definition of T (cid:48) and of µ(σ),
we have height(T (cid:48)) = µ(σ). Thus td(G) ≤ µ(σ).
We proceed to the second claim. Assume that T is a problematic minimal
tree for G and assume that Z,{V (Mi)}i∈[a+b] are defined by any problematic
vertex v in this tree. We adopt the notation used for T in this section. Let σ0 be
the order of vertices of Z on the path Pv(cid:48). For a tree Mi, for i ∈ [a + b], let zi ∈ Z
be the parent of the root of Mi; hence, for i > a we have zi = v(cid:48). Observe that,
then td(G) = height(T ) = max(Z, maxi∈[a+b] σ0(zi) + hi). Since G ⊆ clos(T ),
we infer that σ0(zi) ≥ σ0(w) for any w ∈ Zi. Hence height(T ) ≥ µ(σ0) by the
definition of µ. Consequently, td(G) ≥ µ(σ0), and so td(G) = minσ µ(σ) by the
(cid:117)(cid:116)
first claim.
We are left with the following scheduling problem. Given a set Z of size at
most n, a family number of subsets Zi ⊆ Z for i ∈ [a + b] and corresponding
integers hi ≤ n, we would like to compute the minimum possible µ(σ) among
orderings σ of Z. Let this problem be called Minimum Ordering with Inde-
pendent Delays (MOID, for short).
Lemma 6. Minimum Ordering with Independent Delays is polynomial-
time solvable.
Proof. Observe that since Z ≤ n and hi ≤ n, for any ordering σ we have that
µ(σ) ≤ 2n. We therefore iterate through all possible values M from Z to 2n,
and for each M we check whether there exists some σ with µ(σ) ≤ M . The first
M for which this test returns a positive outcome is equal to minσ µ(σ).
For a given M , construct an auxiliary bipartite graph H with Z on one
side and {1, 2, . . . ,Z} on the other side. We put an edge between an element
z and an index j if and only if the following holds: for every Zi to which z
belongs, it holds that j + hi ≤ M . It is easy to verify that orderings σ of Z with
µ(σ) ≤ M correspond one-to-one to perfect matchings in H. Indeed, if we are
given an ordering σ with µ(σ) ≤ M , then we have that for every z ∈ Z and
Zi to which z belongs, it holds that σ(z) + hi ≤ M by the definition of µ(σ).
Hence, {z, σ(z)} is an edge in H and {{z, σ(z)} z ∈ Z} is a perfect matching
in H. On the other hand, if we are given a perfect matching {{z, jz} z ∈ Z} in
H, then we may define an ordering σ of Z by putting σ(z) = jz. Then for every
z ∈ Z and Zi to which z belongs, we have that {z, σ(z)} is an edge in H and,
consequently, σ(z) + hi ≤ M . As we chose z and Zi arbitrarily, it follows that
maxi∈[a+b] (max(σ(Zi)) + hi) ≤ M and so µ(σ) ≤ M .
12
also solved in O(n +(cid:80)a+b
Therefore, to solve the MOID problem it suffices to construct H in polyno-
mial time and run any polynomial-time algorithm for finding a perfect matching
(cid:117)(cid:116)
in H.
We remark that Minimum Ordering with Independent Delays can be
i=1 Zi) time using greedy arguments. Since we are not
interested in optimizing polynomial factors, in the proof of Lemma 6 we used
the more concise matching argument to keep the description simple. We leave
finding a faster algorithm for MOID to the reader as an interesting exercise.
Concluding, in every subbranch algorithm A constructs an instance of MOID
and solves it in polynomial time using the algorithm of Lemma 6. Lemma 5
ensures that none of the values found in subbranches will be larger than td(G),
and that if G admits a problematic minimal tree T then td(G) will be found in at
least one subbranch. Therefore, by Corollary 2 we can conclude the algorithm A
by outputting the minimum of td∗(G), computed by Aε, and the values returned
by subbranches.
Let us proceed with the analysis of the running time of algorithm A. First,
we have enumerated Sε and run the algorithm Aε, which took
T1(n) = O∗(cid:18)(cid:18)
(cid:19)(cid:19)
n
2 − ε(cid:1) n
(cid:0) 1
(cid:1) since y <(cid:0) 1
(cid:1) n
n
4 + 3ε
(cid:0) 1
2
is bounded by (n + 1)2 ·(cid:0)
time. Then we created a number of subbranches. For every subbranch with
y ≥ 2εn we have spent polynomial time, and the number of these subbranches
6 . Hence, on
4 + 3ε
2
( 1
these subbranches we spent
(cid:1) n and ε < 1
(cid:19)(cid:19)
n
2 )n
4 + 3ε
T2(n) = O∗(cid:18)(cid:18)
(cid:32)
(cid:19)
·
(cid:1),
y
max
y<2εn
2 +ε)n−y
2−ε)n
( 1
T3(n) = O∗
If we now let ε = 1
time in total. Finally, for every subbranch with y < 2εn we have spent at most
(cid:1)) time. As the number of such branches is bounded by (n+1)·(cid:0)n
O∗((cid:0)( 1
the total time spent on these branches is
also easily shown that for any y < 1
(cid:32)(cid:18)n
(cid:18)( 1
2 + ε)n − y
10 , then T1(n), T2(n) = O∗((cid:0) n
2 − ε)n
( 1
5 n, it holds that(cid:0)n
(cid:19)(cid:33)(cid:33)
(cid:1)) = O∗(1.9602n). It can be
(cid:1) ·(cid:0) 3
(cid:1) = O∗(1.9602n).
To prove this, we can use the following simple combinatorial bound:(cid:0)n1
(cid:1)·(cid:0)n2
(cid:1) ≤
(cid:0)n1+n2
(cid:1). This inequality can be proved by combinatorial interpretation as follows:
(cid:18) 3
(cid:19)
(cid:18)n
5 n − y
2
5 n
k1+k2
every choice of k1 elements from a set of size n1 and of k2 elements from a set
of size n2, defines uniquely a choice of k1 + k2 elements from the union of these
(cid:18) 8
sets, which is of size n1 + n2. Therefore, we obtain:
5 n − y
1
5 n
(cid:18) 3
5 n − y
5 n − y
= O∗(1.828n).
(cid:18)n
(cid:19)
5 n−y
2
5 n
(cid:18) 8
5 n
1
5 n
=
y
·
1
·
y
(cid:19)
≤
(cid:19)
≤
2
5 n
y
(cid:19)
k1
k2
.
(cid:19)
y
13
Consequently, T1(n), T2(n), T3(n) = O∗(1.9602n), and the whole algorithm runs
in O∗(1.9602n) time.
4 Conclusion
In this work we gave the first exact algorithm computing the tree-depth of a
graph faster than O∗(2n). As Bodlaender et al. [3] observe, both pathwidth and
treewidth can be reformulated as vertex ordering problems and thus computed
by a simple dynamic programming algorithm similar to the classical Held-Karp
algorithm in time O∗(2n) [9]. For example, computing the optimum value of
treewidth is equivalent to finding an elimination ordering which minimizes the
sizes of cliques created during the elimination process. As far as tree-depth is
concerned, Nesetril and Ossona de Mendez [17] give an alternative definition of
tree-depth in terms of weak-colorings, which in turn are defined also via ver-
tex orderings; however, it is unclear whether this definition can be used for
an algorithm working in O∗(2n) time. Interestingly enough, for many of vertex
ordering problems, like Hamiltonicity, treewidth, or pathwidth, an explicit al-
gorithm working in time O∗(cn) for some c < 2 can be designed, see [1,7,21].
On the other hand, for several other vertex permutation problems no such algo-
rithms are known. The two natural problems to attack are (i) the computation of
cutwidth, and (ii) the Minimum Feedback Arc Set in Digraph problem; see [3,5]
for definitions and details. It is known that the cutwidth of a graph can be com-
puted in time O∗(2t), where t is the size of a vertex cover in the graph [5]; thus
the problem is solvable in time O∗(2n/2) on bipartite graphs. We leave existence
of faster exponential algorithms for these problems as an open question.
References
1. A. Bjorklund. Determinant sums for undirected hamiltonicity. In Proceedings of
the 51st Annual IEEE Symposium on Foundations of Computer Science (FOCS
2010), pages 173 -- 182. IEEE, 2010.
2. H. L. Bodlaender, J. S. Deogun, K. Jansen, T. Kloks, D. Kratsch, H. Muller, and
Z. Tuza. Rankings of graphs. SIAM J. Discrete Math., 11(1):168 -- 181, 1998.
3. H. L. Bodlaender, F. V. Fomin, A. M. C. A. Koster, D. Kratsch, and D. M. Thi-
likos. A note on exact algorithms for vertex ordering problems on graphs. Theory
Comput. Syst., 50(3):420 -- 432, 2012.
4. H. L. Bodlaender, J. R. Gilbert, H. Hafsteinsson, and T. Kloks. Approximat-
ing treewidth, pathwidth, frontsize, and shortest elimination tree. J. Algorithms,
18(2):238 -- 255, 1995.
5. M. Cygan, D. Lokshtanov, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. On
cutwidth parameterized by vertex cover. Algorithmica, pages 1 -- 14, 2012.
6. J. S. Deogun, T. Kloks, D. Kratsch, and H. Muller. On the vertex ranking problem
for trapezoid, circular-arc and other graphs. Discrete Applied Mathematics, 98(1-
2):39 -- 63, 1999.
7. F. V. Fomin, D. Kratsch, I. Todinca, and Y. Villanger. Exact algorithms for
treewidth and minimum fill-in. SIAM J. Comput., 38(3):1058 -- 1079, 2008.
14
8. F. V. Fomin and Y. Villanger. Treewidth computation and extremal combinatorics.
Combinatorica, 32(3):289 -- 308, 2012.
9. M. Held and R. M. Karp. A dynamic programming approach to sequencing prob-
lems. Journal of SIAM, 10:196 -- 210, 1962.
10. M. Katchalski, W. McCuaig, and S. M. Seager. Ordered colourings. Discrete
Mathematics, 142(1-3):141 -- 154, 1995.
11. K. Kitsunai, Y. Kobayashi, K. Komuro, H. Tamaki, and T. Tano. Computing
directed pathwidth in O(1.89n) time. In Proceedings of the 7th International Sym-
posium on Parameterized and Exact Computation (IPEC), volume 7535 of Lecture
Notes in Computer Sci., pages 182 -- 193. Springer, 2012.
12. T. Kloks, H. Muller, and C. K. Wong. Vertex ranking of asteroidal triple-free
graphs. Inf. Process. Lett., 68(4):201 -- 206, 1998.
13. J. Nesetril and P. O. de Mendez. Tree-depth, subgraph coloring and homomorphism
bounds. Eur. J. Comb., 27(6):1022 -- 1041, 2006.
14. J. Nesetril and P. O. de Mendez. Grad and classes with bounded expansion I.
Decompositions. Eur. J. Comb., 29(3):760 -- 776, 2008.
15. J. Nesetril and P. O. de Mendez. Grad and classes with bounded expansion II.
Algorithmic aspects. Eur. J. Comb., 29(3):777 -- 791, 2008.
16. J. Nesetril and P. O. de Mendez. Grad and classes with bounded expansion III.
Restricted graph homomorphism dualities. Eur. J. Comb., 29(4):1012 -- 1024, 2008.
17. J. Nesetril and P. O. de Mendez. Sparsity - Graphs, Structures, and Algorithms,
volume 28 of Algorithms and combinatorics. Springer, 2012.
18. N. Robertson and P. D. Seymour. Graph Minors. XIII. The Disjoint Paths Problem.
J. Comb. Theory, Ser. B, 63(1):65 -- 110, 1995.
19. N. Robertson and P. D. Seymour. Graph Minors. XX. Wagner's conjecture. J.
Comb. Theory, Ser. B, 92(2):325 -- 357, 2004.
20. A. A. Schaffer. Optimal Node Ranking of Trees in Linear Time. Inf. Process. Lett.,
33(2):91 -- 96, 1989.
21. K. Suchan and Y. Villanger. Computing pathwidth faster than 2n. In Proceed-
ings of the 4th International Workshop on Parameterized and Exact Computation
(IWPEC 2009), volume 5917, pages 324 -- 335. Springer, 2009.
15
|
1411.7346 | 3 | 1411 | 2018-12-07T01:30:55 | A Chasm Between Identity and Equivalence Testing with Conditional Queries | [
"cs.DS",
"cs.CC",
"cs.LG",
"math.PR",
"math.ST",
"math.ST"
] | A recent model for property testing of probability distributions (Chakraborty et al., ITCS 2013, Canonne et al., SICOMP 2015) enables tremendous savings in the sample complexity of testing algorithms, by allowing them to condition the sampling on subsets of the domain. In particular, Canonne, Ron, and Servedio (SICOMP 2015) showed that, in this setting, testing identity of an unknown distribution $D$ (whether $D=D^\ast$ for an explicitly known $D^\ast$) can be done with a constant number of queries, independent of the support size $n$ -- in contrast to the required $\Omega(\sqrt{n})$ in the standard sampling model. It was unclear whether the same stark contrast exists for the case of testing equivalence, where both distributions are unknown. While Canonne et al. established a $\mathrm{poly}(\log n)$-query upper bound for equivalence testing, very recently brought down to $\tilde O(\log\log n)$ by Falahatgar et al. (COLT 2015), whether a dependence on the domain size $n$ is necessary was still open, and explicitly posed by Fischer at the Bertinoro Workshop on Sublinear Algorithms (2014). We show that any testing algorithm for equivalence must make $\Omega(\sqrt{\log\log n})$ queries in the conditional sampling model. This demonstrates a gap between identity and equivalence testing, absent in the standard sampling model (where both problems have sampling complexity $n^{\Theta(1)}$).
We also obtain results on the query complexity of uniformity testing and support-size estimation with conditional samples. We answer a question of Chakraborty et al. (ITCS 2013) showing that non-adaptive uniformity testing indeed requires $\Omega(\log n)$ queries in the conditional model. For the related problem of support-size estimation, we provide both adaptive and non-adaptive algorithms, with query complexities $\mathrm{poly}(\log\log n)$ and $\mathrm{poly}(\log n)$, respectively. | cs.DS | cs |
A Chasm Between Identity and Equivalence Testing with
Conditional Queries
Jayadev Acharya∗
Cl´ement L. Canonne†
Gautam Kamath‡
December 10, 2018
Abstract
A recent model for property testing of probability distributions (Chakraborty et al., ITCS
2013, Canonne et al., SICOMP 2015) enables tremendous savings in the sample complexity of
testing algorithms, by allowing them to condition the sampling on subsets of the domain. In
particular, Canonne, Ron, and Servedio (SICOMP 2015) showed that, in this setting, testing
identity of an unknown distribution D (i. e., whether D = D∗ for an explicitly known D∗)
can be done with a constant number of queries (i. e., samples), independent of the support
size n -- in contrast to the required Ω(√n) in the standard sampling model. However, it was
unclear whether the same stark contrast exists for the case of testing equivalence, where both
distributions are unknown. Indeed, while Canonne et al. established a poly(log n)-query upper
bound for equivalence testing, very recently brought down to O(log log n) by Falahatgar et al.
(COLT 2015), whether a dependence on the domain size n is necessary was still open, and
explicitly posed by Fischer at the Bertinoro Workshop on Sublinear Algorithms (2014). In this
article, we answer the question in the affirmative, showing that any testing algorithm for equiv-
alence must make Ω(cid:0)√log log n(cid:1) queries in the conditional sampling model. Interestingly, this
demonstrates a gap between identity and equivalence testing, absent in the standard sampling
model (where both problems have sampling complexity nΘ(1)).
We also obtain results on the query complexity of uniformity testing and support-size es-
timation with conditional samples. In particular, we answer a question of Chakraborty et al.
(ITCS 2013) showing that non-adaptive uniformity testing indeed requires Ω(log n) queries in
the conditional model. This is an exponential improvement on their previous lower bound of
Ω(log log n), and matches up to polynomial factors their poly(log n) upper bound. For the prob-
lem of support-size estimation, we provide both adaptive and non-adaptive algorithms, with
query complexities poly(log log n) and poly(log n), respectively, and complement them with a
qualitatively tight lower bound of Ω(log n) conditional queries for non-adaptive algorithms.
∗Cornell University. [email protected]. Supported by a Cornell University Start Up, and NSF CRII-CIF-
1657471. Part of this work was done when the author was a postdoctoral researcher at MIT.
†Stanford University. [email protected]. Supported by a Motwani Postdoctoral Fellowship. Part of
this work was performed while the author was a graduate student at Columbia University, and supported by NSF
grants CCF-1115703 and NSF CCF-1319788.
‡Simons Institute for the Theory of Computing. [email protected]. Supported as a Microsoft Research Fellow,
as part of the Simons-Berkeley Research Fellowship program. This work was supported by ONR N00014-12-1-0999,
and NSF grants CCF-0953960 (CAREER) and CCF-1101491.
1
1
Introduction
"No, Virginia, there is no constant-query tester."1
Understanding properties and characteristics of an unknown probability distribution is a funda-
mental problem in statistics, and one that has been thoroughly studied. However, it is only since the
work of Goldreich and Ron [GR00] and Batu et al. [BFR+00] that the problem has been considered
through the lens of theoretical computer science, more particularly in the setting of property testing.
In this framework, an unknown "huge object" -- here a probability distribution over a humongous
domain -- can be accessed only by making a few local inspections, and the goal is to decide whether
the object satisfies some prespecified property. While most of the literature focuses on the large
sample regime and studies error exponents and rates of convergence, more recent testing algorithms
look at these problems in the small-sample regime focusing on the probability of errors and sample
complexity. (We refer the reader to [Fis01, Ron08, Ron10, Gol10] for an introduction and surveys
on the general field of property testing. Moreover, due to the specificities of our model we will
interchangeably use the terms sample and query complexity, referring to conditional samples as
"queries.")
Over the subsequent decade, a flurry of work explored this new area, resulting in a better and
often complete understanding of a number of questions in property testing of distributions, or
distribution testing (see, e. g. [GR00, BFF+01, BKR04, Pan08, RS09, ADJ+11, BFRV11, Rub12,
ILR12, ADJ+12, CDVV14, VV17] or [Can15] for a survey).
In many cases, these culminated
in provably sample-optimal algorithms, all of which required at least an nΩ(1) dependence on
the domain size n in the sample complexity -- a dependence which, while sublinear, can still be
prohibitively large. However, the standard setting of distribution testing, where one only obtains
independent samples from an unknown distribution D, does not encompass all scenarios one may
encounter. In recent years, alternative models have thus been proposed to capture more specific
situations [GMV06, CFGM16, CRS15, LRR13, CR14]. Among these is the conditional oracle
model [CFGM16, CRS15] which will be the focus of our work. In this setting, the testing algorithm
is given the ability to sample from conditional distributions: that is, to specify a subset S of the
domain and obtain samples from DS, the distribution induced by D on S (the formal definition of
the model can be found in Definition 2.1). The hope is that allowing a richer set of queries to the
unknown underlying distributions might significantly reduce the number of samples the algorithms
need, thereby sidestepping the strong lower bounds that hold in the standard sampling model.
1.1 Motivation for the conditional model
A recent trend in testing and learning circumvents these impossibility results by providing a more
flexible type of queries than independent samples. One example can be found in the recent paradigm
of active learning [Das05, Set09] (and its testing counterpart, active testing [BBBY12]), which
modifies and generalizes the usual unsupervised learning paradigm. In this setting, the algorithm
is provided with unlabeled examples only, but can then adaptively request the label of any of these
examples. While not directly comparable to these two frameworks (which are not applicable to
the study of probability distributions), the conditional sampling model we shall work on shares
some similarity in spirit. Specifically, it provides the algorithm with some additional power, and
the ability to perform (some type of) more powerful queries to the object to be learned or tested.
The setting of conditional sampling is related to that of group testing, where the objective is to
identify a set of defective individuals among a large population, by querying whether suitably chosen
1The curious reader is referred to [Wik17].
2
subsets contain at least one defective individual. Group testing has been a field of interest since the
40's and has remained an active area of research since (see, e. g. [Dor43, DH00, ND00, CJSA14]).
The type of queries allowed in this framework is reminiscent of conditional sampling, where one
obtains samples conditioned on a subset. This connection between group testing and conditional
sampling is explored in greater detail in [ACK15].
Of course, a crucial aspect of designing and studying these new models of learning and testing
is to understand how justified and natural they are, and argue that they do indeed capture natural
situations. In the case of distribution testing, the conditional access model does meet these criteria,
as discussed in [CFGM16] and [CRS15]. Namely, besides the purely theoretical aspect of helping
understand the key aspects and limitations of the underlying statistical problems, this framework
is characteristic of situations which arise in natural and social sciences. At a very high-level, any
scenario where an experimenter or practitioner is able to restrict the set of possible outcomes of
an experiment or poll -- e. g., in chemistry, where one might control some factor such as the acidity
of a solution; or in sociology by performing stratified polling -- provides this experimenter with the
sort of access granted by the conditional model.
1.2 Background and previous work
We focus in this paper on proving lower bounds for testing two extremely natural properties of
distributions, namely equivalence testing ("are these two distributions identical?") and support-size
estimation ("how many different outcomes can actually be observed?"). Along the way, we use some
of the techniques we develop to obtain an upper bound on the query complexity of the latter. We
state below an informal definition of these two problems, along with closely related ones (uniformity
and identity testing). Hereafter, "oracle access" to a distribution D over [n] = {1, . . . , n} means
access to samples generated independently from D, and "far" is with respect to the total variation
distance dTV(D1, D2) = supS⊆[n](D1(S) − D2(S)) between probability distributions. Moreover, as
in usual in property testing, in all the problems below we allow the algorithms to be randomized.
In the description of the results below, we will often consider the distance parameter ε as a constant
(and focus on the domain size n as the key parameter).
Uniformity testing: granted oracle access to D, decide whether D equals U[n] (the uniform dis-
tribution on [n]) or is far from it;
Identity testing: granted oracle access to D and the full description of a fixed D∗, decide whether
D equals D∗ or is far from it;
Equivalence (closeness) testing: granted independent oracle accesses to D1, D2 (both un-
known), decide whether D1 and D2 are equal or far from each other.
Support-size estimation: granted oracle access to D, return a multiplicative approximation of
the size of the support2 supp(D) ={ x : D(x) > 0 }.
It is not difficult to see that each of the second and third problems generalizes the previous, and
therefore has query complexity at least as big. All of these tasks are known to require sample
complexity nΩ(1) in the standard sampling model (SAMP); yet, as prior work [CFGM16, CRS15]
shows, their complexity decreases tremendously when one allows the more flexible type of access
to the distribution(s) provided by a conditional sampling oracle (COND). For the problems of
2For this problem, it is typically assumed that all points in the support have probability mass at least Ω(1)/n, as
without such guarantee it becomes impossible to give any non-trivial estimate (consider for instance a distribution
D such that D(i) ∝ 1/2in).
3
uniformity testing and identity testing, the sample complexity even becomes a constant provided
the testing algorithm is allowed to be adaptive (i. e., when the next queries it makes can depend on
the samples it previously obtained).
for the tight bounds on these problems).
The uniformity testing problem exemplifies the savings granted by conditional sampling -- as
Testing uniformity and identity. The identity testing question is a generalization of uniformity
testing, where D∗ is taken to be the uniform distribution over [n]. The complexity of both tasks
is well-understood in the sampling model; in particular, it is known that for both uniformity and
identity testing Θ(cid:0)√n/ε2(cid:1) samples are necessary and sufficient (see [GR00, BFR+10, Pan08, VV17]
Canonne, Ron, and Servedio [CRS15] showed, in this setting only O(cid:0)1/ε2(cid:1) adaptive queries3 are
has constant sample complexity as well, namely O(cid:0)1/ε4(cid:1) -- very recently improved to a near-optimal
O(cid:0)1/ε2(cid:1) by Falahatgar et al. [FJO+15]. The power of the COND model is evident from the fact
that a task requiring polynomially many samples in the standard model can now be achieved with
a number of samples independent of the domain size n.
sufficient (and this is optimal, up to logarithmic factors). They further prove that identity testing
The aforementioned algorithms crucially leverage the ability to make adaptive conditional
queries to the probability distributions. Restricting the study to non-adaptive algorithms, Chakraborty
et al. [CFGM16] describe a poly(log n, 1/ε)-query non-adaptive tester for uniformity, showing that
even without the full power of conditional queries one can still get an exponential improvement over
the standard sampling setting. They also obtain an Ω(log log n) lower bound for this problem, and
leave open the possibility of improving this lower bound to a logarithmic dependence. We answer
this question, establishing in Theorem 1.3 that any non-adaptive uniformity tester must perform
Ω(log n) conditional queries.
Testing equivalence. The equivalence testing problem has been extensively studied over the past
decade, and its sample complexity is now known to be Θ(max(n2/3/ε4/3,√n/ε2)) in the sampling
model [BFR+10, Val11, CDVV14].
In the COND setting, Canonne, Ron, and Servedio showed that equivalence testing is possible
with only poly(log n, 1/ε) queries. Concurrent to our work, Falahatgar et al. [FJO+15] brought this
upper bound down to O(cid:0)(log log n)/ε5(cid:1), a doubly exponential improvement over the nΩ(1) samples
needed in the standard sampling model. However, these results still left open the possibility of
a constant query complexity -- given that both uniformity and identity testing admit constant-
query testers, it is natural to wonder where equivalence testing lies.4 This question was posed by
Fischer at the Bertinoro Workshop on Sublinear Algorithms 2014 [Sublinear.info, Problem 66]. We
make decisive progress in answering it, ruling out the possibility of any constant-query tester for
equivalence. Along with the upper bound of Falahatgar et al. [FJO+15], our results nearly settle
the dependence on the domain size, showing that (log log n)Θ(1) samples are both necessary and
sufficient.
Support-size estimation. Raskhodnikova et al. [RRSS09] showed that obtaining additive es-
timates of the support size requires sample complexity almost linear in n. Subsequent work by
Valiant and Valiant [VV11, VV10a] settles the question, establishing that Θ(n/ log n) samples are
3Here and throughout, we use the notation O(f ) to hide polylogarithmic dependencies on the argument, i. e., for
expressions of the form O(f logc f ) (for some absolute constant c).
4It is worth noting that an Ω(logc n) lower bound was known for equivalence testing in a weaker version of
the conditional oracle, PAIRCOND (where the tester's queries are restricted to being either [n] or subsets of size
2 [CRS15]).
4
both necessary and sufficient. Note that the proof of the Valiants' lower bound translates to multi-
plicative approximations as well, as it hinges on the hardness of distinguishing a distribution with
support s ≤ n from a distribution with support s + εn ≥ (1 + ε)s. To the best of our knowledge, the
question of getting a multiplicative-factor estimate of the support size of a distribution given con-
ditional sampling access has not been previously considered. We provide upper and lower bounds
for both the adaptive and non-adaptive versions of this problem.
1.3 Our results
We make significant progress in each of the problems introduced in the previous section, yielding a
better understanding of their query complexities. We prove four results pertaining to the sample
complexity of equivalence testing, support-size estimation, and uniformity testing in the COND
framework.
Our main result gives a lower bound on the sample complexity of testing equivalence with
adaptive queries under the COND model, resolving in the negative the question of whether constant-
query complexity was achievable [Sublinear.info, Problem 66].
4 , must have query complexity Ω(cid:0)√log log n(cid:1).
Theorem 1.1 (Testing Equivalence). Any adaptive algorithm which, given COND access to un-
known distributions D1, D2 on [n], distinguishes with probability at least 2/3 between (a) D1 = D2
and (b) dTV(D1, D2) ≥ 1
Combined with the recent O(log log n) upper bound of Falahatgar et al. [FJO+15], this almost set-
tles the sample complexity of this question. Furthermore, as the related task of identity testing can
be performed with a constant number of queries in the conditional sampling model, this demon-
strates an intriguing and intrinsic difference between the two problems. Our result also shows an
interesting contrast with the usual sampling model, where both identity and equivalence testing
have polynomial sample complexity. Next, we establish a logarithmic lower bound on non-adaptive
support-size estimation, for any (large enough) constant factor. This improves on the result of
Chakraborty et al. [CFGM16], which gave a doubly logarithmic lower bound for constant factor
support-size estimation.
Theorem 1.2 (Non-Adaptive Support-Size Estimation). Any non-adaptive algorithm which, given
COND access to an unknown distribution D on [n], estimates the size of its support up to a factor
γ ≥ √2 must have query complexity Ω(cid:16) log n
log2 γ(cid:17).
Moreover, the approach used to prove this theorem also implies an analogous lower bound on non-
adaptive uniformity testing in the conditional model, answering a conjecture of Chakraborty et
al. [CFGM16].
Theorem 1.3 (Non-Adaptive Uniformity Testing). Any non-adaptive algorithm which, given COND
access to an unknown distribution D on [n] and parameter ε ∈ (0, 1/4), distinguishes with prob-
ability at least 2/3 between (a) D = U[n] and (b) dTV(D,U[n]) ≥ ε, must have query complexity
Ω((log n)/ε).
We note that these results complement polylog(n)-query upper bounds on non-adaptive support-
size estimation and uniformity testing, the former of which we sketch in this paper, and the latter
obtained by Chakraborty et al. [CFGM16]. This shows that both of these problems have query
complexity logΘ(1) n in the non-adaptive case.
Finally, we conclude with an upper bound for adaptive support-size estimation. Specifically, we
provide a O(log log n)-query algorithm for support-size estimation. This shows that the question
becomes double exponentially easier when conditional samples are allowed.
5
Theorem 1.4 (Adaptive Support-Size Estimation). Let τ > 0 be any constant. There exists an
adaptive algorithm which, given COND access to an unknown distribution D on [n] (guaranteed to
have probability mass at least τ /n on every element of its support) and accuracy parameter ε ∈ (0, 1),
makes O(cid:0)(log log n)/ε3(cid:1) queries to the oracle5 and returns a value ω such that the following holds.
With probability at least 2/3, ω ∈ [ 1
1+ε · ω, (1 + ε) · ω], where ω = supp(D).
Problem
COND model
Standard model
Testing equivalence
Estimating support size
(adaptive)
Estimating support size
(non-adaptive)
Testing uniformity
(non-adaptive)
ε3
O(poly(log n, 1/ε)) (Section 5.4)
ε5
O(cid:16) log log n
(cid:17) [FJO+15]
Ω(cid:0)√log log n(cid:1) (Theorem 1.1)
O(cid:16) log log n
(cid:17) (Theorem 1.4)
Ω(cid:0)√log log n(cid:1) [CFGM16] (†)
O(cid:16) log5 n
ε6 (cid:17) [CFGM16]
Ω(cid:16) log n
ε (cid:17) (Theorem 1.3)
Ω(log n) (Theorem 1.2)
Θ(cid:16)max(cid:16) n2/3
ε4/3 , n1/2
ε2 (cid:17)(cid:17) [CDVV14]
Θ(cid:16) n
log n(cid:17) [VV10a]
Θ(cid:16)√n
ε2 (cid:17) [Pan08]
Table 1: Summary of results. Note that the lower bound (†) can also be easily derived from our
lower bound on testing equivalence.
1.3.1 Relation to the Ron-Tsur model
Recent work of Ron and Tsur [RT16] studies a model which is slightly different -- and more favorable
to the algorithm -- than ours. In their setting, the algorithm still performs queries consisting of a
subset of the domain, as in our case. However, the algorithm is also given the promise that the
distribution is uniform on a subset of the domain, and whenever a query set contains 0 probability
mass the oracle explicitly indicates this is the case. Their paper provides a number of results for
support-size estimation in this model.
We point out two connections between our work and theirs. First, our Ω(log n) lower bound
for non-adaptive support-size estimation (Theorem 1.2) holds in the model of Ron and Tsur. Al-
though lower bounds in the conditional sampling setting do not apply directly to their model, our
construction and analysis do carry over, and provide a nearly tight answer to a question left unan-
swered in their paper. Also, our O(log log n)-query algorithm for adaptive support-size estimation
(Theorem 1.4) can be seen as generalizing their result to the weaker conditional sampling model
(most significantly, when we are not given the promise that the distribution be uniform).
1.4 Techniques and proof ideas
Lower bound on adaptive equivalence testing.
we have to deal with one main conceptual issue: adaptivity. While the standard sampling model
does not, by definition, allow any choice on what the next query to the oracle should be, this is
no longer the case for COND algorithms. Quantifying the power that this grants an algorithm
makes things much more difficult. To handle this point, we follow the approach of Chakraborty et
[CFGM16] and focus on a restricted class of algorithms they introduce, called "core adaptive
al.
In order to prove our main lower bound, Theorem 1.1,
5We remark that the constant in the O depends polynomially on 1/τ .
6
testers" (see Section 2.2 for a formal definition). They show that this class of testers is equivalent
to general algorithms for the purpose of testing a broad class of properties, namely those which are
invariant to any permutation of the domain. Using this characterization, it remains for us to show
that none of these structurally much simpler core testers can distinguish whether they are given
conditional access to (a) a pair of random identical distributions (D1, D1), or (b) two distributions
(D1, D2) drawn according to a similar process, which are far apart.
At a high level, our lower bound works by designing instances where the property can be tested
if and only if the support size is known to the algorithm. Our construction randomizes the support
size by embedding the instance into a polynomially larger domain. Since the algorithm is only
allowed a small number of queries, Yao's Minimax Principle allows us to argue that, with high
probability, a deterministic algorithm is unable to "guess" the support size. This separates queries
into several cases. First, in a sense we make precise, it is somehow "predictable" whether or not a
query will return an element previously observed. If this occurs, it is similarly predictable which
element the query will return. On the other hand, if a fresh element is observed, the query set
is either "too small" or "too large." In the former case, the query will entirely miss the support,
and the sampling process is identical for both types of instance. In the latter case, the query will
hit a large portion of the support, and the amount of information gleaned from a single sample is
minimal.
At a lower level, this process itself is reminiscent of the "hard" instances underlying the lower
bound of Canonne, Ron, and Servedio [CRS15] for testing identity (with a PAIRCOND oracle), with
one pivotal twist. As in their work, both D1 and D2 are uniform within each of ω(1) "buckets"
whose size grows exponentially and are grouped into "bucket-pairs." Then, D2 is obtained from
D1 by internally redistributing the probability mass of each pair of buckets, so that the total mass
of each pair is preserved but each particular bucket has mass going up or down by a constant
factor (see Section 3.1 for details of the construction). However, we now add a final step, where
in both D1 and D2 the resulting distribution's support is scaled by a random factor, effectively
reducing it to a (randomly) negligible fraction of the domain. Intuitively, this last modification has
the role of "blinding" the testing algorithm. We argue that unless its queries are on sets whose
size somehow match (in a sense formalized in Section 3.2) this random size of the support, the
sequences of samples it will obtain under D1 and D2 are almost identically distributed. The above
discussion crucially hides many significant aspects and technical difficulties which we address in
Section 3. Moreover, we observe that the lower bound we obtain seems to be optimal with regard
to our proofs techniques (specifically, to the decision tree approach), and not an artifact of our
lower bound instances. Namely, there appear to be conceptual barriers to strengthening our result,
which would require new ideas.
Lower bound on non-adaptive support-size estimation. Turning to the (non-adaptive)
lower bound of Theorem 1.2, we define two families of distributions D1 and D2, where an instance
is either a draw (D1, D2) from D1 × D2, or simply (D1, D1). Any distribution in D2 has support
size γ times that of its corresponding distribution in D1. Yet, we argue that no non-adaptive
deterministic tester making too few queries can distinguish between these two cases, as the tuple of
samples it will obtain from D1 or (the corresponding) D2 is almost identically distributed (where
the randomness is over the choice of the instance itself). To show this last point, we analyze
separately the case of "small" queries (conditioning on sets which turn out to be much smaller than
the actual support size, and thus with high probability will not even intersect it) and the "large"
ones (where the query set S is so big compared to the support T that a uniform sample from S ∩ T
is essentially indistinguishable from a uniform sample from S). We conclude the proof by invoking
7
Yao's Principle, carrying the lower bound back to the setting of non-adaptive randomized testers.
Interestingly, this argument essentially gives us Theorem 1.3 "for free." Indeed, the big-query-
set case above is handled by proving that the distribution of samples returned on those queries is
indistinguishable, both for D1 and D2, from samples obtained from the actual uniform distribu-
tion. Considering again the small-query-set case separately, this allows us to argue that a random
distribution from (say) D1 is indistinguishable from uniform.
Upper bound on support-size estimation. Our algorithm for estimating the support size
within a constant factor (Theorem 1.4) is simple in spirit, and follows a guess-and-check strategy.
In more detail, it first obtains a "reference point" outside the support, to check whether subsequent
samples it may consider belong to the support. Then, it attempts to find a rough upper bound on
the size of the support, of the form 22j
(so that only log log n many options have to be considered);
by using its reference point to check if a uniform random subset of this size contains, as it should,
at least one point from the support. Once such an upper bound has been obtained using this
double-exponential strategy, a refined bound is then obtained via a binary search on the new range
of values for the exponent, {2j−1, . . . , 2j}. Not surprisingly, our algorithm draws on similar ideas
as in [RT16, Sto85], with some additional machinery to supplement the differences in the models.
Interestingly, as a side-effect, this upper bound shows our analysis of Theorem 1.1 to be tight up to
a quadratic improvement. Indeed, the lower bound construction we consider (see Section 3.1) can
be easily "defeated" if an estimate of the support size is known, and therefore cannot yield better
than a Ω(log log n) lower bound. Similarly, this further shows that the adaptive lower bound for
support-size estimation of Chakraborty et al. [CFGM16] is also tight up to a quadratic improvement.
Organization. The rest of the paper describes details and proofs of the results mentioned in the
above discussion. In Section 2, we introduce the necessary definitions and some of the tools we shall
use. Section 3 covers our main result on adaptive equivalence testing, Theorem 1.1. In Section 4 we
prove our lower bounds for support-size estimation and uniformity testing, and Section 5 details our
upper bounds for support-size estimation. The corresponding sections may be read independently.
2 Preliminaries
2.1 Notation and sampling models
All throughout this paper, we denote by [n] the set {1, . . . , n}, and by log the logarithm in base 2.
A probability distribution over a (countable) domain [n] is a non-negative function D : [n] → [0, 1]
such that Px∈[n] D(x) = 1. We denote by US the uniform distribution on a set S. Given a
distribution D over [n] and a set S ⊆ [n], we write D(S) for the total probability mass Px∈S D(x)
assigned to S by D. Finally, for S ⊆ [n] such that D(S) > 0, we denote by DS the conditional
distribution of D restricted to S, that is DS(x) = D(x)/D(S) for x ∈ S and DS(x) = 0 otherwise.
As is usual in distribution testing, in this work the distance between two distributions D1, D2
on [n] will be the total variation distance.
dTV(D1, D2) def=
1
2kD1 − D2k1 =
1
2 Xx∈[n]
D1(i) − D2(i) = max
S⊆[n]
(D1(S) − D2(S))
(1)
which takes value in [0, 1].
8
In this work, we focus on the setting of conditional access to the distribution, as introduced and
studied in [CFGM16, CRS15]. We reproduce below the corresponding definition of a conditional
oracle, henceforth referred to as COND.
Definition 2.1 (Conditional access model). Fix a distribution D over [n]. A COND oracle for D,
denoted CONDD, is defined as follows: the oracle takes as input a query set S ⊆ [n], chosen by the
algorithm, that has D(S) > 0. The oracle returns an element i ∈ S, where the probability that
element i is returned is DS (i) = D(i)/D(S), independently of all previous calls to the oracle.
Note that as described above the behavior of CONDD(S) is undefined if D(S) = 0, i. e., the set
S has zero probability under D. Various definitional choices could be made to deal with this. These
choices do not make a significant difference in most situations, as adaptive algorithms can include
in their next queries a sample previously obtained; while our lower bounds can be thought of as
putting exponentially small probability mass of elements outside the support. For this reason, and
for convenience, we shall hereafter assume, following Chakraborty et al., that the oracle returns in
this case a sample uniformly distributed in S. Furthermore, as in [CRS15, CFGM16] we do not
take the complexity of specifying the set S to the oracle into account, and indeed allow arbitrary
sets as queries.6
Finally, recall that a property P of distributions over [n] is a set consisting of all distributions
that have the property. The distance from D to a property P, denoted dTV(D,P), is then de-
fined as infD′∈P dTV(D,P). We use the standard definition of testing algorithms for properties of
distributions over [n], tailored for the setting of conditional access to an unknown distribution.
Definition 2.2 (Property tester). Let P be a property of distributions over [n]. A t-query COND
testing algorithm for P is a randomized algorithm T which takes as input n, ε ∈ (0, 1], as well as
access to CONDD. After making at most t(ε, n) calls to the oracle, T either returns ACCEPT or
REJECT, such that the following holds:
• if D ∈ P, T returns ACCEPT with probability at least 2/3;
• if dTV(D,P) ≥ ε, T returns REJECT with probability at least 2/3.
We observe that the above definitions can be straightforwardly extended to the more general
setting of pairs of distributions, where given independent access to two oracles CONDD1, CONDD2
the goal is to test whether (D1, D2) satisfies a property (now a set of pairs of distributions). This
will be the case in Section 3, where we will consider equivalence testing, that is the property
Peq ={ (D1, D2) : D1 = D2 }.
2.2 Adaptive Core Testers
In order to deal with adaptivity in our lower bounds, we will use ideas introduced by Chakraborty
et al. [CFGM16]. These ideas, for the case of label-invariant properties7 allow one to narrow down
5Recall that a non-adaptive tester is an algorithm whose queries do not depend on the answers obtained from
previous ones, but only on its internal randomness. Equivalently, it is a tester that can commit "upfront" to all the
queries it will make to the oracle.
6We further observe that, besides the general CONDD oracle which allows these arbitrary query sets, the authors
of [CRS15] introduce two weaker variants of the conditional model (the "pair-cond" PAIRCONDD and "interval-cond"
INTCONDD oracles) which restrict algorithms to "simple" queries.
7Recall that a property is label-invariant (or symmetric) if it is closed under relabeling of the elements of the
support. More precisely, a property of distributions (resp. pairs of distributions) P is label-invariant if for any
distribution D ∈ P (resp. (D1, D2) ∈ P) and permutation σ of [n], one has D ◦ σ ∈ P (resp. (D1 ◦ σ, D2 ◦ σ) ∈ P).
9
the range of possible testers and focus on a restricted class of such algorithms called adaptive core
testers. These core testers do not have access to the full information of the samples they draw,
but instead only get to see the relations (inclusions, equalities) between the queries they make and
the samples they get. Yet, Chakraborty et al. [CFGM16] show that any tester for a label-invariant
property can be converted into a core tester with same query complexity; thus, it is enough to
prove lower bounds against this -- seemingly -- weaker class of algorithms.
We here rephrase the definitions of a core tester and the view they have of the interaction with
the oracle (the configuration of the samples), tailored to our setting.
Definition 2.3 (Atoms and partitions). Given a family A = (A1, . . . , At) ⊆ [n]t, the atoms gener-
ated by A are the (at most) 2t distinct sets of the form Tt
r=1 Cr, where Cr ∈ {Ar, [n] \ Ar}. The
family of all atoms, denoted At(A), is the partition generated by A.
This definition essentially captures "all sets (besides the Ai) about which something can be
learned from querying the oracle on the sets of A." Now, given such a sequence of queries
A = (A1, . . . , At) and pairs of samples s = ((s(1)
t , we
would like to summarize "all the label-invariant information available to an algorithm that obtains
((s(1)
1 , s(2)
t )) upon querying A1, . . . , At for D1 and D2." This calls for the following
definition.
, s(2)
t )) ∈ A2
1 × ··· × A2
1 ), . . . , (s(1)
1 ), . . . , (s(1)
1 , s(2)
, s(2)
t
t
Definition 2.4 (t-configuration). Given A = (A1, . . . , At) and s = ((s(1)
t-configuration of s consists of the 6t2 bits indicating, for all 1 ≤ i, j ≤ t, whether
, s(2)
j
j ))1≤j≤t as above, the
• s(k)
• s(k)
i = s(ℓ)
j
, for k, ℓ ∈ {1, 2}; and
i ∈ Aj, for k ∈ {1, 2}.
(relations between samples)
(relations between samples and query sets)
In other terms, it summarizes which is the unique atom Si ∈ At(A) that contains s(k)
collisions between samples have been observed.
i
, and what
As aforementioned, the key idea is to argue that, without loss of generality, one can restrict
one's attention to algorithms that only have access to t-configurations, and generate their queries
in a specific (albeit adaptive) fashion.
Definition 2.5 (Core adaptive tester). A core adaptive distribution tester for pairs of distributions
is an algorithm T that acts as follows.
• In the i-th phase, based only on its own internal randomness and the configuration of the
i−1) -- whose labels
previous queries A1, . . . , Ai−1 and samples obtained (s(1)
it does not actually know, T provides:
1 ), . . . , (s(1)
i−1, s(2)
1 , s(2)
-- a number kA
i
(How
many fresh, not-already-seen elements of each particular atom A should be included in
the next query.)
for each A ∈ At(A1, . . . , Ai−1), between 0 and(cid:12)(cid:12)(cid:12)
i ⊆ {1, . . . , i − 1} (Which of the samples s(k)
i−1 will be included in
the next query. The labels of these samples are unknown, but are indexed by the index
of the query which returned them.)
j }1≤j≤i−1(cid:12)(cid:12)(cid:12)
A \ {s(1)
1 , . . . , s(k)
, K (2)
, s(2)
j
i
-- sets K (1)
• based on these specifications, the next query Ai is drawn (but not revealed to T ) by
10
-- drawing uniformly at random a set Λi in
n Λ ⊆ [n] \ {s(1)
j
, s(2)
j }1≤j≤i−1 : ∀A ∈ At(A1, . . . , Ai−1), Λ ∩ A = kA
i o .
(2)
That is, among all sets, containing only "fresh elements," whose intersection with each
atom contains as many elements as T requires.
-- adding the selected previous samples to this set:
Γi
def
=n s(1)
j
: j ∈ K (1)
i o ∪n s(2)
j
: j ∈ K (2)
i o ;
Ai
def
= Λi ∪ Γi .
(3)
This results in a set Ai, not fully known to T besides the samples it already got and decided
to query again; in which the labels of the fresh elements are unknown, but the proportions of
elements belonging to each atom are known.
i ∼ (D1)Ai and s(2)
• samples s(1)
i ∼ (D2)Ai are drawn (but not disclosed to T ). This defines
the i-configuration of A1, . . . , Ai and (s(1)
i ), which is revealed to T . Put
differently, the algorithm only learns (i) to which of the Aℓ the new sample belongs, and (ii)
if it is one of the previous samples, in which stage(s) and for which of D1, D2 it has already
seen it.
1 ), . . . , (s(1)
1 , s(2)
, s(2)
i
After t = t(ε, n) such stages, T returns either ACCEPT or REJECT, based only on the configuration
of A1, . . . , At and (s(1)
t ) (which is all the information it ever had access to).
1 ), . . . , (s(1)
1 , s(2)
, s(2)
t
Note that in particular, T does not know the labels of samples it got, nor the actual queries it
makes: it knows all about their sizes and sizes of their intersections, but not the actual "identity"
of the elements they contain.
2.3 On the use of Yao's Principle in our lower bounds
We recall Yao's Principle (e. g., see Chapter 2.2 of [MR95]), a technique which is ubiquitous in
the analysis of randomized algorithms. Consider a set S of instances of some problem: what this
principle states is that the worst-case expected cost of a randomized algorithm on instances in S
is lower-bounded by the expected cost of the best deterministic algorithm on an instance drawn
randomly from S.
As an example, we apply it in a standard way in Section 4: instead of considering a randomized
algorithm working on a fixed instance, we instead analyze a deterministic algorithm working on
a random instance. (We note that, importantly, the randomness in the samples returned by the
COND oracle is "external" to this argument, and these samples behave identically in an application
of Yao's Principle.)
On the other hand, our application in Section 3 is slightly different, due to our use of adaptive
core testers. Once again, we focus on deterministic algorithms working on random instances, and
the randomness in the samples is external and therefore unaffected by Yao's Principle. However, we
stress that the randomness in the choice of the set Λi is also external to the argument, and therefore
unaffected -- similar to the randomness in the samples, the algorithm has no control here. Another
way of thinking about this randomness is via another step in the distribution over instances: after
an instance (which is a pair of distributions) is randomly chosen, we permute the labels on the
elements of the distribution's domain uniformly at random. We note that since the property in
question is label-invariant, this does not affect its value. We can then use the model as stated in
Section 2.2 for ease of analysis, observing that this can be considered an application of the principle
of deferred decisions (as in Chapter 3.5 of [MR95]).
11
2.4 Chernoff bounds for Binomials and Hypergeometrics
We will make extensive use of Chernoff-style bounds in this work. Recall that the Binomial(n, p)
distribution describes the distribution of the number of successes when we run n independent
Bernoulli trials, each with success probability p.
Lemma 2.6 (Chernoff Bound for Binomials). Let X ∼ Binomial(n, p) and µ = E[X] = np. Then
∀δ ∈ (0, 1),
Pr[X − µ ≥ δµ ] ≤ 2 exp −
δ2µ
3 ! .
We will also need a similar Chernoff-style bound for the hypergeometric distribution. The
Hypergeometric(n, K, N ) distribution describes the distribution of the number of successes when
we draw n times without replacement from a population of size N , in which K objects have
the pertinent feature (and thus count as successes). Note that if the drawing were done with
replacement, and K/N = p, then this would be equivalent to Binomial(n, p). Sampling without
replacement introduces negative correlation between the probability of each draw being successful.
This type of negative correlation generally "helps" with concentration, allowing one to prove similar
concentration bounds (see, e. g., [Chv79, DR96], Theorem 1.17 of [AD11]).
Lemma 2.7 (Chernoff Bound for Hypergeometrics). Let X ∼ Hypergeometric(n, K, N ) and µ =
E[X] = nK/N . Then,
∀δ ∈ (0, 1),
Pr[X − µ ≥ δµ ] ≤ 2 exp −
δ2µ
3 ! .
3 A Lower Bound for Equivalence Testing
We prove our main lower bound on the sample complexity of testing equivalence between unknown
distributions. We construct two priors Y and N over pairs of distributions (D1, D2) over [n]. Y is a
distribution over pairs of distributions of the form (D, D), namely the case when the distributions
are identical. Similarly, N is a distribution over (D1, D2) with dTV(D1, D2) ≥ 1
4 . We then show
that no algorithm T making O(cid:0)√log log n(cid:1) queries to CONDD1, CONDD2 can distinguish between
a draw from Y and N with constant probability (over the choice of (D1, D2), the randomness in
the samples it obtains, and its internal randomness).
We describe the construction of Y and N in Section 3.1, and provide a detailed analysis
in Section 3.2.
3.1 Construction
We now summarize how a pair of distribution is constructed under Y and N . (Each specific step
will be described in more detail in the subsequent paragraphs.)
1. Effective Support
(a) Pick kb from the set {0, 1, . . . , 1
(b) Let b = 2kb and m def= b · n1/4.
2 log n} at random.
2. Buckets
12
(a) Choose ρ and r such that P2r
(b) Divide {1, . . . , m} into intervals B1, . . . , B2r with Bi = b · ρi.
i=1 ρi = n1/4.
3. Distributions
D1.
(a) For each i ∈ [2r], assign probability mass 1
(b) For each i ∈ [r] independently, pick πi to be a Bernoulli trial with Pr(πi = 0) = 1
2r uniformly over Bi to generate distribution
2 ; if
4r over B2i−1 and B2i, respectively, else 3
4r
πi = 0 then assign probability mass 1
and 1
4r , respectively. This generates a distribution D2.
4r and 3
4. Support relabeling
(a) Pick a permutation σ ∈ Sn of the total support n.
(b) Relabel the symbols of D1 and D2 according to σ.
5. Output: Generate (D1, D1) for Y, and (D1, D2) otherwise.
Dj(i)
B1B2 B3
B4
(. . . )
m
i
n
Figure 1: A no-instance (D1, D2) (before permutation).
We now describe the various steps of the construction in greater detail.
Effective support. Both D1 and D2, albeit distributions on [n], will have (common) sparse
support. The support size is taken to be m def= b · n1/4. Note that, from the above definition,
m is chosen uniformly at random from products of n1/4 with powers of 2, resulting in values in
[n1/4, n3/4].
In this step b will act as a random scaling factor. The objective of this random scaling is to
induce uncertainty in the algorithm's knowledge of the true support size of the distributions, and
to prevent it from leveraging this information to test equivalence. In fact one can verify that the
class of distributions induced for a single value of b, namely all distributions have the same value of
m, then one can distinguish the Y and N cases with only O(1) conditional queries. The test would
13
(roughly) go as follows. Since Bi is known, one can choose a random subset S of the domain which
(with high probability) has no intersection with Bi for i ≤ 2r − 2, a O(1) size intersection with
B2r−1, and a O(ρ) size intersection with B2r. Perform O(1) conditional queries over the set S, for
both distributions. Given these queries, we can then identify which elements of S belong to B2r−1
or B2r -- namely, those which occur at most once belong to B2r, and those which occur at least
twice belong to B2r−1. In a Y instance, then in both distributions, a 1/2 fraction of queries will
belong to B2r−1, whereas in a N instance, one distribution will have either a 1/4 or 3/4 fraction
of queries in B2r−1, allowing us to distinguish the two cases.
Buckets. Our construction is inspired by the lower bound of Canonne, Ron, and Servedio [CRS15,
Theorem 8] for the more restrictive PAIRCOND access model. We partition the support into 2r
consecutive intervals (henceforth referred to as buckets) B1, . . . , B2r, where the size of the i-th
i=1 bρi = bn1/4, i. e., the buckets fill
bucket is bρi. We note that r and ρ will be chosen such that P2r
the effective support.
Distributions. We output a pair of distributions (D1, D2). Each distribution that we construct
is uniform within any particular bucket Bi. In particular, the first distribution assigns the same
mass 1/2r to each bucket. Therefore, points within Bi have the same probability mass 1/2rbρi.
For the Y case, the second distribution is identical to the first. For the N case, we pair buckets in
r consecutive bucket-pairs Π1, . . . , Πr, with Πi = B2i−1 ∪ B2i. For the second distribution D2, we
consider the same buckets as D1, but repartition the mass 1/r within each Πi. More precisely, in
each pair, one of the buckets gets now total probability mass 1/4r while the other gets 3/4r (so
that the probability of every point is either decreased by a factor 1/2 or increased by 3/2). The
choice of which goes up and which goes down is done uniformly and independently at random for
each bucket-pair determined by the random choices of the πi.
Random relabeling. The final step of the construction randomly relabels the symbols, namely
is a random injective map from [m] to [n]. This is done to ensure that no information about the
individual symbol labels can be used by the algorithm for testing. For example, without this the
algorithm can consider a few symbols from the first bucket and distinguish the Y and N cases. As
mentioned in Section 2.3, for ease of analysis, the randomness in the choice of the permutation is,
in some sense, deferred to the randomness in the choice of Λi during the algorithm's execution.
Summary. A no-instance (D1, D2) is thus defined by the following parameters: the support size
m, the vector (π1, . . . , πr) ∈ {0, 1}r (which only impacts D2), and the final permutation σ of the
domain. A yes-instance (D1, D1) follows an identical process, however, ~π has no influence on the
final outcome. See Figure 1 for an illustration of such a (D1, D2) when σ is the identity permutation
and thus the distribution is supported over the first m natural numbers.
Values for ρ and r. By setting r = log n/(8 log ρ) + O(1), we have as desired P2r
i=1 Bi = m
and there is a factor (1 + o(1))n1/4 between the height of the first bucket B1 and the one of the
last, B2r. It remains to choose the parameter ρ itself; we shall take it to be 2√log n, resulting in
8√log n + O(1). (Note that for the sake of the exposition, we ignore technical details such
r = 1
as the rounding of parameters, e. g., bucket sizes; these can be easily taken care of at the price of
cumbersome case analyses, and do not bring much to the argument.)
14
3.2 Analysis
We now prove our main lower bound, by analyzing the behavior of core adaptive testers (as
per Definition 2.5) on the families Y and N from the previous section.
In Section 3.2.1, we ar-
gue that, with high probability, the sizes of the queries performed by the algorithm satisfy some
specific properties. Conditioned upon this event, in Section 3.2.2, we show that the algorithm will
get similar information from each query, whether it is running on a yes-instance or a no-instance.
Before moving to the heart of the argument, we state the following fact, straightforward from
the construction of our no-instances.
Fact 3.1. For any (D1, D2) drawn from N , one has dTV(D1, D2) = 1/4.
Moreover, as allowing more queries can only increase the probability of success, we hereafter focus
on a core adaptive tester that performs exactly q = 1
that it can only distinguish between yes and no-instances with probability o(1).
10√log log n (adaptive) queries; and will show
3.2.1 Banning "bad queries"
, K (2)
As mentioned in Section 3.1, the draw of a yes or no-instance involves a random scaling of the size
of the support of the distributions, meant to "blind" the testing algorithm. Recall that a testing
algorithm is specified by a decision tree, which at step i, specifies how many unseen elements from
each atom to include in the query ({kA
i }) and which previously seen elements to include in the query
(sets K (1)
, as defined in Section 2.2), where the algorithm's choice depends on the observed
configuration at that time. Note that, using Yao's Principle (as discussed in Section 2.3), these
choices are deterministic for a given configuration -- in particular, we can think of all {kA
i } and
K (1)
i values satisfy
with high probability some particular conditions with respect to the choice of distribution, where
the randomness is over the choice of the support size.
in the decision tree as being fixed. In this section, we show that all kA
i
i
, K (2)
i
i
First, we recall an observation from [CFGM16], though we modify it slightly to apply to con-
figurations on pairs of distributions and we apply a slightly tighter analysis. This essentially limits
the number of states an algorithm could be in by a function of how many queries it makes.
Proposition 3.2. The number of nodes in a decision tree corresponding to a q-sample algorithm
is at most 26q2+1.
Proof. As mentioned in Definition 2.4, an i-configuration can be described using 6i2 bits, resulting
in at most 26i2
i-configurations. Since each i-configuration leads us to some node on the i-th level
of the decision tree, the total number of nodes can be upper bounded by summing over the number
of i-configurations for i ranging from 0 to q, giving us the desired bound.
For the sake of the argument, we will introduce a few notions applying to the sizes of query sets:
namely, the notions of a number being small, large, or stable, and of a vector being incomparable.
Roughly speaking, a number is small if a uniformly random set of this size does not, in expectation,
hit the largest bucket B2r -- in other words, the set is likely to be disjoint from the support. On
the other hand, it is large if we expect such a set to intersect many bucket-pairs (i. e., a significant
fraction of the support).
The definition of stable numbers is slightly more quantitative: a number β is stable if a random
set of size β, for each bucket Bi, is either disjoint from Bi or has an intersection with Bi of size
close to the expected value. In the latter case, we say the set concentrates over Bi. Finally, a vector
of values (βj) is incomparable if the union of random sets S1, . . . , Sm of sizes β1, . . . , βm contains
15
(with high probability) an amount of mass D(cid:16)Sj Sj(cid:17) which is either much smaller or much larger
than the probability D(s) of any single element s.
We formalize these concepts in the definitions below. To motivate them, it will be useful to bear in
mind that, from the construction described in Section 3.1, the expected intersection of a uniform
random set of size β with a bucket Bi is of size βbρi/n; while the expected probability mass from
Bi it contains (under either D1 or D2) is β/2rn.
Definition 3.3. Let q be an integer, and let ϕ = Θ(q5/2). A number β is said to be small if
β < n
bρ2r ; it is large (with relation to some integer q) if β ≥ n
Note that the latter condition equivalently means that, in expectation, a set of large size will
intersect at least ϕ + 1 bucket-pairs (as it hits an expected 2ϕ + 1 buckets, since β B2r−2ϕ /n ≥ 1).
From the above definitions we get that, with high probability, a random set of any fixed size will
in expectation either hit many or no buckets.
bρ2r−2ϕ .
log n (cid:17).
Proposition 3.4. A number is either small or large with probability 1 − O(cid:16) ϕ log ρ
Proof. A number β is neither large nor small if ρ2ϕn
βρ2r . The ratio of the endpoints of
the interval is ρ2ϕ. Since b = 2kb, this implies that at most log ρ2ϕ = 2ϕ log ρ values of kb could
result in a fixed number falling in this range. As there are Θ(log n) values for kb, the proposition
follows.
βρ2r ≤ b ≤ n
The next definition characterizes the sizes of query sets for which the expected intersection with
any bucket is either close to 0 (less than 1/α, for some threshold α), or very big (more than α). (It
will be helpful to keep in mind that we will eventually use this definition with α = poly(q).)
A vector of numbers is said to be α-stable if all numbers it contains are α-stable.
Definition 3.5. A number β is said to be α-stable (for α ≥ 1) if, for each j ∈ [2r], β /∈ h n
Proposition 3.6. A number is α-stable with probability 1 − O(cid:16) r log α
log n (cid:17).
Proof. Fix some j ∈ [2r]. A number β does not satisfy the definition of α-stability for this j if
αβρj ≤ b ≤ nα
βρj . Since b = 2kb, this implies that at most log 2α values of kb could result in a fixed
number falling in this range. Noting that there are Θ(log n) values for kb and taking a union bound
over all 2r values for j, the proposition follows.
bρji.
αbρj , αn
n
The following definition characterizes the sizes of query sets which have a probability mass far
from the probability mass of any individual element. (For the sake of building intuition, the reader
may replace ν in the following by the parameter b of the distribution.)
Definition 3.7. A vector of numbers (β1, . . . , βℓ) is said to be (α, τ )-incomparable with respect to
ν (for τ ≥ 1) if the two following conditions hold.
• (β1, . . . , βℓ) is α-stable.
• Let ∆j be the minimum ∆ ∈ {0, . . . , 2r} such that βj νρ2r−∆
For all i ∈ [2r],
τ 2rνρi ,
n
1
τ
2rνρii.
1
2rn Pℓ
j=1 βj∆j 6∈h
≤ 1
α , or 2r if no such ∆ exists.
16
Recall from the definition of α-stability of a number that a random set of this size either has
essentially no intersection with a bucket or "concentrates over it" (i. e., with high probability, the
probability mass contained in the intersection with this bucket is very close to the expected value).
The above definition roughly captures the following. For any j, ∆j is the number of buckets that
will concentrate over a random set of size βj. The last condition asks that the total probability
mass from D1 (or D2) enclosed in the union of m random sets of size β1, . . . , βℓ be a multiplicative
factor of τ from the individual probability weight 1/2rbρi of a single element from any of the 2r
buckets.
Proposition 3.8. Given that a vector of numbers of length ℓ is α-stable, it is (α, q2)-incomparable
log n (cid:17).
with respect to b with probability at least 1 − O(cid:16) r log q
Proof. Fix any vector (β1, . . . , βℓ). By the definition above, for each value b such that (β1, . . . , βℓ)
is α-stable, we have
αρ2r
n ≤
ρ∆j
b
βj ·
< βj ·
αρ2r+1
n
,
j ∈ [ℓ]
or, equivalently,
log αβj
n
log ρ
+ 2r +
log b
log ρ ≤ ∆j <
log αβj
n
log ρ
+ 2r +
log b
log ρ
+ 1,
j ∈ [ℓ].
Writing λj
def=
log
αβj
n
log ρ + 2r for j ∈ [ℓ], we obtain that
ℓ
ℓ
Xj=1
Xj=1
βj∆jb = b
βj(λj + O(1)) +
Xj=1
j=1 βj(λj + O(1)) ≪ log b · Pℓ
j=1 βj. Then, for any fixed i ∈
j=1 βj∆jb /∈
[n/(200qρi), 200qn/ρi]. This is essentially, with the assumption equation (4) above, requiring
that
• If it is the case that log ρ ·Pℓ
[2r], to meet the second item of the definition of incomparability we need Pℓ
βj.
(4)
ℓ
b log b
log ρ
b log b /∈ "
n log ρ
j=1 βj
2q2ρiPℓ
,
2q2n log ρ
j=1 βj# .
ρiPℓ
Recalling that b log b = kb2kb, this means that O(log q/ log log q) values of kb are to be ruled
out. (Observe that this is the number of possible "bad values" for b without the condition
from the case distinction above; since we add an extra constraint on b, there are at most this
many values to avoid.)
• Conversely, if log ρ ·Pℓ
b /∈"
j=1 βj(λj + O(1)) ≫ log b ·Pℓ
ρiPℓ
2q2ρiPℓ
j=1 βj(λj + O(1))
n log ρ
,
ruling out this time O(log q) values for kb.
j=1 βj the requirement becomes
2q2n log ρ
j=1 βj(λj + O(1))# .
• Finally, the two terms are comparable only if log b = Θ(cid:16) log ρ·Pℓ
given that log b = kb, this rules out this time O(1) values for kb.
j=1 βj(λj+O(1))·(cid:16)Pℓ
j=1 βj(cid:17)−1(cid:17);
17
A union bound over the 2r possible values of i, and the fact that kb can take Θ(log n) values,
complete the proof.
We put these together to obtain the following lemma.
Lemma 3.9. With probability at least 1 − O(cid:18) 26q2 +q(r log α+ϕ log ρ)+26q2
for the decision tree corresponding to a q-query algorithm:
log n
(r log q)
(cid:19), the following holds
• the size of each atom is α-stable and either large or small;
• the size of each atom, after excluding elements we have previously observed,8 is α-stable and
either large or small;
• for each i, the vector (kA
i )A∈At(A1,...,Ai) is (α, q2)-incomparable (with respect to b).
Proof. From Proposition 3.2, there are at most 26q2+1 tree nodes, each of which contains one vector
i )A, and at most 2q atom sizes. The first point follows from Proposition 3.4 and Proposition 3.6
(kA
and applying the union bound over all 26q2+1 · 2 · 2q sizes, where we note the additional fac-
tor of 2 comes from either including or excluding the old elements. The latter point follows
from Proposition 3.8 and applying the union bound over all 26q2+1 nodes in the tree (each containing
a single kA
i vector).
3.2.2 Key lemma: bounding the variation distance between decision trees
In this section, we prove a key lemma on the variation distance between the distribution on leaves
of any decision tree, when given access to either an instance from Y or N . This lemma will in
turn directly yield Theorem 1.1. Hereafter, we set the parameters α (the threshold for stability),
ϕ (the parameter for smallness and largeness) and γ (an accuracy parameter for how well things
def
concentrate over their expected value) as follows.9 α
= 1/ϕ = q−5/2.
(Recall further that q = 1
def
= q5/2 and γ
def
= q7, ϕ
10√log log n.)
Lemma 3.10. Conditioned on the events of Lemma 3.9, consider the distribution over leaves of
any decision tree corresponding to a q-query adaptive algorithm when the algorithm is given a yes-
instance, and when it is given a no-instance. These two distributions have total variation distance
o(1).
Proof. This proof is by induction on 1 ≤ i ≤ q. We will have three inductive hypotheses,
E1(t), E2(t), and E3(t). Assuming all three hold for all t < i, we prove E1(i). Additionally
assuming E1(i), we prove E2(i) and E3(i).
Roughly, the first inductive hypothesis states that the query sets behave similarly to as if we
had picked a random set of that size. It also implies that whether or not we get an element we
have seen before is "obvious" based on past observances and the size of the query we perform. The
second states that we never observe two distinct elements from the same bucket-pair. The third
states that the next sample is distributed similarly in either a yes-instance or a no-instance. Note
that this distribution includes both features which our algorithm can observe (i. e., the atom which
8More precisely, we mean to say that for each i ≤ q, for every atom A defined by the partition of (A1, . . . , Ai),
the values kA
i and A \ {s(1)
i are α-stable and either large or small;
1 , s(2)
1 , . . . , s(1)
i−1, s(2)
i−1} − kA
9This choice of parameters is not completely arbitrary: combined with the setting of q, r and ρ, they ensure a
total bound o(1) on variation distance and probability of "bad events" as well as a (relative) simplicity and symmetry
in the relevant quantities.
18
the sample belongs to and if it collides with a previously seen sample), as well as those which it
can not (i. e., which bucket-pair the observed sample belongs to). It is necessary to show the latter,
since the bucket-pair a sample belongs to may determine the outcome of future queries.
More precisely, the three inductive hypotheses are as follows:
• E1(i): In either a yes-instance or a no-instance, the following occurs: For an atom S in the
partition generated by A1, . . . , Ai, let S′ = S \ {s(1)
i−1, s(2)
i−1}. For every such S′,
let ℓS′
α , or 0 if no such ℓ exists. We
claim that ℓS′ ∈ {0, . . . , 2r− ϕ− 2}∪{2r}, and say S′ is small if ℓS′ = 2r and large otherwise.
Additionally:
1 , s(2)
be the largest index ℓ ∈ {0, . . . , 2r} such that S′bρℓ
1 , . . . , s(1)
n ≤ 1
i
i
i
.
n
p1
p2
, S′ ∩ Bj = 0;
Furthermore, let p1 and p2 be the probability mass contained in Λi and Γi, respectively. Then
-- for j ≤ ℓS′
-- for j > ℓS′, S′ ∩ Bj lies in [1 − iγ, 1 + iγ] S′bρj
p1+p2 ≤ O(cid:16) 1
• E2(i): No two elements from the set {s(1)
• E3(i): Let T yes
q2(cid:17) (that is, either almost all the probability mass comes from
i } belong to the same bucket-pair.
be the random variable representing the atoms and bucket-pairs10 containing
i ), as well as which of the previous samples they intersect with, when the i-th query is
elements which we have not yet observed, or almost all of it comes from previously seen ones).
p1+p2 ≤ O(cid:16) 1
1 , . . . , s(1)
q2(cid:17) or
, s(2)
(s(1)
performed on a yes-instance, and define T no
1 , s(2)
, s(2)
i
i
, T no
O(cid:0)1/q2 + 1/ρ + γ + 1/ϕ(cid:1) = o(1).
similarly for no-instances. Then dTV(cid:0)T yes
i (cid:1) ≤
We will show that E1(i) holds with probability 1 − O(cid:0)2i exp(−2γ2α/3)(cid:1) and E2(i) holds with
probability 1 − O(i/ϕ). Let T yes be the random variable representing the q-configuration and the
bucket-pairs containing each of the observed samples in a yes-instance, and define T no similarly for
a no-instance. We note that this random variable determines which leaf of the decision tree we
reach, which E3(q) bounds. We can then take a union bound over all i ∈ [q] to upper bound the
probability that E1(i) and E2(i) do not hold, and use E3(i) and the coupling interpretation of
total variation distance to upper bound the probability that T yes
ever differ. Any of these
"failure" events happens with probability at most O(2q exp(− 2γ2α
q + q
ϕ ) = o(1)
(from our choice of α, γ, ϕ). This upper bounds the total variation distance between T yes and T no,
giving the desired result.
and T no
i
) + q2
ρ + qγ + q
ϕ + 1
3
i
We proceed with the inductive proofs of E1(i), E2(i), and E3(i), noting that the base cases
hold trivially for all three of these statements. Throughout this proof, recall that Λi is the set
of unseen support elements which we query, and Γi is the set of previously seen support elements
which we query.
Lemma 3.11. Assume that E1(t), E2(t), E3(t) hold for all 1 ≤ t ≤ i − 1. Then E1(i) holds with
probability at least 1 − O(cid:16)2i exp(cid:16)− 2γ2α
Proof. We start with the first part of E1(i) (the statement prior to "Furthermore"). Let S (and
the corresponding S′) be any atom as in E1(i). First, we note that ℓS′ ∈ {0, . . . , 2r − ϕ− 2}∪{2r}
since we are conditioning on Lemma 3.9: S′ is α-stable and either large or small, which enforces
this condition.
3 (cid:17)(cid:17) = 1 − 2i−Ω(q2).
10If a sample s(k)
i
does not belong to any bucket (if the corresponding i-th query did not intersect the support), it
is marked in T yes
i with a "dummy label" to indicate so.
19
Next, suppose S′ is contained in some other atom T generated by A1, . . . , Ai−1, and let T ′ =
T \{s(1)
. We argue about T ′∩ Bj
i−1}. Since S′ ≤ T ′, this implies that ℓT ′ ≤ ℓS′
i−1, s(2)
1 , s(2)
1 , . . . , s(1)
for three regimes of j:
• The first case is j ≤ ℓT ′
probability 1.
. By the inductive hypothesis, T ′ ∩ Bj = 0, so S′ ∩ Bj = 0 with
• The next case is ℓT ′ < j ≤ ℓS′. Recall from the definition of a core adaptive tester that S′ will
be chosen uniformly at random from all subsets of T ′ of appropriate size. By the inductive
hypothesis,
T ′ ∩ Bj
T ′
∈ [1 − (i − 1)γ, 1 + (i − 1)γ]
bρj
n
,
and therefore
E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ∈ [1 − (i − 1)γ, 1 + (i − 1)γ] S′ bρj
n
,
implying E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ≤
2
αρℓS′−j
;
where the inequality is by the definition of ℓS′
and using the fact that (i − 1)γ ≤ 1. Us-
ing a Chernoff bound for hypergeometric random variables (Lemma 2.7), and writing µ def=
E[S′ ∩ Bj] for conciseness,
1 − µ
µ (cid:19) µ(cid:21)
Pr(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12) ≥ 1(cid:3) = Pr(cid:20)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12) ≥ (cid:18)1 +
3µ !
(1 − µ)2
−j(cid:19) ,
αρℓS′
−j and (1− µ)2 ≥ 1/2 for n sufficiently
≤ exp −
≤ exp(cid:18)−
1
12
where the second inequality holds because µ ≤ 2/αρℓS′
large.
• The final case is j > ℓS′. As in the previous one,
E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ∈ [1 − (i − 1)γ, 1 + (i − 1)γ] S′ bρj
n
,
implying E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ≥
where the inequality is by the definition of ℓS′, α-stability, and using the fact that (i − 1)γ ≤
1/2. Using again a Chernoff bound for hypergeometric random variables (Lemma 2.7),
αρj−ℓS′
2
−1
;
Pr"(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12) − E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ≥ γS′ bρj
n
# ≤ Pr(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12) − E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3) ≥ γ2E(cid:2)(cid:12)(cid:12)S′ ∩ Bj(cid:12)(cid:12)(cid:3)(cid:3)
≤ 2 exp −
≤ 2 exp(cid:18)−
(2γ)2E[S′ ∩ Bj]
−1(cid:19) ,
3
γ2αρj−ℓS′
!
2
3
where the first inequality comes from 2(1− (i− 1)γ) ≥ 1, the second is from Chernoff bound,
and the third follows from E[S′ ∩ Bj] ≥ αρj−ℓS′
−1/2.
20
Since we wish to prove the statement for all buckets Bj simultaneously, we take a union bound over
all j. Recall that we want to bound the probability that S′ satisfies two conditions, for j ≤ ℓS′,
S′ ∩ Bj = 0; and for j > ℓS′, S′ ∩ Bj lies in [1 − iγ, 1 + iγ] S′bρj
. Using bounds from all three
of these regimes of j, and a union bound, the probability that S′ does not satisfy the conditions of
E1(i) is at most
n
Xj≤ℓT ′
0 + XℓT ′ <j≤ℓS′
exp(cid:18)−
1
12
αρℓS′
−j(cid:19) + Xj>ℓS′
2 exp(cid:18)−
2
3
γ2αρj−ℓS′
−1(cid:19) .
This probability is maximized when ℓS′ = ℓT ′ = 0, in which case it is
2r
Xj=1
2 exp(cid:18)−
2
3
γ2αρj−1(cid:19) ≤
∞
Xj=1
2 exp(cid:18)−
2
3
γ2αρj−1(cid:19) ≤ 3 exp(cid:18)−
2
3
γ2α(cid:19) .
Taking a union bound over at most 2i sets gives us the desired probability bound.
Finally, we prove the remainder of E1(i) (the statement following "Furthermore"); this will
follow from the definition of incomparability (Definition 3.7).
• First, we focus on Γi. Suppose that Γi contains at least one element with positive probability
mass (if not, the statement trivially holds). Let p′2 be the probability mass of the heaviest
element in Γi. Since our inductive hypothesis implies that Γi has no elements in the same
bucket pair, the maximum possible value for p2 is
p2 ≤ p′2 +
3p′2
ρ
+
3p′2
ρ3 + ··· ≤ p′2 +
3p′2
ρ
≤ (1 + o(1))p′2
∞
Xk=0
1
ρ2k = 1 +
3
ρ
ρ2
ρ2 − 1! p′2
Therefore, p2 ∈ [p′2, (1 + o(1))p′2]. Supposing this heaviest element belongs to bucket j, we
can say that p2 ∈ [ 1
2 , (1 + o(1)) 3
2 ]
2rbρj .
1
• Next, we focus on Λi. Consider some atom A, from which we selected kA elements which
have not been previously observed: call the set of these elements A′.
In the first part
of this proof, we showed that for each bucket Bk, either A′ ∩ Bk = 0 or A′ ∩ Bk ∈
[1 − iγ, 1 + iγ]A′ bρk/n. In the latter case, noting that iγ ≤ 1
2 and that the probability of
4rbρk , the probability mass contained by A′ ∩ Bk
an individual element in Bk is within [1, 3]
belongs to [1, 9]A′8rn . Recalling the definition of ∆A as stated in Definition 3.7, as shown earlier
in this proof, this non-empty intersection happens for exactly ∆A buckets. Therefore, the
4i 1
total probability mass in Λi is in the interval h 1
4 , 9
2rn PA∈At(A1,...,Ai) kA
Recall that we are conditioning on Lemma 3.9 which states that the vector (kA
i )A∈At(A1,...,Ai) is
(α, q2)-incomparable with respect to b. Applying this definition to the bounds just obtained on the
probability masses in Λi and Γi gives Lemma 3.11.
i ∆A.
1
Lemma 3.12. Assume that E1(t), E2(t), E3(t) hold for all 1 ≤ t ≤ i − 1, and additionally E1(i)
holds. Then E2(i) holds with probability at least 1 − O(cid:16) i
ϕ(cid:17).
Proof. We focus on s(1)
small atom intersects any of the buckets, so let us condition on the fact that s(1)
i ∈ Γi, the conclusion is trivial, so suppose s(1)
i ∈ Λi. From E1(i), no
belongs to some
. If s(1)
i
i
21
i
large atom S. Since we want s(1)
to fall in a distinct bucket-pair from 2(i − 1) + 1 other samples,
there are at most 2i − 1 bucket-pairs which s(1)
should not land in. Using E1(i), the maximum
probability mass contained in the intersection of these bucket-pairs and S is (1 + iγ)(2i− 1)S/rn.
Similarly, using the definition of a large atom, the minimum probability mass contained in S is
(1 − iγ)ϕS/rn. Taking the ratio of these two terms gives an upper bound on the probability of
breaking this invariant, conditioned on landing in S, as O(i/ϕ), where we note that 1+iγ
1−iγ = O(1).
Since the choice of which large atom was arbitrary, we can remove the conditioning. Taking a union
bound for s(1)
gives the result.
and s(2)
i
i
i
Lemma 3.13. Assume that E1(t), E2(t), E3(t) hold for all 1 ≤ t ≤ i − 1, and additionally E1(i)
holds. Then E3(i) holds.
Proof. We fix some setting of the intersection history, i. e., the configuration and the bucket-pairs
the past elements belong to, and show that the results of the next query will behave similarly,
whether the instance is a yes-instance or a no-instance. We note that, since we are assuming the
inductive hypotheses hold, certain settings which violate these hypotheses are not allowed. We also
note that s(1)
for the remainder
of this proof.
First, we condition that, based on the setting of the past history, si will either come from Λi or Γi
is distributed identically in both instances, so we focus on si
def= s(2)
i
i
-- this event happens with probability 1 − O(cid:0)1/q2(cid:1).
probability 1−O(cid:16) 1
q2(cid:17), or Γi with probability 1−O(cid:16) 1
Proposition 3.14. In either a yes-instance or a no-instance, si will either come from Λi with
based on the fixed configuration and choice for the bucket-pairs of previously seen elements.
q2(cid:17), where the choice of which one is deterministic
Proof. This is simply a rephrasing of the portion of E1(i) following "Furthermore."
We now try to bound the total variation distance between T yes
conditioning on this
In the case when it does not hold, we trivially bound the total variation distance by 1,
and T no
i
i
event.
incurring a cost of O(cid:0)1/q2(cid:1) to the total variation distance between the unconditioned variables.
Since our target was for this quantity was O(cid:0)1/q2 + 1/ρ + γ + 1/ϕ(cid:1), it remains to show, in the
conditioned space, that the total variation distance in either case is at most O(1/ρ + γ + 1/ϕ) =
O(1/q5/2). We break this into two cases, the first being when s comes from Γi. In this case, we
incur a cost in total variation distance which is O(1/ρ):
Proposition 3.15. In either a yes-instance or a no-instance, condition that si comes from Γi.
Then one of the following holds:
• Γi ∩ Bj = 0 for all j ∈ [2r], in which case si is distributed uniformly at random from the
elements of Γi;
• or Γi∩Bj 6= 0 for some j ∈ [2r], in which case si will be equal to some s ∈ Γi with probability
1 − O(1/ρ), where the choice of s is deterministic based on the fixed configuration and choice
for the bucket-pairs of previously seen elements.
Proof. The former case follows from the definition of the sampling model. For the latter case, let
p be the probability mass of the heaviest element in Γi. Since our inductive hypothesis implies
that Γi has no elements in the same bucket-pair, the maximum possible value for the rest of the
elements is
3p
ρ
+
3p
ρ3 +
3p
ρ5 + ··· ≤
3p
ρ
1
ρ2k =
3p
ρ
ρ2
ρ2 − 1
= O(cid:18) p
ρ(cid:19).
∞
Xk=0
22
Since the ratio of this value and p is O(1/ρ), with probability 1 − O(1/ρ) the sample returned is
the heaviest element in Γi.
Finally, we examine the case when s comes from Λi:
Proposition 3.16. Condition that si comes from Λi. Then either:
• Λi ∩ Bj = 0 for all j ∈ [2r], in which case dTV(T yes
• or Λi ∩ Bj 6= 0 for some j ∈ [2r], in which case dTV(T yes
, T no
i
i
i ) = 0;
, T no
i ) ≤ O(cid:16)γ + 1
q5/2(cid:17)
ϕ(cid:17) = O(cid:16) 1
Proof. The former case follows from the definition of the sampling model -- since Λi does not
intersect any of the buckets, the sample will be labeled as such. Furthermore, the sample returned
will be drawn uniformly at random from Λi, and the probability of each atom will be proportional
to the cardinality of its intersection with Λi, in both the yes and the no-instances.
We next turn to the latter case. Let X be the event that, if the intersection of Λi and some atom
A has a non-empty intersection with an odd number of buckets, then si does not come from the
unpaired bucket. Note that E1(i) and the definition of a large atom imply that an unpaired bucket
can only occur if the atom intersects at least ϕ bucket-pairs: conditioned on the sample coming
from a particular atom, the probability that it comes from the unpaired bucket is O(1/ϕ). Since
the choice of A was arbitrary, we may remove the conditioning, and note that Pr(X ) = 1− O(1/ϕ).
Since
dTV(T yes
i
, T no
i ) ≤ dTV(T yes
≤ dTV(T yes
i
i
i
, T no
, T no
i
X ) Pr(X ) + dTV(T yes
X ) + O(1/ϕ),
i
, T no
i
¯X ) Pr( ¯X )
(5)
it remains to show that dTV(T yes
i
, T no
i
X ) ≤ O(γ).
First, we focus on the distribution over atoms, conditioned on X . Let N A be the number of
bucket-pairs with which A intersects both buckets, i. e., conditioned on X , the sample could come
from 2N A buckets, and let N def= PA∈At(A1,...,Ai) N A. By E1(i), the maximum amount of probability
mass that can be assigned to atom A is (1+γ)SN A/rn
(1+γ)SN/rn , so the
total variation distance in the distribution incurred by this atom is at most O(γN A/N ). Summing
over all atoms, we get the desired result of O(γ).
(1−γ)SN/rn , and the minimum is (1−γ)SN A/rn
Finally, we bound the distance on the distribution over bucket-pairs, again conditioned on X .
By E1(i) only large atoms will contain non-zero probability mass, so condition on the sample
coming from some large atom A. Let N A be the number of bucket-pairs with which A intersects
both buckets, i. e., conditioned on X , the sample could come from 2N A buckets. Using E1(i),
the maximum amount of probability mass that can be assigned to any intersecting bucket-pair is
(1 + γ)Arn ((1 − γ)Arn N A)−1, and the minimum is (1 − γ)Arn ((1 + γ)Arn N A)−1, so the total variation
distance in the distribution incurred by this bucket-pair is at most O(γ/N A). Summing this differ-
ence over all N A bucket-pairs, we get
1−γ2 = O(γ). Since the choice of large atom A was arbitrary,
we can remove the conditioning on the choice of atom. The statement follows by applying the
union bound on the distribution over bucket-pairs and the distribution over atoms. This concludes
the proof of Proposition 3.16.
2γ
We note that in both cases, the cost in total variation distance which is incurred is O( 1
ρ + γ + 1
ϕ ),
which implies E3(i) -- proving Lemma 3.13 .
This concludes the proof of Lemma 3.10.
23
With Lemma 3.10 in hand, the proof of the main theorem is straightforward.
Proof of Theorem 1.1. Conditioned on Lemma 3.9, Lemma 3.10 implies that the distribution over
the leaves in a yes-instance vs. a no-instance is o(1). Since an algorithm's choice to accept or
reject depends deterministically on which leaf is reached, this bounds the difference between the
conditional probability of reaching a leaf which accepts. Since Lemma 3.9 occurs with probability
1 − o(1), the difference between the unconditional probabilities is also o(1).
4 Lower Bounds for Non-Adaptive Algorithms
In this section, we prove our lower bounds for non-adaptive support-size estimation and uniformity
testing, restated here.
Theorem 1.2 (Non-Adaptive Support-Size Estimation). Any non-adaptive algorithm which, given
COND access to an unknown distribution D on [n], estimates the size of its support up to a factor
γ ≥ √2 must have query complexity Ω(cid:16) log n
log2 γ(cid:17).
Theorem 1.3 (Non-Adaptive Uniformity Testing). Any non-adaptive algorithm which, given COND
access to an unknown distribution D on [n] and parameter ε ∈ (0, 1/4), distinguishes with prob-
ability at least 2/3 between (a) D = U[n] and (b) dTV(D,U[n]) ≥ ε, must have query complexity
Ω((log n)/ε).
These two theorems follow from the same argument, which we outline below before turning to the
proof itself. (Note that we will, in this section, establish Theorem 1.3 for constant ε, i. e., ε = 1/4;
before explaining in Section 4.1 how to derive the general statement as a corollary.)
Structure of the proof. By Yao's Minimax Principle, we consider deterministic tests and study
their performance over random distributions, chosen to be uniform over a random subset of carefully
picked size. The proof of Theorem 1.2 then proceeds in 3 steps: in Lemma 4.2, we first argue that
all queries made by the deterministic tester will (with high probability over the choice of the
support size s) behave very "nicely" with regard to s, i. e., not be concentrated around it. Then, we
condition on this to bound the total variation distance between the sequence of samples obtained in
the two cases we "oppose," a random distribution from a family D1 and the corresponding one from
a family D2. In Lemma 4.4 we show that the part of total variation distance due to samples from
the small queries is zero, except with probability o(1) over the choice of s. Similarly, Lemma 4.4
allows us to say (comparing both cases to a third "reference" case, a honest-to-goodness uniform
distribution over the whole domain, and applying a triangle inequality) that the remaining part of
the total variation distance due to samples from the big queries is zero as well, except again with
probability o(1). Combining these three lets us to conclude by properties of the total variation
distance, as (since the queries are non-adaptive) the distribution over the sequence of samples is
a product distribution. (Moreover, applying Lemma 4.4 as a stand-alone enables us, with little
additional work, to obtain Theorem 1.3, as our argument in particular implies distributions from
D1 are hard to distinguish from uniform.)
The families D1 and D2. Fix γ ≥ √2 as in Theorem 1.2; writing β def= γ2, we define the set
(6)
=(cid:26) βkn1/4 : 0 ≤ k ≤
S def
log n
2 log β (cid:27) = {n1/4, βn1/4, β2n1/4, . . . , , n3/4}
A no-instance (D1, D2) ∈ D1 × D2 is then obtained by the following process:
24
• Draw s uniformly at random from S.
• Pick a uniformly random set S1 ⊆ [n] of size s, and set D1 to be uniform on S1.
• Pick a uniformly random set S2 ⊆ [n] of size βs, and set D2 to be uniform on S2.
(Similarly, a yes-instance is obtained by first drawing a no-instance (D1, D2), and discarding D2 to
keep only (D1, D1) ∈ D1 × D1.)
We will argue that no algorithm can distinguish with high probability between the cases
(D1, D2) ∼ D1 × D2 and (D1, D2) ∼ D1 × D1, by showing that in both cases D1 and D2 gen-
erate transcripts indistinguishable from those the uniform distribution U[n] would yield. This will
imply Theorem 1.2, as any algorithm to estimate the support within a multiplicative γ would imply
a distinguisher between instances of the form (D1, D1) and (D1, D2) (indeed, the support sizes of
D1 and D2 differ by a factor β = γ2). (As for Theorem 1.3, observe that any distribution D1 ∈ D1
has constant distance from the uniform distribution on [n], so that a uniformity tester must be able
to tell D1 apart from U[n].)
80000
log n
log2 β
Small and big query sets. Let T be any deterministic non-adaptive algorithm making qT ≤
q = 1
queries. Without loss of generality, we can assume T makes exactly q queries, and
denote them by A1, . . . , Aq ⊆ [n]. Moreover, we let ai = Ai, and (again without loss of generality)
write a1 ≥ ··· ≥ aq.
As a preliminary observation, note that for any A ⊂ [n] and 0 ≤ s ≤ n we have
ES S ∩ A = A s
n
where the expectation is over a uniform choice of S among all(cid:0)n
will lead us to divide the query sets Ai in two groups, depending on the expected size of their
intersection with the (random) support.
s(cid:1) subsets of size s. This observation
With this in mind, the following definition will be crucial to our proof. Intuitively, it captures
the distribution of sizes of intersection of various query sets with the randomly chosen set S.
Definition 4.1. Let s ≥ 1, and A = {a1, . . . , aq} be any set of q integers. For a real number t > 0,
define
(7)
Ct(s)
def
= (cid:12)(cid:12)(cid:12)(cid:12)
(cid:26) i ∈ [q] :
ais
n ∈ (cid:16)β−t, βt(cid:17)(cid:27)(cid:12)(cid:12)(cid:12)(cid:12)
to be the number of t-hit points of A (for s).
The next result will be crucial to prove our lower bounds:
it roughly states that if we consider
the ai and scale them by the random quantity s/n, then the distribution of the random variable
generated has an exponential tail with respect to t.
Lemma 4.2 (Hitting Lemma). Fix A as in the previous definition. If s is drawn uniformly at
random from S, then with probability at least 99/100.
Ct(s)
t
<
2
100
.
sup
t>0
(8)
Proof. Without loss of generality, assume that the ai are in decreasing order. We will work in
the logarithmic domain: a number ai contributes to Ct(s) if and only if log s ∈ [log(n/ai) −
Indeed, we can restate the lemma in an additive form. Let A =
t log β, log(n/ai) + t log β].
25
{α1, . . . , αm} be any set of numbers in [0, log n]. These are defined as transformations of ai's:
αi , log(n/ai). In the additive restating, x will play the role of log s, or equivalently, s = 2x. For
a point x ∈ [0, log n], let ℓx
j denote the distance of x from the jth point to its left and right,
respectively, from the set A. More precisely, if αγ ≤ x ≤ αγ+1, ℓx
, x − αγ+1−j. If we consider
j / log β is the least value of t such that Ct(2x) = j. Therefore, if
only points to the left of x, ℓx
tx
j
, 1
Observe that Ct(2x) is a piecewise-constant function which is monotone non-decreasing in t.
jo, then we are guaranteed that j ≤ Ctx
log β minnℓx
(2x) ≤ 2j.
j and rx
j , rx
j
j
Therefore, the supremum of Ct(2x)
t
is attained at one of these discontinuities:
Ct(2x)
sup
t>0
t
=
{t : ∀δ>0,Ct−δ (2x)<Ct(2x)}
We can in turn upper bound this by looking at the set of all tx
discontinuous points:
However, we note that in this case, Crx
for instance, suppose that ℓx
j < rx
max
Ct(2x)
.
t
j . Note that this may ignore some
j / log β will not be considered.
j , then rx
j / log β(2x)
Crx
j / log β ≤
rx
Therefore,
j / log β(2x) = 2Ctx
j
(2x):
j / log β(2x) ≤ 2Cℓx
j / log β(2x)
j / log β(2x)
2Cℓx
Crx
j / log β ≤
ℓx
ℓx
j / log β
=
(2x)
2Ctx
j
tx
j
.
Ct(2x)
t
sup
t>0
≤ max
{tx
j : j∈[q]}
(2x)
2Ctx
j
tx
j
≤ max
{tx
j : j∈[q]}
4j
tx
j
= max
{tx
j : j∈[q]}
4j log β
min{ℓx
j , rx
j }
We would like to upper bound this term by 2/100. Equivalently, we satisfy the lemma conditions if
min
j
j (cid:26) tx
j (cid:27) ≥ 200 log β.
For a constant c > 0, let Sc be the set of all points x (where recall s = 2x is selected according
to the distribution S) such that violate this inequality for c:
min
j
j (cid:26) tx
j (cid:27) ≤ c.
We would like to upper bound the probability that a randomly selected x violates this inequality for
c = 200 log β by 1/100. Equivalently, since x is selected uniformly at random from log n
2 log β different
values, we would like to upper bound the size of S200 log β by log n/200 log β. We do this with the
following claim.
Claim 4.3. Sc ≤ 2cq.
Substituting in c = 200 log β and q = log n/80000 log2 β will give the desired result.
j /j < c. Let Si
Proof of Claim 4.3. We consider the set of points in Sc,ℓ ⊂ Sc that satisfy ℓx
j /j < c for some j,
and show that their measure is at most cq. An identical bound holds for the set of points of Sc for
c,ℓ ⊂ Sc,ℓ be the set of points in Sc,ℓ that satisfy minjn tx
j o < c with respect
which rx
to the set {α1, . . . , αi}. We will show by induction that Si
c,ℓ = [α1, α1 + c]. Suppose by induction that Si
c,ℓ < ci. Let xi be
c,ℓ. Then it is clear that xi > αi, in fact xi ≥ αi + c. Furthermore,
For the first point α1, the set S1
the right-most point in the set Si
c,ℓ < ci.
j
26
j /j = c for some j. Moreover, we claim that [αi, xi] ∈ Si
either xi = log n, or ℓxi
c,ℓ. Indeed, for the
same j that ℓxi
j /j < c, all points in [αi, xi] satisfy the condition. If xi = log n, then the result holds
trivially. We therefore consider the point αi+1 and prove the inductive step for xi < log n. There
are two cases:
If αi+1 ≥ xi: In this case, Si+1
c,ℓ = Si
c,ℓ∪[αi+1, xi+1]. We have to show that xi+1 ≤ αi+1 +c. Suppose
to the contrary that xi+1 > αi+1 + c ≥ xi + c. Then there is a point αh for h ≤ i, such that
i+2−h < c, and then αi+1+c−αh
xi+1−αh
i+2−h < c, so that
however, this implies that αi+1 ∈ Si
c,ℓ, contradicting the assumption of this case.
αi+1 − αh
i + 1 − h
< c,
If αi+1 < xi: In this case, Si+1
on the contrary that xi+1 > xi + c > αi+1 + c. Let h be the index such that xi+1−αh
therefore, xi+c−αh
c,ℓ ∪ [xi, xi+1]. We have to show that xi+1 ≤ xi + c. Suppose
i+2−h < c, and
c,ℓ = Si
i+2−h < c, implying that
xi − αh
i + 1 − h
contradicting that xi is the rightmost point of Si
< c,
c,ℓ.
This concludes the proof of the hitting lemma.
We proceed to show how to use this lemma to bound the contribution of various types of queries
to the distinguishability of D1 and D2. In particular, we will apply Lemma 4.2 to the set of query
sizes {a1, . . . , aq}.
Recall that the ai are non-increasing. If aq′s/n ≤ 1 let q′ def= q + 1, otherwise define q′ as the largest
integer such that aq′s/n > 1. We partition the queries made by T in two: A1, . . . , Aq′ are said to
be big, while Aq′+1, . . . , Aq are small queries.
Lemma 4.4. With probability at least 1 − 2−10, a random distribution from D1 or from D2 does
not intersect with any small query.
Proof. Let s be the random size drawn for the definition of the instances. We first claim that
E[(cid:12)(cid:12)Aq′+j ∩ S(cid:12)(cid:12)] ≤ β−50j for all j ≥ 1, where the expectation is over the uniform choice of set S1 for
D1. Indeed, by contradiction suppose there is a j ≥ 1 such that E[(cid:12)(cid:12)Aq′+j ∩ S(cid:12)(cid:12)] =
By definition of q′, for 1 ≤ i ≤ j,
n > β−50j.
aq′+is
aq′+js
1 ≥
n
> β−50j.
Therefore, the queries Aq′, Aq′+1, . . . , Aq′+j contribute to C50j, and we obtain C50j
contradicting Lemma 4.2. Thus, the expected intersection can be bounded as follows:
50j ≥ j
E[(cid:12)(cid:12)(Aq′+1 ∪ Aq′+2 ··· ∪ Aq) ∩ S(cid:12)(cid:12)] ≤ E[(cid:12)(cid:12)Aq′+1 ∩ S(cid:12)(cid:12)] + E[(cid:12)(cid:12)Aq′+2 ∩ S(cid:12)(cid:12)] + ··· + E[Aq ∩ S]
50j = 2
100 ,
≤ β−50 + β−100 + . . .
≤ 2−12,
since β ≥ 2. From this, we obtain the result holds for D1 by Markov's inequality. The same applies
to D2 with probability of intersection at most 2−10, proving the lemma.
27
We now turn our attention to the sets with large intersections. We will show that under D1 and
D2, the results of querying the sets A1, . . . Aq′ are indistinguishable from simply picking elements
uniformly from the sets A1, . . . , Aq′. More precisely, we establish the following.
Lemma 4.5. Let η∗ = 2−10 and ηs = 1/100; and fix ℓ ∈ {1, 2}. At least an 1 − ηs fraction of
elements s1, . . . , sq′ ∈ A1 × A2, . . . , Aq′ satisfy
,
(9)
Pr
ℓ (cid:2) (s1, . . . , sq′)(cid:3) ∈ [1 − 5η∗, 1 + 5η∗] ·
1
A1 . . .(cid:12)(cid:12)Aq′(cid:12)(cid:12)
under CONDDℓ.
where Prℓ(cid:2) (s1, . . . , sq′)(cid:3) denotes the probability that s1 . . . sq′ are the results of the queries A1, . . . , Aq′
As this holds for most distributions in both D1 and D2, this implies the queries are indistin-
guishable from the product distribution over A1×A2, . . . , Aq′ (which is the one induced by the same
queries on the uniform distribution over [n]) in either case, with probability at least 1 − ηs − 5η∗.
Proof of Lemma 4.5. From standard Chernoff-like concentration bounds for hypergeometric ran-
dom variables (Lemma 2.7), we obtain the claim below.
Claim 4.6. Suppose A is a set of size a, and S is a uniformly chosen random set of size s. Then, for
n (cid:3) < e−η2· as
all η ∈ (0, 1], we have Pr(cid:2)A ∩ S > (1 + η) as
3n .
We use this to show that indeed all the Ai ∩ S concentrate around their expected values for
1 ≤ j ≤ q′. First note that, as a consequence of Lemma 4.2, it is the case that these expected values
satisfy aq′−js/n ≥ β50(j+1) for every 0 ≤ j ≤ q′− 1 (with probability at least 99/100). Conditioning
on this, we first invoke Claim 4.6 on Aj with η = 3 · β20(j+1), and then apply a union bound to
obtain
3n and Pr(cid:2)A ∩ S < (1 − η) as
n (cid:3) < e−η2· as
Pr(cid:20)∃j ∈ [q′] s.t.
Aj ∩ S /∈h1 − 4 · β−20(j+1), 1 + 4 · β−20(j+1)i ·
ajs
n (cid:21) < e−β10
(10)
i. e., with high probability all intersections simultaneously concentrate around their expected values.
Note that since s is at most n3/4, each Ai under consideration has size at least nβ50/n3/4 > n1/4.
Therefore, the probability that a random selection of elements from A1, . . . , Aq′ exhibits no collision
is at least
q′
Yi=1
Ai − q′
Ai ≥ (cid:18)1 −
q′
n1/4(cid:19)q′
≥ 1 −
(q′)2
n1/4 > 1 −
log2 n
n1/4 .
We henceforth condition on this event.
Let N = (cid:0)n
number of such sets for which equation (10) holds. Let sq′
(s1, . . . , sq′) ∈ A1 × ··· × Aq′, let N (sq′
and let N0(sq′
1 ) of them satisfy equation (10).
1 ) = (cid:0)n−q′
s(cid:1) be the number of outcomes for the set S. We write N0 ≥ N (1 − e−β10
) for the
1 denote s1 . . . sq′. For a set of distinct
1 ,
s−q′(cid:1) be the number of sets of size s that contain sq′
, for a randomly chosen sq′
1 we have
By Markov's inequality, with probability at least 1 − e−β9
1 )/N (sq′
N0(sq′
Prh sq′
1 i ≥
. For any such sq′
1 ,
1
)
q′
1 ) > 1 − e2−β9
N0(sq′
1 )
Yi=1
N ·
≥ (1 − 6 · β−19)
Ai ∩ S ≥ (1 − e2−β9
Yi=1
1
ai
q′
,
28
s(s − 1) . . . (s − q′ + 1)
n(n − 1) . . . (n − q′ + 1) · (1 − 4 · β−19)
q′
Yi=1
n
ais
for large n and as S > n1/4. Since the sum of probabilities of elements is at most 1, the other side
of the inequality in Lemma 4.5 follows.
Proof of Theorem 1.2 and Theorem 1.3. Let T1 (resp. T2, TU ) be the distribution over transcripts
generated by the queries A1, . . . , Aq when given conditional access to the distribution D1 from a no-
instance (resp. D2, resp. uniform U[n]); that is, a distribution over q-tuples in A1 × ··· × Aq. Since
the queries were non-adaptive, we can break T1 (and similarly for T2, TU ) in T big
according
to q′, and use Lemma 4.5 and Lemma 4.4 separately to obtain dTV(T1, T2) ≤ ηs + η∗ + 2−10 < 1/50
and dTV(T1, TU ) ≤ ηs + η∗ + 2−10 < 1/50 (for the latter, recalling that queries that do not intersect
the support receive samples exactly uniformly distributed in the query set) -- thus establishing both
theorems.
1 × T small
1
4.1 On the dependence on ε in Theorem 1.3
We remark that Theorem 1.3, by establishing a lower bound of Ω(log n) queries for non-adaptive
testing of uniformity with constant distance parameter 1/4, immediately implies, by a standard
argument, an Ω((log n)/ε) lower bound for distance parameter ε ∈ (0, 1/4). In more detail, this is
a consequence of the following reduction: any q(n, ε)-query non-adaptive tester for uniformity T
can be used, given conditional access to some distribution D on [n], on the mixture distribution
Dε
def= 4εD + (1 − 4ε)U[n] ,
(11)
for which a conditional oracle can be easily simulated given a conditional oracle for D. Moreover,
answering q(n, ε) to Dε can be done with an expected 4εq(n, ε) conditional queries to D. As it
is immediate to see that dTV(Dε,U[n]) = 4ε dTV(D,U[n]), we get that T can be used to obtain
a tester for non-adaptive testing of uniformity with constant distance parameter 1/4, with query
complexity O(εq(n, ε)) for every ε < 1/4 (converting the expected query complexity to a worst-
case one is straightforward via Markov's inequality followed by success probability amplification
by a constant number of repetitions). Therefore, the lower bound of Theorem 1.3 implies that
q(n, ε) = Ω((log n)/ε), as claimed. It is also worth noting that the above argument does not yield
an analogue statement for support-size estimation via Theorem 1.2. Indeed, mixing the distribution
D with the uniform distribution does not preserve the support size in that case (nor the guarantee
that every point of the support has probability mass at least τ /n).
5 An Upper Bound for Support-Size Estimation
In this section, we prove our upper bound for constant-factor support-size estimation, reproduced
below.
Theorem 1.4 (Adaptive Support-Size Estimation). Let τ > 0 be any constant. There exists an
adaptive algorithm which, given COND access to an unknown distribution D on [n] (guaranteed
to have probability mass at least τ /n on every element of its support) and accuracy parameter
ε ∈ (0, 1), makes O(cid:0)(log log n)/ε3(cid:1) queries to the oracle11 and returns a value ω such that the
following holds. With probability at least 2/3, ω ∈ [ 1
1+ε · ω, (1 + ε) · ω], where ω = supp(D).
Before describing and analyzing our algorithm, we shall need the following results, that we
will use as subroutines: the first one will help us detecting when the support is already dense.
11We remark that the constant in the O depends polynomially on 1/τ .
29
The second, assuming the support is sparse enough, will enable us to find an element with zero
probability mass, which can afterwards be used as a "reference" to verify whether any given element
is inside or outside the support. Finally, the last one will use such a reference point to check whether
a candidate support size σ is smaller or significantly bigger than the actual support size.
Lemma 5.1. Given τ > 0 and COND access to a distribution D such that each support element
has probability at least τ /n, as well as parameters ε ∈ (0, 1/2), δ ∈ (0, 1), there exists an algorithm
TestSmallSupport (Algorithm 2) that makes O(cid:0)1/(τ ε2) + 1/τ 2(cid:1) · log(1/δ) queries to the oracle,
and satisfies the following. (i) If supp(D) ≥ (1 − ε/2)n, then it returns ACCEPT with probability
at least 1 − δ; (ii) if supp(D) ≤ (1 − ε)n, then it returns REJECT with probability at least 1 − δ.
Lemma 5.2. Given COND access to a distribution D, an upper bound m < n on supp(D), as
well as parameter δ ∈ (0, 1), there exists an algorithm GetNonSupport (Algorithm 3) that makes
O(cid:16)log2 1
m(cid:17) queries to the oracle, and returns an element r ∈ [n] such that r /∈ supp(D) with
probability at least 1 − δ.
Lemma 5.3. Given COND access to a distribution D, inputs σ ≥ 2 and r /∈ supp(D), as well as
parameters ε ∈ (0, 1/2), δ ∈ (0, 1), there exists an algorithm IsAtMostSupportSize (Algorithm 4)
that makes O(cid:0)1/ε2(cid:1) log(1/δ) queries to the oracle, and satisfies the following. The algorithm returns
either yes or no, and (i) if σ ≥ supp(D), then it returns yes with probability at least 1 − δ; (ii) if
σ > (1 + ε) supp(D), then it returns no with probability at least 1 − δ.
δ log−2 n
We defer the proofs of these 3 lemmata to the next subsections, and now turn to the proof of
the theorem.
Proof. The algorithm is given in Algorithm 1, and at a high-level works as follows: if first checks
whether the support size is big (an 1 − O(ε) fraction of the domain), in which case it can already
stop and return a good estimate. If this is not the case, however, then the support is sparse enough
to efficiently find an element r outside the support, by taking a few uniform points, comparing and
ordering them by probability mass (and keeping the lightest). This element r can then be used as
a reference point in a (doubly exponential) search for a good estimate: for each guess ω, a random
subset S of size roughly ω is taken, a point x is drawn from DS, and x is compared to r to check if
D(x) > 0. If so, then S intersects the support, meaning that ω is an upper bound on ω; repeating
until this is no longer the case results in an accurate estimate of ω.
Algorithm 1 EstimateSupportD
1: if TestSmallSupportD(ε, 1
2: end if
3: Call GetNonSupportD((1 − ε
4: for j from 0 to log1+ε log1+ε n do
5:
Set ω ← (1 + ε)(1+ε)j
Call IsAtMostSupportSizeD(ω, r, ε,
if
.
10 ) returns ACCEPT then return ω ← (1 − ε2)n
2 )n, 1
10 ) to obtain a non-support reference point r.
1
100·(j+1)2 ) to check if ω is an upper bound on ω.
that IsAtMostSupportSizeD((1 + ε)i, r, ε,
the call returned no then
Perform a binary search on {(1 + ε)j−1, . . . , (1 + ε)j} to find i∗, the smallest i ≥ 2 such
return ω ← (1 + ε)i∗−1.
10(j+1) ) returns no.
1
6:
7:
8:
9:
end if
10:
11: end for
30
In the rest of this section, we formalize and rigorously argue the above. Conditioning on each of
the calls to the subroutines TestSmallSupport, GetNonSupport and IsAtMostSupportSize
being correct (which overall happens except with probability at most 1/10+1/10+P∞j=1 1/(100j2)+
1/10 < 1/3 by a union bound), we show that the output ω of EstimateSupport is indeed within
a factor (1 + ε) of ω.
• If the test on Step 1 passes, then by Lemma 5.1 we must have supp(D) > (1 − ε)n. Thus,
the estimate we return is correct, as [(1 − ε)n, n] ⊆ [ω/(1 + ε), (1 + ε)ω].
• Otherwise, if it does not then by Lemma 5.1 it must be the case that supp(D) < (1 − ε/2)n.
Therefore, if we reach Step 3 then (1−ε/2)n is indeed an upper bound on ω, and GetNonSupport
will return a point r /∈ supp(D) as expected. The analysis of the rest of the algorithm is straight-
forward: from the guarantee of IsAtMostSupportSize, the binary search will be performed for
the first index j such that ω ∈ [(1 + ε)(1+ε)j−1
]; and will be on a set of (1 + ε)j−1
values. Similarly, for the value i∗ eventually obtained, it must be the case that (1 + ε)i∗
> ω (by
contrapositive, as no was returned by the subroutine) but (1 + ε)i∗−1 ≤ (1 + ε)ω (again, as the
subroutine returned yes). But then, ω = (1 + ε)i∗−1 ∈ (ω/(1 + ε), (1 + ε)ω] as claimed.
Query complexity. The query complexity of our algorithm originates from the following different
steps:
, (1 + ε)(1+ε)j
• the call to TestSmallSupport, which from Lemma 5.1 costs O(cid:0)1/ε2(cid:1) queries;
• the call to GetNonSupport, on Step 3, that from the choice of the upper bound also costs
O(cid:0)1/ε2(cid:1) queries;
• the (at most) log1+ε log1+ε n = O((log log n)/ε) calls to IsAtMostSupportSize on Step 6.
Observing that the query complexity of IsAtMostSupportSize is only O(cid:0)1/ε2(cid:1) · log(1/δ),
(j+1)2 at the j-th iteration this step costs at most
and from the choice of δ = 1
O(cid:18) 1
ε2(cid:19) ·
log1+ε log1+ε n
Xj=1
queries.
O(cid:16)log(j2)(cid:17) = O(cid:18) 1
ε2 log1+ε log1+ε n(cid:19) = O(cid:18) 1
ε3 log1+ε log1+ε n(cid:19)
• Similarly, Step 8 results in at most j ≤ log log n calls to IsAtMostSupportSize with δ set to
1/(10(j + 1)), again costing O(cid:16) 1
ε3 log log n(cid:17) queries.
ε2(cid:17) · log j = O(cid:16) 1
ε2 log1+ε log1+ε n(cid:17) = O(cid:16) 1
(cid:17), as claimed.
Gathering all terms, the overall query complexity is O(cid:16) log log n
ε3
5.1 Proof of Lemma 5.1
Hereafter, we assume without loss of generality that τ < 2: indeed, if τ ≥ 2 then the support is of
size at most n/2, and it suffices to return REJECT to meet the requirements of the lemma. We will
rely on the (easy) fact below, which ensures that any distribution with dense support and minimum
non-zero probability τ /n put significant mass on "light" elements.
Fact 5.4. Fix any ε ∈ [0, 1). Assume D satisfies both supp(D) ≥ (1 − ε)n and D(x) ≥ τ /n for
def= { x ∈ [n] : D(x) ∈ [τ /n, 2/n] }, we have Lε ≥ (1/2 − ε)n and
x ∈ supp(D). Then, setting Lε
D(Lε) ≥ (1/2 − ε)τ .
31
Proof. As the second claim follows directly from the first and the minimum mass of elements of
Lε, it suffices to prove that Lε ≥ (1/2 − ε)n. This follows from observing that
1 = D([n]) ≥ D([n] \ Lε) ≥ (supp(D) − Lε)
2
n ≥ 2(1 − ε) −
2Lε
n
and rearranging the terms.
it
Description and intuition. The algorithm (as described in Algorithm 2) works as follows:
first takes enough uniformly distributed samples s1, . . . , sm to get (with high probability) an ac-
curate enough fraction of them falling in the support to distinguish between the two cases. The
issue is now to detect those mj which indeed are support elements; note that we do not care about
underestimating this fraction in case (b) (when the support is at most (1− ε)n, but importantly do
not want to underestimate it in case (a) (when the support size is at least (1 − ε/2)n). To perform
this detection, we take constantly many samples according to D (which are therefore ensured to be
in the support), and use pairwise conditional queries to sort them by increasing probability mass
(up to approximation imprecision), and keep only the lightest of them, t.
In case (a), we now
from Fact 5.4 that with high probability our t has mass in [1/n, 2/n], and will therefore be either
much lighter than or comparable to any support element: this will ensure that in case (a) we do
detect all of the mj that are in the support.
This also works in case (b), even though Fact 5.4 does not give us any guarantee on the mass
of t. Indeed, either t turns out to be light (and then the same argument ensures that our estimate
of the number of "support" elements mj is good), or t is too heavy -- and then our estimate will
end up being smaller than the true value. But this is fine, as the latter only means we will reject
the distribution (as we should, since we are in the small-support case).
Correctness. Let η be the fraction of the elements sj that are in the support of the distribution.
By a multiplicative Chernoff bound and a suitable constant in our choice of m, we get that (i)
if supp(D) ≥ 1 − ε/2, then Pr[ η < 1 − 3ε/4 ] ≤ 1/12, while (ii) if supp(D) ≤ 1 − ε/2, then
Pr[ η ≥ 1 − 3ε/4 ] ≤ 1/12. We hereafter condition on this (i. e., η being a good enough estimate).
We also condition on all calls to Compare yielding results as per specified, which by a union bound
overall happens except with probability 1/12 + 1/12 = 1/6, and break the rest of the analysis in
two cases.
(a) Since the support size ω is in this case at least (1−ε/2)n, from Fact 5.4 we get that D(Lε/2) ≥
1−ε
2 τ ≥ τ
4 . Therefore, except with probability at most (1 − τ /4)k < 1/12, at least one of the
tj will belong to Lε/2. When this happens, and by the choice of parameters in the calls to
Compare, we get that t ∈ Lε/2; that is D(t) ∈ [τ /n, 2/n]. But then the calls to the routine
on Step 15 will always return either a value (since t is "comparable" to all x ∈ Lε/2 -- i. e., has
probability within a factor 2/τ of them) or High (possible for those sj that have weight greater
than 2/n), unless sj has mass 0 (that is, is not in the support). Therefore, the fraction of
points marked as support is exactly η, which by the foregoing discussion is at least 1 − 3ε/4:
the algorithm returns ACCEPT at Step 20.
(b) Conversely, if ω ≤ (1 − ε)n, there will be a fraction 1 − η > 3ε/4 of the sj having mass 0.
However, no matter what t is it will still be in the support and therefore have D(t) ≥ τ /n:
for these sj, the call to Compare on Step 15 can thus only return Low. This means that
there can only be less than (1− 3
4 ε)m points marked "support" among the sj, and hence that
the algorithm will return REJECT as it should.
32
Algorithm 2 TestSmallSupportD
Require: COND access to D; accuracy parameter ε ∈ (0, 1/2), threshold τ > 0, probability of
1: Repeat the following O(log(1/δ)) times and return the majority vote.
2: loop
failure δ
Draw m def= Θ(cid:16) 1
Draw k def= Θ(cid:16) 1
for all 1 ≤ i < j ≤ k do
ε2(cid:17) independent samples s1, . . . , sm ∼ U[n].
τ(cid:17) independent samples t1, . . . , tk ∼ D.
Call Compare({ti},{tj}, η = 1
if Compare returned High or a value ρ then
2 , K = 2, 1
4k2 ) to get a 2-approx. ρ of D(tj )
⊲ Order the tj
D(ti) , High or Low.
else
Record ti (cid:22) tj
Record tj ≺ tj
end if
end for
Set t to be (any of the) smallest elements tj according to (cid:22).
for all 1 ≤ j ≤ m do
Call Compare({t},{sj}, η = 1
if Compare returned High or a value ρ ≥ 1/2 then
end if
Record sj as "support."
2 , K = 2
τ , 1
⊲ Find the fraction of support elements among the mj
4m ) to get either a value ρ, High or Low.
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
21:
19:
20:
end for
if the number of elements sj marked "support" is at least (1− 3
else return REJECT
end if
22:
23: end loop
4 ε)m then return ACCEPT
Overall, the inner loop of the algorithm thus only fails with probability at most 1/12+ 1/6+ 1/12 =
1/3, (where the 3 events contributing to the union bound are (i) when η fails to be a good estimate,
(ii) when the calls to Compare fail to yield results as claimed, and (iii) when no tj hits Lε/2 in case
(a)). Repeating independently log(1/δ) times and taking the majority vote boosts the probability
of success to 1 − δ.
Query complexity. The sample complexity comes from the k2 calls on Step 4 (each costing
ting of m and because of the log(1/δ) repetitions, this results in an overall query complexity
O(log k) queries) and the m calls on Step 15 (each costing O(cid:16) 1
O(cid:16)(cid:16) 1
ε(cid:17) log 1
δ(cid:17).
τ ε2 log 1
τ 2 log 1
τ + 1
τ log m(cid:17) queries). By the set-
5.2 Proof of Lemma 5.2
As described in Algorithm 3, the subroutine is fairly simple: using its knowledge of an upperbound
on the support size, it takes enough uniformly distributed samples to have (with high probability)
at least one falling outside the support. Then, it uses the conditional oracle to "order" these
samples according to their probability mass, and returns the lightest of them -- i. e., one with zero
probability mass.
33
Algorithm 3 GetNonSupportD(m, δ)
Require: COND access to D; upper bound m on supp(D), probability of failure δ
Ensure: Returns r ∈ [n] such that, with probability at least 1 − δ, r /∈ supp(D)
1: Set k def=llog 2
mm.
δ log−1 n
2: Draw independently k points s1, . . . , sk ∼ U[n]
3: for all 1 ≤ i < j ≤ k do
Call Compare({si},{sj}, η = 1
if Compare returned High or a value ρ then
2 , K = 2,
δ
2k2 ) to get a 2-approx. ρ of D(sj )
D(si) , High or Low.
4:
5:
6:
7:
8:
else
Record si (cid:22) sj
Record sj ≺ sj
end if
9:
10: end for
11: return arg min(cid:22){s1, . . . , sk}
⊲ Return (any) minimal element for (cid:22).
Correctness.
It is straightforward to see that provided at least one of the sj falls outside the
support and that all calls to Compare behave as expected, then the procedure returns one of
the "lightest" elements sj, i. e. a non-support element. By a union bound, the latter holds with
probability at least 1 − δ/2; as for the former, since m is by assumption an upper bound on the
support size it holds with probability at least 1− (m/n)k ≥ 1− δ/2 (from our setting of k). Overall,
the procedure's output is correct with probability at least 1 − δ, as claimed.
Query complexity. The query complexity of GetNonSupport is due to the(cid:0)k
2(cid:1) calls to Compare,
m(cid:17)).
and is therefore O(cid:16)k2 log k
(In our case, we shall eventually take m = (1 − ε/2)n and δ = 1/10, thus getting k = O(1/ε) and
a query complexity of O(cid:0)1/ε2(cid:1).)
δ(cid:17) because of our setting for η and K (which is in turn O(cid:16)log2 1
δ log−2 n
5.3 Proof of Lemma 5.3
Our final subroutine, described in Algorithm 4, essentially derives from the following observation:
a random set S of size (approximately) σ obtained by including independently each element of the
domain with probability 1/σ will intersect the support on ω/σ points on expectation. What we
can test given our reference point r /∈ supp(D), however, is only whether S ∩ supp(D) = ∅. But
this is enough, as by repeating sufficiently many times (taking a random S and testing whether
it intersects the support at all) we can distinguish between the two cases we are interested in.
Indeed, the expected fraction of times S includes a support element in either cases is known to
the algorithm and differs by roughly Ω(ε), so O(cid:0)1/ε2(cid:1) repetitions are enough to tell the two cases
apart.
34
δ
∈ [ 1
Algorithm 4 IsAtMostSupportSizeD(σ, r, ε, δ)
Require: COND access to D; size σ ≥ 2, non-support element r, accuracy ε, probability of failure
Ensure: Returns, with probability at least 1−δ, yes if σ ≤ supp(D) and no if σ > (1+ε)supp(D).
σ(cid:17)σ
1: Set α ← (cid:16)1 − 1
τ 2(cid:17) times do
for m = O(cid:16) 1
2: Repeat the following O(log(1/δ)) times and return the majority vote.
3: loop
4:
4 , e−1], τ ← α(α− ε
2 − 1) = Θ(ε).
Draw a subset S ⊆ [n] by including independently each x ∈ [n] with probability 1/σ.
Draw x ∼ DS.
Call Compare({x},{r}, η = 1
Record yes if Compare returned Low, no otherwise.
100m ) ⊲ Low if S ∩ supp(D) 6= ∅; ρ ∈ [ 1
2 , K = 1,
2 , 2) o.w.
1
5:
6:
7:
8:
9:
end for
10:
11: end loop
return yes if at least m(cid:0)α + τ
2(cid:1) "yes"'s were recorded, no otherwise.
⊲ Thresholding.
Correctness. We condition on all calls to Compare being correct: by a union bound, this overall
happens with probability at least 99/100. We shall consider the two cases σ ≤ ω and σ > (1 + ε)ω,
and focus on the difference of probability p of recording yes on Step 8 between the two, in any fixed
of the m iterations. In both cases, note p is exactly (1 − 1/σ)ω.
σ(cid:17)σ
• If σ ≤ ω, then we have p ≤ (cid:16)1 − 1
σ(cid:17)σ/(1+ε)
• If σ > (1 + ε)ω, then p > (cid:16)1 − 1
= α.
> (cid:16)1 − 1
σ(cid:17)σ(1−ε/2)
= α1−ε/2.
As α ∈ [ 1
4 , e−1], the difference between the two is τ = α(α−ε/2 − 1) = Θ(ε). Thus, repeating the
atomic test of Step 4 O(cid:0)1/τ 2(cid:1) before thresholding at Step 10 yields the right answer with constant
probability, then brought to 1 − δ by the outer repeating and majority vote.
Query complexity. Each call to Compare at Step 7 costs O(log m) queries, and is in total
repeated O(m log(1/δ) times. By the setting of m and τ , the overall query complexity is therefore
ε2 log 1
O(cid:16) 1
δ(cid:17).
ε log 1
5.4 A Non-Adaptive Upper Bound
In this section, we sketch how similar -- yet less involved -- ideas can be used to derive a non-
adaptive upper bound for support-size estimation. For simplicity, we describe the algorithm for
2-approximation: adapting it to general (1 + ε)-approximation is straightforward.
The high-level idea is to perform a simple binary search (instead of the double exponential
search from the preceding section) to identify the greatest lower bound on the support size of the
form k = 2j. For each guess k ∈ {2, 4, 8 . . . , n}, we pick uniformly at random a set S ⊆ [n] of
cardinality k, and check whether DS is uniform using the non-adaptive tester of Chakraborty et
al. [CFGM16, Theorem 4.1.2]. If DS is found to be uniform for all values of k, we return n as our
estimate (as the distribution is close to uniform on [n]); otherwise, we return n/k, for the smallest
k on which DS was found to be far from uniform. Indeed, DS can only be far from uniform if S
contains points from the support of D, which intuitively only happens if n/k = Ω(1).
To be more precise, the algorithm proceeds as follows, where τ > 0 is an absolute constant.
35
for all k ∈ {2, 4, . . . , n} do
Set a counter ck ← 0.
for m = O(log log n) times do
Pick uniformly at random a set S ⊆ [n] of k elements.
Test (non-adaptively) uniformity of DS on S, with the tester of [CFGM16].
if the tester rejects then increment ck.
end if
end for
if ck > τ · m then return ω ← n
k .
end if
end for
return ω ← n.
The query complexity is easily seen to be poly log n, from the O(log n) calls to the poly(log n) tester
of [CFGM16, Theorem 4.1.2]. As for correctness, it follows from the fact that for any set S with
mass D(S) > 0 which contains at least an η fraction of points outside the support, it holds that
DS is η-far from US.
Acknowledgments. Cl´ement Canonne would like to thank Dana Ron and Rocco Servedio for the
many helpful discussions and remarks that influenced the lower bound construction of Section 3.
The authors are grateful to L´aszl´o Babai, Robert Krauthgamer, and the anonymous reviewers for
their helpful and detailed comments.
References
[ACK15] Jayadev Acharya, Cl´ement L. Canonne, and Gautam Kamath. Adaptive estimation in
weighted group testing. In IEEE Internat. Symp. Information Theory (ISIT'15), 2015.
[AD11] Anne Auger and Benjamin Doerr. Theory of Randomized Search Heuristics: Founda-
tions and Recent Developments. Series on theoretical computer science. World Scientific,
2011.
[ADJ+11] Jayadev Acharya, Hirakendu Das, Ashkan Jafarpour, Alon Orlitsky, and Shengjun
Pan. Competitive closeness testing.
In Proc. 24th Ann. Conf. on Learning Theory
(COLT'11), volume 19 of JMLR Proceedings, pages 47 -- 68. JMLR.org, 2011. Accessible
at JMLR.
[ADJ+12] Jayadev Acharya, Hirakendu Das, Ashkan Jafarpour, Alon Orlitsky, Shengjun Pan, and
In Proc.
Ananda Theertha Suresh. Competitive classification and closeness testing.
25th Ann. Conf. on Learning Theory (COLT'12), pages 22.1 -- 22.18, 2012. Accessible at
JMLR.
[BBBY12] Maria-Florina Balcan, Eric Blais, Avrim Blum, and Liu Yang. Active property testing.
In Proc. 53rd FOCS, pages 21 -- 30. IEEE Comp. Soc. Press, 2012.
[BFF+01] Tugkan Batu, Eldar Fischer, Lance Fortnow, Ravi Kumar, Ronitt Rubinfeld, and
Patrick White. Testing random variables for independence and identity. In Proc. 42nd
FOCS, pages 442 -- 451. IEEE Comp. Soc. Press, 2001.
36
[BFR+00] Tugkan Batu, Lance Fortnow, Ronitt Rubinfeld, Warren D. Smith, and Patrick White.
Testing that distributions are close. In Proc. 41st FOCS, pages 189 -- 197. IEEE Comp.
Soc. Press, 2000. This is a preliminary version of [BFR+10].
[BFR+10] Tugkan Batu, Lance Fortnow, Ronitt Rubinfeld, Warren D. Smith, and Patrick White.
Testing closeness of discrete distributions. (abs/1009.5397), 2010. This is a long version
of [BFR+00].
[BFRV11] Arnab Bhattacharyya, Eldar Fischer, Ronitt Rubinfeld, and Paul Valiant. Testing
monotonicity of distributions over general partial orders. In Proc. 2nd Conf. on Inno-
vations in Theoret. Comput. Sci. (ITCS'11), pages 239 -- 252, 2011.
[BKR04] Tugkan Batu, Ravi Kumar, and Ronitt Rubinfeld. Sublinear algorithms for testing
In Proc. 36th STOC, pages 381 -- 390. ACM
monotone and unimodal distributions.
Press, 2004.
[Can15] Cl´ement L. Canonne. A Survey on Distribution Testing: your data is Big, but is it
Blue? Electron. Colloq. on Comput. Complexity (ECCC), April 2015.
[CDVV14] Siu-On Chan, Ilias Diakonikolas, Gregory Valiant, and Paul Valiant. Optimal algo-
rithms for testing closeness of discrete distributions. In Proc. 25th Ann. ACM-SIAM
Symp. on Discrete Algorithms (SODA'14), pages 1193 -- 1203. ACM Press, 2014.
[CFGM16] Sourav Chakraborty, Eldar Fischer, Yonatan Goldhirsh, and Arie Matsliah. On the
power of conditional samples in distribution testing. SIAM J. Comput., 45(4):1261 --
1296, 2016. Preliminary version in ITCS'13.
[Chv79] Vasek Chv´atal. The tail of the hypergeometric distribution. Discrete Mathematics,
25(3):285 -- 287, 1979.
[CJSA14] Chun Lam Chan, Sidharth Jaggi, Venkatesh Saligrama, and Samar Agnihotri. Non-
adaptive group testing: Explicit bounds and novel algorithms. IEEE Trans. Inform.
Theory, 60(5):3019 -- 3035, 2014. Preliminary version in ISIT'12.
[CR14] Cl´ement L. Canonne and Ronitt Rubinfeld. Testing probability distributions under-
In Proc. 41st Internat. Colloq. on Automata, Languages and
lying aggregated data.
Programming (ICALP'14), pages 283 -- 295. Springer, 2014.
[CRS15] Cl´ement L. Canonne, Dana Ron, and Rocco A. Servedio. Testing probability distribu-
tions using conditional samples. SIAM J. Comput., 44(3):540 -- 616, 2015.
[Das05] Sanjoy Dasgupta. Analysis of a greedy active learning strategy. In Proc. Advances in
Neural Information Processing Systems (NIPS'05), pages 337 -- 344. MIT Press, 2005.
[DH00] Dingzhu Du and Frank K. Hwang. Combinatorial Group Testing and Its Applications.
Applied Mathematics. World Scientific, 2000.
[Dor43] Robert Dorfman. The detection of defective members of large populations. Ann. Math.
Statist., 14(4):436 -- 440, 12 1943.
[DR96] Devdatt P. Dubhashi and Desh Ranjan. Balls and Bins: A Study in Negative Depen-
dence. Random Structures Algorithms, 13:99 -- 124, 1996.
37
[Fis01] Eldar Fischer. The art of uninformed decisions: A primer to property testing. Bulletin
of the European Association for Theoretical Computer Science, 75:97 -- 126, 2001.
[FJO+15] Moein Falahatgar, Ashkan Jafarpour, Alon Orlitsky, Venkatadheeraj Pichapathi, and
Ananda Theertha Suresh. Faster algorithms for testing under conditional sampling.
In Proc. 28th Ann. Conf. on Learning Theory (COLT'15), JMLR Proceedings, pages
607 -- 636, 2015.
[GMV06] Sudipto Guha, Andrew McGregor, and Suresh Venkatasubramanian. Streaming and
sublinear approximation of entropy and information distances.
In Proc. 17th Ann.
ACM-SIAM Symp. on Discrete Algorithms (SODA'06), pages 733 -- 742. ACM Press,
2006.
[Gol10] Oded Goldreich, editor. Property Testing: Current Research and Surveys. Springer,
2010. LNCS 6390.
[GR00] Oded Goldreich and Dana Ron. On testing expansion in bounded-degree graphs. Tech-
nical report, Electron. Colloq. on Comput. Complexity (ECCC), 2000.
[ILR12] Piotr Indyk, Reut Levi, and Ronitt Rubinfeld. Approximating and Testing k-Histogram
Distributions in Sub-linear Time. In Proc. 31st Symp. on Principles of Database Systems
(PODS'12), pages 15 -- 22, 2012.
[LRR13] Reut Levi, Dana Ron, and Ronitt Rubinfeld. Testing properties of collections of distri-
butions. Theory of Computing, 9(8):295 -- 347, 2013.
[MR95] Rajeev Motwani and Prabhakar Raghavan. Randomized Algorithms. Cambridge Uni-
versity Press, New York, NY, USA, 1995.
[ND00] Hung Q. Ngo and Ding-Zhu Du. A Survey on Combinatorial Group Testing Algorithms
with Applications to DNA Library Screening. In DIMACS Ser. in Discr. Math. and
Theoret. Comput. Sci., 2000.
[Pan08] Liam Paninski. A coincidence-based test for uniformity given very sparsely sampled
discrete data. IEEE Trans. Inform. Theory, 54(10):4750 -- 4755, 2008.
[Ron08] Dana Ron. Property Testing: A Learning Theory Perspective. Foundations and Trends
in Machine Learning, 1(3):307 -- 402, 2008. Preliminary version in COLT'07.
[Ron10] Dana Ron. Algorithmic and analysis techniques in property testing. Found. Trends
Theor. Comput. Sci., 5:73 -- 205, 2010.
[RRSS09] Sofya Raskhodnikova, Dana Ron, Amir Shpilka, and Adam Smith. Strong lower bounds
for approximating distributions support size and the distinct elements problem. SIAM
J. Comput., 39(3):813 -- 842, 2009. Preliminary version in FOCS'07.
[RS09] Ronitt Rubinfeld and Rocco A. Servedio. Testing monotone high-dimensional distribu-
tions. Random Structures Algorithms, 34(1):24 -- 44, January 2009. Preliminary version
in STOC'05.
[RT16] Dana Ron and Gilad Tsur. The power of an example: Hidden set size approximation
using group queries and conditional sampling. ACM Trans. Comput. Theory, 8(4):15:1 --
15:19, 2016.
38
[Rub12] Ronitt Rubinfeld. Taming Big Probability Distributions. XRDS, 19(1):24 -- 28, Septem-
ber 2012.
[Set09] Burr Settles. Active learning literature survey. Computer Sciences Technical Report
1648, University of Wisconsin -- Madison, 2009.
[Sto85] Larry J. Stockmeyer. On approximation algorithms for #P. SIAM J. Comput.,
14(4):849 -- 861, 1985.
[Sub] List of Open Problems in Sublinear Algorithms: Problem 66. Bertinoro Work-
shop on Sublinear Algorithms 2014 (suggested by Eldar Fischer). Available at
http://sublinear.info/66.
[Val11] Paul Valiant. Testing symmetric properties of distributions.
SIAM J. Comput.,
40(6):1927 -- 1968, 2011. Preliminary version in STOC'08.
[VV10a] Gregory Valiant and Paul Valiant. A CLT and tight lower bounds for estimating entropy.
Electron. Colloq. on Comput. Complexity (ECCC), 2010.
[VV10b] Gregory Valiant and Paul Valiant. Estimating the unseen: A sublinear-sample canonical
estimator of distributions. Electron. Colloq. on Comput. Complexity (ECCC), 2010.
[VV11] Gregory Valiant and Paul Valiant. The power of linear estimators. In Proc. 52nd FOCS,
pages 403 -- 412, October 2011. See also [VV10a] and [VV10b].
[VV17] Gregory Valiant and Paul Valiant. An automatic inequality prover and instance opti-
mal identity testing. SIAM J. Comput., 46(1):429 -- 455, 2017. Preliminary version in
FOCS'14.
[Wik17] Wikipedia. Yes, Virginia, there is a Santa Claus -- Wikipedia, The Free Encyclopedia,
2017. https://en.wikipedia.org/w/index.php?title=Yes,_Virginia,_there_is_a_Santa_Claus&oldid=805832621
[Online; accessed 05-November-2017].
39
|
1612.00958 | 1 | 1612 | 2016-12-03T11:34:53 | Reconfiguring Ordered Bases of a Matroid | [
"cs.DS",
"math.CO"
] | For a matroid with an ordered (or "labelled") basis, a basis exchange step removes one element with label $l$ and replaces it by a new element that results in a new basis, and with the new element assigned label $l$. We prove that one labelled basis can be reconfigured to another if and only if for every label, the initial and final elements with that label lie in the same connected component of the matroid. Furthermore, we prove that when the reconfiguration is possible, the number of basis exchange steps required is $O(r^{1.5})$ for a rank $r$ matroid. For a graphic matroid we improve the bound to $O(r \log r)$. | cs.DS | cs |
Reconfiguring Ordered Bases of a Matroid
Anna Lubiw∗
Vinayak Pathak∗
November 5, 2018
Abstract
For a matroid with an ordered (or "labelled") basis, a basis exchange step removes
one element with label l and replaces it by a new element that results in a new basis,
and with the new element assigned label l. We prove that one labelled basis can be
reconfigured to another if and only if for every label, the initial and final elements
with that label lie in the same connected component of the matroid. Furthermore, we
prove that when the reconfiguration is possible, the number of basis exchange steps
required is O(r1.5) for a rank r matroid. For a graphic matroid we improve the bound
to O(r log r).
1
Introduction
Broadly speaking, "reconfiguration" is about changing one structure to a related one via
discrete steps. Examples include changing one triangulation of a point set in the plane to
another via "flips", or changing one independent set of a graph to another by adding/deleting
single vertices. More broadly, other examples are to sort a list by swapping pairs of adjacent
elements, or to change a basic solution of a linear program to an optimum one via pivots.
Central questions are: to test whether one configuration can be changed to another using
the discrete steps; and to compute, or bound, the number of steps required. Reconfigura-
tion problems have been considered for a long time, and explicit attention was drawn to
complexity issues by Ito et al. [5] and by van den Heuvel [4].
A fundamental result about matroids is that any basis B can be reconfigured to any
other basis B(cid:48) using a sequence of basis exchange steps. In each basis exchange one element
of the current basis is replaced by a new element to yield a new basis. The number of basis
exchange steps needed to reconfigure B to B(cid:48) is the difference B − B(cid:48).
In this paper we explore a "labelled" version of basis reconfiguration where the elements
of each of the two initial bases are labelled with unique labels from {1, . . . , r}, where r is the
rank of the matroid. For a basis exchange in this labelled setting, the new element gets the
same label as the replaced element. More precisely, if l is the labelling function on B and
e ∈ B is replaced with e(cid:48) ∈ E \ B, then e(cid:48) gets assigned the label l(e). In standard matroid
terminology, a "labelled" basis is an "ordered" basis.
We consider two questions:
1
1. When can one labelled basis of a matroid be reconfigured to another labelled basis of
the matroid using a sequence of basis exchanges?
2. What is the number of basis exchanges needed in the worst case?
Note that such labelled reconfiguration is not always possible -- for example, a matroid
may have a single basis B, in which case there is no way to reconfigure from one labelling of
B to a different labelling.
Results. In this paper we prove that one labelled basis can be reconfigured to another if
and only if for every label, the element with that label in the first basis and the element
with that label in the second basis lie in the same connected component of the matroid.
Equivalently, for a given basis, the permutations of labels achievable by sequences of basis
exchange steps is a product of symmetric groups, one for each connected component of the
matroid.
Furthermore, we prove that if one labelled basis can be reconfigured to another then
O(r1.5) basis exchange steps always suffice, where r is the rank of the matroid.
In the special case of graphic matroids, our problem is the following: given two spanning
trees of an n-vertex graph, with the edges of the spanning trees labelled with the labels
{1, 2, . . . , n− 1}, reconfigure the first labelled spanning tree to the second via basis exchange
steps. Reconfiguration is possible if and only if for every label, the edge with that label in
the first spanning tree and the edge with that label in the second spanning tree lie in the
same 2-connected component of the graph. This was proved by Hernando et al. [3]. In this
case we can prove a tighter bound: O(n log n) basis exchange steps are always sufficient.
The rest of the paper is organized as follows. Subsections 1.1 and 1.2 contain notation
and background. In Section 2 we prove the results for graphic matroids, and in Section 3 we
prove the results for general matroids.
1.1 Preliminaries
For a matroid M of rank r, a labelled basis (or "ordered basis") is a tuple T = (B, l) where
B is a basis and l : B → {1, . . . , r} is a function that assigns a unique label to each element
of B. For label i, the element with that label is given by l−1(i). If a basis exchange step
replaces e ∈ B with e(cid:48) ∈ E \ B, then e(cid:48) is assigned the label l(e). Two bases are said to be
the same if they have the same elements and the elements are assigned the same labels.
Matroid Properties. For basic definitions and properties of matroids, see Oxley [8]. In
the remainder of this section, we summarize the properties that we will use. Note that we
are now referring to standard unlabelled matroids.
Recall the basis exchange property of matroids:
Property 1. For any two bases B1 and B2 of a matroid M and an element x ∈ B1 − B2,
there exists y ∈ B2 − B1 such that (B1 − {x}) ∪ {y} is also a basis.
By Property 1 any basis of a matroid can be reconfigured into any other using at most
r basis exchange steps.
2
The pairs of elements that can participate in exchanges can be characterized exactly.
Let B be a basis. Adding an element e /∈ B to B creates a unique circuit C called the
fundamental circuit of e with respect to B.
[8, Exercise 5, Section 1.2] Given a basis B, and an element e /∈ B, let C
Property 2.
be the fundamental circuit of e with respect to B. Then the set of elements e(cid:48) such that
(B − {e(cid:48)}) ∪ {e} is a basis are precisely the elements e(cid:48) ∈ C.
The fundamental graph of basis B, denoted SB, is the bipartite graph with a vertex for
each element of the matroid, and an edge (e, f ) for every e ∈ B and f ∈ E − B such that e, f
form a basis exchange, i.e. (B − {e}) ∪ {f} is a basis. Note that the closed neighbourhood
of f ∈ E − B is f 's fundamental circuit.
For a matroid M with element set E and a set T ⊆ E, the matroid that results from
contracting T is denoted M/T .
Property 3. For matroid M, and sets T ⊆ E and X ⊆ E − T
rM/T (X) = rM (X ∪ T ) − rM (T ).
This property implies that we can augment an independent set as follows:
Property 4. If T is independent in matroid M and A is independent in M/T then T ∪ A
is independent in M .
The concept of graph connectivity generalizes to matroids. For any matroid M , define
the relation ψ on the elements of M by e ψ f if and only if either e = f or there exists a
circuit of M that contains both e and f . It can be shown that the relation ψ is an equivalence
relation and we say that each equivalence class defines a connected component. In the case
of a graphic matroid, the equivalence classes are the 2-connected components or blocks of the
graph.
We will use Menger's theorem for graphs, see [10, chapter 9], which says that the max-
imum number of vertex-disjoint paths from vertex s to vertex t is equal to the minimum
number of vertices whose removal disconnects s and t. We will also use the generalization of
Menger's theorem to matroids, which is known as Tutte's linking theorem [11], see [8]. The
statement of this theorem will be given where we use it in Section 3.2.
1.2 Relationship to Triangulations
Our present study of reconfiguring labelled matroid bases was prompted by related results
on reconfiguring labelled triangulations [2].
The set of triangulations of a point set (or a simple polygon) in the plane has some
matroid-like properties.
In particular, given a point set P , let E be the set of all line
segments d such that the endpoints of d are in P and no other point of P lies on d, and
let I be the set of all subsets of pairwise non-crossing segments of E. Then the set system
(E,I) is an independence system, but fails the third matroid axiom: if I1 and I2 are in I and
3
I2 > I1, there does not always exist an element e of I2 − I1 such that I1 ∪ {e} ∈ I. (As a
simple example consider five points 1, 2, 3, 4, 5 in convex position where I1 contains segments
(1, 3) and (1, 4) and I2 contains segment (2, 5).) However, the maximal sets in I all have
the same cardinality -- this is because maximal sets of non-crossing segments correspond to
triangulations and all triangulations contain the same number of edges.
The failure of the third matroid axiom for triangulations means that the greedy algorithm
does not in general find a minimum weight triangulation of a point set, and in fact the
problem is NP-hard even for Euclidean weights [7].
On the other hand, triangulations share some of the reconfiguration properties of matroids --
any triangulation can be reconfigured to any other triangulation via a sequence of elementary
exchange steps called "flips" where one segment is removed and a new segment (the oppo-
site chord of the resulting quadrilateral, if it is convex) is added [1]. Furthermore, there is
evidence that results analogous to the ones we prove here for reconfiguring labelled matroid
bases carry over to labelled triangulations [2]. One intriguing possibility is a more general
result that encompasses both the case of matroids and of triangulations.
2 Reconfiguring ordered bases of a graphic matroid
In this section we concentrate on graphic matroids. We characterize when one labelled
spanning tree of an n-vertex graph can be reconfigured to another labelled spanning tree of
the graph. Our first result provides a bound of O(n2) exchange steps for the reconfiguration.
In the second subsection we improve this to O(n log n). The first result (with the O(n2)
bound) extends immediately to general matroids, but we give the details for the graphic
matroid case for the sake of readers who may wish to learn only about labelled reconfiguration
of spanning trees.
2.1 When are two ordered bases reconfigurable?
Theorem 1. Given two labelled spanning trees T1 and T2 of an n-vertex graph G, we can
reconfigure one to the other if and only if for each label i, the edge with that label in T1 and
the edge with that label in T2 lie in the same 2-connected component of G. Moreover, when
reconfiguration is possible, it can be accomplished with O(n2) basis exchange steps.
Proof. The 'only if' direction is clear because an exchange of edge e to e(cid:48) can be performed
only if both e and e(cid:48) lie in the same 2-connected component. For the 'if' direction, we will
provide an explicit exchange sequence to reconfigure T1 to T2.
First, pick any (unlabelled) spanning tree B and reconfigure both T1 and T2 to B while
ignoring the labels. This takes O(n) exchange steps. Let σ1 be the sequence of exchanges
that reconfigures T1 to B and σ2 be the sequence that reconfigures T2 to B. Obviously, the
labels of the edges of B obtained from the two sequences will not match in general. Below,
we give an exchange sequence σ of length at most O(n2) to rearrange the labels in B. Thus
performing σ1 followed by σ followed by the reverse of σ2 reconfigures T1 to T2 with O(n2)
exchange steps.
4
The problem is now reduced to the following: given one spanning tree B and two labellings
of it, T1 = (B, l1) and T2 = (B, l2), reconfigure T1 to T2 using O(n2) exchanges. We will do
this by repeatedly swapping labels. More precisely, let i be a label, let e1 = l−1
1 (i) be the
edge with label i in T1 and let e2 = l−1
2 (i) be the edge with label i in T2, and suppose that
e1 (cid:54)= e2. We will show that with O(n) exchanges we can swap the labels of e1 and e2 in T1
while leaving all other labels unchanged. This moves label i to the correct place for T2, and
repeating over all labels solves the problem and takes O(n2) exchanges.
Since e1 and e2 lie in the same 2-connected component of G (by hypothesis), there must
exist a cycle C of G that goes through both e1 and e2. Let t be the number of edges of C \ B.
Then t ≥ 1. We will argue by induction on t that 6t− 3 exchanges suffice to swap the labels
of e1 and e2. Let f be an edge of C \ B.
If f is the only edge of C \ B, then we can perform the swap directly: use Property 2
to exchange e1 with f , e2 with e1, and, finally, f with e2. This sequence returns us to B
and swaps the labels of e1 and e2 while leaving all other labels unchanged. Thus the case of
t = 1 can be solved with 3 exchanges.
More generally, f is not the only edge of C \ B. Adding f to B creates a cycle C(cid:48) that
must contain an edge f(cid:48) ∈ B \ C. Using Property 2 we exchange f with f(cid:48). The result is
a new spanning tree B(cid:48) whose intersection with C is strictly increased, so C \ B(cid:48) is smaller.
Also note that B(cid:48) contains e1 and e2. By induction, we can swap the labels of e1 and e2 in
B(cid:48) with at most 6(t − 1) − 3 exchanges, leaving all other labels unchanged. After that, we
exchange f(cid:48) and f to return to B with original labels except that the labels of e1 and e2 are
now swapped. The total number of exchanges is at most 6(t − 1) − 3 + 6 = 6t − 3
2.2 Tightening the bound
In this section we show that the O(n2) bound on the number of exchanges from the previous
section can be improved. Note that the common spanning tree B we chose in the previous
section, to reconfigure both T1 and T2 to, was completely arbitrary. We could have perhaps
chosen a spanning tree that made the task of swapping labels easier. That is precisely what
we do in this section.
Observe that it is sufficient to consider a 2-connected graph G since for a general graph we
can just repeat the argument inside each of its 2-connected components. We will construct
a spanning tree B of G whose fundamental graph (as defined in Section 1.1) has diameter
O(log n). This proves our result based on:
Lemma 1. Let G be a 2-connected graph, and B be a spanning tree of G whose fundamental
graph SB has diameter d. Then for any two edges of B there is an exchange sequence of
length O(d) that swaps the labels of those two edges while leaving other labels unchanged.
Proof. Let e and f be two edges of B, and suppose the shortest path between them in SB
has length t < d. We will prove by induction on t that we can swap the labels of e and f
with 3t − 3 exchanges. Let e1 and e2 be the two vertices that occur immediately after e on
a shortest path from e to f in SB. Then the cycle formed by adding e1 to B contains both
e and e2 by Property 2, and by the same property, we can perform the following exchanges:
5
e1 with e, then e2 with e1, then e with e2. This exchange sequence returns us to B while
swapping the labels of e and e2, and leaving other labels unchanged. In the basis case of the
induction t = 2 and e2 = f and this completes the swap with 3 exchanges. In the general
case, the distance between e2 and f in SB is t − 2. By induction, we can swap the labels of
e2 and f with 3(t− 2)− 3 exchanges. After that we repeat the first 3 exchanges to complete
the swap of the labels of e and f . All other labels are unchanged. The total number of
exchanges is 6 + 3(t − 2) − 3 = 3t − 3.
Thus it remains to construct a spanning tree B such that the diameter of SB is O(log n).
We will construct our spanning tree by repeatedly contracting cycles. For a 2-connected
graph G with edge set E(G) and for a cycle C ⊆ E(G), let G/C denote the graph obtained
by contracting C. Note that E(G/C) = E(G) − E(C). Contracting C creates blocks that
are the maximal subgraphs of G/C that are 2-connected. In order to get a bound on the
diameter of SB we will need the following:
Lemma 2. Any 2-connected graph G with m edges contains a cycle C such that all blocks
of G/C have at most m/2 edges.
Proof. Let C be the cycle that minimizes the size of the largest block obtained upon con-
tracting it. We claim that all blocks of G/C have at most m/2 edges. Suppose not. Then
there exists a block H of size bigger than m/2. We will derive a contradiction.
Some edges of E(H) are incident on vertices of C in G; let those vertices be {v1, . . . , vk}
in clockwise order along C. There are two paths between v1 and v2 along the cycle C in G,
one clockwise and one counterclockwise. Let P be the one that is counterclockwise and thus
contains all vertices of {v1, . . . , vk}. There also exists a path P (cid:48) between v1 and v2 that uses
only the vertices of H. We define C(cid:48) to be P ∪ P (cid:48). We claim that the size of the largest block
of G/C(cid:48) is smaller than the size of the largest block of G/C, hence reaching a contradiction.
First, note that no block of G/C(cid:48) contains an edge of H and an edge not in H. For
consider edges e and e(cid:48) in G/C(cid:48) with e in H and e(cid:48) not in H. Any path from e to e(cid:48) in G
must go through a vertex of C, and in particular, must go through a vertex of P , since P
contains all vertices of C that have an edge of H incident to them. Because C(cid:48) contains all
of P , therefore e and e(cid:48) are in different blocks of G/C(cid:48).
Now G/C(cid:48) contains two kinds of blocks: those that contain edges of H and those that
do not. Blocks of the first kind must have size at most the size of H from the argument
above. In fact, they must be strictly smaller than H since C(cid:48) contains at least one edge of H.
Blocks of the second kind must also be smaller than H since at worst, such a block contains
all edges of G that are not edges of H, and there are at most m/2 such edges.
We are now ready to construct our spanning tree B.
Lemma 3. Given a 2-connected graph G, there exists a spanning tree B such that the
diameter of SB is O(log n).
Proof. The algorithm for constructing B proceeds in iterations i = 1, 2, . . .. In iteration i we
will add a set Bi to B. In the first iteration we find the cycle C of Lemma 2 such that all
6
Figure 1: Adding a minimal set of edges P (in dashed red) to create a 2-connection between
Bi−1 (thick edges) and C i (thin edges). P may consist of two, one, or zero disjoint paths as
shown in (a), (b), or (c), respectively.
blocks of G/C are of size at most m/2. Let B1 be all edges but one of C, and contract those
edges (equivalently, contract C) thus breaking the graph G into several blocks. In general,
in any iteration, we start with the collection of blocks produced in the previous iteration,
contract a connected set of edges inside each block, and add those edges to Bi. After each
contraction we eliminate loops and parallel edges. Each iteration reduces the number of
vertices of G by contracting a set of edges. The algorithm terminates once the graph G is
left with just one vertex.
We now describe what happens to one of the blocks that is dealt with in iteration i. For
i ≥ 2, let H i−1 be a block at the beginning of iteration i − 1, and let Bi−1 be the edges
of H i−1 that we contract and add to B in the (i − 1)st iteration. Let b be the vertex that
Bi−1 gets contracted to, and let H i be one of the blocks formed by the contraction. Then in
iteration i, we pick the edges of H i to be contracted and added to Bi, as follows. Let C i be
the cycle of Lemma 2 for H i. Observe that C i may or may not include vertex b. Let e be
an edge of C i. As in the first iteration, we will add to Bi all edges of C i except e. However,
we may need to add more edges in order to maintain connectivity of B, and to ensure that
SB has small diameter.
In H i−1 the sets of edges Bi−1 and C i are disjoint. We will use the fact that H i−1 is
2-connected and apply Menger's Theorem, see [10, chapter 9], to find a set of edges that
connect Bi−1 and C i. More precisely, let s be a new vertex adjacent to all endpoints of
edges in Bi−1 and t be a new vertex adjacent to all endpoints of edge in C i. Because H i−1
is 2-connected, we must remove at least 2 vertices to disconnect s and t. Then, by Menger's
Theorem, there are two internally vertex-disjoint paths from s to t. Let P be a minimal set
of edges of H i−1 that form two such paths. Note that if C i includes vertex b, one or both of
the paths will have no edges of H i−1. See Figure 1.
Observe that the two paths of P go from two distinct vertices that are joined by a path
in Bi−1 to two distinct vertices that are joined by a path in C i − {e}. Thus, a cycle, D, is
formed by P together with a non-empty subset of Bi−1 and a non-empty subset of C i −{e}.
Let f be an edge of P (if P is non-empty). By minimality of P , there is no cycle in P −{f}.
Add to Bi the set (C i − {e}) ∪ (P − {f}). This completes the description of how we handle
7
Bi -1Ci Bi -1Ci e ffe Bi -1Ci e (a)(b)(c) one block in iteration i. We handle other blocks the same way, adding further edges to Bi
for each other block. This completes the description of the algorithm.
To establish the correctness of the algorithm, first note that after each iteration B =
∪i
j=1Bj is connected and contains no cycle. Thus, when the algorithm terminates, B spans
all vertices of G and contains no cycle.
It remains to prove that the diameter of SB is
O(log n). In each iteration of the algorithm, we reduce the size of each block by at least half,
and thus the algorithm terminates in at most O(log n) iterations. To complete the proof we
will show that for each i, every edge in Bi has a path of length O(1) in SB to some edge in
Bi−1.
Referring to the step of the construction described above, note that the fundamental
circuit of e in Bi contains all of C i. Thus e is joined by an edge of SB to every edge of C i.
If P is empty, then the fundamental circuit of e also includes an edge of Bi−1 and we are
done. Otherwise, the fundamental circuit of f in B contains all of D, which includes all of
P together with at least one edge of C i and at least one edge of Bi−1. Thus f is joined by
an edge of SB to every edge of P , and to at least one edge of C i and to at least one edge of
Bi−1. Therefore, in SB every edge of Bi is within distance 4 of some edge of Bi−1.
Lemmas 1, 2, and 3 give us the following strengthened form of Theorem 1.
Theorem 2. If T1 and T2 are two labelled spanning trees of an n-vertex graph G and for
each label i, the edge with that label in T1 and the edge with that label in T2 lie in the same
2-connected component of G then we can reconfigure T1 to T2 using O(n log n) basis exchange
steps.
3 Reconfiguring ordered bases of a general matroid
In this section we turn to general matroids. We generalize the result of the previous sec-
tion that characterizes when one labelled basis of a rank r matroid can be reconfigured to
another labelled basis of the matroid and prove a bound of O(r2) exchange steps for the
reconfiguration. In the second subsection we improve this bound to O(r1.5), a weaker bound
than was possible for graphic matroids.
3.1 Connectivity
Our goal is to follow the proof of Theorem 1, which used edge contraction, cycles, and
2-connectivity in a graph. Observe that contraction of edges in a graph corresponds to
contraction of elements in a matroid, cycles in a graph correspond to circuits in a matroid,
and 2-connectivity in a graph corresponds to connectivity in a matroid (every pair of elements
is contained in some circuit).
With these correspondences, it is easy to check that every step of the proof of Theorem 1
goes through for matroids and thus we get the following theorem, which we state without
proof.
8
Theorem 3. Given two labelled bases T1 and T2 of a rank r matroid M , we can reconfigure
one to the other if and only if for each label i, the element with that label in T1 and the
element with that label in T2 lie in the same block of M . Moreover, when reconfiguration is
possible, it can be accomplished with O(r2) basis exchange steps.
3.2 A tighter bound for general matroids
In order to tighten the O(r2) bound on the number of exchanges needed for labelled basis
reconfiguration in general matroids, we would like to follow the approach we used for graphic
matroids in Section 2.2. Lemma 1 carries over directly so it suffices to build a basis B whose
fundamental graph SB has small diameter. For graphic matroids of rank r, we achieved
√
diameter O(log r) but for general matroids we will only achieve a weaker bound of diameter
r). Our starting point for graphic matroids was Lemma 2 which proved that there is a
O(
cycle whose contraction cuts the size of a block in half. For general matroids the analogous
result is conjectured to be true, but we must rely on the following weaker result, attributed
to Seymour, and with an explicit proof in [6, Corollary 1.4].
Lemma 4. Let C be the biggest circuit of a connected matroid M . Then the biggest circuit
of M/C is strictly smaller than C.
Using this lemma we can follow the approach we used for graphic matroids to prove the
following bound.
Theorem 4. If T1 and T2 are two labelled bases of a rank r matroid M and for each label
i, the element with that label in T1 and the element with that label in T2 lie in the same
connected component of M then we can reconfigure T1 to T2 using O(r1.5) basis exchange
steps.
Proof. Following the idea of the proof of Theorem 2, we will prove the bound by showing
√
that every connected component of the matroid has a basis B whose fundamental graph SB
r). We do this using almost exactly the algorithm of Lemma 3 with one
has diameter O(
difference: instead of picking the cycle of Lemma 2 in each block in each iteration, we will
pick the biggest circuit and use Lemma 4.
As before, the algorithm for constructing B proceeds in iterations i = 1, 2, . . .. In iteration
i we will add a set Bi to B. In the first iteration we find the the largest circuit C in M
and let B1 be all elements but one of C. We contract those elements, thus breaking the
matroid into several connected components.
In general, in any iteration i, we start with
the collection of connected components produced in the previous iteration, contract some
elements inside each component, and add those elements to Bi. After each contraction we
eliminate loops and parallel elements. Each iteration reduces the number of elements of M
and the algorithm terminates when no elements are left.
We now describe what happens to one of the connected components that is dealt with
in iteration i. For i ≥ 2, let H i−1 be a connected component at the beginning of iteration
i − 1, and let Bi−1 be the elements of H i−1 that we contract and add to B in the (i − 1)st
9
iteration. Let H i be one of the connected components formed by the contraction. Then in
iteration i, we pick the elements of H i to be contracted and added to Bi, as follows.
Let C i be the biggest circuit of H i and let e be an element of C i. As before, we will put
C i − {e} into Bi, but, as before, we may need to add elements to connect the independent
set Bi−1 with the circuit C i. We will be working in the matroid H i−1 which we abbreviate
as H.
Since C i is a circuit in H i, we have C i−1 = rH i(C i). Now, H i is a connected component
of H/Bi−1, so by Property 3, rH i(C i) = rH(C i ∪ Bi−1) − rH(Bi−1). Thus rH(C i ∪ Bi−1) =
Bi−1 + C i − 1, which means that C i ∪ Bi−1 has co-rank 1 and contains a unique circuit.
If C i is independent in H then C i ∪ Bi−1 has a circuit formed by the elements of C i
In this case we add to Bi the set C i − {e}.
together with at least one element of Bi−1.
Observe that this set is independent in H and that the fundamental circuit of e contains all
elements of C i−1 and at least one element of Bi−1. In the graphic case, this corresponds to
Figure 1 (right). We will prove below that Bi has the properties we need.
Otherwise, C i is not independent in H. In this case, the circuit in C i ∪ Bi−1 is C i, and
we will need to add more elements to Bi in order to connect Bi−1 with C i. We will use the
matroid analogue of Menger's Theorem which is known as Tutte's Linking Theorem [11].
This theorem applies to two disjoint sets of elements in a matroid. In our case the matroid
is H, and the disjoint sets are Bi−1 and C i − {e}, both of which are independent in H. To
ease notation, let A = C i − {e}.
The analogue of a separating set of vertices is κH(Bi−1, A), defined as the minimum,
over sets X that contain Bi−1 and exclude A, of rH(X) + rH(E − X) − rH(E). Since H is
connected, this minimum is at least 1. In notation, we have:
κH(Bi−1, A) =
min
Bi−1⊆X⊆E\A
rH(X) + rH(E − X) − rH(E) ≥ 1.
The analogue of vertex-disjoint paths in Menger's theorem is (cid:117)H/P (Bi−1, A), defined as
the maximum over sets P , of rH/P (Bi−1) + rH/P (A) − rH/P (Bi−1 ∪ A).
According to the version of Tutte's Linking Theorem stated as equation (8.16) in Ox-
ley [8], there exists a set of elements P such that
(cid:117)H/P (Bi−1, A) = κH(Bi−1, A).
(1)
As in the proof of the generalization of Tutte's Linking Theorem due to Geelen, Gerards,
and Whittle [8, proof of Theorem 8.5.7], we will choose P to be a minimal set such that
equation (1) holds. As shown in their proof, such a minimal P is independent, and is skew
to Bi−1 and A. Two sets are skew if the rank of their union is the sum of their ranks. In
our case, Bi−1 and A are independent, so skewness implies that Bi−1 ∪ P and A ∪ P are
independent.
Applying Property 3 to (cid:117)H/P (Bi−1, A) yields
(cid:117)H/P (Bi−1, A) = rH(Bi−1 ∪ P ) + rH(A ∪ P ) − rH(P ) − rH(Bi−1 ∪ A ∪ P ).
(2)
10
When P is independent and skew to Bi−1 and A this becomes:
(cid:117)H/P (Bi−1, A) = Bi−1 + A + P − rH(Bi−1 ∪ A ∪ P ).
(3)
From this, it is clear that if the value of equation (1) is greater than 1, then we can delete
elements of P to obtain a minimal P with (cid:117)H/P (Bi−1, A) = 1.
For the remainder of the proof we define P to be a minimal set with (cid:117)H/P (Bi−1, A) = 1.
From equation (3) we know that Bi−1 ∪ A∪ P has co-rank 1. There must be a unique circuit
D in Bi−1∪A∪P and D must contain all elements of P (by minimality of P ) and at least one
element of Bi−1 (since A ∪ P is independent) and at least one element of A (since Bi−1 ∪ P
is independent).
Let f be an element of P and add to Bi the set (C i − {e}) ∪ (P − {f}). Observe that
this set is independent in H. The fundamental circuit of e contains all the elements of C i,
and the fundamental circuit of f contains all the elements of P and at least one element of
Bi−1 and at least one element of C i.
Before proceeding with the proof we will mention why the above analysis was separated
into two cases depending on whether or not C i is independent in H. Observe that if C i is
independent in H, then the minimal set P that connects Bi−1 and C i−{e} is in fact P = {e}
itself, and when we choose f to be an element of P then f = e. It would be rather strange
in this case (though strictly speaking, correct) to say that we add (C i − {e}) ∪ (P − {f}) to
Bi. That is why we treated the case where C i is independent in H as a separate case.
To complete the proof of the Theorem we must show that the final B, defined as ∪Bi, is a
√
r). Since we continue until everything is contracted
basis and that the diameter of SB is O(
away, it is clear that B spans the matroid. Because each Bi is independent after contracting
all Bj for j < i, Property 4 implies that B is independent in the matroid M . Thus B is a
basis.
√
We now analyze the diameter of SB. We first observe that the algorithm terminates in
√
r) iterations. This is because the number of possible iterations for which there exists
r) and once the size of the
r) more
√
O(
√
a block containing a circuit of size Ω(
biggest circuit in each block has been reduced to O(
iterations.
To complete the proof we will show that for each i, every edge in Bi has a path of length
O(1) in SB to some edge in Bi−1. We will refer to the step of the construction described
above. As noted above, the fundamental circuit of e in B contains all of C i. Thus e is joined
by an edge of SB to every element of C i. If P is empty, then the fundamental circuit of e also
includes an element of Bi−1 and we are done. Otherwise, as noted above, the fundamental
circuit of f in B contains all of P together with at least one element of C i and at least one
element of Bi−1. Thus f is joined by an edge of SB to every element of P , and to at least
one element of C i and to at least one element of Bi−1. Therefore, in SB every element of Bi
is within distance 4 of some element of Bi−1.
r) can be at most O(
√
r), there can be at most O(
11
4 Conclusion
We studied the reconfiguration of labelled bases of a rank r matroid and provided an upper
bound of O(r log r) on the worst-case reconfiguration distance for graphic matroids, and a
bound of O(r1.5) for general matroids. The obvious next question is whether this is tight.
The only lower bound we have so far is Ω(r).
Another natural question is to find the minimum number of basis exchange steps needed
to transform one given labelled basis to another. It an open question whether this problem
is NP-hard or polynomial-time solvable.
Acknowledgements
We could not have proved these results without major help from Jim Geelen. The results
first appeared in the second author's PhD thesis [9]. Theorem 1 was proved jointly with
Sander Vendonschot and Prosenjit Bose.
References
[1] P. Bose and F. Hurtado. Flips in planar graphs. Computational Geometry: Theory and
Applications 42:60 -- 80, 2009, doi:10.1016/j.comgeo.2008.04.001.
[2] P. Bose, A. Lubiw, V. Pathak, and S. Verdonschot. Flipping edge-labelled triangula-
tions. CoRR, 2013, arXiv:1310.1166v2.
[3] C. Hernando, F. Hurtado, M. Mora, and E. Rivera-Campo. Grafos de ´arboles etique-
tados y grafos de ´arboles geom´etricos etiquetados. Proc. X Encuentros de Geometra
Computacional, pp. 13-19, 2003.
[4] J. van den Heuvel. The complexity of change. Surveys in Combinatorics, vol. 409,
pp. 127 -- 160. Cambridge University Press, 2013.
[5] T. Ito, E. D. Demaine, N. J. A. Harvey, C. H. Papadimitriou, M. Sideri, R. Uehara, and
Y. Uno. On the complexity of reconfiguration problems. Theoretical Computer Science
412(12):1054 -- 1065, 2011, doi:10.1016/j.tcs.2010.12.005.
[6] N. McMurray, T. J. Reid, B. Wei, and H. Wu. Largest circuits in matroids. Advances
in Applied Mathematics 34(1):213 -- 216, 2005, doi:10.1016/j.aam.2004.09.002.
[7] W. Mulzer and G. Rote. Minimum-weight triangulation is NP-hard. J. ACM 55(2):11:1 --
11:29, 2008, doi:10.1145/1346330.1346336.
[8] J. Oxley. Matroid Theory, second edition. Oxford University Press, 2011.
[9] V. Pathak. Reconfiguring Triangulations. Ph.D. thesis, University of Waterloo, 2014.
12
[10] A. Schrijver. Combinatorial Optimization: Polyhedra and Efficiency. Springer, 2002.
[11] W. T. Tutte. Menger's theorem for matroids. J. Res. Nat. Bur. Standards Sect. B
69:49 -- 53, 1965.
13
|
1802.09333 | 1 | 1802 | 2018-02-20T04:06:33 | The Cut and Dominating Set Problem in A Steganographer Network | [
"cs.DS",
"cs.MM"
] | A steganographer network corresponds to a graphic structure that the involved vertices (or called nodes) denote social entities such as the data encoders and data decoders, and the associated edges represent any real communicable channels or other social links that could be utilized for steganography. Unlike traditional steganographic algorithms, a steganographer network models steganographic communication by an abstract way such that the concerned underlying characteristics of steganography are quantized as analyzable parameters in the network. In this paper, we will analyze two problems in a steganographer network. The first problem is a passive attack to a steganographer network where a network monitor has collected a list of suspicious vertices corresponding to the data encoders or decoders. The network monitor expects to break (disconnect) the steganographic communication down between the suspicious vertices while keeping the cost as low as possible. The second one relates to determining a set of vertices corresponding to the data encoders (senders) such that all vertices can share a message by neighbors. We point that, the two problems are equivalent to the minimum cut problem and the minimum-weight dominating set problem. | cs.DS | cs | The Cut and Dominating Set Problem in A
Steganographer Network
Hanzhou Wu1†, Wei Wang1‡, Jing Dong1‡, Hongxia Wang2 and Lizhi Xiong3
1Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
2School of Inf. Sci. & Technol., Southwest Jiaotong U., Chengdu 611756, China
3School of Comp. & Softw., Nanjing U. of Inf. Sci. & Technol., Nanjing 210044, China
†
[email protected],
‡{wwang,jdong}@nlpr.ia.ac.cn
8
1
0
2
b
e
F
0
2
]
S
D
.
s
c
[
1
v
3
3
3
9
0
.
2
0
8
1
:
v
i
X
r
a
Abstract. A steganographer network corresponds to a graphic struc-
ture that the involved vertices (or called nodes) denote social entities
such as the data encoders and data decoders, and the associated edges
represent any real communicable channels or other social links that could
be utilized for steganography. Unlike traditional steganographic algo-
rithms, a steganographer network models steganographic communication
by an abstract way such that the concerned underlying characteristics of
steganography are quantized as analyzable parameters in the network.
In this paper, we will analyze two problems in a steganographer network.
The first problem is a passive attack to a steganographer network where
a network monitor has collected a list of suspicious vertices correspond-
ing to the data encoders or decoders. The network monitor expects to
break (disconnect) the steganographic communication down between the
suspicious vertices while keeping the cost as low as possible. The second
one relates to determining a set of vertices corresponding to the data
encoders (senders) such that all vertices can share a message by neigh-
bors. We point that, the two problems are equivalent to the minimum
cut problem and the minimum-weight dominating set problem.
Keywords: Steganographer network, steganography, steganalysis, cut,
dominating set, graph, social networks.
1
Introduction
Steganography [1] has been extensively studied in past twenty years. As a means
to secret communication, different from cryptography, steganography refers to
the art of hiding a message into an innocent digital object also called cover by
slightly modifying the noise-like components of the host. The resulting object
also called stego will not introduce any noticeable artifacts and will be sent to
the receiver. As steganography even conceals the presence of communication, it
serves an important role in information security nowadays.
A number of advanced steganographic algorithms and novel perspectives have
been reported in the literature such as [2], [3], [4], [5], [6], [7], [8], [9], [10]. A most
important requirement [8] for any reliable steganographic system is that it should
2
H. Wu et al.
be impossible (or say very hard) for an eavesdropper to distinguish between
ordinary objects and objects containing hidden information. The mainstream
design concept of steganography is to minimize the data embedding impact on
the cover for a required payload, which involves at least two aspects for achieving
superior steganographic performance. The first is to find the most suitable cover
elements for data embedding. For example, complex texture areas within an
image would be quite desirable for steganography. The second is to maximize
data embedding efficiency. For the former, nowadays, we often focus on designing
a suited function for the cover elements that can expose the data embedding
impact. For the latter, cover codes are the key technologies.
Steganography can be also utilized for malicious purposes, which gives the
life of another important research topic called steganalysis. A core work in ste-
ganalysis is to detect whether a given digit object was embedded with a message
or not. The conventional steganalysis algorithms regard this problem as a bi-
nary classification task, which involves feature engineering and classifiers. Due
to the advancement of steganography that tends to alter the cover regions that
are quite hard to detect, conventional featured-based steganalysis algorithms re-
lying on sophisticated manual feature design has been pushed to the limit and
become very hard to improve. To overcome this difficulty, in-deep studies can
be performed on moving the success achieved by deep convolutional neural net-
works (deep CNNs) in computer vision to steganalysis [11], [12], [13], [14]. It
is pointed that, estimating the payload size within a stego, extracting the em-
bedded information and identifying steganographic senders/receivers are all the
important yet quite challenging research topics in steganalysis.
As mentioned in [15], with the rapidly development of social media services
such as Facebook and Twitter, it would be quite desirable to study the social
behaviors, protocols and any other scenarios of steganography. In this sense,
we may not care about the details of the used steganographic algorithms, but
rather quantize the concerned characteristics of steganography as analyzable
parameters for problem optimization. An effective way to analyze steganographic
activities is to model it in a social network, which consists of a set of vertices and
a set of edges. The vertices represent social entities such as the data encoders
and data decoders, and the associated edges represent any real communicable
channels or other social links used for steganography. In [15], the authors model
steganography in an additive-risk social network and present a communication
strategy that aims to minimize the steganographic risk by choosing a set of edges
with the minimum sum of weights.
In this paper, we present a passive attack to steganography in a steganogra-
pher network. We assume that the attacker serves as a network monitor, and has
collected a list of suspicious vertices that are corresponding to the data encoders
and data decoders in advance. Here, the "suspicious vertices" mean that they are
with a high probability representing the steganographic actors though they may
be misjudged in practice. One may use the conventional steganalysis algorithms
or other anomaly detection algorithms to construct the suspicious vertex-set.
We study such a passive attack that, the network monitor hopes to remove a set
The Cut and Dominating Set Problem in A Steganographer Network
3
Fig. 1. An example for steganographer network where the weights represent the cor-
responding removel costs.
of edges so that the suspicious vertices cannot communicate via steganography
while the removal cost could be minimized. In addition to the above-mentioned
attack, we also analyze a new communication scenario in the steganographer net-
work that all the vertices will serve as either a data encoder or a data decoder.
The optimization task is to select a set of vertices (encoders) out such that all
vertices can share a message by edges and the steganographic communication
activities are performed between adjacent vertices in the network. Similarly, we
expect to keep the vertex-selection risk as low as possible.
The rest of this paper are organized as follows. The related work is presented
in Section 2. In Section 3, we introduce the proposed passive attack. We analyze
a new communication scenario in a steganographer network in Section 4. Finally,
we conclude this paper in Section 5.
2 Related Work
A steganographer network corresponds to a graph G(V, E), where V = {v1, v2,
..., vn} represents the set of vertices and E = {e1, e2, ..., em} represents the set of
edges. Every edge ek ∈ E involves with a pair of vertices, i.e., ek = (vi, vj). We
say G is undirected meaning that all e ∈ E have no orientation. Namely, (vi, vj)
is equivalent to (vj, vi). A path (if any) between vi and vj corresponds to such
a vertex sequence (vq1, vq2 , ..., vqt) that vq1 = vi, vqt = vj and (vqk−1, vqk ) ∈ E
for all 2 ≤ k ≤ t. G is connected if and only if for any two vertices in G, there
exists at least one path between them.
In G, a path between vi and vj implies that, vi can share a message with vj
along the path. In [15], the authors present a communication strategy that aims
to minimize the overall steganographic risk by determining a subtree supporting
the required vertices. Mathematically, they use S = {s1, s2, ..., sn1} ⊂ V and
4
H. Wu et al.
Fig. 2. Two ways to remove a set of edges making S and T unconnected where S = {v1}
and T = {v5, v7, v8}.
T = {t1, t2, ..., tn2} ⊂ V to denote the encoder set and decoder set. For each
si ∈ S, its individual decoder set is denoted by Ti ⊂ T , meaning that, si hopes
to send a message to each of Ti. Here, ∪n1
i=1Ti = T . The optimization task is
to find such a subset of E that it enables all si ∈ S to send a message to all
decoders in Ti, and the overall risk can be minimized. Namely,
Eopt(S, T ) = arg min
Eusable⊂E
R(S, T, Eusable),
(1)
where R(S, T, Eusable) denotes the risk over Eusable ⊂ E.
By assuming an additive-risk network, the optimization task is defined as:
(cid:88)
Eopt(S, T ) = arg min
Eusable⊂E
ei∈Eusable
wi,
(2)
where wi > 0 is a real number representing the risk of ei.
The authors in [15] have pointed that Eq. (2) is equivalent to solving the
Steiner Tree Problem (STP) problem [16], which is NP-hard. The STP is seen
as a generalization of two other combinatorial optimization problems, i.e., the
shortest path problem and the minimum spanning tree (MST) problem. If the
STP deals with only two terminals, it reduces to finding a shortest path. If all
vertices are terminals, it is equivalent to the MST problem. Therefore, one may
employ the existing approximation algorithms designed for the STP to find the
suitable strategy for steganographic communication.
The authors also analyze the steganographic communication from a proba-
bilistic graph perspective. They have proved that, a multiplicative probabilistic
graph is equivalent to an additive weighted graph, meaning that, by translat-
ing the probabilistic network into a weighted network, one could find the suited
communication strategy in a similar way.
The Cut and Dominating Set Problem in A Steganographer Network
5
Fig. 3. Vertex insertion and edge insertion in a steganographer network.
3 Passive Attack to Steganographer Network
From a monitor point, even if a steganographer does not use the optimized
communication strategy, he can still communicate a message with the desired
receiver since there may exist multiple communicable paths in the network. This
requires the monitor to remove a set of edges corresponding to the channels
such that the steganographer cannot communicate with the receiver. However,
edge removal may result in a high cost or risk. We study this problem in this
section. Mathematically, we expect to remove a set of edges Eopt such that S
cannot communicate with T via steganography, while the removal operation
could result in the lowest risk (cost), i.e.,
Eopt(S, T ) = arg min
Erem⊂E
R(S, T, Erem).
(3)
We assign a weight wi > 0 to ei ∈ E. wi represents the cost of removing ei
from G. We assume an additive G, i.e.,
R(S, T, Erem) =
(cid:88)
ei∈Erem
wi,
(4)
which makes the problem amenable to mathematical analysis and the additive
assumption is reasonable since two edges far away to each other have ignorable
interaction and interaction between edges near to each other can be scored by
weights.
In the real-world, for the network monitor, it is very likely that he cannot
identify Ti for si, meaning that, he may aim to disconnect the links between S
and T , rather than that between si and Ti for all 1 ≤ i ≤ n1. This assumption has
been utilized in this section. We define the simplest attack for which S = 1 as
single encoder attack. We expect to find a subset of E such that there has no path
between s1 and each vertex in T by removing the selected edges. And, the sum of
weights should be minimum. Fig. 1 shows a steganographer network. Assuming
6
H. Wu et al.
that, S = {v1} and T = {v5, v7, v8}. Fig. 2 provides two ways to divide S and T
into two different connected components. It is seen that, different sets of edges
corresponds to different costs, e.g., the cost for Fig. 2 (a) is 13 + 17 + 31 = 61,
and that for Fig. 2 (b) is 27. We should find a method to separate S and T with
the lowest cost. Notice that, we assume that S ∩ T = ∅.
In graph theory, a cut is defined as a partition of the vertices of a graph into
two disjoint subsets. Specifically, a cut C = (A, B) divides V into two subsets A
and B. The cut-set of C corresponds to an edge-set {(u, v) ∈ Eu ∈ A, v ∈ B}.
Notice that, A∩ B = ∅ and A∪ B = V . If a ∈ V and b ∈ V are specified vertices
of G, we then say a-b cut is a cut in which a ∈ A and b ∈ B. In a weighted
graph, the weight of a cut is defined as the sum of weights of edges belonging to
the cut. A cut is therefore called as minimum cut if its weight is not larger than
that of any other cut. Obviously, the minimum cut is not unique.
A directed graph is called as a flow network if each edge has a capacity and
receives a flow. A flow in a directed graph must satisfy the restriction that the
amount of flow into a vertex is equal to the amount of flow out of the vertex,
unless it is the source vertex which has only outgoing flow, or sink vertex which
has only incoming flow. Mathematically, the capacity of a directed edge vi → vj
is a mapping c : E (cid:55)→ R+, denoted by c(vi, vj). It represents the maximum
amount of flow that can pass through vi → vj. It is noted that, c(vj, vi) is
different from c(vi, vj). If vi → vj /∈ E, one may consider c(vi, vj) = 0. A flow is
a mapping f : E (cid:55)→ R+ subject to two constraints [17]:
and
(cid:88)
vi→vj∈E
f (vi, vj) =
vj→vi∈E
f (vi, vj) ≤ c(vi, vj),∀ vi → vj ∈ E,
(cid:88)
f (vj, vi),∀ vi ∈ V \ {a, b},
(5)
(6)
where a is the source vertex and b is the sink vertex.
Without the loss of generality, an undirected graph can be treated as "di-
rected" since an undirected edge (vi, vj) ∈ E can be decomposed into two di-
rected edges vi → vj and vj → vi. For simplicity, we sometimes use vi → vj
to represent a directed edge from vi to vj and vj → vi from vj to vi though
(vi, vj) ∈ E is undirected in G in this paper. It also means that, vi → vj ∈ E
and vj → vi ∈ E.
With Eqs. (5, 6), the value of a flow is defined by f = (cid:80)
f =(cid:80)
v∈V f (a, v) or
v∈V f (v, b). It means the amount of flow from the source vertex to the
sink vertex. The maximum flow problem is therefore to maximize f, namely,
to route as much flow as possible from a to b. Notice that, if a → v /∈ E, one
may set f (a, v) = 0 ≤ c(a, v) = 0. A lot of practical algorithms can be applied
for determining the maximum flow such as [18], [19], [20]. For a steganographer
network, by regarding the weights as the flow capacity, we can determine the
maximum flow between any two vertices. Therefore, an undirected steganogra-
pher network can be translated as a flow network. A necessary preprocessing
is to decompose each edge into two directed edges, for which both the newly
assigned weights should equal the original one.
The Cut and Dominating Set Problem in A Steganographer Network
7
Fig. 4. Vertex contraction in a steganographer network: (a) vertex contraction and (b)
a new graph.
The max-flow min-cut theorem [17] states that,
Theorem 1. In a flow network, the maximum amount of flow passing through
the source vertex is equal to the weight of the minimum cut.
It indicates that, the smallest weight sum of the edges which if removed
would disconnect the source vertex from the sink vertex is equal to the weight
of the minimum cut, which is also equal to the maximum flow from the source
to the sink. In order to find Eopt(S, T ) shown in Eq. (3), we need to determine
the maximum flow in G(V, E). Then, we construct the minimum cut. The edges
in the minimum cut will constitute Eopt(S, T ). The maximum flow problem
generally deals with only one source vertex and one sink vertex. However, we
have n1 ≥ 1 or n2 ≥ 1, indicating that, we cannot directly use a maximum flow
algorithm in the steganographer network since there may be multiple source
vertices, i.e., s1, s2, ..., sn1, or multiple sink vertices, i.e., t1, t2, ..., tn2 .
A way to address this issue is to insert a new super-source vertex vs and a
super-sink vertex vt into G(V, E), i.e.,
(7)
For each si ∈ S, we insert an edge (vs, si) into E. Then, for each ti ∈ T , we
V = V ∪ {vs, vt}.
insert (ti, vt) into E. Therefore,
E = E ∪ {(vs, si)1 ≤ i ≤ n1} ∪ {(ti, vt)1 ≤ i ≤ n2}.
(8)
Thereafter, we assign an infinite large weight to each of the new edges to
represent the cost of removing it from the new graph. We take Fig. 3 for ex-
planation. In Fig. 3, we have S = {v1, v8} and T = {v3, v7, v9}. A super-source
vertex v10 (i.e., vs) and a super-sink vertex v11 (i.e., vt) are inserted. Five new
edges with the infinite large weight are inserted as well. In applications, one
could assign a very large weight (that is larger than the sum of weights of all
8
H. Wu et al.
original edges) to the new edges, instead of the infinite large weight. Accordingly,
a new graph can be built, as shown in Fig. (3). To find the maximum flow from
v10 to v11, we should decompose all edges shown in Fig. (3) into two directed
edges. By treating the weights of edges as the flow capacity, the maximum flow
from v10 to v11 can be determined as 61. And, the corresponding minimum cut
is {(v1, v3), (v1, v9), (v2, v9), (v8, v7)}. It means that, the minimum cost making
S and T unconnected is 61, and Eopt(S, T ) = {(v1, v3), (v1, v9), (v2, v9), (v8, v7)}.
Another way to determine Eopt(S, T ) is vertex contraction. The contraction
of a set of vertices produces a graph in which all vertices in the set are replaced
with a single vertex such that the single vertex is adjacent to the union of
the vertices to which the vertices in the set were originally adjacent. In vertex
contraction, it does not matter if two vertices in S or T are connected by an
edge, which (if any) is simply removed during contraction.
In detail, we initialize G(cid:48)(V (cid:48), E(cid:48)) as V (cid:48) = ∅ and E(cid:48) = ∅. For each (vi, vj) ∈ E,
we skip it if {vi, vj} ⊂ S or {vi, vj} ⊂ T . Otherwise, if vi /∈ S ∪ T and vj /∈ S ∪ T ,
we update V (cid:48) as V (cid:48) = V (cid:48) ∪ {vi, vj} and E(cid:48) as E(cid:48) = E(cid:48) ∪ {(vi, vj)}. The weights
assigned to edges should be processed as well. If vi ∈ S and vj ∈ T , we first
update V (cid:48) as V (cid:48) = V (cid:48) ∪ {vs, vt} and then update E(cid:48) = E(cid:48) ∪ {(vs, vt)}. The
weight of (vi, vj) will be added to (vs, vt). If vi ∈ S and vj /∈ T , we first update
V (cid:48) as V (cid:48) = V (cid:48) ∪ {vs, vj} and then update E(cid:48) = E(cid:48) ∪ {(vs, vj)}. The weight of
(vi, vj) will be added to (vs, vj). It is similar to process other cases. In this way,
a new graph can be finally constructed. Thereafter, by computing the maximum
flow from vs to vt in the new graph, the minimum cost can be obtained, and
the minimum cut can be constructed as well. Fig. 4 shows an example, in which
{v1, v8} and {v3, v7, v9} are replaced with vs and vt, respectively. It is seen that,
the maximum flow in the new graph is identical to the maximum flow shown
in Fig. 3, which has verified the correctness. Notice that, in Fig. 4 (b), after
determining out the minimum cut {(vs, vt), (v2, vt)} in G(cid:48)(V (cid:48), E(cid:48)). We have to
further construct the minimum cut for G(V, E). Clearly, it is observed that
(vs, vt) is corresponding to {(v1, v3), (v1, v9), (v8, v7)}, and (v2, vt) corresponds
to (v2, v9). Therefore, we have Eopt(S, T ) = {(v1, v3), (v1, v9), (v2, v9), (v8, v7)}.
Remark. For a given directed graph, one can determine the maximum flow
from the source vertex to the sink vertex by any efficient maximum flow algo-
rithm. After calling a maximum flow algorithm, the flow information passing
through each vertex can be identified. Notice that, the flow passing through an
edge will be not larger than its capacity. To determine the cut, one can find the
set of vertices that are reachable from the source vertex in the corresponding
residual graph [17]. All edges involving a reachable vertex and a non-reachable
vertex constitute the minimum cut. This can be completed by applying depth-
first search (DFS) technique with a time complexity of O(V + E).
4 Steganographic Strategy Between Neighbors
It seems to be interesting and reasonable that the steganographic communication
in a steganographer network is only available between adjacent vertices (or say
The Cut and Dominating Set Problem in A Steganographer Network
9
Fig. 5. Three examples of dividing all vertices into two sets S and T : a vertex marked
as "1" means it belongs to S and "0" for T .
between neighbors). This scenario looks closer to reality comparing with the case
that an encoder in G should participate in the path planning for steganographic
communication. The reason is that an edge between any two vertices does not
only represent the steganographic channel between them, but also, to a certain
extent, shows the social relationship between them. Thus, it may be not desirable
sometimes for a vertex to communicate a message along a predetermined network
path with another vertex that is not a neighbor. In this section, we will analyze
steganography between neighbors within the steganographer network.
Mathematically, for a given G(V, E), all vertices will serve as either a data
encoder or a data decoder. One may assume that, G(V, E) is a subnetwork
determined from a complex network in which the vertices may serve as more
complex roles. Therefore, it can be said that, we actually deal with a simplified
model in this paper even though the complex cases are not explicit for us at
present. We are to choose a set of vertices from V to constitute S mentioned
previously and the rest will constitute T , i.e., T = V \S. We define the neighbor-
set of v ∈ V as N (v) = {u(u, v) ∈ E}. It is required that, for any v ∈ T , there
should exist at least one u ∈ S such that u ∈ N (v). It ensures that, any vertex
in V can either hold a message by itself (in S) or receive a message directly from
a neighbor belonging to S. Accordingly, all vertices can share a message.
The determination of S corresponds to a binary network game [21]. Clearly,
each individual vertex v ∈ V must choose an action g(v) ∈ X = {0, 1}, where
action g(v) = 0 indicates it servers as a data decoder, and g(v) = 1 for a data
encoder. Therefore, it is required that
(cid:88)
u∈N (v)
∀v ∈ V,
g(u) ≥ 1 − g(v).
(9)
the number of data encoders, i.e.,(cid:80)
ing(cid:80)
Fig. 5 shows three examples of dividing V into two sets S and T . It is seen
that, there exist lots of legal solutions. An intuitive requirement is to minimize
v∈V g(v), subject to Eq. (9). Simply minimiz-
v∈V g(v), however, indicates that all vertices essentially have no difference
and have the same importance (or say that they have the same risk/cost). From
a generalized viewpoint, we can assign a positive weight w(v) to each v ∈ V to
10
H. Wu et al.
evaluate the cost or risk of marking v as a member of S. By assuming an additive
G(V, E), under the constraint of Eq. (9), our task is to minimize
g(v) · w(v),
(10)
(cid:88)
v∈V
which is a typical 0-1 integer linear programming (ILP) problem.
In graph theory, the family of vertex/edge covering problems involves a broad
range of NP-hard optimization problems, among which the minimum-weight
dominating set (MWDS) problem has played a prominent role in various real-
world domains such as social networks, wireless ad-hoc networks, communication
networks, and industrial applications [22]. For G(V, E), a dominating set D is
a subset of V such that each v ∈ V \ D is adjacent to at least one member of
D. The MWDS problem aims to find a dominating set Dmin that minimizes the
total positive weights assigned to the vertices in the dominating set, namely
(cid:88)
v∈D
Dmin = arg min
D⊂V
w(v),
(11)
(12)
subject to
∀u ∈ V \ D, N (u) ∩ D (cid:54)= ∅.
Obviously, the ILP problem shown in Eqs. (9, 10) is equivalent to the MWDS
problem. Since the MWDS problem is known as a NP-hard problem, generally
we have to use approximation algorithms to find the near-optimal solution unless
all NP problems can be effectively solved [23] or G(V, E) has a small size (or has
some special topological structure). For steganography between neighbors in G,
a core research is therefore to design effective approximation algorithms for the
MWDS problem.
5 Conclusion and Discussion
In this paper, we introduce two simplified optimization problems in a steganog-
rapher network theoretically. Both problems are proven to be equivalent to the
minimum cut problem (or say the maximum flow problem since they are dual
to each other) and the minimum-weight dominating set problem, respectively.
The optimal/near-optimal solutions to the corresponding problems can be found
by exploiting graph-modification techniques (e.g., vertex contraction) as well as
related deterministic/approximation algorithms used in graph theory.
For the passive attack, to identify the suspicious vertices, one may use ste-
ganalysis algorithms or anomaly detection algorithms designed for complex net-
works. On the one hand, these approaches provide the attacker access to finding
both suspicious vertices and edges (corresponding to suspicious channels). On
the other hand, the suspicious edges will help the attacker to evaluate the cost
of removing an edge from the steganographer network. Notice that, the cost
of removing an edge may take into account the local topological characteristics
since a steganographer network may contain community features. For example,
The Cut and Dominating Set Problem in A Steganographer Network
11
the edges connecting two communities may be assigned with a high cost. The
suited definition of cost of removing an edge will be the future work.
Another reasonable explanation for steganography between neighbors is that,
a vertex may have not complete information on the steganographer network. It
may lead each vertex to send a message via steganography to its neighbors or
receive a message via steganography directly from its neighbors. We propose to
find the MWDS in the steganographer network, which, however, has implied that
the vertices know the complete information about the network or there should
exist an ideal "super-vertex " that can access the whole network structure and
can communicate with each vertex in the network. From the viewpoint of game
theory, in case that each vertex has incomplete information about the whole
network, the determination of the suited dominating set is more difficult since a
vertex may only use its local information and may be affected by its neighbors'
actions. Moreover, the determination of weights assigned to vertices in practice
is not explicit to us either. We will focus on this problem in future.
Acknowledgement
This work was supported by the National Natural Science Foundation of China
under Grant Nos. 61502496, U1536120, and U1636201, and the National Key Re-
search and Development Program of China under Grant No. 2016YFB1001003.
References
1. Cox, I., Miller, M., Bloom, J., Fridrich, J., Kalker, T.: Digital watermarking and
steganography. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (2008)
2. Holub, V., Fridrich, J., Denemark, T.: Universal distortion function for steganogra-
phy in an arbitrary domain. EURASIP J. Inf. Security, 2014(1): 1-13 (2014)
3. Guo, L., Ni, J., Su, W., Tang C., Shi, Y.: Using statistical image model for JPEG
steganography: uniform embedding revisited. IEEE Trans. Inf. Forensics Security,
10(12): 2669-2680 (2015)
4. Li, B., Wang, M., Li, X., Tan, S., Huang, J.: A strategy of clustering modification
directions in spatial image steganography. IEEE Trans. Inf. Forensics Security, 10(9):
1905-1917 (2015)
5. Tang, W., Tan, S., Li, B., Huang, J.: Automatic steganographic distortion learning
using a generative adversarial network. IEEE Signal Process. Lett., 24(10): 1547-
1551 (2017)
6. Zhou., W., Zhang, W., Yu, N.: A new rule for cost reassignment in adaptive
steganography. IEEE Trans. Inf. Forensics Security, 12(11): 2654-2667 (2017)
7. Wu, H., Wang, H.: Multibit color-mapping steganography using depth-first search.
In: Proc. IEEE Int. Symp. Biometrics Security Technol., pp. 224-229 (2013)
8. Wu, H., Wang, H., Zhao, H., Yu, X.: Multi-layer assignment steganography using
graph-theoretic approach. Multimed. Tools Appl., 74(18): 8171-8196 (2015)
9. Filler, T., Judas, J., Fridrich, J.: Minimizing additive distortion in steganography us-
ing syndrome-trellis codes. IEEE Trans. Inf. Forensics Security, 6(3): 920-935 (2011)
12
H. Wu et al.
10. Ma, S., Zhao, X., Guan, Q., Zhao, C.: The a priori knowledge based secure pay-
load estimation for additive model. In: IS&T Int. Symp. Electronic Imaging: Media
Watermarking, Security, and Forensics, pp. 16-21 (2017)
11. Xu, G., Wu, H., Shi, Y.: Structural design of convolutional neural networks for
steganalysis. IEEE Signal Process. Lett., 23(5): 708-712 (2016)
12. Xu, G., Wu, H., Shi, Y.: Ensemble of CNNs for steganalysis: an empirical study.
In: Proc. ACM Workshop Inf. Hiding Multimed. Security, pp. 103-107 (2016)
13. Xu, G.: Deep convolutional neural network to detect J-UNIWARD. In: Proc. ACM
Workshop Inf. Hiding & Multimed. Security, pp. 67-73 (2017)
14. Ye, J., Ni, J., Yi, Y.: Deep learning hierarchical representations for image steganal-
ysis. IEEE Trans. Inf. Forensics Security, 12(11): 2545-2557 (2017)
15. Wu, H., Wang, W., Dong, J., Xiong, Y., Wang, H.: A simply study to steganography
on social networks. In: Proc. China Inf. Hiding Workshop, to appear (2018) (online
availble: arXiv:1712.03621, https://arxiv.org/abs/1712.03621)
16. Chlebik, M., Chlebikova, J.: The Steiner tree problem on graphs: inapproximability
results. Theoret. Comput. Sci., 406(3): 207-214 (2008)
17. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to algorithms.
The MIT Press, Cambridge (2009)
18. Ford, L.R., Fulkerson, D.R.: Maximal flow through a network. Classic Papers in
Combinatorics, pp. 243-248 (2009)
19. Edmonds, J., Karp, R.M.: Theoretical improvements in algorithmic efficiency for
network flow problems. J. ACM, 19(2): 248-264 (1972)
20. Dinic, E.A.: Algorithm for solution of a problem of maximum flow in a network
with power estimation. Doklady Akademii Nauk SSSR, 11(5): 1277-1280 (1970)
21. Galeotti, A., Goyal, S., Jackson, M.O., Vega-Redondo, F., Yariv, L.: Network
games. The Review of Economic Studies, 77(1): 218-244 (2010)
22. Wang, Y., Cai, S., Yin, M.: Local search for minimum weight dominating set with
two-level configuration checking and frequency based scoring function. J. Artificial
Intelligence Research, 58(2017): 267-295 (2017)
23. Feige, U.: A threshold of ln(n) for approximating set cover. J. ACM, 45(4): 634-652
(1998)
|
1611.00665 | 1 | 1611 | 2016-11-02T16:09:26 | Combinatorial Prophet Inequalities | [
"cs.DS"
] | We introduce a novel framework of Prophet Inequalities for combinatorial valuation functions. For a (non-monotone) submodular objective function over an arbitrary matroid feasibility constraint, we give an $O(1)$-competitive algorithm. For a monotone subadditive objective function over an arbitrary downward-closed feasibility constraint, we give an $O(\log n \log^2 r)$-competitive algorithm (where $r$ is the cardinality of the largest feasible subset).
Inspired by the proof of our subadditive prophet inequality, we also obtain an $O(\log n \cdot \log^2 r)$-competitive algorithm for the Secretary Problem with a monotone subadditive objective function subject to an arbitrary downward-closed feasibility constraint. Even for the special case of a cardinality feasibility constraint, our algorithm circumvents an $\Omega(\sqrt{n})$ lower bound by Bateni, Hajiaghayi, and Zadimoghaddam \cite{BHZ13-submodular-secretary_original} in a restricted query model.
En route to our submodular prophet inequality, we prove a technical result of independent interest: we show a variant of the Correlation Gap Lemma for non-monotone submodular functions. | cs.DS | cs |
Combinatorial Prophet Inequalities
Aviad Rubinstein∗
Sahil Singla†
November 3, 2016
Abstract
We introduce a novel framework of Prophet Inequalities for combinatorial valuation func-
tions. For a (non-monotone) submodular objective function over an arbitrary matroid
feasibility constraint, we give an O(1)-competitive algorithm. For a monotone subaddi-
tive objective function over an arbitrary downward-closed feasibility constraint, we give
an O(log n log2 r)-competitive algorithm (where r is the cardinality of the largest feasible
subset).
Inspired by the proof of our subadditive prophet inequality, we also obtain an O(log n·
log2 r)-competitive algorithm for the Secretary Problem with a monotone subadditive
objective function subject to an arbitrary downward-closed feasibility constraint. Even
for the special case of a cardinality feasibility constraint, our algorithm circumvents an
Ω(√n) lower bound by Bateni, Hajiaghayi, and Zadimoghaddam [BHZ13] in a restricted
query model.
En route to our submodular prophet inequality, we prove a technical result of inde-
pendent interest: we show a variant of the Correlation Gap Lemma [CCPV07, ADSY12]
for non-monotone submodular functions.
1
Introduction
The Prophet Inequality and Secretary Problem are classical problems in stopping theory. In
both problems a decision maker must choose one of n items arriving in an online fashion. In
the Prophet Inequality, each item is drawn independently from a known distribution, but the
order of arrival is chosen adversarially. In the Secretary Problem, the decision maker has no
prior information about the set of items to arrive (except their cardinality, n), but the items
are guaranteed to arrive in a uniformly random order.
Historically, there are many parallels between the research of those two problems. The
classic (single item) variants of both problems were resolved a long time ago: in 1963 Dynkin
gave a tight e-competitive algorithm for the Secretary Problem [Dyn63]; a little over a decade
later Krengel and Sucheston [KS77] gave a tight 2-competitive algorithm for the Prophet
Inequality. Motivated in part by applications to mechanism design, multiple-choice variants
of both problems have been widely studied for the past decade in the online algorithms com-
munity. Instead of one item, the decision maker is restricted to selecting a feasible subset of
∗([email protected]) Electrical Engineering and Computer Sciences, UC Berkeley, Berkeley, CA.
This work was supported by Microsoft Research PhD Fellowship, NSF grant CCF1408635, and Templeton
Foundation grant 3966.
†([email protected]) Computer Science Department, Carnegie Mellon University, Pittsburgh, PA. This work
was supported by CMU Presidential Fellowship and NSF awards CCF-1319811, CCF-1536002, and CCF-
1617790
1
the items. The seminal papers of [HKP04, Kle05] introduced a secretary problem subject to
a cardinality constraint (and [Kle05] also obtained a 1− O(1/√r)-competitive algorithm). In
2007, Hajiaghayi et al. followed with a prophet inequality subject to cardinality constraint
[HKS07]. In the same year, Babaioff et al. introduced the famous matroid secretary problem
[BIK07]; in 2012 Kleinberg and Weinberg introduced (and solved!) the analogous matroid
prophet inequality. For general downward-closed constraints, O(log n log r)-competitive algo-
rithms were recently obtained for both the problems [Rub16].
In all the works mentioned in the previous paragraph, the goal is to maximize the sum
of selected items' values, i.e. an additive objective is optimized. For the secretary problem,
there has also been significant work on optimizing more general, combinatorial objective
functions. A line of great works [BHZ13, FNS11, BUCM12, FZ15] on secretary problem
with submodular valuations culminated with a general reduction by Feldman and Zenklusen
[FZ15] from any submodular function to additive (linear) valuations with only O(1) loss.
Going beyond submodular is an important problem [FI15], but for subadditive objective
functions there is a daunting Ω(√n) lower bound on the competitive ratio for restricted value
queries [BHZ13].
Surprisingly, this line of work on combinatorial secretary problems has seen no parallels
In this work we break the ice by introducing a new
in the world of prophet inequalities.
framework of combinatorial prophet inequalities.
Combinatorial Prophet Inequalities: Our main conceptual contribution is a general-
ization of the Prophet Inequality setting to combinatorial valuations. Roughly, on each of n
days, the decision maker knows an independent prior distribution over k potential items that
could appear.1 She also has access to a combinatorial (in particular, submodular or monotone
subadditive) function f that describes the value of any subset of the n· k items (see Section 2
for a formal definition and further discussion). We obtain the following combinatorial prophet
inequalities:
Theorem 1.1 (Submodular Prophet; informal). There exists an efficient randomized O(1)-
competitive algorithm for (non-monotone) submodular prophet over any matroid.
Theorem 1.2 (Monotone Subadditive Prophet; informal). There exists an O(log n · log2 r)-
competitive algorithm for monotone subadditive prophet inequality subject to any downward-
closed constraints family.
Subadditive Secretary Problem: Building on the techniques of our subadditive prophet
inequality, we go back to the secretary world and prove a (computationally inefficient) O(log n·
log2 r)-competitive algorithm for the subadditive secretary problem subject to any downward-
closed feasibility constraint. As noted earlier, this algorithm circumvents the impossibility
result of Bateni et al. [BHZ13] for efficient algorithms2.
Theorem 1.3 (Monotone Subadditive Secretary; informal). There exists an O(log n· log2 r)-
competitive algorithm for monotone subadditive secretaries subject to any downward-closed
constraints family.
1Note that some notion of independence assumption is necessary as even for the single choice problem, if
values are arbitrarily correlated then every online algorithm is Ω(n) competitive [HK92].
2In fact, for general downward-closed constraint, even with additive valuations one should not expect
efficient algorithms with membership queries [Rub16].
2
Non-monotone correlation gap: En route to Theorem 1.1, we prove a technical contribu-
tion that is of independent interest: a constant correlation gap for non-monotone submodular
functions. For a monotone submodular function f , [CCPV07] showed that the expected value
of f over any distribution of subsets is at most a constant factor larger than the expectation
over subsets drawn from the product distribution with the same marginals. This bound on
the correlation gap has been very useful in the past decade with applications in optimization
[CVZ14], mechanism design [ADSY12, Yan11, BCK12, BH16], influence in social networks
[RSS15, BPR+16], and recommendation systems [KSS13].
It turns out (see Example 4.1) that when f is non-monotone, the correlation gap is
unbounded, even for n = 2! Instead, we prove a correlation gap for a related function:
fmax(S) , max
T ⊆S
f (T ).
(Note that fmax is monotone, but may not be submodular.)
Theorem 1.4 (Non-monotone correlation gap; informal). For any (non-monotone) submod-
ular function f , the function fmax has a correlation gap of O(1).
1.1 Further related work
Secretary Problem In recent years, following the seminal work of Babaioff, Immorlica,
and Kleinebrg [BIK07], there has been extensive work on the matroid secretary problem, where
the objective is to maximize the sum of values of secretaries subject to a matroid constraint.
For general matroids, there has been sequence of improving competitive ratio with the state of
the art being O(log log r) [Lac14, FSZ15]. Obtaining a constant competitive ratio for general
matroids remains a central open problem, but constant bounds are known for many special
cases (see survey by Dinitz [Din13] and references therein). Other variants have also been
considered, such as a hiring that returns for a second (or potentially k-th) interview [Var15],
or secretaries whose order of arrival has exceptionally low entropy [KKN15]. Of special in-
terest to us is a recent paper by Feldman and Izsak [FI15] that considers matroid secretary
problems with general monotone objective functions, parametrized by the supermodular de-
gree. For objective f with supermodular degree D+
f and general matroid constraint, they
obtain a competitive ratio of O(cid:16)D+
2
f + D+
log r(cid:17). Their results are incomparable
to our Theorem 1.3, but the motivation is related-obtaining secretary algorithms beyond
submodular functions.
f
3
f
log D+
Prophet Inequality The connection between multiple-choice prophet inequalities and
mechanism design was recognized in the seminal paper by Hajiaghayi et al. [HKS07].
In
particular, they proved a prophet inequality for uniform matroids; their bound was later im-
proved by Alaei [Ala11]. Chawla et al.
further developed the connection between prophet
inequalities and mechanism design, and proved, for general matroids, a variant of the prophet
inequality where the algorithm may choose the order in which the items are viewed. The ma-
troid prophet inequality was first explicitly formulated by Kleinberg and Weinberg [KW12],
who also gave a tight 2-competitive algorithm.
In a different direction, Alaei, Hajiaghayi,
and Liaghat [AHL12] considered a variant they call prophet-inequality matching, which is
useful for online ad allocation. More generally, for intersection of a constant number of
matroid, knapsack, and matching constraints, Feldman, Svensson, and Zenklusen [FSZ16]
gave an O(1)-competitive algorithm; this is a corollary of their online contention reslution
3
schemes (OCRS), which we also use heavily (see Section 3.1.2). Azar, Kleinberg, and Wein-
berg [AKW14] considered a limited information variant where the algorithm only has access
to samples from each day's distributions. Esfandiari et al. [EHLM15] considered a mixed
notion of "Prophet Secretary" where the items arrive in a uniformly random order and draw
their values from known independent distributions. Finally, for general downward-closed
constraint, [Rub16] gave O(log n log r)-competitive algorithms for both Prophet Inequality
and Secretary Problem; these algorithms are the basis of our algorithms for the respective
subadditive problems.
Other notions of online submodular optimization Online submodular optimization
has been studied in contexts beyond secretary. In online submodular welfare maximization,
there are m items, n people, and each person has a monotone submodular value function.
Given the value functions, the items are revealed one-by-one and the problem is to immedi-
ately and irrevocably allocate it to a person, while trying to maximize the sum of all the value
functions (welfare). The greedy strategy is already half competitive. Kapralov et al. [KPV13]
showed that for adversarial arrival greedy is the best possible in general (competitive ratio of
1/2), but under a "large capacities" assumption, a primal-dual algorithm can obtain 1− 1/e-
competitive ratio [DHK+13]. For random arrival Korula et al. [KMZ15] showed that greedy
can beat half; obtaining 1 − 1/e in this settings remains open.
Buchbinder et al. [BFS15] considered the problem of (monotone) submodular maximiza-
tion with preemption, when the items are revealed in an adversarial order. Since sublinear
competitive ratio is not possible in general with adversarial order, they consider a relaxed
model where we are allowed to drop items (preemption) and give constant-competitive algo-
rithms. Submodular maximization has also been studied in the streaming setting, where we
have space constraints but are again allowed to drop items [BMKK14, CK15, CGQ15].
The "learning community" has looked into experts and bandits settings for submodular
optimization. In these settings, different submodular functions arrive one-by-one and the al-
gorithm, which is trying to minimize/ maximize its value, has to select a set before seeing the
function. The function is then revealed and the algorithm gets the value for the selected set.
The goal is to perform as close as possible to the best fixed set in hindsight. Since submod-
ular minimization can be reduced to convex function minimization using Lov´asz extension,
sublinear regrets are possible [HK12]. For submodular maximization, the usual benchmark is
a 1 − 1/e multiplicative loss and an additive regret [SG08, GK10, GKS14].
subadditive functions.
Interestingly, to the best of our knowledge none of those problems have been studied for
1.2 Organization
We begin by defining our model for combinatorial prophet inequalities in Section 2; in Sec-
tion 3 we develop some necessary notation and recall known results; in Section 4 we formalize
and prove our correlation gap for non-monotone submodular functions; and in Section 5 we
prove the submodular prophet inequality.
The subadditive prophet and secretary algorithms share the following high level approach:
use a lemma of Dobzinski [Dob07] to reduce subadditive to XOS objective functions; then solve
the XOS case using the respective (prophet and secretary) algorithms of [Rub16] for additive
objective function and general downward closed feasibility constraint. It turns out that the
secretary case is much simpler since we can use the additive downward-closed algorithm of
[Rub16] as black-box; we present it in Section 6. For the prophet inequality, we have to make
4
changes to the already highly non-trivial algorithm of [Rub16] for the additive, downward-
closed case; we defer this proof to Section 7.
2 Combinatorial but Independent Functions
In this section, we define and motivate what we mean by "submodular (or subadditive)
valuations over independent items". Recall that in the classic prophet inequality, a gambler is
asked to choose one of n independent non-negative random payoffs. In multiple choice prophet
inequalities [HKS07], the gambler chooses multiple payoffs, subject to some known-in-advance
feasibility constraint, and receives their sum. Here, we are interested in the case where the
gambler's utility is not additive over the outcomes of the random draws: For example, instead
of monetary payoffs, at each time period the gambler can choose to receive a random item,
say a car, and the utility from owning multiple cars diminishes quickly. As another example,
the gambler receives monetary payoffs, but his marginal utility for the one-millionth dollar is
much smaller than for the first. Formally, we define:
Definition 2.1 (C Valuations over Independent Items). Let C be a class of valuation functions
(in particular, we are interested in C ∈ {submodular, monotone subadditive}). Consider:
• n sets U1, . . . , Un and distributions D1, . . . ,Dn, where Di returns a single item from Ui.
• A function f : {0, 1}U → R+, where U ,Sn
• For X ∈Qi Ui and subset S ⊆ [n], let vX(S) , f(cid:0){Xi : i ∈ S}(cid:1).
Let D be a distribution over valuation functions v (·) : 2n → R+. We say that D is "C
over independent items" if it can be written as the distribution that first samples X ∼×i Di,
and then outputs valuation function vX(·).
i=1 Ui and f ∈ C.
Related notions in the literature
Combinatorial functions over independent items have been considered before. Agrawal et
[ADSY12], for example, consider a different, incomparable framework defined via the
al.
generalization of submodular and monotone functions to non-binary, ordered domains (e.g.
f :×Ui → R+ is submodular if f (max{X, Y}) + f (min{X, Y}) ≤ f (X) + f (Y)). In our
definition, per contra, there is no natural way to define a full order over the set of items
potentially available on each time period. For example, if we select an item on Day 1, an item
X2 on Day 2 may have a larger marginal contribution than item Y2, but a lower contribution
if we did not select any item on Day 1.
Another relevant definition has been considered in probability theory [Sch99] and more
recently in mechanism design [RW15]. The latter paper considers auctioning n items to a
buyer that has a random, "independent" monotone subadditive valuation over the items. Now
the seller knows which items she is selling them, but different types of buyers may perceive
each item differently. This is captured via an attribute of an item, which describes how each
buyer values a bundle containing this item. Formally,
Definition 2.2 (Monotone Subadditive Valuations over Independent Items [Sch99, RW15]).
We say that a distribution D over valuation functions v (·) : 2n → R is subadditive over
independent items if:
5
1. All v (·) in the support of D exhibit no externalities.
Formally, let ΩS = ×i∈S Ωi, where each Ωi is a compact subset of a normed space.
There exists a distribution DX over Ω[n] and functions VS : ΩS → R such that D is the
distribution that first samples X ← DX and outputs the valuation function v (·) with
v (S) = VS (hXiii∈S) for all S.
2. All v (·) in the support of D are monotone and subadditive.
3. The private information is independent across items. That is, the DX guaranteed in
Property 1 is a product distribution.
For monotone valuations, Definition 2.1 is stronger than Definition 2.2 as it assumes that the
valuation function is defined over every subset of U =S Ui, rather than just the support of
D. However, it turns out that for monotone subadditive functions Definitions 2.1 and 2.2 are
equivalent.
Observation 1. A distribution D is subadditive over independent items according to Definition
2.1 if and only if it is subadditive over independent items according to Definition 2.2.
Proof sketch. It's easy to see that Definition 2.1 implies Definition 2.2: We can simply iden-
tify between the set of attributes Ωi on day i and the set of potential items Ui, and let
In the other direction, we again identify between each Ui and Ωi. For feasible set S which
consists of only one item in Ui for each i ∈ R, we can let XS be the corresponding vector in
VS (hXiii∈S) , f(cid:0){Xi : i ∈ S}(cid:1). Observe that the desiderata of Definition 2.2 are satisfied.
Q Ωi, and define f(cid:0)S(cid:1) , VXS (R). Definition 2.1 requires that we define f (·) over any subset
of U ,S Ui. We do this by taking the maximum of f (·) over all feasible subsets. Namely,
for set T ⊆ U , let RT ⊆ [n] denote again the set of i's such that T ∩ Ui ≥ 1. We set:
f (T ) , max
S⊆T s.t.
VXS (RT ).
∀i
S∩Ui≤1
Now f (·) is monotone subadditive because it is maximum of monotone subadditive functions.
3 Preliminaries
3.1 Submodular and Matroid Preliminaries
A set function f : {0, 1}n → R+ is a submodular function if for all S, T ⊆ [n] it satisfies
f (S∪T )+f (S∩T ) ≤ f (S)+f (T ). For any e ∈ [n] and S ⊆ [n], let fS(e) denote f (S∪e)−f (S).
We recall some notation for to extend submodular functions from the discrete hypercube
{0, 1}n to relaxations whose domain is the continuous hypercube [0, 1]n.
For any vector x ∈ [0, 1]n, let S ∼ x denote a random set S that contains each element
i ∈ [n] independently w.p. xi. Moreover, let 1S denote a vector of length n containing 1 for
i ∈ S and 0 for i 6∈ S.
Definition 3.1. We define important continuous extensions of any set function f .
Multilinear extension F :
F (x) , ES∼x[f (S)].
6
Concave closure f +:
f +(x) , max
Continuous relaxation f ∗:
f ∗(x) , min
α n XS⊆[n]
αSf (S) XS
S⊆[n](cid:26)f (S) + Xi∈[n]\S
αS = 1 and XS
fS(e) · xi(cid:27).
αS1S = xo.
3.1.1 Some Useful Results
Lemma 3.2 (Correlation gap [CCPV07]). For any monotone submodular function and x ∈
[0, 1]n,
F (x) ≤ f +(x) ≤ f ∗(x) ≤(cid:18)1 −
1
e(cid:19)−1
F (x).
Lemma 3.3 (Lemma 2.2 of [BFNS14]). Consider any submodular function f and any set
A ⊆ [n]. Let S be a random subset of A that contains each element of S w.p. at most p (not
necessarily independently), then
ES[f (S)] ≥ (1 − p) · f (∅).
Lemma 3.4 (Lemma 2.3 of [FMV11]). For any non-negative submodular function f and any
sets A, B ⊆ [n],
E
S∼1A/2
T ∼1B /2
[f (S ∪ T )] ≥
1
4
(f (∅) + f (A) + f (B) + f (A ∪ B)) .
Lemma 3.5 (Theorem 2.1 of [FMV11]). For any non-negative submodular function f , any
set A ⊆ [n] and p ∈ [0, 1],
F (1A · p) ≥ p(1 − p) · max
T ⊆A
f (T ).
We can now prove the following useful variant of the previous two lemmata (see also an
alternative self-contained proof in Appendix A).
Lemma 3.6. Consider any non-negative submodular function f and 0 ≤ L ≤ H ≤ 1. Let S∗
be a set that maximizes f , and let x ∈ [0, 1]n be such that for all i ∈ [n], L ≤ xi ≤ H. Then,
F (x) ≥ L(1 − H) · f (S∗).
exactly L.
Proof. Imagine a process in which in which we construct the random set T ∼ x, i.e. set T
containing each element e independently w.p. xe, in two steps. In the first step we construct
set T ′ by selecting every element independently w.p.
In the second step we
construct set T containing each element e independently w.p. (xe − L)/(1 − L). It's easy
to verify that the union of the two sets T ′ ∪ T contains each element e independently with
probability exactly xe.
From Lemma 3.5, we know that at the end of first step, the generated set has expected
value E[f (T ′)] ≥ L(1−L)f (S∗). Now, we argue that the second step does not "hurt" the value
by a lot. We note that in the second step each element is added w.p. at most (H − L)/(1− L)
because xe ≤ H. Let g(S) := f (S ∪ T ′) be a non-negative submodular function. We apply
Lemma 3.3 on g to get E[g( T )] ≥ (1 − H)/(1 − L) · g(∅), which implies E[f (T ′ ∪ T )] ≥
(1 − H)/(1 − L) · E[f (T ′)].
Together, we get F (x) = E[T ′ ∪ T ] ≥ L(1− L)(1− H)/(1− L)f (S∗) = L(1− H)f (S∗).
7
3.1.2 Online Contention Resolution Schemes
Given a point x in the matroid polytope P of matroid M, many submodular maximization
applications like to select each element i independently with probability xi and claim that the
selected set S has expected value F (x) [CVZ14]. The difficulty is that S need not be feasible
in M, and we can only select T ⊆ S that is feasible. Chekuri et al. [CVZ14] introduced the
notion of contention resolution schemes (CRS) that describes how, given a random S, one
can find a feasible T ⊆ S such that the expected value f (T ) will be close to F (x).
Recently, Feldman, Svensson, and Zenklusen gave online contention resolution schemes
(OCRS). Informally, it says that the decision of whether to select element i ∈ S into T can be
made online, even before knowing the entire set S [FSZ16]. In particular, we will need their
definition of greedy OCRS; we define it below and state the results from [FSZ16] that we use
in our O(1)-submodular prophet inequality result over matroids.
Definition 3.7 (Greedy OCRS). Let x belong to a matroid polytope P and S ∼ x. A greedy
OCRS defines a downward-closed family Fx of feasible sets in the matroid. All elements reveal
one-by-one if they belong to S, and when element i ∈ [n] reveals, the greedy OCRS selects it
if, together with the already selected elements, the obtained set is in Fx.
Lemma 3.8 (Theorems 1.8 and 1.10 of [FSZ16]). Given a non-negative submodular function
f , a matroid M, and a vector x in the convex hull of independent sets in M, there exists a
deterministic greedy OCRS that outputs a set T satisfying ET [F (1T /2)] ≥ (1/16) · F (x).
3.2 Subadditive and Downward-Closed Preliminaries
A set function f : {0, 1}n → R+ is subadditive if for all S, T ⊆ [n] it satisfies f (S∪T ) ≤ f (S)+
f (T ). It's monotone if f (S) ≤ f (T ) for any S ⊆ T . A set function f : {0, 1}n → R+ is an XOS
(alternately, fractionally subadditive) function if there exist linear functions Li : {0, 1}n → R+
such that f (S) = maxi{Li(S)}. (See Feige [Fei09] for illustrative examples.)
Lemma 3.9 ([Dob07, Lemma 3]). Any subadditive function v : 2n → R+ can be (cid:16) logS
2e (cid:17)-
approximated by an XOS functionbv : 2n → R+; i.e. for every S ⊆ [n],
Furthermore, the XOS function has the formbv (S) = maxT ⊆[n] pT ·T ∩ S for an appropriate
bv (S) ≤ v (S) ≤(cid:18) log S
2e (cid:19)bv (S) .
choice of pT 's.
Theorem 3.10 ([Rub16]). When items take values in {0, 1}, there are O (log n)-competitive
algorithms for (additive) Downward-Closed Secretary and Downward-Closed Prophet.
4 Correlation Gap for non-monotone submodular functions
For monotone submodular functions, [CCPV07] proved that
F (x) ≥ (1 − 1/e)f +(x).
(4.1)
This result was later rediscovered by [ADSY12], who called the ratio between f +(x) and
F (x) correlation gap.
It's useful in many applications since it says that up to a constant
8
factor, picking items independently is as good as the best correlated distribution with the
same element marginals.
What is the correct generalization of (4.1) to non-monotone submodular functions? It is
tempting to conjecture that F (x) ≥ c· f +(x) for some constant c > 0. However, the following
example shows that even for a function as simple as the directed cut function on a two-vertex
graph, this gap may be unbounded.
Example 4.1. Let f be the directed cut function on the two-vertex graph u → v;
f (∅) = 0, f ({u}) = 1, f ({v}) = 0, and f ({u, v}) = 0. Let x = (ǫ, 1 − ǫ). Then,
i.e.
F (x) = ǫ2 ≪ ǫ = f +(x).
It turns out that the right way to generalize (4.1) to non-monotone submodular functions
is to first make them monotone:
Definition 4.2 (fmax).
fmax(S) , max
T ⊆S
f (T ).
For non-monotone submodular f , we have that fmax is monotone, but it may no longer
be submodular, as shown by the following example:
Example 4.3 (fmax is not submodular). Let f be the directed cut function on the four-
vertex graph u → v → w → x. In particular, f ({v}) = 1, f ({u, v}) = 1, f ({v, w}) = 1, and
f ({u, w}) = 2.
fmax({u, v}) − fmax({v}) = 1 − 1 < 2 − 1 = fmax({u, v, w}) − fmax({v, w}).
Finally, we are ready to define correlation gap for non-monotone functions:
Definition 4.4 (Correlation gap). The correlation gap of any set function f is
max
x∈[0,1]n
max
α≥0( f +(x)
Fmax(x)(cid:12)(cid:12)(cid:12)XS
αS = 1 and XS
αS1S = x) ,
where Fmax is the multilinear extension of fmax.
Notice that for monotone f , we have that fmax ≡ f , so Definition 4.4 generalizes the
correlation gap for monotone submodular functions. Furthermore, one could replace f + with
f +
max in Definition 4.4; observe that the resulting definition is equivalent.
Theorem 4.5 (Non-monotone correlation gap). For any (non-monotone non-negative) sub-
modular function f , the correlation gap is at most 200.
While the constant can be improved slightly, we have not tried to optimize it, focusing
instead on clarity of exposition. Our proof goes through a third relaxation, f ∗
1/2.
Definition 4.6 (f ∗
1/2). For any set function f and any x ∈ [0, 1]n,
f ∗
1/2(x) , min
S⊆[n](cid:26)ET ∼1S/2(cid:2)f (T ) + Xi∈[n]\S
fT (e) · xi(cid:3)(cid:27).
9
Below, we will prove (Lemma 4.7) that f +(x) ≤ 4 · f ∗
1/2(x). We then show (Lemma 4.9)
that f ∗
1/2(x) ≤ 50 · F (x/2), which implies
f +(x) ≤ 4 · f ∗
1/2(x) ≤ 200 · F (x/2).
(4.2)
Finally, to finish the proof of Theorem 4.5, it suffices to show that F (x/2) ≤ Fmax(x).
This is easy to see since drawing T according to x/2 is equivalent to drawing S according
to x, and then throwing out each element from S independently with probability 1/2. For
Fmax(x), on the other hand, we draw the same set S and then take the optimal subset.
4.1 Proof that f +(x) ≤ 4 · f ∗
Lemma 4.7. For any x ∈ [0, 1]n and non-negative submodular function f : {0, 1}n → R+,
1/2(x)
f +(x) ≤ 4 · f ∗
1/2(x).
We first prove the following auxiliary claim:
Claim 4.8. For any sets S, T ⊆ [n],
ET1/2∼1T /2[f ((S \ T ) ∪ T1/2] ≥
1
4
f (S).
Proof. Define a new auxiliary function h(U ) , f ((S \ T ) ∪ U ). Observe that h continues to
be non-negative and submodular. We now have,
E
T1/2∼1T /2[f ((S \ T ) ∪ T1/2] = E
T1/2∼1T /2[h(T1/2)]
1
4
1
4
≥
=
h(T )
f (S)
(Lemma 3.6 for L = H = 1/2)
(Definition of h).
Proof of Lemma 4.7. Fix x, and let S∗ = S∗(x) denote the optimal set that satisfies f ∗
ET ∼1S∗ /2(cid:2)f (T ) +Pi∈[n]\S∗ fT (e)xi(cid:3). Let {αS} be the optimal distribution that satisfies
f +(x) =PS αSf (S). Then, 1
4PS αSf (S)
4 f +(x) = 1
1/2(x) =
(Claim 4.8)
αS · E
αS · ET ∼1S∗ /2[f ((S \ S∗) ∪ T )]
αS · ET ∼1S∗ /2 [f (T ) + fT (S \ S∗)]
≤XS
=XS
T ∼1S∗ /2f (T ) + Xi∈S\S∗
fT (e)
≤XS
= ET ∼1S∗ /2f (T )XS
fT (e)
αS Xi∈S\S∗
αS +XS
T ∼1S∗ /2f (T ) + Xi∈[n]\S∗
fT (e)xi = f ∗
1/2(x)
= E
10
(submodularity)
(usingXS
αS1S = x).
4.2 Proof that f ∗
1/2(x) ≤ 50 · F (x/2)
The proof of the following lemma is similar to Lemma 5 in [CCPV07].
Lemma 4.9. .
f ∗
1/2(x) ≤ 50 · F (x/2) .
Proof. Consider an exponential clock running for each element i ∈ [n] at rate xi. Whenever
the clock triggers, we update set S to S∪{i}. For t ∈ [0, 1], let S(t) denote the set of elements
in S by time t. Thus, each element belongs to S(1) w.p. 1 − exp(−xi), which is between
xi(1 − 1
e ) and xi. Let V (t) , ET ∼1S(t)/2[f (T )], i.e. expected value of set that picks each
element in S(t) independently w.p. 1
2 . Our goal is to show that:
f ∗
1/2(x) ≤(cid:18)
2
1 − e−1/2(cid:19) · E[V (1)] ≤(cid:18)
2
1 − e−1/2(cid:19)(cid:18) 4(e − 1)
e − 2 (cid:19) · F (x/2) .
(4.3)
We begin with the second inequality of (4.3). Consider the auxiliary submodular function
g(S) , ET ∼1−exp(−x)[f (S ∩ T )], and let G denote its multilinear extension. Let S∗ be a
maximizer of g, and observe that
V (1) = G(1[n]/2) ≤ g(S∗).
Observe further that, with slight abuse of notation, F (x/2) = G(cid:16)
defined since for any xi ∈ [0, 1], we have
x/2
1−exp(−x)(cid:17); this is well-
1
2 ≤
xi/2
1 − exp(−xi) ≤
e
2(e − 1)
< 1.
Moreover, since
xi/2
1−exp(−xi) is bounded, Lemma 3.6 gives
F (x/2) ≥
1
2 ·(cid:18)1 −
e
2(e − 1)(cid:19) · g(S∗) =
e − 2
4(e − 1)
g(S∗) = Ω(1) · g(S∗).
We now turn to the first inequality of (4.3). Consider an infinitesimal interval interval
(t, t + dt]. For any i /∈ S(t) the exponential clock triggers with probability xi dt, so it
contributes to V (t + dt) with probability xi/2 dt. The probability that two clocks trigger in
the same infinitesimal is negligible (O(dt2)). Therefore,
E[V (t + dt) − V (t)] = E
1
E
S(t)
≥
2(cid:16)f ∗
T ∼1S(t)/2 Xj∈[n]\S
xi
2
fT (j) dt − O(dt2)
(cid:17) dt − O(dt2).
Dividing both sides by dt and taking the limit as dt → 0, we get:
1/2(x) − ES(t)ET ∼1S(t)/2E[f (T )]
}
E[V (t)]
{z
2(cid:16)f ∗
1/2(x) − E[V (t)](cid:17).
d
dt
E[V (t)] ≥
1
11
To solve the differential inequality, let φ(t) = E[V (t)] and ψ(t) = exp( t
2 ) φ(t). We get
. Since ψ(0) = φ(0) = 0,
dφ
2 (f ∗
dt ≥ 1
integration over t gives
1/2(x) − φ(t)) and dψ
dt = exp( t
2 )( dφ
dt + φ(t)
2 ) ≥ exp( t
2 )
f ∗
1/2(x)
2
E[V (t)] = φ(t) = exp(−t/2) ψ(x) ≥
f ∗
1/2(x)
2
(1 − exp(−t/2)).
In particular, plugging in t = 1 completes the proof of the first inequality in (4.3).
5 Submodular Prophets over Matroids
Definition 5.1 (Submodular Matroid Prophet). The offline inputs to the problem are:
• n sets U1, . . . , Un; we denote their union U ,Sn
• a (not necessarily monotone) non-negative submodular function f : {0, 1}U → R+;
• n distributions Di over subset Ui; and
• a matroid M over [n]
i=1 Ui;
On the i-th time period, the algorithm observes an element Xi ∈ Ui drawn according to
Di, independently from outcomes and actions on past and future days. The algorithm must
decide (immediately and irrevocably) whether to add i and Xi to sets W and XW , respectively,
subject to W remaining independent in M. The objective is to maximize f (XW ).
Theorem 5.2. There is a randomized algorithm with a competitive ratio of O(1) for any
Submodular Matroid Prophet
Proof overview The main ingredients in the proof of Theorem 5.2 are known online con-
tention resolution schemes (OCRS) due to Feldman, Svensson, and Zenklusen [FSZ16], and
our new bound on the correlation gap for non-monotone submodular functions (Theorem 4.5).
Let x ∈ [0, 1]U denote the vector of probabilities that each element realizes (i.e. x(i,j) =
Di(j)). A naive proof plan proceeds as follows: Select elements online using the OCRS (w.r.t
x); obtain a constant factor approximation to F (x); use a "correlation gap" to show a constant
factor approximation of f +(x); finally, observe that f +(x) is an upper bound on OP T .
There are two problems with that plan: First, the OCRS of Feldman et al. applies when
elements realize independently. The realization of different elements for the same day is
obviously correlated (exactly one element realizes), so we cannot directly apply their OCRS.
The second problem is that for non-monotone submodular function, it is in general not true
that F (x) approximates f +(x) (see Example 4.1).
The solution to both obstacles is working with x/2 instead of x. In Section 4 we showed
that F (x/2) is a constant factor approximation of f +(x) (Ineq. (4.2)). Then, in Subsection
5.1, we give an algorithm that approximates the selection of the greedy OCRS on x/2. Our
plan is then to show:
ALG = Ω(E
S∼OCRS(x/2)[f (S)])
= Ω(F (x/2))
= Ω(f +(x))
= Ω(OP T ).
(Subsection 5.1)
(Lemma 3.8)
(Ineq. (4.2))
12
5.1 Applying the OCRS to our setting
In this subsection we show an algorithm that obtains, in expectation, 1/2 of the expected
value of the OCRS with probabilities x/2.
Our algorithm uses the greedy OCRS as a black box. On each day, the algorithm (sequen-
tially) feeds the OCRS a subset of the elements Ui that can potentially arrive on that day.
The subset on each day is chosen at random; it is correlated with the element that actually
arrives on that day, and independent from the subsets chosen on other days. The guarantee is
that the distribution over sequences fed into the OCRS is identical to the distribution induced
by x/2.
Reduction
P i
xi,j
x/2({(i,j)})
For each i, let Ui denote the set of elements that can arrive on day i, and fix some (arbitrary)
order over Ui. For a subset Si ⊆ Ui, let P i
x/2(Si) denote the probability that the set Si is
exactly the outcome of sampling from Ui according to x/2. When element (i, j) arrives on day
i, the algorithm feeds into the OCRS a random set Ti drawn from the following distribution.
, the algorithm feeds just element (i, j), i.e. Ti = {(i, j)}; notice
With probability
that this guarantees Pr [Ti = {(i, j)}] = P i
x/2({(i, j)}). Otherwise, the algorithm lets Ti be a
random subset of Ui, drawn according to x/2, conditioned on Ti 6= 1. This guarantees that
the probability mass on subsets of size 6= 1 is also allocated according to x/2.
Now, if the algorithm fed the singleton {(i, j)} and the OCRS selected it, then the al-
gorithm also takes {(i, j)}; otherwise the algorithm does not take {(i, j)}. (In particular, if
Ti 6= 1, the algorithm ignores the decisions of the OCRS.)
Analyzing the reduction
Observe that on each day the distribution over Ti's is identical to the distribution P i
x/2(·).
Since the Ti's are also independent, it means that the distribution of inputs to the OCRS is
indeed distributed according to x/2.
Conditioning on (i, j) is being fed (i.e., with probability xi,j/2), P i
x/2(·) assigns at least
1/2 probability to the event where no other element is also being fed (this is precisely the
reason we divide x by 2):
Pr[Ti = {(i, j)} Ti ∋ (i, j)] ≥ 1/2.
Since the OCRS is greedy, for any history on days 1, . . . , i − 1, if it selects (i, j) when
observing set Ti ∋ (i, j), it would also select (i, j) when observing only this element on day
i. Furthermore, since the OCRS is only allowed to select one element on day i, conditioning
on the OCRS selecting (i, j), the future days (i + 1, . . . , n) proceed independently of whether
the algorithm also selected (i, j). Therefore, conditioning on the greedy OCRS selecting any
set SOCRS, the algorithm selects a subset TALG ⊆ SOCRS where each element appears with
probability at least 1/2.
Finally, to argue that the algorithm obtains at least 1/2 of the expected value of the set
selected by the OCRS, fix the set SOCRS selected by the OCRS, and consider the submodular
function g( ¯T ) , f (SOCRS \ ¯T ). Setting ¯T , TALG \ SOCRS, we have that f (TALG) = g( ¯T ).
Thus by Lemma 3.3,
E[f (TALG)] ≥
1
2
E[g(∅)] =
1
2
E[f (SOCRS)].
13
6 Subadditive Secretary over Downward-Closed Constraints
Definition 6.1 (Monotone Subadditive Downward-Closed Secretary). Consider
n items, a monotone subadditive valuation function from subsets of items to R+, and an
arbitrary downward-closed set system F over the items; both f and F are adversarially
chosen. The algorithm receives as input n (but not F or f ). The items arrive in a uniformly
random order. Initialize W as the empty set. When item i arrives, the algorithm observes all
feasible subsets of items that have already arrived, and their valuation in f . The algorithm
then decides (immediately and irrevocably) whether to add i to the set W , subject to the
constraint that W remains a feasible set in F. The goal is to maximize f (W ).
Theorem 6.2. There is a deterministic algorithm for Monotone Subadditive Downward-
Closed Secretary that achieves a competitive ratio of O(cid:0)log n · log2 r(cid:1).
Proof. Let T ⋆ be the set chosen by the offline algorithm (OP T = f (T ⋆)). By Lemma 3.9
there exists a pT ⋆ such that for every S ⊆ T ⋆:
f (S) ≥ pT ⋆ S ∩ T ⋆ ;
OP T = f (T ⋆) = O (pT ⋆ T ⋆ log T ⋆) = O (pT ⋆ T ⋆) log r.
(6.1)
(6.2)
Assume that we know pT ⋆ (discussed later). We define a new feasibility constraint F ′ as
follows: a set T ⊆ [n] is feasible in F ′ iff it is feasible in F and for every subset S ⊆ T , we
have f (S) ≥ pT ⋆ S. Notice that because we also force the condition on all subsets of T , F ′
is downward-closed and it does not depend on the order of arrival.
We run the algorithm for {0, 1}-valued (additive) Downward-Closed Secretary (as
guaranteed by Theorem 3.10) with feasibility constraint F ′ where all values are 1. By (6.1),
log r(cid:17). Therefore, the additive {0, 1}-values
T ⋆ is feasible in F ′, and by (6.2) pT ⋆ T ⋆ = Ω(cid:16) OP T
algorithm returns a set T ALG of size(cid:12)(cid:12)T ALG(cid:12)(cid:12) = Ω(cid:16)
pT ⋆ log n log r(cid:17). Furthermore, T ALG is also
feasible in F ′, i.e.
f (T ALG) ≥ pT ⋆(cid:12)(cid:12)T ALG(cid:12)(cid:12) = Ω(cid:18) OP T
log n log r(cid:19) .
(6.3)
OP T
Guessing pT ⋆
Finally, we don't actually know pT ⋆, but we can guess it correctly, up to a constant factor,
with probability 1/ log r. We run the classic secretary algorithm over the first n/2 items,
where we use the value of the singleton f ({i}) as "the value of item i": Observe the first n/4
items and select none; then take the next item whose value is larger than every item observed
so far. With constant probability this algorithm selects the item with the largest value, which
we denote by M .
Also, with constant probability the algorithm sees the item with the largest value too
early and does not select it. Assume that this is the case. Since we obtained expected value
of Ω (M ) on the first n/2 items we can, without loss of generality, ignore values less than M/r.
In particular, we know that pT ⋆ ∈ [M/r, M ]. Pick α ∈ {M/r, M/(2r), . . . , M/2, M} uniformly
at random, and use it instead of pT ⋆ to define F ′. With probability 1/ log r, pT ⋆ ∈ [α, 2α], in
which case the algorithm returns a set T ALG satisfying (6.3).
14
7 Subadditive Prophet
Definition 7.1 (Monotone Subadditive Downward-Closed Prophet). The offline
inputs to the problem are:
• n sets U1, . . . , Un; we denote their union U ,Sn
• a monotone non-negative subadditive function f : {0, 1}U → R+;
• n distributions Di over subset Ui; and
• a feasibility constraint F over [n].
i=1 Ui;
On the i-th time period, the algorithm observes an element Xi ∈ Ui drawn according to
Di, independently from outcomes and actions on past and future days. The algorithm must
decide (immediately and irrevocably) whether to add i and Xi to sets W and XW , respectively,
subject to the constraint that W remains feasible in F. The objective is to maximize f (XW ).
Let r denote the maximum cardinality of a feasible set S ∈ F.
Theorem 7.2. There is a deterministic algorithm for Monotone Subadditive Downward-
Closed Prophet that achieves a competitive ratio of O(cid:0)log n · log2 r(cid:1).
The proof of Theorem 7.2 consists of three steps: in Subsection 7.1 we reduce monotone
subadditive valuations over independent items to monotone XOS subadditive valuations over
independent items, with a loss of O (log r), using a lemma of Dobzinski [Dob07]. Then in
Subsection 7.2 we use a standard reduction from general XOS valuations to XOS with {0, 1}
marginal contributions, losing another factor of O (log r). Finally, in Subsection 7.3 we use
techniques from [Rub16] to give an O (log n)-competitive algorithm for monotone XOS with
{0, 1} marginal contributions.
7.1 Subadditive to XOS
Definition 7.3 (Monotone XOS Downward-Closed Prophet). For any set M and
items [n], the offline inputs to the problem are:
• n sets U1, . . . , Un of valuations vectors in RM
• a monotone XOS function bf : {0, 1}U → R+
m∈MXu∈S
bf (S) , max
+ ; we denote their union U ,Sn
um for S ∈ {0, 1}U ;
i=1 Ui;
• n distributions Di over subset Ui; and
• a feasibility constraint F over [n], which is a collection of subsets of [n].
On the i-th time period, the algorithm observes a valuations vector Xi ∈ Ui drawn according
to Di, independently from outcomes and actions on past and future days. The algorithm
must decide (immediately and irrevocably) whether to add i and Xi to sets W and XW ,
respectively, subject to the constraint that W remains feasible in F. The objective is to
maximize bf (XW ).
15
Below (Proposition 7.4) we give an O (log n · log r)-competitive algorithm for Monotone
XOS Downward-Closed Prophet. By Dobzinski's lemma (Lemma 3.9), this implies an
O(cid:0)log n · log2 r(cid:1)-competitive algorithm for Monotone Subadditive Downward-Closed
Prophet.
Proposition 7.4. There is a deterministic algorithm for Monotone XOS Downward-
Closed Prophet that achieves a competitive ratio of O (log n · log r).
7.2 XOS to XOS with {0, 1} coefficients
Below (Proposition 7.5), we give an O (log n)-competitive algorithm for Monotone XOS
Downward-Closed Prophet in the special case where all the vectors v ∈ U are in {0, 1}M .
First, let us show why this would imply Proposition 7.4.
Proof of Proposition 7.4 from Proposition 7.5. We recover separately the contributions from
"tail" events (a single item taking an exceptionally high value) and the "core" contribution
that is spread over many items. Run the better of the following two algorithms:
Tail Let OP T denote the expected offline optimum value. Whenever we see a feasible
item whose valuations vector Xi has value at least 2OP T , we select it. For item i, let
pi = Pr [Xi ≥ 2OP T ]. We have
OP T ≥ 2OP T · Pr [∃i : Xi ≥ 2OP T ] = 2OP T ·(cid:16)1 −Y (1 − pi)(cid:17) .
Dividing by OP T and rearranging, we get
and thus
1/2 ≤Y (1 − pi) ≤ e− P pi,
X pi ≤ ln 2.
Therefore the probability that we want to take an item but can't is at most ln 2, so
this algorithm achieves at least a (1 − ln 2)-fraction of the expected contribution from values
greater than 2OP T .
Core Observe that we can safely ignore values less than OP T /2r, as those can contribute a
total of at most OP T /2. Partition all remaining values into 2+log r intervals [OP T /2r, OP T /r] ,
. . . , [OP T, 2OP T ]. The expected contribution from the values in each interval is Ω (1/ log r)-
fraction of the expected offline optimum without values greater than 2OP T . Pick the interval
with the largest expected contribution, round down all the values in this interval, and run
expected contribution from values less than or equal to 2OP T .
the algorithm guaranteed by Proposition 7.5. This achieves an Ω(cid:16)
7.3 XOS with {0, 1} coefficients
Proposition 7.5. When the Xi's take values in {0, 1}M , there is a deterministic algorithm
for Monotone XOS Downward-Closed Prophet that achieves a competitive ratio of
O (log n).
log n·log r(cid:17)-fraction of the
1
16
7.3.1 A dynamic potential function
At each iteration, the algorithm maintains a target value τ and a target probability π, where
π is the probability (over future realizations) that the current restricted prophet beats τ . We
say that an outcome (i.e. a pair of item and valuations vector) is good if selecting it does
not decrease the probability of beating the target value by a factor greater than n2, and bad
otherwise. Notice that all the bad items together contribute at most a (1/n)-fraction of the
probability of beating τ . A key ingredient is that τ is updated dynamically. If the probability
of observing a good outcome is too low (less than 1/4), we deduct 1 from τ . We show (Lemma
7.9) that this increases π by a factor of at least 2. Since π decreases by at most an n2 factor
when we select an item, and increases by a factor of 2 whenever we deduct 1 from τ : we
balance 2 log n deductions for every item the algorithm selects, and this gives the O (log n)
competitive ratio.
So far our algorithm is roughly as follows: set a target value τ ; whenever the probability
π of reaching the target τ drops below 1/4, decrease τ ; if π > 1/4, sit and wait for a good
outcome - one will arrive with probability at least 1/4 (we actually do this with Pr[A] instead
of π, where A is a closely related event). There is one more subtlety: what should the
algorithm do if no good outcomes arrive? In other words, what if the probability of observing
a good outcome is neither very low nor very close to 1, say 1/2 or even 1− 1
log n ? On one hand,
we can't decrease τ again, because we are no longer guaranteed a significant increase in π; on
the other hand, after, say Θ(cid:0)log2 n(cid:1) iterations, we still have a high probability of having an
iteration where none of the good outcomes arrive. (If no good outcomes are coming, we don't
want the algorithm to wait forever...) Fortunately, there is a simple solution: the algorithm
waits for the last item with a good outcome in its support; if, against the odds, no good
outcomes have yet been observed, the algorithm "hallucinates" that this last item has a good
valuations vector, and selects it. In expectation, at most a constant fraction of the items we
select are "hallucinated", so the competitive ratio is still O (log n).
7.3.2 Notation
We let OP T denote the expected (offline) optimum. W is the set of items selected so far (W
for "Wins"), and ℓW , max {i ∈ W} is the index of the last selected item.
Let F denote the family of all feasible subsets of [n]. For any T ⊆ [n], let FT denote the
family of feasible sets whose intersection with {1, . . . , max (T )} is exactly T .
i )m∈M ∈ {0, 1}M denote the random vector drawn for the i-th item. We use
Let Xi = (X m
zi to refer to the observed realization of Xi. Our algorithm will maintain a subset M ′ ⊆ M .
We let
VM ′(cid:0)F, X[n](cid:1) , max
S∈F
max
m∈M ′Xi∈S
(Xi)m
denote the value of optimum offline solution (note that this is also a random variable).
Let τ = τ (W ) be the current target value, and π = π (τ, W ) denotes the current target
probability:
π (τ, W ) , PrhVM ′(cid:16)FW , X[n](cid:17) > τ X[ℓW ] = z[ℓW ]i .
For each yj ∈ supp(cid:16)Xj(cid:17), we define πj,yj = πj,yj (τ, W ) to be the probability of reaching τ ,
given that:
• zj = yj,
17
• j is the next item we select, and
• item j actually contributes 1 to the offline optimum.
Formally,
πj,yj (τ, W ) , Pr(cid:2)VM ′∩yj(cid:0)FW +j, X[n](cid:1) > τ X[ℓW ]∪[j] =(cid:0)z[ℓW ], yj(cid:1)(cid:3) ,
j = 1.
We say that a future outcome (j, yj) is good if πj,yj ≥ n−2·π and j is feasible (and otherwise
where we slightly abuse notation and also use yj to denote the set of m ∈ M such that ym
it is bad), and let G = {good (j, yj)} denote the set of good future outcomes. Finally,
A , A (π, τ, W ) ,
is the event that at least one of the good outcomes occur.
7.3.3 Updated proof plan and the algorithm
The idea is to always maintain a threshold τ such that probability of one of the good outcomes
to occur is large, i.e. Pr[A] is at least a constant 1
4 . The way we do this is by showing in
Claim 7.6 that at any time during the execution of the algorithm, conditioned on what all has
happened till now, the probability π that the offline algorithm achieves the threshold τ gives
a lower bound on Pr[A]. Hence, whenever Pr[A] goes below 1
4 , we decrease the threshold τ ,
which increases π due to Lemma 7.9 and, indirectly, increases Pr[A] by Claim 7.6.
due to Ledoux to show that in the beginning τ = OP T /2 satisfies π > 1
4 .
Initialize τ ← OP T /2, M ′ ← M , and W ← ∅. Lemma 7.8 uses a concentration bound
After each update to W , decrease τ until Pr [A] ≥ 1/4, or until W > τ . When Pr [A] ≥
1/4, reveal the values of items until observing a good outcome. When we observe a good
outcome zj, add j to W and restrict M ′ to its intersection with zj. Since we restrict M to
M ′, this gives us that at any time
VM ′ (F, XW ) = W.
If we reach the last item with good outcomes in its support, and none of the good outcomes
realize, add this last item to G and subtract 1 from τ (without modifying M ′). See also
pseudocode in Algorithm 1.
We first claim that π gives us a lower bound on Pr[A] because most of the mass in π
comes from good outcomes.
Claim 7.6. At any point during the run of the algorithm,
Pr[A (π, τ, W )] ≥(cid:18)1 −
1
n(cid:19) π(W, τ ).
Proof. For each (j, yj) /∈ G, we have, by definition of G,
πj,yj (W, τ ) < n−2 · π (W, τ ) .
Summing over all (j, yj) /∈ G,
18
Algorithm 1 Prophet
; M ′ ← M ; W ← ∅
2
1. τ ← OP T
2. while τ > W:
# π is the probability that, given the history, the offline optimum can still beat τ .
(a) π ← Pr(cid:2)VM ′(cid:0)FW , X[n](cid:1) > τ X[ℓW ] = z[ℓW ](cid:3)
(b) G ←(cid:8)(j, yj) : j > ℓW AND πj,yj ≥ n−2 · π(cid:9) ∩(cid:16)SS∈FW
# G is the set of good and feasible outcomes.
S(cid:17)
(c) if Pr [A] ≥ 1/4
# A good outcome is likely occur.
i. j∗ ← min{j ∈ G : (j, zj ) ∈ G}
ii. if j∗ = ∞
# Wait for a good and feasible outcome.
# No good outcomes.
A. j∗ ← max G
B. τ ← τ − 1
# Select the last potentially good item.
# Adjust the target value to account for select an item with value 0
iii. else
# j∗ is actually a good item.
A. M ′ ← M ′ ∩ zj
iv. W ← W ∪ {j∗}
(d) else
i. τ ← τ − 1
# decrease target value τ until Pr [A] ≥ 1/4.
19
Xj Xyj:(j,yj) /∈G
Pr [yj] · πj,yj (W, τ ) ≤Xj Xyj :(j,yj) /∈G
Pr [yj] ·(cid:0)n−2 · π (W, τ )(cid:1)
n−2 · π (W, τ )
≤Xj
≤ n−1 · π (W, τ ) .
Thus, most of π comes from good (j, yj)'s:
Pr[A] =Xj Xyj :(j,yj)∈G
Pr [yj] · πj,yj (W, τ ) ≥ (1 − 1/n) π (W, τ ) .
(7.1)
7.3.4 Concentration for the beginning
Theorem 7.7. [Led97, Theorem 2.4] There exists some constant K > 0 such that the fol-
lowing holds. Let Yi's be independent (but not necessarily identical) random variables in
some space S; let C be a countable class of measurable functions f : S → [0, 1]; and let
Z = supf ∈CPn
i=1 f (Yi). Then,
t
t
To make the connection to our setting, let Yi be the vector in [0, 1]F ×M whose (S, m)-th
coordinate is X m
i=1 fS,m (Yi) is simply
i
the value of the set S under the m-th summation in the XOS representation of the valuation
function. Let C , {fS}S∈F . The above concentration inequality can now be written as
K · log(cid:18)1 +
Pr [Z ≥ E [Z] + t] ≤ exp(cid:18)−
E [Z](cid:19)(cid:19) .
if i ∈ S, and 0 otherwise. Let fS,m (Yi) , [Yi]S,m, soPn
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T + t(cid:3) ≤ exp(cid:18)−
K · log(cid:18)1 +
t
OP T(cid:19)(cid:19) .
t
Lemma 7.8. Assume OP T ≥ Ω (log n). Then,
Pr(cid:20)V (cid:0)F, X[n](cid:1) ≥
OP T
2 (cid:21) > 1/4.
Proof. We have,
OP T = ∞
−OP T
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T + t(cid:3) dt,
which can be decomposed as to integrals over [−OP T,−OP T /2],
[OP T,∞].
The first two integrals can be easily bounded as
(7.2)
(7.3)
[−OP T /2, OP T ], and
−OP T /2
−OP T
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T + t(cid:3) dt ≤ −OP T /2
−OP T
1 · dt ≤
OP T
2
20
and
OP T
−OP T /2
∞
OP T
For the third integral we use the concentration bound (7.2):
2
≤
OP T
−OP T /2
3OP T
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T + t(cid:3) dt ≤ OP T
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T /2(cid:3) dt
· Pr(cid:20)V (cid:0)F, X[n](cid:1) >
2 (cid:21) .
K · log(cid:18)1 +
exp(cid:18)−
Pr(cid:2)V (cid:0)F, X[n](cid:1) ≥ OP T + t(cid:3) dt ≤ ∞
exp(cid:18)−
K(cid:19) dt
≤ ∞
= hKe−t/Ki∞
= K · e−OP T /K,
OP T(cid:19)(cid:19) dt
t
t
OP T
OP T
OP T
t
which is negligible since OP T = ω (1).
Plugging into (7.3), we have:
OP T ≤
OP T
2
+
3OP T
2
and after rearranging we get
· Pr(cid:20)V (cid:0)F, X[n](cid:1) >
OP T
2 (cid:21) + o (1) ,
Pr(cid:20)V (cid:0)F, X[n](cid:1) >
OP T
2 (cid:21) ≥ 1/3 − o (1) .
7.3.5 Main lemma
Lemma 7.9. At any point during the run of the algorithm, if Pr [A] ≤ 1/4, then subtracting
1 from τ doubles π; i.e.
π (W, τ − 1) ≥ 2π (W, τ ) .
Proof of Lemma 7.9. Consider the event that the optimum solution (conditioned on the items
W we already selected and the realizations z[ℓW ] we have already seen) reaches τ . We can
write it as a union of disjoint events, depending on the next item j > ℓW that is part of the
optimum solution, and its possible realizations yj:
π (W, τ ) =Xj Xyj
Pr [yj] · Pr(cid:2)VM ′∩yj(cid:0)FW ∪{j}, X[n](cid:1) > τ X[ℓW ]∪[j] =(cid:0)z[ℓW ], yj(cid:1)(cid:3)
}
πj,yj (W,τ )
{z
.
We break the RHS into the sum over (j, yj)'s that are good and the sum over those that are
bad. Now, Claim 7.6 gives
Xj Xyj:(j,yj)∈G
Pr [yj] · πj,yj (W, τ ) ≥ (1 − 1/n) π (W, τ ) .
(7.4)
Since yj ∈ {0, 1}M , each item can contribute at most 1 to the offline optimum. Therefore:
21
Pr(cid:2)VM ′∩yj(cid:0)FW ∪{j}, X[n](cid:1) > τ X[ℓW ]∪[j] =(cid:0)z[ℓW ], yj(cid:1)(cid:3)
}
πj,yj =πj,yj (τ,W )
{z
Plugging into (7.4), we have
≤ πj,0 (W, τ − 1)
(1 − 1/n) π (W, τ ) ≤ Xj Xyj:(j,yj)∈G
Pr [yj] · πj,0 (W, τ − 1)
≤ Xj
≤ Xj
Pr [(j, yj) ∈ G] · πj,0 (W, τ − 1)
Pr [(j, yj) ∈ G] · π (W, τ − 1) ,
(7.5)
where the second inequality follows because πj,0 (W, τ − 1) doesn't depend on yj, and the
third because conditioning on the j-th item being 0 can only decrease the probability of
reaching τ − 1.
Recall that A is the union of all the events (j, yj) ∈ G. Therefore,
Pr [A] ≥Xj
Pr [(j, yj ) ∈ G] (1 − Pr [A])
Plugging in Pr [A] < 1/4, we get that Pj Pr [(j, yj) ∈ G] < 1/3. Plugging into (7.5) and
rearranging, we get
π (W, τ − 1) ≥
π (W, τ ) .
3n
n − 1
7.3.6 Putting it all together
Lemma 7.10. At any point during the run of the algorithm,
τ ≥
OP T
2 − (2 log n + 1) · W − 2
Proof. We prove by induction that at any point during the run of the algorithm,
log π ≥ −2 − (2 log n + 1) · W +(cid:18) OP T
2 − τ(cid:19) .
(7.6)
After initialization, log π ≥ −2 by Lemma 7.8. By definition of G, whenever we add an item
to W , we decrease log π by at most 2 log n - hence the 2 log n·W term. Notice that when the
algorithm "hallucinates" a 1, we also decrease τ by 1 to correct for the hallucination - at any
point during the run of the algorithm, this has happened at most W times. Recall that we
may also decrease τ in the last line of Algorithm 1 (in order to increase π); whenever we do
this, τ decreases by 1, but π doubles (by Lemma 7.9), so log π increases by 1, and Inequality
(7.6) is preserved.
Finally, since π is a probability, we always maintain log π ≤ 0.
We are now ready to complete the proof of Theorem 7.2.
22
OP T
Proof of Proposition 7.5. The algorithm always terminates after at most O (OP T ) decreases
to the value of τ . By Lemma 7.10, when the algorithm terminates, we have W ≥ τ ≥
2 − (2 log n + 1) · W − 2, and therefore in particular W ≥ OP T −4
4 log n+4 .
Finally, recall that sometimes the algorithm "hallucinates" good realizations, i.e. for some
items i ∈ W that we select, Xi = 0. However, each time we add an item, the probability
that we add a zero-value item is at most 3/4 (by the condition Pr [A] > 1/4). Therefore in
expectation the value of the algorithm is at least W /4.
Acknowledgements
The second author thanks Anupam Gupta for introducing him to submodular optimization.
References
[ADSY12] Shipra Agrawal, Yichuan Ding, Amin Saberi, and Yinyu Ye. Price of correlations
in stochastic optimization. Operations Research, 60(1):150–162, 2012.
[AHL12]
Saeed Alaei, MohammadTaghi Hajiaghayi, and Vahid Liaghat. Online prophet-
inequality matching with applications to ad allocation. In ACM Conference on
Electronic Commerce, EC '12, Valencia, Spain, June 4-8, 2012, pages 18–35,
2012.
[AKW14] Pablo Daniel Azar, Robert Kleinberg, and S. Matthew Weinberg. Prophet in-
equalities with limited information. In Proceedings of the Twenty-Fifth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon,
USA, January 5-7, 2014, pages 1358–1377, 2014.
[Ala11]
Saeed Alaei. Bayesian combinatorial auctions: Expanding single buyer mech-
anisms to many buyers.
In IEEE 52nd Annual Symposium on Foundations of
Computer Science, FOCS 2011, Palm Springs, CA, USA, October 22-25, 2011,
pages 512–521, 2011.
[BCK12] Anand Bhalgat, Tanmoy Chakraborty, and Sanjeev Khanna. Mechanism design
for a risk averse seller. In Internet and Network Economics - 8th International
Workshop, WINE 2012, Liverpool, UK, December 10-12, 2012. Proceedings, pages
198–211, 2012.
[BFNS14] Niv Buchbinder, Moran Feldman, Joseph Seffi Naor, and Roy Schwartz. Submod-
ular maximization with cardinality constraints. In Proceedings of the Twenty-Fifth
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1433–1452. SIAM,
2014.
[BFS15]
[BH16]
Niv Buchbinder, Moran Feldman, and Roy Schwartz. Online submodular max-
imization with preemption.
In Proceedings of the Twenty-Sixth Annual ACM-
SIAM Symposium on Discrete Algorithms, pages 1202–1216. SIAM, 2015.
Eric Balkanski and Jason D. Hartline. Bayesian budget feasibility with posted
pricing. In Proceedings of the 25th International Conference on World Wide Web,
WWW 2016, Montreal, Canada, April 11 - 15, 2016, pages 189–203, 2016.
23
[BHZ13] MohammadHossein Bateni, Mohammad Taghi Hajiaghayi, and Morteza Zadi-
moghaddam. Submodular secretary problem and extensions. ACM Trans. Algo-
rithms, 9(4):32, 2013.
[BIK07]
Moshe Babaioff, Nicole Immorlica, and Robert Kleinberg. Matroids, secretary
problems, and online mechanisms. In Proceedings of the Eighteenth Annual ACM-
SIAM Symposium on Discrete Algorithms, SODA 2007, New Orleans, Louisiana,
USA, January 7-9, 2007, pages 434–443, 2007.
[BMKK14] Ashwinkumar Badanidiyuru, Baharan Mirzasoleiman, Amin Karbasi, and An-
dreas Krause. Streaming submodular maximization: Massive data summarization
on the fly. In Proceedings of the 20th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 671–680. ACM, 2014.
[BPR+16] Ashwinkumar Badanidiyuru, Christos H. Papadimitriou, Aviad Rubinstein, Lior
Seeman, and Yaron Singer. Locally adaptive optimization: Adaptive seeding for
monotone submodular functions. In Proceedings of the Twenty-Seventh Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA,
USA, January 10-12, 2016, pages 414–429, 2016.
[BUCM12] Siddharth Barman, Seeun Umboh, Shuchi Chawla, and David L. Malec. Secre-
tary problems with convex costs.
In Automata, Languages, and Programming
- 39th International Colloquium, ICALP 2012, Warwick, UK, July 9-13, 2012,
Proceedings, Part I, pages 75–87, 2012.
[CCPV07] Gruia Calinescu, Chandra Chekuri, Martin P´al, and Jan Vondr´ak. Maximizing a
submodular set function subject to a matroid constraint. In Integer programming
and combinatorial optimization, pages 182–196. Springer, 2007.
[CGQ15] Chandra Chekuri, Shalmoli Gupta, and Kent Quanrud. Streaming algorithms for
submodular function maximization. In International Colloquium on Automata,
Languages, and Programming, pages 318–330. Springer, 2015.
[CK15]
[CVZ14]
Amit Chakrabarti and Sagar Kale. Submodular maximization meets streaming:
Matchings, matroids, and more. Mathematical Programming, 154(1-2):225–247,
2015.
Chandra Chekuri, Jan Vondr´ak, and Rico Zenklusen. Submodular function max-
imization via the multilinear relaxation and contention resolution schemes. SIAM
J. Comput., 43(6):1831–1879, 2014.
[DHK+13] Nikhil R. Devanur, Zhiyi Huang, Nitish Korula, Vahab S. Mirrokni, and Qiqi
Yan. Whole-page optimization and submodular welfare maximization with online
bidders. In ACM Conference on Electronic Commerce, EC '13, Philadelphia, PA,
USA, June 16-20, 2013, pages 305–322, 2013.
[Din13]
[Dob07]
Michael Dinitz. Recent advances on the matroid secretary problem. SIGACT
News, 44(2):126–142, 2013.
Shahar Dobzinski. Two randomized mechanisms for combinatorial auctions.
In Approximation, Randomization, and Combinatorial Optimization. Algorithms
24
and Techniques, 10th International Workshop, APPROX 2007, and 11th Inter-
national Workshop, RANDOM 2007, Princeton, NJ, USA, August 20-22, 2007,
Proceedings, pages 89–103, 2007.
[Dyn63]
E. B. Dynkin. The optimum choice of the instant for stopping a markov process.
Sov. Math. Dokl., 1963.
[EHLM15] Hossein Esfandiari, MohammadTaghi Hajiaghayi, Vahid Liaghat, and Morteza
In Algorithms - ESA 2015 - 23rd Annual
Monemizadeh. Prophet secretary.
European Symposium, Patras, Greece, September 14-16, 2015, Proceedings, pages
496–508, 2015.
[Fei09]
[FI15]
Uriel Feige. On maximizing welfare when utility functions are subadditive. SIAM
Journal on Computing, 39(1):122–142, 2009.
Moran Feldman and Rani Izsak. Building a good team: Secretary problems and
the supermodular degree. CoRR, abs/1507.06199, 2015.
[FMV11] Uriel Feige, Vahab S Mirrokni, and Jan Vondrak. Maximizing non-monotone
submodular functions. SIAM Journal on Computing, 40(4):1133–1153, 2011.
[FNS11] Moran Feldman, Joseph Naor, and Roy Schwartz.
Improved competitive ra-
tios for submodular secretary problems (extended abstract). In Approximation,
Randomization, and Combinatorial Optimization. Algorithms and Techniques -
14th International Workshop, APPROX 2011, and 15th International Workshop,
RANDOM 2011, Princeton, NJ, USA, August 17-19, 2011. Proceedings, pages
218–229, 2011.
[FSZ15]
[FSZ16]
Moran Feldman, Ola Svensson, and Rico Zenklusen. A simple O (log log(rank))-
competitive algorithm for the matroid secretary problem. In Proceedings of the
Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2015, San Diego, CA, USA, January 4-6, 2015, pages 1189–1201, 2015.
Moran Feldman, Ola Svensson, and Rico Zenklusen. Online contention resolution
schemes. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016,
pages 1014–1033, 2016.
[FZ15]
Moran Feldman and Rico Zenklusen. The submodular secretary problem goes
linear. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual
Symposium on, pages 486–505. IEEE, 2015.
[GK10]
Daniel Golovin and Andreas Krause. Adaptive submodularity: A new approach
to active learning and stochastic optimization. In COLT, pages 333–345, 2010.
[GKS14]
Daniel Golovin, Andreas Krause, and Matthew Streeter. Online submodular max-
imization under a matroid constraint with application to learning assignments.
arXiv preprint arXiv:1407.1082, 2014.
[HK92]
Theodore P Hill and Robert P Kertz. A survey of prophet inequalities in optimal
stopping theory. Contemp. Math, 125:191–207, 1992.
25
[HK12]
Elad Hazan and Satyen Kale. Online submodular minimization. Journal of Ma-
chine Learning Research, 13(Oct):2903–2922, 2012.
[HKP04] Mohammad Taghi Hajiaghayi, Robert D. Kleinberg, and David C. Parkes. Adap-
tive limited-supply online auctions. In Proceedings 5th ACM Conference on Elec-
tronic Commerce (EC-2004), New York, NY, USA, May 17-20, 2004, pages 71–
80, 2004.
[HKS07] Mohammad Taghi Hajiaghayi, Robert D. Kleinberg, and Tuomas Sandholm. Au-
tomated online mechanism design and prophet inequalities.
In Proceedings of
the Twenty-Second AAAI Conference on Artificial Intelligence, July 22-26, 2007,
Vancouver, British Columbia, Canada, pages 58–65, 2007.
[KKN15] Thomas Kesselheim, Robert D. Kleinberg, and Rad Niazadeh. Secretary problems
with non-uniform arrival order. In Proceedings of the Forty-Seventh Annual ACM
on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June
14-17, 2015, pages 879–888, 2015.
[Kle05]
Robert D. Kleinberg. A multiple-choice secretary algorithm with applications to
online auctions. In Proceedings of the Sixteenth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2005, Vancouver, British Columbia, Canada, Jan-
uary 23-25, 2005, pages 630–631, 2005.
[KMZ15] Nitish Korula, Vahab Mirrokni, and Morteza Zadimoghaddam. Online submod-
ular welfare maximization: Greedy beats 1/2 in random order. In Proceedings of
the Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages
889–898. ACM, 2015.
[KPV13] Michael Kapralov, Ian Post, and Jan Vondr´ak. Online submodular welfare maxi-
mization: Greedy is optimal. In Proceedings of the Twenty-Fourth Annual ACM-
SIAM Symposium on Discrete Algorithms, pages 1216–1225. Society for Industrial
and Applied Mathematics, 2013.
[KS77]
[KSS13]
[KW12]
[Lac14]
Ulrich Krengel and Louis Sucheston. Semiamarts and finite values. Bull. Amer.
Math. Soc., 83(4):745–747, 07 1977.
Pushmeet Kohli, Mahyar Salek, and Greg Stoddard. A fast bandit algorithm for
recommendation to users with heterogenous tastes. In Proceedings of the Twenty-
Seventh AAAI Conference on Artificial Intelligence, July 14-18, 2013, Bellevue,
Washington, USA., 2013.
Robert Kleinberg and S. Matthew Weinberg. Matroid prophet inequalities. In
Proceedings of the 44th Symposium on Theory of Computing Conference, STOC
2012, New York, NY, USA, May 19 - 22, 2012, pages 123–136, 2012.
Oded Lachish. O(log log rank) competitive ratio for the matroid secretary prob-
lem.
In 55th IEEE Annual Symposium on Foundations of Computer Science,
FOCS 2014, Philadelphia, PA, USA, October 18-21, 2014, pages 326–335, 2014.
[Led97]
Michel Ledoux. On Talagrand's deviation inequalities for product measures.
ESAIM: Probability and Statistics, 1:63–87, 1997.
26
[RSS15]
[Rub16]
[RW15]
Aviad Rubinstein, Lior Seeman, and Yaron Singer. Approximability of adaptive
seeding under knapsack constraints. In Proceedings of the Sixteenth ACM Confer-
ence on Economics and Computation, EC '15, Portland, OR, USA, June 15-19,
2015, pages 797–814, 2015.
Aviad Rubinstein. Beyond matroids: secretary problem and prophet inequal-
ity with general constraints. In Proceedings of the 48th Annual ACM SIGACT
Symposium on Theory of Computing, STOC 2016, Cambridge, MA, USA, June
18-21, 2016, pages 324–332, 2016.
Aviad Rubinstein and S. Matthew Weinberg. Simple mechanisms for a subad-
ditive buyer and applications to revenue monotonicity.
In Proceedings of the
Sixteenth ACM Conference on Economics and Computation, EC '15, Portland,
OR, USA, June 15-19, 2015, pages 377–394, 2015.
[Sch99]
Gideon Schechtman. Concentration, results and applications, 1999.
[SG08]
[Var15]
[Yan11]
Matthew Streeter and Daniel Golovin. An online algorithm for maximizing sub-
modular functions. In Advances in Neural Information Processing Systems, pages
1577–1584, 2008.
Shai Vardi. The returning secretary. In 32nd International Symposium on The-
oretical Aspects of Computer Science, STACS 2015, March 4-7, 2015, Garching,
Germany, pages 716–729, 2015.
Qiqi Yan. Mechanism design via correlation gap. In Proceedings of the twenty-
second annual ACM-SIAM symposium on Discrete Algorithms, pages 710–719.
SIAM, 2011.
A Missing Proofs
Lemma 3.6. Consider any submodular function f and L, H ∈ [0, 1]. Let S∗ be a set that
maximizes f , and let x ∈ [0, 1]n such that for all i ∈ [n], L ≤ xi ≤ H. Then,
F (x) ≥ (1 − H)(1 − L)f (∅) + (1 − H)L · f (S∗)
Proof. Assume, wlog, that S∗ = {1, . . . , k}. By submodularity, at each step after we add
another element, the potential marginal gain of all other elements decreases. In particular,
if we add the elements in S∗ in any order, they all have non-negative marginal contribution
(since each has a non-negative marginal contribution when added last).
Let x≤i denote the restriction of x to [i]. We first show by induction that for every i ≤ k,
F (x≤i) ≥ (1 − L)f (∅) + L · f ([i]). Denote Si , S ∩ [i]. We have that F (x≤i) is at least:
ES∼xhf (Si)i ≥ ES∼xhf (Si−1) + f (Si ∪ [i − 1]) − f ([i − 1])i
= F (x≤i−1) + ES∼xhf (Si ∪ [i − 1]) − f ([i − 1])i
≥ F (x≤i−1) + xi(cid:16)f ([i]) − f ([i − 1])(cid:17)
≥ F (x≤i−1) + L(cid:16)f ([i]) − f ([i − 1])(cid:17)
(Submodularity)
(Submodularity)
(f ([i]) − f ([i − 1]) ≥ 0).
27
Finally by the induction hypothesis, F (x≤i−1) ≥ (1 − L)f (∅) + L · f ([i − 1]). In particular,
we now have that
F (x≤k) ≥ (1 − L)f (∅) + L · f (S∗).
It is left to argue that the rest of the elements do not hurt the value too much. Consider
any Si ⊆ S∗, and let B∗ = B∗(Sk) = {k + 1, . . . , ℓ} be the worst set that we could add to Sk,
i.e. the B that minimizes f (Sk ∪ B). Let Bj , {k + 1, . . . , ℓ} ⊆ B∗, and let Tj denote the
intersection of Bj with a set T ⊆ T ∗ sampled according to x. Now consider two options for
adding elements from T ∗ to Sk:
1. deterministically, or
2. independently at random with probabilities sampled according to x.
Since f is non-negative, we have that even when we add all the bad elements deterministically,
X f (Sk ∪ Bj) − f (Sk ∪ Bj−1) = f (Sk ∪ B∗) − f (Sk) ≥ −f (Sk).
(A.1)
When we add the elements at random, we have (by submodularity) that the marginal contri-
bution of each bad element can only increase compared to its contribution in the first case.
Therefore,
ET [f (Sk ∪ T ) − f (Sk)] =X ET [f (Sk ∪ Tj) − f (Sk ∪ Tj−1)]
≥X xj (f (Sk ∪ Bj) − f (Sk ∪ Bj−1))
≥X H (f (Sk ∪ Bj) − f (Sk ∪ Bj−1))
≥ −Hf (Sk)
(Submodularity)
(f (Sk ∪ Bj) − f (Sk ∪ Bj−1) ≤ 0)
(Inequality (A.1)).
So far we have ET [f (Sk ∪ T )] ≥ (1 − H)f (Sk). Finally, the marginal contribution of the
remaining elments {ℓ + 1, . . . , n} is non-negative by submodularity (if it were negative, we
could get a worse set B′). Therefore, for every Sk, adding the rest of the elements can decrease
the value by at most a factor of 1 − H.
28
|
1804.08236 | 1 | 1804 | 2018-04-23T03:21:10 | How Bad is the Freedom to Flood-It? | [
"cs.DS"
] | Fixed-Flood-It and Free-Flood-It are combinatorial problems on graphs that generalize a very popular puzzle called Flood-It. Both problems consist of recoloring moves whose goal is to produce a monochromatic ("flooded") graph as quickly as possible. Their difference is that in Free-Flood-It the player has the additional freedom of choosing the vertex to play in each move. In this paper, we investigate how this freedom affects the complexity of the problem. It turns out that the freedom is bad in some sense. We show that some cases trivially solvable for Fixed-Flood-It become intractable for Free-Flood-It. We also show that some tractable cases for Fixed-Flood-It are still tractable for Free-Flood-It but need considerably more involved arguments. We finally present some combinatorial properties connecting or separating the two problems. In particular, we show that the length of an optimal solution for Fixed-Flood-It is always at most twice that of Free-Flood-It, and this is tight. | cs.DS | cs | How Bad is the Freedom to Flood-It?
Rémy Belmonte
The University of Electro-Communications, Chofu, Tokyo, Japan
[email protected]
Mehdi Khosravian Ghadikolaei
Université Paris-Dauphine, PSL Research University, CNRS, UMR,
LAMSADE, 75016 Paris, France
[email protected]
Masashi Kiyomi
Yokohama City University, Yokohama, Japan
[email protected]
Michael Lampis
Université Paris-Dauphine, PSL Research University, CNRS, UMR,
LAMSADE, 75016 Paris, France
[email protected]
Yota Otachi
Kumamoto University, Kumamoto, Japan
[email protected]
0000-0002-0087-853X
Abstract
Fixed-Flood-It and Free-Flood-It are combinatorial problems on graphs that generalize a
very popular puzzle called Flood-It. Both problems consist of recoloring moves whose goal is to
produce a monochromatic ("flooded") graph as quickly as possible. Their difference is that in
Free-Flood-It the player has the additional freedom of choosing the vertex to play in each
move. In this paper, we investigate how this freedom affects the complexity of the problem. It
turns out that the freedom is bad in some sense. We show that some cases trivially solvable for
Fixed-Flood-It become intractable for Free-Flood-It. We also show that some tractable
cases for Fixed-Flood-It are still tractable for Free-Flood-It but need considerably more
involved arguments. We finally present some combinatorial properties connecting or separating
In particular, we show that the length of an optimal solution for Fixed-
the two problems.
Flood-It is always at most twice that of Free-Flood-It, and this is tight.
2012 ACM Subject Classification Mathematics of computing → Graph algorithms, Theory of
computation → Parameterized complexity and exact algorithms
Keywords and phrases flood-filling game, parameterized complexity
Funding This work is partially supported by JSPS and MAEDI under the Japan-France Inte-
grated Action Program (SAKURA) Project GRAPA 38593YJ.
1
Introduction
Flood-It is a popular puzzle, originally released as a computer game in 2006 by LabPixies (see
[2]). In this game, the player is presented with (what can be thought of as) a vertex-colored
grid graph, with a designated special pivot vertex, usually the top-left corner of the grid.
In each move, the player has the right to change the color of all vertices contained in the
same monochromatic component as the pivot to a different color of her choosing. Doing this
8
1
0
2
r
p
A
3
2
]
S
D
.
s
c
[
1
v
6
3
2
8
0
.
4
0
8
1
:
v
i
X
r
a
23:2
How Bad is the Freedom to Flood-It?
judiciously gradually increases the size of the pivot's monochromatic component, until the
whole graph is flooded with one color. The goal is to achieve this flooding with the minimum
number of moves. See Figure 1 for an example.
Figure 1 A flooding sequence on a 3 × 3 grid. Each move in this example changes the color of
the top-left monochromatic component. Under such a restriction, the depicted sequence is shortest.
Following the description above, Flood-It immediately gives rise to a natural optimization
problem: given a vertex-colored graph, determine the shortest sequence of flooding moves
that wins the game. This problem has been extensively studied in the last few years (e.g.
[13, 16, 18, 17, 10, 6, 21, 7, 19, 12]; a more detailed summary of known results is given
below), both because of the game's popularity (and addictiveness!), but also because the
computational complexity questions associated with this problem have turned out to be
surprisingly deep, and the problem has turned out to be surprisingly intractable.
The goal of this paper is to add to our understanding of this interesting, puzzle-inspired,
optimization problem, by taking a closer look at the importance of the pivot vertex. As
explained above, the classical version of the game only allows the player to change the
color of a special vertex and its component and has been studied under the name Fixed-
Flood-It [16, 18, 17] (or Flood-It in some papers [2, 21, 6, 7, 12]). However, it is
extremely natural to also consider a version where the player is also allowed to play a
different vertex of her choosing in each turn. This has also been well-studied under the name
Free-Flood-It [2, 13, 16, 18, 17, 6, 21]. See Figure 2.
Figure 2 A flooding sequence with no restriction on selected monochromatic components. This
is shorter than then one in Figure 1.
Since both versions of this problem have been studied before, the question of the impact
of the pivot vertex on the problem's structure has (at least implicitly) been considered.
Intuitively, one would expect Free-Flood-It to be a harder problem; after all, the player
has to choose a color to play and a vertex to play it on, and is hence presented with a
larger set of possible moves. The state of the art seems to confirm this intuition, as only
some of the positive algorithmic results known for Fixed-Flood-It are known also for
Free-Flood-It, while there do exist some isolated cases where Fixed-Flood-It is tractable
and Free-Flood-It is hard, for example co-comparability graphs [8, 10] and grids of height
2 [2, 17]. Nevertheless, these results do not completely pinpoint the added complexity brought
by the task of selecting a vertex to play, as the mentioned algorithms for Fixed-Flood-It
are already non-trivial, and hence the jump in complexity is likely to be the result of the
combination of the tasks of picking a color and a vertex. More broadly, [6] presented a
generic reduction from Fixed-Flood-It to Free-Flood-It that preserves a number of
11122233331122233331112221111222222222211122233311222111111111111111 R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:3
crucial parameters (number of colors, optimal value, etc.) and gives convincing evidence that
Free-Flood-It is always at least as hard as Fixed-Flood-It, but not necessarily harder.
Our Results We investigate the complexity of Free-Flood-It, mostly from the point of
view of parameterized complexity,1 as well as the impact on the combinatorics of the game
of allowing moves outside the pivot.
Our first result is to show that Free-Flood-It is W[2]-hard parameterized by the number
of moves in an optimal solution. We recall that for Fixed-Flood-It this parameterization
is trivially fixed-parameter tractable: when a player has only k moves available, then we can
safely assume that the graph uses at most (roughly) k colors, hence one can easily consider
all possible solutions in FPT time. The interest of our result is, therefore, to demonstrate
that the task of deciding which vertex to play next is sufficient to make Free-Flood-It
significantly harder than Fixed-Flood-It. Indeed, the W[2]-hardness reduction we give,
implies also that Free-Flood-It is not solvable in no(k) time under the ETH. This tightly
matches the complexity of a trivial algorithm which considers all possible vertices and colors
to be played. This is the first concrete example showing a case where Fixed-Flood-It is
essentially trivial, but Free-Flood-It is intractable.
Motivated by this negative result we consider several other parameterizations of the
problem. We show that Free-Flood-It is fixed-parameter tractable when parameterized
by the number of possible moves and the clique-width. This result is tight in the sense that
the problem is hard when parameterized by only one of these parameters. It also implies
the fixed-parameter tractability of the problem parameterized by the number of colors and
the modular-width. In a similar vein, we present a polynomial kernel when Free-Flood-It
is parameterized by the input graph's neighborhood diversity and number of colors. An
analogous result was shown for Fixed-Flood-It in [7], but because of the freedom to
select vertices, several of the tricks used there do not apply to Free-Flood-It, and our
proofs are slightly more involved. Our previously mentioned reduction also implies that
Free-Flood-It does not admit a polynomial kernel parameterized by vertex cover, under
standard assumptions. This result was also shown for Fixed-Flood-It in [7], but it does
not follow immediately for Free-Flood-It, as the reduction of [6] does not preserve the
graph's vertex cover.
Motivated by the above results, which indicate that the complexity of the problem can be
seriously affected if one allows non-pivot moves, we also study some more purely combinatorial
questions with algorithmic applications. The main question we pose here is the following.
It is obvious that for all instances the optimal number of moves for Free-Flood-It is
upper-bounded by the optimal number of moves for Fixed-Flood-It (since the player has
strictly more choices), and it is not hard to construct instances where Fixed-Flood-It needs
strictly more moves. Can we bound the optimal number of Fixed-Flood-It moves needed
as a function of the optimal number of Fixed-Flood-It moves? Somewhat surprisingly, this
extremely natural question does not seem to have been explicitly considered in the literature
before. Here, we completely resolve it by showing that the two optimal values cannot be
more than a factor of 2 apart, and constructing a family of simple instances where they are
exactly a factor of 2 apart. As an immediate application, this gives a 2-approximation for
Free-Flood-It for every case where Fixed-Flood-It is known to be tractable.
We also consider the problem's monotonicity: Fixed-Flood-It has the nice property
that even an adversary that selects a single bad move cannot increase the optimal (that is, in
1 For readers unfamiliar with the basic notions of this field, we refer to standard textbooks [4, 9].
23:4
How Bad is the Freedom to Flood-It?
the worst case a bad move is a wasted move). We construct minimal examples which show
that Free-Flood-It does not have this nice monotonicity property, even for extremely
simple graphs, that is, making a bad move may not only waste a move but also make the
instance strictly worse. Such a difference was not explicitly stated in the literature, while
the monotonicity of Fixed-Flood-It was seem to be known or at least assumed. The only
result we are aware of is the monotonicity of Free-Flood-It on paths shown by Meeks and
Scott [16].
Known results
In 2009, the NP-hardness of Fixed-Flood-It with six colors was sketched by Elad Verbin
as a comment to a blog post by Sariel Har-Peled [23]. Independently to the blog comment,
Clifford et al. [2] and Fleischer and Woeginger [8] started investigations of the complexity of
the problem, and published the conference versions of their papers at FUN 2010. Here we
mostly summarize some of the known results on Free-Flood-It. For more complete lists of
previous result, see e.g. [10, 13, 7].
Free-Flood-It is NP-hard if the number of colors is at least 3 [2] even for trees with
only one vertex of degree more than 2 [13, 6], while it is polynomial-time solvable for general
graphs if the number of colors is at most 2 [2, 16, 13]. Moreover, it is NP-hard even for
height-3 grids with four colors [16]. Note that this result implies that Free-Flood-It with
a constant number colors is NP-hard even for graphs of bounded bandwidth. If the number
of colors is unbounded, then it is NP-hard for height-2 grids [17], trees of radius 2 [6], and,
proper interval graphs and caterpillars [10]. Also, it is known that there is no constant-factor
approximation with a factor independent of the number of colors unless P = NP [2].
There are a few positive results on Free-Flood-It. Meeks and Scott [18] showed that
every colored graph has a spanning tree with the same coloring such that the minimum
number of moves coincides in the graph and the spanning tree. Using this property, they
showed that if a graph has only a polynomial number of vertex subsets that induce connected
subgraphs, then Free-Flood-It (and Fixed-Flood-It) on the graph can be solved in
polynomial time. This in particular implies the polynomial-time solvability on subdivisions
of a fixed graph. It is also known that Free-Flood-It for interval graphs and split graphs
is fixed-parameter tractable when parameterized by the number of colors [10].
Preliminaries
2
For a positive integer k, we use [k] to denote the set {1, . . . , k}. Given a graph G = (V, E), a
coloring function col: V → [cmax], where cmax is a positive integer, and u ∈ V , we denote by
Comp(col, u) the maximal set of vertices S such that for all v ∈ S, col(u) = col(v) and there
exists a path from u to v such that for all its internal vertices w we have col(w) = col(u).
In other words, Comp(col, u) is the monochromatic connected component that contains u
under the coloring function col.
Given G, col, a move is defined as a pair (u, i) where u ∈ V , i ∈ [cmax]. The result
of the move (u, c) is a new coloring function col0 defined as follows: col0(v) = c for all
v ∈ Comp(col, u); col0(v) = col(v) for all other vertices. In words, a move consists of
changing the color of u, and of all vertices in the same monochromatic component as u,
to c. Given the above definition we can also define the result of a sequence of moves
(u1, c1), (u2, c2), . . . , (uk, ck) on a colored graph with initial coloring function col0 in the
natural way, that is, for each i ∈ [k], coli is the result of move (ui, ci) on coli−1.
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:5
The Free-Flood-It problem is defined as follows: given a graph G = (V, E), an
integer k, and an initial coloring function col0, decide if there exists a sequence of k moves
(u1, c1), (u2, c2), . . . , (uk, ck) such that the result colk obtained by applying this sequence of
moves on col0 is a constant function (that is, ∀u, v ∈ V we have colk(u) = colk(v)).
In the Fixed-Flood-It problem we are given the same input as in the Free-Flood-It
problem, as well as a designated vertex p ∈ V (the pivot). The question is again if there
exists a sequence of moves such that colk is monochromatic, with the added constraint that
we must have ui = p for all i ∈ [k].
We denote by OPTFree(G, col), OPTFixed(G, col, p) the minimum k such that for the
input (G, col) (or (G, col, p) respectively) the Free-Flood-It problem (respectively the
Fixed-Flood-It problem) admits a solution.
2.1 Graph parameters
Vertex cover number: A set S ⊆ V is a vertex cover of a graph G = (V, E) if each edge
in E has at least one end point in S. The minimum size of a vertex cover of a graph is its
vertex cover number. By vc(G), we denote the vertex cover number of G.
Neighborhood diversity: Let G = (V, E) be a graph. For each vertex v ∈ V , we denote
the neighborhood of v by NG(v). The closed neighborhood of v is the set NG[v] := {v}∪NG(v).
We omit the subscript G when the underlying graph is clear from the context. Two vertices
u, v ∈ V are true twins in G if N[u] = N[v] and are false twins in G if N(u) = N(v). Two
vertices are twins if they are true twins or false twins. Note that true twins are adjacent
and false twins are not. The neighborhood diversity of G, denoted nd(G), is the minimum
number k such that V can be partitioned into k sets of twin vertices. It is known that
nd(G) ≤ 2vc(G) + vc(G) for every graph G [14]. Given a graph, its neighborhood diversity
and the corresponding partition into sets of twins can be computed in polynomial time [14];
in fact, using fast modular decomposition algorithms, the neighborhood diversity of a graph
can be computed in linear time [15, 22].
Modular-width: Let H be a graph with k ≥ 2 vertices v1, . . . , vk, and let H1, . . . , Hk
be k graphs. The substitution H(H1, . . . , Hk) of the vertices of H by H1, . . . , Hk is the graph
1≤i≤k V (Ei)∪{{u, w} u ∈ V (Hi), w ∈
V (Hj),{vi, vj} ∈ E(H)}. For each i, the set V (Hi) is called a module of H(H1, . . . , Hk).
The modular-width of G, denoted mw(G), is defined recursively as follows:
1≤i≤k V (Hi) and the edge setS
with the vertex setS
If G has only one vertex, then mw(G) = 1.
If G is the disjoint union of graphs G1, . . . , Gh, then mw(G) = max1≤i≤h mw(Gi).
If G is a connected graph with two or more vertices, then
mw(G) =
min
H,H1,...,HV (H)
max{V (H), mw(H1), . . . , mw(HV (H))},
where the minimum is taken over all tuples of graphs (H, H1, . . . , Hk) such that G =
H(H1, . . . , HV (H)).
A recursive substitution structure giving the modular-width can be computed in linear-
time [15, 22]. It is known that mw(G) ≤ nd(G) for every graph G [11].
Clique-width: A k-expression is a rooted binary tree such that
each leaf has label ◦i for some i ∈ {1, . . . , k},
each non-leaf node with two children has label ∪, and
each non-leaf node with only one child has label ρi,j or ηi,j (i, j ∈ {1, . . . , k}, i 6= j).
a ◦i-node represents a graph with one i-vertex;
Each node in a k-expression represents a vertex-labeled graph as follows:
23:6
How Bad is the Freedom to Flood-It?
a ∪-node represents the disjoint union of the labeled graphs represented by its children;
a ρi,j-node represents the labeled graph obtained from the one represented by its child
by replacing the labels of the i-vertices with j;
an ηi,j-node represents the labeled graph obtained from the one represented by its child
by adding all possible edges between the i-vertices and the j-vertices.
A k-expression represents the graph represented by its root. The clique-width of a graph G,
denoted by cw(G), is the minimum integer k such that there is a k-expression representing a
graph isomorphic to G. From their definitions, cw(G) ≤ mw(G) holds for every graph G.
Figure 3 shows relationships among the graph parameters introduced above together with
the well-known treewidth and pathwidth (see [4] for definitions of these two parameters).
Figure 3 Graph parameters. Each segment implies that the one above is more general than the
one below. For example, bounded modular-width implies bounded clique-width but not vice versa.
3 W[2]-hardness of Free-Flood-It
The main result of this section is that Free-Flood-It is W[2]-hard when parameterized by
the minimum length of any valid solution (the natural parameter). The proof consists of a
reduction from Set Cover, a canonical W[2]-complete problem.
Before presenting the construction, we recall two basic observations by Meeks and Vu [19],
both of which rest on the fact that any single move can (at most) eliminate a single color
from the graph, and this can only happen if a color induces a single component.
(cid:73) Lemma 3.1 ([19]). For any graph G = (V, E), and coloring function col that uses cmax
distinct colors, we have OPTFree(G, col) ≥ cmax − 1.
(cid:73) Lemma 3.2 ([19]). For any graph G = (V, E), and coloring function col that uses cmax
distinct colors, such that for all c ∈ [cmax], G[col−1(c)] is a disconnected graph, we have
OPTFree(G, col) ≥ cmax.
The proof of Theorem 3.6 relies on a reduction from a special form of Set Cover, which
we call Multi-Colored Set Cover (MCSC for short). MCSC is defined as follows:
(cid:73) Definition 3.3. In Multi-Colored Set Cover (MCSC) we are given as input a set of
elements R and k collections of subsets of R, S1, . . . ,Sk. We are asked if there exist k sets
S1, . . . , Sk such that for all i ∈ [k], Si ∈ Si, and ∪i∈[k]Si = R.
Observe that MCSC is just a version of Set Cover where the collection of sets is given
to us pre-partitioned into k parts and we are asked to select one set from each part to form
a set cover of the universe. It is not hard to see that any Set Cover instance (S, R) where
we are asked if there exists a set cover of size k can easily be transformed to an equivalent
MCSC instance simply by setting Si = S for all i ∈ [k], since the definition of MCSC does
clique-widthtreewidthpathwidthmodular-widthneighborhooddiversityvertexcovernumber R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:7
...
...
...
...
...
...
L1
I1
L2
I2
Lk
Ik
...
...
...
...
...
...
...
...
R
...
...
...
u
Figure 4 The graph G = (V, E) of Free-Flood-It constructed from the given MCSC instance.
All the vertices in each Ii have color i and all black vertices have color k + 1. Boxes containing black
vertices have size 3k. Also each vertex in Li has k neighbors with degree 1 colored 1, ..., k.
not require that the sub-collections Si be disjoint. We conclude that known hardness results
for Set Cover immediately transfer to MCSC, and in particular MCSC is W[2]-hard when
parameterized by k.
Construction
We are now ready to describe our reduction which, given a MCSC instance with universe R
and k collections of sets Si, i ∈ [k], produces an equivalent instance of Free-Flood-It, that
is, a graph G = (V, E) and a coloring function col on V . We construct this graph as follows:
for every set S ∈ Si, construct a vertex in V . The set of vertices in V corresponding
to sets of Si is denoted by Ii and col(v) = i for each v ∈ Ii. I1 ∪ ... ∪ Ik induces an
independent set colored {1, ..., k}.
for each i ∈ [k], construct 3k new vertices, denoted by Li and connect all of them to all
vertices of Ii such that Li ∪ Ii induces a complete bipartite graph of size 3k × Ii. Then
set col(v) = k + 1 for each v ∈ Li, for all i ∈ [k].
for each vertex v ∈ Li for 1 ≤ i ≤ k, construct k new leaf vertices connected to v with
distinct colors 1, ..., k.
for each element e ∈ R, construct a vertex e. For each S ∈ Si such that e ∈ S we connect
e to the vertex of Ii that represents S.
add a special vertex u with col(u) = k + 1 which is connected it to all vertices in Ii for
i ∈ [k].
An illustration of G is shown in Fig.4. In the following we will show that (G, col) as
an instance of Free-Flood-It is solvable with at most 2k moves if and only if the given
MCSC instance has a set cover of size k which contains one set of each Si.
(cid:73) Lemma 3.4. If (S1, . . . ,Sk, R) is a YES instance of MCSC then OPTFree(G, col) ≤ 2k.
23:8
How Bad is the Freedom to Flood-It?
Proof. Suppose that there is a solution S1, . . . , Sk of the given MCSC instance, with Si ∈ Si,
for i ∈ [k] and ∪i∈[k]Si = R. Recall that for each Si there is a vertex in Ii in the constructed
graph representing Si. Our first k moves consist of changing the color of each of these k
vertices to k + 1 in some arbitrary order.
Observe that in the graph resulting after these k moves the vertices with color k + 1 form
a single connected component: because ∪Si is a set cover, all vertices of R have a neighbor
with color k + 1; all vertices with color k + 1 in some Ii are in the same component as u; and
all vertices of ∪i∈[k]Li are connected to one of the vertices we played. Furthermore, observe
that this component dominates the graph: all remaining vertices of ∪Ii, as well as all leaves
attached to vertices of ∪i∈[k]Li are dominated by the vertices of ∪i∈[k]Li. Hence, we can
select an arbitrary vertex with color k + 1, say u, and cycle through the colors 1, . . . , k on
(cid:74)
this vertex to make the graph monochromatic.
Now we establish the converse of Lemma 3.4.
(cid:73) Lemma 3.5. If OPTFree(G, col) ≤ 2k, then (S1, . . . ,Sk, R) is a YES instance of MCSC.
Proof. Suppose that there exists a sequence of at most 2k moves solving (G, col). We can
assume without loss of generality that the sequence has length exactly 2k, since performing a
move on a monochromatic graph keeps the graph monochromatic. Let (u1, c1), . . . , (u2k, c2k)
be a solution, let col0 = col, and let coli denote the coloring of G obtained after the first i
moves. The key observation that we will rely on is the following:
(i) For all i ∈ [k], there exist j ∈ [k], v ∈ Ii such that colj(v) = k + 1.
In other words, we claim that for each group Ii there exists a vertex that received color
k + 1 at some point during the first k moves. Before proceeding, let us prove this claim.
Suppose for contradiction that the claim is false. Then, there exists a group Ii such that no
vertex in that group has color k + 1 in any of the colorings col0, . . . , colk. We now consider
the vertices of Li and their attached leaves. Since Li contains 3k > k + 2 vertices, there
exist two vertices v1, v2 of Li such that {u1, . . . , uk} contains neither v1, v2, nor any of their
attached leaves. In other words, there exist two vertices of Li on which the winning sequence
does not change colors by playing them or their private neighborhood directly. However,
since v1, v2 only have neighbors in I1 (except for their attached leaves), and no vertex of
I1 received color k + 1, we conclude that colk(v1) = colk(v2) = k + 1, that is, the colors
of these two vertices have remained unchanged, and the same is true for their attached
leaves. Consider now the graph G with coloring colk: we observe that this coloring uses
k + 1 distinct colors, and that each color induces a disconnected graph. This is true for colors
1, . . . , k because of the leaves attached to v1, v2, and true of color k + 1 because of v1, v2 and
the fact that no vertex of Ii has color k + 1. We conclude that OPTFree(G, colk) ≥ k + 1 by
Lemma 3.2, which is a contradiction, because the whole sequence has length 2k.
Because of claim (i) we can now conclude that for all i ∈ [k] there exists a j ∈ [k] such
that colj−1(uj) = i. In other words, for each color i there exists a move among the first k
moves of the solution that played a vertex which at that point had color i. To see that this
is true consider again for contradiction the case that for some i ∈ [k] this statement does not
hold: this implies that vertices with color i in col0 still have color i in col1, . . . , colk, which
means that no vertex of Ii has received color k + 1 in the first k moves, contradicting (i).
As a result of the above, we therefore claim that for all j ∈ [k], we have colj−1(uj) 6= k+1.
In other words, we claim that none of the first k moves changes the color of a vertex that
at that point had color k + 1. This is because, as argued, for each of the other k colors,
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:9
there is a move among the first k moves that changes a vertex of that color. We therefore
conclude that for all vertices v for which col0(v) = k + 1 we have colj(v) = k + 1 for all
j ∈ [k]. In addition, because in col0 all colors induce independent sets, each of the first k
moves changes the color of a single vertex. Because of claim (i), this means that for each
i ∈ [k] one of the first k moves changes the color of a single vertex from Ii to k + 1. We
select the corresponding set of Si in our MCSC solution.
We now observe that, since all vertices of ∪i∈[k]Li retain color k + 1 throughout the first
k moves, colk is a coloring function that uses k + 1 distinct colors, and colors 1, . . . , k induce
disconnected graphs (because of the leaves attached to the vertices of each Li). Thanks to
Lemma 3.2, this means that col−1
k (k + 1) must induce a connected graph. Hence, all vertices
of R have a neighbor with color k + 1 in colk, which must be one of the k vertices played
in the first k moves; hence the corresponding element is dominated by our solution and we
have a valid set cover selecting one set from each Si.
(cid:74)
We are now ready to combine Lemmas 3.4 and 3.5 to obtain the main result of this
section.
(cid:73) Theorem 3.6. Free-Flood-It is W[2]-hard parameterized by OPTFree, that is, param-
eterized by the length of the optimal solution. Furthermore, if there is an algorithm that
decides if a Free-Flood-It instance has a solution of length k in time no(k), then the ETH
is false.
Proof. The described construction, as well as Lemmas 3.4 and 3.5 give a reduction from
MCSC, which is W[2]-hard parameterized by k, to an instance of Free-Flood-It with k +1
colors, where the question is to decide if OPTFree(G, col) ≤ 2k. Furthermore, it is known
that MCSC generalizes Dominating Set, which does not admit an algorithm running in
time no(k), under the ETH [4]. Since our reduction only modifies k by a constant, we odtain
(cid:74)
the same result for Free-Flood-It.
We note that because of Lemma 3.1 we can always assume that the number of colors
of a given instance is not much higher than the length of the optimal solution. As a
result, Free-Flood-It parameterized by OPTFree is equivalent to the parameterization
of Free-Flood-It by OPTFree + cmax and the result of Theorem 3.6 also applies to this
parameterization.
3.1 Kernel lower bound for Free-Flood-It
As a byproduct of the reduction above, we can show a kernel lower bound for Free-Flood-It
parameterized by the vertex cover number.
Let P and Q be parameterized problems. A polynomial-time computable function
f : Σ∗ × N → Σ∗ × N is a polynomial parameter transformation from P to Q if there is a
polynomial p such that for all (x, k) ∈ Σ∗ × N,
(x, k) ∈ P if and only if (x0, k0) = f(x, k) ∈ Q, and
k0 ≤ p(k).
If such a function exits, then P is polynomial parameter reducible to Q.
(cid:73) Proposition 3.7 ([1]). Let P and Q be parameterized problems, and P 0 and Q0 be unpa-
rameterized versions of P and Q, respectively. Suppose P 0 is NP-hard, Q0 is in NP, and
P is polynomial parameter reducible to Q. If Q has a polynomial kernel, then P also has a
polynomial kernel.
23:10
How Bad is the Freedom to Flood-It?
(cid:73) Theorem 3.8. Free-Flood-It parameterized by the vertex cover number admits no
polynomial kernel unless PH = Σp
3.
Proof. The reduction in this section can be seen as a polynomial parameter transformation
from MCSC parameterized by the solution size k and the size R of the universe to Free-
Flood-It parameterized by the vertex cover number with a polynomial p(k,R) = 3k2 +R.
To see this observe that the black vertices in Figure 4 form a vertex cover of size 3k2 + R.
Since MCSC is NP-hard and the decision version of Free-Flood-It is in NP, Proposi-
tion 3.7 implies that if Free-Flood-It parameterized by the vertex cover number has a
polynomial kernel, then MCSC parameterized by k and R also has a polynomial kernel.
It is known that Set Cover (and thus MCSC) parameterized simultaneously by k and
3 [5]. This completes the proof. (cid:74)
R does not admit a polynomial kernel unless PH = Σp
Clique-width and the number of moves
4
In this section, we consider as a combined parameter for Free-Flood-It the length of an
optimal solution and the clique-width. We show that this case is indeed fixed-parameter
tractable by using the theory of the monadic second-order logic on graphs. As an application
of this result, we also show that combined parameterization by the number of colors and the
modular-width is fixed-parameter tractable.
To prove the main claim, we show that Free-Flood-It with a constant length of optimal
solutions is an MSO1-definable decision problem. The syntax of MSO1 (one-sorted monadic
second-order logic) of graphs includes (i) the logical connectives ∨, ∧, ¬, ⇔, ⇒, (ii) variables
for vertices and vertex sets, (iii) the quantifiers ∀ and ∃ applicable to these variables, and
(iv) the following binary relations:
u ∈ U for a vertex variable u and a vertex set variable U;
adj(u, v) for two vertex variables u and v, where the interpretation is that u and v are
adjacent;
equality of variables.
If G models an MSO1 formula ϕ with an assignment X1, . . . , Xq ⊆ V (G) to the q free
variables in ϕ, then we write hG, X1, . . . , Xqi = ϕ.
It is known that, given a graph of clique-width at most w, an MSO1 formula ϕ, and an
assignment to the free variables in ϕ, the problem of deciding whether G models ϕ with the
given assignment is solvable in time O(f(ϕ, w) · n3), where f is a computable function and
ϕ is the length of ϕ [3, 20].
(cid:73) Theorem 4.1. Given an instance (G, col) of Free-Flood-It such that G has n vertices
and clique-width at most w, it can be decided in time O(f(k, w)·n3) whether OPTFree(G, col) ≤
k, where f is some computable function.
Proof. Let Vi (1 ≤ i ≤ cmax) be the set of color i vertices in the input graph. We construct
an MSO1 formula ϕ with cmax free variables X1, . . . , Xcmax such that OPTFree(G, col) ≤ k if
and only if G models ϕ with the assignment Xi := Vi for 1 ≤ i ≤ cmax. We can define the
desired formula ϕ(X1, . . . , Xcmax) as follows:
ϕ(X1, . . . , Xcmax) :=W
1≤c1,...,ck≤k ∃v1,v2,...,vk∈V (G)W
1≤c≤k ∀u∈V (G) colorc,k(u),
where colorc,i(u) for 0 ≤ i ≤ k implies that the color of u is c after the moves (v1, c1), . . . (vi, ci).
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:11
We define colorc,i(u) recursively as follows. We first set colorc,0(u) := (u ∈ Xi). This is
correct as we assign Vi to Xi. For 1 ≤ i ≤ k, we set
(colorc,i−1(u) ∨ SameCCCi−1(u, vi)
colorc,i(u) =
colorc,i−1(u) ∧ ¬SameCCCi−1(u, vi)
c = ci,
c 6= ci,
where SameCCCi(u1, u2) implies that u1 and u2 are in the same monochromatic component
after the moves (v1, c1), . . . (vi, ci). The formula precisely represent the recursive nature of
the color of the vertices. That is, a vertex u is of color c after the ith move (vi, ci) if and
only if either its color is changed to c by the i move, or it was already of color c before the
ith move and its color is not changed by the i move.
Given that colorc,i(u) is defined for all c and u, defining SameCCCi(u1, u2) is a routine:
SameCCCi(u1, u2) =W
1≤c≤k ∃S⊆V (G)(u1, u2 ∈ S) ∧ (∀u∈S colorc,i(u))
∧ (∀T⊆S (T = ∅) ∨ ∃x∈T∃y∈S\T adj(x, y)).
Since k ≥ cmax − 1 by Lemma 3.1, it holds that ϕ is bounded by a function of k. (cid:74)
(cid:73) Corollary 4.2. Given an integer k and an instance (G, col) of Free-Flood-It such that
G has n vertices and modular-width at most w, it can be decided in time O(f(cmax, w) · n3)
whether OPTFree(G, col) ≤ k, where f is some computable function.
Proof. Observe that for every connected graph G of modular-width at most w, it holds that
OPTFree(G, col) ≤ w + cmax − 2: we pick one vertex v from a module M; we next color one
vertex in each module except M with col(v); we then play at v with the remaining cmax − 1
colors. Thus we can assume that k ≤ w + cmax − 2. Since the modular-width of a graph is at
most its clique-width by their definitions, Theorem 4.1 gives an O(g(w + cmax, w) · n3)-time
algorithm for some computable g, which can be seen as an O(f(cmax, w) · n3)-time algorithm
(cid:74)
for some computable f.
5 Neighborhood diversity and the number of colors
Since the modular-width of a graph is upper bounded by its neighborhood diversity, Corol-
lary 4.2 in the previous section implies that Free-Flood-It is fixed-parameter tractable
when parameterized by both the neighborhood diversity and the number of colors. Here
we show that Free-Flood-It admits a polynomial kernel with the same parameterization.
This section is devoted to a proof of the following theorem.
(cid:73) Theorem 5.1. Free-Flood-It admits a kernel of at most nd(G)· cmax ·(nd(G)+ cmax −1)
vertices.
Fellows et al. [7] observed that for Fixed-Flood-It, a polynomial kernel with the same
parameterization can be easily obtained since twin vertices of the same color can be safely
contracted. In Free-Flood-It, this is true for true twins but not for false twins. See
Figure 5a.
Though it might be still possible to show something like "if there are more than some
constant number of false twins with the same color, then one can remove one of them without
changing the minimum number of moves," here we show a weaker claim. Our reduction rules
are as follows:
Rule TT: Let u and v be true twins of the same color in (G, col). Remove v.
23:12
How Bad is the Freedom to Flood-It?
(a) Removing a false twin is not safe.
(b) Removing a twin color is not safe.
Figure 5 Simple reductions that do not work for Free-Flood-It.
Rule FT: Let F be a set of false-twin vertices of the same color in (G, col) such that
F = nd(G) + cmax. Remove arbitrary one vertex in F.
Observe that after applying TT and FT exhaustively in polynomial time, the obtained
graph can have at most nd(G) · cmax · (nd(G) + cmax − 1) vertices. This is because each set of
twin vertices can contain at most nd(G) + cmax − 1 vertices. Hence, to prove Theorem 5.1, it
suffices to show the safeness of the rules.
(cid:73) Lemma 5.2. Rule TT is safe.
Proof. Let u and v be true twins of the same color. Observe that removing v is equivalent
to contracting the edge {u, v}. Since u and v are in the same monochromatic component,
(cid:74)
the lemma holds.
To guarantees the safeness of FT, we need the following technical lemmas.
(cid:73) Lemma 5.3. Let (G, col) be an instance of Free-Flood-It and x, y ∈ V (G) be false-twin
vertices of the same color c. A sequence (u1, c1), . . . , (uk, ck) with ui /∈ {x, y} for 1 ≤ i ≤ k
is a valid flooding sequence for (G, col) if and only if it is a valid flooding sequence for
(G − x, colG−x).
Proof. Let (G0, col0) = (G − x, colG−x). If a neighbor of x and y has color c, then the
lemma trivially holds. Hence, in what follows, we assume that none of the vertices adjacent
to x and y has color c. Assume that (u1, c1), . . . , (uk, ck) is valid for at least one of (G, col)
and (G0, col0). Then there is a move that changes the color of a neighbor of y to c since
ui 6= y for 1 ≤ i ≤ k. Let (ui, ci) be the first such move.
The first part (u1, c1), . . . , (ui−1, ci−1) of the sequence has the same effect to (G, col) and
(G0, col0). That is, the monochromatic components and connection among them are the
same in (G, coli−1) and (G0, col0
i−1) except that (G, coli−1) contains the monochromatic
component {x}. Note that {y} is a monochromatic component of color c in both (G, coli−1)
i−1), and {y} and {x} have the same adjacent monochromatic components in
and (G0, col0
(G, coli−1).
Let C be the set of monochromatic components of color c in (G, coli−1) that are adjacent
to Comp(coli−1, ui). Similarly, let C0 be the set of monochromatic components of color c
in (G0, col0
i−1, ui). Observe that Comp(coli−1, ui) =
Comp(col0
Now we apply the move (ui, ci). It follows that Comp(coli, ui) = Comp(coli−1, ui) ∪
C0∈C0 C0. Since Comp(coli, ui) \
C∈C C and Comp(col0
Comp(col0
i, ui) include y, the compo-
nents Comp(coli, ui) and Comp(col0
i, ui) have the same adjacent monochromatic compo-
nents. Also, for each u /∈ Comp(coli, ui), Comp(coli−1, u) = Comp(col0
the sequence, i.e. (ui+1, ci+1), . . . , (uk, ck), has the same effect to (G0, col0
Therefore, (G, colk) is constant if and only if so is (G0, col0
i) are equivalent and thus the remaining of
i) and (G, coli).
(cid:74)
i−1) that are adjacent to Comp(col0
i−1, ui), C0 = C \ {{x}}, and {y} ∈ C ∩ C0.
This implies that (G, coli) and (G0, col0
i−1, u) holds.
S
i, ui) = Comp(col0
i, ui) = {x} and both Comp(coli, ui) and Comp(col0
i−1, ui) ∪S
k).
222112221211232112113 R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:13
(cid:73) Lemma 5.4. Let S be a set of false-twin vertices with the same color in (G, col). If
S ≥ OPTFree(G, col) + 2, then OPTFree(G, col) = OPTFree(G − x, colG−x) for every
x ∈ S.
Proof. We first show that OPTFree(G, col) ≤ OPTFree(G − x, colG−x) for every x ∈ S. Let
(u1, c1), . . . , (uk, ck) be an optimal valid flooding sequence for (G − x, colG−x). Assume that
OPTFree(G, col) ≥ k (otherwise we are done). Since S \{x} ≥ OPTFree(G, col) + 1 ≥ k + 1,
there is a vertex y ∈ S\{x} such that ui 6= y for 1 ≤ i ≤ k. By Lemma 5.3, (u1, c1), . . . , (uk, ck)
is valid for (G, col) as well.
Next we show the other direction. Since S is a set of monochromatic false-twin
vertices, it suffices to show that OPTFree(G, col) ≥ OPTFree(G − x, colG−x) for some
x ∈ S. Let (u1, c1), . . . , (uk, ck) be an optimal valid flooding sequence of (G, col). Since
S ≥ OPTFree(G, col) + 2 = k + 2, there are two vertices x, y ∈ S such that ui /∈ {x, y} for
1 ≤ i ≤ k. By Lemma 5.3, (u1, c1), . . . , (uk, ck) is valid for (G − x, colG−x) as well.
(cid:74)
(cid:73) Corollary 5.5. Rule FT is safe.
Proof. Let G be a connected graph and col a coloring of G with cmax colors. Observe that
(G, col) admits a flooding sequence of length nd(G)+ cmax −2 as follows. Let T be a maximal
set of twin vertices of G and c be a color used in T. For each maximal set of twin vertices
T 0 6= T of G, pick a vertex u ∈ T 0 and play the move (u, c). After these nd(G) − 1 moves,
the vertices of color c form a connected dominating set of G. Now pick a vertex v of color c
and play the move (v, c0) for each c0 ∈ [cmax] \ {c}. These cmax − 1 moves make the coloring
constant.
Now for some color class C and a false-twin class I of (G, col), if C ∩ I ≥ cmax + nd(G),
then we can remove an arbitrary vertex in C ∩ I while preserving the optimal number of
(cid:74)
steps by Lemma 5. This implies the safeness of FT.
Note that using the concept of twin colors, Fellows et al. [7] further reduced the number
of colors in instances of Fixed-Flood-It and obtained a (nonpolynomial-size) kernel
parameterized by the neighborhood diversity. They say that two colors are twin if the colors
appear in the same family of the maximal sets of twin vertices. They observed that, in
Fixed-Flood-It, removing one of twin colors reduced the fewest number of moves exactly
by 1. Unfortunately, this is not the case for Free-Flood-It. See Figure 5b.
Relation Between Fixed and Free Flood-It
6
The main theorem of this section is the following:
(cid:73) Theorem 6.1. For any graph G = (V, E), coloring function col on G, and p ∈ V we have
OPTFree(G, col) ≤ OPTFixed(G, col, p) ≤ 2OPTFree(G, col).
Theorem 6.1 states that the optimal solutions for Free-Flood-It and Fixed-Flood-It
can never be more than a factor of 2 apart. It is worthy of note that we could not hope to
obtain a constant smaller than 2 in such a theorem, and hence the theorem is tight.
(cid:73) Theorem 6.2. There exist instances of Fixed-Flood-It such that OPTFixed(G, col, p) =
2OPTFree(G, col)
Proof. Consider a path on 2n + 1 vertices properly colored with colors 1, 2. If we set the
pivot to be one of the endpoints then OPTFree = 2n. However, it is not hard to obtain a
Free-Flood-It solution with n moves by playing every vertex at odd distance from the
(cid:74)
pivot.
23:14
How Bad is the Freedom to Flood-It?
Before we proceed to give the proof of Theorem 6.1, let us give a high-level description of
our proof strategy and some general intuition. The first inequality is of course trivial, so
we focus on the second part. We will establish it by induction on the number of non-pivot
moves performed by an optimal Free-Flood-It solution. The main inductive argument is
based on observing that a valid Free-Flood-It solution will either at some point play a
neighbor u of the component of p to give it the same color as p, or if not, it will at some
point play p to give it the same color as one of its neighbors. The latter case is intuitively
easier to handle, since then we argue that the move that changed p's color can be performed
first, and if the first move is a pivot move we can easily fall back on the inductive hypothesis.
The former case, which is the more interesting one, can be handled by replacing the single
move that gives u the same color as p, with two moves: one that gives p the same color as u,
and one that flips p back to its previous color. Intuitively, this basic step is the reason we
obtain a factor of 2 in the relationship between the two versions of the game.
The inductive strategy described above faces some complications due to the fact that
rearranging moves in this way may unintentionally re-color some vertices, which makes it
harder to continue the rest of the solution as before. To avoid this we define a somewhat
generalized version of Free-Flood-It, called Subset-Free-Flood-It.
(cid:73) Definition 6.3. Given G = (V, E), a coloring function col on G, and a pivot p ∈ V ,
a set-move is a pair (S, c), with S ⊆ V and S = Comp(col, u) for some u ∈ V , or
{p} ⊆ S ⊆ Comp(col, p). The result of (S, c) is the coloring col0 that sets col0(v) = c for
v ∈ S; and col0(v) = col(v) otherwise.
We define Subset-Free-Flood-It as the problem of determining the minimum number
of set-moves required to make a graph monochromatic, and Subset-Fixed-Flood-It as
the same problem when we impose the restriction that every move must change the color of
p, and denote as OPTS-Free, OPTS-Fixed the corresponding optimum values.
Informally, a set-move is the same as a normal move in Free-Flood-It, except that
we are also allowed to select an arbitrary connected monochromatic set S that contains p
(even if S is not maximal) and change its color. Intuitively, one would expect moves that
set S to be a proper subset of Comp(col, p) to be counter-productive, since such moves
split a monochromatic component into two pieces. Indeed, we prove below in Lemma 6.4
that the optimal solutions to Fixed-Flood-It and Subset-Fixed-Flood-It coincide, and
hence such moves do not help. The reason we define this version of the game is that it gives
us more freedom to define a solution that avoids unintentionally recoloring vertices as we
transform a given Free-Flood-It solution to a Fixed-Flood-It solution.
(cid:73) Lemma 6.4. For any graph G = (V, E), coloring function col on G, and pivot p ∈ V we
have OPTFixed(G, col, p) = OPTS-Fixed(G, col, p).
Proof. First, observe that OPTS-Fixed(G, col, p) ≤ OPTFixed(G, col, p) is trivial, as any
solution of Fixed-Flood-It is a solution to Subset-Fixed-Flood-It by playing the same
sequence of colors and always selecting all of the connected monochromatic component of p.
of Subset-Fixed-Flood-It, where by definition we have p ∈ Si for all i ∈ [k]. We would
like to prove that (p, c1), (p, c2), . . . , (p, ck) is a valid solution for Fixed-Flood-It. Let coli
be the result of the first i set-moves of the former solution, and col0
i be the result of the first
i moves of the latter solution. We will establish by induction the following:
1. For all i ∈ [k] we have Comp(coli, p) ⊆ Comp(col0
2. For all i ∈ [k], u ∈ V \ Comp(col0
Let us also establish the converse inequality. Consider a solution (S1, c1), (S2, c2), . . . , (Sk, ck)
i, p) we have coli(u) = col0
i(u).
i, p).
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:15
The statements are true for i = 0. Suppose that the two statements are true after i − 1
moves. The first solution now performs the set-move (Si, ci) with Si ⊆ Comp(coli−1, p) ⊆
Comp(col0
i−1, p). We now have that Comp(coli, p) contains Si plus the neighbors of Si
which have color ci in coli−1. Such vertices either also have color ci in col0
i−1, or are contained
in Comp(col0
i, p), which establishes
the first condition. To see that the second condition continues to hold observe that every
vertex for which coli−1(u) 6= coli(u) or col0
i, p); the
colors of other vertices remain unchanged. Since in the end Comp(colk, p) = V the first
condition ensures that Comp(col0
(cid:74)
i−1, p); in both cases they are included in Comp(col0
i−1(u) 6= col0
i(u) belongs in Comp(col0
k, p) = V .
We are now ready to state the proof of Theorem 6.1.
Proof of Theorem 6.1. As mentioned, we focus on proving the second inequality as the first
inequality follows trivially from the definition of the problems. Given a graph G = (V, E), an
initial coloring function col = col0, and a pivot p ∈ V , we suppose we have a solution to Free-
Flood-It (u1, c1), (u2, c2), . . . , (uk, ck). In the remainder, we denote by coli the coloring
that results after the moves (u1, c1), . . . , (ui, ci). We can immediately construct an equivalent
solution to Subset-Free-Flood-It from this, producing the same sequence of colorings:
(Comp(col0, u1), c1), (Comp(col1, u2), c2), . . . , (Comp(colk−1, uk), ck). We will transform
this solution to a solution of Subset-Fixed-Flood-It of length at most 2k, and then invoke
Lemma 6.4 to obtain a solution for Fixed-Flood-It of length at most 2k. More precisely,
we will show that for any G, col, p we have OPTS-Fixed(G, col, p) ≤ 2OPTS-Free(G, col, p).
For a solution S = (S1, c1), (S2, c2), . . . , (Sk, ck) to Subset-Free-Flood-It we define
the number of bad moves of S as b(S) = {(Si, ci) p 6∈ Si}. We will somewhat more strongly
prove the following statement for all G, col, p: for any valid Subset-Free-Flood-It solution
S, we have
OPTS-Fixed(G, col, p) ≤ S + b(S)
Since S + b(S) ≤ 2S, the above statement will imply the promised inequality and the
theorem.
We prove the statement by induction on S + 2b(S). If S + 2b(S) ≤ 2 then S is already
a Subset-Fixed-Flood-It solution, so the statement is trivial. Suppose then that the
statement holds when S + 2b(S) ≤ n and we have a solution S with S + 2b(S) = n + 1.
We consider the following cases:
• The first move (S1, c1) has p ∈ S1. By the inductive hypothesis there is a Subset-
Fixed-Flood-It solution of length at most S + b(S) − 1 for (G, col1, p). We build a
solution for Subset-Fixed-Flood-It by appending this solution to the move (S1, c1), since
this is a valid move for Subset-Fixed-Flood-It.
• There exists a move (Si, ci) with Si = Comp(coli−1, u), for some u ∈ N(Comp(coli−1, p))\
Comp(coli−1, p) such that ci = coli−1(p). That is, there exists a move that plays a vertex
u that currently has a different color than p, and as a result of this move the component of u
and p merge, because u receives the same color as p and u has a neighbor in the component
of p.
Consider the first such move. We build a solution S0 as follows: we keep moves
(S1, c1) . . . (Si−1, ci−1); we add the moves (Comp(coli−1, p), coli−1(u)), (Comp(coli−1, p)∪
Comp(coli−1, u), coli−1(p)); we append the rest of the previous solution (Si+1, ci+1), . . ..
To see that S0 is still a valid solution we observe that Comp(coli−1, p)∪Comp(coli−1, u)
is monochromatic and connected when we play it, and that the result of the first i − 1 moves,
plus the two new moves is exactly coli. We also note that S0 + b(S0) = S + b(S) because
we replaced one bad move with two good moves. However, S0 + 2b(S0) < S + 2b(S), hence
23:16
How Bad is the Freedom to Flood-It?
by the inductive hypothesis there exists a Subset-Fixed-Flood-It solution of the desired
length.
• There does not exist a move as specified in the previous case. We then show that this
reduces to the first case. If no move as described in the previous case exists and the initial
coloring is not already constant, S must have a move (Si, ci) where {p} ⊆ Si ⊆ Comp(col0, p)
and ci = coli−1(u) for u ∈ N(Comp(col0, p))\Comp(col0, p). In other words, this is a good
move (it changes the color of p), that adds a new vertex u to the connected monochromatic
component of p. Such a move must exist, since if the initial coloring is not constant, the
initial component of p must be extended, and we assumed that no move that extends it by
recoloring one of its neighbors exists.
Consider the first such good move (Si, ci) as described above. We build a solution S0
as follows: the first move is (Comp(col0, p), col0(u)), where u is, as described above, the
neighbor of Comp(col0, p) with coli−1(u) = ci. For j ∈ [i − 1] we add the move (Sj, cj) if
u 6∈ Sj, or the move (Comp(colj−1, u) ∪ Comp(col0, p), cj) if u ∈ Sj. In other words, we
keep other moves unchanged if they do not affect u, otherwise we add to them Comp(col0, p).
We observe that these moves are valid since we maintain the invariant that Comp(col0, p)
and u have the same color and since none of the first i − 1 moves of S changes the color of p
(since we selected the first such move). The result of these i moves is exactly coli. We now
append the remaining move (Si+1, ci+1), . . ., and we have a solution that starts with a good
move, has the same length and the same (or smaller) number of bad moves as S and is still
(cid:74)
valid. We have therefore reduced this to the first case.
As we mentioned before, this combinatorial theorem implies 2-approximability of Free-
Flood-It for the cases where Fixed-Flood-It is polynomial-time solvable. Also, as Fixed-
Flood-It admits a (cmax − 1) approximation [2],2 we have a 2(cmax − 1) approximation for
Free-Flood-It.
(cid:73) Corollary 6.5. Free-Flood-It admits a 2(cmax − 1) approximation, where cmax is the
number of used colors.
7 Non-monotonicity of Free-Flood-It
We now consider the (non-)monotonicity of the problem. A game has the monotonicity
property if no legal move makes the situation worse. That is, if Fixed-Flood-It (or
Free-Flood-It) has the monotonicity property, then no single move increases the minimum
number of steps to make the input graph monotone. We believe that the monotonicity of
Fixed-Flood-It was known as folklore and used implicitly in the literature. On the other
hand, we are not sure that the non-monotonicity of Free-Flood-It was widely known.
The only result we are aware of is by Meeks and Scott [16] who showed that on paths Free-
Flood-It has the monotonicity property. In the following, we show that Free-Flood-It
loses its monotonicity property as soon as the underlying graph becomes a path with one
attached vertex.
To be self-contained, we start with proving the following folklore, which says that Fixed-
Flood-It is monotone.
(cid:73) Lemma 7.1. Let (G, col) is an instance of Fixed-Flood-It with pivot p. For every
color c, it holds that OPTFixed(G, col0) ≤ OPTFixed(G, col), where (G, col0) is the result of
the move (p, c).
2 Their proof was only for grids, but it just works for the general case.
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:17
Figure 6 Non-monotonicity of Free-Flood-It.
Proof. Let (p, c1), . . . , (p, ck) be an optimal solution for (G, col). We show that this se-
quence is valid for (G, col0) too. Let col = col0 and for i ≥ 1, let coli be the coloring
obtained from coli−1 by applying the ith move (p, ci). We define col0
i in the analogous
way. Observe that since all moves are played on the pivot p, we have Comp(coli, v) ∈
{Comp(col, v), Comp(coli, p)} and Comp(col0
i, p)} for
every v ∈ V (G) and for every i.
i, v) ∈ {Comp(col0, v), Comp(col0
It suffices to show that Comp(coli, p) ⊆ Comp(col0
i, p) for all i. This obviously holds
i, p) for some i ≥ 0. Let v ∈ V (G) \
when i = 0. Assume that Comp(coli, p) ⊆ Comp(col0
{p} such that Comp(col, v) is not contained in but adjacent to Comp(coli, p), and thus
contained in Comp(coli+1, p). Since Comp(coli, p) ⊆ Comp(col0
i, p), the monochromatic
component Comp(col, v) is either contained in or adjacent to Comp(col0
i, p). Therefore,
i+1, p). This implies that Comp(coli+1, p) ⊆
Comp(col, v) is contained in Comp(col0
Comp(col0
(cid:74)
i+1, p).
Meeks and Scott [16] showed the following monotonicity of Free-Flood-It on paths.
(cid:73) Proposition 7.2 ([16, Lemma 4.1]). Let G be a path, col a vertex coloring of G, (v, c) a
move, and (G, col0) the result of the move. Then, OPTFree(G, col0) ≤ OPTFree(G, col).
One may wonder whether the monotonicity property holds in general for Free-Flood-It.
The following example, however, shows that it does not hold even for some graphs very close
to paths. See the two instances in Figure 6. The instance (G, col0) is obtained from (G, col)
by playing the move (v, 3). We show that OPTFree(G, col) < OPTFree(G, col0).
Observe that OPTFree(G, col) = 3: by Lemma 3.2, OPTFree(G, col) ≥ 3, and the sequence
(u, 2), (u, 1), (u, 1) floods the graph. Suppose that OPTFree(G, col0) ≤ 3. By Lemma 3.2,
the first move in each optimal solution of (G, col0) has to make the subgraph induced by
some color connected, and then the second move has to remove the connected color. We can
see that 2 is the only color that can play this role. If the first move is not played on u, then
it is played on one of the two color-2 vertices and then the second move is played on the
other color-2 vertex. Such a sequence of two moves cannot make either of color-1 or color-3
vertices connected. Thus by Lemma 3.2, it still needs at least two moves. Hence we can
conclude that the first move is (u, 2). Now the second move has to remove the color 2, and
thus has to be played on u (or equivalently on any vertex in the monochromatic component
including u). No matter which color we choose, we end up with an instance with at least
two colors that are not connected. Again by Lemma 3.2, this instance needs more than one
step, and thus OPTFree(G, col0) > 3.
References
1 Hans L. Bodlaender, Rodney G. Downey, Michael R. Fellows, and Danny Hermelin. On
problems without polynomial kernels. J. Comput. Syst. Sci., 75(8):423 -- 434, 2009. doi:
10.1016/j.jcss.2009.04.001.
2122113321231133vv(G,col)(G,col(cid:48))uu 23:18
How Bad is the Freedom to Flood-It?
2 Raphaël Clifford, Markus Jalsenius, Ashley Montanaro, and Benjamin Sach. The com-
plexity of flood filling games. Theory Comput. Syst., 50(1):72 -- 92, 2012. doi:10.1007/
s00224-011-9339-2.
3 Bruno Courcelle, Johann A. Makowsky, and Udi Rotics. Linear time solvable optimization
problems on graphs of bounded clique-width. Theory Comput. Syst., 33(2):125 -- 150, 2000.
doi:10.1007/s002249910009.
4 Marek Cygan, Fedor V. Fomin, Lukasz Kowalik, Daniel Lokshtanov, Dániel Marx, Marcin
Pilipczuk, Michal Pilipczuk, and Saket Saurabh. Parameterized Algorithms. Springer, 2015.
doi:10.1007/978-3-319-21275-3.
5 Michael Dom, Daniel Lokshtanov, and Saket Saurabh. Kernelization lower bounds through
colors and ids. ACM Trans. Algorithms, 11(2):13:1 -- 13:20, 2014. doi:10.1145/2650261.
6 Michael R. Fellows, Uéverton dos Santos Souza, Fábio Protti, and Maise Dantas da Silva.
Tractability and hardness of flood-filling games on trees. Theor. Comput. Sci., 576:102 -- 116,
2015. doi:10.1016/j.tcs.2015.02.008.
7 Michael R. Fellows, Fábio Protti, Frances A. Rosamond, Maise Dantas da Silva, and Uéver-
ton dos Santos Souza. Algorithms, kernels and lower bounds for the Flood-It game pa-
rameterized by the vertex cover number. Discrete Applied Mathematics, 2017.
in press.
doi:10.1016/j.dam.2017.07.004.
8 Rudolf Fleischer and Gerhard J. Woeginger. An algorithmic analysis of the honey-bee game.
Theor. Comput. Sci., 452:75 -- 87, 2012. doi:10.1016/j.tcs.2012.05.032.
Jörg Flum and Martin Grohe. Parameterized Complexity Theory. Texts in Theoretical
Computer Science. An EATCS Series. Springer, 2006.
9
10 Hiroyuki Fukui, Yota Otachi, Ryuhei Uehara, Takeaki Uno, and Yushi Uno. On com-
plexity of flooding games on graphs with interval representations.
In TJJCCGG 2012,
volume 8296 of Lecture Notes in Computer Science, pages 73 -- 84, 2013. doi:10.1007/
978-3-642-45281-9_7.
Jakub Gajarský, Michael Lampis, and Sebastian Ordyniak. Parameterized algorithms for
modular-width. In IPEC 2013, volume 8246 of Lecture Notes in Computer Science, pages
163 -- 176, 2013. doi:10.1007/978-3-319-03898-8_15.
12 Wing-Kai Hon, Ton Kloks, Fu-Hong Liu, Hsiang Hsuan Liu, and Hung-Lung Wang. Flood-
11
it on AT-free graphs. CoRR, abs/1511.01806, 2015. arXiv:1511.01806.
13 Aurélie Lagoutte, Mathilde Noual, and Eric Thierry. Flooding games on graphs. Discrete
Applied Mathematics, 164:532 -- 538, 2014. doi:10.1016/j.dam.2013.09.024.
14 Michael Lampis. Algorithmic meta-theorems for restrictions of treewidth. Algorithmica,
64(1):19 -- 37, 2012.
15 Ross M. McConnell and Jeremy P. Spinrad. Modular decomposition and transitive orien-
tation. Discrete Mathematics, 201(1-3):189 -- 241, 1999.
16 Kitty Meeks and Alexander Scott. The complexity of flood-filling games on graphs. Discrete
Applied Mathematics, 160(7 -- 8):959 -- 969, 2012. doi:10.1016/j.dam.2011.09.001.
17 Kitty Meeks and Alexander Scott. The complexity of Free-Flood-It on 2 × n boards.
Theor. Comput. Sci., 500:25 -- 43, 2013. doi:10.1016/j.tcs.2013.06.010.
18 Kitty Meeks and Alexander Scott. Spanning trees and the complexity of flood-filling games.
Theory Comput. Syst., 54(4):731 -- 753, 2014. doi:10.1007/s00224-013-9482-z.
19 Kitty Meeks and Dominik K. Vu. Extremal properties of flood-filling games. CoRR,
abs/1504.00596, 2015. arXiv:1504.00596.
Sang-il Oum. Approximating rank-width and clique-width quickly. ACM Transactions on
Algorithms, 5, 2008. Article No. 10. doi:10.1145/1435375.1435385.
21 Uéverton dos Santos Souza, Fábio Protti, and Maise Dantas da Silva. An algorithmic
analysis of Flood-it and Free-Flood-it on graph powers. Discrete Mathematics & Theoretical
Computer Science, 16(3):279 -- 290, 2014. URL: http://dmtcs.episciences.org/2086.
20
R. Belmonte, M. Khosravian Ghadikolaei, M. Kiyomi, M. Lampis, and Y. Otachi
23:19
22 Marc Tedder, Derek G. Corneil, Michel Habib, and Christophe Paul. Simpler linear-
In ICALP 2008
doi:
time modular decomposition via recursive factorizing permutations.
(1), volume 5125 of Lecture Notes in Computer Science, pages 634 -- 645, 2008.
10.1007/978-3-540-70575-8_52.
Elad Verbin. Comment to "Is this game NP-Hard? by Sariel Har-Peled. http://sarielhp.
org/blog/?p=2005#comment-993, 2009. Accessed: 2018-01-18.
23
|
1806.06430 | 1 | 1806 | 2018-06-17T19:05:49 | Subspace Embedding and Linear Regression with Orlicz Norm | [
"cs.DS",
"cs.LG"
] | We consider a generalization of the classic linear regression problem to the case when the loss is an Orlicz norm. An Orlicz norm is parameterized by a non-negative convex function $G:\mathbb{R}_+\rightarrow\mathbb{R}_+$ with $G(0)=0$: the Orlicz norm of a vector $x\in\mathbb{R}^n$ is defined as $
\|x\|_G=\inf\left\{\alpha>0\large\mid\sum_{i=1}^n G(|x_i|/\alpha)\leq 1\right\}. $
We consider the cases where the function $G(\cdot)$ grows subquadratically. Our main result is based on a new oblivious embedding which embeds the column space of a given matrix $A\in\mathbb{R}^{n\times d}$ with Orlicz norm into a lower dimensional space with $\ell_2$ norm. Specifically, we show how to efficiently find an embedding matrix $S\in\mathbb{R}^{m\times n},m<n$ such that $\forall x\in\mathbb{R}^{d},\Omega(1/(d\log n)) \cdot \|Ax\|_G\leq \|SAx\|_2\leq O(d^2\log n) \cdot \|Ax\|_G.$ By applying this subspace embedding technique, we show an approximation algorithm for the regression problem $\min_{x\in\mathbb{R}^d} \|Ax-b\|_G$, up to a $O(d\log^2 n)$ factor. As a further application of our techniques, we show how to also use them to improve on the algorithm for the $\ell_p$ low rank matrix approximation problem for $1\leq p<2$. | cs.DS | cs | Subspace Embedding and Linear Regression with Orlicz Norm
Alexandr Andoni 1 Chengyu Lin 1 Ying Sheng 1 Peilin Zhong 1 Ruiqi Zhong 1
8
1
0
2
n
u
J
7
1
]
S
D
.
s
c
[
1
v
0
3
4
6
0
.
6
0
8
1
:
v
i
X
r
a
Abstract
We consider a generalization of the classic lin-
ear regression problem to the case when the
loss is an Orlicz norm. An Orlicz norm is
parameterized by a non-negative convex func-
tion G : R+ → R+ with G(0) = 0:
the
Orlicz norm of a vector x ∈ Rn is defined
as kxkG = inf {α > 0 Pn
i=1 G(xi/α) ≤ 1} .
We consider the cases where the function G(·)
grows subquadratically. Our main result is based
on a new oblivious embedding which embeds the
column space of a given matrix A ∈ Rn×d with
Orlicz norm into a lower dimensional space with
ℓ2 norm. Specifically, we show how to efficiently
find an embedding matrix S ∈ Rm×n, m < n
such that ∀x ∈ Rd, Ω(1/(d log n)) · kAxkG ≤
kSAxk2 ≤ O(d2 log n) · kAxkG. By applying
this subspace embedding technique, we show an
approximation algorithm for the regression prob-
lem minx∈Rd kAx−bkG, up to a O(d log2 n) fac-
tor. As a further application of our techniques,
we show how to also use them to improve on the
algorithm for the ℓp low rank matrix approxima-
tion problem for 1 ≤ p < 2.
1. Introduction
Numerical linear algebra problems play a significant role
in machine learning, data mining, and statistics. One of
the most important such problems is the regression prob-
lem, see some recent advancements in, e.g., (Zhong et al.,
2016; Bhatia et al., 2015; Jain & Tewari, 2015; Liu et al.,
2014; Dhillon et al., 2013).
In a linear regression prob-
lem, given a data matrix A ∈ Rn×d with n data points
A1, A2,··· , An in Rd and the response vector b ∈ Rn, the
goal is to find a set of coefficients x∗ ∈ Rd such that
x∗ = arg minx∈Rd l(Ax − b),
(1)
*Equal contribution
1Computer Science Department,
Columbia University, New York City, NY 10027, U.S.A.. Cor-
respondence to: Peilin Zhong <[email protected]>.
Proceedings of the 35 th International Conference on Machine
Learning, Stockholm, Sweden, PMLR 80, 2018. Copyright 2018
by the author(s).
i=1 y2
2 = Pn
where l : Rn → R+ is the loss function. When l(y) =
kyk2
i , then the problem is the classic least
square regression problem (ℓ2 regression). While there has
been extensive research on efficient algorithms for solving
ℓ2 regression, it is not always a suitable loss function to use.
sion - with l(y) = kyk1 = Pn
In many settings, alternative choices for a loss function lead
to qualitatively better solutions x∗. For example, a popular
such alternative is the least absolute deviation (ℓ1) regres-
i=1 yi - which leads to
solutions that are more robust than those of ℓ2 regression
In a nutshell, the ℓ2 re-
(see (Wikipedia; Gorard, 2005).
gression is suitable when the data contains Gaussian noise,
whereas ℓ1 - when the noise is Laplacian or sparse.
A further popular class of loss functions l(·) arises from
M-estimators, defined as l(y) = Pn
i=1 M (yi) where M (·)
is an M-estimator function (see (Zhang, 1997) for a list of
M-estimators). The benefit of (some) M-estimators is that
they enjoy advantages of both ℓ1 and ℓ2 regression. For
example, when M (·) is the Huber function (Huber et al.,
1964), then the regression looks like ℓ2 regression when yi
is small, and looks like ℓ1 regression otherwise. However,
these loss functions come with a downside: they depend
on the scale, and rescaling the data may give a completely
different solution!
Our contributions. We introduce a generic algorithmic
technique for solving regression for an entire class of loss
functions that includes the aforementioned examples, and
in particular, a "scale-invariant" version of M-estimators.
Specifically, our class consists of loss functions l(y) that
are Orlicz norms, defined as follows: given a non-negative
convex function G : R+ → R+ with G(0) = 0, for x ∈
Rn, we can define kxkG to be an Orlicz norm with respect
to G(·): kxkG , inf {α > 0 Pn
i=1 G(xi/α) ≤ 1} .
Note that ℓp norm, for p ∈ [1,∞), is a special case of Or-
licz norm with G(x) = xp. Another important example is
the following "scale-free" version of M-estimator. Taking
f (·) to be a Huber function, i.e.
f (x) = (cid:26) x2/2
x ≤ δ
δ(x − δ/2) otherwise
for some constant δ, we take G(x) = f (f−1(1)x). Then
the norm kxkG looks like ℓ2 norm when x is flat, and looks
like ℓ1 norm when x is sparse. Figure 1 shows the unit
norm ball of this kind of Orlicz norm.
Subspace Embedding and Linear Regression with Orlicz Norm
Figure 1. Unit norm balls of Orlicz norm induced by normalized
Huber functions with different δ.
Our main result is a generic algorithm for solving any re-
gression problem Eqn. (1) with any loss function that is a
"nice" Orlicz norm; see Section 2 for a formal definition of
"nice", and think of it as subquadratic for now.
Our main result employs the concept of subspace em-
beddings, which is a powerful tool for solving numeri-
cal linear algebra problems, and as such has many appli-
cations beyond regression. We say that a subspace em-
bedding matrix S ∈ Rm×n embeds the column space
of A ∈ Rn×d (n > m) with u-norm into a sub-
space with v-norm, if ∀x ∈ Rd, we have kAxku/α ≤
kSAxkv ≤ βkAxku where αβ is called distortion (approx-
imation). A long line of work studied ℓ2 regression prob-
lem based on ℓ2 subspace embedding techniques; see, e.g.,
(Clarkson & Woodruff, 2009; 2013; Nelson & Nguyen,
2013). Furthermore, there are works on ℓp regression
problem based on ℓp subspace embedding techniques (see,
e.g. (Sohler & Woodruff, 2011; Meng & Mahoney, 2013;
Clarkson et al., 2013; Woodruff & Zhang, 2013)), and sim-
ilarly for M-estimators (Clarkson & Woodruff, 2015).
Our overall results are composed of four parts:
1. We develop the first subspace embedding method for
all "nice" Orlicz norms. The embedding obtains a
distortion factor polynomial in d, which was recently
shown necessary (Wang & Woodruff, 2018).
2. Using the above subspace embedding, we obtain the
first approximation algorithm for solving the linear re-
gression problem with any "nice" Orlicz norm loss.
3. As a further illustration of the power of the subspace
embedding method, we employ it towards improving
on the best known result for another problem: ℓp low
rank approximation for 1 ≤ p < 2 from (Song et al.,
2017), which is the "ℓp-version of PCA".
4. Finally, we complement our theoretical results with
experimental evaluation of our algorithms. Our exper-
iments reveal that that the solution of regression un-
der the Orlicz norm induced by Huber loss is much
better than the solution given by regression under ℓ1
or ℓ2 norms, under natural noise distributions in prac-
tice. We also perform experiments for Orlicz regres-
sion with different Orlicz functions G and show their
behavior under different noise settings, thus exhibiting
the flexibility of our framework.
To the best of our knowledge, our algorithms are the first
low distortion embedding and regression algorithms for
general Orlicz norm. For the problem of low rank approx-
imation under ℓp norm, p ∈ [1, 2), our algorithms achieve
simultaneously the best approximation and the best running
time. In contrast, all the previous algorithms achieve either
the best approximation, or the best running time, but not
both at the same time.
Our algorithms for subspace embedding and regression are
simple, and in particular are not iterative. In particular, for
the subspace embedding, the embedding matrix S is gener-
ated independently of the data. In the regression problem,
we multiply the input with the embedding matrix, and thus
reduce to the ℓ2 regression problem, for which we can use
any of the known algorithm.
Technical discussion. Next we highlight some of our
techniques used to obtain the theoretical results.
Subspace embedding. Our starting point is a technique
introduced in (Andoni et al., 2017) for the Orlicz norms,
which can be seen as an embedding that has guarantees for
a fixed vector only. In contrast, our main challenge here
ized exponential" random variable, i.e., drawn from a dis-
is to obtain an embedding for all vectors x ∈ Rn in a cer-
tain d-dimensional subspace. Consider a random diagonal
matrix D ∈ Rn×n with each diagonal entry is a "general-
tribution with cumulative distribution function 1 − e−G(x).
Then, for a fixed x ∈ Rd, (Andoni et al., 2017) show that
kD−1Axk∞ is not too small with high probability. We can
the dilation bound on kD−1AxkG, to argue that ∀x ∈ Rd,
kD−1Axk∞ is not too small.
combine this statement together with a net argument and
The other direction is more challenging - to show that
for a given matrix A ∈ Rn×d, and any fixed x ∈ Rd,
kD−1AxkG cannot be too large. Once we show this "di-
lation bound", we combine it with the well-conditioned ba-
sis argument (similar to (Dasgupta et al., 2009)), and prove
that ∀x ∈ Rd, kD−1AxkG cannot be too large. Overall, we
have that ∀x ∈ Rd, kD−1AxkG ≤ O(d2 log n) · kAxkG,
and kD−1Axk∞ ≥ Ω(1/(d log n)) · kAxkG. Since ℓ2
norm is sandwiched by k · kG and ℓ∞ norm, we have
that ∀x ∈ Rd, Ω(1/(d log n)) · kAxkG ≤ kD−1Axk2 ≤
O(d2 log n) · kAxkG. Then, the remaining part is to use
standard techniques (Woodruff & Zhang, 2013; Woodruff,
2014) to perform the ℓ2 subspace embedding for the col-
umn space of D−1A. See Theorem 16 for details.
The actual proof of the dilation bound is the most techni-
cally intricate result. Traditionally, since the pth power of
the ℓp norm is the sum of the pth power of all the entries,
it is easy to bound the expectation by using linearity of the
expectation. However it is impossible to apply this analysis
to Orlicz norm directly since Orlicz norm is not an "entry-
wise" norm. Instead, we exploit a key observation that the
Orlicz norm of vectors which are on the unit ball can be
Subspace Embedding and Linear Regression with Orlicz Norm
seen as the sum of contribution of each coordinate. Thus,
we propose a novel analysis for any fixed vector by analyz-
ing the behavior of the normalized vector which is on the
unit Orlicz norm ball. To extend the dilation bound for a
fixed vector to all the vectors in a subspace, we general-
ize the definition of ℓp norm well-conditioned basis to the
Orlicz norm case, and then show that the Auerbach basis
provides a good basis for Orlicz norm. To the best of our
knowledge, this is the first time Auerbach basis are used to
analyze the dilation bound of an embedding method for a
norm distinct from an ℓp norm. See Section 3 for details.
Regression with Orlicz norm. Here, given a matrix A ∈
Rn×d, a vector b ∈ Rn, the goal is to solve Equation 1 with
Orlicz norm loss function. We can now solve this prob-
lem directly using the subspace embedding from above,
in particular by applyingit to the column space of [A b].
We obtain an O(d3 log2 n) approximation ratio, which we
can further improve by observing that it actually suffices
to have the dilation bound on kD−1Ax∗kG only for the
optimal solution x∗ (as opposed to for an arbitrary x). Us-
ing this observation, we improve the approximation ratio
to O(d log2 n). See Theorem 18 for details. We evaluate
the algorithm's performance and show that it performs well
(even when n is not much larger than d). See Section 5.
This connection allows us to consider the following prob-
ℓp low rank matrix approximation. The ℓp norm is a spe-
cial case of the Orlicz norm k · kG, where G(x) = xp.
lem: given A ∈ Rn×d, n ≥ d ≥ k ≥ 1, find a rank-k
matrix B ∈ Rn×d such that kA − Bkp is minimized. Here
we consider the case of 1 ≤ p < 2 and k = ω(log n). The
best known algorithm for this problem is from (Song et al.,
2017), which uses the dense p-stable transform to achieves
k2 · poly(log n) approximation ratio.
It has the down-
side that its runtime does not compare favorably to the
golden standard of runtime linear in the sparsity of the
input. To improve the runtime, one can apply the sparse
p-stable transform and achieve input sparsity runtime, but
that comes at the cost of an Ω(k6) factor loss in the approx-
imation ratio.
Using the above techniques, we develop an algorithm with
best of both worlds: k2 · poly(log n) approximation ratio
and the input sparsity running time at the same time. In par-
ticular, the main inefficiency of the algorithm (Song et al.,
2017) is the relaxation from ℓp norm to ℓ2 norm, which
incurs a further poly(k) approximation factor. In contrast,
the embedding based on exponential random variables em-
beds ℓp norm to ℓ2 norm directly, without further approxi-
mation loss. Our embedding also comes with its own pit-
falls - as we need to deal with mixed norms - thus re-
quiring a new analysis. See Theorem 23 for details.
2. Notations and preliminaries
In this paper, we denote R+ to be the set of nonnegative
reals. Define [n] = {1, 2,··· , n}. Given a matrix A ∈
Rn×d, ∀i ∈ [n], j ∈ [d], Ai and Aj denotes the ith row and
the jth column of A respectively. nnz(A) denotes the num-
ber of nonzero entries of A. The column space of A ∈ Rd is
{y ∃x ∈ Rd, y = Ax}. ∀p 6= 2, kAkp , (P Ai,jp)1/p,
i.e. entrywise p-norm. kAkF defines the Frobenius norm
of A, i.e. (P A2
i,j)1/2. A† denotes the Moore-Penrose
pseudoinverse of A. Given an invertible function f (·), let
f−1(·) be the inverse function of f (·).
If f (·) is not in-
vertible in (−∞, +∞) but it is invertible in [0, +∞), then
we denote f−1(·) to be the inverse function of f (·) in do-
main [0, +∞). inf and sup denote the infimum and supre-
mum respectively. f′(x), f′+(x), f′
−(x) denote the deriva-
tive, right derivative and left derivative of f (x), respec-
tively. Similarly, define f′′(x) for the second derivatives,
and we define f′′+(x) = limh→0+ (f′(x + h) − f′+(x))/h.
In the following, we give the definition of Orlicz norm.
0 ≤ x ≤ 1
x > 1
Definition 1 (Orlicz norm). For any nonzero monotone
nondecreasing convex function G : R+ → R+ with
G(0) = 0. Define Orlicz norm k · kG as: ∀n ∈ Z, n ≥
1, x ∈ Rn,kxkG = inf {α > 0 Pn
i=1 G(xi/α) ≤ 1} .
For any function G1(·) which is valid to define an Orlicz
norm, we can always "simplify/normalize" the function to
get another function G2 such that computingk·kG1 is equiv-
alent to computing k · kG2.
: R+ → R+ which
Fact 2. Given a function G1
can induce an Orlicz norm k · kG1 (Definition 1), de-
: R+ → R+ as the following:
fine function G2
G2(x) = (cid:26) G1(G−1
1 (1)x)
where s =
sx − (s − 1)
sup{(G2(y) − G2(x)) /(y − x) 0 ≤ x ≤ y ≤ 1} . Then
k · kG2 is a valid Orlicz norm. Furthermore, ∀n ∈ Z, n ≥
1, x ∈ Rn, we have kxkG1 = kxkG2/G−1
Thus, without loss of generality, in this paper we con-
sider the Orlicz norm induced by function G which satis-
fies G(1) = 1, and G(x) is a linear function for x > 1.
In addition, we also require that G(x) grows no faster
than quadratically in x. Thus, we define the property P
of a function G : R → R+ as the following: 1) G
is a nonzero monotone nondecreasing convex function in
[0,∞); 2) G(0) = 0, G(1) = 1,∀x ∈ R, G(x) = G(−x);
3) G(x) is a linear function for x > 1, i.e. ∃s > 0,∀x >
1, G(x) = sx + (1 − s); 4) ∃δG > 0 such that G is twice
differentiable on interval (0, δG). Furthermore, G′+(0) and
G′′+(0) exist, and either G′+(0) > 0 or G′′+(0) > 0; 5)
∃CG > 0,∀0 < x < y, G(y)/G(x) ≤ CG(y/x)2.
1 (1).
The condition 1 is required to define an Orlicz norm. The
conditions 2,3 are required because we can always do the
simplification/normalization (see Fact 2). The condition 4
is required for the smoothness of G. The condition 5 is
due to the subquadratic growth condition. Subquadratic
Subspace Embedding and Linear Regression with Orlicz Norm
Table 1. Some of M-estimators.
HUBER (cid:26) x2/2
c(x − c/2)
x ≤ c
x > c
ℓ1 − ℓ2
"FAIR"
2(p1 + x2/2 − 1)
c2 (x/c − log(1 + x/c))
i=1 G(xi)
with sketch size sub-polynomial in the dimension n, as
shown by (Braverman & Ostrovsky, 2010). For example,
growth condition is necessary for sketching Pn
if G(x) = xp for some p > 2, then k · kG is the same as
k · kp. It is necessary to take Ω(n1−2/p) space to sketch
ℓp norm in n-dimensional space. Condition 5 is also nec-
essary for 2-concave property, (Kwapien & Schuett, 1985;
Kwapie & Schtt, 1989) shows that k · kG can be embedded
into ℓ1 space if and only if G is 2-concave. Although (Schtt,
1995) gives an explicit embedding to ℓ1, it cannot be com-
puted efficiently.
There are many potential choices of G(·) which satisfies
property P, the following are some examples: 1) G(x) =
xp for some 1 ≤ p ≤ 2.
In this case k · kG is ex-
actly the ℓp norm k · kp; 2) G(x) can be a normalized M-
estimator function (see (Zhang, 1997)), i.e. define f (x)
to be one of the functions in Table 1. and let G(x) =
(cid:26) f (f−1(1)x)
−(1)x − (G′
G′
−(1) − 1)
x ≤ 1
x > 1
.
The following presents some useful properties of function
G with property P. See Appendix for details of proofs of
the following Lemmas.
Lemma 3. Given a function G(·) with property P, then
∀0 ≤ x ≤ 1, x2/CG ≤ G(x) ≤ x.
Lemma 4. Given a function G(·) with property P, then
∀x ∈ Rn,kxk2/√CG ≤ kxkG ≤ kxk1.
Lemma 5. Given a function G(·) with property P, then
∀0 < x < y, we have y/x ≤ G(y)/G(x).
Lemma 6. Given a function G(·) with property P, there
exist a constant αG > 0 which may depend on G, such that
∀0 ≤ a, b, if ab ≤ 1, then G(a)G(b) ≤ αGG(ab).
3. Subspace embedding for Orlicz norm using
exponential random variables
In this section, we develop the subspace embedding under
the Orlicz norms which are induced by functions G with
the property P. We first show how to embed the subspace
with k·kG norm into a subspace with ℓ2 norm, and then we
use dimensionality reduction techniques for the ℓ2 norm.
Overall, we will prove Theorem 16 stated at the end of this
section. Before discussing the details, we give formal defi-
nitions of subspace embedding.
Definition 7 (Subspace embedding for Orlicz norm).
Given a matrix A ∈ Rn×d, if S ∈ Rm×n satisfies ∀x ∈
Rd,kAxkG/α ≤ kSAxkv ≤ βkAxkG where α, β ≥
1,k · kv is a norm (can still be k · kG), then we say S
embeds the column space of A with Orlicz norm into the
column space of SA with v-norm. The distortion is αβ.
If the distortion and the v-norm are clear from the context,
we just say S is a subspace embedding matrix for A.
Definition 8 (Subspace embedding for ℓ2 norm). Given a
matrix A ∈ Rn×d, if S ∈ Rm×n satisfies ∀x ∈ Rd, (1 −
ε)kAxk2
2, then we say S is a
subspace embedding of column space of A.
2 ≤ (1 + ε)kAxk2
2 ≤ kSAxk2
There are many choices of ℓ2 subspace embedding matrix
A satisfying the above definition. Examples are: random
sign JL matrix (Achlioptas, 2003; Clarkson & Woodruff,
2009),
fast JL matrix (Ailon & Chazelle, 2009), and
sparse embedding matrices (Clarkson & Woodruff, 2013;
Meng & Mahoney, 2013; Nelson & Nguyen, 2013).
The main technical thrust is to embed k · kG into ℓ2 norm.
As the embedding matrix, we use S = ΠD−1 where Π
is one of the above ℓ2 embedding matrices and D is a di-
agonal matrix of which diagonal entries are i.i.d. random
variables draw from the distribution with CDF 1 − e−G(t).
Equivalently, each entry on the diagonal of D is G−1(u),
where u is an i.i.d. sample from the standard exponential
distribution, i.e. CDF is 1 − e−t. In Section 3.1, we will
prove that ∀x ∈ Rd, kD−1AxkGwill not be too large. In
Section 3.2, we will show that ∀x ∈ Rd, kD−1Axk∞ can-
kD−1Axk2 is a good estimator to kAxkG. In Section 3.3,
not be too small. Then due to Lemma 4, we know that
we show how to put all the ingredients together.
3.1. Dilation bound
f
We construct a randomized linear map f : Rn → Rn:
7−→ (x1/u1, x2/u2, ..., xn/un) where each
(x1, x2, ..., xn)
ui is drawn from a distribution with CDF 1 − e−G(t). No-
tice that for proving the dilation bound, we do not need to
assume ui are independent.
Theorem 9. Given x ∈ Rn, let k · kG be an Orlicz
norm induced by function G(·) which has property P, and
let f (x) = (x1/u1, x2/u2, ..., xn/un), where each ui is
drawn from a distribution with CDF 1 − e−G(t). Then
with probability at least 1 − δ − O(1/n19), kf (x)kG ≤
O(αGδ−1 log(n))kxkG, where αG is a constant may de-
pend on function G(·).
Proof sketch: By taking union bound, we have ∀i ∈
[n], ui ≥ G−1(1/n20) with high probability. Let α =
kxkG. For γ ≥ 1, we want to upper bound the prob-
ability that kf (x)kG ≥ γα. This is equivalent to upper
bound the probability that kf (x)/(γα)kG ≥ 1. Notice that
Pr(kf (x)/(γα)kG ≥ 1) = Pr(P G(xi/α·1/(γui)) ≥ 1).
By Markov inequality, it suffices to bound the expectation
ofP G(xi/α·1/(γui)) conditioned on ui are not too small.
By lemma 6,P G(xi/α· 1/(γui)) ≤ αG/γ ·P G(xi/α)·
1/G(ui). Because ui is not too small, the conditional ex-
pectation of 1/G(ui) is roughly O(log n). So the probabil-
Subspace Embedding and Linear Regression with Orlicz Norm
ity that kf (x)kG ≥ γα is bounded by O(αG log n/γ), set
γ = O(log n)αG/δ, we can complete the proof. See ap-
pendix for the details of the whole proof.
The final step is to use a well-conditioned basis; see details
in appendix.We then obtain the following theorem.
Theorem 10. Let G(·) be a function which has property P.
Given a matrix A ∈ Rn×m with rank d ≤ n, let D ∈ Rn×n
be a diagonal matrix of which each entry on the diagonal
is drawn from a distribution with CDF 1 − e−G(t). Then,
with probability at least 0.99, ∀x ∈ Rm,kD−1AxkG ≤
O(αGd2 log n)kAxkG, where αG ≥ 1 is a constant which
only depends on G(·).
3.2. Contraction bound
f
As in Section 3.1, we construct a randomized linear map f :
Rn → Rn: (x1, x2, ..., xn)
7−→ (x1/u1, x2/u2, ..., xn/un)
where each ui is an i.i.d. random variable drawn from a
distribution with CDF 1−e−G(t). Notice that the difference
from proving the dilation bound is that we need ui to be
independent here. We use the following theorem:
Theorem 11 (Lemma 3.1 of (Andoni et al., 2017)). Given
x ∈ Rn,
let k · kG be an Orlicz norm induced by
function G(·) which has property P, and let f (x) =
(x1/u1, x2/u2, ..., xn/un) , where each ui is an i.i.d ran-
dom variable drawn from a distribution with CDF 1 −
e−G(t). Then for α ≥ 1, with probability at least 1 − e−α,
kf (x)k∞ ≥ kxkG/α.
By combining the result with the net argument (see ap-
pendix), and Theorems 11, 10, we get the following:
Theorem 12. G(·) is a function with property P. Given
a matrix A ∈ Rn×m with rank d ≤ n, let D ∈ Rn×n
be a diagonal matrix of which each entry on the diagonal
is an i.i.d. random variable drawn from the distribution
with CDF 1 − e−G(t). Then, with probability at least 0.98,
∀x ∈ Rm, Ω(1/(α′Gd log n))kAxkG ≤ kD−1Axk∞,
where α′G ≥ 1 is a constant which only depends on G(·).
Proof sketch: Set ε = 1/poly(nd), we can build an ε-net
(see Appendix) N for the column space of A. By taking
the union bound over all the net points, we have ∀x ∈ N,
kD−1xk∞ is not too small. Due to Theorem 10, we have
∀x in the column space of A, kD−1xkG is not too large.
Now, for any unit vector y in the column space of A, we
can find the closest point x ∈ N, and kx − yk2 ≤ ε. Since
kD−1yk∞ ≥ kD−1xk∞ −kD−1(y− x)k∞, kD−1xk∞ is
not too small, and kD−1(y − x)k∞ is not too large, we can
get a lower bound for kD−1yk∞. See appendix for details.
3.3. Putting it all together
We now combine Theorem 12, Theorem 10, and Lemma 4,
to get the following theorem.
Theorem 13. Let G(·) be a function which has property P.
Given a matrix A ∈ Rn×m with rank d ≤ n, let D ∈ Rn×n
be a diagonal matrix of which each entry on the diagonal
is an i.i.d. random variable drawn from the distribution
with CDF 1 − e−G(t). Then, with probability at least 0.98,
∀x ∈ Rm, Ω(1/(α′Gd log n))kAxkG ≤ kD−1Axk2 ≤
O(α′′Gd2 log n)kAxkG, where α′′G, α′G ≥ 1 are two con-
stants which only depend on G(·).
The above theorem successfully embeds k·kG into ℓ2 space.
We now use ℓ2 subspace embedding to reduce the dimen-
sion. The following two theorems provide efficient ℓ2 sub-
space embeddings.
Theorem 14 ( (Clarkson & Woodruff, 2013)). Given ma-
trix A ∈ Rn×m with rank d. Let t = Θ(d2/ε2), S = ΦY ∈
Rt×n, where Y ∈ Rn×n is a diagonal matrix with each
diagonal entry independently uniformly chosen to be ±1,
Φ ∈ Rt×n is a binary matrix with Φh(i),i = 1,∀i ∈ [n],
and remaining entries 0. Here h : [n] → [t] is a random
hashing function such that for each i ∈ [n], h(i) is uni-
formly distributed in [t]. Then with probability at least 0.99,
∀x ∈ Rm, (1− ε)kAxk2
2. Fur-
thermore, SA can be computed in nnz(A) time.
Theorem 15 (See e.g. (Woodruff, 2014)). Given matrix
2 ≤ (1 + ε)kAxk2
2 ≤ kSAxk2
be a random matrix of i.i.d. standard Gaussian variables
A ∈ Rn×m with rank d. Let t = Θ(d/ε2), S ∈ Rt×n
scaled by 1/√t. Then with probability at least 0.99, ∀x ∈
Rm, (1 − ε)kAxk2
2 ≤ (1 + ε)kAxk2
2.
2 ≤ kSAxk2
We conclude the full theorem for our subspace embedding:
Theorem 16. Let G(·) be a function which has property
P. Given a matrix A ∈ Rn×d, d ≤ n, let D ∈ Rn×n
be a diagonal matrix of which each entry on the diagonal
is an i.i.d. random variable drawn from the distribution
with CDF 1 − e−G(t). Let Π1 ∈ Rt1×n be a sparse em-
bedding matrix (see Theorem 14) and let Π2 ∈ Rt2×t1 be
a random Gaussian matrix (see Theorem 15) where t1 =
Ω(d2), t2 = Ω(d). Then, with probability at least 0.9, ∀x ∈
Rd, Ω(1/(α′Gd log n))kAxkG ≤ kΠ2Π1D−1Axk2 ≤
O(α′′Gd2 log n)kAxkG, where α′′G, α′G ≥ 1 are two
constants which only depend on G(·).
Furthermore,
Π2Π1D−1A can be computed in nnz(A) + poly(d) time.
4. Applications
In this section, we discuss regression problem with Orlicz
norm error measure, and low rank approximation problem
with ℓp norm, which is a special case of the Orlicz norms.
4.1. Linear regression under Orlicz norm
We give the definition of regression problem as follow.
Definition 17. Function G(·) has property P. Given A ∈
Rn×d, b ∈ Rn, the goal is to solve the following minimiza-
tion problem minx∈Rd kAx − bkG.
Theorem 18. Let G(·) have property P. Given A ∈
Rn×d, b ∈ Rn, Algorithm 1 can output a solution x ∈ Rd
Subspace Embedding and Linear Regression with Orlicz Norm
Algorithm 1 Linear regression with Orlicz norm k · kG
1: Input: A ∈ Rn×d, b ∈ Rn.
2: Output: x.
3: Let t1 = Θ(d2), t2 = Θ(d).
4: Let Π1 ∈ Rt1×n be a random sparse embedding matrix,
Π2 ∈ Rt2×t1 be a random gaussian matrix, and D ∈ Rn×n
be a random diagonal matrix with each diagonal entry inde-
pendently drawn from distribution whose CDF is 1 − e−G(t).
(See Theorem 16.)
5: Compute x = (Π2Π1D−1A)†Π2Π1D−1b.
such that with probability at least 0.8, kAx − bkG ≤
O(βGd log2 n) minx∈Rd kAx − bkG, where βG is a con-
stant which may depend on G(·). In addition, the running
time of Algorithm 1 is nnz(A) + poly(d).
Proof sketch:
space embedding for column space of
Let S = Π2Π1D−1 be the sub-
Let
[A b].
x∗ = arg minx∈Rd kAx − bkG. Due to Theorem 16,
kAx − bkG is bounded by O(d log n)kS(Ax − b)k2 ≤
O(d log n)kS(Ax∗ − b)k2 ≤ O(d log n)kD−1(Ax∗ −
b)k2. Due to Theorem 9, kD−1(Ax∗ − b)k2 ≤
O(1)kD−1(Ax∗ − b)kG ≤ O(log n)kAx∗ − bkG.
4.2. Regression with combined loss function
In this section, we want to point out that our technique can
be used on solving regression problem with more general
cost function. Recall that the goal is to solve the minimiza-
tion problem minx∈Rd kAx− bkG. Now, we consider there
are multiple goals, and we want to minimize a linear com-
bination of the costs. Now we give the definition of regres-
sion problem with combined cost function.
Definition 19. Suppose function G1(·), G2(·), ..., Gk(·)
Given A1 ∈ Rn1×d, A2 ∈
satisfies property P.
Rn2×d, ..., Ak ∈ Rnk×d, b1 ∈ Rn1 , b2 ∈ Rn2 , ..., bk ∈
Rnk , the goal is to solve the following minimization prob-
lem minx∈Rd Pk
The idea of solving this problem is that we can embed every
term into l1 space, and then merge them into one term. By
the standard technique, there is a way to embed l2 space to
l1 space. We show the embedding as below. For the com-
pleteness, we put the proof of this lemma to the appendix.
i=1 kAix − bikGi .
Lemma 20. Let Q ∈ Rt×n be a random matrix with each
entry drawn uniformly from i.i.d. N (0, 1) Gaussian distri-
bution. Let B = (pπ/2/t) · Q. If t = Ω(ǫ−2n log(nǫ−1)),
then with probability at least 0.98, ∀x ∈ Rn,kBxk1 ∈
((1 − ǫ)kxk2, (1 + ǫ)kxk2).
Theorem 21. Let k > 0 be a constant, and
G1(·), G2(·), ..., Gk(·)
Given
A1 ∈ Rn1×d, A2 ∈ Rn2×d, ..., Ak ∈ Rnk×d, b1 ∈
Rnk , Algorithm 2
Rn1 , b2
∈
can output a solution x ∈ Rd
such that with
least 0.7, Pk
i=1 kAi x − bikGi ≤
O(β′Gd log2 n) minx∈Rd Pk
i=1 kAix − bikGi, where β′G is
∈
probability at
satisfy property P.
Rn2, ..., bk
Algorithm 2 Linear regression with combined loss functions
1: Input: A1 ∈ Rn1×d, A2 ∈ Rn2×d, ..., Ak ∈ Rnk ×d, b1 ∈
Rn1 , b2 ∈ Rn2 , ..., bk ∈ Rnk
2: Output: x.
3: Let t1 = Θ(d2), t2 = Θ(d), t3 = Θ(t2 log(t2)).
4: Let Π(1)
2 , · · · , Π(k)
1 ∈ Rt1×n1 , · · · Π(k)
1 ∈ Rt1×nk be k random sparse
embedding matrices, Π(1)
∈ Rt2×t1 be k ran-
dom Gaussian matrices, and D(1) ∈ Rn1×n1 , · · · , D(k) ∈
Rnk ×nk be k random diagonal matrices where each diag-
onal entry of D(i) is independently drawn from distribu-
tion whose CDF is 1 − e−Gi(t).
(See Theorem 16.) Let
Q(1), · · · , Q(k) ∈ Rt3×t2 be random matrices with each
entry drawn uniformly from i.i.d. N (0, 1) Gaussian dis-
2
j=1 nj , D ∈ RPk
tribution. ∀i ∈ [k], let B(i) = (pπ/2/t3) · Q(i) (see
Lemma 20.) Let B ∈ Rkt3×kt2 , Π2 ∈ Rkt2×kt1 , Π1 ∈
Rkt1×Pk
j=1 nj be four block diago-
nal matrices such that ∀i ∈ [k], the ith block of B, Π2, Π1, D
is B(i), Π(i)
5: Let A = [A⊤
BΠ2Π1D−1.
1 , D(i) respectively.
2 , ..., A⊤
k ]⊤, b = [b⊤
2 , Π(i)
1 , A⊤
k ]⊤ and S =
j=1 nj ×Pk
1 , b⊤
2 , ..., b⊤
6: Use classical method of solving l1 regression to get x =
arg minx∈Rd kS(Ax − b)k1.
[A b].
i=1 nnz(Ai) + poly(d).
a constant which may depend on G1(·), G2(·), ..., Gk(·).
the running time of Algorithm 2 is
In addition,
Pk
Proof Sketch: Let A = [A⊤1 , A⊤2 , ..., A⊤k ]⊤, b =
[b⊤1 , b⊤2 , ..., b⊤k ]⊤, and S = BΠ2Π1D−1 be the sub-
space embedding for column space of
Let
Si = B(i)Π(i)
b)k1 = Pk
arg minx∈Rd Pk
and Lemma 20, Pk
i=1 kAi x − bikGi
O(d log n)Pk
b)k1 ≤ O(d log n)kS(Ax∗ − b)k1 = Pk
bi)k1. Due to Theorem 16, Pk
O(log n)Pk
i=1 kAix∗ − bikGi .
2 Π(i)
1 (D(i))−1. Notice that ∀x,kS(Ax −
Let x∗ =
i=1 kSi(Aix − bi)k1.
i=1 kAix − bikGi . Due to Theorem 16
is bounded by
i=1 kSi(Ai x − bi)k1 = O(d log n)kS(Ax −
i=1 kSi(Aix∗ −
i=1 kSi(Aix∗ − bi)k1 ≤
One application of the above Theorem is to solve the
LASSO (Least Absolute Shrinkage Sector Operator) re-
gression. In LASSO regression problem, the goal is to min-
imize kAx − bk2
2 + λkxk1, where λ is a parameter of reg-
ularizer. It is easy to show that it is equivalent to minimize
kAx − bk2 + λ′kxk1 for some other parameter λ′. When
we look at kAx − bk2 + λ′kxk1, we can set A1 = A, b1 =
b, A2 = λ′I, b2 = 0, G1(·) ≡ x2, G2(·) ≡ x, then we are
able to apply Theorem 21 to give a good approximation.
The merit of our algorithm is that it is very simple, and can
be computed very fast.
4.3. ℓp norm low rank approximation using exponential
random variables
We discuss a special case of Orlicz norm k · kG, ℓp
norm, i.e. G(x) ≡ xp for p ∈ [1, 2]. When rank pa-
Subspace Embedding and Linear Regression with Orlicz Norm
rameter k is ω(log n + log d), by using exponential ran-
dom variables, we can significantly improve the approx-
imation ratio of input sparsity time algorithms shown
by (Song et al., 2017). The high level ideas combine the re-
sults of (Woodruff & Zhang, 2013; Song et al., 2017) and
the dilation bound in Section 3. We define the problem in
the following. See Appendix for the proof of Theorem 23.
Definition 22. Let p ∈ [1, 2]. Given A ∈ Rn×d, n ≥ d, k ∈
Z, 1 ≤ k ≤ min(n, d), the goal is to solve the following
minimization problem: minU∈Rn×k,V ∈Rk×d kU V − Akp
p.
Algorithm 3 ℓp norm low rank approximation using exponential
random variables.
1: Input: A ∈ Rn×d, k ∈ Z, min(n, d) ≥ k ≥ 1.
2: Output: U ∈ Rn×k, V ∈ Rk×d.
3: Let t1 = Θ(k2), t2 = Θ(k), t3 = Θ(k log k).
4: Let Π1, S1 ∈ Rt1×n be two random sparse embedding matri-
ces, Π2, S2 ∈ Rt2×t1 be two random gaussian matrices, and
D1, D2 ∈ Rn×n be two random diagonal matrices with each
diagonal entry independently drawn from distribution whose
CDF is 1 − e−tp
. (See Theorem 16.)
5: Let T2, R ∈ Rd×t3 be two random matrix, with i.i.d. entries
drawn from standard p-stable distribution.
6: Let S = S2S1D−1
7: Solve X, Y = arg minX∈Rt2×k ,Y ∈Rk×t3 kT1ARXY SAT2−
1 , T1 = Π2Π1D−1
2 .
T1AT2k2
F .
8: U = AR X, V = Y SA.
Theorem 23. Let 1 ≤ p ≤ 2. Given A ∈ Rn×d, n ≥
d, k ∈ Z, 1 ≤ k ≤ min(n, d), with probability at least 2/3,
U , V outputted by Algorithm 3 satisfies: k U V − Akp
p ≤
α minU∈Rn×k,V ∈Rk×d kU V − Akp
=
O(min((k log k)4−p log2p+2 n, (k log k)4−2p log4+p n)).
In addition,
the running time of Algorithm 3 is
nnz(A) + (n + d)poly(k).
p, where α
5. Experiments
Implementation setups can be seen in appendix.
5.1. Orlicz Norm Linear Regression
In this section, we show that our algorithm i) has reason-
able and predictable performance under different scenarios
and ii) is flexible, general and easy to use. We perform
3 sets of experiments. The first is to compare its perfor-
mance with the standard ℓ1 and ℓ2 regression under dif-
ferent noise assumptions and dimensions of the regression
problem; the second is to compare the performance of Or-
licz regression with different G under different noise as-
sumptions; the third is to experiment with Orlicz function
G that is different from standard ℓp and Huber function. We
evaluate the performance of our Orlicz norm linear regres-
sion algorithm on simulated data.
Comparison with ℓ1 and ℓ2 regression We would like
to see whether Orlicz norm linear regression leads to ex-
pected performance relative to ℓ1 and ℓ2 regression. We
choose our Orlicz norm k · kG to be induced by the normal-
ized Huber function where the Huber function is defined
Table 2. Comparisons of different regressions in different noise
and dimension settings; each entry is the error of ℓ1, ℓ2, Orlicz
norm regression. As expected, ℓ2/ℓ1 regression lead to best perfor-
mance under Sparse/Gaussian noise setting, and the performance
of Orlicz norm regression lies in between.
Gaussian
Sparse
Mixed
balance
overconstraint
211.2/194.5/197.3
25.3/30.7/30.0
37.9/37.8/37.5
25.3/20.0/24.9
2e-9/1.6/1.5
8.7/7.6/7.5
as f (x) = (cid:26) x2/2
δ · (x − δ/2) o.w.
x ≤ δ
. We chose the pa-
rameter δ to be 0.75. Intuitively, it is between ℓ1 and ℓ2
norm (see Figure 1). In all the simulations, we generate ma-
trix A ∈ Rn×d, ground truth x∗ ∈ Rd, and b to be Ax∗ plus
some particular noise. We evaluate the performance of each
algorithm by the ℓ2 distances between the output x and the
ground truth x∗. In terms of algorithm details, since n, d
are not too large in our simulation, we did not apply the ℓ2
subspace embedding to reduce the dimension; we only use
reciprocal exponential random diagonal embedding matrix
to embed k · kG to ℓ2 norm (see Theorem 13)1.
i) n =
We experiment with two n, d combinations,
200, d = 10 ii) n = 100, d = 75, and 3 noise setting with i)
Gaussian noise ii) sparse noise and iii) mixed noise (addi-
tion of i) and ii)), altogether 2× 3 = 6 setting. The detail of
data simulation can be seen in appendix. For each experi-
ment we repeat 50 times and compute the mean. The results
are shown in Table 2. Orlicz norm regression has better per-
formance than ℓ1 and ℓ2 when the noise is mixed. When the
noise is Gaussian or sparse, Orlicz norm regression works
better than ℓ1 and ℓ2 respectively. We did not experiment
with Huber loss regression, since if we rescale the data and
make it small/large in absolute values, the Huber regres-
sion will degenerate into respectively ℓ2/ℓ1 regression (see
Introduction). See appendix for results on approximation
ratio.
Choice of δ for G as a normalized Huber function We
compare the performance of Orlicz norm regression in-
duced by G as normalized Huber loss function with dif-
ferent δ under different noise assumptions. We fix n =
500, d = 30 and generate A and x as in the first set of ex-
periments (see appendix). The noise is a mixture of N (0, 5)
Gaussian noise and sparse noise on 1% entries with dif-
ferent scale of uniform noise from [−skAx∗k2, skAx∗k2],
where scale s is chosen from [0, 0.5, 1, 2]. Under each
noise assumptions with different scale s, we compare the
performance of Orlicz norm regression induced by G with
δ from [0.05, 0.1, 0.2, 0.4, 1, 2]. We repeat each experiment
50 times and report the mean of the ℓ2 distance between
output x and the ground truth x∗. The result is shown in
Figure 2. When the scale is 0/2, the noise is almost Gaus-
sian/sparse and we expect ℓ2/ℓ1 norm and thus large/small
1We use MATLAB's linprog to solve ℓ1 regression.
Subspace Embedding and Linear Regression with Orlicz Norm
Table 3. Orlicz regression with different choices of G, mean of
the ℓ2 distances between the output and the ground truth in 50
repetitions of experiments.
ℓ1
ℓ1.5
ℓ2
Gδ=0.25
Gδ=0.75
17.0
45.0
909.8
60.2
405.7
Gℓ1.5
14.7
δ to perform the best; anything scale lying in between these
extremes will have an optimal δ in between. We observe the
expected trend: as s increases, the performance is optimal
with smaller δ.
x ≤ δ
δ0.5 · (x − δ/3) o.w.
Beyond Huber function - A General Framework We
explore a variant Orlicz function G and evaluate it under
a particular setting; the evaluation criteria is the same as
the first set of the experiments. The G is of the same
form aforementioned, except that it now grows at the or-
der of x1.5 when x is small. We denote it by Gℓ1.5 ,
which is the normalization of function f , and f is de-
fined as: f (x) = (cid:26) x1.5/1.5
. We
generate a 500 × 30 matrix A and the ground truth vec-
tor x∗ in the same way as before, and then add N (0, 5)
Gaussian noises and 1 sparse outlier with scale s = 100.
We find that the modified Gℓ1.5 under this settings out-
performs ℓ1, ℓ2, ℓ1.5, Gδ=0.25, Gδ=0.75 regression by a sig-
nificant amount where Gδ=0.25, Gδ=0.75 are Orlicz norm
induced by regular normalized Huber function with δ =
0.25, 0.75 respectively. The results are shown in Table 3.
This experiment demonstrates that our algorithm is i) flexi-
ble enough to combine the advantage of norm functions, ii)
general for any function that satisfies the nice property, and
iii) easy to experiment with different settings, as long as we
can compute G and G−1.
5.2. ℓ1 low rank matrix approximation
In this section, we evaluate the performance of the ℓ1 low
rank matrix approximation algorithm. We mainly com-
pare the ℓ1 norm error of our algorithm with the error of
(Song et al., 2017) and standard PCA. Inputs are a matrix
A ∈ Rn×d and a rank parameter k; the goal is to output
a rank k matrix B such that kA − Bk1 is as small as pos-
sible. The details of implementations are in the appendix.
For each input, we run the algorithm 50 times and pick the
best solution.
Figure 2. Performance of Orlicz regression with G induced by dif-
ferent δ under different scale of sparse noise. The larger the sparse
noise, the smaller the δ that leads to the best performance, which
makes the norm closer to ℓ1
Datasets. We first run experiment on synthetic data:
we randomly choose two matrices U ∈ R2000×5, V ∈
R5×2000 with each entry drawn uniformly from (0, 1) Then
we randomly choose 100 entries of U V , and add random
outliers uniformly drawn from (−100, 100) on those en-
tries, thus we can get a matrix A ∈ R2000×2000.
In
our experiment, kAk1 is about 5.0 × 106. Then, we run
experiments on real datasets diabetes and glass in UCI
repository(Bache & Lichman, 2013). The data matrix of
diabetes has size 768 × 8, and the data matrix of glass has
size 214 × 9. For each data matrix, we randomly add out-
liers on 1% number of entries.
For each dataset, we evaluate the kA − Bk1. The result
for the experiment on synthetic data is shown in Table 4,
and the results for diabetes and glass are shown in Figure 3.
The running time of algorithm in (Song et al., 2017) on dia-
betes and on glass are 5.69 and 11.97 seconds respectively,
with ours being 3.18 and 3.74 seconds respectively. We
also find that our algorithm consistently outperforms the
other two alternatives (the y-coordinates are at log scale
with base 10).
106
105
104
1
2
3
Cauchy embedding only
Ours, Cauchy+Exponential
PCA
4
5
diabetes
6
105
104
103
102
1
Cauchy embedding only
Ours, Cauchy+Exponential
PCA
2
3
4
glass
5
6
7
8
Figure 3. ℓ1 norm error v.s. target rank.
Table 4. ℓ1 rank-5 approximation on the synthetic data.
Opt
PCA
(Song et al., 2017)
Ours
ℓ1 loss (×104 )
0.50
3.53
1.36
1.04
6. Conclusion and Future Work
We presented an efficient subspace embedding algorithm
for orlicz norm and demonstrated its usefulness in regres-
sion/low rank approximation problem on synthetic and real
datasets. Nevertheless, O(d log2 n) is still a large theo-
retical approximation factor, and hence it is worth i) in-
vestigating whether the theoretical approximation ratio can
be smaller if input are under some statistical distribution
ii) calculating the actual approximation ratio with ground
truth obtained by some slower but more accurate opti-
mization algorithm.
It is also worth examining whether
our exponential embedding sketching method preserves the
statistical properties of the regression error, since we as-
sumed a different noise distribution from Gaussian/double-
exponential as a starting point (Raskutti & Mahoney, 2014;
Lopes et al., 2018).
Subspace Embedding and Linear Regression with Orlicz Norm
Acknowledgements
Research supported in part by Simons Foundation
(#491119 to Alexandr Andoni), NSF (CCF-1617955, CCF-
1740833), and Google Research Award.
References
Achlioptas, Dimitris. Database-friendly random projec-
tions: Johnson-lindenstrauss with binary coins. Journal
of computer and System Sciences, 66(4):671–687, 2003.
Ailon, Nir and Chazelle, Bernard. The fast johnson–
lindenstrauss transform and approximate nearest neigh-
bors.
SIAM Journal on Computing, 39(1):302–322,
2009.
Andoni, Alexandr, Nikolov, Aleksandar, Razenshteyn, Ilya,
and Waingarten, Erik. Approximate near neighbors for
general symmetric norms. In Proceedings of the Forty-
ninth annual ACM symposium on Theory of computing,
2017.
Bache, Kevin and Lichman, Moshe. Uci machine learn-
ing repository. URL http://archive. ics. uci. edu/ml, 901,
2013.
Bhatia, Kush, Jain, Prateek, and Kar, Purushottam. Ro-
bust regression via hard thresholding.
In Cortes, C.,
Lawrence, N. D., Lee, D. D., Sugiyama, M., and Gar-
nett, R. (eds.), Advances in Neural Information Process-
ing Systems 28, pp. 721–729. Curran Associates, Inc.,
2015.
Braverman, Vladimir and Ostrovsky, Rafail. Zero-one fre-
quency laws. In Proceedings of the forty-second ACM
symposium on Theory of computing, pp. 281–290. ACM,
2010.
Cavazza, Jacopo and Murino, Vittorio. People counting by
huber loss regression.
Cavazza,
Jacopo and Murino, Vittorio.
gression with adaptive huber loss.
arXiv:1606.01568, 2016.
Active re-
arXiv preprint
Chierichetti, Flavio, Gollapudi, Sreenivas, Kumar, Ravi,
Lattanzi, Silvio, Panigrahy, Rina, and Woodruff, David P.
Algorithms for ℓp low rank approximation.
arXiv
preprint arXiv:1705.06730, 2017.
Clarkson, Kenneth L and Woodruff, David P. Numerical
linear algebra in the streaming model. In Proceedings
of the forty-first annual ACM symposium on Theory of
computing, pp. 205–214. ACM, 2009.
Clarkson, Kenneth L and Woodruff, David P. Low rank
approximation and regression in input sparsity time. In
Proceedings of the forty-fifth annual ACM symposium on
Theory of computing, pp. 81–90. ACM, 2013.
Clarkson, Kenneth L and Woodruff, David P. Sketching for
m-estimators: A unified approach to robust regression.
In Proceedings of the Twenty-Sixth Annual ACM-SIAM
Symposium on Discrete Algorithms, pp. 921–939. Soci-
ety for Industrial and Applied Mathematics, 2015.
Clarkson, Kenneth L, Drineas, Petros, Magdon-Ismail,
Malik, Mahoney, Michael W, Meng, Xiangrui, and
Woodruff, David P. The fast cauchy transform and faster
robust linear regression. In Proceedings of the Twenty-
Fourth Annual ACM-SIAM Symposium on Discrete Algo-
rithms, pp. 466–477. Society for Industrial and Applied
Mathematics, 2013.
Dasgupta, Anirban, Drineas, Petros, Harb, Boulos, Kumar,
Ravi, and Mahoney, Michael W. Sampling algorithms
and coresets for ℓp regression. SIAM Journal on Com-
puting, 38(5):2060–2078, 2009.
Dhillon, Paramveer, Lu, Yichao, Foster, Dean P, and Ungar,
Lyle. New subsampling algorithms for fast least squares
regression. In Burges, C. J. C., Bottou, L., Welling, M.,
Ghahramani, Z., and Weinberger, K. Q. (eds.), Advances
in Neural Information Processing Systems 26, pp. 360–
368. Curran Associates, Inc., 2013.
Drineas, Petros, Mahoney, Michael W, Muthukrishnan, S,
and Sarl´os, Tam´as. Faster least squares approximation.
Numerische Mathematik, 117(2):219–249, 2011.
Geppert, Leo N, Ickstadt, Katja, Munteanu, Alexander,
Quedenfeld, Jens, and Sohler, Christian. Random projec-
tions for bayesian regression. Statistics and Computing,
pp. 1–23, 2015.
Gorard, Stephen. Revisiting a 90-year-old debate: the ad-
vantages of the mean deviation. British Journal of Edu-
cational Studies, 53(4):417–430, 2005.
Huber, Peter J et al. Robust estimation of a location pa-
rameter. The Annals of Mathematical Statistics, 35(1):
73–101, 1964.
Jain, Prateek and Tewari, Ambuj. Alternating minimiza-
tion for regression problems with vector-valued outputs.
In Cortes, C., Lawrence, N. D., Lee, D. D., Sugiyama,
M., and Garnett, R. (eds.), Advances in Neural Informa-
tion Processing Systems 28, pp. 1126–1134. Curran As-
sociates, Inc., 2015.
Kwapie, Stanisaw and Schtt, Carsten. Some combinato-
rial and probabilistic inequalities and their application to
banach space theory ii. Studia Mathematica, 95(2):141–
154, 1989.
Subspace Embedding and Linear Regression with Orlicz Norm
Kwapien, Stanislaw and Schuett, Carsten. Some combina-
torial and probabilistic inequalities and their application
to banach space theory. Studia Mathematica, 82:91–106,
1985.
Liu, Han, Wang, Lie, and Zhao, Tuo. Multivariate re-
gression with calibration. In Ghahramani, Z., Welling,
M., Cortes, C., Lawrence, N. D., and Weinberger, K. Q.
(eds.), Advances in Neural Information Processing Sys-
tems 27, pp. 127–135. Curran Associates, Inc., 2014.
Lopes, Miles E, Wang,
Shusen,
and Mahoney,
Error estimation for randomized least-
arXiv preprint
Michael W.
squares algorithms via the bootstrap.
arXiv:1803.08021, 2018.
Mahoney, Michael W et al. Randomized algorithms for ma-
trices and data. Foundations and Trends R(cid:13) in Machine
Learning, 3(2):123–224, 2011.
Mangasarian, Olvi L and Musicant, David R. Robust linear
and support vector regression. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 22(9):950–955,
2000.
Meng, Xiangrui and Mahoney, Michael W. Low-distortion
subspace embeddings in input-sparsity time and applica-
tions to robust linear regression. In Proceedings of the
forty-fifth annual ACM symposium on Theory of comput-
ing, pp. 91–100. ACM, 2013.
Nelson, Jelani and Nguyen, Huy L. Osnap: Faster numer-
ical linear algebra algorithms via sparser subspace em-
beddings. In Foundations of Computer Science (FOCS),
2013 IEEE 54th Annual Symposium on, pp. 117–126.
IEEE, 2013.
Owen, Art B. A robust hybrid of lasso and ridge regression.
Contemporary Mathematics, 443:59–72, 2007.
Raskutti, Garvesh and Mahoney, Michael. A statistical
perspective on randomized sketching for ordinary least-
squares. arXiv preprint arXiv:1406.5986, 2014.
Rasmussen, Carl Edward. Gaussian processes for machine
learning. 2006.
Schtt, Carsten. On the embedding of 2-concave orlicz
spaces into l1. Studia Mathematica, 113(1):73–80, 1995.
Sohler, Christian and Woodruff, David P. Subspace embed-
dings for the ℓ1-norm with applications. In Proceedings
of the forty-third annual ACM symposium on Theory of
computing, pp. 755–764. ACM, 2011.
Song, Zhao, Woodruff, David P, and Zhong, Peilin.
Low rank approximation with entrywise ℓ1-norm er-
ror.
In Proceedings of the 49th Annual Symposium
on the Theory of Computing. ACM, arXiv preprint
arXiv:1611.00898, 2017.
Wang, Ruosong and Woodruff, David P. Tight bounds
for ℓp oblivious subspace embeddings. arXiv preprint
arXiv:1801.04414, 2018.
Wikipedia.
https://en.wikipedia.org/wiki/
Least_absolute_deviations.
Woodruff, David and Zhang, Qin. Subspace embeddings
and ℓp regression using exponential random variables. In
Conference on Learning Theory, pp. 546–567, 2013.
Woodruff, David P. Sketching as a tool for numerical lin-
ear algebra. Foundations and Trends in Theoretical Com-
puter Science, 10(1-2):1–157, 2014.
Zhang, Zhengyou. Parameter estimation techniques: A tu-
torial with application to conic fitting. Image and vision
Computing, 15(1):59–76, 1997.
Zhong, Kai, Jain, Prateek, and Dhillon, Inderjit S. Mixed
In Lee,
linear regression with multiple components.
D. D., Sugiyama, M., Luxburg, U. V., Guyon, I., and
Garnett, R. (eds.), Advances in Neural Information Pro-
cessing Systems 29, pp. 2190–2198. Curran Associates,
Inc., 2016.
A. Related Works
Existing literature studied the robust regression with re-
spect to Huber loss function (Mangasarian & Musicant,
2000; Owen, 2007).
Such regression can be applied
to solve many problems like the people counting prob-
lem (Cavazza & Murino). To speed up the regression
process, some dimensional reduction techniques can be
used to reduce the number of observations (Geppert et al.,
2015), also faster algorithms have been proposed to
address the robust regression with reasonable assump-
tion (Bhatia et al., 2015). Besides, different models of re-
gression were explored, such as Gaussian process regres-
sion (Rasmussen, 2006), active regression with adaptive
Huber loss (Cavazza & Murino, 2016).
Recent years, there are lots of randomized sketching and
embedding techniques developed for solving numerical
linear algebra problems. There is a long line of works,
e.g.
(Achlioptas, 2003; Clarkson & Woodruff, 2013;
Nelson & Nguyen, 2013) for ℓ2 subspace embedding, and
works, e.g. (Sohler & Woodruff, 2011; Meng & Mahoney,
2013; Woodruff & Zhang, 2013; Wang & Woodruff, 2018)
for ℓp subspace embedding. For more related works, we
refer readers to the book (Woodruff, 2014). Based on
sketching/embedding techniques, there is a line of works
studied ℓ2 and ℓp regressions, e.g. (Clarkson & Woodruff,
Subspace Embedding and Linear Regression with Orlicz Norm
2013; Drineas et al., 2011; Meng & Mahoney, 2013;
Nelson & Nguyen,
2013).
(Clarkson & Woodruff, 2015) studied linear regression
with M-estimator error measure. We refer to the survey
(Mahoney et al., 2011) for more details.
2013; Woodruff & Zhang,
Frobenius norm low rank matrix approximation problem is
also known as PCA problem. This problem is well studied.
The fastest algorithm is shown by (Clarkson & Woodruff,
2013). For the entrywise ℓp norm low rank approximation
problem, there is no known algorithm with theoretical guar-
antee until the work (Song et al., 2017). (Song et al., 2017)
works only for 1 ≤ p ≤ 2. Recently, (Chierichetti et al.,
2017) gives algorithms for all p ≥ 1. But either the run-
ning time is not in polynomial or the rank of the output is
not exact k.
B. Proof of Fact 2.
Proof. Notice that G1 is a nonzero nondecreasing con-
thus G−1
1 (1) exists, and G2 is a
vex function on R+,
nonzero nondecreasing function.
In addition because
y−x (G2(y) − G2(x)) 0 ≤ x ≤ y ≤ 1o , G2
s = supn 1
is also convex. Thus k · kG2 is Orlicz norm. Let x ∈ Rn.
Notice that if α > 0 satisfies Pn
i=1 G1(xi/α) ≤ 1, then
∀i ∈ [n], G1(xi/α) ≤ 1. It means that xi ≤ G−1
1 (1)α,
thus Pn
i=1 G1(xi/α) ≤
1. Similarly if α satisfies Pn
i=1 G2(xi/α) ≤ 1, then
1 (1)xi/α) = Pn
Pn
i=1 G1(G−1
i=1 G2(xi/α) ≤ 1.
Therefore, kxkG1 = kxkG2/G−1
1 (1).
1 (1)α)) = Pn
i=1 G2(xi/(G−1
C. Proof of Lemma 3.
Due to convexity of G and G(1) = 1, G(0) = 0, ∀x ∈
[0, 1], G(x) ≤ xG(1) + (1 − x)G(0) = x. Since x ≤ 1,
G(1)/G(x) ≤ CG(1/x)2, we have G(x) ≥ x2/CG.
D. Proof of Lemma 4.
With out loss of generality, we can assume ∀i ∈ [n], xi ≥
0. Let x ∈ Rn, α = kxkG. We have Pn
i=1 G(xi/α) = 1.
If xi/α ≤ 1, due to the convexity of G, G(xi/α) ≤
G(1) · xi/α + G(0) · (1 − xi/α) = G(1) · xi/α = xi/α.
If xi/α > 1, then G(xi/α) > 1 which contradicts to
Pn
i=1 G(xi/α) = 1. Thus, kxkG ≤ kxk1.
kx/αk2
i=1 CGG(xi/α) = CG.
2 = Pn
Then
i=1(xi/α)2 ≤ Pn
kxk2 ≤ pCGα.
E. Proof of Lemma 5.
Due to the convexity of G(·) and G(0) = 0, ∀0 < x < y,
we have G(x) ≤ G(y)x/y + G(0)(1 − x/y) = G(y)x/y.
Thus, y/x ≤ G(y)/G(x).
F. Proof of Lemma 6.
x−0
− c < c
G(ab) ≤ G(b)/G(ab) ≤ CG/a2 ≤ 4CG.
It is easy to see that ∀x > 0, G(x) 6= 0, since otherwise for
y > x, the condition G(y)/G(x) < CG(y/x)2 would be
violated. Let s = G′+(1). There are several cases.
If a ≥ 1 or b ≥ 1. Without loss of generality, assume
a ≥ 1.G(a)G(b)/G(ab) = (sa − (s − 1))G(b)/G(ab) ≤
saG(b)/G(ab). since ab ≥ b, G(ab)/G(b) ≥ a. Therefore,
G(a)G(b)/G(ab) ≤ saG(b)/G(ab) ≤ s.
If a, b ≤ 1, 0.5 ≤ a ≤ 1 or 0.5 ≤ b ≤ 1, we want to show
G(a)G(b)/G(ab) is still bounded. Without loss of general-
ity, assume that 0.5 ≤ a ≤ 1. Then G(a)G(b)/G(ab) =
G(a) G(b)
If a, b ≤ 0.5 and G′(0) > 0. Let G′(0) = c > 0.
Therefore, there is a constant δ1 which may depend on G
such that ∀x ∈ (0, δ1), G(x)−G(0)
2 . Therefore,
∀x ∈ (0, δ1], G(x) > c
2 x. Due to Lemma 5, ∀x > δ1,
G(x)/x > G(δ1)/δ1 > c/2. Therefore, ∀x, G(x) ≥ c
2 x.
Since b ≤ 0.5, ab ≤ a ≤ 1. Since G is convex, G(a) ≤
1−ab G(ab) + a−ab
1−a
1−ab . Therefore,
G(a) ≤ G(ab) + (1 − a)/(1 − ab) ≤ G(ab) + 2a. Simi-
larly, G(b) ≤ 2b+G(ab). Then we have G(a)G(b) ≤ (2b+
G(ab))(2a + G(ab)) ⇒ G(a)G(b)/G(ab) ≤ ab/G(ab) +
(2a + 2b) + G(ab) ≤ 2/c + 2 + 1 ≤ 2/c + 3.
If a, b ≤ 0.5, G′+(0) = 0, G′′+(0) = c2 > 0. Let
ǫ = c2/4. Since G is twice differentiable in (0, δG) and
G′+(0), G′′+(0) exist, by Taylor's Theorem, there is a con-
stant δ2 > 0 which may depend on G such that G(x) −
(G(0) + G′+(0)x + c2x2/2) ≤ ǫx2. Therefore, ∀x ∈
(0, δ2), G(x) ≥ c2x2/4, G(x) ≤ c2x2. Hence, ∀a, b ∈
(0, δ2], G(a)G(b)/G(ab) ≤ G(a)G(b)
b2 ≤
4
c2
2 = 4c2. Consider a or b > δ2. Without loss of
c2
generality, assume a > δ2. Similar to the previous argu-
ment, G(a)G(b)/G(ab) ≤ G(a) G(b)
G(ab) ≤ G(b)/G(ab) ≤
CGb2/(ab)2 ≤ CG/δ2
2. Thus G(a)G(b)/G(ab) is bounded
by CG/δ2
2 in this case.
1−ab G(ab) + a−ab
1−ab G(1) = 1−a
c2a2b2/4 ≤ 4
c2
G(a)
G(b)
a2
G. Proof of Theorem 9
Without loss of generality, we assume ∀i ∈ [n], xi ≥
0. Fix i ∈ [n], we have Pr(ui ≥ G−1(1/n20)) =
e−G(G−1(1/n20)) ≥ 1− 1/n20. Define E to be the event that
∀i ∈ [n], ui ≥ G−1(1/n20). By taking union bound over n
coordinates, E happens with probability at least 1 − 1/n19.
Subspace Embedding and Linear Regression with Orlicz Norm
Let α = kxkG. Then, for any γ ≥ 1, we have
H. Proof of Theorem 10
Pr(kf (x)kG ≥ γα)
= Pr(kf (x)kG ≥ γα E ) Pr(E )
+ Pr(kf (x)kG ≥ γα ¬E ) Pr(¬E )
≤ Pr(kf (x)kG ≥ γα E ) Pr(E ) + Pr(¬E )
= Pr(kf (x)/(γα)kG ≥ 1 E ) Pr(E ) + Pr(¬E )
≤ E n
G(cid:18)xi/α ·
Xi=1
E(cid:18)G(cid:18)xi/α ·
Xi=1
=
n
1
γui(cid:19) E! Pr(E ) + 1/n19
γui(cid:19) E(cid:19) Pr(E ) + 1/n19.
1
1
Let r = G−1(1/n20). For a fixed i ∈ [n],
E(cid:18)G(cid:18)xi/α ·
=Z ∞
G(cid:18) xi/α
G(xi/α)Z ∞
γui(cid:19) E(cid:19) Pr(E )
uγ (cid:19) e−G(u)G′(u)du
γ Z 1
G(cid:18) xi/α
u (cid:19) e−G(u)dG
e−G(u)dG +
G(xi/α) +
1
γ
1
γ
≤
≤
1
1
r
r
r
1
γ Z 1
G(cid:18) xi/α
αGG(xi/α)Z 1
1
γ
r
≤
1
γ
G(xi/α) +
≤O(log n)
αG
γ
G(xi/α),
u (cid:19) e−G(u)dG
1
G(u)
e−G(u)dG
where αG is a constant may depend on G(·). The
first inequality follows by G(xi/α · 1/(γu)) ≤ 1/γ ·
G(xi/α · 1/u) + (1 − 1/γ) · G(0) ≤ G(xi/α ·
1/u)/γ ≤ G(xi/α)/γ. The second inequality follows
by R ∞
r e−xdx ≤ 1. The third inequality follows by
Lemma 6. Since xi/α ≤ 1, then there is an αG such that
G(u)G(xi/α· 1/u) ≤ αGG(xi/α). The last inequality fol-
lows byR ∞
1/n20 e−x/xdx = O(log n).
Thus, we have
Similar to (Dasgupta et al., 2009), we can define a well con-
ditioned basis for Orlicz norm.
Definition 24 (Well conditioned basis for Orlicz norm).
Given a matrix A ∈ Rn×m with rank d, let U ∈ Rn×d be
a matrix which has the same column space of A. If U satis-
fies 1. ∀x ∈ Rd, kxk∞ ≤ βkU xkG, 2. Pd
i=1 kUikG ≤ α ,
then U is an (α, β, G)-well conditioned basis of A.
Fortunately, the such good basis exists for Orlicz norm.
Theorem 25 (See Connection to Auerbach basis in Section
3.1 of (Dasgupta et al., 2009)). Given a matrix A ∈ Rn×m
with rank d and norm k·kG, there exist a matrix U ∈ Rn×d
which is a (d, 1, G) well conditioned basis of A.
Proof of Theorem 10. Notice that D−1Ax is exactly the
same as f (Ax). There is a matrix U ∈ Rn×d which is
(d, 1, G)-well conditioned basis of A. Since ∀x ∈ Rm,
there is always a vector y ∈ Rd such that Ax = U y, we
only need to prove that with probability at least 0.99,
∀x ∈ Rd,kD−1U xkG ≤ O(αGd2 log n)kU xkG,
where D, αG are the same as stated in the Theorem. Ac-
cording to Theorem 9, if we look at a fixed i ∈ [d],
then with probability at least 1 − 0.01/d, kD−1UikG ≤
O(αGd log(n)). By taking union bound, with probability
at least 0.99, ∀i ∈ [d], kD−1UikG ≤ O(αGd log(n)). Now
we have, for any x ∈ Rd,
Xi=1
xikD−1UikG
kD−1U xkG ≤
d
≤ kxk∞
d
Xi=1
kD−1UikG
d
Xi=1
≤ O(αGd log(n))kxk∞
≤ O(αGd2 log(n))kU xkG.
kUikG
1
γui(cid:19) E(cid:19) Pr(E )
n
Xi=1
E(cid:18)G(cid:18)xi/α ·
Xi=1
αG
γ
n
≤O(log n)
G(xi/α) ≤ O(log n)
Then,
αG
γ
.
The first inequality follows by triangle inequality. The
third inequality follows by ∀i ∈ [d], kD−1UikG ≤
O(αGd log(n)). The forth inequality follows by (d, 1, G)-
well conditioned basis.
Pr(kf (x)kG ≥ γα) ≤ O(log n)
αG
γ
+ 1/n19.
I. Proof of Theorem 12
It is equivalent to
Pr(kf (x)kG ≤ γα) ≥ 1 − O(log n)
αG
γ − 1/n19.
Set γ = O(log n) αG
δ , we complete the proof.
Now, in the following, we present the concept of ε-net.
Definition 26 (ε-net for ℓ2 norm). Given A ∈ Rn×m with
rank d, let S be the ℓ2 unit ball in the column space of A,
i.e. S = {y kyk2 = 1,∃x ∈ Rm, y = Ax}. Let N ⊂ S,
if ∀x ∈ S,∃y ∈ N such that kx − yk2 ≤ ε, then we say N
is an ε-net for S.
Subspace Embedding and Linear Regression with Orlicz Norm
The following theorem gives an upper bound of the size of
ε-net.
Theorem 27 (Lemma 2.2 of (Woodruff, 2014)). Given A ∈
Rn×m with rank d, let S be the ℓ2 unit ball in the column
space of A. There exist an ε-net N for S, such that N ≤
(1 + 4/ε)d.
It suffices to prove ∀x ∈ Rm,kAxk2 = 1 we have
Ω(1/(α′Gd log n))kAxkG ≤ kD−1Axk∞. Let D ∈ Rn×n
be a diagonal matrix of which each entry on the diago-
random variable drawn from the distribu-
nal is an i.i.d.
tion with CDF 1 − e−G(t). Let α′G ≥ 1 be a sufficiently
large constant. Let S be the ℓ2 unit ball in the column
space of A. Let t1 = Θ(α′Gd log n), t2 = Θ(αGd2 log n),
where αG is the parameter stated in Theorem 10. Set
ε = O(1/(√nCGt1t2)). There exist an ε-net N for S,
and
N = eO(d(log n+log(CGα′
GαG))).
By taking union bound over the net points, according to
Theorem 11, with probability at least 0.99,
∀x ∈ N,kD−1xk∞ ≥ Ω(1/(α′Gd log n))kxkG.
(2)
Also due to Theorem 10, with probability at least 0.99,
J. Proof of Theorem 13
Due to Theorem 12 and Theorem 10, with probability
at least 0.98, ∀x ∈ Rm, Ω(1/(α′Gd log n))kAxkG ≤
kD−1Axk∞ ≤ kD−1Axk2. And kD−1Axk2 ≤
√CGkD−1AxkG ≤ O(√CGα′Gd2 log n)kAxkG.
K. Proof of Theorem 16
Due to Theorem 14 and Theorem 15, with probability at
least 0.95, ∀x, kΠ2Π1D−1Axk2 is a constant approxima-
tion to kΠ1D−1Axk2 and kΠ1D−1Axk2 is a constant ap-
proximation to kD−1Axk2. Combining with Theorem 13,
we complete the proof.
L. Proof of Theorem 18
Let x∗ = arg minx∈Rd kAx−bkG. Due to Theorem 9, with
probability at least 0.99,
kD−1(Ax∗ − b)kG ≤ O(αG log n)kAx∗ − bkG.
(4)
Now let A′ = [A b] ∈ Rn×(d+1). Due to Theorem 16, with
probability at least 0.9, we have
∀x ∈ Rd+1, Ω(1/(α′Gd log n))kA′xkG ≤ kΠ2Π1D−1A′xk2.
(5)
∀x ∈ S,kD−1xkG ≤ O(αGd2 log n)kxkG.
(3)
Then,
By taking union bound, with probability at least 0.98, the
above two events will happen. Then, in this case, consider
a y ∈ S, let x ∈ N such that kx − yk2 ≤ ε, let z = x − y,
we have
t1 kzkG − t2pCGkzkG
1
t2
kD−1yk∞ ≥ kD−1xk∞ − kD−1zk∞
t1kxkG − t2pCGkzkG
≥
1
t1kykG −
≥
1
t1kykG − 2t2pCGkzkG
≥
1
t1kykG − O(
≥
≥ Ω(1/t1)kykG
= Ω(1/(α′Gd log n))kykG.
√CGt1
2
)
The first inequality follows by triangle inequality. The
second inequality follows by Equation 2, Equation 3, and
Lemma 4, i.e. kD−1zk∞ ≤ kD−1zk2 ≤ √CGkD−1zkG.
The third inequality follows by triangle inequality. The
forth inequality follows by t1, CG ≥ 1. The fifth inequality
follows by Lemma 4: kzkG ≤ kzk1 ≤ √nkzk2 = √nε =
O(1/(CGt1)). The sixth inequality follows by Lemma 4:
kykG ≥ 1√CGkyk2 = 1/√CG.
kAx − bkG ≤ O(α′Gd log n)kΠ2Π1D−1(Ax − b)k2
≤ O(α′Gd log n)kΠ2Π1D−1(Ax∗ − b)k2
≤ O(α′Gd log n)kD−1(Ax∗ − b)k2
≤ O(α′GpCGd log n)kD−1(Ax∗ − b)kG
≤ O(αGα′GpCGd log2 n)kAx∗ − bkG.
follows
by
x
follows
Equation
inequality
inequality
first
second
5.
The
=
The
(Π2Π1D−1A)†Π2Π1D−1b, which is the optimal so-
lution for minx∈Rd kΠ2Π1D−1(Ax − b)k2. The third
inequality follows by Theorem 14 and Theorem 15. The
forth inequality follows by Lemma 4. The last inequality
follows by Equation 4. Let βG = α′G√CG, we complete
by
the proof of the correctness of Algorithm 1.
For the running time, according to Theorem 16, comput-
ing Π2Π1D−1A and Π2Π1D−1b needs nnz(A) + poly(d)
time. Since Π2Π1D−1A has size poly(d), computing
x = (Π2Π1D−1A)†Π2Π1D−1b needs poly(d) running
time. The total running time is nnz(A) + poly(d).
M. proof of Lemma 20
Before we prove the Lemma, we need following tools.
Subspace Embedding and Linear Regression with Orlicz Norm
k
k
π
Xi=1
, (1 + ε)r 2
zi ∈ (1 − ε)r 2
Lemma 28 (Concentration bound for sum of half normal
random variables). For any k i.i.d. random Gaussian vari-
ables z1, z2,··· , zk, we have that
π!! ≥ 1 − e−Ω(kε2).
Pr 1
Lemma 29. Let G ∈ Rk×m be a random matrix with each
entry drawn uniformly from i.i.d. N (0, 1) Gaussian distri-
bution. With probability at least 0.99, G2 ≤ 10√km.
Proof. Since E(cid:0)kGk2
F = km(cid:1) , we have that Pr(kGk2
F ≥
100km) ≤ 0.01. Thus, with probability at least 0.99, we
have kGk2 ≤ kGkF ≤ 10√km.
Now, let us prove the lemma.
Proof of Lemma 20. Without loss of generality, we only
need to prove ∀x ∈ Rn with kxk2 = 1, we have kBxk1 ∈
(1 − ε, 1 + ε). Let set S = {v v ∈ Rn,kvk2 = 1}.
Due to Theorem 27, we can find a set G ⊂ S which
satisfies that ∀u ∈ S there exists v ∈ G such that
ku − vk2 ≤ (ε/(1000n))10 and G ≤ (4000n/ε)20n.
Let k ≥ cε−2n ln(n/ε) where c is a sufficiently large con-
stant. By Lemma 28, we have that for a fixed v ∈ G,
with probability at least 1 − e−1000n ln(4000n/ε),kBvk1 ∈
(1 − ε, 1 + ε). By taking union bound over all the points
in G, we have Pr (∀v ∈ G,kBvk1 ∈ (1 − ε, 1 + ε)) ≥
1 − e−980n ln(4000n/ε) ≥ 0.99. Now, consider ∀x ∈ Rn
with kxk2 = 1, i.e. x ∈ S, we can find v ∈ G such that
kv − xk2 ≤ (ε/(1000n))10, and let u = v − x. Then, con-
ditioned on
kBk2 ≤ 10√tn ·pπ/2/t,
we have
kBxk1
∈(kBvk1 − kBuk1,kBvk1 + kBuk1)
⊆(1 − (ε + √tkBk2kuk2), 1 + (ε + √tkBk2kuk2))
⊆(1 − 2ε, 1 + 2ε)
where the first relation follows by triangle inequality, the
second relation follows by
kBuk1 ≤ √tkBuk2 ≤ √tkBk2kuk2,
and the last relation follows by
kuk2 ≤ (ε/(1000n))10,kBk2 ≤ 10√tn ·pπ/2/t.
According to Lemma 29, we know that with probability at
least 0.99, we have
kBk2 ≤ 10√tn ·pπ/2/t.
By taking union bound, we have with probability at least
0.98,
∀x ∈ S,kBxk1 ∈ (1 − 2ε, 1 + 2ε).
By adjusting the ε, we complete the proof.
N. Proof of Theorem 21
Without loss of generality, we assume constant k ≤ 2.
Otherwise, we can always adjust constants in all the re-
lated theorems and lemmas to make larger k work. Let
i=1 kAix−bikGi. By Theorem 9 and
taking union bound, we have that with probability at least
0.98,
x∗ = arg minx∈Rd Pk
∀i ∈ {1, 2, · · · , k},
k(D(i))−1(Aix∗ − bi)kGi ≤ O(αGi log n)kAix∗ − bikGi .
(6)
Now let A′i = [Ai bi] ∈ Rni×(d+1). Due to Theorem 16
and union bound, with probability at least 0.8, we have
∀x ∈ Rd+1,
Ω(1/(α′
Gi d log ni))kA′
ixkGi ≤ kΠ(i)
2 Π(i)
1 (D(i))−1A′
ixk2.
(7)
Then,
k
k
Xi=1
Xi=1
Xi=1
k
≤
≤
kAi x − bikGi
O(α′
Gi d log n)kΠ(i)
2 Π(i)
1 (D(i))−1(Ai x − bi)k2
O(α′
Gi d log n)kB(i)Π(i)
2 Π(i)
1 (D(i))−1(Ai x − bi)k1
≤O(max
i∈[k]
Gi d log n)kBΠ2Π1D−1(Ax − b)k1
α′
≤O(max
i∈[k]
Gi d log n)kBΠ2Π1D−1(Ax∗ − b)k1
α′
k
kB(i)Π(i)
2 Π(i)
1 (D(i))−1(Aix∗ − bi)k1
kΠ(i)
2 Π(i)
1 (D(i))−1(Aix∗ − bi)k2
k(D(i))−1(Aix∗ − bi)k2
≤O(max
i∈[k]
≤O(max
i∈[k]
≤O(max
i∈[k]
≤O((max
k
α′
α′
Gi d log n)
Gi d log n)
Xi=1
Xi=1
Xi=1
i∈[k]pCGi )(max
Gi d log n)
i∈[k]
α′
k
α′
Gi )d log n)
k(D(i))−1(Aix∗ − bi)kGi
k
Xi=1
≤O((max
i∈[k]
αGi )(max
i∈[k]pCGi )(max
i∈[k]
Gi )d log2 n)
α′
kAix∗ − bikGi .
k
Xi=1
The first
inequality follows by Equation 7.
The
second inequality follows by Lemma 20.
The forth
inequality follows by x is the optimal solution for
Subspace Embedding and Linear Regression with Orlicz Norm
minx∈Rd kBΠ2Π1D−1(Ax − b)k1. The sixth inequal-
ity follows by Lemma 20.
ity follows by Theorem 14 and Theorem 15.
eighth inequality follows by Lemma 4.
inequality follows by Equation 6.
The seventh inequal-
The
The last
Let β′G =
(maxi∈[k] αGi)(maxi∈[k]pCGi )(maxi∈[k] α′Gi), we com-
plete the proof of the correctness of Algorithm 2.
ing Π2Π1D−1A and Π2Π1D−1b needs Pk
For the running time, according to Theorem 16, comput-
i=1 nnz(Ai) +
poly(d) time. Due to Lemma 20,
the size of B is
poly(d). To compute BΠ2Π1D−1A and BΠ2Π1D−1b,
we need additional poly(d) time. Since BΠ2Π1D−1A
has size poly(d), computing the optimal solution of
minx∈Rd kBΠ2Π1D−1(Ax−b)k1 by using linear program-
ming needs poly(d) running time. The total running time
is Pk
i=1 nnz(Ai) + poly(d).
O. Proof of Theorem 23
Before we prove the Theorem, we need to show following
Lemmas.
Lemma 30 ((Song et al., 2017)). Let A ∈ Rn×d, R ∈
Rd×t3, k be the same as in the Algorithm 3, then with prob-
ability at least 0.9,
min
X∈Rt3×k ,Y ∈Rk×d
kARXY − Akp
p
≤O((k log k)1−p/2 log n)
min
U ∈Rn×k,V ∈Rk×d
kU V − Akp
p.
Lemma 31 ((Woodruff & Zhang, 2013)). Let 1 ≤ p ≤ 2.
Given a matrix A ∈ Rn×d, d ≤ n, let D ∈ Rn×n be
a diagonal matrix of which each entry on the diagonal is
an i.i.d. random variable drawn from the distribution with
CDF 1 − e−tp
. Let Π1 ∈ Rt1×n be a sparse embedding
matrix (see Theorem 18) and let Π2 ∈ Rt2×t1 be a random
Gaussian matrix (see Theorem 19) where t1 = Ω(d2), t2 =
Ω(d). Then, with probability at least 0.9,
∀x ∈ Rd,
Ω(1/ min{(d log d)1/p, (d log d log n)1/p−1/2})kAxkp
≤kΠ2Π1D−1Axk2.
Lemma 32. Let A ∈ Rn×d, S ∈ Rt2×n, R ∈ Rd×t3, k
be the same as in the Algorithm 3, then with probability at
least 0.9,
min
X∈Rt2×k ,Y ∈Rk×t2
kARXY SA − Akp
p
≤β
min
U ∈Rn×k,V ∈Rk×d
kU V − Akp
p,
Let U∗ = ARX∗, V = (SU∗)†SA.
min{k log k, (k log k log n)1−p/2}. We have
Let γ =
kU ∗ V − Akp
p
≤2p−1kU ∗( V − V ∗)kp
p + 2p−1kU ∗V ∗ − Akp
p
d
kSU ∗( V − V ∗)ikp
2 + 2p−1kU ∗V ∗ − Akp
p
(kS(U ∗ V − A)ik2 + kS(U ∗V ∗ − A)ik2)p
≤O(γ)
≤O(γ)
d
Xi=1
Xi=1
d
d
Xi=1
Xi=1
+ 2p−1kU ∗V ∗ − Akp
p
≤O(γ)
≤O(γ)
(2kS(U ∗V ∗ − A)ik2)p + 2p−1kU ∗V ∗ − Akp
p
kD−1
1 (U ∗V ∗ − A)ikp
2 + 2p−1kU ∗V ∗ − Akp
p
1 (U ∗V ∗ − A)kp
≤O(γ)kD−1
p + 2p−1kU ∗V ∗ − Akp
≤O(γ) logp(nd)kU ∗V ∗ − Akp
p
=O(γ logp(n))kU ∗V ∗ − Akp
p.
p + 2p−1kU ∗V ∗ − Akp
p
The first inequality follows by convexity of xp. The sec-
ond inequality follows by Lemma 31. The third inequality
follows by triangle inequality. The forth inequality follows
by V = (SU∗)†SA. The fifth inequality follows by The-
orem 14 and Theorem 15. The sixth inequality follows by
p ≤ 2. The seventh inequality follows by Theorem 9.
Due to Lemma 30, we have
kU ∗V ∗ − Akp
p
≤O((k log k)1−p/2 log n)
min
U ∈Rn×k,V ∈Rk×d
kU V − Akp
p.
Thus, we have
kARXY SA − Akp
p ≤ kU ∗ V − Akp
p
min
X,Y
≤O(min((k log k)2−p/2 logp+1 n, (k log k)2−p log2+p/2 n))
· min
U,V
kU V − Akp
p.
Lemma 33 ((Song et al., 2017)). Let A ∈ Rn×d, S ∈
Rt2×n, R ∈ Rd×t3, k, T2 ∈ Rd×t3 be the same as in the
Algorithm 3, then with probability at least 0.9, if for α ≥ 1,
X, Y satisfy
kAR X Y SAT2 − AT2kp
X,Y kARXY SAT2 − AT2kp
p,
p ≤ α min
where
then
β = O(min((k log k)2−p/2 logp+1 n, (k log k)2−p log2+p/2 n)).
Proof. Let
X ∗, V ∗ = arg
min
X∈Rt3×k ,V ∈Rk×d
kARXV − Akp
p.
p ≤ αO(log n) min
kAR X Y SA − Akp
Lemma 34. Let A ∈ Rn×d, S ∈ Rt2×n, R ∈
Rd×t2, k, T1 ∈ Rt2×n, T2 ∈ Rd×t3 be the same as in the
X,Y kARXY SA − Akp
p.
Subspace Embedding and Linear Regression with Orlicz Norm
Algorithm 3, then with probability at least 0.9, if for α ≥ 1
t3
Xi=1
kT1(AR X Y SAT2 − AT2)ikp
2
≤ α min
X,Y
t3
Xi=1
kT1(ARXY SAT2 − AT2)ikp
2,
then
kAR X Y SAT2 − AT2kp
where β = O(min(k log k logp n, (k log k)1−p/2 log1+p/2 n)).
X,Y kARXY SAT2 − AT2kp
p,
p ≤ αβ min
inequality follows by p ≤ 2. The sixth inequality follows
by the property of X, Y . The seventh inequality follows by
Theorem 14 and Theorem 15. The eighth inequality fol-
lows by p ≤ 2. Then the ninth inequality follows by Theo-
rem 9.
Now let us prove Theorem:
Proof. Notice that
X, Y = arg
min
X∈Rt2×k ,Y ∈Rk×t3
kT1ARXY SAT2 − T1AT2k2
F ,
we have
Proof. Let
X∗, Y ∗ = arg min
X,Y
t3
Xi=1
kT1(ARXY SAT2 − AT2)ikp
2.
≤O((k log k)1/p−1/2)(min
X,Y
kT1(ARXY SAT2 − AT2)ikp
2)
1
p .
t3
Xi=1
kT1(AR X Y SAT2 − AT2)ikp
2)1/p
(
t3
Xi=1
It means
(
t3
Xi=1
kT1(AR X Y SAT2 − AT2)ikp
2)
≤O((k log k)1−p/2)(min
X,Y
kT1(ARXY SAT2 − AT2)ikp
2).
t3
Xi=1
According to Lemma 34, we have
kAR X Y SAT2 − AT2kp
p ≤ β1 min
X,Y kARXY SAT2 − AT2kp
p,
where
β1 = O(min((k log k)2−p/2 logp n, (k log k)2−p log1+p/2 n)).
Due to Lemma 33, we have
kAR X Y SA − Akp
p ≤ O(β1 log n) min
Then, according to Lemma 32, we have
X,Y kARXY SA − Akp
p.
kAR X Y SA − Akp
p ≤ β2
min
U∈Rn×k,V ∈Rk×d kU V − Akp
p,
where
β2 = O(min((k log k)4−p log2p+2 n, (k log k)4−2p log4+p n)).
For the running time: SA, T1A can be computed in nnz(A)
time. Thus, total running time is nnz(A) + (n + d)poly(k).
Let L = AR, N = SAT2, M = AT2. Let γ =
min{k log k, (k log k log n)1−p/2}. Let H = X Y and let
H∗ = X∗Y ∗. We have
kL HN − M kp
p
≤2p−1kL HN − LH ∗Nkp
p + 2p−1kLH ∗N − M kp
p
t3
≤O(γ)
≤O(γ)
kT1(L HN − LH ∗N )ikp
2 + 2p−1kLH ∗N − M kp
p
(kT1(L HN − M )ik2 + kT1(LH ∗N − M )ik2)p
≤O(γ)
(kT1(L HN − M )ikp
2 + kD−1
2 (LH ∗N − M )ikp
2)
≤O(γ)(
kT1(L HN − M )ikp
2 + kD−1
2 (LH ∗N − M )kp
p)
t3
Xi=1
Xi=1
t3
Xi=1
t3
Xi=1
t3
Xi=1
t3
Xi=1
+ 2p−1kLH ∗N − M kp
p
+ 2p−1kLH ∗N − M kp
p
+ 2p−1kLH ∗N − M kp
p
+ 2p−1kLH ∗N − M kp
p
+ 2p−1kLH ∗N − M kp
p
≤O(γ)(α
kT1(LH ∗N − M )ikp
2 + kD−1
2 (LH ∗N − M )kp
p)
≤O(γ)(α
kD−1
2 (LH ∗N − M )ikp
2 + kD−1
3 (LH ∗N − M )kp
p)
≤O(γ)αkD−1
2 (LH ∗N − M )kp
p + 2p−1kLH ∗N − M kp
p
≤O(γ logp(n))αkLH ∗N − M kp
p.
The first inequality follows by convexity of xp. The sec-
ond inequality follows by Lemma 31. The third inequality
follows by triangle inequality. The forth inequality follows
by convexity of xp, Theorem 14 and Theorem 15. The fifth
P. Implementation Setups
We implement all the algorithms in MATLAB. We ran ex-
periments on a machine with 16G main memory and Intel
Core [email protected] CPU. The operating system
is Ubuntu 14.04.5 LTS. All the experiments were in single
threaded mode.
Subspace Embedding and Linear Regression with Orlicz Norm
Q. Data Simulation for Comparison with ℓ1
S. Implementation Detail for Low Rank
Approximation
• For our algorithm, set t1 = 4k, t2 = 8t1, set S ∈
Rt1×n, T1 ∈ Rt2×n to be two random cauchy matri-
ces, and set R ∈ Rd×t1, T2 ∈ Rd×t2 to be two em-
bedding matrices with exponential random variables
(see Theorem 16.) We solve the minimization prob-
F , and set
lem minX,Y kT1ARXY SAT2 − T1AT2k2
B = ARXY SA.
• For algorithm in (Song et al., 2017), we set t1 =
4k, t2 = 8t1. We set S ∈ Rt1×n, T1 ∈
Rt2×n, R ∈ Rd×t1, T2 ∈ Rd×t2 to be four random
cauchy matrices. We solve the minimization prob-
lem minX,Y kT1ARXY SAT2 − T1AT2k2
F , and set
B = ARXY SA.
• For PCA, we project A onto the space spanned by top
k singular vectors to get B.
and ℓ2 Regression
We generate a matrix A ∈ Rn×d, x∗ ∈ Rd as following:
set each entry of the first d + 5 rows of A as i.i.d. standard
random Gaussian variable, each entry of x∗ as i.i.d. stan-
dard random Gaussian variable. For n ≥ i ≥ d + 6, we
uniformly choose p ∈ [d + 5], and set Ai = Ap, bi = bp.
We perform experiments under 3 different noise assump-
tions and 2 dimension combinations of N, d and in total
3 × 2 = 6 experiments. The 3 different noise assumptions
are, respectively i) N (0, 50) Gaussian noise with on all the
entries of Ax∗; ii) sparse noise, where we randomly pick
3% number of entries of Ax∗, and add uniform random
noise from [−kAx∗k2,kAx∗k2] on each entry to get b; iii)
mixed noise, which is N (0, 5) Gaussian noise plus sparse
noise. The 2 different dimension combinations are i) bal-
ance, where n = 100 ≈ d = 75; ii) overconstraint, where
n = 200 ≫ d = 10.
R. Experiments on Approximation Ratio
Here is a documentation of our preliminary experiments
on calculating the actual approximation ratio for the experi-
ment settings mentioned in Section 5.1, Comparison with
ℓ1 and ℓ2 regression. The approximation ratio of interest
is calculated as follows: kAx′−bkG
, where x′ is the output
kAx∗−bkG
of our novel embedding based algorithm and x∗ is the opti-
mal solution. Since k · kG is convex, we can formulate this
problem as a convex optimization problem and use a vanilla
gradient descent algorithm to calculate the optimal solution.
We heuristically stop our gradient descent algorithm when
the one step brings less than 10−7 improvement on the loss
function and set the learning rate to be 0.001. Admittedly,
we have not yet thoroughly and rigidly examined the con-
vergence of the vanilla gradient descent algorithm (a direc-
tion of future work), and hence such calculation of approx-
imation ratio is only a preliminary attempt.
Under the mixed noise setting, we varied different scale
s of the uniform noise to be 0, 1, 2, 3 and delta to be
0.1, 0.25, 0.5, 0.75. With n = 200, d = 10, for each of
these 4 * 4 = 16 settings, we run the algorithm repeat-
edly for 50 times, and the worst approximation ratio is 1.06
among these 800 runs. Experimentally, it is far below the
theoretical guarantee d log2(n) ≈ 584 ≫ 1.06, and the ap-
proximation ratio is robust among different noise settings.
For n = 100, d = 75, due to time limit, we only run each
of the 16 settings for 5 times, and the worst approximation
ratio is 1.31.
|
1504.04650 | 2 | 1504 | 2015-11-09T16:37:48 | A Faster FPTAS for the Unbounded Knapsack Problem | [
"cs.DS"
] | The Unbounded Knapsack Problem (UKP) is a well-known variant of the famous 0-1 Knapsack Problem (0-1 KP). In contrast to 0-1 KP, an arbitrary number of copies of every item can be taken in UKP. Since UKP is NP-hard, fully polynomial time approximation schemes (FPTAS) are of great interest. Such algorithms find a solution arbitrarily close to the optimum $\mathrm{OPT}(I)$, i.e. of value at least $(1-\varepsilon) \mathrm{OPT}(I)$ for $\varepsilon > 0$, and have a running time polynomial in the input length and $\frac{1}{\varepsilon}$. For over thirty years, the best FPTAS was due to Lawler with a running time in $O(n + \frac{1}{\varepsilon^3})$ and a space complexity in $O(n + \frac{1}{\varepsilon^2})$, where $n$ is the number of knapsack items. We present an improved FPTAS with a running time in $O(n + \frac{1}{\varepsilon^2} \log^3 \frac{1}{\varepsilon})$ and a space bound in $O(n + \frac{1}{\varepsilon} \log^2 \frac{1}{\varepsilon})$. This directly improves the running time of the fastest known approximation schemes for Bin Packing and Strip Packing, which have to approximately solve UKP instances as subproblems. | cs.DS | cs | A Faster FPTAS for the Unbounded Knapsack Problem∗
Department of Computer Science, Kiel University, 24098 Kiel, Germany
Klaus Jansen
Stefan E. J. Kraft
{kj,stkr}@informatik.uni-kiel.de
5
1
0
2
v
o
N
9
]
S
D
.
s
c
[
2
v
0
5
6
4
0
.
4
0
5
1
:
v
i
X
r
a
Abstract
The Unbounded Knapsack Problem (UKP) is a well-known variant of the famous 0-1 Knapsack
Problem (0-1 KP). In contrast to 0-1 KP, an arbitrary number of copies of every item can be
taken in UKP. Since UKP is NP-hard, fully polynomial time approximation schemes (FPTAS)
are of great interest. Such algorithms find a solution arbitrarily close to the optimum OPT(I),
i.e. of value at least (1 − ε)OPT(I) for ε > 0, and have a running time polynomial in the input
length and 1
ε. For over thirty years, the best FPTAS was due to Lawler with a running time
in O(n + 1
ε2 ), where n is the number of knapsack items.
We present an improved FPTAS with a running time in O(n + 1
ε) and a space bound in
ε log2 1
O(n + 1
ε). This directly improves the running time of the fastest known approximation
schemes for Bin Packing and Strip Packing, which have to approximately solve UKP instances
as subproblems.
ε3 ) and a space complexity in O(n + 1
ε2 log3 1
Introduction
1
An instance I of the Knapsack Problem (KP) consists of a list of n items a1, . . . , an, n ∈ N, where
every item has a profit pj ∈ (0, 1] and a size sj ∈ (0, 1]. Moreover, we have the knapsack size c = 1.
In the 0-1 Knapsack Problem (0-1 KP), a subset V ⊂ {a1, . . . , an} has to be chosen such that the
total profit of V is maximized and the total size of the items in V is at most c. Mathematically, the
j=1 sjxj ≤ c; xj ∈ {0, 1} ∀j}. In this paper, we focus on
the unbounded variant (UKP) where an arbitrary number of copies of every item is allowed, i.e. we
problem is defined by max{Pn
want to determine max{Pn
j=1 pjxjPn
j=1 pjxjPn
j=1 sjxj ≤ c; xj ∈ N ∀j}.
1.1 Known Results
The 0-1 Knapsack Problem and other variants of KP are well-known NP-hard problems [5]. They
can be optimally solved in pseudo-polynomial time by dynamic programming [1, 18]. Furthermore,
fully polynomial time approximation schemes (FPTAS) are known for different variants of KP. An
FPTAS is a family of algorithms (Aε)ε>0, where for every ε > 0 the algorithm Aε finds for a given
instance I a solution of profit Aε(I) ≥ (1− ε)OPT(I). The value OPT(I) denotes the optimal value
for I. FPTAS have a running time polynomial in 1
ε and the input length.
∗Research supported by DFG project JA612/14-2, "Entwicklung und Analyse von effizienten polynomiellen
Approximationsschemata für Scheduling- und verwandte Optimierungsprobleme"
1
ε) + 1
ε log( 1
ε2 log( 1
ε} + 1
ε3 log2( 1
ε) · min{n, 1
ε)}). Assuming that n ∈ Ω( 1
The first FPTAS for 0-1 KP was presented by Ibarra and Kim [8] with a running time in
ε), n}) and a space complexity in O(n + 1
ε2 · min{ 1
O(n log n + 1
ε2 log( 1
ε3 ). Lawler [21] improved the
running time to O( 1
ε4 + log( 1
ε)n). In 1981, Magazine and Oguz [23] presented a method to decrease
the space complexity of the dynamic program so that their FPTAS runs in time O(n2 log(n) 1
ε2 ) and
ε ). (The paper focuses on the improved space complexity without a partitioning
needs space in O( n
and reduction of the items as done e.g. by Lawler. Without it, Lawler's basic algorithm has in
fact a time and space complexity in O( n2
ε ).) The currently fastest known algorithm is due to
Kellerer and Pferschy [16, 17, 18, pp. 166 -- 183] with a space bound in O(n + 1
ε2 ) and a running
time in O(n min{log n, log 1
ε), this is
in O(n log( 1
ε)).
For UKP, Ibarra and Kim [8] presented the first FPTAS by extending their 0-1 KP algorithm.
ε) and a space complexity in O(n + 1
Their UKP algorithm has a running time in O(n + 1
ε3 ).
Kellerer et al. [18, pp. 232 -- 234] have moreover described an FPTAS with a running time in
O(n log(n) + 1
ε2 ). In 1979, Lawler [21] presented his
FPTAS with a running time in O(n + 1
ε), this
is still the best known FPTAS.
The study of KP is not only interesting in itself, it is moreover motivated by column generation
for optimization problems like the famous Bin Packing Problem and Strip Packing Problem. In
the former problem, a set J of n items of size in (0, 1] has to be packed in as few unit-sized bins as
possible. In the latter problem, a set J of n rectangles of width (0, 1] and height (0, 1] has to be
packed in a strip of unit width such that the height of the packing is minimized. Many algorithms
for optimization problems like Bin Packing have to solve linear programs (LPs), but enumerating all
columns of the linear programs would take too much time. One way to avoid this is the consideration
of the dual of the LP and to (approximately or exactly) solve a separation problem, e.g. KP, to find
violated inequalities of the dual. These inequalities correspond to columns in the primal LP: the
columns needed for solving the LP are therefore generated and added dynamically. Examples can
be found in [6, 14].
ε4 log 1
ε)) and a space bound in O(n + 1
ε3 ) and a space complexity in O(n + 1
ε2 ). For n ∈ Ω( 1
ε2 (n + log 1
ε log 1
2 for the absolute approximation ratio c. The bound 3
Since Bin Packing and Strip Packing are NP-complete [5], several approximation algorithms
have been found for both problems. However, no efficient (i.e. polynomal-time) algorithm A for BP
or SP can achieve A(J) ≤ c · OPT(J) for c < 3
2 and all problem instances J unless P = NP [5]: we
have c ≥ 3
2 is due to the fact that a polynomial
algorithm could otherwise distinguish between the optimum of 2 or 3 for BP instances and therefore
solve the NP-complete Partition Problem in polynomial time [5]. Since only such small instances
prevent an absolute ratio better than 3
2, larger instances may allow for a better approximation ratio.
So-called asymptotic fully polynomial-time approximation schemes (AFPTAS) (Aε)ε>0 are
therefore especially interesting. They find for every ε > 0 and instance J a solution of value at most
(1 + ε)OPT(J) + f( 1
ε. Roughly
speaking, AFPTAS achieve an approximation ratio of c = (1 + ε) for large problem instances.
For Bin Packing, the first AFPTAS was presented by Karmarkar and Karp [14] with f( 1
ε) =
ε2 ). In 1991, Plotkin et al. [24] described an improved algorithm with a smaller additive term
ε) = O( 1
ε)n). The AFPTAS by Shachnai and
ε)n) for general
ε) and the
O( 1
f( 1
ε6 log3( 1
Yehezkely [25] has the same additive term and a running time in O( 1
instances. Currently, the AFPTAS in [10] has the smallest additive term f( 1
fastest running time in O( 1
ε), and have a running time polynomial in the input length and 1
ε)) and running time in O( 1
ε) + log( 1
ε) = O(log2 1
ε) + log( 1
ε6 log6( 1
ε log( 1
ε6 log 1
ε + log( 1
ε)n).
2
The first AFPTAS for Strip Packing was presented by Kenyon and Rémila [19] with f( 1
Bougeret et al. [2] and Sviridenko [26] independently improved the additive term to f( 1
The algorithm in [2] needs time in O( 1
AFPTAS.
ε) = O( 1
ε2 ).
ε log 1
ε).
ε) + n log n), which is the currently fastest known
Both algorithms in [2, 10] solve UKP instances for column generation. A faster FPTAS for UKP
ε) = O( 1
ε6 log( 1
therefore directly yields faster AFPTAS for Bin Packing and Strip Packing.
1.2 Our Result
We have derived an improved FPTAS for UKP that is faster and needs less space than Lawler's
algorithm.
Theorem 1. There is an FPTAS for UKP with a running time in O(n + 1
complexity in O(n + 1
ε)) and a space
ε2 log3( 1
ε log2( 1
ε)).
Not only the improved running time, but also the improved space complexity is interesting
because "for higher values of 1
ε the space requirement is usually considered to be a more serious
bottleneck for practical applications than the running time" [18, p. 168]. Nevertheless, the improved
time complexity has direct practical consequences. Let KP(d, ε) be the running time to find a
(1 − ε) approximate solution to a UKP instance with d items. The Bin Packing algorithm in [10]
has the running time O(KP(d, ¯ε
ε2 ) (where
¯ε ∈ Θ(ε) and d ∈ O( 1
Corollary 2. There is an AFPTAS (Aε)ε>0 for Bin Packing that finds for ε ∈ (0, 1
J in Aε(J) ≤ (1 + ε)OPT(J) + O(log2( 1
ε)). By using the new FPTAS for UKP, we get the following result:
ε)n) if we assume that KP(d, ¯ε
2] a packing of
6) ∈ Ω( 1
ε + log( 1
ε3 log 1
6) · 1
ε log 1
(cid:18) 1
ε)) bins. Its running time is in
ε5 log4 1
(cid:18)1
+ log
(cid:19)
(cid:19)
n
ε
ε
.
O
6), d ln ln( d
Similarly, the Strip Packing algorithm in [2] (see also [9]) has a running time in O(d( 1
ε)} + n log n) where again d ∈ O( 1
ln d) max{KP(d, ¯ε
FPTAS yields the following improved AFPTAS:
Corollary 3. There is an AFPTAS (Aε)ε>0 for Strip Packing that finds a packing for J of total
height Aε(J) ≤ (1 + ε)OPT(J) + O( 1
ε2 +
ε) and ¯ε ∈ Θ(ε). The new
ε)). Its running time is in
ε log 1
(cid:18) 1
ε log( 1
ε5 log4 1
ε
O
(cid:19)
(cid:18)1
(cid:19)
ε
+ log
n
.
The result in this paper was first presented at IWOCA 2015 [12]. The final publication will be
available at link.springer.com.
For readers acquainted with column generation or linear programs, it should be noted that the
LP solved has the form min{cT x Ax ≥ b, x ≥ 0}. It is indeed a fractional covering problem where
the columns of A represent configurations: a configuration assigns item slots to one bin (for Bin
Packing) or to one shelf of the strip (for Strip Packing) such that the slots fit into the bin or the
strip. The primal LP is then approximately solved with a method by Grigoriadis et al. [7] (see also
[9]). The columns (i.e. configurations) are generated by solving so-called block problems, which are
UKP instances in this case. When the LP has been solved, each item is placed in a slot that has at
3
least the size of the item. As a feasible solution to the LP has been found, there are enough slots
for all items. Because of the unboundedness, some configurations may indeed assign more item
slots of a certain size to the strip or to one or several bins than there are items in the considered
Strip or Bin Packing instance. This does not represent a problem because the supernumerary item
slots are simply left empty in the final solution. For comparison, Plotkin et al. [24] solve the LP
with a decomposition method where the block problem has additional constraints on the knapsack
variables: it is a Bounded Knapsack Problem where a limited number dj ∈ N of copies for every
item aj may be taken.
1.3 Techniques
Most algorithms for UKP [8, 18, 21] rely on 0-1 KP algorithms. The 0-1 KP algorithms determine a
first lower bound P0 for OPT(I). Based on a threshold T depending on P0, the items are partitioned
into large(-profit) items with pj ≥ T and small(-profit) items with pj < T. A subset of the large
items is taken, which is sufficient for an approximate solution. Its profits are then scaled and the
well-known dynamic programming by profits applied to the subset. All combinations of large items
(packed by the dynamic program) and small items (which are greedily added) are checked and the
best one returned. For UKP, copies of the items in the reduced large item set are taken to transform
the UKP instance into a 0-1 KP instance.
Our algorithm also first reduces the number of large items. However, we further preprocess
the remaining large items by taking advantage of the unboundedness: large items of similar profit
[2kT, 2k+1T) are iteratively combined ("glued") together to larger items. Apart from two special cases
that can be easily solved, we prove for this new set G a structure property: there are approximate
solutions where at most one large item from every interval [2kT, 2k+1T) is used, i.e. only O(log 1
ε)
items in total. As a next step, a large item aeff−c that consists of several copies of the most efficient
small item aeff is introduced. We prove that there are now approximate solutions to the large items
G ∪ {aeff−c} of cardinality O(log 1
4 P0.
Instead of exact dynamic programming, we use approximate dynamic programming: the profits in
[ 1
4 P0, 2P0] are divided into intervals of equal length. During the execution of the dynamic program,
we eliminate dominated solutions and store for each interval at most one solution of smallest size.
The combination of approximate dynamic programming with the structure properties yields the
considerable improvement in the running time and the space complexity. The algorithm then returns
the best combination of large items (packed by the dynamic program) and copies of the small item
aeff (added greedily).
ε) and that additionally use at least one item of profit at least 1
2 Preliminaries
We introduce some useful notation. The profit of an item a is denoted by p(a) and its size by s(a).
If a = aj, we also write p(aj) = pj and s(aj) = sj. Let V = {xa : a a ∈ I, xa ∈ N} be a multiset
of items, i.e. a subset of items in I with their multiplicities. We naturally define the total profit
Let v ≤ c = 1 be a part of the knapsack. The corresponding optimum profit for the volume
a∈I s(a)xa ≤ v; a ∈ I; xa ∈ N}. Obviously,
xa>0 p(a)xa and the total size s(V ) :=P
a∈I p(a)xa P
p(V ) :=P
v is denoted by OPT (I, v) = max{P
OPT(I) = OPT (I, c) holds.
xa>0 s(a)xa.
4
We assume throughout the paper that basic arithmetic operations as well as computing the
logarithm can be performed in O (1).
Finally, we have a remark about the use of "item" and "item copy" when we consider a solution
to a UKP instance.
Remark 4. Let I, I be two sets of knapsack items with I ⊆ I. In the 0-1 Knapsack Problem, a
sentence like "the solution to I uses at most one item in I" is obvious: if the solution uses one item
in I, all other items of the solution are in I \ I.
Consider now UKP. When we talk about solutions, we would formally have to distinguish between
an item a0 ∈ I in the instance and the item copies of a0 that a solution V = {xa : a a ∈ I, xa ∈ N}
uses. In this paper, we however use the expressions "item" and "item copy" interchangeably when
talking about solutions. As an example, let us consider the sentence "the solution to I uses at most
one item in I." It means that the solution contains item copies of items in I, but at most one item
copy whose corresponding item is in I. To be more precise, the multiset V uses only one item a ∈ I
a0∈I xa0 ≤ 1.
Similarly, "the solution V uses at most two items in I" means that there are only two item copies
with a multiplicity xa > 0. We have xa ≤ 1, but xa00 = 0 for all other a00 ∈ I, i.e. P
whose corresponding item(s) are in I: we haveP
The interchangeable use of "item" and "item copy" allows for shorter sentences. Moreover, it is
a0∈I xa0 ≤ 2.
based upon 0-1 KP where "item" and "item copy" are in fact identical.
1
s(a). Fill the knapsack with as many copies of ameff as possible, i.e. take b
p(a)
2.1 A First Approximation
We present a simple approximation algorithm for OPT(I). Take the most efficient item ameff :=
s(ameff)c c=1=
arg maxa∈I
b
s(ameff)c copies of ameff. Then the following holds:
s(ameff)c ≥ 1
Theorem 5. We have P0 := p(ameff) · b
O (n) and space O (1).
Proof. Suppose first that ameff can greedily fill the knapsack completely. Then p(ameff) · b
OPT(I). Otherwise, one additional item ameff exceeds the capacity c. Then p(ameff) · b
p(ameff) ≥ OPT(I).
2OPT(I), and the theorem follows. Otherwise p(ameff) · b
1
also proves the theorem.
s(ameff)c =
s(ameff)c +
s(ameff)c ≥ OPT(I) − p(ameff) ≥
2OPT(I), which
To determine P0, we only have to check all items (which can be done in O (n)) and to save the
2OPT(I). The value P0 can be found in time
2OPT(I), then p(ameff) · b
s(ameff)c ≥ p(ameff) ≥ 1
If p(ameff) ≤ 1
c
c
c
c
c
c
most efficient item (which only needs time in O (1)).
(The proof is taken from [18, p. 232, 21])
Assumption 1. From now on, we assume without loss of generality that ε ≤ 1
κ ∈ N. Otherwise, we replace ε by the corresponding
log2( 2
ε) = κ holds.
Similar to Lawler [21], we introduce the threshold T and a constant K:
1
2κ−1 such that
2κ−1 ≤ ε < 1
1
4 and ε = 1
2κ−1 for
2κ−2 . Note that
1
2κ−1 P0
T := 1
2 εP0 = 1
2
5
(1)
and
K := 1
4
1
log2( 2
ε) + 1 εT = 1
4
1
κ + 1
1
2κ−1 T = 1
8
1
κ + 1
(cid:19)2
(cid:18) 1
2κ−1
P0 .
(2)
We will see later that these values are indeed the right choice for the algorithm. (A derivation of
these values is presented in [20, Subsection 5.8.1].)
3 Reducing the Items
We first partition the items into large(-profit) and small(-profit) items, and only keep the most
efficient small item:
IL := {a ∈ I p(a) ≥ T} , IS := I \ IL, and aeff := arg max
(cid:26) p(a)
s(a)
(cid:27)
(cid:12)(cid:12)(cid:12) p(a) < T
.
Theorem 6. The sets IL, IS and the item aeff can be found in time O (n) and space O (n). This is
also the space needed to save IL.
Proof. Obvious.
Similar to Lawler, we now reduce the item set IL. Note that we have OPT(I) ≤ 2P0 according
to Theorem 5, and one item cannot have a profit larger than OPT(I) ≤ 2P0. Hence, the large item
profits are in the interval [T, 2P0]. We partition this interval into
Note that
L(κ) =h2κT, 2κ+1T
(cid:17) =
L(k) := [2kT, 2k+1T) for k ∈ {0, . . . , κ + 1} .
(cid:19)
1
2κ−1 P0, 2κ+1 1
2
1
2κ−1 P0
(cid:20)
2κ 1
2
= [P0, 2P0) .
For convenience, we directly set L(κ+1) := {2P0}.
We further split the L(k) into disjoint sub-intervals, each of length 2kK:
:=h2kT + γ · 2kK, 2kT + (γ + 1)2kK
(cid:17) for γ ∈n0, . . . , 2κ+1(κ + 1) − 1o
L(k)
γ
Note that indeed L(k) =S
(k)
γ holds because
γ L
(3)
.
(4)
2kT + (γ + 1)2kKγ=2κ+1(κ+1)−1 = 2kT + 2κ+1(κ + 1)2kK
(2)= 2kT + 2κ+1(κ + 1)2k 1
4
= 2kT + 2kT = 2k+1T .
1
κ + 1
1
2κ−1 T
Similar to above, we set L
(κ+1)
0
:= {2P0}.
The idea is to keep only the smallest item a for every profit interval L
items are sufficient to determine an approximate solution.
(k)
γ . We will see that these
6
Definition 7. For an item a with p(a) ≥ T, let k(a) ∈ N be the interval such that p(a) ∈ L(k(a))
and γ(a) ∈ N be the sub-interval such that p(a) ∈ L
be the smallest item for the profit
interval L
(k)
. Let a
γ
(k(a))
γ(a)
(k)
γ , i.e.
:= arg minn
a(k)
γ
s(a) a ∈ IL and p(a) ∈ L(k)
γ
o for all k and γ .
Algorithm 1 shows the algorithm to determine the a
IL,red :=[
[
(k)
γ . They form the reduced set of large items
γ } .
{a(k)
As in [21], we now prove that IL,red is sufficient for an approximation.
k
γ
Algorithm 1: The algorithm to determine the a
(k)
γ .
for k = 0, . . . , κ do
for γ = 0, . . . , 2κ+1(κ + 1) − 1 do
:= ∅;
(k)
γ
a
:= ∅;
(κ+1)
a
0
for a ∈ IL do
Determine (k(a), γ(a));
(k(a))
if s(a
γ(a) ) > s(a) or a
(k(a))
S
a
γ(a)
Output: IL,red :=S
:= a;
γ{a
k
γ }
(k)
γ(a) = ∅ then
(k(a))
Lemma 8. Let 0 ≤ v ≤ c = 1. Then
and
OPT ({aeff}, c − v) ≥ OPT (IS, c − v) − T
OPT (IL,red, v) ≥
1 − ε
4
!
1
log2( 2
ε) + 1
OPT (IL, v) .
Proof. For the first inequality, there are two possibilities: either copies of aeff can be taken such
that the entire capacity c − v is used. Then obviously OPT ({aeff}, c − v) = OPT (IS, c − v)
holds. Otherwise, we have similar to the proof of Theorem 5 that OPT({aeff}, c − v) + p(aeff) =
s(aeff)c · p(aeff) + p(aeff) ≥ OPT (IS, c − v). Thus, OPT({aeff}, c − v) ≥ OPT(IS, c − v) − p(aeff) ≥
b c−v
For the second inequality, take an optimal solution (xa)a∈I such that OPT (IL, v) =P
OPT(IS, c − v) − T. The first inequality follows.
p(a)xa.
in IL,red. Obviously, the solution stays feasible,
Replace now every item a by its counterpart a
a∈IL
(k(a))
γ(a)
7
i.e. the volume v will not be exceeded, because an item may only be replaced by a smaller one. This
(cid:19)
1
κ + 1
1
2κ−1
(5)
(cid:19)
1
2κ−1
· p(a)xa
solution has the total profitP
p(a
(k(a))
γ(a) )
≥
p(a)≥2k(a)T≥
by the definition of the L
p(a
a∈IL
p(a) − 2k(a)K
1
(cid:18)1
p(a) −
1
κ + 1
(k(a))
γ(a) )xa. Moreover, we have
(cid:18)
(2)= p(a) − 1
(cid:19)
4
p(a) = p(a) ·
(cid:18)
1
2κ−1
κ + 1
4
1
2κ−1 2k(a)T
1 − 1
4
(k)
γ . We get
OPT (IL,red, v) ≥ X
p(a
(cid:18)
a∈IL
1 − 1
4
1 − ε
4
=
=
(5)≥ X
(cid:19)
!
(k(a))
γ(a) )xa
1
1
2κ−1
κ + 1
1
log2( 2
ε) + 1
1 − 1
4
1
κ + 1
a∈IL
OPT (IL, v)
OPT (IL, v) .
(The reasoning is partially taken directly from or close to the one by Lawler in [21].)
Theorem 9. The set IL,red has O( 1
ε) items. Algorithm 1 needs time in O(n + 1
space in O( 1
ε) for the construction and for saving IL,red.
ε log2 1
ε log2 1
ε log2 1
ε) and
ε ·( 1
ε log2 1
(k)
(κ+1)
γ , including the item a
0
ε)) = O( 1
ε log 1
(k)
γ . Finally, the running time is obviously bounded by O(n + 1
, is bounded by O((κ + 1)· (2κ+1(κ + 1)−
Proof. The number of items a
1+1)) = O(log 1
ε). The space needed is asymptotically bounded by the space
required to save the a
ε): the
values k(a) and γ(a) can be found in O(1) because we assume that the logarithm can be determined
in O (1).
Remark 10. If there is one item a with the profit p(a) = 2P0, i.e. whose profit attains the upper
bound, one optimum solution obviously consists of this single item. During the partition of I into
IL and IS, it can easily be checked whether such an item is contained in I. Since the algorithm can
directly stop if this is the case, we will from now on assume without loss of generality that such an
item does not exist and that a
ε log2 1
= ∅.
(κ+1)
0
4 A Simplified Solution Structure
In this section, we will transform IL,red into a new instance G whose optimum OPT(G, v) is only
slightly smaller than OPT(IL,red, v) and where the corresponding solution has a special structure.
This new transformation will allow us later to faster construct the approximate solution. First, we
define
(cid:12)(cid:12) p(a) ∈ L(k)o =n
a ∈ IL,red
(cid:12)(cid:12) p(a) ∈h2kT, 2k+1T
(cid:17)o
.
I(k) :=n
a ∈ IL,red
Note that the items are already partitioned into the I(k) because of the way IL,red has been
constructed.
8
Definition 11. Let a1, a2 be two knapsack items with s(a1) + s(a2) ≤ c. The gluing operation ⊕
combines them into a new item a1⊕ a2 with p(a1⊕ a2) = p(a1)+ p(a2) and s(a1⊕ a2) = s(a1)+ s(a2).
Thus, the gluing operation is only defined on pairs of items whose combined size does not exceed
c.
The basic idea for the new instance G is as follows: we first set G(0) := I(0). Then, we construct
a1 ⊕ a2 for all a1, a2 ∈ G(0) (which also includes the case a1 = a2), which yields the item set
H(1) := {a1 ⊕ a2 a1, a2 ∈ G(0)}. Note that p(a1 ⊕ a2) ∈ [2T, 4T) = L(1). For every profit
γ , we keep only the item of smallest size in I(1) ∪ H(1), which yields the item set G(1).
(1)
interval L
This procedure is iterated for k = 1, . . . , κ − 1: the set G(k) contains the items with a profit in
[2kT, 2k+1T) = L(k) (see Fig. 1(a)). Gluing like above yields the item set H(k+1) with profits in
[2k+1T, 2k+2T) = L(k+1) (see Fig. 1(b)). By taking again the smallest item in H(k+1) ∪ I(k+1) for
(k)
(k+1)
, the set G(k+1) is derived (see Fig. 1(c)). The item in G(k) with a profit in L
every L
is
γ
γ
(k)
denoted by a
γ
We finish when G(κ) has been constructed. We are in the case where I(κ+1) = ∅, i.e. a
(κ+1)
0
= ∅,
and it is explained at the beginning of Section 5 that it is not necessary to construct G(κ+1) from
(κ+1)
the items in G(κ). Hence, we also have a
0
Note that we may glue items together that already consist of glued items. For backtracking, we
(k)
γ has already been an
γ which two items in G(k−1) have formed it or whether a
(k)
save for every a
item in I(k). Algorithm 2 presents one way to construct the sets G(k).
for every k and γ.
= ∅.
(k)
Remark 12. One item a
γ
(k)
size of a
γ
combinations because an arbitrary number of item copies can be taken in UKP.
is in fact the combination of several items in IL,red. The profit and
represent feasible item
is equal to the total profit and size of these items. The a
(k)
γ
The item set
G :=
κ[
G(k)
k=0
has for every 0 ≤ v ≤ c a solution near the original optimum OPT (IL,red, v) as shown below in
Theorem 14. It is additionally proved that at most one item of every G(k) for k ∈ {0, . . . , κ − 1} is
needed. First, we introduce a definition for the proof.
Definition 13. Let I0 be a set of knapsack items with p(a) ≥ T for every a ∈ I0. For a knapsack
volume v ≤ c and k0 ∈ {0, . . . , κ}, a solution is structured for k = k0 if it fits into v and uses for
every k ∈ {0, . . . , k0} at most one item copy with a profit in L(k). We denote by OPT≤k0 (I0, v) the
corresponding optimum profit.
For instance, the solution for
(cid:16)
OPT≤k0
G(0) ∪ . . . ∪ G(k0) ∪ G(k0+1) ∪ I(k0+2) ∪ . . . ∪ I(κ), v
(cid:17)
fits into the volume v, and it uses only one item from every G(k) for k ∈ {0, . . . , k0}. It may however
use an arbitrary number of item copies e.g. in G(k0+1) or I(k0+2).
9
(a) The items in G(k) and I(k+1). The height of every item a corresponds to its size s(a) while its position on the axis
(k)
γ = [2kT +γ2kK, 2kT +(γ+1)2kK).
corresponds to its profit p(a). The axis is partitioned into the profit sub-intervals L
(b) The set I(k+1) together with the newly constructed items in H(k+1)
(c) The new set G(k+1) after keeping only the smallest item with a profit in L
because it is the smallest item in its profit sub-interval L
(k+1)
γ
.
(k+1)
γ
. For instance, a4 ⊕ a4 is kept
Figure 1: Principle of deriving G(k+1) from G(k) and I(k+1)
10
2kT+γ·2kK2kT2k+1TG(k)2k+1T+γ·2k+1K2k+1T2k+2Ta0a3s(a3)a4a0a1a3a4I(k+1)2k+1T+γ·2k+1K2k+1T2k+2Ta0a0a0⊕a0a3⊕a1a0a4⊕a3a3a3⊕a3a4⊕a4a4⊕a42k+1T+γ·2k+1K2k+1T2k+2TG(k+1)a0a1a0a4⊕a3a4a4⊕ Algorithm 2: The construction of the item sets G(k).
for k = 0, . . . , κ do
for γ = 0, . . . , 2κ+1(κ + 1) − 1 do
:= a
(k)
(k)
a
γ ;
γ
Backtrack(a
(k)
γ ) := a
(k)
γ ;
G(0) := I(0);
for k = 0, . . . , κ − 1 do
for γ0 = γ, . . . , 2κ+1(κ + 1) − 1 do
for γ = 0, . . . , 2κ+1(κ + 1) − 1 do
γ0 ) ≤ c then
(k)
(k)
γ ) + s(a
γ ⊕ a
(k)
(k)
a := a
γ0 ;
(k+1)
if s(a) < s(a
γ(a) ) or a
if s(a
γ(a) = ∅ then
(k+1)
(k+1)
a
:= a;
γ(a)
Backtrack(a
(k+1)
γ(a) ) := (a
(k)
γ , a
(k)
γ0 );
(k+1)
0
, . . . , a
(k+1)
2κ+1(κ+1)−1
G(k+1) :=na
k0+1[
OPT≤k0
o;
≥
Theorem 14. For v ≤ c and k0 ∈ {0, . . . , κ − 1}, we have
κ[
G(k) ∪
I(k), v
k=0
k=k0+2
1 − ε
4
1
log2( 2
ε) + 1
!k0+1
OPT (IL,red, v) .
Proof. The proof idea is quite simple: we iteratively replace the items in I(k0+1) by their counterpart
in G(k0+1) and also replace every pair of item copies in G(k0) by the counterpart in G(k0+1). This
directly follows the way to construct the item sets G(k) presented in Algorithm 2.
Formally, the statement is proved by induction over k0. Let k0 = 0. Take an optimum solution
to G(0) ∪ I(1) ∪ . . . ∪ I(κ) = I(0) ∪ I(1) ∪ . . . ∪ I(κ) = IL,red. For ease of notation, we directly write
each item as often as it appears in the solution. We have three sub-sequences:
• Let ¯a1, . . . , ¯aη (η ∈ N) be the items from G(0) = I(0) in the optimal solution for OPT(IL,red, v).
We assume that η is odd (the case where η is even is easier and handled below.)
• Let ¯aη+1, . . . , ¯aη+ξ (ξ ∈ N) be the items from I(1) in the optimal solution for OPT(IL,red, v).
λ (λ ∈ N) be the remaining items from I(2) ∪ . . . ∪ I(κ) in the optimal solution for
• Let ¯a0
OPT(IL,red, v). This set is denoted by Λ. As defined above, the total profit of these items is
written as p(Λ).
1, . . . , ¯a0
Figure 2(a) illustrates the packing. (Figure 2 shows the case for general k.) We have
G(0) ∪ I(1) ∪ . . . ∪ I(κ), v
p(¯ai) +
p(¯aj) + p(Λ) .
(6)
OPT(cid:16)
η+ξX
j=η+1
(cid:17) =
ηX
i=1
11
(1)
γ(¯a2i−1⊕¯a2i) =: a
ρ(i) in G(1) (for i ∈ {1, . . . ,b η
(1)
In the first step, every pair of items ¯a2i−1 and ¯a2i from G(0) for i ∈ {1, . . . ,b η
2c} is replaced
by ¯a2i−1 ⊕ ¯a2i ∈ H(1) (see Fig. 2(b)). In the second step, every item ¯a2i−1 ⊕ ¯a2i is again replaced
2c}). Only item ¯aη
by the corresponding item a
(1)
remains unchanged. Moreover, ¯aj from I(1) is replaced by the corresponding a
ρ(j) for
j ∈ {η + 1, . . . , η + ξ} (see Fig. 2(c)). Note that this new solution is indeed feasible because the
(1)
γ are at most as large as the original ones. Moreover, the corresponding items
replacing items a
(1)
(1)
ρ(j) must exist by the construction of G(1). Thus, we have a (feasible) solution that
a
ρ(i) and a
consists of the item ¯aη ∈ G(0), the items a
1, . . . , ¯a0
in I(2), . . . , I(κ): this solution respects the structure of OPT≤k0(·, v) for k0 = 0. (If η is even, no
item in G(0) is used.)
ρ(j) in G(1), and the remaining items ¯a0
(1)
Let now ¯a be an item ¯a2i−1 ⊕ ¯a2i or ¯aj. It can be proved as for Inequality (5) that
(1)
γ(¯aj) =: a
(1)
ρ(i) and a
λ
1 − ε
4
p(a
γ(¯a)) ≥
(1)
!
1
log2( 2
ε) + 1
2 cX
b η
i=1
p(¯a2i−1 ⊕ ¯a2i)
p(¯a) .
(7)
η+ξX
j=η+1
p(a
(1)
ρ(i)) +
p(a
(1)
ρ(j)) + p(Λ)
Thus, we have
OPT≤0
+
i=1
(cid:16)
(7)≥ p(¯aη) +
1 − ε
4
(cid:17) ≥ p(¯aη) +
G(0) ∪ G(1) ∪ I(2) ∪ . . . ∪ I(κ), v
! b η
2 cX
1
1 − ε
! η+ξX
log2( 2
4
ε) + 1
1
! ηX
log2( 2
ε) + 1
!
OPT(cid:16)
!
p(¯aj) + p(Λ)
p(¯ai) +
η+ξX
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1 − ε
4
1 − ε
4
1 − ε
4
OPT (IL,red, v) .
≥
(6)=
=
j=η+1
i=1
j=η+1
p(¯aj) + p(Λ)
G(0) ∪ I(1) ∪ . . . ∪ I(κ), v
(cid:17)
The statement for k0 = 1, . . . , κ − 1 now follows by induction. The proof is almost identical to the
case k0 = 0 above, the only difference is that there are additionally the items in G(0), . . . , G(k0−1) that
remain unchanged like the items I(k0+2), . . . , I(κ). Only the items in G(k0) and I(k0) are replaced.
Lemma 15. OPT(G ∪ {aeff}) ≤ OPT(IL,red ∪ IS) ≤ OPT(IL ∪ IS) = OPT(I) ≤ 2P0 holds.
Proof. G consists of items in IL,red or of items that can be obtained by gluing several items in
IL,red together. Every combination of items in G can therefore be represented by items in IL,red.
Moreover, we have aeff ∈ IS. The first inequality follows. Since IL,red ⊆ IL, the second inequality is
obvious. The last inequality follows from Theorem 5.
12
(a) The current solution to
G(0) ∪ ··· ∪ G(k) ∪ I(k+1) ∪
··· ∪ I(κ). The structure of
OPT≤k−1(·, v) is respected,
i.e. at most one item from ev-
ery G(0), . . . , G(k−1) is used.
(b) The items
in G(k) are pair-
wise glued to-
gether with the
possible
excep-
tion of one item.
(c) The items in H(k+1) ∪
I(k+1) are replaced by their
counterparts in G(k+1). Now,
at most one item in G(k) is
part of the solution, and the
structure for OPT≤k(·, v) is
respected.
Figure 2: The principle of the proof for Theorem 14
13
a4a4a4a3a0a1a4G(k)I(k+1)G(0),...,G(k−1)I(k+2),...,I(κ)a4a4a4⊕a3a0⊕a1a4G(k)H(k+1)I(k+1)a4a4a4⊕a1a1a4a4⊕G(k)G(k+1)G(0),...,G(k−1)I(k+2),...,I(κ) Up to now, we have (only) reduced the original item set I to G ∪ {aeff}.
= ∅. Consider the optimum structured
(κ+1)
Lemma 16. Assume as mentioned in Remark 10 that a
0
solutions to G ∪ {aeff} for k0 = κ − 1 (see Definition 13). This means that at most one item is used
from every G(k) for k ∈ {0, . . . , κ − 1}. (The item aeff has a profit p(aeff) < T such that it does not
have to satisfy any structural conditions.) Then there are two possible cases:
G∪{aeff} is 2P0, and the solution consists of two item copies of the item a
• One solution uses (at least) two items in G(κ). This is the case if and only if the optimum for
(κ)
0 ) = P0.
• Every solution uses at most one item in G(κ). Then, OPT≤κ−1(G, v0) = OPT≤κ(G, v0) holds
(κ)
0 with p(a
for all values 0 ≤ v0 ≤ c, and there is a value 0 ≤ v ≤ c such that
OPT≤κ (G, v) + OPT ({aeff} , c − v) = OPT≤κ−1 (G, v) + OPT ({aeff} , c − v)
1 − ε
4
≥
1
log2( 2
ε) + 1
!κ+1
OPT(I) − T .
Moreover, OPT≤κ(G, v) uses at least one item in G(κ−2) ∪ G(κ−1) ∪ G(κ), and/or we have
OPT ({aeff} , c − v) ≥ 1
4 P0.
can be used, and we have p(a
Proof. Note that IL,red does not contain any item with the profit 2P0 (see Remark 10). By
construction, this is still the case for G. Suppose now that one solution to G ∪ {aeff} uses more
than one item in G(κ). Since items in G(κ) have profits in [P0, 2P0), only two copies of the item
(κ)
(κ)
a
0 ) = P0. In fact, 2P0 is the maximum possible profit because
0
OPT(G∪{aeff}) ≤ OPT(I) ≤ 2P0 holds as we have seen in Lemma 15. Thus, the "only if" direction
has been proved. The "if"-direction is obvious.
Suppose now that every structured solution to G ∪ {aeff} for k0 = κ − 1 uses at most one item
in G(κ). Thus, OPT≤κ−1(G, v0) = OPT≤κ(G, v0) holds for all 0 ≤ v0 ≤ c.
Let v ≤ c now be the volume the large items IL occupy in an optimum solution to I. Then
obviously OPT(I) = OPT (IL, v) + OPT (IS, c − v) holds. We have the following inequality:
OPT≤κ (G, v) + OPT ({aeff} , c − v) = OPT≤κ−1 (G, v) + OPT ({aeff} , c − v)
1 − ε
4
1 − ε
4
1 − ε
4
1 − ε
4
Thm. 14≥
Lem. 8≥
≥
=
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
!κ
OPT (IL,red, v) + OPT ({aeff} , c − v)
!κ+1
!κ+1
!κ+1
OPT (IL, v) + OPT (IS, c − v) − T
(OPT (IL, v) + OPT (IS, c − v)) − T
OPT(I) − T .
For the final property, suppose that no item in G(κ−2) ∪ G(κ−1) ∪ G(κ) is used in a solution for
OPT≤κ(G, v). Then we have
OPT≤κ (G, v) ≤ κ−3X
maxn
p(a) a ∈ G(k)o ≤ κ−3X
k=0
2 · 2kT < 2κ−1T
(1)= 1
2 P0 .
k=0
14
(8)
On the other hand, Inequality (8) together with (1 − δ)k ≥ (1 − k · δ) for δ < 1 yields
!
1 − ε
(cid:19)
log2( 2
4
ε) + 1
OPT(I) − 1
κ + 1
1 − ε
4
2 εOPT(I)
OPT(I) − T
(cid:19)
OPT≤κ (G, v) + OPT ({aeff} , c − v) ≥
(cid:18)
2 εP0 ≥
(cid:18)
(1)=
1 − ε
4
ε≤1/4≥ 3
4OPT(I) ≥ 3
OPT(I) − 1
4 P0 .
Hence, OPT({aeff} , c − v) ≥ 1
Definition 17. Take d P0/4
aeff−c.
4 P0 holds. The final property of the second case follows.
p(aeff)e items aeff. If their total size is at most c, they are glued together to
Obviously, aeff−c consists of the smallest number of items aeff whose total profit is at least P0
4 .
Moreover, aeff−c is a large item.
Definition 18. Take a knapsack volume v ≤ c. Consider the following solutions to G ∪ {aeff−c} of
size at most v:
• They are structured for k = κ, i.e. they use for every k ∈ {0, . . . , κ} at most one item in G(k).
• They additionally use the item aeff−c at most once and at least one item a ∈ G(κ−2) ∪ G(κ−1) ∪
G(κ) ∪ {aeff−c}.
4 P0. These special solutions are
The optimal profit for such solutions of total size at most v is denoted by OPTSt(G∪{aeff−c}, v).
Hence, these solutions have a profit of at least p(a) ≥ 2κ−2T = 1
called structured solutions with a lower bound (on the profit).
If v is too small such that such a solution does not exist, we set OPTSt(G ∪ {aeff−c}, v) = 0.
Theorem 19. In the second case of Lemma 16, there is a value 0 ≤ v ≤ c such that
1 − ε
4
1
log2( 2
ε) + 1
!κ+1
OPTSt (G ∪ {aeff−c} , v) + OPT ({aeff} , c − v) ≥
OPT(I) − T .
Proof. Like in the proof of Lemma 16, let v0 be the volume the large items IL occupy in an optimum
solution to I so that OPT(IL, v0) + OPT(IS, c − v0) = OPT(I). Consider an optimum solution
for OPT≤κ(G, v0) and suppose that it does not use any item in G(κ−2) ∪ G(κ−1) ∪ G(κ). Lemma
16 states that OPT({aeff}, c − v0) has a profit of at least 1
4 P0. Thus, a subset of the item copies
of aeff can be replaced by aeff−c, and c − v0 ≥ s(aeff−c). We set v := v0 + s(aeff−c). Note that
OPTSt(G ∪ {aeff−c}, v) ≥ OPT≤κ(G, v0) + p(aeff−c). Moreover, OPT≤κ(G, v0) = OPT≤κ−1(G, v0)
15
holds because we are in the second case of Lemma 16. We get the following inequalities:
OPTSt (G ∪ {aeff−c} , v) + OPT ({aeff} , c − v)
≥ OPT≤κ
= OPT≤κ−1(cid:0)G, v0(cid:1) + OPT(cid:0){aeff} , c − v0(cid:1)
(cid:0)G, v0(cid:1) + p(aeff−c) + OPT(cid:0){aeff} , c − v0 − s(aeff−c)(cid:1)
!κ
OPT(cid:0)IL,red, v0(cid:1) + OPT(cid:0){aeff} , c − v0(cid:1)
!κ+1
OPT(cid:0)IL, v0(cid:1) + OPT(cid:0)IS, c − v0(cid:1) − T
!κ+1(cid:0)OPT(cid:0)IL, v0(cid:1) + OPT(cid:0)IS, c − v0(cid:1)(cid:1) − T
!κ+1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
OPT(I) − T .
1 − ε
4
1 − ε
4
1 − ε
4
1 − ε
4
Thm. 14≥
Lem. 8≥
≥
=
Note that OPT({aeff}, c − v) is well-defined -- and therefore the entire chain of inequalities feasible --
because c − v = c − v0 − s(aeff−c) ≥ 0.
Suppose now that the optimal solution uses at least one item in G(κ−2) ∪ G(κ−1) ∪ G(κ). We can
then directly set v := v0, and the proof is similar to the first case above.
Roughly speaking, a solution in the first case of this proof satisfies the lower bound of the
theorem and uses at most one item in every G(k), but no item in G(κ−2), G(κ−1) or G(κ). This
implies that enough items aeff are part of the solution such that a subset of them can be replaced
by aeff−c.
So far, we have not constructed an actual solution. We have only shown in Theorem 19 that
there is a solution to G ∪ {aeff−c} ∪ {aeff} that is close to OPT(I) and that is a structured solution
with a lower bound.
Theorem 20. The cardinality of G(k) is in O( 1
constructs G in time O( 1
ε)) and space O( 1
and the backtracking information. The item aeff−c can be constructed in time O (1).
Proof. The statement for aeff−c is trivial: the number of items aeff−c to glue together can be
determined by division.
ε) items. Algorithm 2
ε), which also includes the space to store G
ε), i.e. G has O( 1
ε log 1
ε log2 1
ε2 log3( 1
The number of items in G(k) and G can be derived like the number of items in IL,red in Theorem
9. The running time of Algorithm 2 is obviously dominated by the second for-loop. It is in
ε log2 1
(cid:18)
κ ·(cid:16)2κ+1(κ + 1)(cid:17)2(cid:19)
O
= O
log
(cid:18)1
(cid:19)
ε
(cid:18)1
ε
·
log 1
ε
(cid:19)2!
(cid:18) 1
ε2 log3(cid:18)1
ε
(cid:19)(cid:19)
.
= O
(k)
The space complexity is dominated by the space to save the a
γ
which is again asymptotically equal to the number of items in G.
and the backtracking information,
16
5 Finding an Approximate Structured Solution by Dynamic Pro-
gramming
The previous section has presented three cases:
1. The instance I has one item of profit 2P0: return this item for an optimum solution, and
OPT(I) = 2P0 (see Remark 10).
2. If this is not the case, and G has one item of profit P0 and size at most c
2, two copies of this
item are an optimum solution to G ∪ {aeff} (see Lemma 16). Undoing the gluing returns an
optimum solution with OPT(I) = 2P0.
3. Otherwise, there is an approximate structured solution to G ∪ {aeff−c} ∪ {aeff} with a lower
bound (see Theorem 19).
The first two cases can be easily checked, which is the reason why it has not been necessary to
construct the set G(κ+1). We will from now on assume that we are in the third case: a solution uses
at most one item from every G(k) for k ∈ {0, . . . , κ} as well as aeff−c at most once. At the same
time, at least one item a ∈ G(κ−2) ∪ G(κ−1) ∪ G(κ) ∪ {aeff−c} is chosen. (See Definition 18.)
We use dynamic programming to find for all 0 ≤ v ≤ c the corresponding set of large items
V ⊆ G ∪ {aeff−c} with s(V ) ≤ v. For convenience, let G(κ+1) := {aeff−c}. We introduce tuples
(p, s, k) similar to Lawler [21]. For profit p with 0 ≤ p ≤ 2P0 and size 0 ≤ s ≤ c, the tuple (p, s, k)
states that there is an item set of size s whose total profit is p. Moreover, the set has only items in
G(k) ∪ ··· ∪ G(κ+1) and respects the structure above.
The dynamic program is quite simple: start with the dummy tuple set F (κ+2) := {(0, 0, κ + 2)}.
For k = κ + 1, . . . , κ − 2, the tuples in F (k) are recursively constructed by
F (k) :=n(p, s, k) (p, s, k + 1) ∈ F (k+1)o
∪n(p + p(a), s + s(a), k) (p, s, k + 1) ∈ F (k+1), a ∈ G(k), s + s(a) ≤ c
o
.
Note that (0, 0, k + 1) ∈ F (k+1), which guarantees that F (k) also contains the tuples (p(a), s(a), k)
for a ∈ G(k) if k ∈ {κ + 1, . . . , κ − 2}. For k ∈ {κ − 3, . . . , 0}, this tuple (0, 0, k + 1) is no longer
considered to form the new tuples, which guarantees that tuples of the form (p + p(a), s + s(a), k)
for a ∈ G(k) have p, s 6= 0. The recursion becomes
∪n(p + p(a), s + s(a), k) (p, s, k + 1) ∈ F (k+1) \ {(0, 0, k + 1)} , a ∈ G(k), s + s(a) ≤ c
o
.
F (k) :=n(p, s, k) (p, s, k + 1) ∈ F (k+1)o
The actual item set corresponding to (p, s, k) can be reconstructed by saving backtracking information.
Definition 21. A tuple (p2, s2, k) is dominated by (p1, s1, k) if p2 ≤ p1 and s2 ≥ s1.
As in [21], dominated tuples (p, s, k + 1) are now removed from F (k+1) before F (k) is constructed.
This does not affect the outcome: dominated tuples only stand for sets of items with a profit not
larger and a size not smaller than non-dominated tuples. A non-dominated tuple (p, s, k) is therefore
optimal, i.e. the profit p can only be obtained with items of size at least s if items in G(k), . . . , G(κ+1)
are considered.
17
Lemma 22. A tuple (p, s, k) ∈ F (k) stands for a structured solution with a lower bound (see
Definition 18). Therefore, we have p ≥ 2κ−2T if p > 0. For every v ≤ c, there is a tuple
(p, s, 0) ∈ F (0) with p = OPTSt(G ∪ {aeff−c}, v) and s ≤ v.
Proof. This lemma directly follows from the dynamic program: tuples use at most one item from
every G(k). For k ∈ {κ−2, . . . , κ+1}, a tuple with p > 0 represents an item set that uses at least one
item in G(k), . . . , G(κ+1), and such an item has a profit of at least 2κ−2T. Tuples for k ≤ κ − 3 with
p > 0 are only derived from tuples that use at least one item in G(κ−2), . . . , G(κ+1). If dominated
tuples are not removed, the dynamic program obviously constructs tuples for all possible structured
solutions with a lower bound, especially the optimum combinations for every 0 ≤ v ≤ c. Removing
dominated tuples does not affect the tuples that stand for the optimum item combinations so that
the second property still holds.
While the dynamic program above constructs the desired tuples, their number may increase
dramatically until F (0) is obtained. We therefore use approximate dynamic programming for the
tuples with profits in [ 1
4 P0, 2P0]. This method is inspired by the dynamic programming used in [15]
(see also [18, pp. 97 -- 112]).
Definition 18 and Lemma 22 state that a tuple (p, s, k) with p > 0 satisfies p ≥ 2κ−2T.
4 P0, 2P0] =
Apart from (0, 0, k), all tuples have therefore profits in the interval [2κ−2T, 2P0] (1)= [ 1
[2κ−2T, . . . , 2κ+1T]. We partition this interval into sub-intervals of length 2κ−2K. We get
[2κ−2T, 2P0] =
h2κ−2T + ξ · 2κ−2K, 2κ−2T + (ξ + 1)2κ−2K
(cid:17) ∪ {2P0}
ξ0[
ξ0[
ξ=0
ξ=0
=:
(κ−2)
L
ξ
∪ L
(κ−2)
ξ0+1
(κ−2)
ξ
(κ−2)
ξ
for ξ0 := 7(κ + 1)2κ+1 − 1. (A short calculation shows that 2κ−2T + (ξ0 + 1)2κ−2K = 2P0.) The
approximate dynamic program keeps for every ξ ∈ {0, . . . , ξ0 + 1} only the tuple (p, s, k) with
p ∈ L
that has the smallest size s. The dominated tuples are removed when all tuples for k have
been constructed. The modified dynamic program is presented in Algorithm 3 and shown in Figure 3.
The sets of these non-dominated tuples are denoted by D(k). For convenience, (p(ξ), s(ξ), k) ∈ D(k)
denotes the smallest tuple with a profit in L
. We again save the backtracking information
during the execution of the algorithm.
Lemma 23. Let D(k) be the set D(k) from Algorithm 3 before the dominated tuples are removed.
A tuple (p, s, k) ∈ D(k) for k = κ + 1, . . . , 0 stands for a structured solution with a lower bound.
Therefore, we have p ≥ 2κ−2T if p > 0. This is also true for (p, s, k) ∈ D(k).
Proof. The proof is almost identical to the one of Lemma 22. In fact, the proof is not influenced by
keeping only the tuple of smallest size in every profit interval L
Theorem 24. Let k ∈ {0, . . . , κ + 1}. For every (non-dominated) tuple (¯p, ¯s, k) ∈ F (k), there is a
tuple (p, s, k) ∈ D(k) such that
1 − ε
4
!κ−k+1
s ≤ ¯s .
(κ−2)
ξ
.
¯p
and
p ≥
1
log2( 2
ε) + 1
18
Algorithm 3: The approximate dynamic programming
D(κ+2) := {(0, 0, κ + 2)};
Backtrack(0, 0, κ + 2) := ∅;
for k = κ + 1, . . . , 0 do
D(k) := ∅;
for (p(ξ), s(ξ), k + 1) ∈ D(k+1) do
D(k) := D(k) ∪ {(p(ξ), s(ξ), k)};
Backtrack(p(ξ), s(ξ), k) := Backtrack(p(ξ), s(ξ), k + 1);
for a ∈ G(k) do
for (p, s, k + 1) ∈ D(k+1) \ {(0, 0, k + 1)} do
(p0, s0, k) := (p + p(a), s + s(a), k);
(κ−2)
Determine ξ0 for (p0, s0, k) such that p0 ∈ L
ξ0
if s0 < s(ξ0) or (p(ξ0), s(ξ0), k) = ∅ then
// Construction of new tuples
;
// Only new tuples of smaller size are kept
D(k) := D(k) \ {(p(ξ0), s(ξ0), k)};
(p(ξ0), s(ξ0), k) := (p0, s0, k);
Backtrack(p(ξ0), s(ξ0), k) := ((p, s, k + 1), a);
D(k) := D(k) ∪ {(p(ξ0), s(ξ0), k)};
if k ≥ κ − 2 then
// Construction of (possible) tuples (p(a), s(a), k) for k ≥ κ − 2
(κ−2)
Determine ξ0 for p(a) such that p(a) ∈ L
ξ0
if s(a) < s(ξ0) or (p(ξ0), s(ξ0), k) = ∅ then
;
D(k) := D(k) \ {(p(ξ0), s(ξ0), k)};
(p(ξ0), s(ξ0), k) := (p(a), s(a), k);
Backtrack(p(ξ0), s(ξ0), k) := (a);
D(k) := D(k) ∪ {(p(ξ0), s(ξ0), k)};
Remove dominated tuples from D(k);
19
(a) The tuples in D(k+1)
(b) The new tuples are constructed with the items
in G(k).
(c) Only the tuple of smallest size is kept for every
(κ−2)
L
ξ
, which yields D(k),. . .
(d) . . . and removing the dominated tuples yields
D(k).
Figure 3: The principle of the approximate dynamic programming
20
sp2κ−2K(p(ξ),s(ξ),k+1)(p(ξ+1),s(ξ+1),k+1)sp+s(a)+p(a)(p(ξ)+p(a),s(ξ)+s(a))sps2p2s1p1<2κ−2K≥sps2p2(p,s)=(p(ξ0),s(ξ0),k)sp>< Proof. This statement is trivial for (¯p, ¯s, k) = (0, 0, k) because (0, 0, k) ∈ D(k) (this tuple is never
removed in the construction of F (k) and D(k)).
Suppose now that (¯p, ¯s, k) 6= (0, 0, k). The theorem is proved by induction for k = κ + 1, . . . , 0.
The statement is evident for k = κ + 1. If aeff−c exists (i.e. enough copies of aeff can be glued
together without exceeding the capacity c), then we have
F (κ+1) = D(κ+1) = {(0, 0, κ + 1), (p(aeff−c), s(aeff−c), κ + 1)} .
If aeff−c does not exist, then we have F (κ+1) = D(κ+1) = {(0, 0, κ + 1)}.
Suppose that the statement is true for k + 1, . . . , κ + 1. As defined in Lemma 23, D(k) is the set
D(k) before the dominated tuples are removed. Let (¯p, ¯s, k) ∈ F (k).
There are two cases. In the first case, we have (¯p, ¯s, k +1) ∈ F (k+1). By the induction hypothesis,
there is a tuple (p1, s1, k + 1) ∈ D(k+1) such that the inequalities p1 ≥ ¯p(1 − ε
ε )+1)κ−(k+1)+1
4
and s1 ≤ ¯s hold (see Fig. 4(a)). Note that this implies (p1, s1, k + 1) 6= (0, 0, k + 1) and therefore
p1 ≥ 2κ−2T by Lemma 23. Let ξ1 be the index such that p1 ∈ L
. During the execution of
Algorithm 3, (p1, s1, k + 1) yields the tuple (p1, s1, k), which may only be replaced in D(k) by a tuple
. Thus, there must be a tuple (p2, s2, k) ∈ D(k)
of a smaller size, but with a profit still in L
with s2 ≤ s1 and p2 ∈ L
(see Fig. 4(b)). Let now (p, s, k) ∈ D(k) be the tuple that dominates
(p2, s2, k) (which can of course be (p2, s2, k) itself), i.e. p ≥ p2 and s ≤ s2 (see Fig. 4(c)). For the
profit, we have
(κ−2)
ξ1
(κ−2)
ξ1
(κ−2)
ξ1
1
log2( 2
p ≥ p2 ≥ p1 − 2κ−2K
1 − ε
4
p16=0= p1 ·
!
1
log2( 2
ε) + 1
(1),(2)= p1 ·
1 − 2κ−2K
p1
1 − ε
4
≥ ¯p ·
! Lem. 23≥ p1 ·
1 − 2κ−2K
!κ−k+1
2κ−2T
!
1
log2( 2
ε) + 1
.
The lower bound on the profit is therefore true for (p, s, k). We have s ≤ s2 ≤ s1 ≤ ¯s for the bound
on the size (see also Fig. 4(c)).
Consider now the second case where (¯p, ¯s, k) ∈ F (k), but (¯p, ¯s, k + 1) /∈ F (k+1). Therefore,
(¯p, ¯s, k) is a new (non-dominated) tuple with (¯p, ¯s, k) = (p + p(a), s + s(a), k) for the right item
a ∈ G(k) and tuple (p, s, k + 1) ∈ F (k+1). By the induction hypothesis, there must be a tuple
(p1, s1, k + 1) ∈ D(k+1) such that p1 ≥ p(1 − ε
ε )+1)κ−(k+1)+1 and s1 ≤ s (see Fig. 5(a)). Thus,
4
!κ−(k+1)+1
the following inequality holds:
!κ−(k+1)+1
1
log2( 2
≥ (p(a) + p) ·
1 − ε
4
1
log2( 2
ε) + 1
p1 + p(a) ≥ p(a) + p ·
1 − ε
4
= ¯p ·
1 − ε
4
1
log2( 2
ε) + 1
1
!κ−(k+1)+1
log2( 2
ε) + 1
.
There are two possibilities: either k ≥ κ− 2, i.e. p(a) ≥ 2κ−2T holds, and p1 + p(a) ≥ 2κ−2T directly
follows. Otherwise, we have k ≤ κ − 3. Then, the identity (¯p, ¯s, k) = (p + p(a), s + s(a), k) 6= (0, 0, k)
implies that (p, s, k + 1) 6= (0, 0, k + 1) holds because the tuple (0, 0, k + 1) is not used to form
any new tuple in D(k) and therefore in D(k). This again implies that p1 6= 0 and therefore
p1 + p(a) ≥ p1 ≥ 2κ−2T as seen in Lemma 23.
21
(a) Since (¯p, ¯s, k + 1) ∈ F (k+1), there must be a corre-
sponding tuple (p1, s1, k + 1) ∈ D(k+1) by the induc-
tion hypothesis whose profit can be bounded from
below.
(b) By construction, there must be a tuple (p2, s2, k) ∈ D(k)
with a profit in the same interval L
as (p1, s1, k). This
makes it possible to bound p2 from below.
(κ−2)
ξ1
(c) There may be a tuple (p, s, k) ∈ D(k) that domi-
nates (p2, s2, k). Since p ≥ p2, the bound still holds.
Figure 4: The first case of the proof for Theorem 24: we have (¯p, ¯s, k) ∈ F (k) and also (¯p, ¯s, k + 1) ∈
F (k+1). We set Q := (1 − ε
4
1
log2( 2
ε )+1).
22
sp(¯p,¯s,k+1)=(¯p,¯s,k)¯p¯s(p1,s1,k+1)=(p1,s1,k)p1s1≤p1≥(1−Q)κ−(k+1)+1¯psp(¯p,¯s,k)¯p¯s(p1,s1,k)s1(p2,s2,k)p2s2≤≤L(κ−2)ξ1p2≥(1−Q)κ−k+1¯psp(¯p,¯s,k)¯s¯p(p1,s1,k)s1(p2,s2,k)s2(p,s,k)sp≤≤≤p≥p2≥(1−Q)κ−k+1¯p (κ−2)
ξ1
Thus, there is an index ξ1 such that p1 +p(a) ∈ L
. Similar to above, the tuple (p1 +p(a), s1 +
s(a), k) is formed during the construction of D(k) (see Fig. 5(b)). It may only be replaced by a
tuple of smaller size. Hence, there must be (p2, s2, k) ∈ D(k) with p2 ∈ L
. Let (p, s, k) ∈ D(k)
be the tuple that dominates (p2, s2, k) (see Fig. 5(c)). We get
1 − 2κ−2K
p1 + p(a)
1
1 − ε
log2( 2
4
ε) + 1
p1+p(a)6=0=
! (1),(2)= (p1 + p(a)) ·
!κ−k+1
(p1 + p(a)) ·
p1+p(a)≥2κ−2T
(κ−2)
ξ1
!
!
p
p2 ≥ p1 + p(a) − 2κ−2K
1 − 2κ−2K
(p1 + p(a)) ·
2κ−2T
1
1 − ε
log2( 2
4
ε) + 1
¯p ·
≥
≥
≥
.
We have similar to above s ≤ s2 ≤ s1 + s(a) ≤ s + s(a) = ¯s for the bound on the size (see also Fig.
5(c)).
Remark 25. As can be seen, the proof of Theorem 24 is only possible because it is guaranteed
that p1 or p1 + p(a) is at least 2κ−2T. In fact, this is achieved by the construction of the glued
item set G with its structured solution (Theorem 14). Hence, we can prove Lemma 16, and with
the introduction of aeff−c, we have the structure property of Definition 18 with a corresponding
solution (Theorem 19). This shows that p1 ≥ 2κ−2T or p1 + p(a) ≥ 2κ−2T (see also Lemma 22 and
23). Without the structure, a dynamic program like Algorithm 3 would also have to generate tuples
(p, s, k) with p < 2κ−2T for k ≤ κ − 3. Hence, we would need for the same approximation ratio
with a smaller length than 2κ−2K, and we would have to save more
profit sub-intervals like L
tuples. Both would increase the asymptotic running time and space complexity as can be seen in
the proof of Theorem 27.
Corollary 26. For every v ≤ c, there is a tuple (p, s, 0) ∈ D(0) such that s ≤ v and
(κ−2)
ξ
1 − ε
4
p ≥
1
log2( 2
ε) + 1
!κ+1
OPTSt (G ∪ {aeff−c} , v) .
ε log2 1
ε).
Proof. Lemma 22 states that there is a tuple (¯p, ¯s, 0) ∈ F (0) with ¯p = OPTSt(G ∪ {aeff−c} , v) and
¯s ≤ v. Theorem 24 implies that there is a tuple (p, s, 0) ∈ D(0) with the desired property.
ε2 log3 1
Theorem 27. Algorithm 3 constructs all tuple sets D(k) for k = κ + 1, . . . , 0 in time O( 1
ε).
The space needed for the algorithm and to save the D(k) as well as the backtracking information is
in O( 1
Proof. Let us first bound the space complexity. The profit interval [ 1
4 P0, 2P0] is partitioned into
O(ξ0) intervals L
. The set D(k) saves at most one tuple with the corresponding backtracking
information for every L
or the information that a tuple does not exist. Thus, the space
needed for all D(k) and the corresponding backtracking data is in O(κ · ξ0) = O(κ · (κ2κ)) =
ε)· (log( 1
O(log( 1
ε) 1
ε). All other information of the algorithm is only temporarily saved
and needs O (1).
ε)) = O( 1
ε log2 1
(κ−2)
ξ
(κ−2)
ξ
23
(a) Since (¯p, ¯s, k +1) /∈ F (k+1), there must be an item
a such that (¯p, ¯s, k) = (p+p(a), s+s(a), k) for a tuple
(p, s, k + 1) ∈ F (k+1). By the induction hypothesis,
there must be a tuple (p1, s1, k + 1) ∈ D(k+1) whose
profit can be bounded from below.
(b) The tuple (p1 + p(a), s1 + s(a), k) is constructed
during the execution of the dynamic program.
(c) As in the first case, there must be a tuple (p, s, k) ∈ D(k)
whose profit can be bounded as desired. Here, (p2, s2, k)
is not dominated, i.e. (p, s, k) = (p2, s2, k).
Figure 5: The second case of the proof for Theorem 24: we have (¯p, ¯s, k) ∈ F (k), but (¯p, ¯s, k + 1) /∈
F (k+1). We set Q := (1 − ε
4
1
log2( 2
ε )+1).
24
sp(¯p,¯s,k)=(p+p(a),s+s(a),k)(p,s,k+1)+s(a)+p(a)p(p1,s1,k+1)p1s1s≤p1≥(1−Q)κ−(k+1)+1psp(¯p,¯s,k)=(p+p(a),s+s(a),k)s+s(a)(p1,s1,k+1)+s(a)+p(a)(p1+p(a),s1+s(a),k)s1+s(a)≤L(κ−2)ξ1s(¯p,¯s,k)=(p+p(a),s+s(a),k)s+s(a)s1+s(a)≤(p2,s2,k)=(p,s,k)s≤s2≤¯ppp≥p2≥(1−Q)κ−k+1¯pL(κ−2)ξ1 The loops dominate the running time. Apart from removing the dominated tuples, they need in
total
(cid:16)
κ ·(cid:16)
ξ0 + ξ0 · G(k) + G(k)(cid:17)(cid:17) Thm. 20= O
O
(cid:18)1
(cid:19)
ε
(cid:18)
(cid:18)1
(cid:19)(cid:18)1
(cid:18) 1
(cid:19)
log
ε2 log3 1
ε
ε
ε
log
.
(cid:18)1
(cid:19)(cid:19)(cid:19)
· 1
ε
log
ε
= O
As stated in [21] and [11, Lemma 5], non-dominated tuples (p, s, k) can be removed in linear time in
the number of tuples if the tuples are different and sorted by profit. This is the case because every
tuple in D(k) is stored in an array sorted according to the corresponding ξ. The total time to remove
the dominated tuples from all D(k) is therefore in O(κ · ξ0) = O( 1
ε), which is dominated by
the overall running time.
ε log2 1
6 The Algorithm
We can now put together the entire approximation algorithm.
Algorithm 4: The complete algorithm
Input: Item set I
Output: Profit P, solution set J
Determine P0 and define T, K;
Partition the items into IL and IS and find aeff;
if Item a with p(a) = 2P0 found during the partitioning then
return 2P0, {a};
Reduce IL to IL,red with Algorithm 1;
Construct G with Algorithm 2 and the item aeff−c;
if p(a
0 ) ≤ c
(κ)
(κ)
0 ) = P0 and s(a
2 then
(κ)
Recursively undo the gluing of a
0
two copies of every item in J0;
return 2P0, J ;
to get the item set J0. Let J be the set consisting of
Construct with Algorithm 3 the tuple sets D(κ+1), . . . , D(0);
Find (p, s, 0) ∈ D(0) such that
P := p + OPT ({aeff}, c − s) = max(p0,s0,0)∈D(0) p0 + OPT ({aeff}, c − s0);
Backtrack the tuple (p, s, 0) to find the corresponding structured solution with a lower bound
J0 ⊂ G ∪ {aeff−c};
Recursively undo the gluing of all a ∈ J0 and add these items to the solution set J;
Add the items of OPT({aeff}, c − s) to J;
return P, J ;
Theorem 28. Algorithm 4 finds a solution of value at least (1 − ε)OPT(I).
Proof. The algorithm returns a feasible solution: (p, s, 0) represents an item set of size s. If items
a ∈ G derived from gluing are part of the solution, their ungluing does not change the total size nor
the total profit (see Remark 12).
25
We prove the solution quality. First, the algorithm considers the two special cases listed at the
beginning of Section 5. Each of them returns a solution of profit 2P0 so that OPT(I) = 2P0 (see
Theorem 5 and Lemma 15). If the special cases do not yield a solution, we are in the third case.
Let v be the volume from Theorem 19. Corollary 26 guarantees the existence of one (p, s, 0) ∈ D(0)
with s ≤ v such that
1 − ε
4
p ≥
1
log2( 2
ε) + 1
!κ+1
OPTSt (G ∪ {aeff−c} , v) .
Moreover, we have OPT({aeff}, c− s) ≥ OPT({aeff}, c− v) because c− s ≥ c− v. Thus, the following
inequality holds for this (p, s, 0):
p + OPT ({aeff}, c − s)
≥
≥
Thm. 19≥
≥
(1)≥
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
1
log2( 2
ε) + 1
!κ+1
1 − ε
!κ+1
4
1 − ε
!2κ+2
4
1 − ε
!2κ+2
4
1 − ε
!
4
1 − ε
ε) + 1(cid:17)
4 · 2 ·(cid:16)log2( 2
1 − ε
4
ε) + 1
log2( 2
2κ + 2
≥
log2( 2
= (1 − ε) OPT(I) .
ε) + 1
OPTSt (G ∪ {aeff−c} , v) + OPT ({aeff}, c − v)
(OPTSt (G ∪ {aeff−c} , v) + OPT ({aeff}, c − v))
!κ+1
T
1 − ε
4
1
log2( 2
ε) + 1
OPT(I) −
OPT(I) − T
OPT(I) − 1
2 εP0
OPT(I) − 1
2 εOPT(I)
1
log2( 2
Taking the maximum over all (p, s, 0) ∈ D(0) therefore yields the desired solution. Note that we
have used (1 − δ)k ≥ (1 − k · δ) for δ < 1.
Remark 29. The total bound on the approximation ratio is mainly due to the exponent 2κ + 2, i.e.
that we make the multiplicative error of (1 − ε
ε )+1) only 2κ + 2 times. Such an error occurs
4
when I is replaced by IL,red at the beginning (Lemma 8), in each of the κ iterations in which G is
constructed (Theorem 14), and in κ + 1 of the κ + 2 iterations of the dynamic program (Theorem
24 and Corollary 26). The error of the dynamic program can be bounded because the structured
solution with a lower bound has at least one item of profit at least 2κ−2T (see the second property
of Definition 18 and Remark 25).
Theorem 30. The algorithm has a running time in O(n + 1
ε log2 1
1
ε).
Proof. Determining P0, constructing IL and IS as well as finding aeff can all be done in time and
space O (n) as stated in Theorems 5 and 6. The definition of T and K in time and space O (1) is
ε) and needs space in O(n +
ε2 log3 1
26
obvious. It is also clear that an item p(a) = 2P0 can directly be found during the construction of IL
such that the first if-condition does not influence the asymptotic running time.
9).
Algorithm 1 returns the set IL,red in time O(n + 1
ε) (see Theorem
ε) as explained
The second if-condition can be checked in O (1). The running time for undoing the gluing will
in Theorem 20, which clearly dominates the construction of aeff−c in O (1).
Algorithm 2 constructs the G(k) and G in time O( 1
ε) and space O( 1
ε) and space O( 1
ε log2 1
ε2 log3 1
ε log2 1
ε log2 1
ε log2 1
ε2 log3 1
ε) and space O( 1
ε) items in G ∪ {aeff−c}, which bounds the storage space needed.
be determined at the end of the proof.
Algorithm 3 constructs the sets D(k) in time O( 1
ε) (see Theorem 27).
For one tuple (p0, s0, 0), the corresponding OPT ({aeff}, c − s0) can be found in O (1) by computing
b c−s0
s(aeff)c· p(aeff). Thus, finding the best tuple (p, s, 0) can be done in O(D(0)) = O(ξ0) = O( 1
ε log 1
ε).
Since only the currently best tuple (p, s, 0) has to be saved, the space needed is in O (1).
The backtracking for the tuple (p, s, 0) needs time in O(κ) = O(log 1
ε): the backtracking
information Backtrack(p0, s0, k) for k = 0, . . . , κ + 1 states whether the tuple was formed by adding
an item a ∈ G(k) and with which tuple (p00, s00, k + 1) to continue. Hence, the item set J0 also has at
most O(log 1
To conclude, the time and space for the ungluing still have to be bounded. Consider one item a ∈
G. The backtracking information Backtrack(a) returns two items (¯a1, ¯a2) (with ¯a1, ¯a2 ∈ IL,red ∪ G)
on which the backtracking can be recursively applied. The recursive ungluing of the items can be
represented as a binary tree where the root is the original item a and the (two) children of each
node are the items (¯a0, ¯a00) returned by the backtracking information. The leaves of the tree are the
original items in IL,red. This binary tree obviously has a height in O(κ) because the children (¯a0, ¯a00)
for one ¯a ∈ G(k) are in G(k−1) ∪ IL,red. A binary tree of height O(κ) = O(log 1
ε) has at most O( 1
ε)
nodes. The backtracking or ungluing of a can therefore be done in time and space O( 1
ε), which
also includes saving the items ¯a ∈ IL,red of which a is composed. Since J0 has O(log 1
ε) items, the
ε log 1
original items IL,red of the approximate solution can be found in time and space O( 1
ε). This
(κ)
should the body of the second if-condition be
also dominates the time to undo the gluing of a
0
executed.
Similar to above, the number of items aeff for OPT ({aeff}, c − s) can be found in O (1). To sum
up, Algorithm 4 has the stated running time and space complexity.
7 Concluding Remarks
The most important steps in this algorithm are the creation of the item set G by gluing and the
introduction of aeff−c. This guarantees the existence of an approximate structured solution with a
lower bound (see Definition 18). Therefore, the approximate dynamic program has to store less
tuples (p, s, k) than in the case without the structure.
We [20] have extended our algorithm to the Unbounded Knapsack Profit with Inversely Propor-
tional Profits (UKPIP) introduced in [11]. Here, several knapsack sizes 0 < c1 < . . . < cM = 1 are
given, and the profit of an item counts as pj/cl if packed in cl. The goal is to find the best knapsack
size and corresponding solution of maximum profit. UKPIP is used for column generation in our
AFPTAS for Variable-Sized Bin Packing [10] where several bin sizes are given and the goal is to
minimize the total volume of the bins used. The faster FPTAS for UKPIP yields a faster AFPTAS
for Variable-Sized Bin Packing [20].
27
There are interesting open questions. As stated in Subsection 1.2, the space complexity is a
more serious bottleneck than the running time. Recently, Lokshtanov and Nederlof [22] showed
that the 0-1 Knapsack Problem and the Subset Sum Problem have a pseudo-polynomial time
and only polynomial space algorithm. Subset Sum is a special case of the Knapsack Problem
where the profit of an item is equal to its size, i.e. pj = sj. Moreover, it was shown that Unary
Subset Sum is in Logspace [3, 13]. Gál et al. [4] described an FPTAS for Subset Sum whose space
complexity is in O( 1
ε), i.e. which does not depend on the actual input size, and whose running time
is in O( 1
ε)). Can any of these results be further extended to improve the space
complexity of an UKP FPTAS?
Finally, it is open whether the ideas presented in this paper can be extended to the normal 0-1
KP or other KP variants as well as used for column generation of other optimization problems. The
currently fastest known algorithm for 0-1 KP is due to Kellerer and Pferschy [16 -- 18]. We mention
in closing that by using the same approach similar improved approximation algorithms can be
expected for various Packing and Scheduling Problems, e.g. for Bin Covering, Bin Packing with
Cardinality Constraints, Scheduling Multiprocessor Tasks and Resource-constrained Scheduling.
ε n(n + log n + log 1
References
[1] R. E. Bellman. Dynamic Programming. Princeton University Press, 1957.
[2] M. Bougeret, P.-F. Dutot, K. Jansen, C. Robenek, and D. Trystram. "Approximation Algo-
rithms for Multiple Strip Packing and Scheduling Parallel Jobs in Platforms". In: Discrete
Mathematics, Algorithms and Applications 3.4 (2011), pp. 553 -- 586.
[3] M. Elberfeld, A. Jakoby, and T. Tantau. Logspace Versions of the Theorems of Bodlaender and
Courcelle. Tech. rep. 62. Electronic Colloquium on Computational Complexity (ECCC), 2010.
First published in: Proceedings of the 51th Annual Symposium on Foundations of Computer
Science (FOCS 2010). IEEE Computer Society, 2010, pp. 143 -- 152.
[4] A. Gál, J. Jang, N. Limaye, M. Mahajan, and K. Sreenivasaiah. Space-Efficient Approximations
for Subset Sum. Tech. rep. 180. Electronic Colloquium on Computational Complexity (ECCC),
2014.
[5] M. R. Garey and D. S. Johnson. Computers and Intractability. A Guide to the Theory of
NP-Completeness. W. H. Freeman and Company, 1979.
[6] P. C. Gilmore and R. E. Gomory. "A Linear Programming Approach to the Cutting-Stock
Problem". In: Operations Research 9.6 (1961), pp. 849 -- 859.
[7] M. D. Grigoriadis, L. G. Khachiyan, L. Porkolab, and J. Villavicencio. "Approximate Max-Min
Resource Sharing for Structured Concave Optimization". In: SIAM Journal on Optimization
11.4 (2001), pp. 1081 -- 1091.
[8] O. H. Ibarra and C. E. Kim. "Fast Approximation Algorithms for the Knapsack and Sum of
Subset Problems". In: Journal of the ACM 22.4 (1975), pp. 463 -- 468.
[9] K. Jansen. "Approximation Algorithms for Min-Max and Max-Min Resource Sharing Problems,
and Applications". In: Efficient approximation and online algorithms. Vol. 3484. LNCS.
Springer, 2006, pp. 156 -- 202.
28
[10] K. Jansen and S. Kraft. "An Improved Approximation Scheme for Variable-Sized Bin Packing".
In: Proceedings of the 37th International Symposium on Mathematical Foundations of Computer
Science, MFCS 2012. Vol. 7464. LNCS. Springer, 2012, pp. 529 -- 541.
[11] K. Jansen and S. Kraft. "An Improved Knapsack Solver for Column Generation". In: Proceed-
ings of the 8th International Computer Science Symposium in Russia, CSR 2013. Vol. 7913.
LNCS. Springer, 2013, pp. 12 -- 23.
[12] K. Jansen and S. E. J. Kraft. "A Faster FPTAS for the Unbounded Knapsack Problem". In:
Proceedings of the 26th International Workshop on Combinatorial Algorithms, IWOCA 2015.
LNCS. Springer, 2015/2016. In press.
[13] D. M. Kane. Unary Subset-Sum is in Logspace. 2010. arXiv:1012.1336.
[14] N. Karmarkar and R. M. Karp. "An Efficient Approximation Scheme for the One-Dimensional
Bin-Packing Problem". In: Proceedings of the 23rd Annual Symposium on Foundations of
Computer Science (FOCS 1982). IEEE Computer Society, 1982, pp. 312 -- 320.
[15] H. Kellerer, R. Mansini, U. Pferschy, and M. G. Speranza. "An efficient fully polynomial
approximation scheme for the Subset-Sum Problem". In: Journal of Computer and System
Sciences 66.2 (2003), pp. 349 -- 370.
[16] H. Kellerer and U. Pferschy. "A New Fully Polynomial Time Approximation Scheme for the
Knapsack Problem". In: Journal of Combinatorial Optimization 3.1 (1999), pp. 59 -- 71.
[17] H. Kellerer and U. Pferschy. "Improved Dynamic Programming in Connection with an FTPAS
for the Knapsack Problem". In: Journal of Combinatorial Optimization 8.1 (2004), pp. 5 -- 11.
[18] H. Kellerer, U. Pferschy, and D. Pisinger. Knapsack Problems. Springer, 2004.
[19] C. Kenyon and E. Rémila. "A Near-Optimal Solution to a Two-Dimensional Cutting Stock
Problem". In: Mathematics of Operations Research 25.4 (2000), pp. 645 -- 656. First published as
"Approximate Strip Packing". In: Proceedings of the 37th Annual Symposium on Foundations
of Computer Science (FOCS 1996). IEEE Computer Society, 1996, pp. 31 -- 36.
[20] S. E. J. Kraft. "Improved Approximation Algorithms for Packing and Scheduling Problems".
PhD thesis. Christian-Albrechts-Universität zu Kiel, 2015. Forthcoming.
[21] E. L. Lawler. "Fast Approximation Algorithms for Knapsack Problems". In: Mathematics of
Operations Research 4.4 (1979), pp. 339 -- 356.
[22] D. Lokshtanov and J. Nederlof. "Saving space by algebraization". In: Proceedings of the 42nd
ACM Symposium on Theory of Computing, STOC 2010. ACM, 2010, pp. 321 -- 330.
[23] M. J. Magazine and O. Oguz. "A fully polynomial approximation algorithm for the 0-1
knapsack problem". In: European Journal of Operational Research 8.3 (Nov. 1981), pp. 270 --
273.
[24] S. A. Plotkin, D. B. Shmoys, and É. Tardos. "Fast Approximation Algorithms for Fractional
Packing and Covering Problems". In: Mathematics of Operations Research 20.2 (1995), pp. 257 --
301. First published in: Proceedings of the 32nd Annual Symposium on Foundations of Computer
Science (FOCS 1991). IEEE Computer Society, 1991, pp. 495 -- 504.
29
[25] H. Shachnai and O. Yehezkely. "Fast Asymptotic FPTAS for Packing Fragmentable Items with
Costs". In: Proceedings of the 16th International Symposium on Fundamentals of Computation
Theory, FCT 2007. Vol. 4639. LNCS. Springer, 2007, pp. 482 -- 493.
[26] M. Sviridenko. "A note on the Kenyon-Remila strip-packing algorithm". In: Information
Processing Letters 112.1 -- 2 (2012), pp. 10 -- 12.
This bibliography contains information from the DBLP database (www.dblp.org), which is
made available under the ODC Attribution License.
30
|
1711.04506 | 1 | 1711 | 2017-11-13T10:37:36 | Constraint Solving via Fractional Edge Covers | [
"cs.DS"
] | Many important combinatorial problems can be modeled as constraint satisfaction problems. Hence identifying polynomial-time solvable classes of constraint satisfaction problems has received a lot of attention. In this paper, we are interested in structural properties that can make the problem tractable. So far, the largest structural class that is known to be polynomial-time solvable is the class of bounded hypertree width instances introduced by Gottlob et al. Here we identify a new class of polynomial-time solvable instances: those having bounded fractional edge cover number.
Combining hypertree width and fractional edge cover number, we then introduce the notion of fractional hypertree width. We prove that constraint satisfaction problems with bounded fractional hypertree width can be solved in polynomial time (provided that a the tree decomposition is given in the input). Together with a recent approximation algorithm for finding such decompositions by Marx, it follows that bounded fractional hypertree width is now the most general known structural property that guarantees polynomial-time solvability. | cs.DS | cs |
Constraint Solving via Fractional Edge Covers1
Martin Grohe, RWTH Aachen University, Lehrstuhl fur Informatik 7, Aachen, Germany,
[email protected]
D´aniel Marx, Computer and Automation Research Institute, Hungarian Academy of Sciences (MTA
SZTAKI), Budapest, Hungary, [email protected]. 2
0
Many important combinatorial problems can be modeled as constraint satisfaction problems. Hence iden-
tifying polynomial-time solvable classes of constraint satisfaction problems has received a lot of attention.
In this paper, we are interested in structural properties that can make the problem tractable. So far, the
largest structural class that is known to be polynomial-time solvable is the class of bounded hypertree width
instances introduced by Gottlob et al. [2002]. Here we identify a new class of polynomial-time solvable in-
stances: those having bounded fractional edge cover number.
Combining hypertree width and fractional edge cover number, we then introduce the notion of fractional
hypertree width. We prove that constraint satisfaction problems with bounded fractional hypertree width
can be solved in polynomial time (provided that a the tree decomposition is given in the input). Together
with a recent approximation algorithm for finding such decompositions [Marx 2010a], it follows that bounded
fractional hypertree width is now the most general known structural property that guarantees polynomial-
time solvability.
Categories and Subject Descriptors: F.2 [Theory of Computing]: Analysis of Algorithms and Problem
Complexity; G.2.2 [Mathematics of Computing]: Discrete Mathematics-Graph Theory
General Terms: Algorithms
Additional Key Words and Phrases: constraint satisfaction, hypergraphs, hypertree width, fractional edge
covers
1. INTRODUCTION
Constraint satisfaction problems form a large class of combinatorial problems that contains
many important "real-world" problems. An instance of a constraint satisfaction problem
consists of a set V of variables, a domain D, and a set C of constraints. For example,
the domain may be {0, 1}, and the constraints may be the clauses of a 3-CNF-formula.
The objective is to assign values in D to the variables in such a way that all constraints
are satisfied. In general, constraint satisfaction problems are NP-hard; considerable efforts,
both practical and theoretical, have been made to identify tractable classes of constraint
satisfaction problems.
On the theoretical side, there are two main directions towards identifying polynomial-
time solvable classes of constraint satisfaction problems. One is to restrict the constraint
language, that is, the type of constraints that are allowed (see, for example, [Bulatov 2006;
Bulatov 2011; Bulatov et al. 2001; Feder and Vardi 1998; Jeavons et al. 1997; Schaefer 1978]).
Formally, the constraint language can be described as a set of relations on the domain. The
other direction is to restrict the structure induced by the constraints on the variables (see,
for example, [Cohen et al. 2008; Dalmau et al. 2002; Dechter and Pearl 1989; Freuder 1990;
Kolaitis and Vardi 1998; Grohe et al. 2001; Grohe 2007]). The present work goes into this
direction; our main contribution is the identification of a natural new class of structurally
tractable constraint satisfaction problems.
The hypergraph of an instance (V, D, C) has V as its vertex set and for every constraint
in C a hyperedge that consists of all variables occurring in the constraint. For a class H of
1An extended abstract of the paper appeared in the Proceedings of the seventeenth annual ACM-SIAM
Symposium on Discrete Algorithms (SODA 2006).
2Research supported by the European Research Council (ERC) grant "PARAMTIGHT: Parameterized
complexity and the search for tight complexity results," reference 280152.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:2
M. Grohe, D. Marx
hypergraphs, we let Csp(H) be the class of all instances whose hypergraph is contained in H.
The central question is for which classes H of hypergraphs the problem Csp(H) is tractable.
Most recently, this question has been studied in [Chen and Dalmau 2005; Cohen et al. 2008;
Marx 2011; Marx 2010b; Gottlob et al. 2005]. It is worth pointing out that the corresponding
question for the graphs (instead of hypergraphs) of instances, in which two variables are
incident if they appear together in a constraint, has been completely answered in [Grohe
2007; Grohe et al. 2001] (under the complexity theoretic assumption FPT (cid:54)= W[1]): For a
class G of graphs, the corresponding problem Csp(G) is in polynomial time if and only if G
has bounded tree width. This can be generalized to Csp(H) for classes H of hypergraphs of
bounded hyperedge size (that is, classes H for which max{e ∃ H = (V, E) ∈ H : e ∈ E}
exists). It follows easily from the results of [Grohe 2007; Grohe et al. 2001] that for all
classes H of bounded hyperedge size,
Csp(H) ∈ PTIME ⇐⇒ H has bounded tree width
(1)
(under the assumption FPT (cid:54)= W[1]).
It is known that (1) does not generalize to arbitrary classes H of hypergraphs (we will
give a very simple counterexample in Section 2). The largest known family of classes of
hypergraphs for which Csp(H) is in PTIME consists of all classes of bounded hypertree
width [Gottlob et al. 2002; Gottlob et al. 2003; Gottlob et al. 2000]. Hypertree width is a
hypergraph invariant that generalizes acyclicity [Berge 1976; Fagin 1983; Yannakakis 1981].
It is a very robust invariant; up to a constant factor it coincides with a number of other
natural invariants that measure the global connectivity of a hypergraph [Adler et al. 2007].
On classes of bounded hyperedge size, bounded hypertree width coincides with bounded tree
width, but in general it does not. It has been asked in [Chen and Dalmau 2005; Cohen et al.
2008; Gottlob et al. 2005; Grohe 2007] whether there are classes H of unbounded hypertree
width such that Csp(H) ∈ PTIME. We give an affirmative answer to this question.
Our key result states that Csp(H) ∈ PTIME for all classes H of bounded fractional
edge cover number. A fractional edge cover of a hypergraph H = (V, E) is a mapping
e∈E x(e) is the
weight of x. The fractional edge cover number ρ∗(H) of H is the minimum of the weights
of all fractional edge covers of H. It follows from standard linear programming results
that this minimum exists and is rational. Furthermore, it is easy to construct classes H
of hypergraphs that have bounded fractional edge cover number and unbounded hypertree
width (see Example 4.2).
e∈E,v∈e x(e) ≥ 1 for all v ∈ V . The number(cid:80)
x : E → [0,∞) such that(cid:80)
We then start a more systematic investigation of the interaction between fractional covers
and hypertree width. We propose a new hypergraph invariant, the fractional hypertree width,
which generalizes both the hypertree width and fractional edge cover number in a natural
way. Fractional hypertree width is an interesting hybrid of the "continuous" fractional edge
cover number and the "discrete" hypertree width. We show that it has properties that are
similar to the nice properties of hypertree width. In particular, we give an approximative
game characterization of fractional hypertree width similar to the characterization of tree
width by the "cops and robber" game [Seymour and Thomas 1993]. Furthermore, we prove
that for classes H of bounded fractional hypertree width, the problem Csp(H) can be solved
in polynomial time provided that a fractional hypertree decomposition of the underlying
hypergraph is given together with the input instance. We do not know if for every fixed
k there is a polynomial-time algorithm for finding a fractional hypertree decomposition of
width k. However, a recent result [Marx 2010a] shows that we can find an approximate
decomposition whose width is bounded by a (cubic) function of the fractional hypertree
width. This is sufficient to show that Csp(H) is polynomial-time solvable for classes H of
bounded fractional hypertree width, even if no decomposition is given in the input. There-
fore, bounded fractional hypertree width is the so far most general hypergraph property that
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:3
Fig. 1. Hypergraph properties that make CSP polynomial-time solvable.
makes Csp(H) polynomial-time solvable. Note that this property is strictly more general
than bounded hypertree width and bounded fractional edge cover number (see Figure 1).
For classes H of hypergraphs of bounded fractional hypertree width, we also show that
all solutions of an instance of Csp(H) can be computed by a polynomial delay algorithm.
Closely related to the problem of computing all solutions of a Csp-instance is the problem
of evaluating a conjunctive query against a relational database. We show that conjunctive
queries whose underlying hypergraph is in H can be evaluated by a polynomial delay algo-
rithm as well. Finally, we look at the homomorphism problem and the embedding problem
for relational structures. The homomorphism problem is known to be equivalent to the CSP
[Feder and Vardi 1998] and hence can be solved in polynomial time if the left-hand side
structure has an underlying hypergraph in a class H of bounded fractional hypertree width.
This implies that the corresponding embedding problem, parameterized by the size of the
universe of the left-hand side structure, is fixed-parameter tractable. Recall that a problem
is fixed-parameter tractable (FPT) by some parameter k if the problem can be solved in time
f (k) · nO(1) for a computable function f depending only on k. In particular, if a problem is
polynomial-time solvable, then it is fixed-parameter tractable (with any parameter k).
Follow up work. Let us briefly discuss how the ideas presented in the conference version
of this paper [Grohe and Marx 2006] influenced later work. The conference version of this
paper posed as an open question whether for every fixed k there is a polynomial-time
algorithm that, given a hypergraph with fractional hypertree width k, finds a fractional
hypertree decomposition having width k or, at least, having width bounded by a function
of k. This question has been partially resolved by the algorithm of [Marx 2010a] that finds
a fractional hypertree decomposition of width O(k3) if a decomposition of width k exists.
This algorithm makes the results of the present paper stronger, as the polynomial-time
algorithms for CSPs with bounded fractional hypertree width no longer need to assume
that a decomposition is given in the input (see Section 4.3). The problem of finding a
decomposition of width exactly k, if such a decomposition exists, is still open. In the special
case of hypergraphs whose incidence graphs are planar, a constant-factor approximation
of fractional hypertree width can be found by exploiting the fact that for such graphs
fractional hypertree width and the tree width of the incidence graph can differ only by at
most a constant factor [Fomin et al. 2009].
A core combinatorial idea of the present paper is that Shearer's Lemma gives an up-
per bound on the number of solutions of a CSP instance and we use this bound for the
subproblems corresponding to the bags of a fractional hypertree decomposition. This up-
per bound has been subsequently used by Atserias et al. [2008] in the context of database
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
treewidthBoundededgecovernumberBoundedfractionalhypertreewidthBoundedBounded fractional hypertree width 0:4
M. Grohe, D. Marx
queries, where it is additionally shown that a linear programming duality argument implies
the tightness of this bound. We repeat this argument here as Theorem 3.7. An optimal
query evaluation algorithm matching this tight bound was given by Ngo et al. [2012]. The
bound was generalized to the context of conjunctive queries with functional dependencies
by Gottlob et al. [2012].
We show that if H is a class of hypergraphs with bounded fractional hypertree width, then
Csp(H) is fixed-parameter tractable parameterized by the number of variables and, in fact,
polynomial-time solvable (using the algorithm of [Marx 2010a] for finding decompositions).
It is a natural question if there are more general fixed-parameter tractable or polynomial-
time solvable classes of hypergraphs. Very recently, Marx [2010b] gave a strictly more general
such class by introducing the notion of submodular width and showing that Csp(H) is fixed-
parameter tractable for classes H with bounded submodular width. Furthermore, it was
shown in [Marx 2010b] that there are no classes H with unbounded submodular width that
make Csp(H) fixed-parameter tractable (the proof uses a complexity-theoretic assumption
called Exponential Time Hypothesis [Impagliazzo et al. 2001]). However, with respect to
polynomial-time solvability, bounded fractional hypertree width is still the most general
known tractability condition and it is an open question whether there are classes H with
unbounded fractional hypertree width such that Csp(H) is polynomial-time solvable.
2. PRELIMINARIES
2.1. Hypergraphs
A hypergraph is a pair H = (V (H), E(H)), consisting of a set V (H) of vertices and a
set E(H) of nonempty subsets of V (H), the hyperedges of H. We always assume that
hypergraphs have no isolated vertices, that is, for every v ∈ V (H) there exists at least one
e ∈ E(H) such that v ∈ e.
For a hypergraph H and a set X ⊆ V (H), the subhypergraph of H induced by X is the
hypergraph H[X] = (X,{e∩ X e ∈ E(H) with e∩ X (cid:54)= ∅}). We let H \ X = H[V (H)\ X].
The primal graph of a hypergraph H is the graph
H = (V (H),{{v, w} v (cid:54)= w, there exists an
e ∈ E(H) such that {v, w} ⊆ e}).
A hypergraph H is connected if H is connected. A set C ⊆ V (H) is connected (in H) if the
induced subhypergraph H[C] is connected, and a connected component of H is a maximal
connected subset of V (H). A sequence of vertices of H is a path of H if it is a path of H.
A tree decomposition of a hypergraph H is a tuple (T, (Bt)t∈V (T )), where T is a tree and
(Bt)t∈V (T ) a family of subsets of V (H) such that for each e ∈ E(H) there is a node t ∈ V (T )
such that e ⊆ Bt, and for each v ∈ V (H) the set {t ∈ V (T ) v ∈ Bt} is connected in T .
The sets Bt are called the bags of the decomposition. The width of a tree-decomposition
(T, (Bt)t∈V (T )) is max(cid:8)Bt(cid:12)(cid:12) t ∈ V (t)} − 1. The tree width tw(H) of a hypergraph H is the
minimum of the widths of all tree-decompositions of H. It is easy to see that tw(H) = tw(H)
for all H.
It will be convenient for us to view the trees in tree-decompositions as being rooted and
directed from the root to the leaves. For a node t in a (rooted) tree T = (V (T ), E(T )), we
let Tt be the subtree rooted at t, that is, the induced subtree of T whose vertex set is the
set of all vertices reachable from t.
We say that a class H of hypergraphs is of bounded tree width if there is a k such that
tw(H) ≤ k for all H ∈ H. We use a similar terminology for other hypergraph invariants.
2.2. Constraint satisfaction problems
A CSP instance is a triple I = (V, D, C), where V is a set of variables, D is a set called the
domain, and C is a set of constraints of the form (cid:104)(v1, . . . , vk), R(cid:105), where k ≥ 1 and R is a
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:5
k-ary relation on D. A solution to the instance I is an assignment α : V → D such that for
all constraints (cid:104)(v1, . . . , vk), R(cid:105) in C we have (α(v1), . . . , α(vk)) ∈ R.
instance I = (V, D, C) is the number (cid:107)I(cid:107) = V + D +(cid:80)
Constraints are specified by explicitly enumerating all possible combinations of values
for the variables, that is, all tuples in the relation R. Consequently, we define the size of
a constraint c = (cid:104)(v1, . . . , vk), R(cid:105) ∈ C to be the number (cid:107)c(cid:107) = k + k · R. The size of an
c∈C (cid:107)c(cid:107). Of course, there is no
need to store a constraint relation repeatedly if it occurs in several constraints, but this
only changes the size by a polynomial factor.
Let us make a few remarks about this explicit representation of the constraints. There are
important special cases of constraint satisfaction problems where the constraints are stored
implicitly, which may make the representation exponentially more succinct. Examples are
Boolean satisfiability, where the constraint relations are given implicitly by the clauses of
a formula in conjunctive normal form, or systems of arithmetic (in)equalities, where the
constraints are given implicitly by the (in)equalities. However, our representation is the
standard "generic" representation of constraint satisfaction problems in artificial intelli-
gence (see, for example, [Dechter 2003]). An important application where the constraints
are always given in explicit form is the conjunctive query containment problem, which plays
a crucial role in database query optimization. Kolaitis and Vardi [Kolaitis and Vardi 1998]
observed that it can be represented as a constraint satisfaction problem, and the constraint
relations are given explicitly as part of one of the input queries. A related problem from
database systems is the problem of evaluating conjunctive queries (cf. Theorem 4.14). Here
the constraint relations represent the tables of a relational database, and again they are
given in explicit form. The problem of characterizing the tractable structural restrictions of
CSP has also been studied for other representations of the instances: one can consider more
succinct representations such as disjunctive formulas or decision diagrams [Chen and Grohe
2010] or less succinct representations such as truth tables [Marx 2011]. As the choice of rep-
resentation influences the size of the input and the running time is expressed as a function
of the input size, the choice of representation influences the complexity of the problem and
the exact tractability criterion.
Observe that there is a polynomial-time algorithm deciding whether a given assignment
for an instance is a solution.
The hypergraph of the CSP instance I = (V, D, C) is the hypergraph HI with vertex set
V and a hyperedge {v1, . . . , vk} for all constraints (cid:104)(v1, . . . , vk), R(cid:105) in C. For every class H,
we consider the following decision problem:
Csp(H)
Instance: A CSP instance I with HI ∈ H.
Problem: Decide if I has a solution.
If the class H is not polynomial-time decidable, we view this as a promise problem, that is,
we assume that we are only given instances I with HI ∈ H, and we are only interested in
algorithms that work correctly and efficiently on such instances.
width such that Csp(H) is tractable.
We close this section with a simple example of a class of hypergraphs of unbounded tree
Example 2.1. Let H be that class of all hypergraphs H that have a hyperedge that
contains all vertices, that is, V (H) ∈ E(H). Clearly, H has unbounded tree width, because
the hypergraph (V,{V }) has tree width V − 1. We claim that Csp(H) ∈ PTIME.
To see this, let I = (V, D, C) be an instance of Csp(H). Let (cid:104)(v1, . . . , vk), R(cid:105) be a con-
straint in C with {v1, . . . , vk} = V . Such a constraint exists because HI ∈ H. Each tuple
¯d = (d1, . . . , dk) ∈ R completely specifies an assignment α ¯d defined by α ¯d(vi) = di for
1 ≤ i ≤ k. If for some i, j we have vi = vj, but di (cid:54)= dj, we leave α ¯d undefined.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:6
M. Grohe, D. Marx
Observe that I is satisfiable if and only if there is a tuple ¯d ∈ R such that α ¯d is (well-
defined and) a solution for I. As R ≤ (cid:107)I(cid:107), this can be checked in polynomial time.
3. A POLYNOMIAL-TIME ALGORITHM FOR CSPS WITH BOUNDED FRACTIONAL
COVER NUMBER
In this section we prove that if the hypergraph HI of a CSP instance I has fractional edge
cover number ρ∗(HI ), then it can be decided in (cid:107)I(cid:107)ρ∗(HI )+O(1) time whether I has a solution.
Thus if H is a class of hypergraphs with bounded fractional edge cover number (that is,
there is a constant r such that ρ∗(H) ≤ r for every H ∈ H), then Csp(H) ∈ PTIME.
Actually, we prove a stronger result: A CSP instance I has at most (cid:107)I(cid:107)ρ∗(HI ) solutions and
all the solutions can be enumerated in time (cid:107)I(cid:107)ρ∗(HI )+O(1).
The proof relies on a combinatorial lemma known as Shearer's Lemma. We use Shearer's
Lemma to bound the number of solutions of a CSP instance; our argument resembles an
argument that Friedgut and Kahn [Friedgut and Kahn 1998] used to bound the number of
subhypergraphs of a certain isomorphism type in a hypergraph. The second author applied
similar ideas in a completely different algorithmic context [Marx 2008].
The entropy of a random variable X with range U is
h[X] := −(cid:88)
x∈U
Pr(X = x)log Pr(X = x)
Shearer's lemma gives an upper bound of a distribution on a product space in terms of its
marginal distributions.
Lemma 3.1 (Shearer's Lemma [Chung et al. 1986]). Let X = (Xi i ∈ I) be a
random variable, and let Aj, for j ∈ [m], be (not necessarily distinct) subsets of the index
set I such that each i ∈ I appears in at least q of the sets Aj. For every B ⊆ I, let
XB = (Xi i ∈ B). Then
m(cid:88)
h[XAj ] ≥ q · h[X].
Lemma 3.1 is easy to see in the special case when q = 1 and {A1, . . . , Ap} is a partition
of V . The proof of the general case in [Chung et al. 1986] is based on the submodularity of
entropy. See also [Rhadakrishnan ] for a simple proof.
j=1
Lemma 3.2.
If I = (V, D, C) is a CSP instance where every constraint relation contains
at most N tuples, then I has at most N ρ∗(HI ) ≤ (cid:107)I(cid:107)ρ∗(HI ) solutions.
Proof. Let x be a fractional edge cover of HI with(cid:80)
Let pe and q be nonnegative integers such that x(e) = pe/q. Let m =(cid:80)
e∈E(HI ) x(e) = ρ∗(HI ); it follows
from the standard results of linear programming that such an x exists with rational values.
e∈E(HI ) pe, and let
A1, . . . , Am be a sequence of subsets of V that contains precisely pe copies of the set e, for
all e ∈ E(HI ). Then every variable v ∈ V is contained in at least
x(e) ≥ q
pe = q · (cid:88)
(cid:88)
e∈E(HI ):v∈e
e∈E(HI ):v∈e
of the sets Ai (as x is a fractional edge cover). Let X = (Xv v ∈ V ) be uniformly distributed
on the solutions of I, which we assume to be non-empty as otherwise the claim is obvious.
That is, if we denote by S the number of solutions of I, then we have Pr(X = α) = 1/S
for every solution α of I. Then h[X] = log S. We apply Shearer's Lemma to the random
variable X and the sequence A1, . . . , Am of subsets of V . Assume that Ai corresponds to
some constraint (cid:104)(v(cid:48)
k) is 0
k), R(cid:105). Then the marginal distribution of X on (v(cid:48)
1, . . . , v(cid:48)
1, . . . , v(cid:48)
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:7
on all tuples not in R. Hence the entropy of XAi is is bounded by the entropy of the uniform
distribution on the tuples in R, that is, h[XAi ] ≤ log N . Thus by Shearer's Lemma, we have
(cid:88)
pe · log N ≥ (cid:88)
m(cid:88)
i=1
pe · h[Xe] =
h[XAi] ≥ q · h[X] = q · log S.
e∈E(HI )
It follows that
e∈E(HI )
(cid:80)
e∈E(HI )(pe/q)·log N = 2ρ∗(HI )·log N = N ρ∗(HI ).
S ≤ 2
We would like to turn the upper bound of Lemma 3.2 into an algorithm enumerating all
the solutions, but the proof of Shearer's Lemma is not algorithmic. However, a very simple
algorithm can enumerate the solutions, and Lemma 3.2 can be used to bound the running
time of this algorithm. Starting with a trivial subproblem consisting only of a single variable,
the algorithm enumerates all the solutions for larger and larger subproblems by adding one
variable at a time. To define these subproblems, we need the following definitions:
Definition 3.3. Let R be an r-ary relation over a set D. For 1 ≤ i1 < ··· < i(cid:96) ≤ r,
the projection of R onto the components i1, . . . , i(cid:96) is the relation Ri1,...,i(cid:96) which contains an
(cid:96)-tuple (d(cid:48)
j = dij
for 1 ≤ j ≤ (cid:96).
Intuitively, a tuple is in Ri1,...,i(cid:96) if it can be extended into a tuple in R.
(cid:96)) ∈ D(cid:96) if and only if there is a k-tuple (d1, . . . , dk) ∈ R such that d(cid:48)
1, . . . , d(cid:48)
Definition 3.4. Let I = (V, D, C) be a CSP instance and let V (cid:48) ⊆ V be a nonempty
subset of variables. The CSP instance I[V (cid:48)] induced by V (cid:48) is I(cid:48) = (V (cid:48), D, C(cid:48)), where C(cid:48)
is defined in the following way: for each constraint c = (cid:104)(v1, . . . , vk), R(cid:105) having at least
one variable in V (cid:48), there is a corresponding constraint c(cid:48) in C(cid:48). Suppose that vi1, . . . , vi(cid:96)
are the variables among v1, . . . , vk that are in V (cid:48). Then the constraint c(cid:48) is defined as
(cid:104)(vi1 , . . . , vi(cid:96)), Ri1,...,i(cid:96)(cid:105), that is, the relation is the projection of R onto the components
i1, . . . , i(cid:96).
Thus an assignment α on V (cid:48) satisfies I[V (cid:48)] if for each constraint c of I, there is an
assignment extending α that satisfies c (however, it is not necessarily true that there is an
assignment extending α that satisfies every constraint of I simultaneously). Note that that
the hypergraph of the induced instance I[V (cid:48)] is exactly the induced subhypergraph HI [V (cid:48)].
Theorem 3.5. The solutions of a CSP instance I can be enumerated in time
(cid:107)I(cid:107)ρ∗(HI )+O(1).
Proof. Let V = {v1, . . . , vn} be an arbitrary ordering of the variables of I and let Vi
be the subset {v1, . . . , vi}. For i = 1, 2, . . . , n, the algorithm creates a list Li containing the
solutions of I[Vi]. Since I[Vn] = I, the list Ln is exactly what we want.
For i = 1, the instance I[Vi] has at most D solutions, hence the list Li is easy to construct.
Notice that a solution of I[Vi+1] induces a solution of I[Vi]. Therefore, the list Li+1 can
be constructed by considering the solutions in Li, extending them to the variable vi+1 in
all the D possible ways, and checking whether this assignment is a solution of I[Vi+1].
Clearly, this can be done in Li · D · (cid:107)I[Vi+1](cid:107)O(1) = Li · (cid:107)I(cid:107)O(1) time. By repeating this
procedure for i = 1, 2, . . . , n − 1, the list Ln can be constructed.
i=1 Li · (cid:107)I(cid:107)O(1). Ob-
serve that ρ∗(HI[Vi]) ≤ ρ∗(HI ): HI[Vi] is the subhypergraph of HI induced by Vi, thus
any fractional cover of the hypergraph of I gives a fractional cover of I[Vi] (for every edge
e ∈ E(HI[Vi]), we set the weight of e to be the sum of the weight of the edges e(cid:48) ∈ E(HI )
The total running time of the algorithm can be bounded by (cid:80)n−1
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:8
M. Grohe, D. Marx
with e(cid:48) ∩ Vi = e). Therefore, by Lemma 3.2, Li ≤ (cid:107)I(cid:107)ρ∗(HI ), and it follows that the total
running time is (cid:107)I(cid:107)ρ∗(HI )+O(1).
We note that the algorithm of Theorem 3.5 does not actually need a fractional edge cover:
the fact that the hypergraph has small fractional edge cover number is used only in proving
the time bound of the algorithm. By a significantly more complicated algorithm, Ngo et
al. [Ngo et al. 2012] improved Theorem 3.5 by removing the O(1) term from the exponent.
Corollary 3.6. Let H be a class of hypergraphs of bounded fractional edge cover num-
ber. Then Csp(H) is in polynomial time.
We conclude this section by pointing out that Lemma 3.2 is tight: there are arbitrarily
large instances I where every constraint relation contains at most N tuples and the number
of solutions is exactly N ρ∗(HI ). A similar proof appeared first in [Atserias et al. 2008] in
the context of database queries, but we restate it here in the language of CSPs for the
convenience of the reader.
such that (cid:80)
Theorem 3.7. Let H be a hypergraph. For every N0 ≥ 1, there is a CSP instance
I = (V, D, C) with hypergraph H where every constraint relation contains at most N ≥ N0
tuples and I has at least N ρ∗(H) solutions.
v∈e y(v) ≤ 1 for every e ∈ E(H). The weight of y is (cid:80)
Proof. A fractional independent set of hypergraph H is an assignment y : V (H) → [0, 1]
v∈V (H) y(H). The
fractional independent set number α∗(H) is the maximum weight of a fractional independent
set of H. It is a well-known consequence of linear-programming duality that α∗(H) = ρ∗(H)
for every hypergraph H, since the two values can be expressed by a pair of primal and dual
linear programs [Schrijver 2003, Section 30.10].
Let y be a fractional independent set of weight α∗(H). By standard results of linear
programming, we can assume that y is rational, that is, there is an integer q ≥ 1 such
that for every v ∈ V (H), y(v) = pv/q for some nonnegative integer pv. We define a CSP
0 ] such that for every e ∈ E(H) where
instance I = (V, D, C) with V = V (H) and D = [N q
e = {v1, . . . , vr}, there is a constraint (cid:104)(v1, . . . , vr), Re(cid:105) with
Re = {(a1, . . . , ar) ai ∈ [N pv
0 ] for every 1 ≤ i ≤ r}.
0 . We claim that Re contains at most N tuples. Indeed, the number of tuples in
(cid:80)
q·(cid:80)
(cid:80)
N pv
0 = N
0
v∈e pv
= N
0
v∈e pv/q
= (N q
0 )
v∈e y(v) ≤ N q
0 = N,
Let N = N q
Re is exactly(cid:89)
v∈e
since y is a fractional independent set. Observe that α : V (H) → D is a solution if and only
if α(v) ∈ [N pv
0 ] for every v ∈ V (H). Hence the number of solutions is exactly
N pv
0 )α∗(H) = N α∗(H) = N ρ∗(H),
(cid:89)
0 = N
0
q·(cid:80)
v∈V (H) pv/q
= (N q
v∈V (H) pv
= N
0
(cid:80)
v∈V (H)
as required.
The significance of this result is that it shows that there is no "better" measure than frac-
tional edge cover number that guarantees a polynomial bound on the number of solutions,
in the following formal sense. Let w(H) be a width measure that guarantees a polynomial
bound: that is, if I is a CSP instance where every relation has at most N tuples, then I has
at most N w(H) solutions for some function f . Then by Theorem 3.7, we have ρ∗(H) ≤ w(H).
This means the upper bound on the number of solutions given by w(H) already follows from
the bound given by Lemma 3.2 and hence ρ∗(H) can be considered a stronger measure.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:9
4. FRACTIONAL HYPERTREE DECOMPOSITIONS
Let H be a hypergraph. A generalized hypertree decomposition of H [Gottlob et al. 2002] is
a triple (T, (Bt)t∈V (T ), (Ct)t∈V (T )), where (T, (Bt)t∈V (T )) is a tree decomposition of H and
(Ct)t∈V (T ) is a family of subsets of E(H) such that for every t ∈ V (T ) we have Bt ⊆(cid:83) Ct.
Here (cid:83) Ct denotes the union of the sets (hyperedges) in Ct, that is, the set {v ∈ V (H)
∃e ∈ Ct : v ∈ e}. We call the sets Bt the bags of the decomposition and the sets Ct the
guards. The width of (T, (Bt)t∈V (T ), (Ct)t∈V (T )) is max{Ct t ∈ V (T )}. The generalized
hypertree width ghw(H) of H is the minimum of the widths of the generalized hypertree
decompositions of H. The edge cover number ρ(H) of a hypergraph is the minimum number
of edges needed to cover all vertices; it is easy to see that ρ(H) ≥ ρ∗(H). Observe that the
size of Ct has to be at least ρ(H[Bt]) and, conversely, for a given Bt there is always a
suitable guard Ct of size ρ(H[Bt]). Therefore, ghw(H) ≤ r if there is a tree decomposition
where ρ(H[Bt]) ≤ r for every t ∈ V (T ).
additional special condition: ((cid:83) Ct) ∩(cid:83)
For the sake of completeness, let us mention that a hypertree decomposition of H is a
generalized hypertree decomposition (T, (Bt)t∈V (T ), (Ct)t∈V (T )) that satisfies the following
u∈V (Tt) Bu ⊆ Bt for all t ∈ V (T ). Recall that Tt
denotes the subtree of the T with root t. The hypertree width hw(H) of H is the minimum
of the widths of all hypertree decompositions of H. It has been proved in [Adler et al.
2007] that ghw(H) ≤ hw(H) ≤ 3· ghw(H) + 1. This means that for our purposes, hypertree
width and generalized hypertree width are equivalent. For simplicity, we will only work with
generalized hypertree width.
Observe that for every hypergraph H we have ghw(H) ≤ tw(H) + 1. Furthermore, if H
is a hypergraph with V (H) ∈ E(H) we have ghw(H) = 1 and tw(H) = V (H) − 1.
We now give an approximate characterization of (generalized) hypertree width by a game
that is a variant of the cops and robber game [Seymour and Thomas 1993], which character-
izes tree width: In the robber and marshals game on H [Gottlob et al. 2003], a robber plays
against k marshals. The marshals move on the hyperedges of H, trying to catch the robber.
Intuitively, the marshals occupy all vertices of the hyperedges where they are located. In
each move, some of the marshals fly in helicopters to new hyperedges. The robber moves on
the vertices of H. She sees where the marshals will be landing and quickly tries to escape,
running arbitrarily fast along paths of H, not being allowed to run through a vertex that is
occupied by a marshal before and after the flight (possibly by two different marshals). The
marshals' objective is to land a marshal via helicopter on a hyperedge containing the vertex
occupied by the robber. The robber tries to elude capture. The marshal width mw(H) of a
hypergraph H is the least number k of marshals that have a winning strategy in the robber
and marshals game played on H (see [Adler 2004] or [Gottlob et al. 2003] for a formal
definition).
It is easy to see that mw(H) ≤ ghw(H) for every hypergraph H. To win the game
on a hypertree of generalized hypertree width k, the marshals always occupy guards of a
decomposition and eventually capture the robber at a leaf of the tree. Conversely, it can be
proved that ghw(H) ≤ 3 · mw(H) + 1 [Adler et al. 2007].
Observe that for every hypergraph H, the generalized hypertree width ghw(H) is less than
or equal to the edge cover number ρ(H): hypergraph H has a generalized hypertree decom-
position consisting of a single bag containing all vertices and having a guard of size ρ(H).
On the other hand, the following two examples show that hypertree width and fractional
edge cover number are incomparable.
Example 4.1. Consider the class of all graphs that only have disjoint edges. The tree
width and hypertree width of this class is 1, whereas the fractional edge cover number is
unbounded.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:10
M. Grohe, D. Marx
Example 4.2. For n ≥ 1, let Hn be the following hypergraph: Hn has a vertex vS for
every subset S of {1, . . . , 2n} of cardinality n. Furthermore, for every i ∈ {1, . . . , 2n} the
hypergraph Hn has a hyperedge ei = {vS i ∈ S}.
Observe that the fractional edge cover number ρ∗(Hn) is at most 2, because the mapping
x that assigns 1/n to every hyperedge ei is a fractional edge cover of weight 2. Actually, it
is easy to see that ρ∗(Hn) = 2.
We claim that the hypertree width of Hn is n. We show that Hn has a hypertree de-
composition of width n. Let S1 = {1, . . . , n} and S2 = {n + 1, . . . , 2n}. We construct a
generalized hypertree decomposition for Hn with a tree T having two nodes t1 and t2. For
i = 1, 2, we let Bt1 contain a vertex VS if and only if S ∩ Si (cid:54)= ∅. For each edge ej ∈ E(Hn),
there is a bag of the decomposition that contains ej: if j ∈ Si, then Bti contains every
vertex of ej. We set the guard Cti to contain every ej with j ∈ Si. It is clear that Cti = n
and Cti covers Bti: vertex vS is in Bti only if there is a j ∈ S ∩ Si, in which case ej ∈ Cti
covers vS. Thus this is indeed a generalized hypertree decomposition of width n for Hn and
ghw(Hn) ≤ n follows.
To see that ghw(Hn) > n − 1, we argue that the robber has a winning strategy against
(n − 1) marshals in the robber and marshals game. Consider a position of the game where
the marshals occupy edges ej1, . . . , ejn−1 and the robber occupies a vertex vS for a set S
with S ∩{j1, . . . , jn−1} = ∅. Suppose that in the next round of the game the marshals move
to the edges ek1, . . . , ekn−1. Let i ∈ S \ {k1, . . . , kn−1}. The robber moves along the edge ei
to a vertex vR for a set R ⊆ {1, . . . , 2n} \ {k1, . . . , kn−1} of cardinality n that contains i. If
she plays this way, she can never be captured.
For a hypergraph H and a mapping γ : E(H) → [0,∞), we let
γ(e) ≥ 1}.
B(γ) = {v ∈ V (H) (cid:88)
weight(γ) =(cid:80)
e∈E γ(e).
e∈E(H),v∈e
We may think of B(γ) as the set of all vertices "blocked" by γ. Furthermore, we let
Definition 4.3. Let H be a hypergraph. A fractional hypertree decomposition of H is
a triple (T, (Bt)t∈V (T ), (γt)t∈V (T )), where (T, (Bt)t∈V (T )) is a tree decomposition of H and
(γt)t∈V (T ) is a family of mappings from E(H) to [0,∞) such that for every t ∈ V (T ) we
have Bt ⊆ B(γt).
guards.
We call the sets Bt the bags of the decomposition and the mappings γt the (fractional)
The width of (T, (Bt)t∈V (T ), (γt)t∈V (T )) is max{weight(γt) t ∈ V (T )}. The fractional
hypertree width fhw(H) of H is the minimum of the widths of the fractional hypertree de-
compositions of H. Equivalently, fhw(H) ≤ r if H has a tree decomposition where ρ∗(Bt) ≤ r
for every bag Bt.
It is easy to see that the minimum of the widths of all fractional hypertree decompositions
of a hypergraph H always exists and is rational. This follows from the fact that, up to an
obvious equivalence, there are only finitely many tree decompositions of a hypergraph.
Clearly, for every hypergraph H we have
fhw(H) ≤ ρ∗(H)
and fhw(H) ≤ ghw(H).
Examples 4.1 and 4.2 above show that there are families of hypergraphs of bounded frac-
tional hypertree width, but unbounded fractional edge cover number and unbounded gen-
eralized hypertree width.
It is also worth pointing out that for every hypergraph H,
fhw(H) = 1 ⇐⇒ ghw(H) = 1.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:11
To see this, note that if γ : E(H) → [0,∞) is a mapping with weight(γ) = 1 and B ⊆ B(γ),
then B ⊆ e for all e ∈ E(H) with γ(e) > 0. Thus instead of using γ as a guard in a
fractional hypertree decomposition, we may use the integral guard {e} for any e ∈ E(H)
with γ(e) > 0. Let us remark that ghw(H) = 1 if and only if H is acyclic [Gottlob et al.
2002].
4.1. The robber and army game
As robbers are getting ever more clever, it takes more and more powerful security forces
to capture them. In the robber and army game on a hypergraph H, a robber plays against
a general commanding an army of r battalions of soldiers. The general may distribute his
soldiers arbitrarily on the hyperedges. However, a vertex of the hypergraph is only blocked
if the number of soldiers on all hyperedges that contain this vertex adds up to the strength
of at least one battalion. The game is then played like the robber and marshals game.
Definition 4.4. Let H be a hypergraph and r a nonnegative real. The robber and army
game on H with r battalions (denoted by RA(H, r)) is played by two players, the robber and
the general. A position of the game is a pair (γ, v), where v ∈ V (H) and γ : E(H) → [0,∞)
with weight(γ) ≤ r. To start a game, the robber picks an arbitrary v0, and the initial position
is (0, v0), where 0 denote the constant zero mapping.
In each round, the players move from the current position (γ, v) to a new position (γ(cid:48), v(cid:48))
as follows: The general selects γ(cid:48), and then the robber selects v(cid:48) such that there is a path
If a position (γ, v) with v ∈ B(γ) is reached, the play ends and the general wins. If the
from v to v(cid:48) in the hypergraph H \ (B(γ) ∩ B(γ(cid:48))(cid:1).
play continues forever, the robber wins.
The army width aw(H) of H is the least r such that the general has winning strategy for
the game RA(H, r).
Again, it is easy to see that aw(H) is well-defined and rational (observe that two positions
(γ1, v) and (γ2, v) are equivalent if B(γ1) = B(γ2) holds).
Theorem 4.5. For every hypergraph H,
aw(H) ≤ fhw(H) ≤ 3 · aw(H) + 2.
The rest of this subsection is devoted to a proof of this theorem. The proof is similar
to the proof of the corresponding result for the robber and marshal game and generalized
hypertree width in [Adler et al. 2007], which in turn is based on ideas from [Reed 1997;
Seymour and Thomas 1993].
Let H be a hypergraph and γ, σ : E(H) → [0,∞). For a set W ⊆ V (H), we let
A mapping σ : E(H) → [0,∞) is a balanced separator for γ if for every connected component
R of H \ B(σ),
weight(γR) ≤ weight(γ)
.
2
Lemma 4.6. Let H be a hypergraph with aw(H) ≤ r for some nonnegative real r. Then
every γ : E(H) → [0,∞) has a balanced separator of weight r.
Proof. Suppose for contradiction that γ : E(H) → [0,∞) has no balanced separator
of weight r. We claim that the robber has a winning strategy for the game RA(H, r).
The robber simply maintains the invariant that in every position (σ, v) of the game, v is
contained in the connected component R of H \ B(σ) with weight(γR) > weight(γ)/2.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
weight(γW ) =
γ(e).
(cid:88)
e∈E(H)
e∩W(cid:54)=∅
0:12
M. Grohe, D. Marx
To see that this is possible, let (σ, v) be such a position. Suppose that the general moves
from σ to σ(cid:48), and let R(cid:48) be the connected component of H \ B(σ(cid:48)) with weight(γR(cid:48)) >
weight(γ)/2. Then there must be some e ∈ E(H) such that e ∩ R (cid:54)= ∅ and e ∩ R(cid:48)
(cid:54)= ∅,
because otherwise we had
weight(γ) = weight(γ)/2 + weight(γ)/2 < weight(γR) + weight(γR(cid:48)) ≤ weight(γ),
which is impossible. Thus the robber can move from R to R(cid:48) via the edge e.
Let H be a hypergraph and H(cid:48) an induced subhypergraph of H. Then the restriction of
a mapping γ : E(H) → [0,∞) to H(cid:48) is the mapping γ(cid:48) : E(H(cid:48)) → [0,∞) defined by
(cid:88)
γ(cid:48)(e(cid:48)) =
γ(e).
e∈E(H)
e∩V (H(cid:48))=e(cid:48)
Note that weight(γ(cid:48)) ≤ weight(γ) and B(γ(cid:48)) = B(γ) ∩ V (H(cid:48)). The inequality may be
strict because edges with nonempty weight may have an empty intersection with V (H(cid:48)).
Conversely, the canonical extension of a mapping γ(cid:48) : E(H(cid:48)) → [0,∞) to H is the mapping
γ : E(H) → [0,∞) defined by
γ(e) =
{e1 ∈ E(H) e1 ∩ V (H(cid:48)) = e ∩ V (H)}
γ(cid:48)(e ∩ V (H(cid:48)))
if e∩ V (H(cid:48)) (cid:54)= ∅ and γ(e) = 0 otherwise. Intuitively, for every e(cid:48) ∈ E(H(cid:48)), we distribute the
weight of e(cid:48) equally among all the edges e ∈ E(H) whose intersection with V (H(cid:48)) is exactly
e(cid:48). Note that weight(γ) = weight(γ(cid:48)) and B(γ(cid:48)) = B(γ) ∩ V (H(cid:48)).
Proof (of Theorem 4.5). Let H be a hypergraph. To prove that aw(H) ≤ fhw(H),
let (T, (Bt)t∈V (T ), (γt)t∈V (T )) be a fractional hypertree decomposition of H having width
fhw(H). We claim that the general has a winning strategy for RA(H, r). Let (0, v0) be the
initial position. The general plays in such a way that all subsequent positions are of the
form (γt, v) such that v ∈ Bu for some u ∈ V (Tt). Intuitively, this means that the robber is
trapped in the subtree below t. Furthermore, in each move the general reduces the height
of t. He starts by selecting γt0 for the root t0 of T . Suppose the game is in a position (γt, v)
such that v ∈ Bu for some u ∈ V (Tt). If u = t, then the robber has lost the game. So let
us assume that u (cid:54)= t. Then there is a child t(cid:48) of t such that u ∈ V (Tt(cid:48)). The general moves
to γt(cid:48). Suppose the robber escapes to a v(cid:48) that is not contained in Bu(cid:48) for any u(cid:48) ∈ Tt(cid:48).
Then there is a path from v to v(cid:48) in H \ (B(γt) ∩ B(γt(cid:48))) and hence in H \ (Bt ∩ Bt(cid:48)).
However, it follows easily from the fact that (T, (Bt)t∈T ) is a tree decomposition of H that
every path from a bag in Tt(cid:48) to a bag in T \ Tt(cid:48) must intersect Bt ∩ Bt(cid:48). This proves that
aw(H) ≤ fhw(H).
For the second inequality, we shall prove the following stronger claim:
Claim: Let H be a hypergraph with aw(H) ≤ r for some nonnegative real r. Furthermore,
let γ : E(H) → [0,∞) such that weight(γ) ≤ 2r +2. Then there exists a fractional hypertree
decomposition of H of width at most 3r + 2 such that B(γ) is contained in the bag of the
root of this decomposition.
Note that for γ = 0, the claim yields the desired fractional hypertree decomposition of
H.
Proof of the claim: The proof is by induction on the cardinality of V (H) \ B(γ).
By Lemma 4.6, there is a balanced separator of weight at most r for γ in H. Let σ be such
a separator, and define χ : E(H) → [0,∞) by χ(e) = γ(e) + σ(e). Then weight(χ) ≤ 3r + 2,
and B(γ) ∪ B(σ) ⊆ B(χ).
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:13
If V (H) = B(χ) (this is the induction basis), then the 1-node decomposition with bag
V (H) and guard χ is a fractional hypertree decomposition of H of width at most 3r + 2.
Otherwise, let R1, . . . , Rm be the connected components of H\B(χ). Note that we cannot
exclude the case m = 1 and R1 = V (H) \ B(χ).
For 1 ≤ i ≤ m, let ei be an edge of H such that ei ∩ Ri (cid:54)= ∅, and let Si be the unique
connected component of H \ B(σ) with Ri ⊆ Si. Note that weight(γSi) ≤ r + 1, because σ
is a balanced separator for γ. Let χi : E(H) → [0,∞) be defined by
1
χi(e) =
σ(e) + γ(e)
σ(e)
if e = ei,
if e (cid:54)= ei and Si ∩ e (cid:54)= ∅,
otherwise.
Then
weight(χi) ≤ 1 + weight(σ) + weight(γSi) ≤ 2r + 2
and B(χi) \ Ri ⊆ B(χ) (as ei cannot intersect any Rj with i (cid:54)= j). Let Hi = H[Ri ∪ B(χi)]
and observe that
V (Hi) \ B(χi) ⊆ Ri \ ei ⊂ Ri ⊆ V (H) \ B(γ)
(the first inclusion holds because χi(ei) = 1). Thus the induction hypothesis is applica-
ble to Hi and the restriction of χi to Hi. It yields a fractional hypertree decomposition
(T i, (Bi
t)t∈V (T i)) of Hi of weight at most 3r + 2 such that B(χi) is contained in
the bag Bi
ti
0
t)t∈V (T i), (γi
of the root ti
0 of T i.
t to H for all t ∈ V (T i).
0 of the T i. Let Bt0 = B(χ) and Bt = Bi
the roots ti
and let γt be the canonical extension of γi
Let T be the disjoint union of T 1, . . . , T m together with a new root t0 that has edges to
t for all t ∈ V (T i). Moreover, let γt0 = χ,
It remains to prove that (T, (Bt)t∈V (T ), (γt)t∈V (T )) is a fractional hypertree decomposition
of H of width at most 3r + 2. Let us first verify that (T, (Bt)t∈V (T )) is a tree decomposition.
- Let v ∈ V (H). To see that {v ∈ V (T ) v ∈ Bt} is connected in T , observe that {t ∈
V (T i) v ∈ Bti} is connected (maybe empty) for all i. If v ∈ Ri for some i, then
v (cid:54)∈ V (Hj) = Rj ∪ B(χj) for any i (cid:54)= j (as Ri and Rj are disjoint and we have seen that
B(χj) \ Rj ⊆ B(χ)) and this this already shows that {t ∈ V (T ) v ∈ Bt} is connected.
Otherwise, v ∈ B(χi) \ Ri ⊆ B(χ) = Bt0 for all i such that v ∈ V (Hi). Again this shows
that {v ∈ V (T ) v ∈ Bt} is connected.
- Let e ∈ E(H). Either e ⊆ B(χ) = Bt0 , or there is exactly one i such that e ⊆ Ri ∪ B(χi).
In the latter case, e ⊆ Bt for some t ∈ V (T i).
It remains to prove that Bt ⊆ B(γt) for all t ∈ T . For the root, we have Bt0 = B(γt0 ). For
t ∈ V (T i), we have Bt ⊆ B(γi
t) = B(γt) ∩ V (Hi) ⊆ B(γt). Finally, note that weight(γt) ≤
3r + 2 for all t ∈ V (T ). This completes the proof of the claim.
Remark 4.7. With respect to the difference between hypertree decompositions and gener-
alized hypertree decompositions, it is worth observing that the fractional tree decomposition
(T, (Bt)t∈V (T ), (γt)t∈V (T )) of width at most 3r + 2 constructed in the proof of the theorem
u∈V (Tt) Bu ⊆ Bt for all t ∈ V (T ). This
satisfies the following special condition: B(γt) ∩(cid:83)
implies that a hypergraph of fractional hypertree width at most r has a fractional hypertree
decomposition of width at most 3r + 2 that satisfies the special condition.
4.2. Finding decompositions
For the algorithmic applications, it is essential to have algorithms that find fractional hy-
pertree decompositions of small width. The question is whether for any fixed r > 1 there
is a polynomial-time algorithm that, given a hypergraph H with fhw(H) ≤ r, computes
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:14
M. Grohe, D. Marx
a fractional hypertree decomposition of H of width at most r or maybe of width at most
f (r) for some function f . Similarly to hypertree width, one way of obtaining such an al-
gorithm would be through the army and robber game characterization. The idea would be
to inductively compute the set of all positions of the game from which the general wins in
0, 1, . . . rounds. The problem is that, as opposed to the robber and marshals game, there is
no polynomial bound on the number of positions.
Using a different approach (approximately solving the problem of finding balanced sep-
arators with small fractional edge cover number), Marx [2010a] gave an algorithm that
approximates fractional hypertree width in the following sense:
Theorem 4.8 ([Marx 2010a]). For every r ≥ 1, there is an nO(r3) time algorithm
that, given a hypergraph with fractional hypertree width at most r, finds a fractional hypertree
decomposition of width O(r3).
The main technical challenge in the proof of Theorem 4.8 is finding a separator with bounded
fractional edge cover number that separates two sets X, Y of vertices. An approximation
algorithm is given in [Marx 2010a] for this problem, which finds a separator of weight O(r3)
if a separator of weight r exists. This algorithm is used to find balanced separators, which
in turn is used to construct a tree decomposition (by an argument similar to the proof of
the second part of Theorem 4.5).
It is shown in [Marx 2010a; Fomin et al. 2009] that deciding whether H has fractional
hypertree width r is NP-hard if r is part of the input. However, the more relevant question
of whether for every fixed r, a fractional hypertree decomposition of width r can be found in
polynomial-time (i.e., if the width bound O(r3) in Theorem 4.8 can be improved to r) is still
open. Given that it is NP-hard to decide whether a hypergraph has generalized hypertree
width at most 3 [Gottlob et al. 2009], it is natural to expect that a similar hardness result
holds for fractional hypertree width as well.
4.3. Algorithmic applications
In this section, we discuss how problems can be solved by fractional hypertree decompo-
sitions of bounded width. First we give a basic result, which formulates why tree decom-
positions and width measures are useful in the algorithmic context: CSP can be efficiently
solved if we can polynomially bound the number of solutions in the bags. Recall that HI
denotes the hypergraph of a CSP instance I. If (T, (Bt)t∈V (T )) is a tree decomposition of
HI , then I[Bt] denotes the instance induced by bag Bt, see Definition 3.4.
Lemma 4.9. There is an algorithm that, given a CSP instance, a hypertree decompo-
sition (T, (Bt)t∈V (T )) of HI , and for every t ∈ V (t) a list Lt of all solutions of I[Bt],
decides in time C · (cid:107)I(cid:107)O(1) if I is satisfiable (and computes a solution if it is), where
C := maxt∈V (T )Lt.
Proof. Define Vt :=(cid:83)
t0
is not empty for the root t0 of the tree decomposition.
t∈V (Tt) Bt. For each t ∈ V (T ), our algorithm constructs the list
t ⊆ Lt of those solutions of I[Bt] that can be extended to a solution of I[Vt]. Clearly, I
L(cid:48)
has a solution if and only if L(cid:48)
The algorithm proceeds in a bottom-up manner: when constructing the list L(cid:48)
t, we assume
that for every child t(cid:48) of t, the lists L(cid:48)
t(cid:48) are already available. If t is a leaf node, then Vt = Bt,
and L(cid:48)
t = Lt. Assume now that t has children t1, . . . , tk. We claim that a solution α of
I[Bt] can be extended to I[Vt] if and only if for each 1 ≤ i ≤ k, there is a solution αi of
I[Vti] that is compatible with α (that is, α and αi assign the same values to the variables
in Bt ∩ Vti = Bt ∩ Bti). The necessity of this condition is clear: the restriction of a solution
of I[Vt] to Vti is clearly a solution of I[Vti ]. For sufficiency, suppose that the solutions αi
exist for every child ti. They can be combined to an assignment α(cid:48) on Vt extending α in a
well-defined way: every variable v ∈ Vti ∩ Vtj is in Bt, thus αi(v) = αj(v) = α(v) follows for
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:15
Therefore, L(cid:48)
such a variable. Now α(cid:48) is a solution of I[Vt]: for each constraint of I[Vt], the variables of
the constraint are contained either in Bt or in Vti for some 1 ≤ i ≤ k, thus α or αi satisfies
the constraint, implying that α(cid:48) satisfies it as well.
t can be determined by first enumerating every solution α ∈ Lt, and then for
each i, checking whether L(cid:48)
contains an assignment αi compatible with α. This check can
be efficiently performed the following way. Recall that solutions α and αi are compatible
if their restriction to Bt ∩ Bti is the same assignment. Therefore, after computing L(cid:48)
, we
restrict every αi ∈ L(cid:48)
to Bt ∩ Bti and store these restrictions in a trie data structure for
easy membership tests. Then to check if there is an αi ∈ L(cid:48)
compatible with α, all we
need to do is to check if the restriction of α to Bt ∩ Bti is in the trie corresponding to
L(cid:48)
t (and the corresponding trie structure) can
be computed in time Lt · (cid:107)I(cid:107)O(1). As every other part of the algorithm can be done in
time (cid:107)I(cid:107)O(1), it follows that the total running time can be bounded by C · (cid:107)I(cid:107)O(1). Using
standard bookkeeping techniques, it is not difficult to extend the algorithm such that it
actually returns a solution if one exists.
, which can be checked in (cid:107)I(cid:107)O(1). Thus L(cid:48)
ti
ti
ti
ti
ti
Lemma 4.9 tells us that if we have a tree decomposition where we can give a polynomial
bound on the number of solutions in the bags for some reason (and we can enumerate all
these solutions), then the problem can be solved in polynomial time. Observe that in a
fractional hypertree decomposition every bag has bounded fractional edge cover number
and hence Theorem 3.5 can be used to enumerate all the solutions. It follows that if a
fractional hypertree decomposition of bounded width is given in the input, then the problem
can be solved in polynomial time. Moreover, if we know that the hypergraph has fractional
hypertree width at most r (but no decomposition is given in the input), then we can use brute
force to find a fractional hypertree decomposition of width at most r by trying every possible
decomposition and then solve the problem in polynomial time. This way, the running time
is polynomial in the input size times an (exponential) function of the number of variables.
This immediately shows that if we restrict CSP to a class of hypergraphs whose fractional
hypertree width is at most a constant r, then the problem is fixed-parameter tractable
parameterized by the number of variables. To get rid of the exponential factor depending
on the number of variables and obtain a polynomial-time algorithm, we can replace the brute
force search for the decomposition by the approximation algorithm of Theorem 4.8 [Marx
2010a].
Theorem 4.10. Let r ≥ 1. Then there is a polynomial-time algorithm that, given a CSP
instance I of fractional hypertree width at most r, decides if I is satisfiable (and computes
a solution if it is).
Proof. Let I be a CSP instance of fractional hypertree width at most r, and let
(T, (Bt)t∈V (T ), (γt)t∈V (T )) be the fractional hypertree decomposition of HI of width O(r3)
computed by the algorithm of Theorem 4.8. By the definition, the hypergraph of I[Bt] has
fractional edge cover number O(r3) for every bag Bt. Thus by Theorem 3.5, the list Lt of
the solutions of I[Bt] has size at most (cid:107)I(cid:107)O(r3) and can be determined in time (cid:107)I(cid:107)O(r3).
Therefore, we can find a solution in time (cid:107)I(cid:107)O(r3) using the algorithm of Lemma 4.9.
In the remainder of this section, we sketch further algorithmic applications of fractional
hypertree decompositions. As these results follow from our main results with fairly standard
techniques, we omit a detailed and technical discussion.
It is has been observed by Feder and Vardi [1998] that constraint satisfaction problems
can be described as homomorphism problems for relational structures (see [Feder and Vardi
1998] or [Grohe 2007] for definitions and details). A homomorphism from a structure A
to a structure B is a mapping from the domain of A to the domain of B that preserves
membership in all relations. With each structure A we can associate a hypergraph HA whose
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:16
M. Grohe, D. Marx
vertices are the elements of the domain of A and whose hyperedges are all sets {a1, . . . , ak}
such that (a1, . . . , ak) is a tuple in some relation of A. For every class H of hypergraphs,
we let Hom(H) be the problem of deciding whether a structure A with HA ∈ H has a
homomorphism to a structure B. As an immediate corollary to Theorem 4.10, we obtain:
Corollary 4.11. Let H be a class of hypergraphs of bounded fractional hypertree width.
Then Hom(H) is solvable in polynomial time.
An embedding is a homomorphism that is one-to-one. Note that there is an embedding
from a structure A to a structure B if and only if B has a substructure isomorphic to A.
Analogously to Hom(H), we define the problem Emb(H) of deciding whether a structure
A with HA ∈ H has an embedding into a structure B. Observe that Emb(H) is NP-
complete even if H is the class of paths, which has fractional hypertree width 1, because
the Hamiltonian Path problem is a special case. However, we obtain a fixed-parameter
tractability result for Emb(H) parameterized by the size A of the input structure A:
Theorem 4.12. Let H be a class of hypergraphs of bounded fractional hypertree width.
Then Emb(H) parameterized by the size of the input structure A is fixed-parameter tractable.
More precisely, there is an algorithm that, given a structure A with HA ∈ H and a structure
B, decides if there is an embedding of A into B in time 2AO(1)BO(1).
This follows from Corollary 4.11 with Alon, Yuster, and Zwick's [Alon et al. 1995] color
coding technique.
In some situations it is necessary to not only decide whether a CSP-instance has a solution
or to compute one solution, but to enumerate all solutions. As the number of solutions may
be exponential in the instance size, we can rarely expect a polynomial-time algorithm for
this problem. Instead, we may ask for a polynomial-delay algorithm, which is required to
compute the first solution in polynomial time and after returning a solution is required to
return the next solution (or determine that no other solution exists) in polynomial time.
Polynomial-delay algorithms for CSPs have been studied more systematically in [Bulatov
et al. 2012].
Theorem 4.13. Let r ≥ 1. Then there is a polynomial-delay algorithm that, given a
CSP instance I of fractional hypertree width at most r, enumerates all solutions of I.
Proof (sketch). Given an instance I, the algorithm first computes a fractional hyper-
tree decomposition (T, (Bt)t∈V (T ), (γt)t∈V (T )) of HI of width O(r3) using the algorithm of
Theorem 4.8. Then it orders the nodes of T by a preorder traversal and then orders the
variables of I in such a way that if v ∈ Bt and w ∈ Bu \ Bt and t comes before u in the
preorder traversal, v is smaller than w. For every node t of T the algorithm computes a list
of all solutions for I[Bt] and sorts it lexicographically. It is easy to modify the algorithm of
Lemma 4.9 to compute the lexicographically first solution and, for every given solution, the
lexicographically next solution.
A problem closely related to the problem of enumerating all solutions to a given CSP-
instance is the problem of computing the answer for a conjunctive query in a relational
database (see [Kolaitis and Vardi 1998]). To be precise, answering conjunctive queries is
equivalent to computing projections of solution sets of CSP-instances to a given subset V (cid:48)
of variables. With each conjunctive query, we can associate hypergraph in a similar way
as we did for CSP-instances. Then answering queries with hypergraphs in H is equivalent
to computing projections of solution sets of Csp(H)-instances. We define the fractional
hypertree width of a conjunctive query to be the fractional hypertree width of its hypergraph.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:17
Theorem 4.14. Let r ≥ 1. Then there is a polynomial-delay algorithm that, given a
conjunctive query Q of fractional hypertree width at most r and a database instance D,
computes the answer Q(D) of Q in D.
The proof is a straightforward refinement of the proof of the previous theorem.
5. CONCLUSIONS
In this paper we have considered structural properties that can make a constraint satis-
faction problem polynomial-time solvable. Previously, bounded hypertree width was the
most general such property. Answering an open question raised in [Chen and Dalmau 2005;
Cohen et al. 2008; Gottlob et al. 2005; Grohe 2007], we have identified a new class of
polynomial-time solvable CSP instances: instances having bounded fractional edge cover
number. This result suggests the definition of fractional hypertree width, which is always
at most as large as the hypertree width (and in some cases much smaller). It turns out that
CSP is polynomial-time solvable for instances having bounded fractional hypertree width, if
the hypertree decomposition is given together with the instance. This immediately implies
that CSP is fixed-parameter tractable parameterized by the number of variables for hyper-
graphs with bounded fractional hypertree width. Furthermore, together with the algorithm
of [Marx 2010a] finding approximate fractional hypertree decompositions, it also follows
that CSP is polynomial-time solvable for this class. Currently, bounded fractional hyper-
tree width is the most general known structural property that makes CSP polynomial-time
solvable.
The most natural open question we leave open regarding fractional hypertree width is
whether for every fixed r there is a polynomial-time algorithm that decides if a hypergraph
has fractional hypertree width at most r (and if so, constructs a decomposition). As the
analogous problem for generalized hypertree width is NP-hard for r = 3, we expect this to
be the case for fractional hypertree width as well.
Another open question is whether there are polynomial-time solvable or fixed-parameter
tractable families of CSP instances having unbounded fractional hypertree width. Very
recently, [Marx 2010b] showed that, under suitable complexity assumptions, bounded sub-
modular width (a condition strictly more general that bounded fractional hypertree width)
exactly characterizes the fixed-parameter tractability of the problem. However, the exact
condition characterizing polynomial-time solvability is still an open problem. In light of the
results of [Marx 2010b], one needs to understand the complexity of the problem for classes
that have bounded submodular width, but unbounded fractional hypertree width.
REFERENCES
Adler, I. 2004. Marshals, monotone marshals, and hypertree-width. Journal of Graph Theory 47, 275–296.
Adler, I., Gottlob, G., and Grohe, M. 2007. Hypertree width and related hypergraph invariants. Eur.
J. Comb. 28, 8, 2167–2181.
Alon, N., Yuster, R., and Zwick, U. 1995. Color-coding. Journal of the ACM 42, 844–856.
Atserias, A., Grohe, M., and Marx, D. 2008. Size bounds and query plans for relational joins. In 49th
Annual IEEE Symposium on Foundations of Computer Science (FOCS 2008). 739–748.
Berge, C. 1976. Graphs and Hypergraphs. North Holland.
Bulatov, A. A. 2006. A dichotomy theorem for constraint satisfaction problems on a 3-element set. J.
ACM 53, 1, 66–120.
Bulatov, A. A. 2011. Complexity of conservative constraint satisfaction problems. ACM Trans. Comput.
Log. 12, 4, 24.
Bulatov, A. A., Dalmau, V., Grohe, M., and Marx, D. 2012. Enumerating homomorphisms. J. Comput.
Syst. Sci. 78, 2, 638–650.
Bulatov, A. A., Krokhin, A. A., and Jeavons, P. 2001. The complexity of maximal constraint languages.
In Proceedings of the 33rd ACM Symposium on Theory of Computing (STOC 2001). 667–674.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
0:18
M. Grohe, D. Marx
Chen, H. and Dalmau, V. 2005. Beyond hypertree width: Decomposition methods without decompositions.
In Principles and Practice of Constraint Programming (CP 2005), P. van Beek, Ed. Lecture Notes in
Computer Science Series, vol. 3709. Springer Berlin / Heidelberg, 167–181.
Chen, H. and Grohe, M. 2010. Constraint satisfaction with succinctly specified relations. Journal of
Computer System Sciences 76, 847–860.
Chung, F., Frankl, P., Graham, R., and Shearer, J. 1986. Some intersection theorems for ordered sets
and graphs. Journal of Combinatorial Theory, Series A 43, 23–37.
Cohen, D. A., Jeavons, P., and Gyssens, M. 2008. A unified theory of structural tractability for constraint
satisfaction problems. J. Comput. Syst. Sci. 74, 5, 721–743.
Dalmau, V., Kolaitis, P. G., and Vardi, M. Y. 2002. Constraint satisfaction, bounded treewidth, and
finite-variable logics. In Proceedings of the 8th International Conference on Principles and Practice of
Constraint Programming (CP 2002), P. V. Hentenryck, Ed. Lecture Notes in Computer Science Series,
vol. 2470. Springer-Verlag, 310–326.
Dechter, R. 2003. Constraint Processing. Morgan Kaufmann.
Dechter, R. and Pearl, J. 1989. Tree clustering for constraint networks. Artificial Intelligence 38, 353–
366.
Fagin, R. 1983. Degrees of acyclicity for hypergraphs and relational database schemes. Journal of the
ACM 30, 3, 514–550.
Feder, T. and Vardi, M. 1998. The computational structure of monotone monadic SNP and constraint
satisfaction: A study through datalog and group theory. SIAM Journal on Computing 28, 57–104.
Fomin, F. V., Golovach, P. A., and Thilikos, D. M. 2009. Approximating acyclicity parameters of sparse
hypergraphs. In 26th International Symposium on Theoretical Aspects of Computer Science (STACS
2009), S. Albers and J.-Y. Marion, Eds. Leibniz International Proceedings in Informatics (LIPIcs)
Series, vol. 3. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, Dagstuhl, Germany, 445–456.
Freuder, E. 1990. Complexity of k-tree structured constraint satisfaction problems. In Proceedings of the
8th National Conference on Artificial Intelligence. 4–9.
Friedgut, E. and Kahn, J. 1998. On the number of copies of a hypergraph in another. Israel Journal of
Mathematics 105, 251–256.
Gottlob, G., Grohe, M., Musliu, N., Samer, M., and Scarcello, F. 2005. Hypertree decompositions:
Structure, algorithms, and applications. In Graph-Theoretic Concepts in Computer Science (WG 2005),
D. Kratsch, Ed. Lecture Notes in Computer Science Series, vol. 3787. Springer Berlin / Heidelberg, 1–15.
Gottlob, G., Lee, S. T., Valiant, G., and Valiant, P. 2012. Size and treewidth bounds for conjunctive
queries. J. ACM 59, 3, 16:1–16:35.
Gottlob, G., Leone, N., and Scarcello, F. 2000. A comparison of structural CSP decomposition meth-
ods. Artif. Intell. 124, 2, 243–282.
Gottlob, G., Leone, N., and Scarcello, F. 2002. Hypertree decompositions and tractable queries. Jour-
nal of Computer and System Sciences 64, 579–627.
Gottlob, G., Leone, N., and Scarcello, F. 2003. Robbers, marshals, and guards: Game theoretic and
logical characterizations of hypertree width. Journal of Computer and System Sciences 66, 775–808.
Gottlob, G., Mikl´os, Z., and Schwentick, T. 2009. Generalized hypertree decompositions: NP-hardness
and tractable variants. J. ACM 56, 6.
Grohe, M. 2007. The complexity of homomorphism and constraint satisfaction problems seen from the
other side. J. ACM 54, 1.
Grohe, M. and Marx, D. 2006. Constraint solving via fractional edge covers. In Proceedings of the sev-
enteenth annual ACM-SIAM symposium on Discrete algorithm (SODA 2006). ACM Press, New York,
NY, USA, 289–298.
Grohe, M., Schwentick, T., and Segoufin, L. 2001. When is the evaluation of conjunctive queries
tractable? In Proceedings of the 33rd ACM Symposium on Theory of Computing (STOC 2001). 657–
666.
Impagliazzo, R., Paturi, R., and Zane, F. 2001. Which problems have strongly exponential complexity?
J. Comput. System Sci. 63, 4, 512–530.
Jeavons, P., Cohen, D. A., and Gyssens, M. 1997. Closure properties of constraints. Journal of the
ACM 44, 4, 527–548.
Kolaitis, P. and Vardi, M. 1998. Conjunctive-query containment and constraint satisfaction. In Proceed-
ings of the 17th ACM Symposium on Principles of Database Systems (PODS 1998). 205–213.
Marx, D. 2008. Closest substring problems with small distances. SIAM Journal on Computing 38, 4,
1382–1410.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
Constraint Solving via Fractional Edge Covers
0:19
Marx, D. 2010a. Approximating fractional hypertree width. ACM Trans. Algorithms 6, 2, 1–17.
Marx, D. 2010b. Tractable hypergraph properties for constraint satisfaction and conjunctive queries. In
Proceedings of the 42nd ACM Symposium on Theory of Computing (STOC 2010). 735–744.
Marx, D. 2011. Tractable structures for constraint satisfaction with truth tables. Theory of Computing
Systems 48, 444–464.
Ngo, H. Q., Porat, E., R´e, C., and Rudra, A. 2012. Worst-case optimal join algorithms. In Proceedings
of the 31st symposium on Principles of Database Systems (PODS 2012). ACM, New York, NY, USA,
37–48.
Reed, B. 1997. Tree width and tangles: A new connectivity measure and some applications. In Surveys in
Combinatorics, R. Bailey, Ed. LMS Lecture Note Series Series, vol. 241. Cambridge University Press,
87–162.
Rhadakrishnan, J. Entropy and counting. Available at http://www.tcs.tifr.res.in/∼jaikumar/mypage.html.
Schaefer, T. 1978. The complexity of satisfiability problems. In Proceedings of the 10th ACM Symposium
on Theory of Computing (STOC 1978). 216–226.
Schrijver, A. 2003. Combinatorial optimization. Polyhedra and efficiency. Algorithms and Combinatorics
Series, vol. 24. Springer, Berlin.
Seymour, P. and Thomas, R. 1993. Graph searching and a min-max theorem for tree-width. Journal of
Combinatorial Theory, Series B 58, 22–33.
Yannakakis, M. 1981. Algorithms for acyclic database schemes. In 7th International Conference on Very
Large Data Bases (VLDB 1981). 82–94.
ACM Transactions on Algorithms, Vol. 0, No. 0, Article 0, Publication date: 0.
|
1412.2379 | 2 | 1412 | 2015-05-04T04:49:47 | An Algorithm for $L_\infty$ Approximation by Step Functions | [
"cs.DS"
] | An algorithm is given for determining an optimal $b$-step approximation of weighted data, where the error is measured with respect to the $L_\infty$ norm. For data presorted by the independent variable the algorithm takes $\Theta(n + \log n \cdot b(1+\log n/b))$ time and $\Theta(n)$ space. This is $\Theta(n \log n)$ in the worst case and $\Theta(n)$ when $b = O(n/\log n \log\log n)$. A minor change determines an optimal reduced isotonic regression in the same time and space bounds, and the algorithm also solves the $k$-center problem for 1-dimensional weighted data. | cs.DS | cs |
An Algorithm for L∞ Approximation by Step Functions
Quentin F. Stout
Computer Science and Engineering
University of Michigan
Ann Arbor, MI 48109-2121
[email protected]
+1 734.763.1518
Abstract
An algorithm is given for determining an optimal b-step approximation of weighted data, where the error is
measured with respect to the L∞ norm. For data presorted by the independent variable the algorithm takes
Θ(n + log n · b(1 + log n/b)) time and Θ(n) space. This is Θ(n log n) in the worst case and Θ(n) when
b = O(n/ log n log log n). A minor change determines an optimal reduced isotonic regression in the same
time and space bounds, and the algorithm also solves the k-center problem for 1-dimensional weighted data.
Keywords: step function approximation; reduced isotonic regression; variable width histogram; weighted
k-center; interval tree of bounded envelopes
1 Introduction
Step functions are a fundamental form of approximation, arising in variable width histograms, databases,
segmentation, approximating sets of planar points, piecewise constant approximations, etc. Here we are
interested in L∞ stepwise approximation of weighted data. By weighted data (y, w) on 1 . . . n we mean
values (yi, wi), 1 ≤ i ≤ n, where yi is an arbitrary real number and wi (the weight) is a positive real
number. For integers i ≤ j let [i : j] denote i . . . j. A function f on [1 : n] is a b-step function iff there
are indices j1 = 1 < j2 < . . . < jb+1 = n + 1 and real values Ck, k ∈ [1 : b], such that fi = Ck for
i ∈ [jk : jk+1 − 1]. f is an optimal L∞ b-step approximation of (y, w) iff it minimizes the weighted L∞
error, max{wi · fi − yi : i ∈ [1 : n]}, among all b-step functions. Since a step can be split into smaller ones,
we do not differentiate between "b steps" and "no more than b steps".
Many algorithms have been developed for L∞ b-step regression [2, 3, 4, 5, 6, 8, 9, 11, 12, 13, 14].
The first Θ(n log n) time deterministic algorithms [6, 11] were decidedly impractical, relying on parametric
search. A more feasible Θ(n log n) algorithm appeared in [3]. However, the time of these algorithms does
not improve when b is small, which is the typical case of interest. We present a faster algorithm that is
Θ(n + log n · b(1 + log n/b)) when the data is presorted by independent coordinate.
2 L∞ b-Step Approximation
At a high-level overview, our algorithm shares aspects of those in [2, 5, 9], with important differences:
1. Build an interval tree to determine the regression error of an arbitrary interval if it is a single step.
2. Use "search in a sorted matrix" to find the minimal possible error for a b-step approximation.
The search uses a feasibility test which is given ǫ and decides if there is a b-step approximation with error
≤ ǫ. If there is such an approximation then the test produces one. We incorporate important improvements
to this approach: feasibility tests are used during tree construction, not just during the search; tests do
not determine the minimal regression error of an interval, merely that it is sufficiently small or too large;
previous searches, except the randomized version in [13], were not search in a sorted matrix; and we exploit
the fact that calculations at one stage of the search are related to those of the previous stage. We will show:
1
e
r
r
o
r
εhigh
new
εlow
old
εlow
a
b
y
x
c
regression value
essential segment endpoint
new ǫlow makes y and c inessential
b−1(new ǫlow) < x−1(new ǫlow), hence any
regression value in this range has error < new ǫlow
and exact value is not needed
Figure 1: Downward and upward bounded envelopes
Theorem 1 Given weighted data (y, w), sorted by the independent coordinate, and number of steps b, one
can determine an optimal L∞ b-step approximation in Θ(n + log n · b(1 + log n/b)) time and Θ(n) space.
Given a set (y, w) of weighted values and k ∈ [1 : n], the 1-dimensional weighted k-center problem is
to find a set S = {s1, . . . , sk} of real numbers that minimizes max{d(yi, S) : i ∈ [1 : n]}, where d(·, S) is
the weighted distance to S, i.e., d(yi, S) = min{wi · yi − sj : j ∈ [1 : k]}. Note that a set S is an optimal
k-center iff it is the step values of an optimal L∞ k-step approximation of the values in sorted order.
From now on we generally omit mention of "L∞" and "optimal" since they are implied.
2.1 Interval Tree of Bounded Envelopes
For a weighted value (y, w), in the y-z plane the error of using z ≥ y as its regression value is given by the
ray in the upper half-plane that starts at (y, 0) with slope w. Given a set of weighted data (y, w), its upward
error envelope is the topmost sequence of line segments corresponding to all such rays. For each z, it gives
the maximum error of using z as the regression value for all points (yi, wi) where z ≥ yi. The downward
error envelope uses rays in the upper half-plane starting at (yi, 0) with slope −wi, representing the error
of using a regression value ≤ yi. The intersection of the downward and upward error envelopes gives the
regression value minimizing the error over the entire set, i.e., the weighted L∞ mean, and its error.
To simplify exposition we assume that n is an integral power of 2. A binary interval tree has a root
corresponding to the interval [1 : n], its two children correspond to [1 : n/2] and [n/2 +1 : n], their children
represent intervals of length n/4, etc. The leaves are [1 : 1], [2 : 2], . . . [n : n]. The intervals corresponding to
nodes will be called binary intervals.
Some authors [2, 5, 9] used interval trees where each node contains the upward and downward error
envelopes of the data in its interval, but most of the envelopes' segments are unnecessary. Let ǫopt be the
(unknown) error of an optimal b-step approximation, and suppose bounds ǫlow < ǫopt ≤ ǫhigh are known.
Then ǫopt can be determined using only the segments representing errors in (ǫlow, ǫhigh). These will be called
essential segments, and they form bounded envelopes. All others, the inessential segments, are discarded.
Figure 1 shows how essential segments can become inessential when a better error bound is determined.
Initially ǫlow = 0 and ǫhigh = ∞, and the algorithm continually improves these bounds. In our interval
tree each node contains its upward and downward bounded envelopes. Throughout, the essential segments
are precisely the segments in the standard, unbounded, envelopes that represent errors in (ǫlow, ǫhigh). Thus
correctness depends on properties of the standard envelopes, though timing does not.
Bounded envelopes are stored as a doubly-linked list ordered by slope. Whenever a node is visited, by
starting at both ends, inessential segments (i.e., those with no errors in (ǫlow, ǫhigh)) are discarded. The time
is charged to the segments removed, not the search visiting the node. Only Θ(n) segments are ever created,
hence the total time to remove inessential ones is Θ(n). Whenever the number of remaining segments is
counted the count is only of the essential segments given the current values of ǫlow and ǫhigh.
2
2.2 Feasibility Tests
j=1Ij for some k ≥ 1, where each Ij is a binary interval. Then U (S)−1(ǫ) = mink
Given an arbitrary interval let U (S) denote the upward bounded envelope of the data in S, and D(S) the
downward bounded envelope. For ǫ ∈ (ǫlow, ǫhigh) the error of making S a single step is <, =, > ǫ iff
D(S)−1(ǫ) <, =, > U (S)−1(ǫ) (see Figure 1). Since ǫ ∈ (ǫlow, ǫhigh) this can be calculated as follows: let
S = ∪k
j=1 U (Ij)−1(ǫ)
and D(S)−1(ǫ) = maxk
j=1 D(Ij)−1(ǫ). For a binary interval I, to determine U (I)−1(ǫ), and similarly
D(I)−1(ǫ), go through the segments of its bounded envelope until the segment r containing ǫ is found.
This search alternates back and forth starting at the topmost and bottommost essential segments. This is
only performed during a feasibility test, which will result in either the segments above r, or those below
r, becoming inessential. Thus the time to find r is at most a constant plus a term linear in the number of
segments that become inessential. Here too the linear term is charged to the inessential segments.
S (ǫ), and max values used for D−1
Any interval [i : j] can be decomposed into ≤ 2⌊lg(j − i + 1)⌋ + 1 binary intervals where the sizes
increase and then decrease, with perhaps two intervals of the same size in the middle. E.g., [2 : 13] =
[2 : 2] ∪ [3 : 4] ∪ [5 : 8] ∪ [9 : 12] ∪ [13 : 13]. These can be visited in O(log n) time by a tree traversal starting
at the leaf [i : i], moving upward to the least common ancestor of i and j, and then downward to the leaf
[j : j]. Suppose, given i and ǫ ∈ (ǫlow, ǫhigh), we want to find the largest j such that the error of making [i : j]
a single step is ≤ ǫ. We do this by a traversal to locate j + 1. By incrementally updating the min values
to determine U −1
S (ǫ), when moving upward at node p one can determine
if j + 1 is in p's subtree (and hence the traversal should start going downward) by using p's envelopes to
decide if adding the entire subtree gives an error > ǫ. When moving downward, j + 1 is in p's left subtree
iff adding the left subtree gives error > ǫ, otherwise it is in p's right subtree. Not counting the queries of
children's envelopes, the nodes visited are the same as those in going from i to j + 1 when j + 1 is known.
Given b and ǫ ∈ (ǫlow, ǫhigh), a feasibility test determines if there is a b-step function with regression
error ≤ ǫ. This can be accomplished by starting at 1 and determining the largest j1 for which the error of
making [1 : j1] a single step is ≤ ǫ, then starting at j1 + 1 and determining the largest j2 for which the error of
making [j1 + 1 : j2] a single step is ≤ ǫ, etc. If the bth step is finished before n is reached then ǫ is infeasible,
the test stops, and ǫlow = ǫ. Otherwise, ǫ is feasible, the steps have been identified, and ǫhigh = ǫ.
To count the number of nodes visited, for each step the traversal visits nodes at a given height at most
twice, once moving upward and once moving downward. Thus at any height at most 2b nodes are visited.
The top ⌊lg b⌋ levels have a total of Θ(b) nodes. There are ⌈lg n⌉ − ⌊lg b⌋ = Θ(log n/b) lower levels, so in
total Θ(b(1 + log n/b)) nodes are visited. Each visit takes Θ(1) time, so this is also the time required.
2.3 Constructing the Tree
We reduce the time to construct the interval tree of bounded envelopes by continually shrinking (ǫlow, ǫhigh].
At the end, (ǫlow, ǫhigh] is so small that each bounded envelope is a single segment. See Figure 2.
First a feasibility test with ǫ = 0 is performed using only the base level. If it passes then the algorithm is
done. Otherwise, set ǫlow = 0, ǫhigh = ∞, and R = ∅. Throughout, R is an unordered multiset containing
the errors of all segment endpoints (e.g., the error of the joint endpoint of a and b in Fig. 1) remaining in
(ǫlow, ǫhigh). At height 0 each interval is a singleton, with single rays in its upward and downward envelopes
for a total of 2n rays. In general, at the end of height h, R < n/2h and the total number of essential
segments in the envelopes at height h is < 3n/2h. Moving upward, bounded envelopes from height h are
merged to form those at height h + 1, creating < 3n/2h segments and < 2n/2h segment endpoints (the
number of segment endpoints is the number of segments minus one per envelope). Add the errors of the
segment endpoints to those already in R, resulting in R < 3n/2h. Take the median error in R and do a
3
initialize envelopes of leaf nodes, ǫlow = 0, ǫhigh = ∞, R = ∅
for h=1 to lg n {h is height}
for every binary interval I at height h, make I's envelopes by merging children's envelopes,
and add any segment endpoint errors in (ǫlow, ǫhigh) to R
repeat 3 times {reducing R to < n/2h and total essential segments at height h < 3n/2h}
feasibility test using median of remaining essential segment endpoint errors in R
do 2 more feasibility tests, reducing R to ∅, i.e., all envelopes are single essential segments
Figure 2: Constructing the Tree of Bounded Envelopes, Feasibility Tests Continually Shrink (ǫlow, ǫhigh]
feasibility test. Depending on the outcome, one of ǫlow and ǫhigh is adjusted and 1/2 the entries in R can be
eliminated. Doing this 3 times results in R < n/2h+1. The number of essential segments is ≤ R+ 1 per
envelope (since R included segment endpoints at level h + 1), and hence is < 3n/2h+1.
When the top is finished R ≤ 2 and 2 feasibility tests are used to eliminate the remaining endpoint
errors, i.e., at every node of the interval tree the upward and downward bounded envelopes have only one
segment. Complete the tree construction by removing all inessential segments, taking Θ(n) time. Feasibility
tests during tree construction have a slight change from standard traversals in that when height h is being
constructed, when the test's traversals reach height h they go sideways, not upwards, from one node to the
next since nodes at higher levels haven't yet been constructed. This increases the total number of nodes
visited per test by at most n/2h. Since only 3 tests are done per height (see Figure 2), this adds Θ(n) total
time over all heights. Thus the total time to construct the tree is Θ(n + log n · b(1 + log n/b)).
2.4 Search for Minimal Feasible Error
The L∞ error of a stepwise approximation is the maximum of the L∞ errors of its steps, thus there is an
interval [i : j] such that the error of an optimal b-step approximation is the error of using the weighted L∞
mean as the step value on [i : j]. Thus searching through such errors can determine the minimal feasible
error. "Parametric search" was used in [6, 11] but this is only of theoretical interest since parametric search
is completely impractical, involving very complex data structures and quite large constants.
Search in a sorted matrix provides a practical approach [7]. Let E be the n × n matrix where E(i, j) is
the error of using the L∞ mean on [i : j] if i ≤ j, and is 0 if i > j. E is not actually created, but serves as
a conceptual guide. Few of its entries are ever determined. Its rows are nondecreasing and the columns are
nonincreasing, so for any submatrix its largest entry is in the upper right and the smallest is in the lower left.
The algorithm has stages 0 . . . lg n − 1, where at the start of stage s there is a collection of disjoint
square submatrices of size n/2s. Stage 0 starts with all of E. At each stage, divide each of the matrices into
quadrants, and let ǫ1 be a median of the smallest value from each quadrant and ǫ2 a median of their largest
values. Values determined to be outside (ǫlow, ǫhigh), as in Figure 1, are not calculated exactly and are set
to ǫlow − 1 or ǫhigh + 1, as appropriate. Feasibility tests for ǫ1 and ǫ2 are done, resulting in improvements
to ǫlow and/or ǫhigh. Quadrants with smallest value ≥ ǫhigh, or largest value ≤ ǫlow, are eliminated. The
remaining quadrants are the matrices that start the next stage. Note that if ǫopt < ǫhigh then any quadrant
with an entry of ǫopt is not eliminated, hence at the end of each stage, either ǫopt = ǫhigh or one of the entries
in the remaining submatrices is ǫopt.
After the last stage the remaining matrices are 1 × 1 and a standard binary search on these values is
used to finish the determination of ǫopt. The search uses Θ(log n) feasibility tests, and the less obvious fact,
proven in [7], is that only Θ(n) entries of E are evaluated.
4
E(I,J): intervals with one endpoint in I, other in J, gap is K
I
K
J
I1
I2
a
b
quadrant E(I , J )11
J1
2J
c
d
smallest entry in E(I , J ): E(b, c)
largest entry in E(I , J ):
1
1
1
1 E(a, d)
: one endpoint in I , other in J , gap is K I
∪
1
1
2
Figure 3: Evaluating smallest, largest entries in E(I1, J1) using gap from E(I, J)
2.5 Evaluating E During The Search
For intervals I, J ⊆ [1 : n] let E(I, J) denote the submatrix {E(i, j) : i ∈ I, j ∈ J}, i.e., the submatrix
corresponding to intervals starting at some i ∈ I and ending at some j ∈ J. At the start of stage s of
the search there is a collection of submatrices of the form E(I, J) for binary intervals I, J of size n/2s.
Either I = J, or I is to the left of J and there is a (perhaps empty) gap between them with length an
integral multiple of n/2s (K in Figure 3). During stage s, the quadrants of E(I, J) are formed by cutting
I and J in half into I1, I2 and J1, J2, respectively, creating quadrants E(I1, J1), E(I1, J2), E(I2, J1), and
E(I2, J2). The smallest and largest value in each quadrant needs to be determined, and as Figure 3 shows,
one can evaluate the smallest entry in E(I1, J1) (i.e., E(b, c)), by using the envelopes from K and the binary
intervals [b : b], I2, and [c : c]. The largest entry in the quadrant, E(a, d), uses envelopes from K and the
binary intervals I and J1. Similar results hold for all of the other quadrants of E(I, J). Exact values for
entries outside (ǫlow, ǫhigh] are irrelevant and ǫlow − 1 or ǫhigh + 1 is used, as appropriate.
The bounded envelopes for gap K are associated with E(I, J), and if, say, E(I1, J1) is kept for stage
s+1 then the envelopes for I2∪K are associated with it. Just as for the tree construction, as search in a sorted
matrix is proceeding the number of segments in the gap envelopes is reduced by interleaving feasibility tests
based on the segment endpoint errors with tests for the basic search. At the end of stage s each gap envelope
is copied at most 4 times, and each quadrant passed on to the next level adds ≤ 3 binary intervals which
have envelopes that are only single segments. As shown in [7] there are ≤ 2s+3 − 1 such quadrants, so by
using 2 additional feasibility tests at each stage the total number of segments in the bounded envelopes at
stage s is O(2s). Since the time to evaluate an entry of E is linear in the number of segments involved, the
total time to evaluate entries of E over all stages of the algorithm is Θ(n), and the total time of the Θ(log n)
feasibility tests is Θ(log n · b(1 + log n/b)) plus the Θ(n) time to remove inessential segments.
This completes the proof of Theorem 1.
3 Final Comments
We have shown how to find an L∞ b-step approximation of weighted data, presorted by its independent
coordinate, in Θ(n + log n · b(1 + log n/b)) time. No previous algorithm [2, 3, 4, 5, 6, 8, 9, 11, 12, 13, 14]
was o(n log n) whenever b = o(n), nor Θ(n) whenever b = O(n/ log n log log n). For sorted data the
algorithm solves the 1-dimensional weighted k-center problem in the same time.
With a small change the algorithm also produces a "reduced isotonic" b-step function. f is an isotonic
function iff f (1) ≤ f (2) ≤ . . . ≤ f (n), and is an optimal L∞ b-step reduced isotonic regression of (y, w)
iff it minimizes the L∞ error among all isotonic b-step functions. Isotonic regression is an important form
of nonparametric regression that allows researchers to replace parametric assumptions with weaker shape
5
constraints [1, 15]. Some researchers were concerned that it can overfit the data and/or be too compli-
cated [10, 16, 17] and resorted to reduced isotonic regression. However, they used approximations because
previous exact algorithms were too slow. Merely changing the feasibility test to insure increasing steps finds
b-step reduced isotonic regression in the same time bounds as b-step approximation.
Acknowledgement: Research partially supported by NSF grant CDI-1027192
References
[1] Barlow, RE, Bartholomew, DJ, Bremner, JM, and Brunk, HD, Statistical Inference Under Order Re-
strictions: The Theory and Application of Isotonic Regression, John Wiley, 1972.
[2] Chen, DZ and Wang, H, "Approximating points by a piecewise linear function", Algorithmica 66
(2013), pp. 682 -- 713.
[3] Chen, DZ and Wang, H, "A note on searching line arrangements and applications", Info. Proc. Let. 113
(2013), pp. 518 -- 521.
[4] D´ıaz-B´annez, JM and Mesa, JA, "Fitting rectilinear polygonal curves to a set of points in the plane",
Eur. J. Operational Res. 130 (2001), pp. 214 -- 222.
[5] Fournier, H and Vigneron, A, "Fitting a step function to a point set", Algor. 60 (2011), pp. 95 -- 109.
[6] Fournier, H and Vigneron, A, "A deterministic algorithm for fitting a step function to a weighted point-
set", Info. Proc. Let. 113 (2013), pp. 51 -- 54.
[7] Frederickson, G and Johnson, D, "Generalized selection and ranking: Sorted matrices", SIAM J. Comp.
13 (1984), pp. 14 -- 30.
[8] Fulop, J and Prill, M, "On the minimax approximation in the class of the univariate piecewise constant
functions", Oper. Res. Let. 12 (1992), pp. 307 -- 312.
[9] Guha, S and Shim, K, "A note on linear time algorithms for maximum error histograms", IEEE Trans.
Knowledge and Data Engin. 19 (2007), pp. 993 -- 997.
[10] Haiminen, N, Gionis, A, and Laasonen, K, "Algorithms for unimodal segmentation with applications
to unimodality detection", Knowl. Info. Sys. 14 (2008), pp. 39 -- 57.
[11] Hardwick, J and Stout, QF, "Optimal reduced isotonic reduction", Proc. Interface 2012.
[12] Karras, P, Sacharidis, D., and Mamoulis, N, "Exploiting duality in summarization with deterministic
guarantees", Proc. Int'l. Conf. Knowledge Discovery and Data Mining (KDD) (2007), pp. 380 -- 389.
[13] Liu, J-Y, "A randomized algorithm for weighted approximation of points by a step function", COCOA
1 (2010), pp. 300 -- 308.
[14] Mayster, Y and Lopez, MA, "Approximating a set of points by a step function", J. Vis. Commun. Image
R. 17 (2006), pp. 1178 -- 1189.
[15] Robertson, T, Wright, FT, and Dykstra, RL, Order Restricted Statistical Inference, Wiley, 1988.
[16] Salanti, G and Ulm, K, "A nonparametric changepoint model for stratifying continuous variables under
order restrictions and binary outcome", Stat. Methods Med. Res. 12 (2003), pp. 351 -- 367.
[17] Schell, MJ and Singh, B, "The reduced monotonic regression method", J. Amer. Stat. Assoc. 92 (1997),
pp. 128 -- 135.
6
|
1009.3502 | 1 | 1009 | 2010-09-17T20:25:56 | Simultaneous Interval Graphs | [
"cs.DS"
] | In a recent paper, we introduced the simultaneous representation problem (defined for any graph class C) and studied the problem for chordal, comparability and permutation graphs. For interval graphs, the problem is defined as follows. Two interval graphs G_1 and G_2, sharing some vertices I (and the corresponding induced edges), are said to be `simultaneous interval graphs' if there exist interval representations R_1 and R_2 of G_1 and G_2, such that any vertex of I is mapped to the same interval in both R_1 and R_2. Equivalently, G_1 and G_2 are simultaneous interval graphs if there exist edges E' between G_1-I and G_2-I such that G_1 \cup G_2 \cup E' is an interval graph.
Simultaneous representation problems are related to simultaneous planar embeddings, and have applications in any situation where it is desirable to consistently represent two related graphs, for example: interval graphs capturing overlaps of DNA fragments of two similar organisms; or graphs connected in time, where one is an updated version of the other.
In this paper we give an O(n^2*logn) time algorithm for recognizing simultaneous interval graphs,where n = |G_1 \cup G_2|. This result complements the polynomial time algorithms for recognizing probe interval graphs and provides an efficient algorithm for the interval graph sandwich problem for the special case where the set of optional edges induce a complete bipartite graph. | cs.DS | cs | Simultaneous Interval graphs
Krishnam Raju Jampani ∗
Anna Lubiw †
0
1
0
2
p
e
S
7
1
]
S
D
.
s
c
[
1
v
2
0
5
3
.
9
0
0
1
:
v
i
X
r
a
Abstract
In a recent paper, we introduced the simultaneous representation problem (defined for any graph
class C) and studied the problem for chordal, comparability and permutation graphs. For interval
graphs, the problem is defined as follows. Two interval graphs G1 and G2, sharing some vertices I (and
the corresponding induced edges), are said to be "simultaneous interval graphs" if there exist interval
representations R1 and R2 of G1 and G2, such that any vertex of I is mapped to the same interval
in both R1 and R2. Equivalently, G1 and G2 are simultaneous interval graphs if there exist edges E ′
between G1 − I and G2 − I such that G1 ∪ G2 ∪ E ′ is an interval graph.
Simultaneous representation problems are related to simultaneous planar embeddings, and have
applications in any situation where it is desirable to consistently represent two related graphs, for
example:
interval graphs capturing overlaps of DNA fragments of two similar organisms; or graphs
connected in time, where one is an updated version of the other.
In this paper we give an O(n2log n) time algorithm for recognizing simultaneous interval graphs,
where n = G1 ∪ G2. This result complements the polynomial time algorithms for recognizing probe
interval graphs and provides an efficient algorithm for the interval graph sandwich problem for the spe-
cial case where the set of optional edges induce a complete bipartite graph.
Keywords: Simultaneous Graphs, Interval Graphs, Graph Sandwich Problem, Probe Graphs,
PQ-trees
1
Introduction
Let C be any intersection graph class (such as interval graphs or chordal graphs) and let G1 and G2 be two
graphs in C, sharing some vertices I and the edges induced by I. G1 and G2 are said to be simultaneously
representable C graphs or simultaneous C graphs if there exist intersection representations R1 and R2 of G1
and G2 such that any vertex of I is represented by the same object in both R1 and R2. The simultaneous
representation problem for class C asks whether G1 and G2 are simultaneous C graphs. For example, Figures
1(a) and 1(b) show two simultaneous interval graphs and their interval representations with the property
that vertices common to both graphs are assigned to the same interval. Figure 1(c) shows two interval
graphs that are not simultaneous interval graphs.
Simultaneous representation problems were introduced by us in a recent paper [9] and have application
in any situation where two related graphs should be represented consistently. A main instance is for
temporal relationships, where an old graph and a new graph share some common parts. Pairs of related
graphs also arise in many other situations, e.g: two social networks that share some members; overlap
graphs of DNA fragments of two similar organisms, etc.
∗David R. Cheriton School of Computer Science, University of Waterloo, Email:[email protected]
†David R. Cheriton School of Computer Science, University of Waterloo, Email:[email protected]
1
e
d
G1
f
G2
a
b
c
(a)
e
a
d
c
b
f
(b)
e
d
G1
f
G2
a
b
c
(c)
Figure 1: Graphs in (a) are simultaneous interval graphs as shown by the representations in (b). Graphs
in (c) are not simultaneous interval graphs.
Simultaneous representations are related to simultaneous planar embeddings: two graphs that share
some vertices and edges (not necessarily induced) have a simultaneous geometric embedding [3] if they have
planar straight-line drawings in which the common vertices are represented by common points. Thus edges
may cross, but only if they are in different graphs. Deciding if two graphs have a simultaneous geometric
embedding is NP-Hard [4].
In [9], we showed that the simultaneous representation problem can be solved efficiently for chordal,
comparability and permutation graphs. We also showed that for any intersection class C, the simultaneous
representation problem for G1 and G2 is equivalent to the following problem: Do there exist edges E′
between G1 − I and G2 − I so that the augmented graph G1 ∪ G2 ∪ E′ belongs to class C.
The graph sandwich problem [7] is a more general augmentation problem defined for any graph class
C: given graphs H1 = (V, E1) and H2 = (V, E2), is there a set E of edges with E1 ⊆ E ⊆ E2 so that the
graph G = (V, E) belongs to class C. This problem has a wealth of applications but is NP-complete for
interval graphs, comparability graphs, and permutation graphs [7].
The simultaneous representation problem (for class C) is the special case of the graph sandwich problem
(for C) where E2 − E1 forms a complete bipartite subgraph. A related special case where E2 − E1 forms a
clique is the problem of recognizing probe graphs: a graph G with a specified independent set N is a probe
graph for class C if there exist edges E′ ⊆ N × N so that the augmented graph G ∪ E′ belongs to class C.
Probe graphs have several applications [13, 8] and have received much attention recently. The first
polynomial-time algorithm for recognizing probe interval graphs was due to Johnson and Spinrad [10].
They used a variant of PQ-trees and achieved a run-time of O(n2). Techniques from modular decomposition
provided more speed up [12], but the most recent algorithm by McConnell and Nussbaum [11] reverts to
PQ-trees and achieves linear time.
We note that there has also been work [14] on a concept of simultaneous intersection called "polysemy"
where two graphs are represented as intersections of sets and their complements.
In this paper, we give an O(n2log n) algorithm for solving the simultaneous representation problem for
interval graphs. We use PQ-trees, which were developed by Booth and Lueker for the original linear time
interval graph recognition algorithm. They used a PQ-tree to capture the orderings of the maximal cliques
of the graph (see [6] for an introduction to interval graphs and PQ-trees).
In the probe interval recognition problem, there is a single PQ-tree (of the graph induced by the probes)
and a set of constraints imposed by the non-probes. However in our situation we have two PQ-trees, one
for each graph, that we want to re-order to "match" on the common vertex set I. We begin by "reducing"
each PQ-tree to contain only vertices from I. This results in PQ-trees that store non-maximal cliques, and
our task is to modify each PQ-tree by inserting non-maximal cliques from the other tree while re-ordering
the trees to make them the same.
2
2 Reduction to PQ-trees
In this section we transform the simultaneous interval graph problem to a problem about "compatibility"
of two PQ-trees arising from the two graphs.
Recall that an interval graph is defined to be the intersection graph of intervals on the real line. For
any point on the line, the intervals containing that point form a clique in the graph. This leads to the
fundamental one-to-one correspondence between the interval representations of an interval graph and its
clique orderings, defined as follows: A clique ordering of G is a sequence of (possibly empty) cliques
S = Q1, Q2, · · · , Ql that contains all the maximal cliques of G and has the property that for each vertex
v, the cliques in S that contain v appear consecutively. Note that we allow cliques to be empty.
The standard interval graph recognition algorithm attempts to find a clique order of the maximal cliques
of a graph by making the maximal cliques into leaves of a PQ-tree, and imposing PQ-tree constraints to
ensure that the cliques containing each vertex v appear consecutively. This structure is called the PQ-tree
of the graph. Note that the children of a P-node may be reordered arbitrarily and the children of a Q-node
may only be reversed. We consider a node with 2 children to be a Q-node. In the figures, we use a circle
to denote a P-node and a rectangle to denote a Q-node. A leaf-order of a PQ-tree is the order in which
its leaves are visited in an in-order traversal of the tree, after children of P and Q-nodes are re-ordered as
just described.
Note that ignoring non-maximal cliques is fine for recognizing interval graphs; for our purposes, however,
we want to consider clique orders and PQ-trees that may include non-maximal cliques. We say that a PQ-
tree whose leaves correspond to cliques of a graph is valid if for each of its leaf orderings and for each vertex
v, the cliques containing v appear consecutively.
Let S = Q1, Q2, · · · , Ql be a clique ordering of interval graph G and let the maximal cliques of G be
Qi1 , Qi2 , · · · , Qim (appearing in positions i1 < i2 < · · · < im respectively). Note that all the cliques in
S between Qij and Qij+1 contain B = Qij ∩ Qij+1. We say that B is the boundary clique or boundary
between Qij and Qij+1. Note that B may not necessarily be present in S. The sequence of cliques between
Qij and Qij+1 that are subsets of Qij is said to be the right tail of Qij . The left tail of Qij+1 is defined
analogously. Observe that the left tail of a clique forms an increasing sequence and the right tail forms
a decreasing sequence (w.r.t set inclusion). Also note that all the cliques that precede Qi1 are subsets of
Qi1 and this sequence is called the left tail of Qi1 and all the cliques that succeed Qim are subsets of Qim
and this sequence is called the right tail of Qim. Thus any clique ordering of G consists of a sequence of
maximal cliques, with each maximal clique containing a (possibly empty) left and right tail of subcliques.
Let Q0 and Ql+1 be defined to be empty sets. An insertion of clique Q′ between Qi and Qi+1 (for some
i ∈ {0, · · · , l}) is said to be a subclique insertion if Q′ ⊇ Qi ∩ Qi+1 and either Q′ ⊆ Qi or Q′ ⊆ Qi+1. It is
clear that after a subclique insertion the resulting sequence is still a clique ordering of G. A clique ordering
S ′ is an extension of S if S ′ can be obtained from S by subclique insertions. We also say that S extends
to S ′. Furthermore, we say that a clique ordering is generated by a PQ-tree, if it can be obtained from a
leaf order of the PQ-tree with subclique insertions. The above definitions yield the following Lemma.
Lemma 1. A sequence of cliques S is a clique ordering of G if and only if S can be generated from the
PQ-tree of G.
Let G1 and G2 be two interval graphs sharing a vertex set I (i.e. I = V (G1) ∩ V (G2)) and its induced
edges. Note that G1[I] is isomorphic to G2[I]. A clique ordering of G1[I] is said to be an I-ordering.
The I-restricted PQ-tree of Gj is defined to be the tree obtained from the PQ-tree of Gj by replacing
each clique Q (a leaf of the PQ-tree) with the clique Q ∩ I. Thus there is a one-to-one correspondence
between the two PQ-trees, and the leaves of the I-restricted PQ-tree are cliques of G1[I].
3
Let I = X1, X2, · · · , Xl be an I-ordering. I is said to be Gj-expandable if there exists a clique ordering
O = Q1, Q2, · · · , Ql of Gj such that Xi ⊆ Qi for i ∈ {1, · · · , l}. Further, we say that I expands to O. By
the definition of clique-ordering it follows that, if I is Gj-expandable then it remains Gj-expandable after
a subclique insertion (i.e. any extension of I is also Gj-expandable). We first observe the following.
Lemma 2. The set of Gj-expandable I-orderings is same as the set of orderings that can be generated
from the I-restricted PQ-tree of Gj.
Proof. Let T be the PQ-tree of Gj and T ′ be the I-restricted PQ-tree of Gj.
Let I be a Gj-expandable I-ordering of Gj. Then there exists a clique ordering O of Gj such that
I expands to O. But by Lemma 1, O can be generated from T (from a leaf order with subclique inser-
tions). This in turn implies that I can be generated from T ′ (from the corresponding leaf order with the
corresponding subclique insertions).
Now for the other direction, let I ′ = X1, · · · , Xl be any leaf order of T ′. Then there exists a corre-
sponding leaf order O′ = Q1, · · · , Ql of T such that Xi ⊆ Qi for i ∈ {1, · · · , l}. This implies that I ′ is a
Gj-expandable I-ordering. Finally, observe that if I ′′ is generated from I ′ by subclique insertions than I ′′
is also a Gj-expandable I-ordering. Thus the Lemma holds.
Two I-orderings I1 and I2 are said to be compatible if both I1 and I2 (separately) extend to a common I-
ordering I. For e.g. the ordering {1}, {1, 2}, {1, 2, 3, 4} is compatible with the ordering {1}, {1, 2, 3}, {1, 2, 3, 4},
as they both extend to the common ordering: {1}, {1, 2}, {1, 2, 3}, {1, 2, 3, 4}. Note that the compatibility
relation is not transitive. Two PQ-trees T1 and T2 are said to be compatible if there exist orderings O1 and
O2 generated from T1 and T2 (respectively) such that O1 is compatible with O2. The following Lemma is
our main tool.
Lemma 3. G1 and G2 are simultaneous interval graphs if and only if the I-restricted PQ-tree of G1 is
compatible with the I-restricted PQ-tree of G2.
Proof. By Lemma 2, it is enough to show that G1 and G2 are simultaneous interval graphs if and only if
there exists a G1-expandable I-ordering I1 and a G2-expandable I-ordering I2 such that I1 is compatible
with I2. We now show this claim.
Let I1 and I2 be as defined in the hypothesis. Since I1 and I2 are compatible, they can be extended to
a common I-ordering I. Let I expand to clique orderings O1 and O2 in G1 and G2 respectively. Since each
vertex of I appears in the same positions in both O1 and O2, it is possible to obtain interval representations
R1 and R2 of G1 and G2 (from O1 and O2 respectively) such that each vertex in I has the same end points
in both R1 and R2. This implies that G1 and G2 are simultaneous interval graphs.
For the other direction, let G1 and G2 be simultaneous interval graphs. Then there exists an augmenting
set of edges A′ ⊆ V1 − I × V2 − I such that G = G1 ∪ G2 ∪ A′ is an interval graph. Let O = Q1, Q2, · · · , Ql
be a clique-ordering of G. For each i ∈ {1, · · · , l} and j ∈ {1, 2}, by restricting Qi to Vj (i.e. replacing Qi
with Qi ∩ Vj), we obtain a clique ordering Oj of Gj. Now for j ∈ 1, 2, let Ij be the I-ordering obtained
from Oj by restricting each clique in Oj to I. It follows that I1 is a G1-expandable I-ordering and I2 is a
G2-expandable I-ordering. Further I1 = I2 and hence I1 and I2 are compatible.
Our algorithm will decide if the I-restricted PQ-tree of G1 is compatible with the I-restricted PQ-tree
of G2. We first show how the I-restricted PQ-trees can be simplified in several ways. Two I-orderings I1
and I2 are said to be equivalent if for any I-ordering I ′, I1 and I ′ are compatible if and only if I2 and
I ′ are compatible. Note that this is an equivalence relation. The Lemma below follows directly from the
definitions of equivalent orderings and subclique insertions.
4
Lemma 4. Let I = X1, X2, · · · , Xl be an I-ordering in which Xi = Xi+1 for some i ∈ 1, · · · , l − 1. Let I ′
be the I-ordering obtained from I by deleting Xi+1. Then I is equivalent to I ′.
Further, because equivalence is transitive, Lemma 4 implies that an I-ordering I is equivalent to the
I-ordering I ′ in which all consecutive duplicates are eliminated. This allows us to simplify the I-restricted
PQ-tree of Gj. Let T be the I-restricted PQ-tree of Gj. We obtain a PQ-tree T ′ from T as follows.
1. Initialize T ′ = T .
2. As long as there is a non-leaf node n in T ′ such that all the descendants of n are the same, i.e. they are
all duplicates of a single clique X, replace n and the subtree rooted at n by a leaf node representing X.
3. As long as there is a (non-leaf) Q-node n in T ′ with two consecutive child nodes na and nb (among
others) such that all the descendants of na and nb are the same i.e. they are all duplicates of a single clique
X, replace na, nb and the subtrees rooted at these vertices by a single leaf node representing the clique X.
Note that the resulting T ′ is unique. We call T ′ the I-reduced PQ-tree of Gj.
Lemma 5. G1 and G2 are simultaneous interval graphs if and only if the I-reduced PQ-tree of G1 is
compatible with the I-reduced PQ-tree of G2.
Proof. For j ∈ {1, 2}, let Tj and T ′
j be the I-restricted and I-reduced PQ-trees of Gj respectively. Let I
be any I-ordering. Observe that by Lemma 4, I is compatible with a leaf ordering of Tj if and only if I is
compatible with a leaf ordering of T ′
j. Thus the conclusion follows from Lemma 3.
3 Labeling and Further Simplification
In section 2, we transformed the simultaneous interval graph problem to a problem of testing compatibility
of two I-reduced PQ-trees where I is the common vertex set of the two graphs. These PQ-trees may have
nodes that correspond to non-maximal cliques in I. In this section we prove some basic properties of such
I-reduced PQ-trees, and use them to further simplify each tree.
Let T be the I-reduced PQ-tree of Gj. Recall that each leaf l of T corresponds to a clique X in Gj[I].
If X is maximal in I, then X is said to be a max-clique and l is said to be a max-clique node, otherwise
X is said to be a subclique and l is said to be a subclique node. When the association is clear from the
context, we will sometimes refer to a leaf l and its corresponding clique X interchangeably, or interchange
the terms "max-clique" and "max-clique node" [resp. subclique and subclique node]. A node of T is said
to be an essential node if it is a non-leaf node or if it is a leaf node representing a max-clique.
Given a node n of T , the descendant cliques of n are the set of cliques that correspond to the leaf-
descendants of n. Because our algorithm operates by inserting subcliques from one tree into the other, we
must take care to preserve the validity of a PQ-tree. For this we need to re-structure the tree when we
do subclique insertions. The required restructuring will be determined based on the label U (n) that we
assign to each node n as follows.
U (n) or the Universal set of n is defined as the set of vertices v such that v appears in all descendant
cliques of n.
Note that for a leaf node l representing a clique X, U (l) = X by definition. Also note that along any
path up the tree, the universal sets decrease. The following Lemma gives some useful properties of the
I-reduced PQ-tree.
5
Lemma 6. Let T be the I-reduced PQ-tree of Gj. Let n be a non-leaf node of T (n is used in properties
2–6). Then we have:
0. Let l1 and l2 be two distinct leaf nodes of T , containing a vertex t ∈ I. Let y be the least common
ancestor of l1 and l2. Then: (a) If y is a P-node then all of its descendant cliques contain t. (b) If y is
a Q-node then t is contained in all the descendant cliques of all children of y between (and including) the
child of y that is the ancestor of l1 and the child that is the ancestor of l2.
1. Each max-clique is represented by a unique node of T .
2. A vertex u is in U (n) if and only if for every child n1 of n, u ∈ U (n1).
3. n contains a max-clique as a descendant.
4. If n is a P-node, then for any two child nodes n1 and n2 of n, we have U (n) = U (n1) ∩ U (n2).
5. If n is a P-node, then any child of n that is a subclique node represents the clique U (n).
6. If n is a Q-node and n1 and n2 are the first and last child nodes of n then U (n) = U (n1) ∩ U (n2).
Proof. (0) Observe that in any leaf ordering of T , all the nodes that appear between l1 and l2 must also
contain the vertex t, otherwise T would be invalid. Now let l3 be a leaf descendant of y, that doesn't
contain t.
If y is a P-node, then we can reorder the children of y in such a way that in the leaf-ordering of
the resulting tree l3 appears between l1 and l2. But this contradicts the validity of T . This proves (a).
Similarly, if y is a Q-node, then l3 cannot be equal to l1 or l2 or any node between them. Thus (b) also holds.
(1) Note that by definition of I-reduced PQ-tree of Gj, each max-clique must be present in T . Now
assume for the sake of contradiction that a max-clique X is represented by two leaf nodes, say l1 and l2. Let
y be the least common ancestor of l1 and l2. Let c1 and c2 be the child nodes of y that contain n1 and n2
(respectively) as descendants. Now by (0), if y is a P-node then all of its descendant cliques must contain
all the vertices of X. But as X is maximal, all these cliques must be precisely X. However this is not
possible, as we would have replaced y with a leaf node representing X in the construction of T . Similarly,
if y is a Q-node then the descendant cliques of c1, c2 and all the nodes between them must represent the
max-clique X. But then we would have replaced these nodes with with a leaf node representing X in the
construction of T . This proves (1).
(2) If u ∈ U (n), then all the descendant cliques of n contain u. This implies that for any child n1 of
n, all the descendant cliques of n1 contain u. Hence u ∈ U (n1). On the other hand, if each child n1 of n
contains a vertex u in its universal set, then u is present in all the descendant cliques of n and thus u ∈ U (n).
(3) Note that if each descendant clique of n contains precisely U (n) (and no other vertex), then we
would have replaced the subtree rooted at n with a leaf node corresponding to the clique U (n), when
constructing T . Thus there exists a clique Q2 that is a descendant of n, such that Q2 − U (n) is non-empty.
If Q2 is a max-clique then we are done. Otherwise let t ∈ Q2 − U (n) and let Q1 be a max-clique containing
t. Suppose Q1 is a not a descendant of n1. Applying (0) on Q1 and Q2, we infer that irrespective of
whether n is a P-node or a Q-node, all the descendant cliques of n must contain t. But then t ∈ U (n), a
contradiction. Thus Q1 is a descendant of n1.
(4) By (2) we observe that U (n) ⊆ U (n1)∩U (n2). Thus it is enough to show that U (n1)∩U (n2) ⊆ U (n).
Let u ∈ U (n1) ∩ U (n2), then u is present in all the descendant cliques of n1 and n2. By (0), u must be
present in all the descendant cliques of n and hence u ∈ U (n). Therefore U (n1) ∩ U (n2) ⊆ U (n).
6
(5) Consider any child n1 of n. Suppose n1 is a leaf-node and is not a max-clique. It is enough to
show that n1 represents the clique U (n) i.e. U (n1) = U (n). Suppose not. Then there exists a vertex
t ∈ U (n1) − U (n). Let Q1 be a max-clique containing t. Note that the common ancestor of Q1 and n1 is
either n or an ancestor of n. Applying (0) on Q1 and n1, we infer that all the descendant cliques of n must
contain t. But then t ∈ U (n), a contradiction.
(6) This follows from (2) and (0).
Let T be the I-reduced PQ-tree of Gj. Recall that an essential node is a non-leaf node or a leaf node
representing a maximal clique. Equivalently (by Lemma 6.3), an essential node is a node which contains
a max-clique as a descendant. The following Lemma shows that in some situations we can obtain an
equivalent tree by deleting subclique child nodes of a P-node n. Recall that by Lemma 6.5, such subclique
nodes represent the clique U (n).
Lemma 7. Let T be the I-reduced PQ-tree of Gj and n be a P-node in T . Then
1. If n has at least two essential child nodes, then T is equivalent to the tree T ′, obtained from T by
deleting all the subclique children of n.
2. If n has at least two subclique child nodes, then T is equivalent to the tree T ′, obtained from T by
deleting all except one of the subclique children of n.
Proof. We give the proof of (1) below. The proof of (2) is very similar and hence omitted.
Let O1 be any I-ordering. It is enough to show that there exists a leaf ordering O of T that is compatible
with O1 if and only if there exists a leaf ordering O′ of T ′ that is compatible with O1.
Let O be any leaf ordering of T , compatible with O1. Consider the ordering O′ obtained from O by
deleting the cliques U (n) that correspond to the child nodes of n in T . Clearly O′ is a leaf ordering of T ′.
Further O′ can be extended to O by adding copies of the cliques U (n) at appropriate positions. Thus O′
is compatible with O1.
Now for the other direction, let O′ be a leaf order of T ′, compatible with O1 and let O′ and O1 extend to
a common ordering OF . From the hypothesis, we can assume that there exist two essential child nodes n1
and n2 of n in T ′ such that the clique descendants of n1, appear immediately before the clique descendants
of n2 in O′. Also let S(n1) and S(n2) be the two subsequences of O′ containing the clique descendants
of n1 and n2 respectively. Since n1 and n2 are essential nodes, S(n1) and S(n2) each contain at least one
max-clique. Let Q1 be the last max-clique in S(n1) and Q2 be the first max-clique in S(n2). By Lemma
6.0, Q1 ∩ Q2 = U (n1) ∩ U (n2) = U (n). Since O′ is compatible with O1, in each of the two orderings O′
and O1, Q2 occurs after Q1 and no other max-clique appears between them. Further the same holds for
OF (as it is an extension of O′). Let k be the number of subclique children of n (that represent the clique
U (n)). Then obtain a leaf ordering O of T , from O′, by inserting k copies of U (n) between S(n1) and
S(n2). Now extend OF to O′
F by inserting k copies of U (n) between Q1 and Q2 (there is a unique way
of adding a subclique between two max-cliques). It is clear that O′
F is an extension of both O and O1.
Therefore O is compatible with O1. This proves (1).
We will simplify T as much as possible by applying Lemma 7 and by converting nodes with two
children into Q-nodes. We call the end result a simplified I-reduced PQ-tree, but continue to use the term
"I-reduced PQ-tree" to refer to it. Note that the simplification process does not change the universal sets
and preserves the validity of the PQ-tree so Lemma 5 and all the properties given in Lemma 6 still hold.
Because we consider nodes with 2 children as Q-nodes Lemma 7 implies:
Corollory 1. In a [simplified] I-reduced PQ-tree, any P-node has at least 3 children, and all the children
are essential nodes.
7
4 Algorithm
For k ∈ {1, 2}, let Tk be the [simplified] I-reduced PQ-tree of Gk. By Lemma 5, testing whether G1 and
G2 are simultaneous interval graphs is equivalent to testing whether T1 and T2 are compatible. We test
this by modifying T1 and T2 (e.g. inserting the sub-clique nodes from one tree into the other) so as to make
them identical, without losing their compatibility. The following is a high level overview of our approach
for checking whether T1 and T2 are compatible.
Our algorithm is iterative and tries to match essential nodes of T1 with essential nodes of T2 in a
bottom-up fashion. An essential node n1 of T1 is matched with an essential node n2 of T2 if and only if the
subtrees rooted at n1 and n2 are the same i.e. their essential children are matched, their subclique children
are the same and furthermore (in the case of Q-nodes) their child nodes appear in the same order. If n1
is matched with n2 then we consider n1 and n2 to be identical and use the same name (say n1) to refer
to either of them. Initially, we match each max-clique node of T1 with the corresponding max-clique node
of T2. Note that every max-clique node appears uniquely in each tree by Lemma 6.1. A sub-clique node
may appear in only one tree in which case we must first insert it into the other tree. This is done when
we consider the parent of the subclique node.
In each iteration, we either match an unmatched node u of T1 to an unmatched node v of T2 (which
may involve inserting subclique child nodes of v as child nodes of u and vice versa) or we reduce either T1
or T2 without losing their compatibility relationship. Reducing a PQ-tree means restricting it to reduce
the number of leaf orderings. Finally, at the end of the algorithm either we have modified T1 and T2 to
a "common" tree TI that establishes their compatibility or we conclude that T1 is not compatible with
T2. The common tree TI is said to be an intersection tree (of T1 and T2) and has the property that any
ordering generated by TI can also be generated by T1 and T2. If T1 and T2 are compatible, there may be
several intersection trees of T1 and T2, but our algorithm finds only one of them.
We need the following additional notation for the rest of this paper. A sequence of subcliques S =
X1, X2, · · · , Xl is said to satisfy the subset property if Xi ⊆ Xi+1 for i ∈ {1, · · · , l − 1}. S is said to satisfy
the superset property if Xi ⊇ Xi+1 for each i. Note that S satisfies the subset property if and only if
¯S = Xl, · · · , X2, X1 satisfies the superset property.
Let d be an essential child node of a Q-node in Tk. We will overload the term "tail" (previously defined
for a max clique in a clique ordering) and define the tails of d as follows. The left tail (resp. right tail) of
d is defined as the sequence of subcliques that appear as siblings of d, to the immediate left (resp. right)
of d, such that each subclique is a subset of U (d). Note that the left tail of d should satisfy the subset
property and the right tail of d should satisfy the superset property (otherwise Tk will not be valid). Also
note that since the children of a Q-node can be reversed in order, "left" and "right" are relative to the
child ordering of the Q-node. We will be careful to use "left tail" and "right tail" in such a way that this
ambiguity does not matter. Now suppose d is a matched node. Then in order to match the parent of d in
T1 with the parent of d in T2, our algorithm has to "merge" the tails of d.
Let L1 and L2 be two subclique sequences that satisfy the subset property. Then L1 is said to be
mergable with L2 if the union of subcliques in L1 and L2 can be arranged into an ordering L′ that satisfies
the subset property. Analogously, if L1 and L2 satisfy the superset property, then they are said to be
mergable if the union of their subcliques can be arranged into an ordering L′ that satisfies the superset
property. In both cases, L′ is said to be the merge of L1 and L2 and is denoted by L1 + L2.
A maximal matched node is a node that is matched but whose parent is not matched. For an unmatched
essential node x, the MM-descendants of x, denoted by M M D(x) are its descendants that are maximal
matched nodes.
If x is matched then we define M M D(x) to be the singleton set containing x. Note
that the MM-descendants of an essential node is non-empty (since every essential node has a max-clique
8
descendant).
Our algorithm matches nodes from the leaves up, and starts by matching the leaves that are max-
cliques. As the next node n1 that we try to match, we want an unmatched node whose essential children
are already matched. To help us choose between T1 and T2, and also to break ties, we prefer a node with
larger U set. Then, as a candidate to match n1 to, we want an unmatched node in the other tree that has
some matched children in common with n1. With this intuition in mind, our specific rule is as follows.
Among all the unmatched essential nodes of T1 union T2 choose n1 with maximal U (n1), minimal
M M D(n1), and maximal depth, in that preference order. Assume without loss of generality that n1 ∈ T1.
Select an unmatched node n2 from T2 with maximal U (n2), minimal M M D(n2) and maximal depth (in
that order) satisfying the property that M M D(n1) ∩ M M D(n2) 6= ∅. The following Lemma captures
certain properties of n1 and n2, including why these rules match our intuitive justification.
Lemma 8. For n1 and n2 chosen as described above, let M1 = M M D(n1), M2 = M M D(n2) and X =
M1 ∩ M2. Also let C1 and C2 be the essential child nodes of n1 and n2 respectively. Then we have:
1. M1 = C1 and X ⊆ C2.
Further when T1 is compatible with T2, we have:
2. For every (matched) node l in M1 − X of T1, its corresponding matched node l′ in T2 is present outside
the subtree rooted at n2. Analogously, for every (matched) node r′ in M2 − X of T2, its corresponding
matched node r in T1 is present outside the subtree rooted at n1.
3. If n1 [resp. n2] is a Q-node, then in its child ordering, no node of C1 − X [resp. C2 − X] can be present
between two nodes of X.
4. If n1 and n2 are Q-nodes, then in the child ordering of n1 and n2, nodes of X appear in the same
relative order i.e. for any three nodes x1, x2, x3 ∈ X, x1 appears between x2 and x3 in the child ordering of
n1 if and only if x1 also appears between x2 and x3 in the child ordering of n2.
5. If C1 − X (resp.C2 − X) is non-empty then U (n1) ⊆ U (n2) (resp. U (n2) ⊆ U (n1)). Further, if C1 − X
is non-empty then so is C2 − X and hence U (n1) = U (n2).
6. Let C1 − X be non-empty. If n1 [resp. n2] is a Q-node, then in its child-ordering either all nodes of
C1 − X [resp. C2 − X] appear before the nodes of X or they all appear after the nodes of X.
Proof. (1) If there exists an unmatched child c of n1, then as U (c) ⊇ U (n1), M M D(c) ⊆ M M D(n1) and
c has a greater depth than n1, we would have chosen c over n1. Thus every node in C1 is matched and
hence by the definition of MM-descendants C1 = M1.
For the second part, suppose there exists a node x ∈ X that is not a child of n2. Let c2 be the child of
n2 that contains x as a descendant. c2 must be an unmatched node. (Otherwise M M D(n2) would have
contained c2 and not x). But then we would have picked c2 over n2.
(2) Let l be a (matched) node in M1 − X (in T1) such that the corresponding matched node l′ in T2
is a descendant of n2. Note that l′ cannot be a child of n2. Otherwise l′ ∈ M2 and thus l = l′ is in X.
Let p′ be the parent of l′. Now p′ cannot be a matched node. (Otherwise p′ would have been matched
to n1, a contradiction that n1 is unmatched). Also p′ is a descendant of n2 and hence U (p′) ⊇ U (n2),
M M D(p′) ⊆ M M D(n2) and p′ has greater depth than n2. Further l = l′ is a common MM-descendant of
n1 and p′. This contradicts the choice of n2.
Now let r′ be a (matched) node in M2 − X (in T2), such that the corresponding matched node r is a
descendant of n1. Note that r is not a child of n1, otherwise r = r′ is a common MM-descendant of n1
and n2 and hence r′ = r ∈ X. Let p be the parent of r in T1. Since p is a proper descendant of n1, p is a
matched node. Let p be matched to a node p′ in T2. Now p′ is a parent of r′ and a descendant of n2. But
then the MM-descendants of n2 should not have contained r′.
9
(3) Suppose in the child ordering of n1, node y ∈ C1 − X is present between nodes xa ∈ X and xb ∈ X.
Let Y, Xa and Xb be any max-cliques that are descendants of y, xa and xb respectively. Then in any
ordering of T1, Y appears between Xa and Xb. But by (2), the corresponding matched node y′ of y in T2
appears outside the subtree rooted at n2. Thus in any ordering of T2, Y appears either before or after both
Xa and Xb. Thus T1 and T2 are not compatible. This shows the claim for n1. The proof for n2 is similar.
(4) This follows from the fact that T1 and T2 are compatible and by observing that each matched node
(in particular any node in X) contains a max-clique as a descendant.
(5) Let xa ∈ X be a common child of n1 and n2. Let Xa be a max-clique descendant of xa. Suppose
C1 − X is non-empty. Then let Ya be any max-clique descendant of a node in C1 − X. Note that by (2),
Ya is present outside the subtree rooted at n2. Now by Lemma 6.0 and observing that the least common
ancestor of Xa and Ya (in T2) is an ancestor of n2, we get U (n2) ⊇ Xa ∩ Ya ⊇ U (n1). Thus U (n1) ⊆ U (n2).
Using an analogous argument we can show that if C2 − X is non-empty then U (n2) ⊆ U (n1). This proves
the first part of the property.
For the second part we once again assume that C1 − X is non-empty and hence U (n1) ⊆ U (n2). Now
if C2 − X is empty then M M D(n2) = X ⊂ M M D(n1). But this contradicts the choice of n1 (we would
have selected n2 instead).
(6) By (5), C2 − X is non-empty and U (n1) = U (n2). Let xa ∈ X and suppose ya, yb ∈ C1 − X are
any two nodes on different sides of X. Let za ∈ C2 − X. Note that by (2), the matched nodes of ya, yb in
T2 appear outside the subtree rooted at n2 and the matched node of za in T1 appears outside the subtree
rooted at n1. Now let Xa, Ya, Yb and Za be any descendant max-cliques of xa, ya, yb and za respectively. In
any leaf-ordering of T1, Xa appears between Ya and Yb, and Za doesn't appear between Ya and Yb. But in
any leaf-ordering of T2, either Za and Xa both appear between Ya and Yb or they both appear before or
after Ya and Yb. This contradicts that T1 and T2 are compatible. Therefore all nodes of X appear before or
after all nodes of C1 − X in the child ordering of n1. Similarly, the claim also holds for the child ordering
of n2 in T2.
We now describe the main step of the algorithm. Let n1, n2, M1, M2, C1, C2 and X be as defined in
the above Lemma. We have four cases depending on whether n1 and n2 are P or Q-nodes. In each of
these cases, we make progress by either matching two previously unmatched essential nodes of T1 and T2
or by reducing T1 and/or T2 at n1 or n2 while preserving their compatibility. We show that our algorithm
requires at most O(nlog n) iterations and each iteration takes O(n) time. Thus our algorithm runs in
O(n2 log n) time.
During the course of the algorithm we may also insert subcliques into a Q-node when we are trying to
match it to another Q-node. This is potentially dangerous as it may destroy the validity of the PQ-tree.
When the Q-nodes have the same universal set, this trouble does not arise. However, in case the two
Q-nodes have different universal sets, we need to re-structure the trees. Case 4, when n1 and n2 are both
Q-nodes, has subcases to deal with these complications.
Case 1: n1 and n2 are both P-nodes.
By Corollary 1, the children of n1 and n2 are essential nodes, so C1 and C2 are precisely the children of n1
and n2 respectively. Let X consist of nodes {x1, · · · , xk0}. If C2 − X is empty, then by Lemma 8.5, C1 − X
10
T1
n1
T2
n2
T1
n1
T2
n2
x1
xk0
x1
xk0
r1
rk2
x1
xk0
l1
lk1
x1
xk0
r1
rk2
n2
n1
n2
n1
r1
rk2
ml
m
m
mr
x1
xk0
l1
lk1
x1
xk0
x1
xk0
r1
rk2
(a) C1 − X is empty
(b) C1 − X is non-empty
Figure 2: Reduction templates for Case 1
is also empty and hence n1 and n2 are the same. So we match n1 with n2 and go to the next iteration.
Suppose now that C2 − X is non empty. Let C2 − X = {r1, · · · , rk2}. If C1 − X is empty, then we use the
reduction template of Figure 2(a) to modify T2, matching the new parent of X in T2 to n1. It is easy to
see that T1 is compatible with T2 if and only if T1 is compatible with the modified T2.
1 and T ′
Now let C1 − X = {l1, · · · , lk1} be non-empty. In this case we use the reduction template of Figure
2(b) to modify T1 and T2 to T ′
2 respectively. Note that it is possible to have ki = 1 for some i's, in
which case the template is slightly different because we do not make a node with one child, however, the
reduction always makes progress as each ni has at least 3 children.
We now claim that T1 is compatible with T2 if and only if T ′
2 . The reverse
direction is trivial. For the forward direction, let O1 and O2 be two compatible leaf orderings of T1 and
T2 respectively. Recall that by Lemma 8.2, for every [matched] node of C1 − X in T1, the corresponding
matched node in T2 appears outside the subtree rooted at n2. This implies that the descendant nodes
of {x1, x2, · · · , xk0} all appear consecutively in O1. Hence the descendant nodes of {x1, x2, · · · , xk0} also
appear consecutively in O2. Thus we conclude that T1 and T2 are compatible if and only if the reduced
trees T ′
1 is compatible with T ′
1 and T ′
2 are also compatible. Note that both the template reductions take at most O(n) time.
Case 2: n1 is a P-node and n2 is a Q-node.
If C1 − X = ∅, we reduce T1 by ordering the children of n1 as they appear in the child ordering of n2, and
changing n1 into a Q-node (and leading to Case 4). This reduction preserves the compatibility of the two
trees.
Now suppose C1 − X 6= ∅. Lemma 8.5 implies that, C2 − X 6= ∅ and U (n1) = U (n2). By Lemma 8.6,
we can assume that the nodes in X appear before the nodes in C2 − X in the child ordering of n2. Now let
X = x1, · · · , xk0, C1 − X = l1, · · · , lk1 and C2 − X = r1, · · · , rk2. For i ∈ 2, · · · , k0, let Si be the sequence
of subcliques that appear between xi−1 and xi in the child ordering of n2. Note that Si consists of the
right tail of xi−1 followed by the left tail of xi. We let S1 and Sk0+1 denote the left and right tails of x1
and xk0 respectively. We now reduce the subtree rooted at n1 as shown in Figure 3, changing it into a
Q-node. Clearly U (n1) is preserved in this operation. The correctness of this operation follows by Lemma
11
T1
n1
T2
n2
x1
x2
xk0
l1
lk1
x1
S1
S2
x2
xk0
r1
rk2
Sk0+1
T ′
1
n1
m
x1
S1
S2
x2
xk0
Sk0+1
l1
lk1
Figure 3: Reduction template for Case 2, when C1 − X 6= ∅
8.2. It is easy to see that both the template reductions run in O(n) time.
Case 3: n1 is a Q-node and n2 is a P-node.
If C2 − X is empty, then we reduce T2 by ordering the child nodes of n2 (i.e. X) as they appear in the
child ordering of n1, and changing n2 into a Q-node.
Now let C2 − X be nonempty. By Lemma 8.5, U (n2) ⊆ U (n1). Let X = {x1, x2, · · · , xk0}, C2 − X =
{r1, · · · , rk2} and S1, · · · , Sk0+1 be defined as in the previous case: S1 is the left tail of x1 (in T1), Si is the
concatenation of the right tail of xi−1 and the left tail of xi, for i ∈ {2, · · · , k0} and Sk0+1 is the right tail
of xk0.
Now if C1 − X is empty, then we use the template of Figure 4 to reduce T2, grouping all nodes of X
into a new Q-node w, ordering them in the way they appear in T1 and inserting the subclique children of
n1 into w. Note that since U (n2) ⊆ U (n1), this operation doesn't change U (n2) and hence it preserves the
validity of T2. Further n1 is identical to w and hence we match these nodes. Thus we make progress even
when X = 1.
If C1 − X is non-empty, we use the template similar to Figure 3 (to reduce T2) in which the roles of n1
and n2 have been switched. Note that the template reductions of this case run in O(n) time.
Case 4: n1 and n2 are both Q-nodes
Let X = {x1, · · · , xk0} appear in that order in the child ordering of n1 and n2. (They appear in the same
order because of Lemma 8.4.) Let p1 and p2 be the parents of n1 and n2 respectively.
If n1 and n2 have no other children than X, we match n1 with n2 and proceed to the next iteration.
More typically, they have other children. These may be essential nodes to one side or the other of X (by
Lemma 8.6) or subclique nodes interspersed in X as tails of the nodes of X. We give a high-level outline
of Case 4, beginning with a discussion of subclique nodes.
For i ∈ {1, · · · , k0}, let Li and Ri be the left and right tails of xi in T1 and, L′
i be the left and
right tails of xi in T2. The only way to deal with the subclique nodes is to do subclique insertions in both
trees to merge the tails. This is because in any intersection tree TI obtained from T1 and T2, the tails of xi
in TI must contain the merge of the tails of xi in T1 and T2. So long as X ≥ 2, the ordering x1, · · · , xk0
i and R′
12
T1
n1
x1
S1
S2
x2
xk0
Sk0+1
T2
n2
x1
x2
xk0
r1
rk2
n2
w
rk2
r1
x1
S1
S2
x2
xk0
Sk0+1
Figure 4: Reduction template for Case 3, when n1 is a Q-node, n2 is a P-node and C1 − X is empty.
completely determines which pairs of tails must merge: Li must merge with L′
R′
i.
i and Ri must merge with
The case X = 1 is more complicated because the intersection tree may merge L1 with L′
1 and R1 with
1 or merge L1 with ¯R1 and R1 with ¯L1. This decision problem is referred to as the alignment problem.
R′
We prove (at the beginning of Case 4.3) that in case both choices give mergable pairs, then either choice
yields an intersection tree, if an intersection tree exists.
This completes our high-level discussion of subclique nodes. We continue with a high-level description
of the subcase structure for Case 4. We have subcases depending on whether U (n1) = U (n2) and whether
n1 and n2 have the same essential children. If both these conditions hold, then we merge the tails of the
nodes of X and match n1 with n2. (In other words we replace Li and L′
i, and replace Ri and
R′
i). The cost of matching any two nodes x and y is (mx + my)I), where mx and my are
the number of subclique children of x and y respectively. Once a node is matched its subclique children
will not change. Hence the total amortized cost of matching all the nodes is O(n · I) = O(n2).
i with Ri + R′
i with Li + L′
When U (n1) 6= U (n2) or when n1 and n2 do not have the same essential children then we have three
subcases. Case 4.1 handles the situation when U (n1) 6⊇ U (n2). In this case we either insert subcliques
of one tree into another and match n1 with n2 or we do some subclique insertions that will take us to
the case when U (n1) ⊇ U (n2). The remaining cases handle the situation when U (n1) ⊇ U (n2), Case 4.2
when C1 − X is non-empty and Case 4.3 when it is empty. In both cases, we reduce T1 but the details
vary. However in both cases our reduction templates depend on whether p1 is a P-node or a Q-node. If
p1 is a P-node, we reduce T1 by grouping some of the child nodes of p1 into a single node, deleting them
and adding the node as a first or last child of n1. If p1 is a Q-node then there are two ways of reducing:
delete n1 and reassign its children as children of p1 or reverse the children of n1, delete n1 and reassign its
children as children of p1. We refer to this operation as a collapse. We now give the details of each case.
Case 4.1: U(n1) 6⊇ U(n2)
Since n1 was chosen so that U (n1) is maximal, we also have U (n2) 6⊇ U (n1). Now using Lemma 8.5, we
infer that C1 − X is empty and C2 − X is empty. Thus the difference between U (n1) and U (n2) arises due
13
to the subcliques. Let L be a subclique that is either the first or the last child of n1, with the property
that L 6⊆ U (n2). Such a subclique exists since by Lemma 6.6, the intersection of the universal sets of the
first and last child nodes of n1 is U (n1). Also let R be a subclique that is either the first or the last child
of n2, with the property that R 6⊆ U (n1).
Note that even if X = 1, the alignment is unique since L and R cannot appear in the same tail of x1
in any intersection tree. Further, we can assume without loss of generality that L is present in the left tail
of x1 in T1 and R is present in the right tail of xk0 in T2.
Let X1 be any max-clique descendant of x1. If p1 is a P-node then we claim that U (p1) ⊆ U (n2). To
see this, let Z be a max-clique descendant of p1, that is not a descendant of n1. In T2, Z appears outside
the subtree rooted at n2. Now by applying Lemma 6.0 on X1 and Z, we conclude that every descendant
of n2 must contain the vertex set Z ∩ X1. Thus we have U (p1) ⊆ (Z ∩ X1) ⊆ U (n2). Note that if T1 and
T2 are compatible, then in any intersection tree of T1 and T2, the nodes of L1 and L′
1 appear in the left tail
of x1 and the nodes of Rk0 and R′
1 is non-empty, then we insert
k0
the left most subclique of L′
1 into L1 (at the appropriate location so that the resulting sequence is still a
subclique ordering), as a child of n1. Also if R′
is non-empty, then we insert the right most subclique of
k0
R′
into Rk0, as a child of n1. These insertions change U (n1) to U (n1) ∩ U (n2) ⊇ U (p1) and we would be
k0
in case 4.3 with the roles of n1 and n2 being reversed. (Note that since the universal set of the modified
n1 is a superset of the universal set of p1, the resulting reduced tree of T1 is valid). Although this doesn't
constitute a progress step since the number of leaf orderings of n1 doesn't change, we will make progress
in the Case 4.3.
appear in the right tail of xk0. Now if L′
Similarly if p2 is a P-node then we insert the first subclique of L1 (if it exists) into L′
1 and the last
subclique of Rk0 (if it exists) into R′
k0
. After this we would be in Case 4.3.
Now if the parents of n1 and n2 are both Q-nodes then we look at the tails of n1 and n2. If all the
i and Ri and
i. This changes U (n1) and U (n2) to U (n1) ∩ U (n2) and makes n1 identical to n2. Thus we
subcliques in these tails are subsets of U (n1) ∩ U (n2), then we replace Li and L′
R′
match n1 with n2 and iterate.
i with Ri + R′
i with Li + L′
Otherwise without loss of generality let the subclique S 6⊆ U (n2) be present in the (say left) tail of n1.
Observe that in any intersection tree S and R cannot be present in the same tail of xk0 (since neither is
a subset of the other). This implies that we can reduce the tree T1 by collapsing n1 i.e. by removing n1,
inserting the sequence of child nodes of n1 after S (S and L are now in the left tail of x1), and assigning
p1 as their parent. This completes case 4.1. Note that all the steps in this case take O(n) time, except the
matching step (recall that all the matching steps take O(n2) amortized time).
Case 4.2: U(n1) ⊇ U(n2) and C1 − X is non-empty.
By Lemma 8.5, C2 − X is also non-empty and further U (n1) is equal to U (n2). In this case we will reduce
T1 depending on whether p1 is a P-node or a Q-node. Further when p1 is a Q-node, our reduction template
also depends on whether n1 has sibling essential nodes.
Let l1, l2, · · · , lk1 be the essential nodes in C1 − X appearing in that order and appearing (without loss
of generality) before the nodes of X in T1. Note that by Lemma 8.2, for each node in C1 − X, the corre-
sponding matched node in T2 appears outside the subtree rooted at n2. Thus if T1 and T2 are compatible,
then all the nodes of C2 − X must appear after the nodes of X in the child ordering of n2. Let these nodes
be r1, r2, · · · , rk2.
Case 4.2.1: p1 is a P-node
Let Y = {y1, y2, · · · , yk3} be the child nodes of p1 other than n1 . Also, let TI be any intersection tree of
T1 and T2. We first observe that for i ∈ {1, · · · , k0} and j ∈ {1, · · · , k2}, M M D(yi) ∩ M M D(rj) 6= ∅, if
14
and only if yi and rj have a max-clique descendant.
For any such pair yi and rj, let M M D(yi) ∩ M M D(rj) 6= ∅ and let Y be a common max-clique
descendant of yi and rj. Then note that because of the constraints imposed by the child ordering of n2,
in any leaf ordering of TI , the descendant cliques of l1 do not appear between the descendant cliques
of x1 and Y . Thus yi must appear after xk1, and so we reduce T1, by grouping all nodes yi satisfying
M M D(yi)∩M M D(rj) 6= ∅ for some rj into a P-node and adding it as a child node of n1 to the (immediate)
right of Rk0 as shown in Figure 5(top).
Now if M M D(yi) ∩ M M D(rj) = ∅ for all yi and rj, then the above reduction doesn't apply. But
in this case (because of the constraints on n2), for every yi and every leaf ordering of TI , no max-clique
descendant of yi appears between the max-clique descendants of n2. Thus we group all the nodes of Y into
a P-node and add it as a child of n1 to the left of l1 as shown in Figure 5(bottom).
Note that for any two distinct nodes ya, yb ∈ {y1, · · · , yk3} we have: M M D(ya) ∩ M M D(yb) = ∅.
Similarly, for any two distinct nodes ra, rb ∈ {r1, · · · , rk2} we have: M M D(ra) ∩ M M D(rb) = ∅. This
implies that we can first compute the MM-Descendants of all yi and rj in O(n) time and further we can
compute all nodes yi that satisfy M M D(yi) ∩ M M D(rj) 6= ∅ for some rj, in O(n) time. Thus the template
reductions of Figure 5 run in O(n) time.
Case 4.2.2: p1 is a Q-node and n1 is its only essential child.
Since the only essential child of p1 is n1, all of its remaining children are subcliques that are present as
tails of n1. Thus each of these subcliques is a subset of U (n1). Now let Z and R be any two max-clique
descendants of xk0 and r1 respectively. By Lemma 8.2, R appears outside the subtree rooted at n1 (in T1)
and hence outside the subtree rooted at p1. By Lemma 6.0, we conclude that each descendant clique of p1
must contain Z ∩ R. Thus we have U (p1) ⊇ Z ∩ R ⊇ U (n2) = U (n1) ⊇ U (p1). Hence all of these sets must
be equal and hence we infer the following: Z ∩ R = U (n1) and hence U (xk0) ∩ U (r1) = U (n1). Further,
each subclique child of p1 must precisely be the clique U (n1).
Since we have eliminated adjacent duplicates from all Q-nodes, there can be at most one such subclique
in each tail of n1. Now if the subclique (U (n1)) appears on both sides of n1, then there is a unique way
of collapsing n1 (see Figure 6(top)). Otherwise we collapse n1 in such a way that U (n1) is present in the
tail of xk0 as shown in Figure 6(bottom). This is justified (i.e. it preserves compatibility between T1 and
T2) because U (n1) can be inserted into the right tail of xk0 in both T1 and T2. In other words, if T1 and
T2 are compatible, then there exists an intersection tree in which U (n1) is present in the right tail of xk0.
The template reductions of this case, clearly run in O(n) time.
Case 4.2.3: p1 is a Q-node and has more than one essential child.
Let y be an essential child of p1, such that all the nodes between n1 and y are subcliques. Without
loss of generality, we assume that y appears to the right of n1. We collapse n1, depending on whether
M M D(y) ∩ M M D(r1) is empty or not, as shown in Figure 7. Thus the template reduction runs in O(n)
time.
If M M D(y) ∩ M M D(r1) is non-empty, there exists a max-clique Y that is a descendant of both r1
and y. Now if T1 and T2 are compatible, then in the leaf ordering of any intersection tree, the max-clique
descendants of xk0 appear in between the max-clique descendants of l1 and Y . Thus we collapse the node
n1, by deleting n1, and reassigning p1 as the parent of all the children of n1. (Thus no essential node
appears between xk0 and y).
On the other hand if M M D(y) ∩ M M D(r1) is empty, we observe the following: In the leaf ordering
of any intersection tree TI no max-clique appears in between the max-clique descendants of xk0 and the
max-clique descendants of r1. Therefore, in this case we collapse n1, by reversing its children, deleting it,
15
T2
n2
x1
x2 xk0
R′
k0
L′
1
r1
r2
rk2
p1
n1
yk3
yi+1
T1
p1
y1
yk3
yi
yi+1
n1
l1
l2
lk1
x1
x2
xk0
L1
Rk0
l1
l2
lk1
x1
x2
xk0
L1
Rk0
y1
yi
T1
p1
yk3
y1
n1
p1
l1
l2
lk1
x1
x2
xk0
L1
Rk0
l1
l2
lk1
x1
x2
xk0
L1
Rk0
y1
yk3
Figure 5: Reduction template of T1 for Case 4.2.1. A node ya has horizontal stripes if M M D(ya) ∩
M M D(rb) 6= ∅ for some rb and no stripes otherwise.
16
T1
p1
U (n1)
U (n1)
n1
p1
l1
lk1
x1
S1
S2
x2
xk0
S ′
1
Sk0+1
U (n1)
S ′
1
l1
lk1
x1
S1
S2
x2
xk0
U (n1)
Sk0+1
T1
p1
U (n1)
n1
p1
l1
lk1
x1
S1
S2
x2
xk0
Sk0+1
S ′
1
l1
lk1
x1
S1
S2
x2
xk0
S ′
1
U (n1)
Sk0+1
Figure 6: Reduction templates of T1 for Case 4.2.2.
and reassigning p1 as the parent of all the children of n1. (Thus no essential node appears between l1 and y).
Case 4.3: U(n1) ⊇ U(n2) and C1 − X is empty
As before we have three cases depending on whether p1 is a P-node or a Q-node and whether p1 has more
than one essential child. In each of these cases, when X = 1, we need to first solve the alignment problem
(as a preprocessing step). Also when p1 is a Q-node, unlike in Case 4.2, both ways of collapsing n1 may
lead to a valid intersection tree.
Alignment Problem
Recall that when X = 1 (and C1 − X = ∅), the alignment may not be unique i.e. one of the following
might happen in the intersection tree TI.
1. Left and right tails of x1 (in TI) contain L1 + L′
1 and R1 + R′
1 respectively.
2. Left and right tails of x1 (in TI) contain ¯R1 + L′
1 and ¯L1 + R′
1 respectively.
If one of the merges in (1) or (2) is invalid, then there is only a single way of aligning the tails, otherwise
we show in the following Lemma that if T1 and T2 are compatible, then choosing either one of the two
alignments will work.
1, R′
1, R1 + R′
Lemma 9. Let U (n1) ⊇ U (n2), X = 1 and C1 − X be empty. Let L1, R1 be the left and right tails of
x1 in T1 and L′
1 be the left and right tails of x1 in T2. If both ways of alignment are mergable i.e. (a)
L1 + L′
1 are valid, then there exists an intersection tree TI
1 and R1 + R′
(of T1 and T2) with L1 + L′
1 contained in the left and right tails of x1 (respectively) if and
I with ¯R1 + L′
only if there exists an intersection tree T ′
1 contained in the left and right tails
of x1 (respectively).
1 are valid and (b) ¯R1 + L′
1 and ¯L1 + R′
1, ¯L1 + R′
17
T1
p1
Ep1
Lp
Rp
y Ep2
n1
T2
n2
l1
lk1
x1
xk0
Rn
Ln
x1 xk0
r1
rk2
M M D(y) ∩ M M Dr1) 6= ∅ ?
YES
NO
T ′
1
p1
T ′
1
p1
Ep1
Lp
l1
lk1
x1
xk0
Rn
Ln
Rp
y Ep2
¯Ep2
y
¯Rp
l1
lk1
x1
xk0
Ln
¯Ep1
¯Lp
Rn
Figure 7: Reduction template of T1 for Case 4.2.3
Proof. Let L, R be the left and right tails of x1 in an intersection tree TI . Each subclique S in L or R
appears as a subclique in T1 or T2. In particular we observe the following:
Property 1: If S is a subclique in L or R then in T1 or T2, S is present in a tail of x1 or in a tail of an
ancestor of x1.
Note that since n contains at least two children, L and R both cannot be empty. If one of them, say
L is empty then the last clique in R must be U (n1). If both L and R are non-empty then the intersection
of the first subclique of L1 with the last subclique of R1 is U (n1). In either case we observe that, since
L1 + L′
1 is either a superset of
U (n1) or a subset of U (n1). Similarly, since R1 + R′
1 are both valid (superclique orderings),
each subclique in R′
1 is either a superset of U (n1) or a subset of U (n1). Further for any ancestor na of n2,
U (na) ⊆ U (n2) ⊆ U (n1) and hence the tails of any such na would consist of subcliques that are subsets of
U (n1). Note that this condition also holds for any ancestor of n1 in T1.
1 are both valid (subclique orderings), each subclique in L′
1 and ¯R1 + L′
1 and ¯L1 + R′
By above conditions and (1) we infer that for any subclique S in L or R, S is either a superset of U (n1)
1 or
1 and R contains
1, then replacing L with L − L1 + ¯R1 and R with R − R1 + ¯L1 also results in a valid intersection
or a subset of U (n1). Furthermore, if S is a superset of U (n1) then it is present in one of L1, R1, L′
R′
R1 + R′
tree.
1. This implies that if there exists an intersection tree TI in which L contains L1 + L′
Note that the amortized cost of doing the mergability checks (a) and (b) of Lemma 9 (over all iterations
of the algorithm) is O(n · I) = O(n2). For the rest of the cases, we can assume that L1 is aligned with L2
18
and R1 is aligned with R2. In other words if T1 and T2 are compatible, then there exists an intersection
tree that contains L1 + L2 and R1 + R2 as the tails of x1.
Case 4.3.1: p1 is a P-node.
If C2 − X = ∅, then using the same argument as before (Lemma 6.0), we get U (p1) ⊆ U (n2). Hence we
replace Li and L′
i in T1 and T2 changing U (n1) to U (n1) ∩ U (n2) ⊇ U (p1), and we match n1
with n2.
i with Li + L′
Now we look at the case when C2 − X is non-empty. Let L = {l1, l2, · · · , lk1} be the set of essential
nodes appearing to the left of X and R = {r1, · · · , rk2} be the set of essential nodes appearing to the
right of X in T2. Let Y = {y1, · · · , yk3} be all the remaining child nodes of p1 other than n1. For all
i ∈ {1, · · · , k3}, if M M D(yi) ∩ M M D(lj) 6= ∅ for some j ∈ {1, · · · , k1}, then yi and lj both have a common
max-clique descendant say Y , and further in any leaf order of a common intersection tree L1 + L′
1 must
appear between Y and the descendants of x1.
Thus we group all yi such that M M D(yi) ∩ M M D(lj) 6= ∅ into a new P-node and add it to the
(immediate) left of L1 (see Figure 8). Similarly, we group all yi such that M M D(yi) ∩ M M D(rj) 6= ∅, for
some j ∈ {1, · · · , k2} into a new P-node and add it to the (immediate) right of Rk0.
Note that if for some y ∈ Y , there exists li and rj such that both M M D(y)∩M M D(li) and M M D(y)∩
M M D(rj) 6= ∅, then we can conclude that T1 and T2 are incompatible.
Also if L and R are both non-empty and for all y ∈ Y , M M D(y) doesn't intersect with any M M D(li)
for i ∈ {1, · · · , k1} and with any M M D(rj) for j ∈ {1, · · · , k2} then once again we conclude that T1 and
T2 are incompatible.
On the other hand if one of L or R is empty, say L, and M M D(yi) ∩ M M D(rj) is empty for all yi ∈ Y
and rj ∈ R, then the above template would not reduce T1. But then note that in any leaf-ordering of any
intersection tree, L1 + L′
1 should appear between the descendants of yi and x1 for all yi ∈ Y (because of
the constraints imposed by T1 and T2). Hence in this case we group all the nodes of Y into a P-node and
add it a child node of p1 to the (immediate) left of L1 as shown in Figure 9.
Note that since the MM-Descendents of any two sibling nodes are disjoint, both of the above templates
can be implemented in O(n) time.
Case 4.3.2: p1 is a Q-node and n1 is its only essential child.
Let Lp and Rp be the left and right tails of p1. Note that in this case all the siblings of n1 are subcliques
that are present in its tails. We have three subcases depending on how U (p1) intersects U (n2).
Suppose U (p1) properly intersects U (n2). We have U (p1) − U (n2) 6= ∅ and U (n2) − U (p1) 6= ∅. We first
claim that C2 − X is empty. Suppose not. Let Z be a max-clique descendant of a node in C2 − X and X1
be a max-clique descendant of x1. By Lemma 6.2, in T1, Z appears outside the subtree rooted at n1, and
hence outside the subtree rooted at p1. Thus using Lemma 6.0, we conclude that each descendant of p1
must contain all the vertices in Z ∩ X1 ⊇ U (n2). A contradiction. Hence C2 − X is empty.
Now by Lemma 6.6, there exists a subclique S1 6⊇ U (n2) such that S1 is the first clique of Lp or the
last clique of Rp. Similarly there exists a subclique S2 6⊇ U (p1) such that S1 is the first clique of L′
1 or the
last clique of R′
. Without loss of generality let S1 be the first clique of Lp and S2 be the last clique of
k0
R′
. Observe that S1 ⊇ U (p1) and S2 ⊇ U (n2). This implies that S1 and S2 cannot be in the same tail
k0
(of x1 or xk0) in any intersection tree of T1 and T2. Thus we reduce T1 by collapsing n1 i.e. by deleting n1,
changing the parent of child nodes of n1 to p1 and arranging the child nodes such that Lp appears to the
left of L1 and Rp appears to the right of Rk0. Clearly, this reduction can be done in O(n) time.
Now we have to deal with the case when either U (n2) ⊆ U (p1) or U (p1) ⊆ U (n2). Note that in any
1 appear in the left tail of x1 and the cliques of
intersection tree TI (of T1 and T2), the cliques of L1 + L′
19
T1
p1
y1
yi
yi+1
yj
yj+1
yk3
n1
T2
n2
x1
L1
x2
xk0
Rk0
l1
l2
lk1
x1
x2 xk0
R′
k0
L′
1
r1
r2
rk2
p1
n1
yk3
yj+1
x1
L1
x2
xk0
Rk0
y1
yi
yi+1
yi+1
Figure 8: First reduction template of T1 for Case 4.3.1. A node ya has vertical stripes if M M D(ya) ∩
M M D(lb) 6= ∅ for some lb, horizontal stripes if M M D(ya) ∩ M M D(rb) 6= ∅ for some rb and no stripes
otherwise.
T1
p1
yk3
y1
n1
T2
n2
x1
L1
x2
xk0
Rk0
x1
x2 xk0
R′
k0
L′
1
r1
r2
rk2
p1
x1
L1
x2
xk0
Rk0
y1
yk3
Figure 9: Second reduction template of T1 for Case 4.3.1. The stripes on the y nodes are defined as before
20
Rk0 + R′
appear in the right tail of xk0. Further either (a) the cliques of Lp appear in the left tail of x1
k0
and the cliques of Rp appear in the right tail of xk0 or (b) the cliques of Lp appear in the right tail of xk0
and the cliques of Rp appear in the left tail of x1. In the first case L1 + L′
+ Rp are
+ ¯Lp are both valid. If neither of these
both valid and in the second case L1 + L′
is valid then we conclude that T1 and T2 are incompatible. If exactly one of the above merges is valid,
then there is a unique way of collapsing n1. When both of the above merge pairs are valid, we use the
reduction template shown in Figure 10. The justification (given below) depends on whether U (n2) ⊆ U (p1)
or U (p1) ⊆ U (n2).
1 + ¯Rp and Rk0 + R′
k0
1 + Lp and Rk0 + R′
k0
Let U (n2) ⊆ U (p1). Note that by Lemma 6.6, the intersection of the universal nodes of the first and
+ ¯Lp are
last child nodes of p1 is U (p1). Hence if L1 + L′
all valid then any subclique in L′
is either a superset or a subset of U (p1). Thus in any intersection
tree TI , any subclique S in the left tail of x1 or the right tail of xk0 is either a superset or a subset of
U (p1). Further if S ⊇ U (p1), then S must appear in one of {L′
, Lp, Rp, L1, Rk0}. This implies that
an intersection tree satisfying condition (a) exists if and only if an intersection tree satisfying condition (b)
exists. This justifies the use of our template in Figure 10, for reducing T1.
1 + ¯Rp and Rk0 + R′
k0
1 + Lp, Rk0 + R′
k0
+ Rp, L1 + L′
1 or R′
k0
1, R′
k0
Similarly, if U (p1) ⊆ U (n2), we infer that any clique in Lp or Rp is either a subset of U (n2) or a superset
of U (n2). This in turn implies that in TI, any subclique S in the left tail of x1 or the right tail of x1, is either
a subset or a superset of U (n2). Further, if S ⊇ U (n2) then it must appear in {L1, Rk0 , L′
, Lp, Rp}.
This implies that an intersection tree satisfying condition (a) exists if and only if an intersection tree
satisfying (b) exists. This justifies the use of our template in Figure 10, for reducing T1.
1, R′
k0
We now show that the amortized cost of executing the reduction template in Figure 10, over all instances
of the algorithm takes O(n2) time. Note that we use the same template for Case 4.3.3 when C2 − X is
empty. It is enough to show that the amortized time of all the mergability checks: (whether L′
1 + Lp and
R′
k0
) be the (consecutive) subsequences of L′
) contains U (p1) but not U (n1). c(L′
1 and R′
k0
1) and c(R′
k0
1) and c(R′
k0
(respectively) such that each
) are said to be the core tails
+ Rp are both valid) take O(n2) time.
Let c(L′
1) and c(R′
k0
subclique in c(L′
of n2.
Similarly let c(Lp) and c(Rp) be the (consecutive) subsequences of Lp and Rp (respectively) such that
each subclique in c(Lp) and c(Rp) contains U (n2). c(Lp) and c(Rp) are said to be the core tails of p1.
Note that the core tails are only defined for p1 and n2, for the current case and Case 4.3.3, when C2 − X
is empty. We define the core tails of all other nodes to be empty. Observe that when n1 is collapsed, the
(new) core tails of any node in T1 (resp. T2) are disjoint from the core tails of p1 (resp. n2) before the
collapse.
We observe that checking the validity of L′
Similarly, checking the validity of R′
k0
1 + Lp reduces to checking the validity of c(L′
) + c(Rp).
+ Rp reduces to checking the validity of c(R′
k0
1) + c(Lp).
Now computing the cores over all executions of this template , takes O(n · I) = O(n2) amortized time.
Also, computing the mergability of the cores, over all executions of the template takes Pi(mi + ti)I,
where mi, ti are the number of subcliques in the cores of n2 and p1 (respectively), in the ith execution
of the template. Since Pi mi = O(n) and Pi ti = O(n), the total running time of template 10 over all
executions is O(n2).
Case 4.3.3: p1 is a Q-node with more than one essential child.
Now let y be an essential child of p1, such that all the nodes between n1 and y are subcliques. Without
loss of generality, we assume that y appears to the right of n1.
We first consider the subcase when C2 − X is empty.
In this case observe that all the max-clique
descendants of y appear outside the subtree rooted at n1 in T1. Applying Lemma 6.0 on a max-clique
21
T1
p1
Lp
Rp
n1
T2
n2
x1
x2
xk0
L1
Rk0
x1
x2
xk0
L′
1
E2
R′
k0
L′
1 + Lp and R′
k0
+ Rp are valid ?
YES
NO
p1
p1
x1
x2
xk0
Lp
L1
Rk0
Rp
x1
x2
xk0
¯Rp
L1
Rk0
¯Lp
Figure 10: Reduction template of T1 for Case 4.3.2. Note that E2 denotes the (possibly empty) sequence
of children of n2 that appear after R′
k0
.
22
T1
p1
Lp
Rp
y Ep1
n1
T2
n2
En1
x1
L1
x2
xk0
Rk0
x1
x2 xk0
R′
k0
L′
1
r1
r2
rk2
M M D(y) ∩ M M D(r1) 6= ∅ ?
YES
NO
p1
p1
Lp
En1
x1L1
x2
xk0
Rk0
Rp
y Ep1
¯Ep1
y
En1
x1L1
¯Rp
x2
xk0
Rk0
¯Lp
Figure 11: Reduction template of T1 for Case 4.3.3.
descendant of y and a max-clique descendant of x1, we infer that each descendant clique of n2 must
In other words we get U (p1) ⊆ U (n2). Now the template (and the
contain all the vertices in U (p1).
argument) in this case is analogous to case 4.3.2, when U (p1) ⊆ U (n2) (Figure 10).
Now suppose C2 − X is non empty. Let r1 be the first essential child to the right of xk0 in T2.
We use the templates in Figure 11, depending on whether M M D(y) ∩ M M D(r1) is empty or not.
If
M M D(y) ∩ M M D(r1) is non-empty, then by Lemma 6.2, there exists a max-clique Y that is a descendant
of both r1 and y. Now if T1 and T2 are compatible, then in any leaf ordering of an intersection tree TI, the
subcliques of Rk0 + R′
appear in between the descendants of xk and Y (because of T2). This justifies the
k0
reduction template of T1 in Figure 11.
On the other hand, if M M D(r1) ∩ M M D(y) = ∅, then by the constraints of T2, in any leaf ordering of
TI, the subcliques of Rk0 + R′
appear between the descendants of xk0 and l1 and further no max-clique
k0
appears between them. This justifies the reduction template of T1 in Figure 11. Moreover the template
reduction takes O(n) time.
Run time of the Algorithm
In this Section, we show that the run time of our algorithm is O(n2log n), where n is the total number of
vertices in G1 ∪ G2.
Observe that reducing T1 [resp. T2] decreases the number of leaf orderings of T1 [resp. T2] by at least
half. Moreover the total number of nodes in T1 and T2 is at most n. Thus the number of leaf orderings of
T1 and T2 is at most n! and hence the algorithm requires at most nlog n reductions.
We begin by showing that selecting the nodes n1 and n2 takes O(n) time in any iteration. We first note
that computing the number of MM-Descendants for all the nodes takes O(n) time (they can be computed
in a bottom-up fashion). With each node x, we store U (x) and the cardinality of M M D(x).
23
Recall that as we go down the tree the universal sets increase and the MM-Descendants sets decrease.
Thus n1 must have maximal depth among all unmatched essential nodes. Hence we can select n1 in O(n)
time by looking at the unmatched essential nodes of maximal depth in T1 and T2, and selecting a node
with the greatest universal set size and the least number of MM-Descendants in that order. Note that by
property 1 of Lemma 8, the MM-descendants of n1 are same as the essential child nodes of n1. Now we can
select n2 from the other tree T2 in O(n) time as follows: Let S be the set of [matched] essential children of
n1 and S′ be the corresponding set of matched nodes in T2. Let p(S′) be the set of parent nodes of nodes
in S′. We select n2 to be a node of maximum depth among p(S′).
Also recall that at the high-level our algorithm has 4 cases depending on whether n1, n2 are P-nodes
or Q-nodes. We showed that each step in cases 1,2 or 3 takes O(n) time. For case 4, we showed that
each of reduction steps, excluding the mergability checks in Cases 4.3.2 and 4.3.3 take O(n) time. We also
showed that the mergability checks of Cases 4.3.2 and 4.3.3 take O(n2) amortized time over all steps of the
algorithm. Further, in the beginning of Case 4, we showed that the matching steps, which involve inserting
the subcliques of one tree into the other take at most O(n2) amortized time. Thus the total time taken
by our algorithm is O(n2log n + n2 + n2) = O(n2log n). At each node y of T1 (resp. T2) we explicitly store
the set U (y) and the cardinality of M M D(y). Since the number of internal nodes is less than the number
of leaf nodes, this additional storage still takes O(n + m). Thus the space complexity of our algorithm is
O(n + m).
5 Open Problem
Simultaneous graphs can be generalized in a natural way to more than two graphs: when G1 = (V1, E1), G2 =
(V2, E2), · · · , Gk = (Vk, Ek) are k graphs in class C, sharing a vertex set I and its induced edges i.e. Vi ∩Vj =
I for all i, j ∈ {1, · · · , k}. In this version of the problem the set of optional edges induces a complete k-
partite graph and hence this also generalizes probe graphs. This generalized version can be solved in
polynomial time for comparability and permutation graphs [9]. We conjecture that it can be solved in
polynomial time for interval graphs.
References
[1] Berry, A., Golumbic, M.C., Lipshteyn, M. Recognizing chordal probe graphs and cycle-bicolorable
graphs. SIAM J. Discret. Math., 21(3):573–591, 2007.
[2] Booth, K.S and Lueker, G.S. Testing for the consecutive ones property, interval graphs, and graph
planarity using PQ-tree algorithms. J. Comput. System Sci., 13:335–379, 1976.
[3] Brass, P., Cenek, E., Duncan, C.A., Efrat, A., Erten, C., Ismailescu, D.P., Kobourov, S.G., Lubiw,
A., Mitchell J.S.B. On simultaneous planar graph embeddings. Comput. Geom. Theory Appl., 36(2),
117-130 (2007).
[4] Estrella-Balderrama, A., Gassner, E., Junger, M., Percan, M., Schaefer, M., Schulz, M., Simultaneous
geometric graph embeddings. GD 2007. LNCS, vol. 4875, pp.280-290.
[5] Chandler, D.B., Chang, M., Kloks, T., Liu, J., and Peng, S. Partitioned probe comparability graphs.
Theor. Comput. Sci., 396(1-3):212–222, 2008.
[6] Golumbic, M.C. Algorithmic Graph Theory And Perfect Graphs. Academic Press, New York, 1980.
24
[7] Golumbic, M.C., Kaplan, H., and Shamir, R. Graph sandwich problems. J. Algorithms, 19(3):449–473,
1995.
[8] Golumbic M.C, and Lipshteyn, M. Chordal probe graphs. Discrete Appl. Math., 143(1-3):221–237,
2004.
[9] Jampani, K.R., Lubiw, A. The simultaneous representation problem for chordal, comparability and
permutation graphs. In WADS '09, pages 387–398, 2009.
[10] Johnson, J.L., and Spinrad, J.P. A polynomial time recognition algorithm for probe interval graphs.
In SODA '01, pages 477–486, Philadelphia, PA, USA, 2001.
[11] McConnell, R.M., and Nussbaum, Y. Linear-time recognition of probe interval graphs. In ESA 09,
pages 349-360.
[12] McConnell, R.M., and Spinrad, J.P. Construction of probe interval models.
In SODA '02, pages
866–875, Philadelphia, PA, USA, 2002.
[13] McMorris, F. R., Wang, C., Zhang, P. On probe interval graphs. Discrete Appl. Math., 88(1-3):315–
324, 1998.
[14] Tanenbaum, Paul J. Simultaneous intersection representation of pairs of graphs. J. Graph Theory.,
32(2):171–190, 1999.
25
|
1210.1890 | 2 | 1210 | 2013-03-04T16:28:24 | Local Search is Better than Random Assignment for Bounded Occurrence Ordering k-CSPs | [
"cs.DS"
] | We prove that the Bounded Occurrence Ordering k-CSP Problem is not approximation resistant. We give a very simple local search algorithm that always performs better than the random assignment algorithm. Specifically, the expected value of the solution returned by the algorithm is at least Alg > Avg + a(B,k) (Opt - Avg), where "Opt" is the value of the optimal solution; "Avg" is the expected value of the random solution; and a(B,k)=Omega_k(B^{-(k+O(1))} is a parameter depending only on "k" (the arity of the CSP) and "B" (the maximum number of times each variable is used in constraints). The question whether bounded occurrence ordering k-CSPs are approximation resistant was raised by Guruswami and Zhou (APPROX 2012) who recently showed that bounded occurrence 3-CSPs and "monotone" k-CSPs admit a non-trivial approximation. | cs.DS | cs |
Local Search is Better than Random Assignment
for Bounded Occurrence Ordering k-CSPs∗
Konstantin Makarychev
Microsoft Research, Redmond, WA 98052
Abstract
We prove that the Bounded Occurrence Ordering k-CSP Problem is not approximation
resistant. We give a very simple local search algorithm that always performs better than the
random assignment algorithm (unless, the number of satisfied constraints does not depend on
the ordering). Specifically, the expected value of the solution returned by the algorithm is at
least
Alg ≥ Avg + α(B, k)(Opt − Avg),
where Opt is the value of the optimal solution; Avg is the expected value of the random solution;
and α(B, k) = Ωk(B−(k+O(1))) is a parameter depending only on k (the arity of the CSP) and
B (the maximum number of times each variable is used in constraints).
The question whether bounded occurrence ordering k-CSPs are approximation resistant was
raised by Guruswami and Zhou (2012), who recently showed that bounded occurrence 3-CSPs
and "monotone" k-CSPs admit a non-trivial approximation.
1
Introduction
In this work, we give a very simple local search algorithm for ordering constraints
Overview.
satisfaction problems that works better than the random assignment for those instances of the
ordering k-CSP problem, where each variable is used only a bounded number of times. To motivate
the study of the problem, we first overview some known results for regular constraint satisfaction
problems.
An instance of a constraint satisfaction problem consists of a set of variables V = {x1, . . . , xn}
taking values in a domain D and a set of constraints C. Each constraint C ∈ C is a function from
Dk to R+ applied to k variables from V . Given an instance of a CSP, our goal is to assign values
to the variables to maximize the total payoff of all constraints:
max
x1,...,xn∈DnXC∈C
C(x1, . . . , xn).
In fact, C may depend on at
Note, that we write C(x1, . . . , xn) just to simplify the notation.
most k variables. The parameter k is called the arity of the CSP. In specific CSP problems,
constraints C come from a specific family of constraints. For example, in Max Cut, the domain
is D = {−1, 1}, and all constraints have the form C(x1, . . . , xn) = 1(xi 6= xj); in Max 3LIN-2,
the domain D = {0, 1}, and all constraints have the form C(x1, . . . , xn) = 1(xi ⊕ xj ⊕ xl = 0) or
C(x1, . . . , xn) = 1(xi ⊕ xj ⊕ xl = 1).
∗The conference version of this paper appeared at STACS 2013.
1
Various approximation algorithms have been designed for CSPs. The most basic among them,
the "trivial" probabilistic algorithm simply assigns random values to the variables. It turns out,
however, that in some cases this algorithm is essentially optimal. Hastad (1997) showed that for
some CSPs e.g., 3LIN-2 and E3-SAT, beating the approximation ratio of the random assignment
algorithm (by any positive constant ε) is NP-hard. Such problems are called approximation re-
sistant. That is, a constraint satisfaction problem is approximation resistant, if for every positive
ε > 0, it is NP-hard to find a (Atrivial + ε) approximation, where Atrivial is the approximation ratio
of the random assignment algorithm. If there exists an algorithm with the approximation ratio
(Atrivial + ε) for some positive ε, we say that the problem admits a non-trivial approximation. It is
still not known which constraint satisfaction problems are approximation resistant and which admit
a non-trivial approximation. This is an active research question in approximation algorithms.
Suppose now that in our instance of k-CSP, each variable is used by at most B constraints. (For
example, for Max Cut, this means that the maximum degree of the graph is bounded by B.) Hastad
(2000) proved that such instances (which we call B-bounded occurrence k-CSPs) admit a non-trivial
approximation. Let Opt denote the value of the optimal solution; Avg denote the expected value
of the random assignment; and Alg denote the expected value returned by the algorithm. Hastad
(2000) showed that there exists an approximation algorithm such that1
Alg ≥ Avg +
Opt − Avg
Ok(B)
.
Here the hidden constant in Ok(·) may depend on k. Trevisan (2001) showed a hardness of approx-
imation lower bound of Avg + (Opt − Avg)/(Ωk(√B)).
In this work, we study ordering constraints satisfaction problems. A classical example of an
ordering k-CSP is the Maximum Acyclic Subgraph problem. Given a directed graph G = (V, E),
the goal is to find an ordering of the vertices π : V → {1, . . . , n} (π is a bijection; n = V ), so
as to maximize the number of forward edges. In this case, the edges of the graph are constraints
on the ordering π. An edge (u, v) corresponds to the constraint π(u) < π(v). Another example
is the Betweenness problem. We are given a set of vertices V and a set of constraints {Cu,v,w}.
Each Cu,v,w is defined as follows: Cu,v,w(π) = 1, if u < v < w or w < v < u, and Cu,v,w(π) = 0,
otherwise. The goal again is to find an ordering satisfying the maximum number of constraints.
More generally, in an ordering k-CSP, each constraint C is a function of the ordering that
depends only on the relative order of k vertices. The goal is given a set of vertices V and a set of
constraints C, to find an ordering π : V → [n] to maximize the total value of all constraints:
max
π:V →[n]XC∈C
C(π).
Here π is a bijection and n = V . If all constraints take values {0, 1}, then the objective is simply
to maximize the number of satisfied constraints. Note, that an ordering k-CSP is not a k-CSP.
Surprisingly, we know more about ordering CSPs than about regular CSPs. Guruswami,
Hastad, Manokaran, Raghavendra, and Charikar (2011) showed that every ordering CSP prob-
lem is approximation resistant assuming the Unique Games Conjecture (special cases of this result
were obtained by Guruswami et al. (2008) and Charikar et al. (2009)). On the positive side,
1The quantity ("value of the solution" − Avg) is called the advantage over random. The algorithm of Hastad
(2000) is Ok(B) approximation algorithm for the advantage over random:
Alg − Avg ≥
Opt − Avg
Ok(B)
.
2
Berger and Shor (1990) showed that bounded degree Maximum Acyclic Subgraph, and thus every
bounded occurrence ordering 2CSP, admits a non-trivial approximation. Their result implies that
Alg ≥ Avg + (Opt − Avg)/O(√B). Charikar, Makarychev, and Makarychev (2007) showed that
a slight advantage over the random assignment algorithm can be also achieved for instances of
Maximum Acyclic Subgraph (Alg ≥ Avg + (Opt− Avg)/O(log n)) whose maximum degree is not
bounded. Gutin, van Iersel, Mnich, and Yeo (2012) showed that the "advantage over the random
assignment" for ordering 3CSPs is fixed -- parameter tractable (we refer the reader to the paper for
definitions and more details). Finally, Guruswami and Zhou (2012) proved that all bounded oc-
currence ordering 3CSPs admit a non-trivial approximation (Alg ≥ Avg + (Opt − Avg)/Ok(B)).
They also proved that there exists an approximation algorithm for monotone k-CSP (i.e., ordering
CSPs, where all constraints are of the form π(ui1) < π(ui2) < ··· < π(uik )) with approximation
ratio 1/k! + 1/Ok(B).
Our results. We show that a very simple randomized local search algorithm finds a solution
of expected value:
Alg ≥ Avg +
Opt − Avg
Ok(Bk+2)
.
(1)
This algorithm works for every bounded occurrence ordering k-CSP. Consequently, all bounded
occurrence ordering k-CSPs admit a non-trivial approximation. The running time of the algorithm
is O(n log n). We do not know whether the dependence on B is optimal. However, the result of
Trevisan (2001) implies a hardness of approximation upper bound of Alg+(Opt−Avg)/Ωk(√B)2.
Techniques. Our algorithm works as follows: first, it permutes all vertices in a random order.
Then, n times, it picks a random vertex and moves it to the optimal position without changing the
positions of other vertices. We give an elementary proof that this algorithm performs better than
the random assignment. However, the bound we get is exponentially small in B.
Then, we improve this bound. Roughly speaking, instead of the original problem we consider the
"D-ordering" problem, where the algorithm puts vertices in D ≈ Bk buckets (possibly, in a clever
way), then it randomly permutes vertices in each of the buckets, and finally outputs vertices in
the first bucket, second bucket, etc. This idea was previously used by Charikar, Makarychev,
and Makarychev (2007), Guruswami, Hastad, Manokaran, Raghavendra, and Charikar (2011),
Gutin, van Iersel, Mnich, and Yeo (2012) and Guruswami and Zhou (2012). The transition to "D-
orderings" allows us to represent the payoff function as a Fourier series with relatively few terms.
We prove that the L1 weight of all coefficients of the payoff function is at least Avg+Ωk(Opt−Avg)
(Note, that the optimal value of the "D-ordering" problem may be less than Opt). Then we show
that (a) for each vertex we can find one "heavy" Fourier coefficient fS; and (b) when the original
local search algorithm moves a vertex it increases the value of the solution in expectation by at
least Ωk( fS/B). This concludes the proof.
Correction. In the preliminary version of the paper that appeared at arXiv, we proved the
main result of the paper, Theorem 3.1. We also gave an alternative, more complicated algorithm in
the Appendix. We erroneously claimed that the performance guarantee of the alternative algorithm
is slightly better than (1). This is not the case. So the best bound known to the author is (1).
2Every k-CSP can be encoded by an ordering 2k-CSP by replacing every boolean variable x with two variables
u←
x and u→
x , and letting x = 1 if and only if π(u←
x ) < π(u→
x ).
3
2 Preliminaries
An instance of an ordering k-CSP problem (V,C) consists of a set of vertices V of size n, and a set
of constraints C. An ordering of vertices π : V → {1, . . . , n} is a bijection from V to {1, . . . , n}.
Each constraint C ∈ C is a function from the set of all ordering SV = {π : V → {1, . . . , n}} to R+
that depends on the relative order of at most k vertices. That is, for every C there exists a set
TC ⊂ V of size at most k such that if for two orderings π1 and π2, π1(u) < π1(v) ⇔ π2(u) < π2(v)
for all u, v ∈ TC , then C(π1) = C(π2). The value of an ordering π equals
value(π,C) = XC∈C
C(π).
We will sometimes write value((u1, . . . un),C) to denote the value(π,C) for π : ui 7→ i. We denote
the optimal value of the problem by Opt(V,C) ≡ maxπ∈SV value(π,C), the average value -- the
value returned by the random assignment algorithm -- by Avg(V,C) = 1/n! Pπ∈SV value(π,C).
3 Algorithm
We now present the algorithm.
Randomized Local Search Algorithm
Input: a set of vertices V , and a set of constraints C.
Output: an ordering of vertices (v1, . . . , vn).
1. Randomly permute all vertices.
2. Repeat n times:
• Pick a random vertex u in V .
• Remove u from the ordering and insert it at a new location to maximize the payoff.
I.e., if v1, . . . , vn−1 is the current ordering of all vertices but the vertex u, then find a
location i that maximizes the value(v1, . . . , vi−1, u, vi+1, . . . vn−1,C), and put u in the
i-th position.
3. Return the obtained ordering.
Theorem 3.1. Given an instance (V,C) of a B-bounded occurrence ordering k-CSP problem, the
Randomized Local Search Algorithm returns a solution πAlg of expected value
E value(πAlg,C) ≥ Avg(V,C) +
Opt(V,C) − Avg(V,C)
Ok(Bk+2)
.
(2)
Remark 3.1. In fact, our proof implies a slightly stronger bound: The second term on the right
hand side of the inequality (2) can be replaced with (Opt(V,C) − Worst(V,C))/Ok(Bk+2), where
Worst(V,C) is the value of the worst possible solution.
4
Proof. I. We first show using an elementary argument that
E value(πAlg,C) ≥ Avg(V,C) + α(B, k)(Opt(V,C) − Avg(V,C)),
for some function α(B, k) depending only on B and k. This immediately implies that every bounded
occurrence ordering k-CSP admits a non-trivial approximation. Then, using a slightly more involved
argument we prove the bound (2).
Observe, that the expected value of the solution after step 1 is exactly equal to Avg(V,C).
So we need to estimate how much local moves at step 2 improve the solution. Let ∆u be the
maximum possible increase in the value of an ordering π, when we move u to another position. In
other words, ∆u = maxπ+,π−(value(π+,C) − value(π−,C)), where the orderings π+ and π− differ
only in the position of the vertex u. Let π∗ be the optimal ordering, and π∗ be the worst possible
ordering. We can transition from π∗ to π∗ by moving every vertex u at most once. Thus,
∆u ≥ value(π∗,C) − value(π∗,C) = Opt(V,C) − Worst(V,C) ≥ Opt(V,C) − Avg(V,C).
Xu∈V
Now, our goal is to show that when the algorithm moves a vertex u, the value of the solution
increases in expectation by at least α(B, k)∆u for some function α depending only on B and k.
Fix a vertex u. Let π+ and π− be the orderings that differ only in the position of the vertex u
such that ∆u = value(π+,C) − value(π−,C). It may happen that the random permutation chosen
by the algorithm at step 1 is π−, and u is chosen first among all vertices in V at step 2. In this case,
the algorithm can obtain the permutation π+ by moving u, and thus it can increase the value of
the solution by ∆u. However, the probability of such event is negligible. It is 1/n · 1/n!. The main
observation is that the increase in the value of the ordering, when we move u, depends only on the
order of the neighbors of u i.e., those vertices that share at least one common constraint C ∈ C
with u (including u itself). We denote the set of neighbors by N (u). Since each vertex participates
in at most B constraints, and every constraint depends on at most k variables, N (u) ≤ kB.
Consider an execution of the algorithm. We say that u is fresh if u was chosen at least once in
the "repeat" loop of the algorithm, and none of the neighbors were chosen before u was chosen the
first time. The probability that a variable u is fresh is at least 1/2N (u)−1. Indeed, the probability
that at least one vertex in N (u) is chosen is 1 − (1 − N (u)/n)n > 1 − 1/e; the probability that the
first vertex chosen in N (u) is u is 1/N (u) (since all vertices in N (u) have the same probability of
being chosen first).
If u is fresh, then when it is chosen, its neighbors are located in a random order (since none
of them was moved by the algorithm). Thus, with probability 1/N (u)! ≥ 1/(kB)!, the order of
neighbors of u is the same as in π−. Then, by moving u we can increase the value of the ordering
by ∆u.
Therefore, when the algorithm moves the vertex u, the value of the ordering increases in expec-
tation by at least
Pr(u is fresh) · Pr(N (u) is ordered as π− after step 1) · ∆u =
∆u
2N (u)N (u)! ≥
∆u
kB (kB)!
.
This finishes the elementary proof that a positive α(B, k) exists.
II. We now improve the lower bound on α(B, k). We show that for a fresh variable u, the
value of the ordering increases in expectation by at least Ωk(B−(k+1))∆u, and thus (2) holds. Let
L = ⌈log2(N (u) + 1)⌉ and D = 2L. Consider D buckets [D] = {1, . . . , D}. For every mapping x of
the vertices to the buckets v 7→ xv ∈ [D], we define a distribution Ux on orderings of V . A random
5
ordering from Ux is generated as follows: put each vertex v in the bucket xv; then randomly and
uniformly permute vertices in each bucket; and finally output vertices in the first bucket, second
bucket, etc (according to their order in those buckets). Let Cu be the set of constraints that
depend on the vertex u. Since every variable participates in at most B constraints, Cu ≤ B. Let
f (x) be the expected total value of constraints in Cu on a random ordering π sampled from the
distribution Ux:
f (x) = Eπ∼Uxh XC∈Cu
C(π)i.
Since the number of buckets D is greater than or equals to N (u)+1, we may put every vertex in
N (u) in its own bucket and keep one bucket empty. Let π+ and π− be the orderings as in part I of the
proof: π+ and π− differ only in the position of the vertex u, and value(π+,C)− value(π−,C) = ∆u.
Consider mappings x+ : V → [D] and x− : V → [D] that put only one vertex from N (u) in every
v for every v 6= u, and x+ orders vertices in N (u) according to π+, x−
bucket and such that x+
orders vertices in N (u) according to π−. For example, if π+ arranges vertices in the order (a, b, u, c),
and π− arranges vertices in the order (a, u, b, c), then x+ = (a 7→ 1,∗, b 7→ 3, u 7→ 4, c 7→ 5) and
x− = (a 7→ 1, u 7→ 2, b 7→ 3,∗, c 7→ 5). Since the order of all vertices in N (u) is fixed by x+ and x−,
we have f (x+) = value(π+,Cu) and f (x−) = value(π−,Cu). Then
v = x−
f (x+) − f (x−) = value(π+,Cu) − value(π−,Cu)
= value(π+,C) − value(π−,C) = ∆u.
We now use Theorem 4.1, which we prove in Section 4. Let Xv (for v ∈ V ) be independent random
variables uniformly distributed in [D]. By Theorem 4.1,
E[max
xu∈D
f (xu,{Xv}v6=u) − f (Xu,{Xv}v6=u)] ≥ Ωk(B−1D−k)(f (x+) − f (x−))
= Ωk(B−(k+1))∆u.
Here, (xu,{Xv}v6=u) denotes the mapping u 7→ xu and v 7→ Xv for v 6= u; and (Xu,{Xv}v6=u)
denotes the mapping v 7→ Xv for all v.
Observe, that when we sample random variables Xv, and then sample π according to UX , we
get a random uniform ordering π of all vertices in V . Thus,
E[f (Xu,{Xv}v6=u)] = Eπ∈SVh XC∈Cu
C(π)i = Eπ[value(π,Cu)].
Similarly, when we sample random variables Xv, set xu = argmaxxu∈D f (xu,{Xv}v6=u), and then
sample π′ according to U(xu,{Xv}v6=u), we get a random uniform ordering of all vertices except for
the vertex u. Denote by LS(π, u) the ordering obtained from the ordering π by moving the vertex
u to the optimal position. It is easy to see that if π is a random uniform ordering, then LS(π, u)
has the same distribution as LS(π′, u), since the new optimal position of u depends only on the
relative order of other vertices v, and not on the old position of u. Hence,
E[max
xu∈D
f (xu,{Xv}v6=u)] ≡ Eπ′[value(π′,Cu)]
≤ Eπ′[value(LS(π′, u),Cu)]
= Eπ[value(LS(π, u),Cu)].
Hence,
Eπ[value(LS(π, u),C) − value(π, u,C)] = Eπ[value(LS(π, u),Cu) − value(π, u,Cu)]
≥ Ωk(B−(k+1))∆u.
6
4 Theorem 4.1
Theorem 4.1. Let D be a set of size 2L (for some L). Consider a function f : Dn+1 → R that
can be represented as a sum of T functions ft : Dn+1 → R:
f (x0, x1, . . . , xn) =
ft(x0, x1, . . . , xn)
T
Xt=1
such that each function ft depends on at most k variables xu. Here, x0, . . . , xn ∈ D. Then, the
following inequality holds for random variables X0, . . . , Xn uniformly and independently distributed
in D:
E[max
x∈D
f (x, X1, . . . , Xn) − f (X0, X1, . . . , Xn)] ≥
Ωk(T −1D−k)
max
x+,x−,x1,...,xn∈D
(f (x+, x1, . . . , xn) − f (x−, x1, . . . , xn)).
Remark 4.1. The variable x0 corresponds to xu from the proof of Theorem 3.1. The functions
ft(x) are equal to Eπ∼UxCt(π), where Ct is the t-th constraint from Cu.
Proof. Without loss of generality we assume that elements of D are vertices of the boolean cube
{−1, 1}L. We denote the i-th coordinate of x ∈ D by x(i). We now treat f as a function of (n + 1)L
boolean variables xu(i). We write the Fourier series of the function f . The Fourier basis consists
of functions
χS(x0, . . . , xn) = Y(u,i)∈S
xu(i),
which are called characters. Each index S ⊂ {0, 1, . . . , n} × {1, . . . , L} corresponds to the set of
boolean variables {xu(i) : (u, i) ∈ S}. Note, that χ∅(x0, . . . , xn) = 1. The Fourier coefficients of f
equal
fS = E[f (X0, . . . , Xn) χS(X0, . . . , Xn)],
and the function f equals
f (x0, . . . , xn) =XS
fS χS(x0, . . . , xn).
Remark 4.2. In the proof, we only use the very basic facts about the Fourier transform. The main
property we need is that the characters form an orthonormal basis, that is,
E[χS1(X0, . . . , Xn)χS2(X0, . . . , Xn)] =(1,
0,
if S1 = S2;
if S1 6= S2.
Particularly, for S 6= ∅,
E[χS(X0, . . . , Xn)] = E[χS(X0, . . . , Xn)χ∅(X0, . . . , Xn)] = 0.
We will also need the following property: if f does not depend on the variable xu(i), then all Fourier
coefficients fS with (u, i) ∈ S are equal to 0.
Here is a brief overview of the proof: We will show that the L1 weight of Fourier coefficients
of f is at least f (x+, x1, . . . , xn) − f (x−, x1, . . . , xn), and the weight of one of the coefficients fS∗
is at least Ω(T −1D−k(f (x+, x1, . . . , xn) − f (x−, x1, . . . , xn))). Consequently, if we flip a single bit
7
X0(i∗) in X0 to make the term fS∗χS∗(X ′
of f by fS∗.
Observe, that since each function ft depends on at most kL boolean variables, it has at most
2kL = Dk nonzero Fourier coefficients. Thus, f has at most T Dk nonzero Fourier coefficients fS.
0, X1, . . . , Xn) positive, we will increase the expected value
Pick x+, x−, x∗
1, . . . , x∗
n that maximize f (x+, x∗
1, . . . , x∗
1, . . . , x∗
n). We have
n) − f (x−, x∗
1, . . . , x∗
fS(χS(x+, x∗
n) − χS(x−, x∗
1, . . . , x∗
n)).
f (x+, x∗
1, . . . , x∗
n) − f (x−, x∗
1, . . . , x∗
n) =XS
If S does not contain pairs (0, i) corresponding to the bits of the variable x0, then the character
χS(x0, x1, . . . , xn) does not depend on x0, and χS(x+, x∗
n) = 0, hence
1, . . . , x∗
1, . . . , x∗
n) − χS(x−, x∗
f (x+, x∗
1, . . . , x∗
n) − f (x−, x∗
= XS:∃i s.t. (0,i)∈S
n) =
1, . . . , x∗
fS · (χS(x+, x∗
1, . . . , x∗
n) − χS(x−, x∗
1, . . . , x∗
n)) ≤ 2 XS:∃i s.t. (0,i)∈S
fS.
Pick a character fS∗ with maximum absolute value and pick one of the elements (0, i∗) ∈ S∗. Since
the number of nonzero characters fS is at most T Dk,
n) − f (x−, x∗
2T Dk
Let σ = sgn( fS∗). Define a new random variable X ′
0,
fS∗ ≥
1, . . . , x∗
f (x+, x∗
1, . . . , x∗
n)
.
X ′
0(i) =(X0(i),
σ χS∗(X0, . . . , Xn) X0(i),
for i 6= i∗;
for i = i∗.
Consider a character χS.
If (0, i∗) /∈ S, then χS does not depend on the bit x0(i∗), hence
E[χS(X ′
0, X1, . . . , Xn)] = E[χS(X0, X1, . . . , Xn)]. On the other hand, if (0, i∗) ∈ S, then
0, X1, . . . , Xn)] = E(cid:2)
χS(X0, X1, . . . , Xn)(cid:3) =
0(i∗)
X ′
X0(i∗)
E[χS(X ′
E[σχS∗(X0, X1, . . . , Xn)χS(X0, X1, . . . , Xn)] =(0,
σ,
if S 6= S∗;
if S = S∗.
The last equality holds because characters χS form an orthonormal basis. Therefore,
E[f (X ′
0, X1, . . . , Xn) − f (X0, X1, . . . , Xn)] = σ fS∗ = fS∗.
We get
E[max
x∈D
f (x, X1, . . . , Xn) − f (X0, X1, . . . , Xn)] ≥ E[f (X ′
= fS∗ ≥ Ωk(T −1D−k)
max
x∗,x∗,x1,...,xn∈D
0, X1, . . . , Xn) − f (X0, X1, . . . , Xn)]
(f (x+, x1, . . . , xn) − f (x−, x1, . . . , xn)).
8
5 Concluding remarks
We can guarantee that the algorithm finds a solution of value (2) with high probability by repeating
the algorithm Θk(Bk+2) times (since the maxim possible value of the solution is Opt).
We note that our local search algorithm works not only for ordering k-CSPs, but also for
(regular) k-CSPs. The algorithm first assigns random values to all variable xi, and then, n times,
picks a random i ∈ {1, . . . , n}, and changes the value of the variable xi to the optimal value for
fixed other variables. The approximation guarantee of the algorithm is Avg(V,C) + (Opt(V,C) −
Avg(V,C))/Ok,D(B), here k is the arity, and D is the domain size of the CSP. The approximation
guarantee has the same dependence on B as the approximation guarantee of Hastad's (2000) original
algorithm. The analysis relies on Theorem 4.1.
References
B. Berger and P. Shor (1990). Approximation Algorithms for the Maximum Acyclic Subgraph
Problem. SODA 1990.
M. Charikar, V. Guruswami, and R. Manokaran (2009). Every Permutation CSP of arity 3 is
Approximation Resistant. IEEE Conference on Computational Complexity 2009.
M. Charikar, K. Makarychev, and Y. Makarychev. On the Advantage over Random for Maximum
Acyclic Subgraph. FOCS 2007.
V. Guruswami, J. Hastad, R. Manokaran, P. Raghavendra, and M. Charikar (2011). Beating the
Random Ordering Is Hard: Every Ordering CSP Is Approximation Resistant. SIAM J. Com-
put. 40(3).
V. Guruswami, R. Manokaran, and P. Raghavendra (2008) Beating the Random Ordering is Hard:
Inapproximability of Maximum Acyclic Subgraph. FOCS 2008.
V. Guruswami and Y. Zhou (2012). Approximating Bounded Occurrence Ordering CSPs.
APPROX-RANDOM 2012.
G. Gutin, L. van Iersel, M. Mnich and A. Yeo (2012). Every ternary permutation constraint satisfac-
tion problem parameterized above average has a kernel with a quadratic number of variables.
J. Comput. Syst. Sci., vol 78 (1).
J. Hastad (1997). Some optimal inapproximability results. STOC 1997.
J. Hastad. On bounded occurrence constraint satisfaction. Inf. Process. Lett. (IPL) 74(1-2):1-6
(2000).
J. Hastad (2005). Every 2-CSP allows nontrivial approximation. STOC 2005.
A. Newman (2001). The Maximum Acyclic Subgraph Problem and Degree-3 Graphs. RANDOM-
APPROX 2001.
L. Trevisan (2001). Non-approximability results for optimization problems on bounded degree in-
stances. STOC 2001.
9
|
1905.13492 | 1 | 1905 | 2019-05-31T10:11:29 | Majorisation-minimisation algorithms for minimising the difference between lattice submodular functions | [
"cs.DS"
] | We consider the problem of minimising functions represented as a difference of lattice submodular functions. We propose analogues to the SupSub, SubSup and ModMod routines for lattice submodular functions. We show that our majorisation-minimisation algorithms produce iterates that monotonically decrease, and that we converge to a local minimum. We also extend additive hardness results, and show that a broad range of functions can be expressed as the difference of submodular functions. | cs.DS | cs |
Majorisation-minimisation algorithms for minimising the
difference between lattice submodular functions
Conor McMeel
Panos Parpas
[email protected]
[email protected]
June 3, 2019
Abstract
We consider the problem of minimising functions represented as a difference of lattice sub-
modular functions. We propose analogues to the SupSub, SubSup and ModMod routines for
lattice submodular functions. We show that our majorisation-minimisation algorithms produce
iterates that monotonically decrease, and that we converge to a local minimum. We also extend
additive hardness results, and show that a broad range of functions can be expressed as the
difference of submodular functions.
1
Introduction
In discrete optimisation, many objectives are expressible as submodular minimisation or maximisa-
tion problems. This is useful, as submodular functions can be minimised in strongly polynomial time
[9], and have strong constant approximation factors under a variety of constraints for maximisation
[11] [4]. Applications for these problems have included medical imaging [8], document summarisation
[12] and online advertising [1].
We refer to V = {1, . . . , n} as the ground set, and a function f : 2V → R is then called submodular
if the following inequality holds for all A, B ⊆ V :
f (A) + f (B) ≥ f (A ∪ B) + f (A ∩ B).
(1)
We note that a subset A can also be represented as a 0 − 1 vector x, where a component takes on
the value 1 if the element is present in the subset A. In this way our inequality can be written as
f (x) + f (y) ≥ f (min(x, y)) + f (max(x, y))
(2)
for all x, y ∈ {0, 1}n. A popular and often-studied application of submodular optimisation which we
shall use as our running example is that of sensor placement. Suppose we wish to detect the level of
air pollution in a town. We have a finite set of locations that we can install air pollution sensors in,
but we are constrained by some budget. Knowing the geography of our town, we wish to select the
set of locations such that the information we receive is maximised. As shown in [16], this problem
is submodular.
In the classic 0 − 1 case, this corresponds to only having one type of sensor. However, we can
suppose that we have a number of different types of sensors available, each with a strength level
1
that can be parameterised by an integer variable x ∈ {0, . . . , k}. Lattice submodularity is then
determined by satisfying Equation (2), but for all x, y ∈ {0, . . . , k}V . Note that in general, we can
allow k to be different for different elements.
Submodular functions in the set case also lend themselves to a useful diminishing returns prop-
erty, which is often considered as an equivalent definition.
Definition 1.1 (Diminishing returns property). A submodular set function can also be defined by
the following. For any A ⊂ B ⊂ V and e /∈ A, we have
f (A ∪ e) − f (A) ≥ f (B ∪ e) − f (B).
(3)
However, for lattice functions we unfortunately do not get that for free [17], so we define a proper
subclass known as DR-submodular functions as follows:
Definition 1.2 (DR-Submodular). Let f be a lattice submodular function on Qi{0, . . . , k}V . Then
f is called a DR-submodular function if for all x, y ∈ Qi{0, . . . , k}V with x ≤ y componentwise,
c > 0 and basis vector ej, we have
f (x + cej) − f (x) ≥ f (y + cej) − f (y).
(4)
As shown in [2], many typical objectives we come across are in fact DR-submodular, so it makes
sense to consider this subclass. The specific problem of optimising DR-submodular functions has
been studied before [3] [17]. Across these papers, the problem of maximisation has been studied in
monotone and non-monotone settings across a variety of constraints.
In this paper, we will carry out our majorisation-minimisation algorithm by discretising the
lattice submodular functions, giving us submodular functions on some lattice. In [6], the authors
showed that DR-submodular functions on a lattice can be reduced to the set function submodular
case, where the ground set will be smaller. Due to this, we do not consider the specific DR subclass
here, as it would be more efficient to carry out this reduction and then use the work of [10].
To motivate the problem of considering the difference of submodular functions, we turn back
to the sensor information problem. Here, we may wish to also factor in the costs of sensors also.
As is typical in real-world applications, we expect to get a bulk discount if we buy many sensors.
Additionally, we may expect that for any one sensor, as the sensor gets stronger the unit price
increases slower. This leads us to the diminishing returns property for submodularity. Specifically,
if we obtain information f (x), and spend c(x) to get that information, for some sensor placement x,
we wish to minimise f (x) − λc(x), for some tradeoff parameter λ. This formulation also applies to
any other problem of optimising a submodular function with some cost.
In the set function case, this problem was considered by [10], who proposed several different
majorisation minimisation techniques. Here, we will use the work of [2] to come up with our
algorithm, who worked on discretised versions of continuous submodular function, hence lattice
submodular functions. We will derive majorisation-minimisation algorithms on the difference v(x) =
f (x) − g(x). It can be considered the extension to the lattice case of the SupSub procedure in [10],
originally introduced in [14], as well as the SubSup and ModMod procedures. In fact by discretising
continuous functions, we can also consider it an extension to the continuous domain much like [2].
We will show that we converge to a local minimum of the function, and consider the class of functions
we can represent as the difference of submodular functions.
2
1.1 Outline
We begin by discussing some preliminaries Section 2. We consider the theory of continuous and
lattice submodular functions, and show how subgradients are obtained. We show modular and tight
lower and upper bounds, along with a decomposition that allows us to utilise the theory of DR-
submodular functions effectively. Section 3 presents and discusses the three algorithms, along with
a discussion of a stopping criterion. In Section 4, we discuss complexity, deriving the functions f, g
for the representation v(x) = f (x) − g(x), along with other theoretical issues. We conclude with a
brief discussion in Section 5.
2 Preliminaries
We are considering the problem of minimising the difference between lattice submodular functions.
In particular, we seek to minimise
v(x) = f (x) − g(x)
(5)
where f, g are lattice submodular. For our algorithms, our plan will be to in fact use an extension
of the functions f that was defined in [2]. While doing this, we will also give the main ingredient
for computing the lower bound.
2.1 An Extension
Our extension will be from the set Qi{0, . . . , ki − 1} = Qi Xi to the product of measures of Xi.
The motivation for this can be thought of as follows. In the Lovasz extension, the value x ∈ [0, 1]
taken on by a component can be thought of as the probability that an element x is included in a
set. Similarly, we will extend here to probability mass functions over the domain {0, . . . , ki − 1}.
We note that any PMF on this domain can be represented by its reverse cumulative probability
density function ρi(xi) = µi(xi) + . . . + µi(ki − 1) = Fµi (xi). We then see that ρi(0) = 1 always,
and the only constraint on it is that its elements are non-decreasing, and all belong to [0, 1]. As in
[2], we denote this set of vectors as [0, 1]ki−1
.
↓
Next, we think of this set of vectors just defined, and note that if we mark the locations of ρi(xi),
we will divide the number line into ki segments. This induces a map θ(ρi, t) from R to 0, . . . , ki − 1
with θ(ρi, t) = ki − 1 if t ≤ ρi(ki − 1), θ(ρi, t) = xi if t ∈ (ρi(xi+1, ρi(xi)] for xi ∈ 1, . . . , ki − 2, and
θ(ρi, t) = 0 for t ≥ ρi(1).
Our extension is then thus defined, for a submodular function f
f↓(ρ) = Z 1
0
f (θ(ρ1, t), . . . , θ(ρn, t))dt.
(6)
For a more thorough exposition, see [2]. We now detail how the extension is evaluated using a greedy
algorithm.
2.2 A lower bound
We now give the greedy algorithm from [2].
3
Theorem 2.1. Consider the extension of f↓(ρ) of a submodular function f . Order all values of ρ in
decreasing order, breaking ties arbitrarily but ensuring that all components for a given ρi are in the
correct order. Assume the value at position s is equal to t(s) and corresponds to ρi(s)(j(s)). Then
we have the following expression for f↓
r
f↓(ρ) = f (0) +
Xi=1
which we write in the form f↓(ρ) = f (0) +Pn
i=1Pki−1
xi=1 wi(xi)ρi(xi), where wi(xi) corresponds to the
difference in value between two function evaluations whose arguments differ by a single basis vector.
t(s)(f (y(s)) − f (y(s − 1)))
(7)
In this section, we will first remind ourselves of the different characterisations of the base polyhe-
dron, how the greedy algorithm is used to construct extreme points for one of these characterisations,
and how subgradients are computed. Using this, we will compute a lower bound that will be used
for our algorithm.
We remind ourselves first of the characterisation of the base polyhedron as given initially in [2]:
Definition 2.1 (Base Polyhedron). Let f (x1, . . . , xn) be a discrete submodular function WLOG
with f (0, 0, . . . , 0) = 0, in each argument from 0 to the integer ki − 1 respectively. Then the
Rki−1 such that for all (x1, . . . , xn):
submodular polyhedron can be defined by arguments wi ∈ Qi
n
xi
Xi=1
Xyi=1
wi(yi) ≤ f (x1, . . . , xn),
n
Xi=1
ki−1
Xyi=1
wi(yi) = f (k1 − 1, . . . , kn − 1).
(8)
(9)
As mentioned in [2], this polyhedron is in fact not a polyhedron, and is unbounded, if there is
any ki > 2. This can be made explicit in the following example.
Example 2.1 (Base Polyhedron Unbounded). Let ki = 3 for some i. Then for that i, we can add
to wi any ui such that
u1(1) = −1, u1(2) = 1,
(10)
and we see this won't violate any of the equations in the definition. This argument extends straight-
forwardly to any ki > 2.
Because of this, they instead define the base polyhedron as the convex hull of outputs of a greedy
algorithm. The following result shows that the base polyhedron still behaves in the same way:
Lemma 2.1. Let f be some submodular function. Then for any ρ ∈ Qi
Rki−1
↓
we have that
max
w
hw, ρi = f↓(ρ) − f (0),
(11)
where we take the max over either characterisation of the base polyhedron.
We see that as long as we take a w compatible with the ordering ρ, we will get a subgradient
for the function f↓(ρ). Restricting to the ρ that will give us points in the domain of our original
submodular function f , we get an element of the subdifferential of f .
4
Now we can construct a lower bound. First, we construct a set A with r elements, each of which
corresponds to an increment of one of the n basis vectors. There are ki − 1 copies of the increment
of element i. Note that such a set A corresponds to a chain defined by
0 = p0 < p1 < . . . < pr = (k1 − 1, . . . , kn − 1).
(12)
Take a permutation of A denoted by σ and form its corresponding chain Cσ. Ensure σ is such that
the chain pi contains z, something we define now.
Definition 2.2 (Chain containing an element). Let pi be a chain as defined in Equation (12). We
say the chain contains x, for some vector x = (x1, . . . , xn), if we have px1+...+xn = x.
Note that any chain that contains a vector y is compatible with the ordering of the ρ that
corresponds to y. In particular, we'll take the chain Cσ where we increment the first element to y1,
then the second to y2 and so on. After it reaches y, we let it have any behaviour. Then we can form
w by taking differences of successive elements of the chain.
Now we can form the function corresponding to the lower bound by making w as described
earlier. This will be denoted by
hf,y(i, j) = wi(j).
To extend this definition to an entire point x = (x1, x2, . . . , xn) we can do the following:
hf,y(x) =
n
xi
Xi=1
Xj=1
hf,y(i, j) =
n
xi
Xi=1
Xj=1
wi(j).
Now note that for each pi in the chain, we have that
hf,y(pi) = f (pi).
In particular, note that as y is contained in this chain, we have that
hf,y(y) = f (y).
Then due to the fact that this is a subgradient, we have that
for all x. This is parameterised by y, σ and is tight at y.
hf,y(x) ≤ f (x)
2.3 Upper Bound
(13)
(14)
(15)
(16)
(17)
In [10], an upper bound was derived for a submodular set function. This can be extended, but it
does require DR-submodularity. We show now how to get around that as follows if we know one
particular quantity:
Lemma 2.2. Let f be a lattice submodular function. Then f can be represented as the sum of a
modular function g and a DR-submodular function h.
Proof. Let λ be the largest second difference of the function f that violates the DR property, or
at least an upper bound of it. Because of submodularity, we know the difference will be a second
difference entirely within one basis element. Namely, take:
5
λ ≤ max
x,i
(f (x + 2ei) − 2f (x + ei) + f (x)) .
(18)
Then let g = λ(x2
1 + . . . + x2
n). This will give us h = f − g DR-submodular as required.
Computing the exact tight value of λ is often hard, but there are some cases where it will be
easier to derive upper bounds:
1. If the function f is the discretisation of a continuous function, we can compute the largest
positive eigenvalue, or an upper bound of it, via a number of eigenvalue algorithms.
2. If the function f is also L♮-convex (or midpoint convex), the function is convex-extensible [13],
and so proceed similarly to above.
3. If the function f is a quadratic program, with f (x) = xT Ax + bT x + c, then we have that λ is
the maximum positive diagonal element of A (or 0 if none exist).
The idea now here is that we will let f = g + h, and then derive a modular and tight upper
bound for h, thus giving our modular and tight upper bound for f as g is already modular. The
aim for that is to generalise the bounds found in [10], given as
f (Y ) ≤ f (X) − Xj∈X\Y
f (Y ) ≤ f (X) − Xj∈X\Y
f (jX \ j) + Xj∈Y \X
f (jV \ j) + Xj∈Y \X
f (j∅),
f (jX).
(19)
(20)
In this section, we assume that f is a lattice submodular function given on the product of sets
{0, . . . , ki − 1}. The extension of these bounds is given in the following lemma:
Lemma 2.3. Let f be a DR-submodular function as described. Let m(x) = max(x, 0) with m
extending its arguments to vectors componentwise. Let x, y be vectors in Qi{0, . . . , ki}. Let
(a1, . . . , an) = m(x − y),
(b1, . . . , bn) = m(y − x).
Then we have the following:
f (y) ≤ f (x) −
f (y) ≤ f (x) −
n
Xi=1
Xi=1
n
[f (x) − f (x − aiei)] +
n
Xi=1
f (biei),
[f (kmax) − f (kmax − aiei)] +
n
Xi=1
f (f (x + biei) − f (x)),
where ei are the usual basis vectors.
Proof. We first show the following bounds:
(21)
(22)
(23)
(24)
(25)
f (y) ≤ f (x) −
f (y) ≤ f (x) −
n
Xi=1
Xi=1
n
[f (x) − f (x − aiei)] +
[f (biei + min(x, y)) − f (min(x, y))],
Xi=1
n
[f (max(x, y)) − f (max(x, y) − aiei)] +
n
Xi=1
f (f (x + biei) − f (x)).
(26)
6
The proof proceeds similarly to the derivation in [10], but with unions and intersections replaced
with mins and maxes respectively. We start with the second statement. Take an arbitrary x, y with
ai, bi as before. Then note we have
f (max(x, y)) − f (x) =
=
[f (x +
i
Xj=1
ρbi,ei (x +
n
Xi=1
n
Xi=1
bjej) − f (x +
i−1
Xj=1
bjej)]
i−1
Xj=1
bjej),
(27)
(28)
where we take a0, e0 = 0. Here ρa,b(c) denotes the marginal return on adding the value a in the
basis vector b when the function already has argument c. Using the DR-submodular property, we
then see
n
Xi=1
ρbi,ei(x +
i−1
Xj=1
bjej) ≤
n
Xi=1
ρbi,ei(x) =
n
Xi=1
f (x + biei) − f (x).
(29)
Similarly, we'll now consider the following expression:
f (max(x, y)) − f (y) =
=
[f (y +
i
Xj=1
ρai,ei (y +
n
Xi=1
n
Xi=1
ajej) − f (y +
i−1
Xj=1
ajej)]
i
Xj=1
ajej − aiei)
(30)
(31)
≥ ρai,ei(max(x, y) − aiei) =
n
Xi=1
[f (max(x, y)) − f (max(x, y) − ai)].
(32)
Subtracting these two gives us the required result. We now proceed to the first statement. We get
the first inequality similarly to just how we proceeded:
f (x) − f (min(x, y)) =
=
[f (x +
i
Xj=1
ρai,ei (y +
n
Xi=1
n
Xi=1
ajej) − f (x +
i−1
Xj=1
ajej)]
i
Xj=1
ajej − aiei)
≤ ρai,ei(x − aiei) =
n
Xi=1
[f (x) − f (x − aiei)].
The second also coming easily:
f (y) − f (min(x, y)) =
=
≤
=
Xi=1
Xi=1y−x
n
[f (y +
i
Xj=1
ρbi,ei (y +
n
Xi=1
n
Xi=1
n
bjej) − f (y +
i−1
Xj=1
bjej)]
i−1
Xj=1
bjej),
ρbi,ei (min(x, y))
[f (biei + min(x, y)) − f (min(x, y))].
7
(33)
(34)
(35)
(36)
(37)
(38)
(39)
And again we subtract to get the first bounds that we wanted. To get to the required result, simply
apply DR-submodularity to the final term of the first bound and the second term of the second
bound.
We note that these bounds are separable. Additionally, we see that these bounds are tight at x,
namely equality is achieved with y = x. This result also applies as written for continuous submodular
functions, where the vectors x, y instead belong to [0, 1]n WLOG.
3 Three algorithms
We describe the integer lattice majorisation-minimisation algorithms here, in Algorithms 1, 2, 3. In
Algorithm 1, note that at every step we are minimising a lattice submodular function, which as shown
in [2] to get this to arbitrary precision we have complexity O(( 2GBn
)), for a continuous
submodular function defined on [0, B]n with Lipschitz constant G (note that the author minimises
a discretised version of the continuous function). For Algorithm 2, we are instead maximising, for
which we have an approximation factor of 1/3 [7] and runs in O(kn) calls. For Algorithm 3, we are
at each point minimising a modular function, which can be done easily in O(k′) function evaluations,
)3 log( 2GBn
ε
ε
where we have k′ = Pi ki, by evaluating each separated function at every point and taking the n
minima. Additionally, we note that Algorithms 2, 3 will require in principle an upper bound on the
quantity λ as described in Lemma 2.2.
We note that we have quite easily:
v(xt+1) = f (xt+1) − hg
xt (xt) − hg
≤ mf
σ(xt+1)
σ(xt)
= f (xt) − g(xt) = v(xt),
(40)
(41)
(42)
the second to last equality coming from the demonstrated tightness of our bounds. However, we
note that the sequence may not strictly decrease, and we seek a convergence condition. For the
upper bound, we can simply say try both of them, and if neither strictly decreases the function,
we're done. For the lower bound, there are many permutations we can change and we want some
sort of a stopping criterion.
Algorithm 1: Integer lattice SubSup algorithm.
input : Function v(x) = f (x) − g(x), where f and g are both submodular.
output: A local minimum of v(x)
1 Discretise the function v(x) in a preferred way;
2 x0 ← (0, . . . , 0); t ← 0;
3 while not converged (xt+1 6= xt) do
4
5
6
7
Form the extension f↓(xt) of f (xt).;
Choose a permutation σ such that the induced chain is compatible with xt;
xt+1 = arg minx f↓(x) − hg
t = t+1;
σ(x);
8
Lemma 3.1. If we choose O(n) permutations each with different increments directly before and
after xt, and attempt to decrease with both upper bounds and we are not successful, then we have
reached a local minimum.
Proof. Proof is functionally identical to the set function case as in [10], except instead of saying we
consider all g(X ∪ j), g(X \ j), we just consider all g(x + ei), g(x − ei) where ei are basis vectors.
Algorithm 2: Integer lattice SupSub algorithm.
input : Function v(x) = f (x) − g(x), where f and g are both submodular and have
bounded Hessian.
output: A local minimum of v(x)
1 Discretise the function v(x) in a preferred way;
2 x0 ← (0, . . . , 0); t ← 0;
3 while not converged (xt+1 6= xt) do
4
5
xt+1 = arg minx mf
t = t+1;
xt (x) − g(x);
Algorithm 3: Integer lattice ModMod algorithm.
input : Function v(x) = f (x) − g(x), where f and g are both submodular and have
bounded Hessian.
output: A local minimum of v(x)
1 Discretise the function v(x) in a preferred way;
2 x0 ← (0, . . . , 0); t ← 0;
3 while not converged (xt+1 6= xt) do
4
5
6
Choose a permutation σ such that the induced chain is compatible with xt;
xt+1 = arg minx mf
t = t+1;
xt (x) − hg
σ(x);
4 Theoretical Analysis
We note that we don't get any multiplicative approximation guarantees, as the hardness results are
inherited from the set function case. However, we would like to extend the additive hardness results
of [10]. While following their proof will rely on f being DR-submodular, we recall we can write
f = g + h for g modular and h DR-submodular via Lemma 2.2.
So now we act just on the DR-submodular function. This requires an extension of the decompo-
sition from [5] which we give in a slightly weaker form:
Lemma 4.1. Let f be any DR-submodular function with f (0) = 0. It can be decomposed into a
modular function g plus a monotone function h with h(0) = 0.
9
Proof. We first construct the modular function g, then show that the function h = f −g is monotone,
and takes the value 0 at 0. For any input y we form:
g(y) =
n
yk
Xk=1
Xj=1
mj,k
j
,
(43)
for 0 < j ≤ yk for each k. We note this decays to the modular function in [5] if we restrict all yk to
{0, 1}. This function is clearly modular. Additionally, it is clear that h = f − g has h(0) = 0. To
show that it is monotone:
f ′(k + ei) = f (k + ei) −
n
y ′
k
Xk=1
Xj=1
mj,k
j
f ′(k) = f (k) −
n
yk
Xk=1
Xj=1
mj,k
j
.
(44)
(45)
Note that in the first equation, there will be one extra term subtracted,
hand side of the first equation here is greater than the second, as:
mi,ki +1
ki+1 . We claim the right
f (k + ei) −
n
y ′
k
Xk=1
Xj=1
mj,k
j
− (f (k) −
n
yk
Xk=1
Xj=1
mj,k
j
)
= f (k + ei) − f (k) −
mi,ki+1
ki + 1
=
1
ki + 1
((ki + 1)(f (k + ei) − f (k)) − (f (kmax) − f (kmax − eiki + 1))).
(46)
(47)
(48)
We then split up the second term to get:
f (kmax) − f (kmax − eiki + 1)) =
k1
Xj=1
f (kmax − (j − 1)ei) − f (kmax − jei),
(49)
and use the DR property to bound each individual term by f (k + ei) − f (k), giving us our result.
Using this, we now have the following combining the two decompositions:
Theorem 4.1. Let f be a lattice submodular function with f (0) = 0. Then f can be represented as
the sum of a modular function g and a monotone submodular function h, with h(0) = 0.
This will allow us to get additive bounds similar to [10]:
Lemma 4.2. Consider the problem of minimising f (x) − g(x). Apply our decomposition to instead
have v(x) = f ′(x) − g′(x) + k(x). Here k is modular, and f, g are monotone. Then we have:
min v(x) ≥ min
x
f ′(x) + k(x) − g′(kmax),
n
Xk=1
(min
y
y
Xi=1
mi,k
i
).
min v(x) ≥ f ′(0) − g′(kmax) +
10
(50)
(51)
(52)
Proof. We have:
min
x
v(x) = min
x
f ′(x) − g′(x) + k(x)
≥ min
(f ′(x) + k(x)) − max
g′(x)
x
x
x
= min
(f ′(x) + k(x)) − g′(kmax),
by monotonicity of g′. Next, we continue from the previous line:
≥ min
x
f ′(x) + min
x
k(x) − g′(kmax)
= f ′(0) − g′(kmax) +
n
Xk=1
(min
y
y
Xi=1
mi,k
i
).
(53)
(54)
(55)
(56)
(57)
The nested optimisation problem here in the second bound is non-trivial, but for our purposes it
doesn't actually need to be solved. All we need to see is that by setting y to zero, we can eliminate
this term and thus by the monotonicity of f ′, g′, we will always get that this second bound is less
than zero.
We also prove some other results that show this algorithm can be broadly applied. The first
results shows that we can efficiently minimise many functions using this approach:
Lemma 4.3. Let v(x) be any function on the lattice {0, . . . , ki − 1}. Then v can be written as the
difference of two submodular functions.
Proof. Let g be any strictly submodular function on the same lattice, that is, one with all second
differences over pairs i 6= j strictly less than zero. We shall denote by m the minimum absolute
value of the second differences over all pairs i 6= j and x, that is:
m = min
i6=j,x
g(x + ei + ej) − g(x + ei) − g(x + ej) + g(x).
(58)
As all second differences of g are less than zero as its submodular, note for any other difference m′
we have m′
m ≤ −1. We shall additionally define n to be the maximum absolute value of the second
difference of v under the same constraint:
n = max
i6=j,x
v(x + ei + ej) − v(x + ei) − v(x + ej) + v(x).
We can now form the function f :
f (x) = v(x) +
n
m
g(x).
(59)
(60)
We claim that f is submodular, which will give our result, as a submodular function scaled by a
constant is submodular. To do this, take any of the second differences over pairs i 6= j of the function
f . Denote the second difference as Di,j(f )(x). We find:
as required.
Di,j(f )(x) = Di,j(v)(x) +
n
m
≤ Di,j(v)(x) − n
Di,j(g)(x)
≤ 0,
11
(61)
(62)
(63)
Of course, n, m is in general hard to find, so another question that may be asked regarding this
problem is how difficult it is to find the functions f, g corresponding to some v. That is answered
with this result, extending [10]:
Lemma 4.4. Suppose that we know n, or a lower bound on n as in the previous lemma. Then
given a function v, we can construct submodular functions f, g such that v(x) = f (x) − g(x).
Proof. Consider the submodular function on Qi{0, . . . , ki − 1}:
p − 4Xi6=j
g(x1, . . . , xp) = x2
1 + . . . + x2
xixj
(64)
We can clearly verify here that g is strictly submodular, and we have m = −4. Thus if we know a
lower bound on n we can form f in the manner of Lemma 4.3 as required.
Note that the choice of g in the proof above is not special, it simply needs to be submodular and
we need to know its m. Thus depending on the particular problem v(x), different choices of g may
give nice or meaningful decompositions.
To do complexity analysis, we will be working on an ǫ-approximate version of the algorithm,
introduced in [15]. This means that we will only proceed to step t+1 if we have v(xt+1) ≤ (1+ǫ)v(xt).
The reason we do this comes from [10], as we know this problem for set functions is PLS-complete.
We now consider the complexity of this procedure. The complexity in minimising the lattice version
can be found from a lemma whose statement and proof can be adapted directly from [10]:
Lemma 4.5. The ǫ-approximate version of each algorithm has a worst case complexity of O( log(M/m)
where T is the complexity of the iterative step, M = f ′(0) − g′(kmax) and m = v(x1).
ǫ
T ),
We also note that as described earlier, the ModMod procedure has the lowest complexity at each
iteration.
4.1 Constrained Optimisation
For the SubSup procedure, we are minimising a submodular function with some constraint at every
iteration. However, we know that this is hard, and also hard to approximate [18] for the set function
case, and so will be for ours also. Therefore this algorithm is not suitable for using constraints.
We know that for some simple constraints such as cardinality, knapsack and polymatroid, the
submodular maximisation problem has been studied on the integer lattice for monotone functions
[17]. So at least for some subclasses, we can use the SupSub procedure to optimise under constraints.
While to the best of our knowledge no-one has explicitly written algorithms for modular min-
imisation on the integer lattice, we know it is easy and can be done exactly at least for cardinality
constraints, where we can just enumerate all marginal gains and take the lowest in each variable.
We will then take the B lowest, where our cardinality constraint says x(V ) ≤ B.
In the set function case, we can optimise easily and exactly over a variety of other constraints
[10]. However, here our separability also gives us linearity, something we lose on our lattice as
we separate to arbitrary discrete functions.
optimisation problems.
It is worth looking into more of these constrained
12
5 Conclusion
In [10], the authors studied the problem of minimising the difference between two submodular set
functions. We have extended this to the case of general lattice submodular functions, without the DR
requirement. Additionally, we note that via discretisation, our method can be applied to continuous
functions also.
In performing the majorisation-minimisation technique, we extended an earlier bound from [10]
for the upper bound, which is valid for DR-submodular functions, and used a decomposition for
submodular functions into DR-submodular functions plus a modular function to take advantage of
it. For the lower bound, we used the method of computing the subgradient as in [2]. The result of
this greedy algorithm gives us our lower bound, and did not require f to be DR-submodular.
After that we formally stated our algorithms, performed some complexity analysis, and analysed
other theoretical properties of it. One clear extension on this work would be finding an alternative
lower bound that can be used on continuous functions, as our upper bound can already be used in
this context.
13
References
[1] Saeed Alaei, Ali Makhdoumi, and Azarakhsh Malekian. Maximizing sequence-submodular func-
tions and its application to online advertising. arXiv preprint arXiv:1009.4153, 2010.
[2] Francis Bach. Submodular functions:
from discrete to continuous domains. Mathematical
Programming, pages 1 -- 41, 2016.
[3] An Bian, Kfir Levy, Andreas Krause, and Joachim M Buhmann. Continuous dr-submodular
maximization: Structure and algorithms. In Advances in Neural Information Processing Sys-
tems, pages 486 -- 496, 2017.
[4] Niv Buchbinder, Moran Feldman, Joseph Seffi Naor, and Roy Schwartz. Submodular maxi-
In Proceedings of the twenty-fifth annual ACM-SIAM
mization with cardinality constraints.
symposium on Discrete algorithms, pages 1433 -- 1452. Society for Industrial and Applied Math-
ematics, 2014.
[5] William H Cunningham. Decomposition of submodular functions. Combinatorica, 3(1):53 -- 68,
1983.
[6] Alina Ene and Huy L Nguyen. A reduction for optimizing lattice submodular functions with
diminishing returns. arXiv preprint arXiv:1606.08362, 2016.
[7] Corinna Gottschalk and Britta Peis. Submodular function maximization on the bounded integer
lattice. In International Workshop on Approximation and Online Algorithms, pages 133 -- 144.
Springer, 2015.
[8] Steven CH Hoi, Rong Jin, Jianke Zhu, and Michael R Lyu. Batch mode active learning and its
application to medical image classification. In Proceedings of the 23rd international conference
on Machine learning, pages 417 -- 424. ACM, 2006.
[9] Satoru Iwata, Lisa Fleischer, and Satoru Fujishige. A combinatorial strongly polynomial al-
gorithm for minimizing submodular functions. Journal of the ACM (JACM), 48(4):761 -- 777,
2001.
[10] Rishabh Iyer and Jeff Bilmes. Algorithms for approximate minimization of the difference be-
tween submodular functions, with applications. arXiv preprint arXiv:1207.0560, 2012.
[11] Jon Lee, Vahab S Mirrokni, Viswanath Nagarajan, and Maxim Sviridenko. Non-monotone
submodular maximization under matroid and knapsack constraints. In Proceedings of the forty-
first annual ACM symposium on Theory of computing, pages 323 -- 332. ACM, 2009.
[12] Hui Lin and Jeff A Bilmes. Learning mixtures of submodular shells with application to document
summarization. arXiv preprint arXiv:1210.4871, 2012.
[13] Satoko Moriguchi and Kazuo Murota. On discrete hessian matrix and convex extensibility.
Journal of the Operations Research Society of Japan, 55(1):48 -- 62, 2012.
[14] Mukund Narasimhan and Jeff A Bilmes. A submodular-supermodular procedure with applica-
tions to discriminative structure learning. arXiv preprint arXiv:1207.1404, 2012.
14
[15] James B Orlin, Abraham P Punnen, and Andreas S Schulz. Approximate local search in
combinatorial optimization. SIAM Journal on Computing, 33(5):1201 -- 1214, 2004.
[16] Goran Radanovic and Boi Faltings. Incentive schemes for participatory sensing. In Proceedings
of the 2015 International Conference on Autonomous Agents and Multiagent Systems, pages
1081 -- 1089. International Foundation for Autonomous Agents and Multiagent Systems, 2015.
[17] Tasuku Soma and Yuichi Yoshida. Maximizing monotone submodular functions over the integer
lattice. Mathematical Programming, 172(1-2):539 -- 563, 2018.
[18] Zoya Svitkina and Lisa Fleischer. Submodular approximation: Sampling-based algorithms and
lower bounds. SIAM Journal on Computing, 40(6):1715 -- 1737, 2011.
15
|
1905.00612 | 1 | 1905 | 2019-05-02T08:20:39 | Online Circle Packing | [
"cs.DS",
"cs.CG"
] | We consider the online problem of packing circles into a square container. A sequence of circles has to be packed one at a time, without knowledge of the following incoming circles and without moving previously packed circles. We present an algorithm that packs any online sequence of circles with a combined area not larger than 0.350389 0.350389 of the square's area, improving the previous best value of {\pi}/10 \approx 0.31416; even in an offline setting, there is an upper bound of {\pi}/(3 + 2 \sqrt{2}) \approx 0.5390. If only circles with radii of at least 0.026622 are considered, our algorithm achieves the higher value 0.375898. As a byproduct, we give an online algorithm for packing circles into a 1\times b rectangle with b \geq 1. This algorithm is worst case-optimal for b \geq 2.36. | cs.DS | cs |
Online Circle Packing
Sven von Hoveling2,[0000−0003−3937−2429], and
S´andor P. Fekete1[0000−0002−9062−4241],
Christian Scheffer1[0000−0002−3471−2706]
1 Department of Computer Science, TU Braunschweig, Germany
{s.fekete,c.scheffer}@u-bs.de
2 Department of Computing Science, University of Oldenburg, Germany
[email protected]
Abstract. We consider the online problem of packing circles into a
square container. A sequence of circles has to be packed one at a time,
without knowledge of the following incoming circles and without moving
previously packed circles. We present an algorithm that packs any online
sequence of circles with a combined area not larger than 0.350389 of the
square's area, improving the previous best value of π/10 ≈ 0.31416; even
√
2) ≈ 0.5390. If
in an offline setting, there is an upper bound of π/(3 + 2
only circles with radii of at least 0.026622 are considered, our algorithm
achieves the higher value 0.375898.
As a byproduct, we give an online algorithm for packing circles into a 1×b
rectangle with b ≥ 1. This algorithm is worst case-optimal for b ≥ 2.36.
Keywords: Circle Packing · Online Algorithms · Packing Density.
1
Introduction
Packing a set of circles into a given container is a natural geometric optimization
problem that has attracted considerable attention, both in theory and practice.
Some of the many real-world applications are loading a shipping container with
pipes of varying diameter [9], packing paper products like paper rolls into one
or several containers [8], machine construction of electric wires [18], designing
control panels [2], placing radio towers with a maximal coverage while minimizing
interference [19], industrial cutting [19], and the study of macro-molecules or
crystals [20]. See the survey paper of Castillo, Kampas, and Pint´er [2] for an
overview of other industrial problems. In many of these scenarios, the circles
have to be packed online, i.e., one at a time, without the knowledge of further
objects, e.g., when punching out a sequence of shapes from the raw material.
Even in an offline setting, deciding whether a given set of circles fits into a
square container is known to be NP-hard [4], which is also known for packing
squares into a square [10]. Furthermore, dealing with circles requires dealing
with possibly complicated irrational numbers, incurring very serious additional
geometric difficulties. This is underlined by the slow development of provably
optimal packings of n identical circles into the smallest possible square. In 1965,
2
S. P. Fekete, C. Scheffer, S. von Hoeveling
Fig. 1. Circles are arriving one at a time and have to be packed into the unit square.
At this stage, the packed area is about 0.33. What is the largest A ≥ 0 for which any
sequence of total area A can be packed?
Schaer [16] gave the optimal solution for n = 7 and n = 8 and Schaer and
Meir [17] gave the optimal solution for n = 9. A quarter of a century later, Wurtz
et al. [3] provided optimal solutions for 10,11,12, and 13 equally sized circles. In
1998, Nurmela and Ostergard [14] provided optimal solutions for n ≤ 27 circles
by making use of computer-aided optimality proofs. Mark´ot and Csendes [12]
gave optimal solutions for n = 28, 29, 30 also by using computer-assisted proofs
within tight tolerance values. Finally, in 2002 optimal solutions were provided
for n ≤ 35 by Locatelli and Raber [11]; at this point, this is still the largest n
for which optimal packings are known. The extraordinary challenges of finding
densest circle packings are also underlined by a long-standing open conjecture
by Erdos and Oler from 1961 [15] regarding optimal packings of n unit circles
into an equilateral triangle, which has only been proven up to n = 15.
These difficulties make it desirable to develop relatively simple criteria for
the packability of circles. A natural bound arises from considering the packing
density, i.e., the total area of objects compared to the size of the container; the
critical packing density δ is the largest value for which any set of objects of total
area at most δ can be packed into a unit square; see Fig. 1.
√
In an offline setting, two equally sized circles that fit exactly into the unit
square show that δ ≤ δ∗ = π/(3 + 2
2) ≈ 0.5390. This is indeed tight: Fekete,
Morr and Scheffer [7] gave a worst-case optimal algorithm that packs any in-
stance with combined area at most δ∗; see Fig. 2 (left). More recently, Fekete,
Keldenich and Scheffer [6] established 0.5 as the critical packing density of circles
in a circular container.
The difficulties of offline circle packing are compounded in an online setting.
This is highlighted by the situation for packing squares into a square, which does
not encounter the mentioned issues with irrational coordinates. It was shown in
1967 by Moon and Moser [13] that the critical offline density is 0.5: Refining
an approach by Fekete and Hoffmann [5], Brubach [1] established the currently
best lower bound for online packing density of 0.4. This yields the previous best
bound for the online packing density of circles into a square: Inscribing circles
into bounding boxes yields a value of π/10 ≈ 0.3142.
? Online Circle Packing
3
Fig. 2. Examples of algorithmic circle packings. (Left) The worst-case optimal offline
algorithm of Fekete et al. [7] for packing circles into the unit square. (Center) The
worst-case optimal offline algorithm of Fekete et al. [6] for packing circles into the unit
circle. (Right) Our online algorithm for packing circles into the unit square.
1.1 Our Results
In this paper, we establish new lower bounds for the online packing density of
circles into a square and into a rectangle. Note that in the online setting, a
packing algorithm has to stop as soon as it cannot pack a circle. We provide
three online circle packing results for which we provide constructive proofs, i.e.,
corresponding algorithms guaranteeing the claimed packing densities.
Theorem 1. Let b ≥ 1. Any online sequence of circles with a total area no larger
can be packed into the 1 × b-rectangle R.
than min
This is worst-case optimal for b ≥ 2.36.
0.528607 · b − 0.457876, π
(cid:16)
(cid:17)
4
We use the approach of Theorem 1 as a subroutine and obtain the following:
Theorem 2. Any online sequence of circles with a total area no larger than
0.350389 can be packed into the unit square.
If the incoming circles' radii are lower bounded by 0.026623, the density
guaranteed by the algorithm of Theorem 2 improves to 0.375898.
Theorem 3. Any online sequence of circles with radii not smaller than 0.026623
and with a total area no larger than 0.375898 can be packed into the unit square.
We describe the algorithm of Theorem 1 in Section 2 and the algorithm of
the Theorems 2 and 3 in Section 3.
2 Packing into a Rectangle
In this section, we describe the algorithm, Double-Sided Structured Lane Pack-
ing (DSLP), of Theorem 1. In particular, DSLP uses a packing strategy called
Structured Lane Packing (SLP) and an extended version of SLP as subroutines.
4
S. P. Fekete, C. Scheffer, S. von Hoeveling
Fig. 3. Comparison of Tight Lane Packing (left) with Structured Lane Packing (right)
for the same input. The former has a smaller packing length.
2.1 Preliminaries for the Algorithms
A lane L ⊂ R2 is an x- and y-axis-aligned rectangle. The length (cid:96)(L) and the
width w(L) of L are the dimensions of L such that w(L) ≤ (cid:96)(L). L is horizontal
if the length of L is given via the extension of L w.r.t. the x-axis. Otherwise, L is
vertical. The distance between two circles packed into L is the distance between
the orthogonal projections of the circles' midpoints onto the longer side of L. A
lane is either open or closed. Initially, each lane is open.
Packing a circle C into a lane L means placing C inside L such that C does
not intersect another circle that is already packed into L or into another lane. A
(packing) strategy for a lane L is a set of rules that describe how a circle has to
be packed into L. The (packing) orientation of a strategy for a horizontal lane
is either rightwards or leftwards and the (packing) orientation of a strategy for
a vertical lane is either downwards or updwards.
Let w be the width of L. Depending on the radius r of the current circle C,
we say: C is medium (Class 1) if w > r ≥ w
4 > r ≥ 0.0841305w
(Class 2), C is tiny (Class 3 or 4) if 0.0841305w > r ≥ 0.023832125w, and C is
very tiny if 0.023832125w > r (Classes 5,6, . . . ). For a more refined classification
of r, we refer to Section 2.3. The general idea is to reach a certain density within
a lane by packing only relatively equally sized circles into a lane with SLP.
For the rest of Section 2, for 0 < w ≤ b, let L be a horizontal w × b lane.
4 , C is small if w
2.2 Structured Lane Packing (SLP) -- The Standard Version
Rightwards Structured Lane Packing (SLP) packs circles into L alternating
touching the bottom and the top side of L from left to right, see Fig. 3 (right).
In particular, we pack a circle C into L as far as possible to the left while
guaranteeing: (1) C does not overlap a vertical lane packed into L, see Section 2.3
for details3. (2) The distance between C and the circle C(cid:48) packed last into L is
at least min{r, r(cid:48)} where r, r(cid:48) are the radii of C, C(cid:48), see Fig. 3 (right).
Leftwards Structured Lane Packing packs circles by alternatingly touching
the bottom and the top side of L from right to left. Correspondingly, upwards
and downwards Structured Lane Packing packs circles alternatingly touching the
left and the right side of L from bottom to top and from top to bottom.
3 Requiring that C does not overlap a vertical lane placed inside L is only important
for the extension of SLP (see Section 2.3), because the standard version of SLP does
not place vertical lanes inside L.
Online Circle Packing
5
2.3 Extension of SLP -- Filling Gaps by (Very) Tiny Circles
Now consider packing medium circles with SLP. We extend SLP for placing tiny
and very tiny circles within the packing strategy, see Fig. 4. Note that small
circles are not considered for the moment, such that (very) tiny circles are relat-
evily small compared to the medium ones. In particular, if the current circle C
is medium, we apply the standard version of SLP, as described in Section 2.2.
If C is (very) tiny, we pack C into a vertical lane inside L, as described next.
Fig. 4. A packing produced by extended SLP: 128 (3 medium, 17 tiny, and 108 very
tiny) circles packed into 15 (1 medium, 4 tiny and 10 very tiny) lanes. A possible input
order of the circles is 1 medium circle, 23 very tiny circles (filling the sparse block A),
1 medium circle, 13 tiny circles (filling the sparse block B), 24 very tiny circles (filling
the vertical lane C), 1 medium circle, 2 tiny circles (filling the sparse block D), 11 very
tiny circles (filling the sparse block E), and 2 tiny circles (filling the vertical lane F ).
We pack (very) tiny circles into vertical lanes inside L, see Fig. 4. The vertical
lanes are placed inside blocks that are the rectangles induced by vertical lines
touching medium circles already packed into L, see Fig. 3 (right).
Blocks that include two halves of medium circles are called dense blocks,
while blocks that include one half of a medium circle are called sparse blocks.
The area of L that is neither covered by a dense block or a sparse block is called
a free block. Packing a vertical lane L(cid:48) into a sparse block B means placing L(cid:48)
inside B as far as possible to the left, such that L(cid:48) does not overlap another
vertical lane already packed into B. Packing a vertical lane L(cid:48) into L means
placing L(cid:48) inside L as far as possible to the left, such that L(cid:48) does neither
overlap another vertical lane packed into L, a dense block of L, or a sparse block
of L.
densesparseblockBsparseblockAblocksparseblockDsparseblockEverticallaneCverticallaneF7verticallanes2verticallanes1verticallane2verticallanes 6
S. P. Fekete, C. Scheffer, S. von Hoeveling
We extend our classification of circles by defining classes i of lane widths wi
and relative lower bounds qi for the circles' radii as described in Table 1. This
means a circle with radius r belongs to class 1 if 0.5w ≥ r > q1w1 and to class i
if qi−1wi−1 ≥ r > qiwi, for i ≥ 2. Only circles of class i are allowed to be packed
into lanes of class i.
Class i
(Relative) lower bound qi Lane width wi
q1 := qM := 0.25
q2 := qS := 0.168261
q3 := 0.371446
q4 := 0.190657
q5 := 0.175592
q6 := 0.170699
q7 := 0.169078
q8 := 0.168354
q9 := 0.168293
q10 := 0.168272
q11 := 0.168265
q12 := 0.168263
q13 := 0.168262
. . .
qk := 0.168262
1 (Medium)
2 (Small)
3 (Tiny)
4 (Tiny)
5 (Very tiny)
6 (Very tiny)
7 (Very tiny)
8 (Very tiny)
9 (Very tiny)
10 (Very tiny)
11 (Very tiny)
12 (Very tiny)
13 (Very tiny)
. . .
k (Very tiny)
Table 1. Circles are classified into the listed classes. Note that the lower bounds to
the circles' radii is relative to the lane width, e.g., the absolute lower bound for circles
inside a small lane is qSw2 = 0.168261w2.
w1 := w
w2 := 2qM w = 0.5 · w
w3 := 4qM qSw = 0.168261 · w
w4 := 8qM qSq3w ≈ 0.125 · w
w5 := 16qM qSq3q4w ≈ 0.047664 · w
w6 := 32qM qSq3q4q5w ≈ 0.016739 · w
w7 ≈ 0.005715 · w
w8 ≈ 0.001932 · w
w9 ≈ 0.000651 · w
w10 ≈ 0.000219 · w
w11 ≈ 0.000074 · w
w12 ≈ 0.000025 · w
w13 ≈ 0.000008 · w
. . .
wk := 2k−1qM qSq3q4 · . . . · qk−1w
A sparse block is either free, reserved for class 3, reserved for class 4, reserved
for all classes i ≥ 5, or closed. Initially, each sparse block is free.
We use SLP in order to pack a circle C of class i ≥ 3, into a vertical lane
Li ⊂ L of class i and width wi by applying the Steps 1-5 in increasing order
as described below. When one of the five steps achieves that C is packed into a
vertical lane Li, the approach stops and returns successful.
-- Step (1): If there is no open vertical lane Li ⊂ L of class i go to Step 2.
Assume there is an open vertical lane Li of class i. If C can be packed into Li,
we pack C into Li. Else, we declare Li to be closed.
-- Step (2): We close all sparse blocks B that are free or reserved for class i
in which a vertical lane of class i cannot be packed into B.
-- Step (3): If there is an open sparse block B that is free or reserved for
class i and a vertical lane of class i can be packed into B:
• (3.1): We pack a vertical lane Li ⊂ L of class i into B. If the circle half
that is included in B touches the bottom of L, we apply downwards SLP
to Li. Otherwise, we apply upwards SLP to Li.
Online Circle Packing
7
and i ≥ 5, we reserve B for all classes i ≥ 5.
• (3.2): If B is free and i ∈ {3, 4}, we reserve B for class i. If B is free
• (3.3): We pack C into Li.
• (4.1): We pack a vertical lane Li of class i into L and apply upwards SLP
• (4.2): We pack C into Li
-- Step (4): If a vertical lane of class i can be packed into L:
to Li.
-- Step (5): We declare L to be closed and return failed.
2.4 Double-Sided Structured Lane Packing (DSLP)
We use SLP as a subroutine in order to define our packing strategy Double-Sided
Structured Lane Packing (DSLP) of Theorem 1. In particular, additionally to L,
we consider two small lanes L1, L2 that partition L, see Fig. 5.
Fig. 5. A packing produced by DSLP: The medium lane is packed from left to right by
medium circles. The two contained small lanes are packed simultaneously in parallel
from right to left by small circles.
Rightwards Double-Sided Structured Lane Packing (DSLP) applies the ex-
tended version of rightwards SLP to L and leftwards SLP to L1, L2. If the cur-
rent circle C is medium or (very) tiny, we pack C into L. If C is small, we pack
C into that lane of L1, L2, resulting in a smaller packing length.
Leftwards DSLP is defined analogously, such that the extended version of
leftwards SLP is applied to L and rightwards SLP to L1, L2. Correspondingly,
upwards and downwards DSLP are defined for vertical lanes.
3 Packing into the Unit Square
We extend our circle classification by the class 0 of large circles and define a
relative lower bound q0 := w
2 and the lane width of corresponding large lanes as
w0 := 1 − w.
We set w to 0.288480 and 0.277927 for Theorem 2 respectively Theorem 3.
In order to pack large circles, we use another packing strategy called Tight Lane
Packing (TLP) defined as SLP, but without restrictions (1) and (2), see Fig. 3.
LL1L2ft(L)fb(L) 8
S. P. Fekete, C. Scheffer, S. von Hoeveling
Fig. 6. Left: The unit square is divided into four lanes L1, L2, L3, and L4, into which
medium, small, tiny, and very tiny circles are packed. Large circles are packed into a
lane L0 that overlaps L1, L2, and L3. Right: An example packing. A medium circle
(dotted) does not fit.
We cover the unit square by the union of one large lane L0 and four medium
lanes L1, . . . , Lk for k = 4, see Fig. 6. We apply TLP to L0 and DSLP to
L1, . . . , L4. The applied orientations for L0, L1, L2 are leftwards, rightwards, and
downwards. For i = 3, 4, the orientation for Li is chosen such that the first circle
packed into Li is placed adjacent to the bottom side of Li.
If the current circle to be packed is large, we pack C into the large lane L0
and stop if C does not fit in L0. Otherwise, in increasing order we try to pack C
into L1, . . . , L4.
4 Analysis of the Algorithms
In this section we sketch the analysis of our approaches and refer to the appendix
for full details. First, we analyze the packing density guaranteed by DSLP. Based
on that, we prove our main results Theorems 1, 2, and 3.
4.1 Analysis of SLP
In this section, we provide a framework for analyzing the packing density guar-
anteed by DSLP for a horizontal lane L of width w. It is important to note that
this framework and its analysis in this subsection deals with the packing of only
one class into a lane.
We introduce some definitions. The packing length p(L) is the maximal differ-
ence of x-coordinates of points from circles packed into L. The circle-free length
f (L) of L is defined as (cid:96)(L)− p(L). We denote the total area of a region R ⊂ R2
by area(R) and the area of an a×b-rectangle by R(a, b). Furthermore, we denote
the area of a semicircle for a given radius r by H(r) := π
2 r2. The total area of
the circles packed into R is called occupied area denoted by occ(R). Finally, the
density den(R) is defined by occ(R)/area(R).
In order to apply our analysis for different classes of lanes, i.e., different lower
bounds, we consider a general (relative) lower bound q for the radii of circles
L2L1L3L4L0w1−wwww allowed to be packed into L with 0 < q ≤ 1/2. The following lemma deduces a
lower bound for the density of dense blocks depending on q.
Online Circle Packing
9
Lemma 1. Consider a dense block D containing two semicircles of radii r1
and r2 such that 0 < qw ≤ r1, r2 ≤ 1/2w. Then den(D) is lower-bounded by
(cid:18)
(cid:21)
1
2
δ :
0,
→ R with q (cid:55)→
≈ 0.6046
πq
√
π
3
3
πq2√
4q−1
3
√
0 < q < 1
3
≤ q ≤ 1
√
1
3 < q ≤ 1
2 .
3
1
3
3
Fig. 7. (Left): A plot of δ(q) for its complete range. It provides the minimal density
of a dense block whose two semicircles have a radius of at least q · w. A lower bound
of q = 1/2 leads to a minimal density of π/4 which is the ratio of a circle to its
minimal bounding square. (Right): (1): δ(0.15) ≈ 0.47123, (2): δ( 1
) ≈ 0.6046, and
√
(3): δ(0.4) ≈ 0.6489.
3
3
We continue with the analysis of sparse blocks. Sparse blocks have a minimum
length qw. Lemma 2 states a lower bound for the occupied area of sparse blocks.
This lower bound consists of a constant summand and a summand that is linear
with respect to the actual length.
Lemma 2. Given a density bound δmin ≤ δ(q) for dense blocks. Let S be
a sparse block and z be the lower bound for (cid:96)(S) with (cid:96)(S) ≥ z ≥ qw. Then
occ(S) ≥ R((cid:96)(S) − z, w) · δmin + H(z).
The occupied area of a sparse block is at least a semicircle of a smallest
possible circle plus the remaining length multiplied by the lane width and by
the minimal density δmin of dense blocks. This composition is shown in Fig. 8.
Next, we combine the results of Lemma 1 and Lemma 2. We define the term
(cid:1) := R(p− 2z, w)· δmin + 2·H(z) for some p, w, z, δmin > 0
(cid:0)p, w, z, δmin
minSLP
and state the following (see also Fig. 8 (4)).
Lower Limit qMinimal densityFunctionpqp(33)(pq2)4q-10.00.10.20.30.40.50.00.20.40.60.81.0(1)(2)(3) 10
S. P. Fekete, C. Scheffer, S. von Hoeveling
Fig. 8. (1)+(2)+(3) Three sparse blocks S1, S2, S3 in order of ascending length. The
grey coloured area (light and dark unified) represents the occupied area. The dashed
area shows the lower bound of Lemma 2, which is composed of the smallest possible
semicircle plus a linear part. The dark grey parts symbolize the area that exceeds the
bound, whereas the red parts symbolize the area missing to the bound. Block S1 has
the minimal length p so that the occupied area and the bound are equal. For blocks of
larger length, represented by S2 and S3, the dark grey area is larger than the red area.
(4) A packing produced by SLP and the lower bound of Lemma 3.
Lemma 3. Given a lane L packed by SLP, a lower bound q, and a density bound
δmin ≤ δ(q) for dense blocks. Let w be the width of L. The occupied area in L is
lower-bounded by minSLP
(cid:0)p(L), w, qw, δmin
(cid:1).
4.2 Analysis of DSLP
Let L be a horizontal lane packed by DSLP. We define pt(L) (pb(L)) as the sum
of the packing lengths of the packing inside L and the length of the packing
inside the top (bottom) small lane inside L. Furthermore, we define ft(L) :=
(cid:96)(L) − pt(L) and fb(L) := (cid:96)(L) − pb(L), see Fig. 5.
By construction, vertical dense blocks packed into L have a density of at
least δ(q2). In fact, the definitions of circle sizes for all classes i ≥ 2 ensure the
common density bound δ := δ(q2) for dense blocks.
We consider mixed dense blocks, that were defined as sparse blocks of L in
which vertical lanes are packed, also as dense blocks and extend the lower bound
den(D) ≥ δ to all kinds of dense blocks by the following Lemma.
Lemma 4. Let D be a dense block of L. Assume all vertical lanes packed into
D to be closed. Then den(D) ≥ δ.
As some vertical lanes may not be closed, we upper bound the error O(L)
that we make by assuming that all vertical lanes L1, . . . , Ln ⊂ L are closed.
Lemma 5. O(L1 ∪ . . . ∪ Ln) < 0.213297 · w2.
We lower bound the occupied area inside L by using the following term:
minDSLP (pt(L), pb(L), w, z, δmin) := R(cid:0)pt(L) + pb(L) − w − 4z, w2
(cid:1) · δmin
+2 · H(cid:0) w
(cid:1) + 4H(z),
4
zzS1S2(1)(2)zS3z(3)(4) Online Circle Packing
11
where z denotes the minimal radius for the circles, i.e., z := q2w2.
Applying Lemma 4 and Lemma 2 separately to L, L1, and L2, analogous to
the combination of Lemma 1 and Lemma 2 in the last subsection, and estimating
the error O(L) with Lemma 5, yields lower bounds for the occupied areas of L,
L1, and L2. Fig. 9 separately illustrates the lower bounds for the occupied areas
of L, L1, and L2 for two example packings and Lemma 6 states the result.
Fig. 9. Two example packings and the compositions of our lower bounds (red) for the
occupied area implied by Lemma 6. Note that O(L) is not visualized.
(cid:0)pt(L), pb(L), w, z, δ(cid:1) − O(L) ≥ occ(L).
Lemma 6. minDSLP
4.3 Analysis of Packing Circles into a Rectangle
Given a 1 × b rectangle R, we apply DSLP for packing the input circles into R.
Theorem 1. Let b ≥ 1. Any online sequence of circles with a total area no larger
can be packed into the 1 × b-rectangle R.
than min
This is worst-case optimal for b ≥ 2.36.
0.528607 · b − 0.457876, π
(cid:16)
(cid:17)
4
4 − q2)· δ + π
The lower bound for the occupied area implied by Lemma 6 is equal to
Hence, the online sequence consisting of one circle with a radius of 1
resulting total area of π
Fig. 10. This concludes the proof of Theorem 1.
2 (q2)2 − 0.213297. This is lower bounded by π
4 for b ≥ 2.36.
2 + and
4 + is a worst case online sequence for b ≥ 2.36, see
16 + π
(cid:0)b− 3
Fig. 10. A worst case for packing circles into an 1 × b-rectangle with b ≥ 2.36 consists
of one circle with radius 1
2 + . The shown circle with radius 1
2 just fits.
4.4 Analysis of Packing Circles into the Unit Square
In this section, we analyze the packing density of our overall approach for online
packing circles into the unit square. In order to prove Theorem 3, i.e., a lower
bound for the achieved packing density, we show that if there is an overlap or if
12
S. P. Fekete, C. Scheffer, S. von Hoeveling
Fig. 11. (Left): Different cases for an overlap. Case 0: A single circle is too large
for L0. Case 1: L0 exceeded. Case 2: Overlap in L2. Case 3: Overlap in L3. Case 4:
Overlap in L4 with large circle being involved. Case 5: Overlap in L4 with no large
circle being involved. (Right): A plot of 24 terms for corresponding 24 (sub-)cases for
qS = 0.191578. The point P = (0.277927, 0.375898) is the highest point of the 0-level.
Its y-value is the highest guaranteed packing density for circles with minimal radii of
0.191578 · 0.277927/2 < 0.0266223.
there is no space in the last lane, then the occupied area must be at least this
lower bound. Our analysis distinguishes six different higher-level cases of where
and how the overlap can happen, see Fig. 11 left.
For each of the six cases and its subcases, we explicitly give a density function,
providing the guaranteed packing density depending on the choice of w, see
Fig. 11 right. The shown functions are constructed for the case of no (very)
tiny circles with an alternative q2 = 0.191578, which was chosen numerically in
order to find a high provable density. The w with the highest guaranteed packing
density of 0.375898 is w = 0.277927. This concludes the proof of Theorem 3.
Theorem 3. Any online sequence of circles with radii not smaller than 0.026623
and with a total area no larger than 0.375898 can be packed into the unit square.
With the same idea but with w = 0.288480 and all circles classes, especially
with classes i ≥ 2 as defined in Table 1, we prove Theorem 2.
Theorem 2. Any online sequence of circles with a total area no larger than
0.350389 can be packed into the unit square.
5 Conclusion
We provided online algorithms for packing circles into a square and a rectangle.
For the case of a rectangular container, we guarantee a packing density which is
worst-case optimal for rectangles with a skew of at least 2.36. For the case of a
square container, we provide a packing density of 0.350389 which we improved
to 0.375898 if the radii of incoming circles are lower-bounded by 0.026622.
(1)(2)(3)(4)(5)(5)(5)(0)Width wMinimal density aCase0123.1.13.1.23.1.33.2.13.2.23.2.43.2.53.3.13.3.23.4.13.4.24.1.14.1.24.2.24.2.34.2.45.15.25.35.4.15.4.20.270.280.290.300.310.320.330.340.360.380.400.420.44P Online Circle Packing
13
References
1. B. Brubach. Improved online square-into-square packing. In Proc. 12th Interna-
tional Workshop on Approximation and Online Algorithms (WAOA), pages 47 -- 58,
2014.
2. I. Castillo, F. J. Kampas, and J. D. Pint´er. Solving circle packing problems by
global optimization: numerical results and industrial applications. European J. of
Operational Research, 191(3):786 -- 802, 2008.
3. C. de Groot, R. Peikert, and D. Wurtz. The optimal packing of ten equal circles
in a square. Technical Report 90-12, ETH Zurich, Switzerland, 1990.
4. E. D. Demaine, S. P. Fekete, and R. J. Lang. Circle packing for origami design is
hard. In Proceedings 5th International Conference on Origami in Science, Mathe-
matics and Education (Origami5), pages 609 -- 626. A. K. Peters/CRC Press, 2010.
5. S. P. Fekete and H. Hoffmann. Online square-into-square packing. Algorithmica,
77(3):867 -- 901, 2017.
6. S. P. Fekete, P. Keldenich, and C. Scheffer. Packing disks into disks with optimal
worst-case density. In Proc. 35th Symposium on Computational Geometry (SoCG),
2019. to appear.
7. S. P. Fekete, S. Morr, and C. Scheffer. Split packing: Algorithms for packing circles
with optimal worst-case density. Discrete & Computational Geometry, 2018. Online
First at https://doi.org/10.1007/s00454-018-0020-2.
8. H. J. Fraser and J. A. George. Integrated container loading software for pulp and
paper industry. European J. of Operational Research, 77(3):466 -- 474, 1994.
9. J. A. George, J. M. George, and B. W. Lamar. Packing different-sized circles into a
rectangular container. European J. of Operational Research, 84(3):693 -- 712, 1995.
10. J. Y. T. Leung, T. W. Tam, C. S. Wong, G. H. Young, and F. Y. Chin. Packing
squares into a square. Journal of Parallel and Distributed Computing, 10(3):271 --
275, 1990.
11. M. Locatelli and U. Raber. Packing equal circles in a square: a deterministic global
optimization approach. Discrete Applied Mathematics, 122(1):139 -- 166, 2002.
12. M. C. Mark´ot and T. Csendes. A new verified optimization technique for the"
packing circles in a unit square" problems. SIAM J. on Optimization, 16(1):193 --
219, 2005.
13. J. W. Moon and L. Moser. Some packing and covering theorems. Colloquium
Mathematicae, 17(1):103 -- 110, 1967.
14. J. K. Nurmela and J. P. R. Ostergard. More optimal packings of equal circles in a
square. Discrete & Computational Geometry, 22(3):439 -- 457, 1998.
15. N. Oler. A finite packing problem. Canad. Mathematical Bulletin, 4:153 -- 155, 1961.
16. J. Schaer. The densest packing of nine circles in a square. Canad. Mathematical
Bulletin, 8:273 -- 277, 1965.
17. J. Schaer and A. Meir. On a geometric extremum problem. Canad. Mathematical
Bulletin, 8(1), 1965.
18. K. Sugihara, M. Sawai, H. Sano, D. S. Kim, and D. Kim. Disk packing for the
estimation of the size of a wire bundle. Japan Journal of Industrial and Applied
Mathematics, 21(3):259 -- 278, 2004.
19. P. G. Szab´o, M. C. Mark´ot, T. Csendes, E. Specht, L. G. Casado, and I. Garc´ıa.
New Approaches to Circle Packing in a Square: With Program Codes, volume 6 of
Springer Optimization and Its Applications. Springer, 2007.
20. D. Wurtz, M. Monagan, and R. Peikert. The history of packing circles in a square.
Maple Technical Newsletter, pages 35 -- 42, 1994.
|
1708.04862 | 1 | 1708 | 2017-08-16T12:47:19 | New Approximations for Coalitional Manipulation in General Scoring Rules | [
"cs.DS"
] | We study the problem of coalitional manipulation---where $k$ manipulators try to manipulate an election on $m$ candidates---under general scoring rules, with a focus on the Borda protocol. We do so both in the weighted and unweighted settings. We focus on minimizing the maximum score obtainable by a non-preferred candidate.
In the strongest, most general setting, we provide an algorithm for any scoring rule as described by a vector $\vec{\alpha}=(\alpha_1,\ldots,\alpha_m)$: for some $\beta=O(\sqrt{m\log m})$, it obtains an additive approximation equal to $W\cdot \max_i \lvert \alpha_{i+\beta}-\alpha_i \rvert$, where $W$ is the sum of voter weights.
For Borda, both the weighted and unweighted variants are known to be $NP$-hard. For the unweighted case, our simpler algorithm provides a randomized, additive $O(k \sqrt{m \log m} )$ approximation; in other words, if there exists a strategy enabling the preferred candidate to win by an $\Omega(k \sqrt{m \log m} )$ margin, our method, with high probability, will find a strategy enabling her to win (albeit with a possibly smaller margin). It thus provides a somewhat stronger guarantee compared to the previous methods, which implicitly implied a strategy that provides an $\Omega(m)$-additive approximation to the maximum score of a non-preferred candidate.
For the weighted case, our generalized algorithm provides an $O(W \sqrt{m \log m} )$-additive approximation, where $W$ is the sum of voter weights. This is a clear advantage over previous methods: some of them do not generalize to the weighted case, while others---which approximate the number of manipulators---pose restrictions on the weights of extra manipulators added.
Our methods are based on carefully rounding an exponentially-large configuration linear program that is solved by using the ellipsoid method with an efficient separation oracle. | cs.DS | cs |
New Approximations for Coalitional
Manipulation in General Scoring Rules
Orgad Keller
[email protected]
Department of Computer Science, Bar-Ilan University, Israel
Avinatan Hassidim
[email protected]
Department of Computer Science, Bar-Ilan University, Israel
Noam Hazon
[email protected]
Department of Computer Science, Ariel University, Israel
Abstract
We study the problem of coalitional manipulation -- where k manipula-
tors try to manipulate an election on m candidates -- under general scoring
rules, with a focus on the Borda protocol. We do so both in the weighted
and unweighted settings.
For these problems, recent approaches to approximation tried to min-
imize k, the number of manipulators needed to make the preferred candi-
date win (thus assuming that the number of manipulators is not limited in
advance), we focus instead on minimizing the maximum score obtainable
by a non-preferred candidate.
√
In the strongest, most general setting, we provide an algorithm for
for some
any scoring rule as described by a vector (cid:126)α = (α1, . . . , αm):
m log m), it obtains an additive approximation equal to W ·
β = O(
maxiαi+β − αi, where W is the sum of voter weights.
In words, this
factor is the maximum difference between two scores in (cid:126)α that are β
entries away, multiplied by W . The unweighted equivalent is provided as
well.
√
For Borda, both the weighted and unweighted variants are known to
√
be NP-hard. For the unweighted case, our simpler algorithm provides
a randomized, additive O(k
m log m) approximation; in other words, if
there exists a strategy enabling the preferred candidate to win by an
m log m) margin, our method, with high probability, will find a strat-
Ω(k
egy enabling her to win (albeit with a possibly smaller margin). It thus
provides a somewhat stronger guarantee compared to the previous meth-
ods, which implicitly implied (with respect to the original k) a strategy
that provides an Ω(m)-additive approximation to the maximum score of
1
a non-preferred candidate: when k is o((cid:112)m/ log m), our strategy thus
provides a stronger approximation.
For the weighted case, our generalized algorithm provides an O(W
√
m log m)-
additive approximation, where W is the sum of voter weights. This is a
clear advantage over previous methods: some of them do not general-
ize to the weighted case, while others -- which approximate the number
of manipulators -- pose restrictions on the weights of extra manipulators
added.
We note that our algorithms for Borda can also be viewed as a (1 +
o(1))-multiplicative approximation since the values we approximate have
natural Ω(km) (unweighted) and Ω(W m) (weighted) lower bounds.
Our methods are novel and adapt techniques from multiprocessor
scheduling by carefully rounding an exponentially-large configuration lin-
ear program that is solved by using the ellipsoid method with an efficient
separation oracle. We believe that such methods could be beneficial in
social choice settings as well.
1
Introduction
Elections are one of the pillars of democratic societies, and are an important
part of social choice theory. In addition they have played a major role in mul-
tiagent systems, where a group of intelligent agents would like to reach a joint
decision [9]. In its essence, an election consists of n agents (also called voters)
who need to decide on a winning candidate among m candidates. In order to
do so, each voter reveals a ranking of the candidates according to his preference
and the winner is then decided according to some protocol.
Ideally in voting, we would like the voters to be truthful, that is, that their
reported ranking of the candidates will be their true one. However, almost
all voting rules are prone to manipulation: Gibbard and Satterthwaite [11, 22]
show that for any reasonable preference-based voting system with at least 3
candidates, voters might benefit from reporting a ranking different than their
true one in order to make sure that the candidate they prefer the most wins.
Furthermore, several voters might decide to collude, to form a coalition and then
to coordinate their votes in such a way that a specific candidate p (hereafter the
preferred candidate) will prevail. Such a setting is reasonable especially when
the voters are agents that are operated by one party of interest.
For some time, the hope for making voting protocols immune to manipula-
tions at least in practice relied on computational assumptions: for several com-
mon voting protocols, it was shown that computing a successful voting strategy
for the manipulators is NP-hard [3, 6, 25, 10]. However, as it is many times the
case, approximation algorithms and heuristics were devised in order to overcome
the NP-hardness albeit with some compromises on the quality of the resulting
strategy. This paper fits within this scheme.
In this paper we focus on general scoring rules R(cid:126)α and in particular on the
Borda voting rule. We first study the problem of (constructive) unweighted
2
coalitional manipulation (UCM )1: assume that k additional voters (hereafter
the manipulators), all of them preferring a specific candidate p, can be added
to the voting system, thus forming a coalition. Also assume that all n original
voters (hereafter the non-manipulators) voted first (or equivalently, that the
non-manipulators are truthful and that their preferences are known). Find a
strategy for the manipulators telling each one of them how to vote so that p
wins, if such strategy exists. We call such a strategy a p-winning strategy. In
the weighted variant (WCM ), the manipulators are weighted; essentially this
means that points awarded by a voter to a candidate are multiplied by the
voter's weight.
R(cid:126)α-WCM, for all positional scoring rules R(cid:126)α, except plurality-like rules, was
shown to be NP-hard when m ≥ 3 [6, 13, 21]. Therefore, Borda-WCM is NP-
hard. Borda-UCM eluded researchers for some time; it was first conjectured
and finally proven to be NP-hard [7, 4].
As a way of overcoming the hardness, recent research [26] focused on an
approximation to the minimum number of manipulators needed to be added to
the system in order to guarantee that the preferred candidate p would win. For
Borda-UCM, they showed that if there exists a p-winning strategy for k manip-
ulators, then they will find a p-winning strategy with at most one additional
manipulator -- besides the k given by the problem definition. For Borda-WCM,
they showed that if there exists a p-winning strategy using the k given weighted
manipulators, they will find a p-winning strategy using additional manipulators,
if the sum of weights of the additional manipulators equals the maximum over
the weights of the k original manipulators.
This kind of approximation might seem a bit problematic: first, the ability
to add a manipulator is a strong operation, perhaps too strong; for instance,
for Borda-UCM, adding a manipulator adds Ω(m) to the difference between
p and its highest-scoring competitor. Second, while in some cases it might be
reasonable that the party behind the manipulators can add another manipulator
to the system, in many cases we do not expect this to be true. Furthermore,
in the weighted case assumptions are needed to be made on the weight of the
additional manipulators -- also a problematic aspect. Instead, it is interesting
to ask what can we assert -- assuming that the number of manipulators cannot
be changed -- on the ability to promote a specific candidate p given the non-
manipulator scores of all candidates and the value k (or equivalently, the length-
k vector of manipulator weights).
We provide a positive result of the following type:
Main Result: If there exists a manipulation strategy enabling p to
win by a large-enough margin, we efficiently find a successful ma-
nipulation strategy making p win.
Take the unweighted case as an example: assume that we can provide, for some
function f (k, m), an f (k, m)-additive approximation to the maximum difference
1The problem was called CCUM, for "constructive coalitional unweighted manipulation",
in [26].
3
obtainable between p's final score and the final score of the highest-scoring non-
preferred candidate. Then, if there exists a p-winning strategy such that this
difference is at least f (k, m), we can be rest-assured that the algorithm will find
a p-winning strategy.
This, in turn, boils down to approximating an upper-bound to the score
of the highest-ranked candidate who is not p. Earlier research of this flavor
for R(cid:126)α-
focused only on cases where the number of candidates is bounded:
WCM, Brelsford et al. [5] provide an FPTAS to that upper bound (to be exact,
they provide an FPTAS to the same exact value we defined, and then use it to
provide another FPTAS to another value, which is their value-of-interest.2) For
R(cid:126)α-UCM, if the number of candidates is bounded, the entire problem becomes
easy and polynomial-time solvable [6, Proposition 1]. Compared to this line of
work, we do not limit ourselves to bounded number of candidates.
1.1 Our Results and Contributions
Consider a general positional scoring rules R(cid:126)α, as is usually described by a
vector (cid:126)α = (α0, . . . , αm) (see Section 2.1 for full definitions). Now let T ∗ =
minS maxc(cid:48)∈C\{p} scoreR(cid:126)α,E,S(c(cid:48)) be the minimum possible score (ranging over
all possible manipulation strategies S) of the highest scoring candidate who is
not p, where scoreR(cid:126)α,E,S(c(cid:48)) is the final score of a candidate c(cid:48) (w.r.t. voting
rule R(cid:126)α, election E, and strategy S) and C is the candidate set.
√
Our main technical contribution is a constructive proof to the following
m log m for some constant d, and let g((cid:126)α) =
two theorems. Let β = d
maxi=0,...,m−βαi+β − αi. In words, g((cid:126)α) is the biggest difference between a
score in (cid:126)α and another score β entries away from it.
Theorem 1. There exists a randomized Monte Carlo algorithm for R(cid:126)α-UCM
which provides a k · g((cid:126)α)-additive approximation to T ∗ with an exponentially-
small failure probability.
Theorem 2. There exists a randomized Monte Carlo algorithm for R(cid:126)α-WCM
which provides a W · g((cid:126)α)-additive approximation to T ∗ with an exponentially-
small failure probability, where W is the sum of voter weights.
These theorems immediately pave the way to the following corollaries:
Corollary 3. There exists a randomized Monte Carlo algorithm for Borda-
m log m)-additive approximation to T ∗ with an
UCM which provides an O(k
exponentially-small failure probability.
√
Corollary 4. There exists a randomized Monte Carlo algorithm for Borda-
m log m)-additive approximation to T ∗ with an
WCM which provides an O(W
exponentially-small failure probability.
√
2They are interested in the difference between the score of the preferred candidate and
the highest-scoring non-preferred candidate when including the manipulator votes, minus the
same difference when not including the manipulator votes. Notice that the upper-bound we
defined is the only non-trivial value in this computation.
4
Taking Borda-UCM as an example, if there exists a p-winning strategy en-
abling p to win by a margin of Ω(k
m log m) compared to the score of the
highest-scoring non-preferred candidate, our method will find a p-winning strat-
egy (albeit with a possibly smaller margin). Similar guarantees apply in the
more general settings.
√
Notice that for Borda, such approximations can also be seen as a (1 + o(1))-
multiplicative approximation on the score of the highest-scoring non-preferred
candidate, and thus is superior to an FPTAS; to see that, notice that for Borda-
WCM the overall 'voting mass' given by the manipulators is Ω(W m2), and so
the highest scoring candidate has score of at least Ω(W m). Therefore O(W
m)
3 is a lower order term. This is an advantage over the previous methods:
√
• Opposed to the heuristics in [8], we provide provable guarantees.
• Also opposed to the heuristics in [8], our algorithm generalizes to the
weighted case.
• Compared to the reverse algorithm of [26] for Borda-UCM, while adding
only a single extra manipulator sounds like a minor operation, it is not;
as mentioned, an extra manipulator implies the addition of Ω(m) points
to the difference between p and its highest-scoring competitor.
• Consider the reverse algorithm of [26], and assume that adding extra
manipulators is not allowed. We will show that their method implies no
better than an Ω(m)-additive approximation to the score of the highest-
scoring non-preferred candidate. Our approximation is thus superior when
k is o((cid:112)m/ log m).
• Compared to previous methods, are results are linear-programming-based,
and not greedy. Thus, they make a decision based on the entire input,
as opposed to repeatedly making a decision based on a greedy estimate
w.r.t. some subset of the problem.
The following claim analyzes reverse according to our metric. It is proven
in Section 3.
Claim 5. For any m, when the addition of more than k extra manipulators
is not allowed, there are families of cases in which the optimal strategy enables
p to win by a margin of at least m/3, but reverse fails to find a p-winning
strategy.
Our techniques are novel: they employ the use of configuration linear pro-
grams (C-LP), a method that is also used in the scheduling literature, namely for
two well-studied problems, the problem of scheduling on unrelated machines [24],
and the so-called Santa Claus problem [2]. See Section 1.2 for a discussion of
these problems. It is important to note that the solutions to the two above prob-
lems and to ours all differ from one another with respect to how the algorithm
proceeds once the C-LP result is computed.
3The O notation suppresses poly-logarithmic factors.
5
C-LPs are used for the generation of an initial, invalid strategy, which is
later modified to become valid. They are unique in the sense that they are
linear programs that have an exponential number of variables, an issue which
we solve by referring to the LP dual and using the ellipsoid method with a
polynomially-computable separation oracle [17, 12]. We have also implemented
our algorithm: as a result of not finding a library which enables solving an LP
this way, we simulated this by an iterative use of a general LP-solving library,
each time adding a violated constraint based on running the separation oracle
externally.
1.2 Related Work
Borda. The Borda voting mechanism was introduced by Jean-Charles de
Borda in 1770. It is used, sometimes with some modifications, by parliaments
in countries such as Slovenia, and competitions such as the Eurovision song
contest, selecting the MVP in major league baseball, Robocup robot soccer
competitions and others. The Borda voting mechanism is described as follows:
every agent ranks the candidates from 1 to m, and awards the candidate ranked
i-th a score of m − i. Notice that this makes the scores given by each single
voter a permutation of 0, . . . , m − 1. Finally, the winning candidate is the one
with the highest aggregate score.
Easiness Results. The computational complexity of coalitional manipulation
problems was studied extensively. For general scoring rules R(cid:126)α, most earlier
work considered the case where the number of candidates is bounded: Conitzer
at al. [6] show that when m is bounded, R(cid:126)α-UCM is solvable in polynomial
time.
Even when m is unbounded, Plurality-UCM and Veto-UCM are still easy
using the greedy algorithm of Zuckerman et al. [26]. This also holds for t-
approval-UCM, which generalizes both [19].
NP-Hardness Results.
In the weighted case, the situation is different: for
all positional scoring rules R(cid:126)α, except plurality-like rules, R(cid:126)α-WCM is NP-hard
when m ≥ 3 [6, 13, 21]. In particular, this holds for Borda-WCM.
However, the computational hardness of Borda-UCM still remained open for
quite some time, until finally shown to be NP-hard as well [7, 4] in 2011, even
for the case of n = 3 and adding 2 manipulators.
Approximating the number of manipulators. Zuckerman et al. [26] present
a greedy algorithm later referred to as reverse4. reverse works as follows:
after the non-manipulators had finished voting, we go over the manipulators
one by one, and each manipulator will rank the candidates (besides p) by the
reversed order of their aggregated score so far (candidate with the highest score
so far gets the lowest ranking). As mentioned, for Borda-UCM, reverse can be
4This name was given in [7].
6
seen as an additive +1 approximation for the objective of finding the minimum
number of manipulators needed.
For Borda-WCM, their approximation has the following flavor. Let S be the
set of the k given weighted manipulators. If there exists a p-winning strategy
using S, they will find a p-winning strategy using additional manipulators, if
the sum of weights of the additional manipulators equals max(cid:96)∈S w(cid:96), where w(cid:96)
is the weight of manipulator (cid:96) ∈ S.
Approximating the Maximum Score of a Non-Preferred Candidate.
Returning to earlier results, when m is bounded, Brelsford et al. [5, Lemma
3] provide an FPTAS with respect to the maximum score of a non-preferred
candidate. As mentioned, this paves the way for an FPTAS on their value-of-
interest.
Heuristics for Borda. Davies et al. [8] present two additional heuristics:
iteratively, assign the largest un-allocated score to the candidate with the largest
gap (Largest Fit), or to the candidate with the largest ratio of gap divided
by the number of scores yet-to-be-allocated to this candidate (Average Fit).
To the best of our knowledge, these algorithms do not have a counterpart for
the weighted case.
Configuration Linear Programs. As discussed, configuration linear pro-
grams were also used in scheduling literature, for example for the following two
problems which were extensively studied before:
• In the so-called Santa Claus problem [2], Santa Claus has t presents that
he wishes to distribute between m kids, and pi,j is the value that kid i has
to present j. The goal is to maximize the happiness of the least happy
kid: mini
pi,j, where Si is the presents allocated to kid i.
(cid:80)
j∈Si
• In the problem of scheduling on unrelated machines [24]. We need to assign
t jobs between m machines, and pi,j is the time required for machine i to
execute job j. The goal is to minimize the makespan maxi
pi,j,
where Si is the jobs assigned to machine i.
(cid:80)
j∈Si
Both papers researched a natural and well-researched 'restricted assignment'
variant of the two problems where pi,j ∈ {pj, 0}.
In [2], they obtained an
O(log log m/ log log log m)-multiplicative approximation to the first problem and
in [24], they obtained a (33/17 + )-multiplicative approximation to the second.
2 Preliminaries
2.1 Problem Definition
Candidate Set. With a slight change of notation, let C = {c0, c1, . . . , cm} be
a candidate set consisting of the preferred candidate p = c0 and the other m
7
candidates c1, . . . , cm. Note that we changed the notation so that the overall
number of candidates is m + 1; this will help streamline the writing.
Election. An election E = (C, V ) is defined by a candidate set C and a set of
voters V where each voter submits a ranking of the candidates according to its
preference. Formally, we define V = {(cid:31)1, . . . ,(cid:31)n}, where each (cid:31)(cid:96)∈ V is a total
order of the candidates. For example, c1 (cid:31)(cid:96) p (cid:31)(cid:96) c2 is one such possible order
if C = {p, c1, c2}. Then, some decision rule R is applied in order to decide on
the winner(s); formally R(E) ⊆ C is the set of winners of the elections. In the
specific case of a positional scoring rule R(cid:126)α, the rule is described by a vector
(cid:126)α = (α0, α1, . . . , αm) for which α0 ≤ α1 ≤ ··· ≤ αm, and αm is polynomial in
m, used as follows: each voter awards αi to the candidate ranked (m − i)-th5.
Finally, the winning candidate is the one with the highest aggregated score. In
the specific case of Borda scoring rule, we have that (cid:126)α = (0, 1, . . . , m − 1, m).
In the R(cid:126)α-(constructive) weighted coalitional manipula-
WCM and UCM.
tion (R(cid:126)α-WCM) problem, we are given as input:
• A score profile vector (σ0, σ1, . . . , σm) representing the aggregated scores
given so far to each candidate in C by the original voters in an election
E under the rule R(cid:126)α. Notice that (σ0, σ1, . . . , σm) eliminates the need
for obtaining E as input: as we have no control on the truthful voters
V , (σ0, σ1, . . . , σm) is thus a sufficient representation for the outcome of
non-manipulator votes.
• A vector (cid:126)w = (w1, . . . , wk) of positive integers representing the weights
of k manipulators who will be added to the election. The weights have
the following meaning: each manipulator (cid:96) = 1, . . . , k is replaced by w(cid:96)
identical but unweighted copies of himself.
It then should be determined if when adding the k manipulators then either
(a) no strategy under R(cid:126)α exists in which p wins, or that (b) there exists a voting
strategy under R(cid:126)α such that p can win. In this case, the algorithm should find
it. R(cid:126)α-(constructive) unweighted coalitional manipulation (R(cid:126)α-UCM) is the
specific case where (cid:126)w is the all-ones vector and therefore can be replaced in the
input by the integer k.
Manipulation Matrices. Note that in case (b), the output is a voting strat-
egy S which can be represented as a k × (m + 1) matrix in which the entry S(cid:96),i
describes the score given by manipulator (cid:96) to candidate ci, and where each row
of S is a permutation of (cid:126)α. Such a representation is also called a manipulation
matrix. We can relax the requirement that each row of S is a permutation,
and replace it by the requirement that each score-type αj, is repeated exactly
k times in S. Such a matrix is called a relaxed manipulation matrix. We can
5Usually (cid:126)α is defined in a descending manner, such that α0 ≥ α1 ≥ · · · ≥ αm, and αi is
awarded to the candidate ranked i-th. Our choice helps streamline the presentation.
8
perform this relaxation as Davies et al. [8, Theorem 7] show that each relaxed
manipulation matrix can be rearranged to become a valid manipulation matrix
while preserving each candidate's final score.
High Probability. Throughout the paper, when we use the term 'with high
probability', we mean an arbitrarily-chosen polynomially-small failure probabil-
ity, i.e., success probability of the form 1 − m−d where d ≥ 1 is a constant that
can be chosen without affecting the asymptotic running time.
'Failure' refers
to the event that the algorithm does not provide the desired approximation
guarantee.
In this paper we will use various forms of the Hoeffding inequalities, which
are variants of the Chernoff inequalities:
Generalized Hoeffding inequality [14, Theorem 2]. Let X1, . . . , Xm be
independent random variables where each Xi is bounded by the interval [ai, bi]
respectively. Let X =(cid:80)m
i=1 Xi and ¯X = X/m. Then
2n2t2
Pr[ ¯X − E[ ¯X] ≥ t] ≤ exp
−
(cid:80)m
i=1(bi − ai)2
By redefining the above inequality in terms of the sum X (instead of the mean
¯X) and defining λ = tm we derive the following equivalent formulation:
(cid:18)
(cid:18)
(cid:19)
(cid:19)
.
.
2λ2
(cid:80)m
i=1(bi − ai)2
(cid:19)
(cid:18)
− 2λ2
m
.
Specifically, when Xi ∈ {0, 1}, we obtain the following 'classic' Hoeffding in-
equality, which is equivalent to [14, Theorem 1]:
Pr[X − E[X] ≥ λ] ≤ exp
−
Pr[X − E[X] ≥ λ] ≤ exp
2.2 Reduction to a Pure Min-Max Problem
Since we know what will be the final score of p (non-manipulator votes are known
and each manipulator will give p the maximum score possible), we can effectively
discard p and treat the problem as a minimization problem on the final scores
of c1, . . . , cm only. In other words, we focus on finding minS maxc(cid:48)∈C\{p} s(c(cid:48)) =
minS maxi=1,...,m s(ci), where s(c(cid:48)) is c(cid:48)'s final score. Thus the output S is
actually a k × m relaxed manipulation matrix.
Another thing to note is that we can not assume anything about the values in
the initial score profile (σ0, σ1, . . . , σm); this follows from [8, Lemma 1], where
it is shown that in the context of Borda, for any given non-negative integer
vector (σ0, σ1, . . . , σm), we can define a set of non-manipulators (along with
their preferences) and an additional candidate cm+1 that will induce an initial
score profile (T (cid:48) + σ1, . . . , T (cid:48) + σm, y) for some values T (cid:48) and y < T (cid:48). Since such
an additive translation and the addition of a candidate that will be awarded less
9
than any other candidate have no influence on the winner (and on the difference
between each two candidates' final scores), and since our results are concerned
with an additive approximation, we should assume no prior limitation on the
nature of values in (σ0, σ1, . . . , σm).
3 Lower Bound for REVERSE
We start by showing that there are cases in which the reverse algorithm for
Borda gives only an Ω(m)-additive approximation to minimum final score of the
highest-scoring non-preferred candidate.
Proof of Claim 5. We provide an infinite family of cases where the claim holds.
Let k = 3 and let m = 3t for some integer t. Consider the case where after
the non-manipulators voted, all candidates (but p) have the same score σi = s
for all i. Effectively this can be normalized to (σ1, . . . , σm) = (cid:126)0.
By the reverse algorithm, the first manipulator can award c1, . . . , cm with
0, . . . , m − 1 respectively, after which the second manipulator will be obliged to
award c1, . . . , cm with m − 1, . . . , 0 respectively. Repeat this process with the
rest of the manipulators, until the final one. It can be verified that cm will end
up with the maximal score of (cid:100)k/2(cid:101)(m − 1) = 2(m − 1).
Conversely, as an upper bound for an optimal solution, consider the following
strategy: place all scores to be given in a descending sequence, that is the
sequence (cid:104)m − 1, m − 1, m − 1, m − 2, m − 2, m − 2, . . . , 0, 0, 0(cid:105). Give the first
m scores in the sequence to c1, . . . , cm respectively, the next m to cm, . . . , c1
respectively, and the last m to c1, . . . , cm respectively. Since every score-type has
3 copies, we have just described a relaxed manipulation matrix and therefore by
Davies et al. [8, Theorem 7] it can be rearranged to become a valid manipulation
matrix without changing the final score of each candidate. Now notice that the
score given to any candidate is of the form (m− r) + (m/3 + r− 1) + (m/3− r) =
5m/3 − r − 1 for some r ∈ {1, . . . , m/3}. As this is at most 5m/3 − 2 (when
r = 1), the difference is thus m/3 = Ω(m).
4 Linear Programming for UCM
We will begin by providing a "natural" way to formulate the min-max version
of the problem as an Integer Program (IP). As solving IPs is NP-hard, we will
relax it to the equivalent Linear Program (LP). However, such a natural LP will
not be useful in our setting, and we will thus introduce a totally different LP
formulation, called Configuration Linear Programming (C-LP). The number of
variables in the C-LP is exponential in the size of the input. Nevertheless, we
show that our C-LP can be solved in polynomial time.
Let [m] = {1, . . . , m} and [m]0 = {0, . . . , m − 1}. We define the variables
xi,j for (i, j) ∈ [m] × [m]0, and the variable T , with the intent that xi,j will
equal the number of times candidate ci received a score of αj, and T will serve
10
as the upper-bound on each candidate final aggregate score. The IP can then
be stated as follows:
T
min
(cid:126)x
subject to:
xi,j = k
i=1
m(cid:88)
m−1(cid:88)
m−1(cid:88)
j=0
xi,j = k
αjxi,j ≤ T − σi
∀j ∈ [m]0 ,
∀i ∈ [m] ,
∀i ∈ [m] ,
(1)
(2)
(3)
(4)
j=0
xi,j ∈ {0, . . . , k}
∀i ∈ [m], j ∈ [m]0 ,
where (1) guarantees that every score was awarded k times, (2) guarantees
that every candidate was given k scores, and (3) guarantees that every candidate
gets at most T points.
It should be noted that when treating the problem as a min-max problem,
we need to take T as a variable that we wish to minimize (this is done by the
objective function). However, if we consider the original definition in which our
aim is to make the preferred candidate p win, T can be set to σ0 + kαm (the
final score of the preferred candidate), and the IP will not have an objective
function.
4.1
Integrality Gap of the Natural LP
While we can relax this IP into an LP by replacing the set in the last constraint
to be the continuous interval [0, k], it will not be as helpful. We will shortly
show why; however note that as this sub-section relates to the deficiencies of
the original LP formulation, it can be safely skipped and is not needed for the
full understanding of our algorithms.
Consider a "pure" LP rounding algorithm, applied w.l.o.g. to a minimization
problem. Such an algorithm works as follows: given an instance of a problem,
it solves its associated relaxed LP (the problem's natural IP where integrality
constraints are replaced with their continuous counterparts) and then rounds
the resulting solution in some way or the other such that a valid (non-necessarily
optimal) solution to the original IP is obtained. The approximation analysis of
such algorithms is based on reasoning about how worse is the objective value
of the rounded solution compared to the fractional one. In other words, what
is the increase -- or "damage done" -- to the optimum objective incurred by the
rounding process. Since the fractional optimum of a relaxed LP is a lower bound
to the integral optimum, i.e., the optimum of the original problem, the same fac-
tor also upper-bounds the difference between the objective value of the rounded
solution and the one of the integral optimum. Thus this process derives an
11
approximation guarantee. We show that in our case, the increase can be Ω(m),
by showing that there are cases in which the difference between the integral
and fractional optimum objective values is Ω(m), and thus an algorithm solely
based on the rounding procedure cannot hope for an o(m) additive approxima-
tion. This kind of reasoning is known as an integrality gap, and is demonstrated
by the following:
Lemma 6. For Borda-UCM, an algorithm based solely on rounding the relaxed
natural LP cannot obtain o(m) additive approximation.
Proof. We show a lower bound on the approximation ratio in the form of an
additive integrality gap. In other words, we show an infinite family of instances
where the integral solution to the LP (and thus, to the original problem) gives
Ω(m) worse objective value when compared to the fractional solution.
Consider the simple case of m candidates, all having equal initial score
(w.l.o.g. σi = 0 for all i = 1, . . . , m) and a single manipulator. When solving the
problem, one candidate will be awarded m − 1 and thus will have final score of
m− 1. However, in the fractional solution, the optimum is obtained by splitting
each score equally, that is, setting xi,j = 1/m for every i ∈ [m] and j ∈ [m]0.
j=0 j = (m − 1)/2.
Therefore notice that the gap between the objective of the integral and frac-
tional solutions is (m − 1)/2 = Ω(m).
Now every candidate obtained a final score of 1/m ·(cid:80)m−1
4.2
Introducing Configuration LPs
In order to work around this we will have to resort to a totally different approach,
in which variables no longer represent score types, and instead represent the set
of scores (configuration) that can be awarded to a candidate.
Formally, a configuration C for some candidate ci is a vector of dimension
m in which Cj represents a number of scores of type αj that i has received, and
j=0 Cj = k, that is, the overall number of scores awarded is k. For
a candidate ci and a bound T , let Ci(T ) be the set of configurations that do not
cause the candidate overall score to surpass T , i.e., the set of configurations C
We formulate the configuration LP as follows:
for which(cid:80)m−1
for which(cid:80)m−1
j=0 Cjαj ≤ T − σi.
(cid:88)
(cid:88)
xi,C ≤ 1
C∈Ci(T )
Cjxi,C ≥ k
i,C
C∈Ci(T )
xi,C ≥ 0
∀i ∈ [m] ,
∀j ∈ [m]0 ,
∀i ∈ [m], C ∈ Ci(T ) .
(5)
(6)
(7)
where we wish that the xi,C's would serve as indicator variables indicating
whether or not ci was awarded with configuration C, (5) guarantees that every
candidate was given at most 1 configuration and (6) guarantees that every score
12
was awarded at least k times. The choice of inequalities over equalities will be
explained soon.
Example 1. Consider the case where k = 2, m = 5, (cid:126)α = (0, 1, 2, 3, 4) (i.e.,
Borda) and (σ1, . . . , σ5) = (5, 6, 6, 6, 7). We are omitting the non-manipulator
votes that provided (cid:126)σ, however recall that there is a possible non-manipulator
voting yielding any (cid:126)σ up to an additive factor and an addition of a candidate.
Now assume T = 10 (this is indeed the optimum). Let us focus on the last
candidate c5. C5(T ) should therefore contain all configurations which award c5
at most T − σ5 = 3 points. Those configurations are (2, 0, 0, 0, 0) (0 points),
(1, 1, 0, 0, 0) (1 point), (0, 2, 0, 0, 0), (1, 0, 1, 0, 0) (2 points), and (1, 0, 0, 1, 0) (3
points).
When solving the C − LP , only two of her configurations will get non-zero
value: x5,(1,0,0,1,0) (cid:117) 0.7 and x5,(0,1,1,0,0) (cid:117) 0.3. We omit the variables corre-
sponding to the rest of the candidates.
After solving the LP, we will execute a rounding procedure that will trans-
form the fractional LP solution into a valid solution for the original problem.
This procedure can increase the score of some of the candidates, and hence we
wish to start with the smallest possible T (so that even after the increase the
final score will hopefully be bounded by σ0 + kαm).
To find the smallest possible T , we perform a one-sided binary search on
the value of T . For this purpose, for each possible value of T that we come
across during the binary search, we redefine the LP and then solve the new LP
from scratch, and see if it has a valid solution. The reason we do not add T
as a variable in an objective function (instead of the binary search) is that the
number of summands in Equations (5,6) depends on T .
This formulation has the obvious drawback that the number of variables is
exponential in k. However, following the approach of [2], if we find a polynomially-
computable separation oracle we can solve the LP by referring to the LP dual
and using the ellipsoid method. Such an oracle will require a solution to the
following seemingly unrelated problem as a subroutine: a variant of the classic
Knapsack problem.
4.3 k-Multiset Knapsack
Let {1, . . . , m} be a set of distinct items, where each item has an associated
value vj and a weight wj. We also obtain a weight upper-bound W and a value
lower-bound V . As opposed to ordinary knapsack, we also obtain an integer k.
We are required to find a multiset S of exactly k items (i.e., we can repeat items
from the item-set), such that S's overall weight is at most W and S's overall
value is greater than V (or to return that no such multiset exists).
Lemma 7. The k-multiset knapsack can be solved in time polynomial in k, m,
and W (which is pseudo-polynomial due to the dependence on W ).
Proof. We fill out a table Q[w, (cid:96)], for w = 0, . . . , W and (cid:96) = 0, . . . , k,
in
which Q[w, (cid:96)] is the highest value obtainable with a size-(cid:96) multiset of items
13
(cid:40)
of aggregate-weight at most w. Notice that Q can be filled using the following
recursion:
Q[w, (cid:96)] =
0
maxj vj + Q(cid:48)(w − wj, (cid:96) − 1)
if (cid:96) = 0,
otherwise,
(8)
where Q(cid:48)(w, (cid:96)) = Q[w, (cid:96)] if it is defined, i.e., w ≥ 0 and (cid:96) ≥ 0, and otherwise is
−∞.
Therefore Q can be filled-out using dynamic programming. Finally, the entry
Q[W, k] contains the highest value obtainable with overall weight at most W .
Therefore, if Q[W, k] > V , we have found a required multiset; otherwise such
does not exists. The resulting multiset itself can be recovered using backtracking
on the table Q. Noticed that the amount of work done is O(W km).
4.4 Solving the UCM C-LP
We return to our problem. The choice of inequalities over equalities is motivated
by our use of the LP dual in Theorem 9. However, they have the same effect as
equalities, as shown by the following lemma:
Lemma 8. In a solution to the above C-LP, Equations (5,6) will actually be
equalities.
Proof. Notice that by Equation (6):
where (12) holds by plugging (5) into (11). We therefore obtain that
which forces both above non-trivial LP inequalities to be equalities.
Theorem 9. Given a fixed value T , the UCM C-LP can be solved in polynomial
time.
14
Cjxi,C
m−1(cid:88)
Cj
xi,C
j=0
xi,C
(cid:88)
km ≤ m−1(cid:88)
m(cid:88)
m(cid:88)
m(cid:88)
(cid:88)
(cid:88)
C∈Ci(T )
j=0
i=1
i=1
=
C∈Ci(T )
= k
i=1
≤ km
C∈Ci(T )
m−1(cid:88)
m(cid:88)
(cid:88)
j=0
i=1
C∈Ci(T )
Cjxi,C = km
(9)
(10)
(11)
(12)
Proof. We need to refer to the LP dual for our C-LP in order to solve it; we
briefly repeat some LP duality concepts here, refer to [23] for complete defini-
tions and discussion.
The dual of a maximization problem is a minimization problem. In order to
define it we can treat our primal program as a maximization problem having
all coefficients 0 in its objective function.
In the dual there is a variable for
every constraint of the primal, and a constraint for every variable of the primal.
Therefore, we define a variable yi for each candidate ci and a variable zj for
each score-type αj (since the primal has a constraint for each candidate ci and
for each score-type αj). However, since our primal has an exponential number
of variables, the dual will have an exponential number of constraints. We will
show how to address this.
In short, the non-trivial constraints are then obtained by transposing the
constraint-coefficient matrix of the primal, using the primal objective function
coefficients as the right-hand side of the dual constraints, and using right-hand
side of the primal constraints as the coefficients of the dual objective function.
The process yields the following dual:
m(cid:88)
min
(cid:126)y,(cid:126)z
m−1(cid:88)
yi − k
zj
i=1
j=0
subject to:
m−1(cid:88)
j=0
Cjzj ≤ yi
yi ≥ 0
zj ≥ 0
∀i ∈ [m], C ∈ Ci(T )
∀i = 1, . . . , m
∀j = 0, . . . , m − 1
As mentioned, the dual has an exponential number of constraints. However
it is solvable; the ellipsoid method [17] is a method for solving an LP which
iteratively tries to find a point inside the feasible region described by the con-
straints. However, we do not need to provide all the constraints in advance.
Instead, the algorithm can be provided with a subroutine, called a separation
oracle, to which it calls with a proposed point, and the subroutine then either
confirms that the point is inside the feasible region or that it returns a violated
constraint [12]. The ellipsoid method algorithm performs a polynomial number
of iterations, therefore if the separation oracle runs in polynomial time as well,
the LP is solved in overall polynomial time. Notice that the polynomial num-
ber of iterations performed by the ellipsoid method implies that the number of
constraints that played a part in finding the optimum (known as active con-
straints) was polynomial as well. In other words, we could effectively discard
all the constraints but a polynomial number of them.
As discussed, a separation oracle for the dual, given a proposed solution
((cid:126)y; (cid:126)z), needs to find in polynomial time a violated constraint, if exists. It remains
to show that such a separation oracle is polynomial-time computable.
Observe that a violated constraint to this program is a pair i, C for which C ∈
15
Ci(T ) (and therefore(cid:80)m−1
j=0 Cjαj ≤ T − σi) and at the same time(cid:80)m−1
violated constraint can be seen as finding a k-multiset (since (cid:80)m−1
j=0 Cjzj >
yi. Fortunately, for a specified i, finding a configuration C that induces a
j=0 Cj = k)
given by a solution to our knapsack variant: [m]0 is the item set (over which j
ranges), the value for the item j is zj, while its weight is αj. The given value
lower bound is yi, and T −σi is the given upper bound on the weight. Effectively,
we use the possibly-tighter weight bound min{kαm−1, T −σi} instead, as kαm−1
bounds the overall weight obtainable with a size-k multiset. As now the weight
bound is polynomial in m and k, the solution to our knapsack variant becomes
polynomial.
We repeat this knapsack-solving step for each i until we find a violated
constraint, or conclude that no constraint is violated. Once we have solved the
dual using the ellipsoid method with the separation oracle, we can discard all
variables in the primal that do not correspond to violated constraints of the dual,
since the inclusion of those constraints (resp. their corresponding variables) did
not have any effect on the dual optimum (resp. the primal optimum).6 The
primal now contains only a polynomial number of variables and can be solved
directly using the ellipsoid method or any other known polynomial solvers for
LP, such as [15].
5 Algorithm for UCM
xi,C, C ∈ Ci(T ). Since(cid:80)
Solve the above mentioned configuration-LP formulation as described in Sec-
tion 4. As mentioned, while both constraints are inequalities, in any solution
they will actually be equalities. For each candidate ci, observe the variables
C∈Ci(T ) xi,C = 1, treat the xi,C's as a distribution over
the configurations for i and randomly choose one according to that distribution.
For the time being, give ci this configuration.
i=1 C i
such that H[j] =(cid:80)m
While every candidate now has a valid configuration (and her score does not
exceed T ), it is possible that the number of scores of a certain type is above or
below k. Formally, if candidate ci received a configuration C i, let the array H
j be the histogram of the scores. It is then possible that
H[j] (cid:54)= k. If we would translate the configuration given to each candidate to
the list of the scores awarded within it, and would write this list as the column
of a matrix, this matrix might not be a relaxed manipulation matrix. In order
to solve this, we need to replace some of scores in this matrix with others such
that the number of scores of each type will be k. On the other hand, we need
to make sure this process does not add much to the score of each candidate.
Let t = (i, j) be a tuple representing the event that candidate ci received a
score of αj in its configuration. Place all such tuples in a single multiset (if αj
is awarded to i more than once, repeat (i, j) as needed). Now sort this multiset
6In other words, the dual of the dual without the discarded constraints is the primal
without their corresponding variables. Another way to explain this is that this is exactly the
complementary slackness condition of the Karush-Kuhn-Tucker conditions [16, 18], a necessary
condition for obtaining the optimum.
16
Algorithm 1: Approximation algorithm.
randomly choose C i ∼ q.
1 Solve the C-LP as described in Section 4
2 foreach i do define distribution q s.t. q(C) = xi,C for all C ∈ Ci(T ) and
3 L ← (cid:104)(cid:105)
4 foreach i ∈ [m], j ∈ [m]0 do
/* L is the empty list */
j copies of (i, j) to L
/* j represents the score type
5
Append C i
αj */
6 Sort L in an ascending order by ScoreIndex(·)
7 Re-index L such that L = (cid:104)t0, . . . , tkm−1(cid:105)
8 for (cid:96) = 0, . . . , km − 1 do
t = (i, j) */
/* ScoreIndex(t) = j if
9
10
Observe tuple t(cid:96) = (i, j)
/* Assign the score α(cid:98)(cid:96)/k(cid:99) to ci, instead of the previous
αj:
j ← C i
C i
C i(cid:98)(cid:96)/k(cid:99) ← C i(cid:98)(cid:96)/k(cid:99) + 1
11
12 return the relaxed manipulation matrix corresponding to C 1, . . . , C m
j − 1
*/
j ← C i
according to the αj value in an non-decreasing manner (break ties between
candidates arbitrarily) thus creating the event-sequence t0, . . . , tkm−1, i.e., the
tuples are now indexed by their rank in this sequence. We now start the actually
fixing of the scores given; for each tuple t(cid:96) = (i, j) having rank (cid:96) in the list, we
change the score awarded to ci (as described by the tuple) from αj to α(cid:98)(cid:96)/k(cid:99). To
j − 1 followed
perform this change in the algorithm, it is enough to set C i
by setting C i(cid:98)(cid:96)/k(cid:99) ← C i(cid:98)(cid:96)/k(cid:99) + 1. This is correct as C i
j(cid:48), for any j(cid:48), represents the
number of αj(cid:48) scores awarded to ci.
Notice that now every score is repeated k times (there are only k possible
(cid:96) values mapping to the same (cid:98)(cid:96)/k(cid:99) value). Finally, the corrected configura-
tions represent the final strategy. This can be easily represented as a relaxed
configuration matrix by referring to the matrix [MS(C 1);··· ; MS(C m)], where
MS(C i) is a column constructed by taking the configuration C i, represented as
an ordered-multiset of scores (each j repeats C i
j times), in some arbitrary order.
m log m for some constant d. Let g((cid:126)α) = maxi=0,...,m−β αi+β−αi.
In words, g((cid:126)α) is the biggest different between a score in (cid:126)α and another score β
entries away from it.
Lemma 10. Let C ∈ {C 1, . . . , C m} be a configuration obtained for some can-
didate by the rounding process, and let C(cid:48) be its corrected version given by the
process described above. Then with arbitrary-chosen polynomially-small failure
The entire process is summarized as Algorithm 1.
Let β = d
√
17
.
αjC(cid:48)
probability,
m−1(cid:88)
αjCj + k · g((cid:126)α) .
j ≤ m−1(cid:88)
let the array G be the array of histogram partial sums, i.e., G[j] =(cid:80)j
In a similar manner, define Di[j] =(cid:80)j
Proof. Let H be the histogram of the original configurations C 1, . . . , C m, and
j(cid:48)=0 H[j(cid:48)].
j(cid:48) to be the partial sums array w.r.t.
each candidate ci. We will show that with high probability, G[j] ≤ (j + 1)k +
dk
j(cid:48)=0 C i
√
j=0
j=0
m ln m.
Fix a specific j. Notice that
j(cid:88)
j(cid:88)
(cid:88)
E[G[j]] =
E[H[j(cid:48)]] =
Cj(cid:48)xi,C = (j + 1)k
according to the LP constraints, and that G[j] =(cid:80)m
C∈Ci(T )
j(cid:48)=0
j(cid:48)=0
i,C
i=1 Di[j], that is, G[j] is a
random variable which is the sum of the independent random variables Di[j]
for i = 1, . . . , m. In addition, for every candidate ci, it holds that Di[j] ∈ [0, k],
as a configuration contains at most k scores. Therefore, using the generalized
Hoeffding inequality [14, Theorem 2]:
Pr [G[j] − E[G[j]] ≥ λ] ≤ exp
= exp
(cid:19)
(cid:18)
(cid:18)
− 2λ2(cid:80)m
(cid:19)
− 2λ2
mk2
i=1 k2
.
√
m ln m] ≤ 1/md(cid:48)
√
m ln m, for some arbitrary constant d(cid:48), we get that Pr[G[j] −
Setting λ = d(cid:48)k
√
E[G[j]] ≥ d(cid:48)k
, that is, the probability that we deviate from
E[G[j]] by more than O(k
m) can be made arbitrarily polynomially small.
Using the union bound, the same can be made to hold for all j = 0, . . . , m − 1
simultaneously.
Now observe a tuple t(cid:96) = (i, j(cid:48)) before being possibly corrected by the al-
gorithm. Since its rank (cid:96) in the sorted sequence is at most the number of
scores whose type is at most j(cid:48), which is by definition G[j(cid:48)], we get that
(cid:96) ≤ G[j(cid:48)] ≤ (j(cid:48) + 1)k + d(cid:48)k
m ln m, where the second inequality holds with
high probability. Therefore by the algorithm changing the score αj(cid:48) to α(cid:98)(cid:96)/k(cid:99),
the score increases by at most α(cid:98)(cid:96)/k(cid:99) − αj(cid:48) ≤ α(j(cid:48)+1)+d(cid:48)√
Now observe some candidate ci with a given configuration C i corrected to
become a configuration C(cid:48) by the algorithm. Since at worst case, all of ci's
k scores where affected as such, her overall score has increased by at most
kg((cid:126)α).
m ln m − αj(cid:48) ≤ g((cid:126)α).
√
Corollary 11. The above algorithm provides an additive kg((cid:126)α) approximation
with high probability. By repeating the randomized rounding procedure a linear
number of times, the failure probability becomes exponentially-small. The overall
running time is polynomial.
18
Proof. Let T ∗ be the optimal value for the original problem, and let TCLP be
the best bound obtainable via the above C-LP combined with the binary search
on T . Notice that TCLP ≤ T ∗, as the optimal solution is also a valid solution
for the C-LP. Now observe the highest scoring candidate in the C-LP. When
the algorithm terminates, we get that with high-probability her score is TCLP +
kg((cid:126)α) ≤ T ∗ + kg((cid:126)α). If we repeat the randomized rounding procedure a linear
number of times and pick the iteration yielding the minimum addition to TCLP,
the probability of not getting a kg((cid:126)α)-approximation becomes exponentially-
small.
As the additional score given by the algorithm to any other candidate is also
kg((cid:126)α), the bound T ∗ + kg((cid:126)α) holds for all candidates. We conclude that this is
indeed an kg((cid:126)α) additive approximation.
As discussed, solving the C-LP is done in polynomial time (by the polynomial
number of iterations of the ellipsoid method and the polynomial runtime of the
k-multiset knapsack separation oracle). The rounding is dominated by going
over a polynomial number of non-zero variables of the C-LP and is therefore
polynomial as well. It is repeated a linear number fo times in order to provide
an exponentially-small failure probability.
From here we can directly obtain Corollary 3:
√
Proof of Corollary 3. By noticing that for Borda, g((cid:126)α) = O(
m log m).
6 Linear Programming for WCM
When turning to the WCM problem, the 'natural' LP still suffers from the
deficiencies described in Section 4. We again resort to using configurations.
However, configurations will now be defined in a different manner, since now,
when each voter has an associated weight, voters are not identical anymore and
therefore our configurations need to capture the identity of the voters.
A configuration C for some candidate ci is now defined as a length-k sequence
in which C(cid:96) = j if the voter (cid:96) awarded αj to ci. For a candidate ci and a bound
T , Ci(T ) is again the set of configurations that do not cause the candidate overall
score to surpass T , which this time is formally(cid:80)k
(cid:96)=1 w(cid:96)αC(cid:96) ≤ T − σi.
The configuration LP is now formulated as follows:
(cid:88)
(cid:88)
C∈Ci(T )
i,C∈Ci(T )
xi,C ≥ 0
C(cid:96)=j
xi,C ≤ 1
∀i ∈ [m] ,
xi,C ≥ 1
∀j ∈ [m]0,∀(cid:96) ∈ [k] ,
∀i ∈ [m], C ∈ Ci(T ) .
(13)
(14)
(15)
Again, we wish that the xi,C's would serve as indicator variables indicating
whether or not ci was awarded with configuration C, (13) guarantees that every
candidate was given at most 1 configuration and (14) guarantees that every
19
score was awarded by every voter at least once. The choice of inequalities over
equalities will be explained soon.
We present another -- much more complex -- Knapsack variant, which will be
used later by the separration orcale needed for solcing the C-LP.
6.1 k-Sequence Knapsack
Let {1, . . . , m} be a set of distinct items. In the k-sequence knapsack problem
we are required to construct a length-k sequence S = s1, . . . , sk of items; we
can repeat items from the item-set, however, we are subject to some additional
constraints as will be specified immediately. The input to the problem is the
following:
• A value v(j, (cid:96)), for every j ∈ [m]0, (cid:96) ∈ [k] where v(j, (cid:96)) is the value obtained
by placing item j at location (cid:96) in the sequence.
• A cost b(j) for each item j, and a penalty p(cid:96) for each location (cid:96) ∈ [k].
Placing an item j at location (cid:96) in the sequence has a penalized-cost p(cid:96)b(j),
i.e., it depends on both the item's weight and the penalty for location (cid:96).
The resulting sequence S should abide the following constraints:
• A value lower-bound V .
• A penalized-cost upper-bound B.
• S's overall value(cid:80)k
• S's overall penalized-cost(cid:80)k
(cid:96)=1 v(s(cid:96), (cid:96)) is greater than V .
(cid:96)=1 p(cid:96)b(s(cid:96)) is at most B.
(cid:40)
Q[b(cid:48), (cid:96)] =
If such sequence S exists, we should return it; otherwise we return that no such
sequence exists.
Lemma 12. The k-sequence knapsack can be solved in time polynomial in k,m,
and B (which is pseudo-polynomial due to the dependence on B).
Proof. Similar to the proof of Lemma 7, we fill out a table Q[b(cid:48), (cid:96)], for b(cid:48) =
0, . . . , B and (cid:96) = 0, . . . , k, in which Q[b(cid:48), (cid:96)] is the highest value obtainable with
a length-(cid:96) sequence of items of penalized-cost at most b. This time Q is filled
using a different recursion:
0
maxj v(j, (cid:96)) + Q(cid:48)(b(cid:48) − p(cid:96)b(j), (cid:96) − 1)
if (cid:96) = 0,
otherwise,
(16)
where Q(cid:48)(b(cid:48), (cid:96)) = Q[b(cid:48), (cid:96)] if it is defined, i.e., b(cid:48) ≥ 0 and (cid:96) ≥ 0, and otherwise is
−∞.
Therefore Q can be filled-out using dynamic programming. Finally, the en-
try Q[B, k] contains the highest value obtainable with overall cost at most B.
Therefore, if Q[B, k] > V , we have found a required sequence; otherwise such
does not exists. The resulting sequence itself can be recovered using backtrack-
ing on the table Q. Noticed that the amount of work done is O(Bkm).
20
6.2 Solving the WCM C-LP
We return to our problem. Again, the choice of inequalities over equalities is
motivated by our use of the LP dual in Theorem 14. However, they have the
same effect as equalities, as shown by the following lemma:
Lemma 13. In a solution to the above C-LP, Equations (13,14) will actually
be equalities.
Proof. Notice that by Equation (14):
xi,C
xi,C
(cid:88)
(cid:96),j
C(cid:96)=j
1
(cid:88)
(cid:88)
C∈Ci(T )
C(cid:96)=j
j,(cid:96)
i=1
km ≤(cid:88)
m(cid:88)
m(cid:88)
m(cid:88)
m(cid:88)
(cid:88)
(cid:88)
(cid:88)
C∈Ci(T )
C∈Ci(T )
i=1
i=1
= k
=
=
(cid:96),j
C(cid:96)=j
xi,C
xi,C
C∈Ci(T )
i=1
≤ km
(cid:88)
m(cid:88)
(cid:88)
j,(cid:96)
i=1
C∈Ci(T )
xi,C = km
(17)
(18)
(19)
(20)
(21)
where (21) holds by plugging (13) into (20). We therefore obtain that
which forces both above non-trivial LP inequalities to be equalities.
Theorem 14. Given a fixed value T , the WCM C-LP can be solved in polyno-
mial time.
Proof. We again refer to the LP-dual, which this time is:
m(cid:88)
yi −(cid:88)
i=1
j,(cid:96)
min
(cid:126)y,(cid:126)z
subject to:(cid:88)
zj,(cid:96) ≤ yi
j,(cid:96)
C(cid:96)=j
yi ≥ 0
zj,(cid:96) ≥ 0
zj,(cid:96)
21
∀i ∈ [m], C ∈ Ci(T )
∀i = 1, . . . , m
∀(j, (cid:96)) ∈ [m]0 × [k]
zC(cid:96),(cid:96) ≤ yi
k(cid:88)
(cid:96)=1 w(cid:96)αC(cid:96) ≤ T − σi) and at the same time(cid:80)k
(cid:96)=1
∀i ∈ [m], C ∈ Ci(T ) .
Furthermore, the above single non-trivial constraint can be more conveniently
re-written as
therefore(cid:80)k
We again turn to the ellipsoid method with a separation oracle; this time, a
violated constraint to this program is a pair i, C for which C ∈ Ci(T ) (and
(cid:96)=1 zC(cid:96),(cid:96) > yi. For a
specified i, finding a configuration C that induces a violated constraint can be
seen as finding a k-sequence given by a solution to our knapsack variant: [m]0
is the item set (over which j ranges), the value for placing item j at location
(cid:96) is zj,(cid:96), item j's cost is αj, and the penalty for location (cid:96) is w(cid:96). The given
value lower bound is yi, and T − σi is the given upper bound on the penalized
cost. Effectively, we use the possibly-tighter weight bound min{W αm−1, T −σi}
instead, as W αm−1 bounds the overall cost obtainable with a length-k sequence.
As now the weight bound is polynomial in αm−1 and W , the solution to our
knapsack variant becomes polynomial.
We repeat this knapsack-solving step for each i until we find a violated
constraint, or conclude that no constraint is violated. Once we have solved the
dual using the ellipsoid method with the separation oracle, we continue in a
similar fashion to the proof of Theorem 9.
7 Algorithm for WCM
xi,C's as a distribution over the configurations for i since (cid:80)
Solve the above mentioned configuration-LP formulation as described in Sec-
tion 6. As both constraints will be equalities, for each candidate ci, we treat the
C∈Ci(T ) xi,C = 1,
and randomly choose one according to that distribution. For the time being,
give i this configuration.
As for UCM, every candidate now has a valid configuration but constraints
may still be violated; it is possible that the number of scores of a certain type
αj given by a specific voter (cid:96) is not exactly 1. Formally, fix a specific voter (cid:96);
we let the array H such that H[j] = {i C i
(cid:96) = j} be histogram of the scores
with respect to (cid:96). It is then possible that H[j] (cid:54)= 1. Our goal, as before, is to
fix this without introducing too much of an addition to the candidates' overall
scores. However, there is now some added complexity due to the necessity to
preserve the identity of the voter when fixing a specific score given.
Let t = (i, (cid:96), j) be a tuple representing the event that candidate ci received
a score of αj from voter (cid:96) in its configuration. Fix a manipulator (cid:96), and place
all tuples having (cid:96) as their respective voter in a set A, that is A = {(i, (cid:96), C i
(cid:96))
i = 1, . . . , m}. Sort A according to each tuple (i, (cid:96), C i
(cid:96), and let
L = (cid:104)t1, . . . , tm(cid:105) be the resulting list. Notice that any tuple tj = (i, (cid:96), j(cid:48)) in L
(cid:96) ← j.
represents the event that currently C i
In words, C i
(cid:96) gets the rank of its respective tuple in the sorted list L. In effect,
the score awarded to ci by (cid:96) changes from αj(cid:48) to αj.
(cid:96) = j(cid:48). Now change C i
(cid:96))'s score-index C i
(cid:96) such that C i
22
Algorithm 2: WCM Approximation algorithm.
1 Solve the C-LP as described in Section 6
2 foreach i do define distribution q s.t. q(C) = xi,C for all C ∈ Ci(T ) and
3 for (cid:96) ← 1 to k do
randomly choose C i ∼ q.
/* tuple (i, (cid:96), j) represnts the
(cid:96)) i = 1, . . . , m}
Let A = {(i, (cid:96), C i
event the (cid:96) awarded αj to ci */
Sort A in an ascending order by ScoreIndex(·) and let
L = (cid:104)t0, . . . , tm−1(cid:105) be the resulting list.
t = (i, (cid:96), j) */
for j ← 0 to m − 1 do
Observe tuple tj = (i, (cid:96), j(cid:48)).
(cid:96) ← j
C i
the previous αj(cid:48) */
8
/* this assigns the score αj to ci, instead of
9 return the resulting manipulation matrix [C 1;··· ; C m] /* place C i as
/* ScoreIndex(t) = j if
4
5
6
7
the i-th column vector of the resulting matrix */
We repeat the above process for each voter. Notice that now every score is
repeated k times, one by each voter. The process is summarized as Algorithm 2.
m log m for some constant d and let g((cid:126)α) = maxi=0,...,m−β αi+β−
As before, let β = d
√
αi.
Lemma 15. Let C ∈ {C 1, . . . , C m} be a configuration obtained for some can-
didate by the rounding process, and let C(cid:48) be its corrected version given by the
process described above. Then with arbitrary-chosen polynomially-small failure
probability,
m−1(cid:88)
j ≤ m−1(cid:88)
αjC(cid:48)
αjCj + W · g((cid:126)α) .
.
j=0
j=0
the array G be the array of histogram partial sums, i.e., G[j] =(cid:80)j
Proof. Fix a specific voter (cid:96). Let H be the histogram with respect to voter (cid:96)
over the original configurations C 1, . . . , C m, i.e. H[j] = {i C i
(cid:96) = j}. Let
j(cid:48)=0 H[j(cid:48)] =
(cid:96) ≤ j}. We also define Di[j] to be a Bernoulli variable which is equal
(cid:96) ≤ j, and 0 otherwise. We will show that with high probability,
{i C i
√
to 1 if C i
G[j] ≤ (j + 1) + d
m ln m.
Fix a specific j. Notice that
j(cid:88)
j(cid:48)=0
j(cid:88)
(cid:88)
j(cid:48)=0
E[G[j]] =
E[H[j(cid:48)]] =
xi,C = j + 1
according to the LP constraints, and that G[j] =(cid:80)m
i,C
C∈Ci(T )
C(cid:96)=j(cid:48)
i=1 Di[j], that is, G[j] is a
random variable which is the sum of the independent random variables Di[j] ∈
23
{0, 1} for i = 1, . . . , m. Therefore, using the 'classic' Hoeffding inequality:
Pr [G[j] − E[G[j]] ≥ λ] ≤ e− 2λ2
m .
m ln m] ≤ 1/md(cid:48)
Setting λ = d(cid:48)√
m ln m, for some arbitrary constant d(cid:48), we get that Pr[G[j] −
E[G[j]] ≥ d(cid:48)√
√
, that is, the probability that we deviate from
E[G[j]] by more than O(
m) can be made arbitrarily polynomially small. Using
the union bound, the same can be made to hold for all j = 0, . . . , m − 1 and
(cid:96) = 1, . . . , k simultaneously.
Now observe a tuple tj = (i, (cid:96), j(cid:48)) before being possibly corrected by the
algorithm. Since its rank j in the sorted sequence is at most the number of
scores given by (cid:96) whose type is at most j(cid:48), which is by definition G[j(cid:48)], we get
that j ≤ G[j(cid:48)] ≤ (j(cid:48) + 1) + d(cid:48)√
m ln m, where the second inequality holds with
high probability. Therefore by the algorithm changing the score αj(cid:48) to αj, the
score increases by at most αj − αj(cid:48) ≤ α(j(cid:48)+1)+d(cid:48)√
Now observe some candidate ci with a given configuration C i corrected to
(cid:80)k
become a configuration C(cid:48) by the algorithm. Since at worst case, all of ci's
k scores where affected as such, her overall score has increased by at most
m ln m − αj(cid:48) ≤ g((cid:126)α).
(cid:96)=1(w(cid:96)g((cid:126)α)) = W g((cid:126)α).
Corollary 16. The above algorithm provides an additive W g((cid:126)α) approximation
with high probability. By repeating the randomized rounding procedure a linear
number of times, the failure probability becomes exponentially-small. The overall
running time is polynomial.
Proof. Identical to the proof of Corollary 11, the only difference being W used
instead of k.
From here we can directly obtain Corollary 4:
√
Proof of Corollary 4. By noticing that for Borda, g((cid:126)α) = O(
m log m).
8
Implementation
We implemented our algorithm for the case for Borda-UCM and Borda-WCM
and uploaded the code to a public repository7. The main subroutine is solving
the LP duals that we have defined. However, the dependency on an LP solver
using the ellipsoid method with a separation oracle proved difficult as to the
best of our knowledge there is no library which enables solving an LP this way.
Instead we simulated this by using a general LP-solving library [1, 20] and
running the separation oracle externally as described in Algorithm 3.
7https://github.com/okeller/BordaManipulation
24
Algorithm 3: Simulating the ellipsoid method with a separation oracle.
Input: A linear program P = (f, S) with an objective function f and a
separation oracle S (instead of an explicit list of constraits)
/* R will be a list of effective constraints */
/* P (cid:48) is P without any constraints */
1 Let R ← ∅
2 Let P (cid:48) ← (f, R)
3 (cid:126)x ← LP-solve(P (cid:48))
4 while S((cid:126)x) returns a violated constraint r do
5
R ← R ∪ {r}
P (cid:48) ← (f, R)
(cid:126)x ← LP-solve(P (cid:48))
6
7
8 return (cid:126)x
8.1 Practical Heuristics
We used the following heuristics in our implementation:
• We obtained a running-time speedup was modifying the separation oracle
to return a set of violated constraints, one for each i (if such exists),
instead of a single violated constraint, and adding all of them to the list
of constraints.
• When sorting by the score index after the rounding part, we broke ties
between tuples having the same score in favor of the candidate with the
higher current score. This way, candidates with lower score are more likely
to have a score index awarded to them increased by the fixing procedure.
• Our practical running time is dominated by solving the C-LP. Therefore,
we can perform the randomized part of rounding process several times,
and pick the best one, while incurring only a negligible overhead to the
runtime.
9 Experiments
9.1 Experimental Setting
To evaluate our algorithms for both UCM and WCM in a well-studied setting,
we have run experiments for Borda-UCM and Borda-WCM on sets of values
for n, k and m (choice of values will be explained shortly). For the unweighted
case, our algorithm was compared to Average Fit that was shown empirically
to outperform reverse [8]. For the weighted case, as we are not aware of a
generalization of Average Fit to a weighted setting, we compared against
reverse.
We have chosen n = 2k, as having one manipulator for every two original
voters represents a sweet-spot where manipulators have enough power to change
outcomes, but not too much; for each n, k, m combination, we have run 20
25
experiments in which we have drawn the non-manipulator votes from a uniform
m or smaller, as our algorithms are suited
for low k values: as mentioned, our algorithm for Borda-UCM is theoretically
distribution. We have chosen k ≈ √
competitive when k = o((cid:112)m/ log m), and we have wished to verify this also in
an empirical setting. Lower k values are also the cases which are more difficult
to Average Fit heuristic of [8].
9.2 Borda-UCM
We are comparing our results to those obtained by Average Fit, and to the
fractional solution, TCLP. The results are summarized in Figure 1. As can be
seen, our algorithm performs well in practice. In the many of the cases, both
the C-LP and Average Fit were known to be identical to the optimal solution,
as depicted in Figure 1a: these are the instances in which the C-LP solution
did not increase when performing the LP rounding. Therefore, in those cases
we know what is the real optimal solution, as TCLP (resp. our final algorithm
result) provides a lower (resp. an upper) bound for it, and both are equal. Thus
our formulation in many times provides a very efficient method for verifying
whether our algorithm (or any other algorithm) finds an optimal solution.
In cases where the C-LP solution increased when performing the LP round-
ing, C-LP tends to be better than Average Fit on small k values, and worse
on larger ones. This is depicted in Figures 1b and 1c. For instance, consider
our example from before, this time with p: k = 2, m = 5 and (σ0, . . . , σ5) =
(0, 5, 6, 6, 6, 7). Obviously both methods will award p with 5 + 5 = 10. However
in ours the top score of a candidate who is not p will be 10, and in Average
Fit it is 12. Therefore, the choice of algorithm determines if p wins or not by
a difference of 2.
9.3 Borda-WCM
For the weighted case, as Average Fit has no weighted generalization, our
competitor is the weighted variant of reverse. Here the results were much
more conclusive: we experimented with the same type of values, choosing the
manipulator weights randomly from {1, 2}. The reasoning behind this was to
explore the weighted case on one hand, but on the other hand not to give any one
manipulator excessive power over others by giving him a large weight. As the
similar tables show (Figure 2), our algorithm beats reverse on any instance.
10 Conclusions
We have presented additive approximations to the score of the highest-scoring
non-preferred candidate in general scoring rules. It enables us to find a winning
strategy for p in cases where other methods would not necessarily have found
one. Our method employs new techniques based on configuration linear pro-
grams. There are several interesting directions for future work, which vary from
26
(a) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where our algorithm provided results equal to the fractional solution
TCLP (and therefore, to the optimum T ∗).
(b) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where the C-LP method provided strictly better results compared to
Average Fit.
(c) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where the Average Fit provided strictly better results compared to
the C-LP method.
Figure 1: Experimental results for Borda-UCM.
27
(a) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where our algorithm provided results equal to the fractional solution
TCLP (and therefore, to the optimum T ∗).
(b) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where the C-LP method provided strictly better results compared to
reverse.
(c) Number of cases for various value of m and k (out of 20 experiments carried out
for each m, k), where reverse provided strictly better results compared to the C-LP
method.
Figure 2: Experimental results for Borda-WCM.
28
the concrete to the general:
1. Understand in which instances different manipulation strategies outper-
form others.
2. Find algorithms that can guarantee victory even if the margin is smaller,
or prove NP-hardness even if there is a solution with margin O(
3. Apply a C-LP to other voting methods.
√
m).
4. Apply a C-LP to other problems in computational social choice, such as
fair division of multiple indivisible goods.
Acknowledgments
We thank Sarit Kraus and Ariel Procaccia for insightful discussions.
This work was supported by the Israel Science Foundation, under Grant No.
1488/14 and Grant No. 1394/16.
References
[1] M. S. Andersen, J. Dahl, and L. Vandenberghe. CVXOPT: A Python
package for convex optimization, version 1.1.9. cvxopt.org, 2016.
[2] N. Bansal and M. Sviridenko. The Santa Claus problem. In Proceedings of
the 38th Annual ACM Symposium on Theory of Computing, pages 31 -- 40,
2006.
[3] J. J. Bartholdi III, C. A. Tovey, and M. A. Trick. The computational
difficulty of manipulating an election. Social Choice and Welfare, 6(3):227 --
241, 1989.
[4] N. Betzler, R. Niedermeier, and G. J. Woeginger. Unweighted coalitional
manipulation under the Borda rule is NP-hard. In IJCAI 2011, Proceedings
of the 22nd International Joint Conference on Artificial Intelligence, pages
55 -- 60, 2011.
[5] E. Brelsford, P. Faliszewski, E. Hemaspaandra, H. Schnoor, and I. Schnoor.
Approximability of manipulating elections. In Proceedings of the Twenty-
Third AAAI Conference on Artificial Intelligence, pages 44 -- 49, 2008.
[6] V. Conitzer, T. Sandholm, and J. Lang. When are elections with few
candidates hard to manipulate? J. ACM, 54(3):14, 2007.
[7] J. Davies, G. Katsirelos, N. Narodytska, and T. Walsh. Complexity of and
In Proceedings of the Twenty-Fifth
algorithms for Borda manipulation.
AAAI Conference on Artificial Intelligence, AAAI, 2011.
29
[8] J. Davies, G. Katsirelos, N. Narodytska, T. Walsh, and L. Xia. Complexity
of and algorithms for the manipulation of Borda, Nanson's and Baldwin's
voting rules. Artificial Intelligence, 217:20 -- 42, 2014.
[9] E. Ephrati and J. S. Rosenschein. Multi-agent planning as a dynamic
search for social consensus. In Proceedings of the 13th International Joint
Conference on Artificial Intelligence, pages 423 -- 431, 1993.
[10] P. Faliszewski and A. D. Procaccia. AI's war on manipulation: Are we
winning? AI Magazine, 31(4):53 -- 64, 2010.
[11] A. Gibbard. Manipulation of voting schemes: a general result. Economet-
rica: journal of the Econometric Society, pages 587 -- 601, 1973.
[12] M. Grotschel, L. Lov´asz, and A. Schrijver. The ellipsoid method and its
consequences in combinatorial optimization. Combinatorica, 1(2):169 -- 197,
1981.
[13] E. Hemaspaandra and L. A. Hemaspaandra. Dichotomy for voting systems.
J. Comput. Syst. Sci., 73(1):73 -- 83, 2007.
[14] W. Hoeffding. Probability inequalities for sums of bounded random vari-
ables. Journal of the American statistical association, 58(301):13 -- 30, 1963.
[15] N. Karmarkar. A new polynomial-time algorithm for linear programming.
Combinatorica, 4(4):373 -- 396, 1984.
[16] W. Karush. Minima of functions of several variables with inequalities as
side conditions. Master's thesis, University of Chicago, 1939.
[17] L. G. Khachiyan. Polynomial algorithms in linear programming. USSR
Computational Mathematics and Mathematical Physics, 20(1):53 -- 72, 1980.
[18] H. Kuhn and A. Tucker. Nonlinear programming. In Proceedings of 2nd
Berkeley Symposium. Berkeley: University of California Press, pages 481 --
492, 1951.
[19] A. P. Lin. Solving hard problems in election systems. PhD thesis, Rochester
Institute of Technology, 2012.
[20] A. Makhorin. GLPK: GNU linear programming kit, version 4.61.
www.gnu.org/software/glpk/glpk.html, 2017.
[21] A. D. Procaccia and J. S. Rosenschein. Junta distributions and the average-
case complexity of manipulating elections. J. Artif. Intell. Res. (JAIR),
28:157 -- 181, 2007.
[22] M. A. Satterthwaite. Strategy-proofness and Arrow's conditions: Existence
and correspondence theorems for voting procedures and social welfare func-
tions. Journal of economic theory, 10(2):187 -- 217, 1975.
30
[23] A. Schrijver. Theory of linear and integer programming. John Wiley &
Sons, 1998.
[24] O. Svensson. Santa Claus schedules jobs on unrelated machines. SIAM J.
Comput., 41(5):1318 -- 1341, 2012.
[25] L. Xia, M. Zuckerman, A. D. Procaccia, V. Conitzer, and J. S. Rosenschein.
Complexity of unweighted coalitional manipulation under some common
voting rules. In IJCAI 2009, Proceedings of the 21st International Joint
Conference on Artificial Intelligence, pages 348 -- 353, 2009.
[26] M. Zuckerman, A. D. Procaccia, and J. S. Rosenschein. Algorithms for the
coalitional manipulation problem. Artificial Intelligence, 173(2):392 -- 412,
2009.
31
|
1702.04307 | 2 | 1702 | 2017-10-13T04:38:45 | Approximating the Held-Karp Bound for Metric TSP in Nearly Linear Time | [
"cs.DS"
] | We give a nearly linear time randomized approximation scheme for the Held-Karp bound [Held and Karp, 1970] for metric TSP. Formally, given an undirected edge-weighted graph $G$ on $m$ edges and $\epsilon > 0$, the algorithm outputs in $O(m \log^4n /\epsilon^2)$ time, with high probability, a $(1+\epsilon)$-approximation to the Held-Karp bound on the metric TSP instance induced by the shortest path metric on $G$. The algorithm can also be used to output a corresponding solution to the Subtour Elimination LP. We substantially improve upon the $O(m^2 \log^2(m)/\epsilon^2)$ running time achieved previously by Garg and Khandekar. The LP solution can be used to obtain a fast randomized $\big(\frac{3}{2} + \epsilon\big)$-approximation for metric TSP which improves upon the running time of previous implementations of Christofides' algorithm. | cs.DS | cs | Approximating the Held-Karp Bound
for Metric TSP in Nearly Linear Time∗
Chandra Chekuri
Kent Quanrud
October 16, 2017
Abstract
We give a nearly linear time randomized approximation scheme for the Held-Karp bound
[Held and Karp, 1970] for Metric-TSP. Formally, given an undirected edge-weighted graph G =
(V,E) on m edges and > 0, the algorithm outputs in O(m log4 n/2) time, with high probability,
a (1 + )-approximation to the Held-Karp bound on the Metric-TSP instance induced by the
shortest path metric on G. The algorithm can also be used to output a corresponding solution
to the Subtour Elimination LP. We substantially improve upon the O(m2 log2(m)/2) running
time achieved previously by Garg and Khandekar. The LP solution can be used to obtain a
2 + (cid:1)-approximation for Metric-TSP which improves upon the running time
fast randomized(cid:0) 3
of previous implementations of Christofides' algorithm.
7
1
0
2
t
c
O
3
1
]
S
D
.
s
c
[
2
v
7
0
3
4
0
.
2
0
7
1
:
v
i
X
r
a
∗Department of Computer Science, University of Illinois, Urbana, IL 61820. {chekuri,quanrud2}@illinois.edu.
Work on this paper partially supported by NSF grant CCF-1526799.
1
Introduction
The Traveling Salesman Problem (TSP) is a central problem in discrete and combinatorial opti-
mization, and has inspired fundamental advances in optimization, mathematical programming and
theoretical computer science. Cook's recent book [2014] gives an introduction to the problem, its
history, and general appeal. See also Gutin and Punnen [2006], Applegate, Bixby, Chvatal, and
Cook [2011], and Lawler, Lenstra, Rinnooy-Kan, and Shmoys [1985] for book-length treatments of
TSP and its variants.
Formally, the input to TSP is a graph G = (V,E) equipped with positive edge costs c : E → R>0.
In this paper we focus on TSP in
The goal is to find a minimum cost Hamiltonian cycle in G.
undirected graphs. Checking whether a given graph has a Hamiltonian cycle is a classical NP-
Complete decision problem, and hence TSP is not only NP-Hard but also inapproximable. For this
theoretical reason, as well as many practical applications, a special case of TSP called Metric-TSP
is extensively studied. In Metric-TSP, G is a complete graph Kn and c obeys the triangle inequality
cuv ≤ cuw +cwv for all u, v, w ∈ V. An alternative interpretation of Metric-TSP is to find a minimum-
cost tour of an edge-weighted graph G; where a tour is a closed walk that visits all the vertices. In
other words, Metric-TSP is a relaxation of TSP in which a vertex can be visited more than once.
The graph-based view of Metric-TSP allows one to specify the metric on V implicitly and sparsely.
Unlike TSP, which is inapproximable, Metric-TSP admits a constant factor approximation. The
classical algorithm of Christofides [1976] yields a 3/2-approximation. On the other hand it is known
that Metric-TSP is APX-Hard and hence does not admit a PTAS (Lampis [2012] showed that there
is no 185
184-approximation unless P = N P ). An outstanding open problem is to improve the bound of
3/2. A well-known conjecture states that the worst-case integrality gap of the Subtour-Elimination
LP formulated by Dantzig, Fulkerson, and Johnson [1954] is 4/3 (see [Goemans, 1995]). There has
been exciting recent progress on this conjecture and several related problems; we refer the reader to
an excellent survey by Vygen [2012]. The Subtour Elimination LP for TSP is described below and
models the choice to take an edge e ∈ E with a variable ye ∈ [0, 1]. In the following, let C(U ) (resp.
C(v)) denotes the set of edges crossing the set of vertices U ⊆ V (resp. the vertex v ∈ V).
SE(G, c) = min (cid:104)c, y(cid:105)
s.t. (cid:80)
(cid:80)
e∈C(v) ye = 2
e∈C(U ) ye ≥ 2 for all ∅ (cid:40) U (cid:40) V,
for all v ∈ V,
and ye ∈ [0, 1]
for all e ∈ E.
The first set of constraints require each vertex to be incident to exactly two edges (in the integral
setting); these are referred to as degree constraints. The second set of constraints force connectivity,
hence the name "subtour elimination". The LP provides a lower bound for TSP, and in order to
apply it to an instance of Metric-TSP defined by G, one needs to apply it to the metric completion
of G.
A problem closely related to Metric-TSP is the 2-edge-connected spanning subgraph problem
(2ECSS). In 2ECSS the input is an edge-weighted graph (G = (V,E), c), and the goal is to find a
minimum cost subgraph of G that is 2-edge-connected. We focus on the simpler version where an
edge is allowed to be used more than once. A natural LP relaxation for 2ECSS is described below
on the left. We have a variable ye for each edge e ∈ E, and constraints which ensure that each cut
has at least two edges crossing it. We also describe the dual LP on the right which corresponds to
a maximum packing of cuts into the edge costs. In the following, let C ⊆ 2E denote the family of
all cuts in G. (For technical reasons, we prefer to treat cuts as sets of edges.)
1
2ECSS(G, c) = min(cid:104)c, y(cid:105)
s. t.
ye ≥ 2
(cid:88)
e∈C
and ye ≥ 0
∀C ∈ C
∀e ∈ E
(cid:88)
C(cid:51)e
2ECSSD(G, c) = max 2(cid:104)1, x(cid:105)
s. t.
xC ≤ ce
and xC ≥ 0
∀e ∈ E
∀C ∈ C
Cunningham (see [Monma, Munson, and Pulleyblank, 1990]) and Goemans and Bertsimas [1993]
observed that for any edge-weighted graph (G, c), the optimum value of the Subtour Elimination
LP for the metric completion of (G, c) coincides with the optimum value of the 2ECSS LP for (G, c).
The advantage of this connection is twofold. First, the 2ECSS relaxation is a pure covering LP, and
its dual is a pure packing LP. Second, the 2ECSS formulation works directly with the underlying
graph (G, c) instead of the metric completion.
On the importance of solving the Subtour-LP: The subtour elimination LP is extensively
studied in mathematical programming both for its application to TSP as well as the many techniques
its study has spawned. It is a canonical example in many books and courses on linear and integer
programming. The seminal paper of Dantzig, Fulkerson and Johnson proposed the cutting plane
method based on this LP as a way to solve TSP exactly. Applegate, Bixby, Chvátal, and Cook [2003]
demonstrated the power of this methodology by solving TSP on extremely large real world instances;
the resulting code named Concorde is well-known [Applegate et al., 2011]. The importance of solving
the subtour elimination LP to optimality has been recognized since the early days of computing. The
Ellipsoid method can be used to solve the LP in polynomial time since the separation oracle required
is the global mincut problem. However, it is not practical. One can also write polynomial-sized
extended formulations using flow variables, but the number of variables and constraints is cubic in n
and this too leads to an impractical algorithm. Held and Karp [1970] provided an alternative lower
bound for TSP via the notion of one-trees. They showed, via Lagrangian duality, that their lower
bound coincides with the one given by SE(G, c). The advantage of the Held-Karp bound is that it
can be computed via a simple iterative procedure relying on minimum spanning tree computations.
In practice, this iterative procedure provides good estimates for the lower bound. However, there is
no known polynomial-time implementation with guarantees on the convergence rate to the optimal
value.
In the rest of the paper we focus on Metric-TSP. For the sake of brevity, we refer to the Held-
Karp bound for the metric completion of (G, c) as simply the Held-Karp bound for (G, c). How fast
can one compute the Held-Karp bound for a given instance? Is there a strongly polynomial-time
or a combinatorial algorithm for this problem? These questions have been raised implicitly and
are also explicitly pointed out, for instance, in [Boyd and Pulleyblank, 1990] and [Goemans and
Bertsimas, 1993]. A fast algorithm has several applications ranging from approximation algorithms
to exact algorithms for TSP.
Plotkin, Shmoys, and Tardos [1995], in their influential paper on fast approximation schemes for
packing and covering LPs via Lagrangian relaxation methods, showed that a (1 + )-approximation
for the Held-Karp bound for Metric-TSP can be computed in O(cid:0)n4 log6 n/2(cid:1) randomized time.
They relied on an algorithm for computing the global minimum cut1. Subsequently, Garg and
Khandekar obtained a (1 + )-approximation in O(m2 log2 m/2) time and they relied on algorithms
for minimum-cost branchings (see [Khandekar, 2004]).
1Their scheme can in fact be implemented in randomized O(m2 log4 m/2) time using subsequent developments
in minimum cut algorithms and width reduction techniques.
2
The main result. In this paper we obtain a near-linear running time for a (1 + )-approximation,
substantially improving the best previously known running time bound.
Theorem 1.1. Let G = (V,E) be an undirected graph with E = m edges and V = n vertices,
and positive edge weights c : E → R>0. For any fixed > 0, there exists a randomized algorithm
that computes a (1 + )-approximation to the Held-Karp lower bound for the Metric-TSP instance
on (G, c) in O(m log4 n/2) time. The algorithm succeeds with high probability.
The algorithm in the preceding theorem can be modified to return a (1+)-approximate solution
to the 2ECSS LP within the same asymptotic time bound. For fixed , the running time we achieve
is asymptotically faster than the time to compute or even write down the metric completion of
(G, c). Our algorithm can be applied low-dimensional geometric point sets to obtain a running-time
that is near-linearly in the number of points.
In typical approximation algorithms that rely on mathematical programming relaxations, the
bottleneck for the running time is solving the relaxation. Surprisingly, for algorithms solving Metric-
TSP via the Held-Karp bound, the bottleneck is no longer solving the relaxation (albeit we only
find a (1 + )-approximation and do not guarantee a basic feasible solution). We mention that the
recent approaches towards the 4/3 conjecture for Metric-TSP are based on variations of the classical
Christofides heuristic (see [Vygen, 2012]). The starting point is a near-optimal feasible solution x
to the 2ECSS LP on (G, c). Using a well-known fact that a scaled version of x lies in the spanning
tree polytope of G, one generates one or more (random) spanning trees T of G. The tree T is then
augmented to a tour via a min-cost matching M on its odd degree nodes. Genova and Williamson
[2017] recently evaluated some of these Best-of-Many Christofides' algorithms and demonstrated
their effectiveness. A key step in this scheme, apart from solving the LP, is to decompose a given
point y in the spanning tree polytope of G into a convex combination of spanning trees. Our recent
work [Chekuri and Quanrud, 2017b] shows how to achieve a (1 − )-approximation for this task
in near-linear time; the algorithm implicitly stores the decomposition in near-linear space. One
remaining bottleneck to achieve an overall near-linear running time is to compute an approximate
min-cost perfect matching on the odd-degree nodes of a given spanning tree T .
In recent work
[Chekuri and Quanrud, 2017a], we have been able to overcome this bottleneck in one way. We
obtain a randomized algorithm which uses a feasible solution x to 2ECSS LP as input, and outputs
a perfect matching M on the odd-degree nodes of T whose expected cost is at most ( 1
2 + ) times
the cost of x. Combined with our algorithm in Theorem 1.1, this leads to a(cid:0) 3
2 + (cid:1)-approximation
for Metric-TSP in O(m/2 + n1.5/3) time. If the metric space is given explicitly, then the overall
run time is O(n2/2) and near-linear in the input size. Previous implementations of Christofides'
algorithm required Ω(n2.5 log 1
2 + )-approximation even when the metric space
is given explicitly.
) time to obtain a ( 3
Integrated design of the algorithm
1.1
Our algorithm is based on the multiplicative weight update framework (MWU), and like Plotkin,
Shmoys, and Tardos [1995], we approximate the pure packing LP 2ECSSD. Each iteration requires
an oracle for computing the global minimum cut in an undirected graph. A single minimum cut
computation takes randomized near-linear-time via the algorithm of [Karger, 2000], and the MWU
framework requires Ω(m/2) iterations. Suprisingly, the whole algorithm can be implemented to
run in roughly the same time as that required to compute one global mincut.
While the full algorithm is fairly involved, the high-level design is directed by some ideas devel-
oped in recent work by the authors [Chekuri and Quanrud, 2017b] that is inspired by earlier work
of Mądry [2010] and Young [2014]. We accelerate MWU-based algorithms for some implicit packing
3
problems with the careful interplay of two data structures. The first data structure maintains a
minimum cost object of interest (here the global minimum cut) by (partially) dynamic techniques,
rather than recompute the object from scratch in every iteration. The second data structure ap-
plies the multiplicative weight update in a lazy fashion that can be amortized efficiently against
the weights of the constraints. The two data structures need to be appropriately meshed to obtain
faster running times, and this meshing depends very much on the problem specifics as well as the
details of the dynamic data structures.
While we do inherit some basic ideas and techniques from this framework, the problem here is
more sophisticated than those considered in [Chekuri and Quanrud, 2017b]. The first component in
this paper is a fast dynamic data structure for maintaining a (1 + )-approximate global minimum
cut of a weighted graph whose edge weights are only increasing. We achieve an amortized poly-
logarithmic update time by a careful adaptation of the randomized near-linear-time minimum cut
algorithm of Karger [2000] that relies on approximate tree packings. This data structure is developed
with careful consideration of the MWU framework; for example, we only need to compute an
approximate tree packing O(cid:0)1/2(cid:1) times (rather than in every iteration) because of the monotonicity
of the weight updates and standard upper bounds on the total growth of the edge weights.
The second technical ingredient is a data structure for applying a multiplicative weight update
to each edge in the approximately minimum cuts selected in each iteration. The basic difficulty
here is that we cannot afford to touch each edge in the cut. While this task suggests a lazy weight
update similar to [Chekuri and Quanrud, 2017b], these techniques require a compact representation
of the minimum cuts. It appears difficult to develop in isolation a data structure that can apply
multiplicative weight updates along any (approximately) minimum cut. However, additional nice
properties of the cuts generated by the first dynamic data structure enable a clean interaction with
lazy weight updates. We develop an efficient data structure for weight updates that is fundamentally
inextricable from the data structure generating the approximately minimum cuts.
Remark 1.2. If we extract the data structure for minimum cuts from the MWU framework, then
we obtain the following.
Given an edge-weighted G graph on m edges there is a dynamic data structure for the
weighted incremental maintenance of a (1 + )-approximate global minimum cut that
updates and queries in constant time plus O(m polylog(n) log(W1/W0)/) total amortized
time, where W0 is the weight of the minimum cut of the initial graph and W1 is the weight
of the minimum cut of the final graph (after all updates). Here, updates in the weighted
incremental setting consist of an edge e and a positive increment α to its weight. The
edges of the approximate minimum cut can be reported in constant time per edge reported.
However, a data structure for dynamic maintenance of minimum cuts is not sufficient on its own, as
there are several subtleties to be handled both appropriately and efficiently. For instance, Thorup
[2007] developed a randomized fully-dynamic data structure for global mincut with a worst-case
update time of O(√n). Besides the slower update time, this data structure assumes a model and
use case that clashes with the MWU framework at two basic points. First, the data structure does
not provide access to the edges of the mincut in an implicit fashion that allows the edge weights
to be updated efficiently, and updating each of the edges of the minimum cut individually leads
to quadratic running times. Second, the randomization in the data structure assumes an oblivious
adversary, which is not suitable for our purposes where the queries of the MWU algorithm to the
data structure are adaptive.
There are many technical details that the above overview necessarily skips. We develop the
overall algorithm in a methodical and modular fashion in the rest of the paper.
4
1.2 Related Work
Our main result has connections to several important and recent directions in algorithms research.
We point out the high-level aspects and leave a more detailed discussion to a later version of the
paper. A reader more interested in the algorithm should feel free to skip this subsection and directly
go to the next section.
In one direction there has been a surge of interest in faster (approximate) algorithms for classical
problems in combinatorial optimization including flows, cuts, matchings and linear programming
to name a few. These are motivated by not only theoretical considerations and techniques but also
practical applications with very large data sizes. Many of these are based on the interplay between
discrete and continuous techniques and there have been several breakthroughs, too many to list
here. One prominent recent example of a problem that admits a near-linear-time approximation
scheme is maximum flow in undirected graphs [Peng, 2016] building on many previous tools and
techniques. Our work adds to the list of basic problems that admit a near-linear-time approximation
scheme.
In another direction there has been exciting work on approximation algorithms for TSP and
its variants including Metric-TSP, Graphic-TSP, Asymmetric-TSP, TSP-Path to name a few. Several
new algorithms, techniques, connections and problems have arisen out of this work.
Instead of
recapping the many results we refer the reader to the survey [Vygen, 2012]. Our result adds to this
literature and suggests that one can obtain not only improved approximation but also much faster
approximations than what was believed possible. As we already mentioned, some other ingredients
in the rounding of a solution such as decomposing a fractional point in a spanning tree polytope
into convex combination of spanning trees can also be sped up using some of the ideas in our prior
work.
In terms of techniques, our work falls within the broad framework of improved implementations
of iterative algorithms. Our recent work showed that MWU based algorithms, especially for implicit
packing and covering problems, have the potential to be substantially improved by taking advantage
of data structures. The key, as we described already, is to combine multiple data structures together
while exploiting the flexibility of the MWU framework. Although not a particularly novel idea
(see the work of Mądry [2010] for applications to multicommodity flow and the work of Agarwal
and Pan [2014] for geometric covering problems), our work demonstrates concrete and interesting
problems to highlight the sufficient conditions under which this is possible. Lagrangian-relaxation
based approximation schemes for solving special classes of linear programs have been studied for
several decades with a systematic investigation started in Plotkin, Shmoys, and Tardos [1995] and
Grigoriadis and Khachiyan [1994] following earlier work on multicommodity flows. Since then there
have been many ideas, refinements and applications. We refer the reader to [Koufogiannakis and
Young, 2014, Young, 2014, Chekuri, Jayram, and Vondrák, 2015] for some recent papers on MWU-
based algorithms which supply further pointers, and to the survey [Arora, Hazan, and Kale, 2012]
for the broader applicability of MWU in theoretical computer science. Accelerated gradient descent
methods have recently resulted in fast randomized near-linear-time algorithms for explicit fractional
packing and covering LPs; these algorithms improved the dependence of the running time on to
1/ from 1/2. See [Allen-Zhu and Orecchia, 2015] and [Wang, Rao, and Mahoney, 2015].
Our work here builds extensively on Karger's randomized nearly linear time mincut algorithm
[2000]. As mentioned already, we adapt his algorithm to a partially dynamic setting informed by
the MWU framework. [Thorup, 2007] also builds on Karger's tree packing ideas to develop a fully
dynamic mincut algorithm. Thorup's algorithm is rather involved and is slower than what we are
able to achieve for the partial dynamic setting; he achieves an update time of (cid:101)O(√n) while we
achieve polylogarithmic time. There are other obstacles to integrating his ideas and data structure
5
≥0 and (cid:88)
e∈E
(cid:41)
(cid:88)
C(cid:51)e
we
zC ≤ (cid:104)w, c(cid:105)
while t < 1
C∈C w(C) // compute the minimum cut
held-karp-1(G = (V,E),c,)
x ← 0C, w ← 1/c, t ← 0, η ← / ln m
(cid:40)
C ← arg min
(cid:104)1, z(cid:105) : z ∈ RC
y ← (cid:104)w, c(cid:105)
w(C) · 1C // y = arg max
(cid:88)
γ ← min
e∈C
γ
δ ←
η
x ← x + δy // x is maintained implicitly
for all e ∈ C
t ← t + δ
end while
return x
// such that δ
yC ≤
η
ce
C(cid:51)e
ce for all e ∈ E
we ← we exp(δη/ce) = we exp(γ/ce) // increase the weight of edge e
Figure 1: An exact (and slow) implementation of the MWU framework applied to packing cuts.
to the needs of the MWU framework as we already remarked. For unweighted incremental mincut,
Goranci, Henzinger, and Thorup [2016] recently developed a deterministic data structure with a
poly-logarithmic amortized update time.
Some of the data structures in this paper rely on Euler tour representations of spanning trees.
The Euler tour representation imposes a particular linear order on the vertices by which the edges of
the underlying graph can be interpreted as intervals. The Euler tour representation was introduced
by Tarjan and Vishkin [1985] and has seen applications in other dynamic data structures, such as
the work by Henzinger and King [1999] among others.
Organization. Our main result is a combination of some high-level ideas and several data struc-
tures. The implementation details take up considerable space. Instead of compressing the details
we organized the paper in a modular fashion with the understanding that the reader may skip
some details here and there. Section 2 gives a high-level overview of a first MWU-based algorithm,
highlighting some important properties of the MWU framework that arise in more sophisticated
arguments later. Section 3 summarizes the key properties of tree packings that underlie Karger's
mincut algorithm, and explains their use in our MWU algorithm via the notion of epochs. Section
4 describes an efficient subroutine, via appropriate data structures, to find all approximate mincuts
(in the partial dynamic setting) induced by the spanning trees of a tree packing. Section 5 describes
the data structure that implement the weights of the edges in a lazy fashion, building on Euler tours
of spanning trees, range trees, and some of our prior work. In Section 6, we outline a formal proof
of Theorem 1.1.
2 MWU based Algorithm and Overview
Given an edge-weighted graph (G = (V,E), c) our goal is to find a (1 − )-approximation to the
optimal value of the LP 2ECSSD. We later describe how to obtain an approximate solution to
the primal LP 2ECSS. Note that the LP is a packing LP of the form max(cid:104)1, x(cid:105) : Ax ≤ c with
6
= (cid:80)
def
guarantees that the output x will have objective value (cid:104)1, x(cid:105) ≥ OPT while satisfying(cid:80)
an exponential number of variables corresponding to the cuts, but only m non-trivial constraints
corresponding to the edges. The MWU framework can be used to obtain a (1 − )-approximation
for fractional packing. Here we pack cuts into the edge capacities c. The MWU framework reduces
packing cuts to finding global minimum cuts as follows. The framework starts with an empty
solution x = 0C and maintains edge weights w : E → R≥0, initialized to 1/c and non-decreasing
over the course of the algorithm. Each iteration, the framework finds the minimum cut C =
arg minC(cid:48)∈C w(C(cid:48)) w/r/t w, where w(C(cid:48))
e∈C(cid:48) we denotes the total weight of a cut. The
framework adds a fraction of C to the solution x, and updates the weight of every edge in the cut
in a multiplicative fashion such that the weight of an edge is exponential in its load x(e)/c(e). The
weight updates steer the algorithm away from reusing highly loaded edges. The MWU framework
C(cid:51)e xC ≤
(1+O())ce for every edge e. Scaling down x by a (1+O())-factor gives a (1−O())-approximation
satisfying all packing constraints.
An actual implementation of the above high-level idea needs to specify the step size in each
iteration and the precise weight-update. The non-uniform increments idea of Garg and Könemann
[2007] gives rise to a width-independent bound on the number of iterations, namely O(m log m/2)
in our setting. Our algorithms follow the specific notation and scheme of Chekuri et al. [2015],
which tracks the algorithm's progress by a "time" variable t from 0 to 1 increasing in non-uniform
steps. A step of size δ in an iteration corresponds to adding δβ1C to the current solution x where
1C is the characteristic vector of the mincut C found in the iteration, and β = (cid:104)w, c(cid:105)/w(C) greedily
takes as much of the cut as can fit in the budget (cid:104)w, c(cid:105). The process is controlled by a parameter η,
which when set to O(ln m/) guarantees both the width-independent running time bound as well
as the (1 − )-approximation.
Both the correctness and the number of iterations are derived by a careful analysis of the edge
weights w. We review the argument that bounds the number of iterations. At the beginning of the
algorithm, when every edge e has weight we = 1/ce, we have (cid:104)w, c(cid:105) = m. The weights monotonically
The edge weight increases are carefully calibrated so that at least one edge increases by a (1 + )-
multiplicative factor in each iteration. As the upper bound on (cid:104)w, c(cid:105) is an upper bound on wece
for any edge e, and the initial weight of an edge e is 1/ce, an edge weight we can increase by a
(1 + )-factor at most log(1+)
increase, and standard proofs show that this inner product is bounded above by (cid:104)w, c(cid:105) ≤ O(cid:0)mO(1/)(cid:1).
(cid:0)mO(1/)(cid:1) = O(cid:0)log(m)/2(cid:1) times. Charging each iteration to an edge
weight increased by a (1 + )-factor, the algorithm terminates in O(cid:0)m/2(cid:1) iterations.
A direct implementation of the MWU framework, held-karp-1, is given in Figure 1. In addition
to the input (G, c, ), there is an additional parameter η > 0 that dampens the step size at each
iteration. Standard analysis shows that the appropriate choice of η is about ln(m)/. Each iteration
requires the minimum global cut w/r/t the edge weights w, which can be found in O(m) time (with
high probability) by the randomized algorithm of Karger [2000]. Calling Karger's algorithm in each
iteration gives a quadratic running time of O(cid:0)m2/2(cid:1).
Approximation in the framework. The MWU framework is robust to approximation in several
ways, especially in the setting of packing and covering where non-negativity is quite helpful. For
example, it suffices to find a (1 + O())-multiplicative approximation of the minimum cut, maintain
the edge weights to a (1 + O())-multiplicative approximation of their true edge weights, and so
forth, while retaining a (1 + O())-multiplicative approximation overall. We exploit this slack in
two concrete ways that we describe at a high-level below.
Approximate mincuts. First, we only look for cuts that are within a (1+O())-approximate factor of
the true minimum cut. To this end, we maintain a target cut value λ > 0, and maintain the invariant
that there are no cuts of value strictly less than λ. We then look for cuts of value ≤ (1 + O())λ
7
held-karp-2(G = (V,E),c,)
x ← 0C, w ← 1/c, t ← 0, η ← / ln m
λ ← size of minimum cut w/r/t edge weights w
while t < 1
while (a) t < 1 and
(cid:40)
(b) there is a cut C ∈ C s.t. w(c) ≤ (1 + )λ
(cid:104)1, z(cid:105) : z ∈ RC
y ← (cid:104)w, c(cid:105)
w(C) · 1C // (cid:104)1, y(cid:105) ≤ (1 + ) max
γ ← min
e∈C
γ
δ ←
η
x ← x + δy
for all e ∈ C
t ← t + δ
end while
λ ← (1 + )λ // start new epoch
we ← exp(δη/ce) · we = exp(γ/ce) · we
ce for all e ∈ E
// such that δ
(cid:88)
C(cid:51)e
yC ≤
η
≥0 and (cid:88)
e∈E
(cid:41)
(cid:88)
C(cid:51)e
we
zC ≤ (cid:104)w, c(cid:105)
ce
end while
return x
Figure 2: A second implementation of the MWU framework applied to packing cuts that divides the iterations
into epochs and seeks only approximately minimum cuts at each iteration.
until we are sure that there are no more cuts of value ≤ (1 + )λ. When we have certified that
there are no cuts of value ≤ (1 + )λ, we increase λ to (1 + )λ, and repeat. Each stretch of
iterations with the same target value λ is called an epoch. Epochs were used by [Fleischer, 2000]
for approximating fractional multicommodity flow. We show that all the approximately minimum
cuts in a single epoch can be processed in O(m) total time (plus some amortized work bounded by
other techniques). If κ is the weight of the initial minimum cut (when w = 1/c), then the minimum
cut size increases monotonically from κ to O(cid:0)m1/κ(cid:1). If the first target cut value is λ = κ, then
have Ω(m) edges, and updating m edge weights in each of O(cid:0)m/2(cid:1) iterations leads to a O(cid:0)m2/2(cid:1)
there are only O(ln m/2) epochs over the course of the entire algorithm.
Lazy weight updates. Even if we can concisely identify the approximate minimum cuts quickly,
there is still the matter of updating the weights of all the edges in the cut quickly. A cut can
running time. Instead of visiting each edge of the cut, we have a subroutine that simulates the
weight increase to all the edges in the cut, but only does work for each edge whose (true) weight
increases by a full (1 + )-multiplicative power. That is, for each edge e, we do work proportional
(up to log factors) to the number of times we increases to the next integer power of (1 + ). By
the previous discussion on weights, the number of such large updates is at most O(log m/2) per
edge. This amortized efficiency comes at the expense of approximating the (true) weight of each
edge to a (1 + O())-multiplicative factor. Happily, an approximate minimum cut on approximate
edge weights is still an approximate minimum cut to the true weights, so these approximations are
tolerable. In [Chekuri and Quanrud, 2017b] we isolated a specific lazy-weight scheme from Young
[2014] into a data structure with a clean interface that we build upon here.
Obtaining a primal solution. Our MWU algorithm computes a (1 − )-approximate solution to
8
2ECSSD. Recall that the algorithm maintains a weight w(e) for each edge e. Standard arguments
show that one can recover from the evolving weights a (1+)-approximate solution to the LP 2ECSS
(which is the dual of 2ECCSD). A sketch is provided in the appendix.
3 Tree packings and epochs
Karger [2000] gave a randomized algorithm that finds the global minimum cut in a weighted and
undirected graph with high probability in O(m) time. Treating Karger's algorithm as a black box
in every iteration leads to a quadratic running time of O(cid:0)m2/2(cid:1), so we open it up and adjust the
techniques to our setting.
Tree packings. Karger [2000] departs from previous algorithms for minimum cut with an approach
based on packing spanning trees. Let G = (V,E) be an undirected graph with positive edge weights
w : E → R>0, and let T ⊆ 2E denote the family of spanning trees in G. A tree packing is a
total weight of trees containing e is at most ce (i.e.,(cid:80)
nonnegatively weighted collection of spanning trees, p : T → R≥0, such that for any edge e, the
T(cid:51)e p(T ) ≤ ce). Classical work of Tutte [1961]
and Nash-Williams [1961] gives an exact characterization for the value of a maximum tree packing
in a graph, which as an easy corollary implies it is at least half of the value of a mincut. If C ∈ C
is a minimum cut and p is a maximum packing, then a tree T ∈ T selected randomly in proportion
to its weight in the packing will share ≤ 2 edges with C in expectation. By Markov's inequality, T
has strictly less than 3 edges in C with constant probability.
For a fixed spanning tree T , a one-cut in G induced by an edge e ∈ T is the cut C(X) where
X is the vertex set of one of the components of T − e. A two-cut induced by two edges e1, e2 ∈ T
is the following. Let X, Y, Z be the vertex sets of the three components of T − {e1, e2} where X is
only incident to e1 and Z is only incident to e2 and Y is incident to both e1, e2. Then the two-cut
induced by e1, e2 is C(Y ) = C(X ∪ Z). Thus, if T is a spanning tree sampled from a maximum tree
packing, and C is a minimum cut, then C is either a one-cut or a two-cut with constant probability.
The probabilistic argument extends immediately to approximations. Let ζ ∈ (0, 1) be sufficiently
small. If C is an approximately minimum cut with w(C) ≤ (1 + ζ)κ, and p is a tree packing of total
weight ≥ (1−ζ)κ/2, then a random tree T sampled from p has 2+O(ζ) edges from C in expectation,
and strictly less than three edges in T with constant probability. Sampling O(log(1/δ)) trees from a
tree packing amplifies the probability of a tree containing ≤ 2 edges of C from a constant to ≥ 1− δ.
For constant ζ > 0, a (1 − ζ)-approximate tree-packing can be computed in O(m) time, either by
applying O(κm)-time tree packing algorithms [Gabow, 1995, Plotkin et al., 1995] to a randomly
sparsified graph [Karger, 1998], or directly and deterministically in O(m) time by a recent algorithm
of Chekuri and Quanrud [2017b].
Another consequence of Karger [2000] is that for ζ < 1/2, the number of (1 + ζ)-approximate
minimum cuts is at most n2. By the union bound, if we select O(log n) trees at random from
an approximately maximum tree packing, then with high probability every (1 + ζ)-approximate
minimum cut is induced by one or two edges in one of the selected trees. In summary, we have the
following.
Theorem 3.1 (Karger, 2000). Let G = (V,E) be an undirected graph with edge capacities c : E →
spanning trees T1, T2, . . . , Th such that with probability ≥ 1− δ, every (1 + )-approximate minimum
cut C has C ∩ Ti ≤ 2 for some tree Ti.
R>0, let δ ∈ (0, 1), and let ∈ (0, 1/2). One can generate, in O(cid:0)m + n log3 n(cid:1) time, h = O(log(n/δ))
Karger [2000] finds the minimum cut by checking, for each tree Ti, the minimum cut in G
obtained by removing one or two edges from Ti in O(m) time. (The details of this subroutine are
9
reviewed in the following section.) Since there are O(log n) trees, this amounts to a near-linear-time
algorithm. Whereas Karger's algorithm finds one minimum cut, we need to find many minimum
cuts. Moreover, each time we find one minimum cut, the edge weights on that cut increase and so
the underlying graph changes. We are faced with the challenge of outputting O(cid:0)m/2(cid:1) cuts from a
dynamically changing graph that in particular adapts to each cut output, all in roughly the same
amount of time as a single execution of Karger's algorithm.
Epochs. We divide the problem into O(log m/2) epochs. Recall that at the start of an epoch
we have a target value λ > 0 and are guaranteed that there are no cuts with value strictly less
than λ. Our goal is to repeatedly output cuts with value ≤ (1 + O())λ or certify that there are
no cuts of value strictly less than (1 + )λ, in O(m) time. (The reason we output a cut with value
≤ (1 + O())λ is because the edge weights are maintained only approximately.) Each epoch is
(conceptually) structured as follows.
• At the start of the epoch we invoke Theorem 3.1 to find h = O(log(n/)) trees T1, . . . , Th such
that with probability 1 − poly(/n), every (1 + )-approximate mincut of G with respect to
the weights at the start of the epoch is a one-cut or two-cut of some tree Ti.
• Let C1, C2, . . . , Cr be an arbitrary enumeration of one and two-cuts induced by the trees.
• For each such cut Cj in the order, if the current weight of Cj is ≤ (1 + O())λ output it to
the MWU algorithm and update weights per the MWU framework (weights only increase).
Reuse Cj as long as its current weight is ≤ (1 + O())λ.
Since weights of cuts only increase, and we consider every one and two-cut of the trees we obtain
the following lemma.
Lemma 3.2. Let w : E → R>0 be the weights at the start of the epoch and suppose mincut of G with
respect to w is ≥ λ; and suppose the trees T1, . . . , Th have the property that every (1+)-approximate
mincut is induced by a one or two-edge cut of one of the trees. Let w(cid:48) : E → R>0 be the weights at
the end of an epoch. Then, the mincut with respect to weights w(cid:48) is ≥ (1 + )λ.
Within an epoch, we must consider all cuts induced by 1 or 2 edges in O(1) spanning trees. We
have complete flexibility in the order we consider them. For an efficient implementation we process
the trees one at a time. For each tree T , we need to repeatedly output cuts with value ≤ (1 + O())λ
or certify that there are no cuts of value strictly less than (1 + )λ induced by one or two edges
of T . Although the underlying graph changes, we do not have to reconsider the same tree again
because the cut values are non-decreasing. If each tree can be processed in O(m) time, then an
entire epoch can be processed in O(m) time and the entire algorithm will run in O(cid:0)m/2(cid:1) time. In
Figure 3, we revise the algorithm to invoke Theorem 3.1 once per epoch and use the trees to look
for approximately minimum cuts, while abstracting out the search for approximately minimum cuts
in each tree.
4 Cuts induced by a tree
Let T ∈ T be a rooted spanning tree of G, and let λ > 0 be a fixed value such that the minimum
cut value w/r/t w is ≥ λ. We want to list cuts C ∈ C induced by 1 or 2 edges in T with weight
w(C) ≤ (1 + )λ, in any order, but with a twist: each time we select one cut C, we increment the
weight of every edge in C, and these increments must be reflected in all subsequent cuts. We are
10
held-karp-3(G = (V,E),c,)
x ← 0C, w ← 1/c, t ← 0, η ← / ln m
λ ← size of minimum cut w/r/t edge weights w
while t < 1
p ← O(log(n/)) spanning trees capturing every (1 + )-apx min cut
for each tree T ∈ p
let CT ⊆ C be the cuts induced by removing ≤ 2 edges from T
while (a) t < 1 and
w/ high probability // by Theorem 3.1
ce
(b) there is a cut C ∈ CT s.t. w(C) < (1 + )λ
y ← (cid:104)w, c(cid:105)
w(C) · C
γ ← min
e∈C
γ
δ ←
η
x ← x + δy
for all e ∈ C
t ← t + δ
end while
we ← exp(γ/ce) · we
end for
λ ← (1 + )λ // start new epoch
end while
return x
Figure 3: A third implementation of the MWU framework applied to packing cuts that samples a tree
packing once at the beginning of each epoch and searches for approximately minimum cuts induced by one
or two edges in the sampled trees.
allowed to output cuts of value ≤ (1 + O())λ. When we finish processing T , we must be confident
that there are no 1-cuts or 2-cuts C ∈ CT induced by T with weight w(C) ≤ (1 + )λ.
The algorithmic challenge here is twofold. First, we need to find a good 1-cut or 2-cut in T
quickly. Second, once a good cut is found, we have to increment the weights of all edges in the cut.
Either operation, if implemented in time proportional to the number of edges in the cut, can take
O(m) time in the worst case. To achieve a nearly linear running time, both operations must be
implemented in polylogarithmic amortized time.
In this section we focus on finding the cuts, and assume that we can update the edge weights along
any 1 or 2-cut efficiently via the lazy-inc-cuts data structure. Since the weights are dynamically
changing, we need to describe how the algorithm finding small 1-cuts and 2-cut in T interacts with
the lazy-inc-cuts data structure.
Interacting with the lazy weight update data structure: The interface for lazy-inc-cuts
is given in Figure 4. The primary function of the lazy-inc-cuts data structure is inc-cut, which
takes as input one or two edges in T , and simulates a weight update along the corresponding 1-cut
or 2-cut. inc-cut returns a list of weight increments (ei, δi), where each ei is an edge and δi > 0 is
a positive increment that we actually apply to the underlying graph. The cumulative edge weight
increments returned by the subroutine will always underestimate the true increase, but by at most
a (1 + O())-multiplicative factor for every edge. That is, if we let we denote the "true weight" of
11
Lazy Incremental Cuts: Interface
Description
Operations
lazy-inc-cuts Given an undirected graph G = (V,E) with capacities c : E → R>0 and edge
weights w : E → R>0, and a rooted spanning tree T , initializes the lazy-inc-
cuts data structure.
Given a 1-cut or 2-cut C ∈ CT induced by T , simulates a weight incre-
ment along the cut per the MWU framework. Returns a list of tuples ∆ =
{(e1, δ1), (e2, δ2), . . .}, where each ei ∈ E is an edge and each δi > 0 is a positive
real value that signals that the weight of the edge ei has increased by an additive
factor of δi. Each increment (ei, δi) satisfies δi ≥ Ω(cid:0)we/ log2 n(cid:1).
inc-cut
Returns a list of tuples ∆ = {(e1, δ1), . . . , (ek, δk)} of all residual weight incre-
ments not accounted for in previous calls to inc-cut and flush.
flush()
Figure 4: The interface for the data structure lazy-inc-cuts, which efficiently simulates weight increments
along cuts induced by 1 or 2 edges in a fixed and rooted spanning tree. Implementation details are provided
later in Section 5.
for e returned by inc-cut, then the lazy-inc-cuts data structure maintains the invariant that
A second routine, flush(), returns a list of increments that captures all residual weight in-
crements not accounted for in previous calls to inc-cut and flush. The increments returned by
≤ (1 + )λ w/r/t the approximated edge weights retains weight ≤ (1 + O())λ in the true underlying
graph. Each increment (e, δ) returned by inc-cut reflects a (1+Ω(/ log2 n))-multiplicative increase
an edge e implied by a sequence of weight increments along cuts, and (cid:101)we the sum of δ increments
(cid:101)we ≤ we ≤ (1 + ρ)(cid:101)we, where ρ is a fixed constant that we can adjust. In particular, a cut of weight
to we. As the weight of an edge is bounded above by O(cid:0)m1/(cid:1), the subroutine returns an increment
for a fixed edge e at most O(cid:0)log3 n/2(cid:1) times total over the entire algorithm.
flush() may not be very large, but restores (cid:101)we = we for all e. Formally, we will refer to the
following.
Lemma 4.1. Let G = (V,E) be an undirected graph with positive edge capacities c : E → R>0 and
positive edge weights w : E → R>0, and T a rooted spanning tree in G. Consider an instance of
we be the "true" weight of edge e if the MWU framework were followed exactly, and let (cid:101)we be the
lazy-inc-cuts(G,c,w,T ) over a sequence of calls to inc-cut and flush. For each edge e, let
(a) After each call to inc-cut, we have (cid:101)we ≤ we ≤ (1 + O())(cid:101)we for all edges e ∈ E.
approximate weight implied by the weight increments returned by calls to inc-cut and flush.
(b) After each call to flush, we have (cid:101)we = we.
(e) Each increment (e, δ) returned by inc-cut satisfies δ ≥ Ω(cid:0)(/ log2 n)we
(c) Each call to inc-cut takes O(log2 n + I), where I is the total number of weight increments
(d) Each call to flush takes O(m log2 n) time and returns at most m weight increments.
The implementation details and analysis of lazy-inc-cuts are given afterwards in Section 5.
An enhanced sketch of the algorithm so far, integrating the lazy-inc-cuts data structure but
abstracting out the search for small 1- and 2-cuts, is given in Figure 5. In this section, we assume
Lemma 4.1 and prove the following.
returned by inc-cut.
(cid:1).
12
held-karp-4(G = (V,E),c,)
x ← 0C, w ← 1E, t ← 0, η ← / ln m
λ ← size of minimum cut w/r/t edge weights w
while t < 1
w/r/t weights w w/ high probability // by Theorem 3.1
p ← O(log(n/)) spanning trees capturing all (1 + )-apx min cuts
for each tree T ∈ p
lic ← lazy-inc-cuts.init(G,c,T )
let CT ⊆ C be the cuts induced by removing ≤ 2 edges from T
while (a) t < 1 and until
s.t. w(C) < (1 + )λ
(b) we have certified there are no cuts C ∈ CT
(cid:18)
(cid:19)
· C
(cid:104)w, c(cid:105)
w(C)
ce
find C ∈ CT w/ w(e) < (1 + O())λ
y ←
γ ← min
e∈C
γ
δ ←
η
x ← x + δy
t ← t + δ
∆ ← lic.inc-cut(C)
for (e, δ) ∈ ∆
we ← we + δ
end for
end while
∆ ← lic.flush()
for (e, δ) ∈ ∆
we ← we + δ
end for
end for
end while
return x
Figure 5: A fourth implementation of the MWU framework applied to the Held-Karp relaxation, that
integrates the lazy-inc-cuts data structure into the subroutine that processes each tree.
Lemma 4.2. Let T be a rooted spanning tree and λ > 0 a target cut value. One can repeatedly
output 1-cuts and 2-cuts C ∈ CT induced by T of weight w(C) ≤ (1 + O())λ and increment the
corresponding edge weights (per the MWU framework) until certifying that there are no 1-cuts or
2-cuts C ∈ CT of value w(C) ≤ (1 + )λ in O(cid:0)m log2 n + K log2 n + I log n(cid:1) time, where K is the
number of 1-cuts and 2-cuts of value ≤ (1 + O())λ output and I is the total number of weight
increments returned by the lazy-inc-cuts data structure.
The rest of the section is devoted to the proof of the preceding lemma. We first set up some
notation. Fixing a rooted spanning tree T induces a partial order on the vertices V. For two
vertices u and v, we write u ≤ v if u is a descendant of v. Two vertices u and v are incomparable,
written u (cid:107) v, if neither descends from the other. For each vertex v ∈ V, let Tv be the subtree
13
1-cut
Incomparable 2-cut
Nested 2-cut
Figure 6: 3 types of tree-cuts
def
def
def
= w(C(S)) =(cid:80)
of T rooted at v. Let D(v)
= V(Tv) = {u : u ≤ v} be the down set of all descendants of v in
T (including v). We work on a graph with edge weights given by a vector w : E → R>0. For a
= {e ∈ E : e ∩ S = 1} be the set of edges cut by S, and let
set of vertices S ⊂ V, we let C(S)
e∈C(S) we denote the weight of the cut induced by S. For two disjoint sets of
C(S)
vertices S, T ⊆ V, we let C(S, T ) = {e = (a, b) ∈ E : a ∈ S, b ∈ T} be the edges crossing from S to
T , and let C(S, T ) = w(C(S, T )) be their sum weight.
Per Karger [2000], we divide the 1-cuts and 2-cuts into three distinct types, drawn in Figure 6.
First there are the 1-cuts of T of the form C(D(s)). Then we have 2-cuts C(D(s) ∪ D(t)) where
s and t are incomparable (i.e., s (cid:107) t). We call these incomparable 2-cuts. Lastly, we have 2-cuts
C(D(t) \ D(s)) where s is a descendant of t (i.e., s < t). We call these nested 2-cuts. For a fixed
tree, we process all 1-cuts, incomparable 2-cuts, and then nested 2-cuts, in this order. For each of
these three cases we prove a lemma similar to Lemma 4.2 but restricted to the particular type of
cut. Throwing in O(m) time to initialize the lazy-inc-cuts data structure at the beginning and
O(m) to call flush at the end then gives Lemma 4.2.
Standard data structures. Following Karger [2000], we rely on the dynamic tree data structure,
link-cut trees, of Sleator and Tarjan [1983]. Link-cut trees store real values at the nodes of a rooted
tree T and support aggregate operations along node-to-root paths in logarithmic time. We employ
two operations in particular. First, given a node v and a value α ∈ R, we can add α to the value
of every node on the v-to-root path (i.e., every node u such that u ≥ v) in O(log n) amortized
time. Second, given a node v, we can find the minimum value of any node on the v-to-root path in
O(log n) amortized time.
We also need to compute least common ancestors (lca) in T for pairs of nodes. Harel and Tarjan
[1984] showed how to preprocess T in linear time and answer lca queries in constant time.
1-cuts
4.1
Let e = (t, s) ∈ T be a tree edge with t a parent of s. Then T − e consists of two components: the
subtree Ts rooted at s, and the remaining tree T \ Ts. The cut C(D(s)) in the underlying graph G
induced by the subtree Ts is a "1-cut" of T . In this section, we want to output a sequence of 1-cuts
C ∈ CT in T with w(C) ≤ (1 + O())λ, repeating if necessary, while incrementing weights of each
listed cut on the fly, until we've determined that there are no 1-cuts C with weight w(C) ≤ (1 + )λ.
Computing and maintaining the 1-cut values. To find the minimum 1-cut induced by a down
14
sstst set D(v), Karger observes that the weight C(D(v)) can be rewritten as
(cid:88)
C(D(v)) =
(cid:88)
C(x) − 2
x∈D(v)
e∈E[D(v)]
we,
x∈D(v)
The quantities(cid:80)
where C(x) is the weighted degree of the vertex x, and E[D(v)] is the set of edges with both endpoints
in D(v).
C(x) over v ∈ V are just tree sums over the weighted degrees, and Karger
[2000] computes them once for all v in O(m) total time by a depth-first traversal. In the dynamic
setting, we can compute and maintain the tree sums in O(log n) time per edge update in a link-cut
tree. Each time we increment the value of an edge e = (u, v) ∈ E by some δ > 0, we add δ to the
value of every node on the paths to the root from u and v.
e∈E[D(v)] we over edges with both endpoints in D(v) is just the sum of
weights of all edges e = (x, y) for which the least common ancestor lca(x, y) of x and y is in D(v);
e:lca(e)∈D(v) we By Harel and Tarjan [1984], we can compute lca(e) for every
lca(e)=v we of all weights
of edges whose least common ancestor is v. The quantities w(E[(D(v))]) can then be computed as
tree-sums by depth-first search, and updated dynamically with link-cut trees, similar to before.
The preceding discussion shows that the sums C(D(v)) for each v can be maintained dynamically
i.e., w(E[D(v)]) =(cid:80)
edge e ∈ E in linear time. We then compute for each vertex v the sum(cid:80)
The sum w(E[D(v)])
= (cid:80)
def
in logarithmic time using link-cut trees. This is summarized in the lemma below.
Lemma 4.3. Given edge-weighted graph G = (V,E) and a rooted spanning tree T , there is a data
structure to maintain C(D(v)) for all v ∈ V as edge-weights are changed. The initialization cost is
O(m), updates and queries take O(log n) amortized time.
Processing all 1-cuts. The weight of all 1-cuts can be computed in O(m) time statically and
O(log n)-time per edge update. To process all 1-cuts of the tree, we consider all downsets D(v) in
any order. For each vertex v, we check if C(D(v)) ≤ (1 + )λ in O(log n) amortized time. If not,
then we move on to the next vertex. Since the edge weights are monotonically increasing, C(D(v))
will never be ≤ (1 + )λ again, so we do not revisit v. Otherwise, if C(D(v)) is good, we take the cut
in our solution and pass v to inc-cut to signal a weight increment along C(D(v)). In turn, inc-
cut returns a list of "official" edge increments {(ei, δi)}. Each update can be incorporated into the
link-cut tree in logarithmic time by Lemma 4.3. After processing the edge increments returned by
inc-cut, we continue to process D(v) until C(D(v)) is no longer a good cut. The total running time
to process all 1-cuts is O(m) plus O(cid:0)log2 n(cid:1) per cut output and O(log n) per edge weight increment
returned by inc-cut.
Lemma 4.4. Let T be a fixed and rooted spanning tree and λ > 0 a fixed target cut value. Employing
the inc-cut routine of the lazy-inc-cuts data structure to approximately increment edge weights,
one can repeatedly find 1-cuts C induced by T of value C(C) ≤ (1 + O())λ and increment the
corresponding edge weights (per the MWU framework) until certifying that there are no 1-cuts C of
value C(C) ≤ (1 + )λ in O(cid:0)m + K log2 n + I log n(cid:1) total running time, where K is the number of
1-cuts of value ≤ (1 + O()) output and I is the total number of weight increments returned by inc-
cut.
2-cuts on incomparable vertices
4.2
In this section, we consider incomparable 2-cuts of the form C(D(s) ∪ D(t)), where s (cid:107) t. We first
consider the case where s is a fixed leaf, and t ranges over all vertices incomparable to s. We then
15
Fixed leaf
Fixed path to a leaf
Induction step
Figure 7: 3 cases for incomparable 2-cuts
extend the leaf case to consider the case where s ranges over a path in the tree down to a leaf, and
t ranges over all incomparable vertices to the path. We then apply an induction step that reduces
processing the whole tree to a logarithmic number of rounds where in each round we process all the
paths to leaves. See Figure 7.
Incomparable 2-cuts with a leaf. We first fix s to be a leaf in T , and consider all incomparable
2-cuts with one side fixed to be {s}. Karger observes that the weight of the cut induced by D(t) + s,
for t (cid:107) s, can be written as
C(D(t) + s) = C(s) + C(D(t)) − 2C(D(t), s)
where C(D(t), s) is the weight of edges with one endpoint s and the other in D(t). We only need to
consider down sets D(t) for which s and D(t) are connected by some edge, since otherwise either
{s} or D(t) induce a smaller cut, and after processing all 1-cuts, neither {s} or D(t) induce small
enough cuts.
For fixed s and ranging t over all t (cid:107) s, Karger finds the weight of the best incomparable
D(t) + s cut statically as follows. As C(s) is fixed, the t's are differentiated only by the values
C(D(t))−2C(D(t), s). Karger's algorithm creates a dynamic tree over T with each vertex t initialized
with the value C(D(t)). (This initial tree only needs to be constructed once, and immutably reused
for other s's). Karger's algorithm immediately adds +∞ to the value of every vertex on the s-to-
root path to eliminate all comparable vertices from consideration. For each edge e = (s, v) with
endpoint s, Karger's algorithm subtracts 2we from every vertex on the v-to-root path, to incorporate
the second term −2C(D(t), s). Then, for each edge e = (s, v) adjacent to s, Karger's algorithm finds
the minimum value vertex on the v to root path. The minimum value vertex z over all such queries
gives the best s + D(z) cut subject to s (cid:107) z, as desired. The running time to process the leaf s,
excluding the construction of the initial tree, is bounded by the time spent subtracting edge weights
and taking minimums over the v-to-root path for each edge (s, v) incident to s. With dynamic trees,
subtracting and finding the minimum along a vertex-to-root path each take O(log n) amortized time.
Thus, up to a logarithmic factor, the time spent processing all incomparable 2-cuts with one side a
fixed leaf s is proportional to the number of edges incident to s.
In the dynamic setting, we do the following. We first compute the sum C(D(t)) − 2C(D(t), s)
for all incomparable t : t (cid:107) s, as above. We then run through the edges (s, v) incident to s in any
fixed order, checking for s + D(t) cuts with t ≥ v that have value ≤ (1 + )λ and updating edge
weights along good cuts on the fly, as follows.
16
stts1s2...s` 1s` Fix an edge e = (s, v) incident to s. We process e until we know that there is no good (D(t)+ s)-
cut for any ancestor t of v. We first find the minimum value vertex on the v-to-root path in O(log n)
amortized time. If the minimum value is > (1 + )λ − C(D(s)), then we move onto the next edge
incident to s. Otherwise, let t be an ancestor of v such that s+D(t) induces a good cut. We pass the
cut C(s + D(t)) (in terms of the roots of s and t) to inc-cut, which simulates a weight update along
C(s + D(t)) and returns a list of weight increments {(ei, δi)}. For each such edge e = (x, y) whose
weight is increased by some δ > 0, we add δ to every vertex on the x-to-root and y-to-root paths.
We then subtract 2δ from the lca(x, y)-to-root path. These additions and subtraction account for
the first term, C(D(t)). If x = s, then we also subtract 2δ from the y-to-root path for the sake of
the second term, −2C(D(t), s). After all official weight increments returned by inc-cut have been
incoporated into the dynamic tree, we continue to process the same edge e = (s, v). Once all edges
e = (s, v) incident to s have been processed, we have certified that all 2-cuts with one side fixed to
the leaf s have weight ≥ (1 + )λ. The total time is O(log n) times the number of edges incident to
Leaf-paths. The approach outlined above for processing leaves extends to processing paths. Let
a leaf-path be a maximal subtree that is a path. Let p = s1, s2, . . . , s(cid:96) be a leaf path, where s1 is a
leaf, s2 is its parent, and so forth. We want to process all 1-cuts of the form D(si) ∪ D(t), where t
is incomparable to any and all of the si's on the path.
s, plus O(cid:0)log2 n(cid:1) time for every cut output and O(log n) for every increment returned by inc-cut.
For each i, we have
C(D(si) ∪ D(t)) = C(D(si)) + C(D(t)) − 2C(D(si), D(t)),
where C(D(si), D(t)) is the weight of all edges between D(si) and D(t). The values C(D(t)) for
t (cid:107) s1, . . . , s(cid:96) and C(D(si)) for each i are easy to compute initially and maintain dynamically, per
the earlier discussion about 1-cuts in Section 4.1. For fixed i, the incomparable t are differentiated
by the value C(D(t))− 2C(D(si), D(t)). The basic idea here is that C(D(si+1), D(t)) can be written
in terms of the preceding cut in the form C(D(si+1), D(t)) = C(D(si), D(t))+ C(si+1, D(t)). That is,
after computing C(D(si), D(t)) for all incomparable t, we can keep these values and just incorporate
the weights of edges incident to si+1 to get the values C(D(si), D(t)) for all incomparable t.
We first review the static case, where we want to find the minimum such cut. Karger's algorithm
first processes the leaf s1 as a leaf, as outlined above, and takes note of the minimum s1 + D(t)
cut. It then processes s2 in the same fashion, except it continues the aggregate data built when
processing s1 rather than starting fresh. By subtracting the weights of all edges incident s2 from the
values of the various incomparable downsets, and having already subtracted the weights of edges
incident to s1, we will have subtracted the weight of all edges incident to D(s2) = {s1, s2}, and so we
have computed C(D(t))− 2C(D(s2), D(t)) for all t (cid:107) s2. In this fashion, Karger's algorithm marches
up the path doing essentially the same work for each vertex as for a fixed leaf, except continuing
the same link-cut tree from one si to the next. The total work processing the entire leaf-path is, up
to a logarithmic factor, proportional to the total number of edges incident to any node si on the
path.
The analogous adjustments are made in the dynamic setting. That is, we process s1 as a
leaf, passing good cuts to inc-cut and incorporating the returned edge increments on the fly. After
processing s1, we keep the accumulated aggregate values, and start processing s2 likewise. Marching
up the leaf-path one vertex at a time. In this manner, we are able to process 2-cuts of the form
D(si) + D(t) over each si along the leaf-path s1, . . . , s(cid:96) (subject to t (cid:107) s1, . . . , s(cid:96)) and at the end
certify there are no such cuts of weight < (1 + )λ. The total work in processing the leaf path is
O(log n) for each edge incident to s1, . . . , s(cid:96), O(cid:0)log2 n(cid:1) for each good cut output, and O(log n) for
each weight increment returned from the subroutine inc-cut.
17
All incomparable 2-cuts. Karger [2000] showed that an efficient subroutine for processing all
incomparable 2-cuts on a single leaf-path leads to an efficient algorithm for processing all incom-
parable 2-cuts in the entire tree as follows. The overall algorithm processes the 2-cuts of a tree in
phases. Each phase processes all the leaf-paths of the current tree, and then contracts the leaf-paths
into their parents and recurses on the new tree in a new phase. Each leaf in the contracted graph
had at least two children in the previous phase, so the number of nodes has halved. After at most
O(log n) phases, we have processed all the incomparable 2-cuts in the tree. Excluding the work for
selecting a cut and incrementing edge weights returned by the lazy-inc-cuts data structure, pro-
cessing a leaf-path takes time proportional to the number of edges incident to the leaf-path. Each
edge is incident to at most 2 leaf-paths. Thus, a single phase takes O(m log n) time, in addition to
O(cid:0)log2 n(cid:1) time for every cut output and O(log n) amortized time for every edge weight increment
returned by the lazy-inc-cuts data structure. In conclusion, we have shown the following.
Lemma 4.5. Let T be a fixed and rooted spanning tree and λ > 0 a fixed target cut value. Employing
the inc-cut routine of the lazy-inc-cuts data structure to approximately increment edge weights,
one can repeatedly find incomparable 2-cuts C induced by T of value C(C) ≤ (1 + O())λ and
increment the corresponding edge weights (per the MWU framework) until certifying that there are no
incomparable 2-cuts C of value C(C) ≤ (1 + )λ in O(cid:0)m log2 n + K log2 n + I log n(cid:1) total amortized
time, where K is the number of incomparable 2-cuts of value ≤ (1 + O())λ output and I is the total
number of weight increments returned by inc-cut.
4.3 Nested 2-cuts
The other type of 2-cuts is of the form D(t) \ D(s), where t is an ancestor of s. We call these cuts
nested 2-cuts. As in the case of incomparable 2-cuts, we first consider the case where s is fixed to
be a leaf, then the case where s is on a leaf-path, and then finally an induction step that processes
all leaf-paths in each pass.
Nested 2-cuts with a leaf. We take the same approach as with incomparable 2-cuts, and start
with the case where s is a fixed leaf. Karger observed that the weight of the cut induced by D(t)− s
can be written as
(1)
C(D(t) − s) = C(D(t)) + 2C(s, D(t)) − C(s),
where C(s, D(t)) is the sum weight of the edges between s and D(t). We start with a link-cut tree
over T where each node v is initialized with C(D(v)). For each edge e = (s, u) incident to s, we
find the least common ancestor t = lca(s, u), and add 2we to every node on the u-to-root path in
O(log n) time. We also subtract we from the value of every node on the s-to-root path. The first
set of updates along the t-to-root path is for the second term of (1), and the second set along the
s-to-root path is for the third term.
Karger's algorithm finds the minimum value on the s-to-root path to get the value of the mini-
mum (D(t)−s)-cut over all t > s. In the dynamic setting, we repeatedly find the minimum value on
the s-to-root path as long as this value is below the threshold (1 + )λ. Each time we find a good cut
C(D(t) − s), we send the cut to the routine inc-cut (compactly represented by s and t). For each
edge weight increment (e = (x, y), δ) returned by inc-cut we do the following. Let t = lca(x)y be
the least common ancestor. First, we add δ to every the value of every vertex on the x-to-root and
y-to-root paths and subtract 2δ from every value of every vertex on the t-to-root-path in O(log n)
time, to account for the first term C(D(t)) (see Section 4.1). If e is also incident to s, then we add
2δ to the value of every vertex on the t-to-root path (for the second term of (1)), and subtract δ
from the value of every vertex on the s-to-root path (for the third term). Thus, incorporating an
18
edge increment (e, δ) consists of a constant number of updates along node-to-root paths and takes
O(log n) amortized time.
fixed leaf s is O(log n) times the number of edges incident to s, plus O(cid:0)log2 n(cid:1) for every cut output
The total amortized running time for processing all nested 2-cuts of the form D(t) − s with a
and O(log n) for every edge weight increment returned by the inc-cut subroutine.
Nested 2-cuts along a leaf-path. We extend the leaf case to a leaf path s1, . . . , s(cid:96), where s1 is
a leaf and each si with i > 1 has exactly one child, and consider all cuts of the form D(t) \ D(si)
over all si in the leaf path and all ancestors t of si. For each i, we have
(2)
When i = 1, this is the same as equation (1) obtained for the leaf case. For i = 1, . . . , (cid:96) − 1, the
difference in (2) between consecutive vertices si and si+1 is
C(D(t) \ D(si)) = C(D(t)) + 2C(D(si), D(t)) − C(D(si)).
C(D(t) \ D(si+1)) − C(D(t) \ D(si))
= 2(cid:2)C(D(si+1), D(t)) − C(D(si), D(t))(cid:3)
= 2C(si+1, D(t)) + C(si+1) − 2C(si+1, D(si)).
(cid:2)C(cid:0)D(si+1) − C(D(si))(cid:1)(cid:3)
−
(3)
These observations lead to the following bottom-up approach. We first process s1 as though it were
a leaf, as described above. For i = 1, . . . , (cid:96) − 1, once we have certified that are no nested D(t) − si
cuts of value < (1 + )λ, we begin to process si+1, continuing the aggregate values computed when
processing si instead of starting over. For each edge e = (si+1, x) incident to si+1, we update the
values along paths up the tree per (3) as follows. For the first term, 2C(si+1, D(t)), we find the
least common ancestor t = lca(si+1, x), and add 2we to every vertex on the t-to-root path. For
the second term, C(si+1), we add we to every vertex on the si+1-to-root path. For the third term,
2C(si+1, D(Si)), if x = sj for some j < i, then we subtract 2we from every vertex on the si+1-to-
root path. After processing the edges incident to si+1, we repeatedly find the minimum value in the
si+1-to-root path so long as the minimum value is less than the threshold (1 + )λ. Each time the
minimum value is below the threshold, we update weights along the edges of the corresponding cut
via inc-cut, and propagate any increments returned by inc-cut (as before) to restore (2) before
querying for the next minimum value.
To process all the nested 2-cuts along the leaf-path, as with incomparable 2-cuts along leaf-paths,
the total amortized running time is O(log n) times the number of edges incident to s1, . . . , s(cid:96), plus
O(cid:0)log2 n(cid:1) work for each cut output and O(log n) work for each edge increment returned by inc-
cut.
All nested 2-cuts. By the same induction step as for incomparable 2-cuts, efficiently processing
leaf-paths leads to a procedure for efficiently processing the whole tree. The processing is broken
into phases. Each phase processes all the leaf-paths of the current tree, and then contracts the
leaf-paths into their parents. Each leaf in the contracted tree had at least two children previously,
so the number of nodes has at least halved. After O(log n) phases, we have processed the entire tree.
Modulo O(cid:0)log2 n(cid:1) work for each cut output and O(log n) work for each edge increment returned
by inc-cut, processing a leaf path takes O(log n) time for each edge incident to any node on
the path, and conversely each edge is incident to at most 2 leaf paths. Thus, excluding the time
spent outputting cuts and incrementing edge weights along the cuts, each phase takes O(m log n)
to complete.
Lemma 4.6. Let T be a fixed and rooted spanning tree and λ > 0 a fixed target cut value. Employing
the inc-cut routing of the lazy-inc-cuts data structure to manage edge weight increments, one
19
can repeatedly find nested 2-cuts C induced by T of value C(C) ≤ (1 + O())λ and increment the
corresponding edge weights (per the MWU framework) until certifying that there are no incomparable
2-cuts C of value C(C) ≤ (1 + )λ, in O(cid:0)m log2 n + K log2 n + I log n(cid:1) total running time, where
K is the number of nested 2-cuts of value ≤ (1 + O())λ output and I is the total number of weight
increments returned by inc-cut.
5 Applying multiplicative weight updates along cuts
We address the remaining issue of implementing the lazy-inc-cuts data structure for a fixed and
rooted spanning tree T . An interface for lazy-inc-cuts was given previously in Figure 4 and we
target the bounds claimed by Lemma 4.1. Recall that the MWU framework takes a cut C ∈ C,
computes the smallest capacity γ = arg mine∈C ce in the cut, and increases the weight of each
cut-edge e ∈ C by a multiplicative factor of exp(γ/ce) (see Figure 1). A subtle point is that the
techniques of Section 4 identify approximate min-cuts without explicitly listing the edges in the cut.
A 1-cut C(D(s)) is simply identified by the root s of the down-set D(s), and likewise 2-cuts of the
form C(D(s) ∪ D(t)) (when s (cid:107) t) and C(D(t) \ D(s)) (when s < t) can be described by the two
nodes s and t. In particular, an approximately minimum cut C is identified without paying for the
number of edges O(C) in C in the running time. When it comes to updating the edge weights, the
natural approach of visiting each edge e ∈ C would be too slow, to say nothing of even identifying
all the edges in C in time proportional to C.
While the general task of incrementing2 weights along a cut appears difficult to execute both
quickly and exactly, we have already massaged the setting to be substantially easier. First, we can
afford to approximate the edge weights we by a small multiplicative error. This means, for example,
that a cut edge e ∈ C with very large capacity ce (cid:29) γ can to some extent be ignored. Second,
we are not incrementing weights along any cut, but just the 1-cuts and 2-cuts of a fixed rooted
spanning tree T . We have already seen in Section 4 that restricting ourselves to 1-cuts and 2-cuts
allows us to (basically) apply dynamic programming to find small cuts, and also allows us to use
dynamic trees to efficiently update and scan various values in the aggregate. Here too we will see
that 1-cuts and 2-cuts are simple enough to be represented efficiently in standard data structures.
In the following, we use the term "cuts" liberally as the set of edges with endpoints in each of two
sets of vertices; in particular, the two sets of vertices may not be disjoint.
Lemma 5.1. Let T be a fixed rooted spanning tree of an undirected graph G = (V,E) with E = m
DT ⊆ 2E such that
(i) every edge e ∈ E appears in at most O(log2 n) cuts D ∈ DT , and
(ii) every 1-cut or 2-cut C ∈ CT (described succinctly by at most 2 roots of subtrees) can be
decomposed into the disjoint union C = D1 (cid:116) ··· (cid:116) D(cid:96) of (cid:96) = O(log2 n) cuts D1, . . . , D(cid:96) ∈ DT
in O(log2 n) time.
By building the collection D ∈ DT once for a tree T , Lemma 5.1 reduces the problem of
incrementing along any 1-cut or 2-cut C ∈ CT to incrementing along a "canonical" cut D ∈ DT
known a priori. This is important because multiplicative weight updates can be applied to a static
set relatively efficiently by known techniques [Young, 2014, Chekuri and Quanrud, 2017b].
It is
edges and V = n vertices. In O(cid:0)m log2 n(cid:1) time, one can construct a collection of nonempty cuts
also important that cuts in DT are sparse, in the sense that(cid:80)
D∈DT D = O(cid:0)m log2 n(cid:1), as this sum
2In the MWU framework the weights increase in a multiplicative fashion. We use the term "incrementing" in place
of "updating" since weights are increasing and also because it is convenient to think in terms of the logarithm of the
weights, which do increase additively.
20
Figure 8: An edge cut e by a subtree Ts, and two possible inequalities in the Euler order.
± (see the picture on the right).
≤ u−
factors directly into the running time guarantees. We first prove Lemma 5.1 in Section 5.1, and then
in Section 5.2 we show how to combine amortized data structures for each D ∈ DT to increment
along any 1-cut or 2-cut C ∈ CT .
5.1 Canonical cuts
Euler tours and orderings. Let T be a fixed and rooted tree on
n vertices V, and fix an Euler tour on a bidirected copy of T that
replaces each edge of T with arcs in both directions, starting from the
root. For each vertex v, we create two symbols: v− means we enter
An Euler tour inducing the or-
dering r− < a− < a+ < b− <
the subtree Tv rooted at v, and v+ means we leave the subtree v. Let
c− < c+ < d− < d+ < b+ < r+.
± = {v−, v+ : v ∈ V(T )} denote the whole collection of these 2n symbols. The Euler tour enters
V
and leaves each subtree exactly once in a fixed order. Tracing the Euler tree induces a unique total
ordering on V
This ordering has a couple of interesting properties. For every vertex v ∈ V(T ), we have v− < v+.
Letting r denote the root of T , r− is the first element in the ordering and r+ is the last element in the
ordering. More generally, for a vertex v and a vertex w ∈ D(v) in the subtree rooted at v, we have
v−
≤ w− < w+ ≤ v+, with all inequalities strict if v (cid:54)= w. That is, the ranges {[v−, v+] : v ∈ V(T )}
between entering and leaving a subtree form a laminar set.
The Euler order on the vertices endows a sort of geometry to the edges. Each edge e = (u, v)
≤ s−
(with u− < v−) can be thought of as an interval [u−, v−]. A downset D(s) cuts e iff u−
≤
v−
} ∩ [s−, s+] = 1. Alternatively,
≤ s+ or s−
D(s) cuts e iff [u−, v−] ∩ {s−, s+} = 1. These observations suggest that this is not a problem about
graphs, but about intervals, and perhaps a problem suited for range trees.
± at its leaves. As a
Range trees on the Euler order. Let R be a balanced range tree with V
balanced tree with 2n leaves, R has O(n) nodes and height O(log n). Each range-node a ∈ V(R)
± at the leaves of the subtree Ra
induces an interval Ia on V
± (induced by a), and let I = {Ia : a ∈ V(R)}
rooted at a. We call Ia a canonical interval of V
denote the collection of all O(n) canonical intervals.
≤ s+ ≤ v− (see Figure 8); that is, iff {u−, v−
±, consisting of all the elements of V
Each element in V
± appears in O(log n) canonical intervals because the
± decomposes into
height of R is O(log n). Moreover, every interval J on V
the disjoint union of O(log n) canonical intervals. The decomposition can be
obtained in O(log n) time by tracing the two paths from the root of R to the
endpoints of J and taking the canonical intervals corresponding to the O(log n) maximal subtrees
between the paths (see the picture on the right). Similarly, the union of a constant number of
intervals or the complement of the union of a constant number of intervals decomposes to O(log n)
canonical intervals in O(log n) time.
21
usveu s v s+usves u s+v rabcdJ The canonical intervals I relate to the 1-cuts and 2-cuts CT induced by T as follows. For any
pair of disjoint canonical intervals I1, I2 ∈ I, let C(I1, I2) = {(u, v) ∈ E : u ∈ I1, v ∈ I2} be the set
of edges with one endpoint in I1 and the other in I2. We call C(I1, I2) a canonical cut induced by
I1 and I2. We claim that any 1-cut or 2-cut C ∈ CT decomposes into the disjoint union of O(log2 n)
canonical cuts.
For starters, let C = C(D(s)) be the 1-cut induced by the downset of a vertex
s. The 1-cut C(D(s)) consists of all edges (u, v) with one (tagged) endpoint u−
[s−, s+] decomposes to the disjoint union [s−, s+] =(cid:70)
in the interval [s−, s+] and the other endpoint v− outside [s−, s+]. The interval
the disjoint union [r−, s−)∪ (s+, r+] =(cid:70)
I1∈I1 I1 of O(log n) canon-
ical intervals I1 ⊆ I, and the complement [r−, s−)∪ (s+, r+] also decomposes to
I2 ⊆ I (see the picture on the right). Together, the disjoint union C(D(s)) =(cid:70)
I2∈I2 I2 of O(log n) canonical intervals
C(I1, I2)
of canonical cuts over the cross product (I1, I2) ∈ I1 × I2 decomposes the 1-cut C(D(s)) into
I1 × I2 = O(log2 n) canonical cuts. Moreover, both decompositions I1 and I2 can be obtained
in O(log n) time.
Any 2-cut decomposes similarly. If s (cid:107) t are incomparable, with (say) s− < s+ < t− < t+, then
the incomparable 2-cut C = C(D(s) ∪ D(t)) is the set of edges with one endpoint in [s−, s+]∪[t−, t+]
and other endpoint in the complement, [r−, s−)∪(s+, t−)∪(t+, r+]. Both of these sets are the disjoint
cross-product of the two decompositions decomposes C(D(s) ∪ D(t)) into O(cid:0)log2 n(cid:1) canonical cuts.
union of two or three intervals, and they each decompose into O(log n) canonical intervals. The
If s < t are two comparable vertices, then t− < s− < s+ < t+ and the nested 2-cut
C(D(t) \ D(s)) is the set of edges with one endpoint in [t−, s−) ∪ (s+, t+] and the other endpoint
into O(cid:0)log2 n(cid:1) disjoint canonical cuts.
in the complement [r−, t−) ∪ [s−, s+] ∪ [t+, r+] and the cross-product breaks C(D(t) \ D(s)) down
Thus, any 1-cut or 2-cut C ∈ CT breaks down into to the disjoint union of O(log2 n) canonical
cuts. An edge e = (u, v) appears in a canonical cut C(I1, I2) iff u ∈ I1 and v ∈ I2. As either
cuts are easily constructed in O(m log2 n) time by adding each edge to the O(cid:0)log2 n(cid:1) canonical cuts
end point u or v appears in O(log n) canonical intervals, e appears in at most O(log2 n) canonical
cuts. In turn, there are at most O(m log2 n) nonempty canonical cuts. The nonempty canonical
I1∈I1,I2∈I2
containing it. Taking DT to be this set of nonempty canonical cuts gives Lemma 5.1.
5.2 Lazy weight increments
Lemma 5.1 identifies a relatively small set of canonical cuts DT such that any 1-cut or 2-cut C ∈ CT
is the disjoint union of O(cid:0)log2 n(cid:1) canonical cuts D1, . . . , DO(log2 n) ∈ DT . While we can now at least
gather an implicit list of all the edges of C in O(cid:0)log2 n(cid:1) time, it still appears necessary to visit every
edge e ∈ C to increment its weight. The challenge is particularly tricky because the multiplicative
weight updates are not uniform. However, for a fixed and static set, recent techniques by Young
[2014] and Chekuri and Quanrud [2017b] show how to approximately apply a multiplicative weight
update to the entire set in polylogarithmic amortized time. In this section, we apply these techniques
to each canonical cut D ∈ DT and show how to combine their outputs carefully to meet the claimed
bounds of Lemma 4.1.
For every canonical cut D ∈ DT , we employ the lazy-inc data structure of Chekuri and
Quanrud [2017b]. The lazy-incs data structure gives a clean interface to similar ideas in Young
[2014], and we briefly describe a restricted version of the data structure in [Chekuri and Quanrud,
2017b] per the needs of this paper3. The lazy-incs data structure is an amortized data structure
3What we refer to as lazy-incs is most similar to lazy-incs(0) in Chekuri and Quanrud [2017b].
22
[s ,s+][r ,s )(s+,r+] that approximates a set of counters increasing concurrently at different rates. For a fixed instance of
lazy-incs for a particular canonical cut D ∈ DT , we will have one counter for each edge tracking
the "additive part" ve = ln(we)/ for each edge e ∈ D. The rate of each counter e is stored as
rate(e), and in this setting is proportional to the inverse of the capacity ce. The primary routine is
inc(ρ), which simulates a fractional increase on each counter proportional to a single increment at
the rate ρ. That is, each counter ve increases by rate(e)/ρ. The argument ρ will always be taken
to be proportional to the inverse of the minimum capacity of any edge in the cut, γ = mine∈C ce.
The output of inc(ρ) is a list of increments {(e, δe)} over some of the edges in D. The routine
does not return an increment (e, δe) unless δe is substantially large. This allows us to charge off the
work required to propagate the increment (e, δe) to the rest of the algorithm to the maximum weight
of e. In exchange for reducing the number of increments, the sum of returned increments for a fixed
edge e underestimates ve by a small additive factor. An underestimate for ve = ln(we)/ within
an additive factor of O(1) translates to an underestimate for we = exp(ve) within a (1 + O())-
multiplicative factor, as desired.
At a high level, lazy-incs buckets the counters by rounding up each rate to the nearest power
of 2. For each power of 2, an auxiliary counter is maintained with rate set to this power (more or
less) exactly and efficiently. Counters with rates that are powers of 2 can be maintained in constant
amortized time for the same reason that a binary number can be incremented in constant amortized
time. When an auxiliary counter increases to the next whole integer, the tracked counters in the
corresponding bucket are increased proportionately. Up to accounting details, this lazy scheme
tracks the increments of each counter to within a constant additive factor at all times. A more
general implementation of lazy-incs and a detailed proof of the following theorem are given in
[Chekuri and Quanrud, 2017b].
Lemma 5.2 (Chekuri and Quanrud, 2017b). Let e1, . . . , ek be k counters with rates rate(e1) ≥
rate(e2) ≥ ··· ≥ rate(ek) in sorted order.
(a) An instance of lazy-incs can be initialized in O(k) time.
(b) Each inc runs in O(1) amortized time plus O(1) for each increment returned.
(c) flush runs in O(k) time.
(d) For each counter ei, let vi be the true value of the counter and let vi be the sum of increments
for counter i in the return values of inc and flush.
(i) After each call to inc, we have vi ≤ vi ≤ vi + O(1) for each counter ei.
(ii) After each call to flush, we have vi = vi for each counter ei.
We note that flush is not implemented in [Chekuri and Quanrud, 2017b], but is easy enough
to execute by just reading off all the residual increments in the data structure and resetting these
values to 0.
The application of lazy-incs to our problem bears some resemblance to its application to
packing intervals in [Chekuri and Quanrud, 2017b], where the weights are also structured by range
trees. We first invoke Lemma 5.1 to generate the family of canonical cuts DT . For each canonical cut
D ∈ DT , by Lemma 5.2, we instantiate an instance of lazy-incs where each edge e ∈ D has rate
rate(e) = log2 n/ce. (Lemma 5.2 takes as an input a collection of counters sorted by rate. Rather
than sort each canonical cut individually, we can save a log(n)-factor by a slight modification to
Lemma 5.1. If the edges are sorted before applying Lemma 5.1, then each canonical cut D ∈ DT
will naturally be in sorted order.) We also compute, for each canonical cut D ∈ DT , the minimum
When incrementing weights along a cut C ∈ CT , we update the weights as follows. Lemma
(cid:48)
5.1 divides C into the union of O(log2 n) disjoint canonical cuts D
T ⊆ DT . We find the minimum
capacity γ = mine∈C ce by taking the minimum of the precomputed minimum capacities of each
capacity γD = mine∈D ce in the cut. These preprocessing steps take O(cid:0)m log2 n(cid:1) time total.
23
(cid:48)
T , i.e., γ = minD∈D(cid:48)
T
γD. For each canonical set D ∈ D
(cid:48)
canonical cut D ∈ D
T , we call inc(log2 n/γ)
on the corresponding lazy-incs instance. For every increment (e, δ) returned in lazy-incs, where
e ∈ C and δ > 0 represents a increment to ve = ln(we)/, we increase ve by an additive factor of
δ/ log2 n, and multiply we by exp(cid:0)δ/ log2 n(cid:1) accordingly.
Fix an edge e ∈ E. The edge e is a member of O(cid:0)log2 n(cid:1) canonical cuts in DT , and receives
its weight increments from O(log2 n) instances of lazy-incs. The slight errors from each instance
of lazy-incs accumulate additively (w/r/t ve = ln(we)/). Although each instance of lazy-incs
promises only a constant additive error in Lemma 5.2, we scaled up the rate of e from 1/ce to
log2 n/ce and divide the returned weight increments by a factor of log2 n. The scaling reduces the
error from each instance of lazy-incs from an additive factor of O(1) to O(1/ log2 n). Consequently,
the sum of all O(log2 n) instances of lazy-incs underestimates ve to an O(1) additive error w/r/t
ve, and underestimates we by at most a (1 + O())-multiplicative factor, as desired.
Increasing the sensitivity of each lazy-incs data structure by a factor of log2 n means every
increment (e, δ) returned by inc may increase ve by as little as 1/ log2 n. An additive increase
in 1/ log2 n corresponds to a multiplicative increase of exp(cid:0)/ log2 n(cid:1) in we, hence property (e) in
Lemma 4.1.
Together, the construction of the canonical cuts DT and the O(m log2 n) carefully calibrated
instances of lazy-inc, one for each canonical cut D ∈ DT , give Lemma 4.1.
6 Putting it all together: Proof of Theorem 1.1
(cid:48) = O(log n).
In this section we combine the ingredients discussed so far and outline the proof of Theorem 1.1.
Let (G, c) be an Metric-TSP instance and (cid:48) > 0 be the error parameter. We assume without loss
of generality that that (cid:48) is sufficiently small, and in particular that (cid:48) < 1/2. We will also assume
that (cid:48) > 1/n2 for otherwise one could use an exact algorithm and achieve the desired time bound.
This implies that log 1
Consider the algorithm in Figure 5 which takes as a parameter. We will choose = ρ(cid:48) for some
sufficiently small ρ. It thus suffices to argue that the algorithm outputs a (1 + O())-approximation
with high probability.
High-level MWU analysis: At a high-level, our algorithm is a standard width-independent
MWU algorithm for pure packing problems with an α-approximate oracle, for α = (1 + O()). Our
implementation follows the "time"-based algorithm of Chekuri et al. [2015]. An exact oracle in this
setting repeatedly solves the global minimum cut problem w/r/t the edge weights w. To implement
an α-approximate oracle, every cut output needs to be an α-approximate minimum cut.
To argue that we are indeed implementing an α-approximate minimum cut oracle, let us identify
the two points at which we deviate from Karger's minimum cut algorithm [2000], which would
otherwise be an exact oracle that succeeds with high probability. In both Karger's algorithm and
our partially dynamic extension, we sample enough spanning trees from an approximately maximum
tree packing to contain all (1+O())-approximate minimum cuts as a 1-cut or 2-cut with probability
1−poly(1/n) = 1−poly(/n) (see Theorem 3.1). The sampled tree packings are the only randomized
component of the algorithm, and we will argue that all the sampled tree packings succeed with high
probability later after establishing a basic correctness in the event that all the sampled packings
succeed.
Karger's algorithm searches all the spanning trees for the minimum 1-cut or 2-cut. We retrace
the same search, except we output any approximate minimum 1-cut or 2-cut found in the search.
More precisely, we maintain a target value λ with the invariant that there are no cuts of value < λ,
and output any 1-cut or 2-cut with weight ≤ (1 + O())λ at that moment (see Lemma 4.2). Since
24
λ is no more than the true minimum cut w/r/t w at any point, any cut with weight ≤ (1 + O())λ
is a (1 + O())-approximation to the current minimum cut.
The second point of departure is that we do not work with the "true" weights implied by the
framework, but a close approximation. By Lemma 4.1, the approximated weights underestimate
the true weights by at most a (1 + O())-multiplicative factor. In turn, we may underestimate the
value of a cut by at most a (1 + O())-multiplicative factor, but crucially any 1-cut or 2-cut will
still have value ≤ (1 + O())λ, for a slightly larger constant hidden in the O().
This firmly establishes that the proposed algorithm in fact implements an α-approximate oracle
for α = 1 + O() (with high probability). The MWU framework then guarantees that the final pack-
ing of cuts output by the algorithm approximately satisfies all the constraints with approximately
optimal value, as desired.
Probability of failure. We now consider the probability of the algorithm failing. Randomiza-
tion enters the algorithm for two reasons. Initially, we invoke Karger's randomized minimum cut
algorithm [2000] to compute the value of the minimum cut w/r/t the initial weights w = 1/c, in
order to set the first value of λ. Second, at the beginning of each epoch, we randomly sample a
subset of an approximately maximum tree packing hoping that the sampled trees contain every
(1 + O())-approximate minimum cut as a 1-cut or 2-cut in one of the sampled trees4.
Karger [2000] showed that the minimum cut algorithm fails with probability at most poly(1/n).
He also showed (see Theorem 3.1) that each random sample of O(log(n/)) = O(log n) spanning
trees of an approximately maximum tree packing (conducted at the beginning of an epoch) fails to
capture all (1 + )-approximate minimum cuts with probability at most poly(1/n) = poly(/n). As
discussed in Section 3, there are a total of O(cid:0)log n/2(cid:1) epochs over the course of the algorithm. By
the union bound, the probability of any random sample of spanning trees failing is O(cid:0)log n/2(cid:1)
The algorithm initializes the edge weights in O(m) time, and invokes Karger's minimum cut
·
poly(/n) = poly(/n). Taking a union bound again with the chance of the initial call to Karger's
minimum cut algorithm failing, the probability of any randomized component of the algorithm
failing is at most poly(1/n), as desired.
It remains to bound the running time of the algorithm. The analysis is not
Running time.
entirely straightforward, as some operations can be bounded directly, while others are charged
against various upper bounds given by the MWU framework.
algorithm [2000] once which runs in O(cid:0)m log3 n(cid:1) time. As established in Section 3, the remaining
algorithm is divided into O(cid:0)log(n)/2(cid:1) epochs. Each epoch invokes Theorem 3.1 once to sample
O(log(n/)) = O(log n) spanning trees from an approximate tree packing in O(cid:0)m log3 n(cid:1) time. We
then invoke Lemma 4.2 for each spanning tree in the sample. Over the course of O(cid:0)log(n)/2(cid:1)
epochs, each with O(log(n)) spanning trees, we invoke Lemma 4.2 at total of O(cid:0)log2(n)/2(cid:1) times.
weight increments. By Lemma 4.2, the total time to process T is O(cid:0)m log2 n + K log2 n + I log n(cid:1).
The first term, O(cid:0)m log2 n(cid:1), comes from following the tree-processing subroutine of Karger
[2000]. Over a total O(cid:0)log2(n)/2(cid:1) trees, the total time spent mimicking Karger's subroutine is
O(cid:0)m log4(n)/2(cid:1).
Suppose we process a tree T , outputting K approximately minimum cuts and making I full edge
The remaining terms of Lemma 4.2, depending on K and I, are introduced by the MWU
framework and can be amortized against standard properties of the MWU framework. The quantity
K represents the number of approximate minimum cuts output while processing a single tree T ,
and the sum of all K across all trees is the total number of approximate minimum cuts output.
Each such cut corresponds to a single iteration, and the total number of iterations in the MWU
4Interestingly, the only randomization in Karger's algorithm also stems from randomly sampling an approximately
maximum tree packing.
25
framework is O(cid:0)m log(n)/2(cid:1). Thus, the sum of the second term O(cid:0)K log2 n(cid:1) over all invocations
of Lemma 4.2 is O(cid:0)m log3(n)/2(cid:1).
O(cid:0)log3(n)/2(cid:1) times (see Section 2). It follows that the total number of edge increments, over all
edges, returned by lazy-inc-cuts is O(cid:0)m log3(n)/2(cid:1). Thus, the sum of the quantities O(I log n)
over all invocations of Lemma 4.2 is O(cid:0)m log4(n)/2(cid:1).
All told, the algorithm runs in O(cid:0)m log4(n)/2(cid:1) = O(cid:0)m/2(cid:1) time total.
The third term of Lemma 4.2, O(I log n) depends on the number of increments I returned by
the inc-cut subroutine of the lazy-inc-cuts data structure. Each increment (e, δ) returned by
inc-cut increases we by at least a (1 + Ω(/ log2 n))-multiplicative factor (i.e., δ ≥ Ω(we/ log2 n)).
The MWU framework shows that a single edge can increase by a (1 + Ω(/ log n)) factor at most
Acknowledgements. We thank Neal Young for suggesting that Metric-TSP may be amenable to
the techniques from [Chekuri and Quanrud, 2017b].
References
P. K. Agarwal and J. Pan. Near-linear algorithms for geometric hitting sets and set covers. In Proc.
30th Annu. Sympos. Comput. Geom. (SoCG), page 271, 2014.
Z. Allen-Zhu and L. Orecchia. Nearly-linear time positive LP solver with faster convergence rate.
In Proc. 47th Annu. ACM Sympos. Theory Comput. (STOC), pages 229–236, 2015.
D. Applegate, R. Bixby, V. Chvátal, and W. Cook. Implementing the dantzig-fulkerson-johnson
algorithm for large traveling salesman problems. Mathematical programming, 97(1):91–153, 2003.
D. L. Applegate, R. E. Bixby, V. Chvatal, and W. J. Cook. The traveling salesman problem: a
computational study. Princeton University Press, 2011.
S. Arora, E. Hazan, and S. Kale. The multiplicative weights update method: a meta-algorithm and
applications. Theory of Computing, 8(1):121–164, 2012.
S. C. Boyd and W. R. Pulleyblank. Optimizing over the subtour polytope of the travelling salesman
problem. Mathematical programming, 49(1-3):163–187, 1990.
C. Chekuri and K. Quanrud. Faster approximation for Metric-TSP via linear programming.
Manuscript, October 2017a.
C. Chekuri and K. Quanrud. Near-linear time approximation schemes for some implicit fractional
packing problems. To appear in Proc. 28th ACM-SIAM Sympos. Discrete Algs. (SODA), 2017,
2017b.
C. Chekuri, T. Jayram, and J. Vondrák. On multiplicative weight updates for concave and sub-
modular function maximization. In Proc. 6th Conf. Innov. Theoret. Comp. Sci. (ITCS), pages
201–210, 2015.
N. Christofides. Worst-case analysis of a new heuristic for the traveling salesman problem. Technical
Report 388, Graduate School of Industrial Administration, Carnegie Mellon University, 1976.
W. J. Cook. In pursuit of the traveling salesman: mathematics at the limits of computation. Prince-
ton University Press, 2014.
26
G. B. Dantzig, D. R. Fulkerson, and S. M. Johnson. Solution of a large-scale traveling-salesman
problem. Operations Research, 2(4):393–410, 1954.
L. K. Fleischer. Approximating fractional multicommodity flow independent of the number of
commodities. SIAM J. Discrete Math., 13(4):505–520, 2000. Preliminary version in Proc. 40th
Annu. IEEE Sympos. Found. Comput. Sci. (FOCS), 1999.
H. N. Gabow. A matroid approach to finding edge connectivity and packing arborescences. J.
Comput. Sys. Sci., 50(2):259–273, 1995. Preliminary version in Proc. 23rd Annu. ACM Sympos.
Theory Comput. (STOC), 1991.
N. Garg and J. Könemann. Faster and simpler algorithms for multicommodity flow and other
fractional packing problems. SIAM J. Comput., 37(2):630–652, 2007. Preliminary version in
Proc. 39th Annu. IEEE Sympos. Found. Comput. Sci. (FOCS), 1998.
K. Genova and D. P. Williamson. An experimental evaluation of the Best-of-Many Christofides'
Algorithm for the Traveling Salesman Problem. Algorithmica, pages 1–22, 2017. Preliminary
version appeared in Proceedings of ESA 2015.
M. X. Goemans. Worst-case comparison of valid inequalities for the TSP. Math. Prog., 69(1-3):
335–349, 1995.
M. X. Goemans and D. Bertsimas. Survivable networks, linear programming relaxations and the
parsimonious property. Math. Prog., 60(1):145–166, June 1993.
G. Goranci, M. Henzinger, and M. Thorup. Incremental exact min-cut in poly-logarithmic amor-
tized update time. CoRR, abs/1611.06500, 2016. URL http://arxiv.org/abs/1611.06500.
Preliminary version in Proceedings of ESA 2016.
M. D. Grigoriadis and L. G. Khachiyan. Fast approximation schemes for convex programs with
many blocks and coupling constraints. SIAM J. Optim., 4(1):86–107, 1994.
G. Gutin and A. P. Punnen. The traveling salesman problem and its variations, volume 12. Springer
Science & Business Media, 2006.
D. Harel and R. E. Tarjan. Fast algorithms for finding nearest common ancestors. SIAM J. Comput.,
13(2):338–355, 1984.
M. Held and R. M. Karp. The traveling-salesman problem and minimum spanning trees. Operations
Research, 18(6):1138–1162, 1970.
M. R. Henzinger and V. King. Randomized fully dynamic graph algorithms with polylogarithmic
time per operation. J. Assoc. Comput. Mach., 46(4):502–516, 1999. Preliminary version in Proc.
27th Annu. ACM Sympos. Theory Comput. (STOC), 1995.
D. R. Karger. Random sampling and greedy sparsification for matroid optimization problems.
Math. Program., 82:41–81, 1998. Preliminary version in Proc. 34th Annu. IEEE Sympos. Found.
Comput. Sci. (FOCS), 1993.
D. R. Karger. Minimum cuts in near-linear time. J. Assoc. Comput. Mach., 47(1):46–76, 2000.
Preliminary version in Proc. 28rd Annu. ACM Sympos. Theory Comput. (STOC), 1996.
R. Khandekar. Lagrangian relaxation based algorithms for convex programming problems. PhD
thesis, Indian Institute of Technology Delhi, Mar. 2004.
27
C. Koufogiannakis and N. E. Young. A nearly linear-time PTAS for explicit fractional packing and
covering linear programs. Algorithmica, 70(4):648–674, 2014. Preliminary version in Proc. 48th
Annu. IEEE Sympos. Found. Comput. Sci. (FOCS), 2007.
M. Lampis. Improved inapproximability for TSP. In Approximation, Randomization, and Combi-
natorial Optimization. Algorithms and Techniques, pages 243–253. Springer, 2012.
E. L. Lawler, J. K. Lenstra, A.-G. Rinnooy-Kan, and D. B. Shmoys. The traveling salesman problem.
John Wiley & Sons, Ltd., 1985.
C. L. Monma, B. S. Munson, and W. R. Pulleyblank. Minimum-weight two-connected spanning
networks. Math. Prog., 46(1):153–171, Jan. 1990.
A. Mądry. Faster approximation schemes for fractional multicommodity flow problems via dynamic
graph algorithms. In Proc. 42nd Annu. ACM Sympos. Theory Comput. (STOC), pages 121–130,
2010.
C. S. J. A. Nash-Williams. Edge-disjoint spanning trees of finite graphs. J. London Math. Soc., 36:
445–450, 1961.
R. Peng. Approximate undirected maximum flows in o(mpolylog(n)) time. In Proc. 27th ACM-
SIAM Sympos. Discrete Algs. (SODA), pages 1862–1867, 2016.
S. A. Plotkin, D. B. Shmoys, and É. Tardos. Fast approximation algorithms for fractional packing
and covering problems. Math. of Oper. Res., 20(2):257–301, 1995.
D. Sleator and R. E. Tarjan. A data structure for dynamic trees. J. Comput. Syst. Sci., 26:362–391,
1983.
R. E. Tarjan and U. Vishkin. An efficient parallel biconnectivity algorithm. SIAM J. Comput., 14
(4):862–874, 1985.
M. Thorup. Fully-dynamic min-cut. Combinatorica, 27(1):91–127, 2007. Preliminary version in
Proc. 33rd Annu. ACM Sympos. Theory Comput. (STOC), 2001.
W. T. Tutte. On the problem of decomposing a graph into n connected components. J. London
Math. Soc., 36:221–230, 1961.
J. Vygen. New approximation algorithms for the TSP. OPTIMA, 90:1–12, 2012.
D. Wang, S. Rao, and M. Mahoney. Unified acceleration method for packing and covering problems
via diameter reduction. CoRR, abs/1508.02439, August 2015. To appear in Proc. 43rd Internat.
Colloq. Automata Lang. Prog. (ICALP), 2016.
N. E. Young. Nearly linear-time approximation schemes for mixed packing/covering and facility-
location linear programs. CoRR, abs/1407.3015, 2014. URL http://arxiv.org/abs/1407.3015.
28
A Computing the primal solution
Our MWU algorithm computes a (1−)-approximate solution to 2ECSSD. Recall that the algorithm
maintains a weight w(e) for each edge e. These weights evolve with time. Using a standard argument
we show how we can recover a (1 + )-approximation to the LP 2ECSS.
1
1+ OPT.
iteration i is ((cid:80)
Let OPT be the optimum solution value of 2ECSSD and 2ECSS. Our algorithm computes an
implicit packing in the LP 2ECSSD of value (1 − ) OPT in near-linear time. It does this via the
MWU method which works in several iterations. In each iteration i the algorithm computes a global
mincut Ci with respects to the edge weights wi(e) where wi(e) is the weight of edge e in iteration i.
Each iteration corresponds to solving a relaxation of the original LP by collapsing the constraints
into a single constraint via the weights in iteration i. The value of the solution to this relaxation in
e wi(e)c(e))/wi(Ci). Since this is a relaxation to the original LP, the value of this
solution is at least OPT. Over the entire algorithm we have a time of one unit, and the integral of
all solutions we compute is at least OPT. Let z be the averaged solution at the end of the algorithm.
Via the MWU analysis z violates each constraint by at most a (1 + )-factor. Thus by scaling down
z by (1 + ), we ensure feasibility of constraints and obtain a feasible solution to 2ECSSD of value
at least
(cid:80)
We claim (and this is standard) that there is an iteration i such that the value of the solution
more than OPT, a contradiction. Given an iteration i where (cid:80)
e wi(e)c(e)/wi(Ci) ≤ (1 + ) OPT. If not, the value of z would be larger than (1 + ) OPT and
we would, after scaling down by a factor of (1 + ), obtain a feasible solution to 2ECSSD of value
e wi(e)c(e)/wi(Ci) ≤ (1 + ) OPT
feasible, consider any cut Cj. The capacity of the cut is (cid:80)
we can obtain a feasible solution to 2ECSS as follows. Simply set y(e) = wi(e)/wi(Ci) for each
is the cheapest cut with respect to the weights wi(e). The cost of the solution y is(cid:80)
e; we claim that this is a feasible solution whose cost is at most (1 + ) OPT. To see that y is
(cid:80)
e∈Cj ye = wi(Cj)/wi(Ci) ≥ 1 since Ci
e y(e)c(e) =
e wi(e)c(e)/wi(Ci) ≤ (1 + ) OPT by choice of i.
to within a (1 + O()) factor as well as the sum (cid:80)
The solution y can be output in nearly linear time. Our algorithm maintains the weights wi(e)
implicitly via data structures. However, during iteration i we know the value of the mincut wi(Ci)
(cid:80)
e wi(e)c(e). In particular, we can keep track
of the iteration i∗ that minimizes this ratio; this iteration i∗ will satisfy the desired condition,
e wi∗(e)c(e)/wi∗(Ci∗) ≤ (1 + O()) OPT. The actual set of weights wi∗(e) needed to create
the solution y(e) can be recovered by either replaying the algorithm (by storing the necessary
randomization from the first run) or simply maintaining the weights by persistent data structures
and keeping a pointer to the persistent copy of the weights from iteration i∗.
29
|
1106.6336 | 1 | 1106 | 2011-06-30T18:43:43 | External-Memory Network Analysis Algorithms for Naturally Sparse Graphs | [
"cs.DS"
] | In this paper, we present a number of network-analysis algorithms in the external-memory model. We focus on methods for large naturally sparse graphs, that is, n-vertex graphs that have O(n) edges and are structured so that this sparsity property holds for any subgraph of such a graph. We give efficient external-memory algorithms for the following problems for such graphs: - Finding an approximate d-degeneracy ordering; - Finding a cycle of length exactly c; - Enumerating all maximal cliques. Such problems are of interest, for example, in the analysis of social networks, where they are used to study network cohesion. | cs.DS | cs |
External-Memory Network Analysis Algorithms
for Naturally Sparse Graphs
Michael T. Goodrich and Pawe(cid:32)l Pszona
Dept. of Computer Science
University of California, Irvine
Abstract. In this paper, we present a number of network-analysis al-
gorithms in the external-memory model. We focus on methods for large
naturally sparse graphs, that is, n-vertex graphs that have O(n) edges
and are structured so that this sparsity property holds for any subgraph
of such a graph. We give efficient external-memory algorithms for the
following problems for such graphs:
1. Finding an approximate d-degeneracy ordering.
2. Finding a cycle of length exactly c.
3. Enumerating all maximal cliques.
Such problems are of interest, for example, in the analysis of social
networks, where they are used to study network cohesion.
1
Introduction
Network analysis studies the structure of relationships between various entities,
with those entities represented as vertices in a graph and their relationships
represented as edges in that graph (e.g., see [11]). For example, such structural
analyses include link-analysis for Web graphs, centrality and cohesion measures
in social networks, and network motifs in biological networks. In this paper,
we are particularly interested in network analysis algorithms for finding various
kinds of small subgraphs and graph partitions in large graphs that are likely to
occur in practice. Of course, this begs the question of what kinds of graphs are
likely to occur in practice.
1.1 Naturally Sparse Graphs
A network property addressing the concept of a "real world" graph that is
gaining in prominence is the k-core number [25], which is equivalent to a graph's
width [16], linkage [19], k-inductivity [18], and k-degeneracy [2,21], and is one
less than its Erdos-Hajnal coloring number [14]. A k-core, G(cid:48), in a graph, G, is
a maximal connected subgraph of G such that each vertex in G(cid:48) has degree at
least k. The k-core number of a graph G is the maximum k such that G has a
non-empty k-core. We say that a graph G is naturally sparse if its k-core number
is O(1). This terminology is motivated by the fact that almost every n-vertex
graph with O(n) edges has a bounded k-core number, since Pittel et al. [23] show
that a random graph with n vertices and cn edges (in the Erdos-R´enyi model)
has k-core number at most 2c + o(c), with high probability. Riordan [24] and
Fernholz and Ramachandran [15] have also studied k-cores in random graphs.
In addition, we also have the following:
-- Every s-vertex subgraph of a naturally sparse graph is naturally sparse,
hence, has O(s) edges.
-- Any planar graph has k-core number at most 5, hence, is naturally sparse.
-- Any graph with bounded arboricity is naturally sparse (e.g., see [10]).
-- Eppstein and Strash [13] verify experimentally that real-world graphs in
four different data repositories all have small k-core numbers relative to
their sizes; hence, these real-world graphs give an empirical motivation for
naturally sparse graphs.
-- Any network generated by the Barab´asi-Albert [4] preferential attachment
process, with m ∈ O(1), or as in Kleinberg's small-world model [20], is
naturally sparse.
Of course, one can artificially define an n-vertex graph, G(cid:48), with O(n) edges that
is not naturally sparse just by creating a clique of O(n1/2) vertices in an n-vertex
graph, G, having O(n) edges. We would argue, however, that such a graph G(cid:48)
would not arise "naturally." We are interested in algorithms for large, naturally
sparse graphs.
1.2 External-Memory Algorithms
One well-recognized way of designing algorithms for processing large data sets
is to formulate such algorithms in the external memory model (e.g., see the
excellent survey by Vitter [27]). In this model, we have a single CPU with
main memory capable of storing M items and that computer is connected to
D external disks that are capable of storing a much larger amount of data.
Initially, we assume the parallel disks are storing an input of size N . A single
I/O between one of the external disks and main memory is defined as either
reading a block of B consecutively stored items into memory or writing a block
of the same size to a disk. Moreover, we assume that this can be done on all D
disks in parallel if need be.
Two fundamental primitives of the model are scanning and sorting. Scanning
is the operation of streaming N items stored on D disks through main memory,
with I/O complexity
(cid:18) N
(cid:19)
DB
logM/B
,
N
B
(cid:19)
,
scan(N ) = Θ
(cid:18) N
DB
and sorting N items has I/O complexity
sort(N ) = Θ
e.g., see Vitter [27].
Since this paper concerns graphs, we assume a problem instance is a graph
G = (V, E), with n = V , m = E and N = G = m + n. If G is d-degenerate,
that is, has k-core number, d, then m ≤ dn and N = O(dn) = O(n) for d = O(1).
We use d to denote the k-core number of an input graph, G, and we use the term
"d-degenerate" as a shorthand for "k-core number equal to d."
2
1.3 Previous Related Work
Several researchers have studied algorithms for graphs with bounded k-core
numbers (e.g., see [1,3,12,17,18]). These methods are often based on the fact
that the vertices in a graph with k-core number, d, can be ordered by repeatedly
removing a vertex of degree at most d, which gives rise to a numbering of the
vertices, called a d-degeneracy ordering or Erdos-Hajnal sequence, such that each
vertex has at most d edges to higher-numbered vertices. In the RAM model, this
greedy algorithm takes O(n) time (e.g., see [5]). Bauer et al. [6] describe methods
for generating such graphs and their d-degeneracy orderings at random.
In the internal-memory RAM model, Eppstein et al. [12] show how to find all
maximal cliques in a d-degenerate graph in O(d3d/3n) time. Alon et al. [3] show
that one can find a cycle of length exactly c, or show that one does not exist, in
a d-degenerate graph in time O(d1−1/km2−1/k), if c = 4k − 2, time O(dm2−1/k),
if c = 4k − 1 or 4k, and time O(d1+1/km2−1/k), if c = 4k + 1.
A closely related concept to a d-degeneracy ordering is a k-core decomposition
of a graph, which is a labeling of each vertex v with the largest k such that
v belongs to a k-core. Such a labeling can also be produced by the simple
linear-time greedy algorithm that removes a vertex of minimum degree with
each iteration. Cheng et al. [9] describe recently an external-memory method for
constructing a k-core decomposition, but their method is unfortunately fatally
flawed1. The challenge in producing a k-core decomposition or d-degeneracy
ordering in external memory is that the standard greedy method, which works
so well in internal memory, can cause a large number of I/Os when implemented
in external memory. Thus, new approaches are needed.
1.4 Our Results
In this paper, we present efficient external-memory network analysis algorithms
for naturally sparse graphs (i.e., degenerate graphs with small degeneracy). First,
we give a simple algorithm for computing a (2 + )d-degeneracy ordering of a
d-degenerate graph G = (V, E), without the need to know the value of d in
advance. The I/O complexity of our algorithm is O(sort(dn)).
Second, we give an algorithm for determining whether a d-degenerate graph
G = (V, E) contains a simple cycle of a fixed length c. This algorithm uses
I/O complexity, where is a
(cid:16)
d1± ·(cid:0)k · sort(m2− 1
O
constant depending on c ∈ {4k − 2, . . . , 4k + 1}.
k ) + (4k)! · scan(m2− 1
k )(cid:1)(cid:17)
Finally, we present an algorithm for listing all maximal cliques of an undi-
rected d-degenerate graph G = (V, E), with O(3δ/3sort(dn)) I/O complexity,
where δ = (2 + )d.
One of the key insights to our second and third results is to show that, for
the sake of designing efficient external-memory algorithms, using a (2 + )d-
degeneracy ordering is almost as good as a d-degeneracy ordering. In addition
to this insight, there are a number of technical details that lead to our results,
which we outline in the remainder of this manuscript.
1 We contacted the authors and they confirmed that their method is indeed incorrect.
3
2 Approximating a d-Degeneracy Ordering
Our method for constructing a (2 + )d-degeneracy ordering for a d-degenerate
graph, G = (V, E), is quite simple and is given below as Algorithm 1. Note
that our algorithm does not take into account the value of d, but it assumes we
are given a constant > 0 as part of the input. Also, note that this algorithm
destroys G in the process. If one desires to maintain G for other purposes, then
one should first create a backup copy of G.
1: L ← ∅
2: while G is nonempty do
3:
4:
5:
6: end while
7: return L
S ← n/(2 + ) vertices of smallest degree in G
L ← LS
remove S from G
// append S to the end of L
Algorithm 1: Approximate degeneracy ordering of vertices
Lemma 1. If G is a d-degenerate graph, then Algorithm 1 computes a (2 + )d-
degeneracy ordering of G.
Proof. Observe that any d-degenerate graph with n vertices has at most 2n/c
vertices of degree at least cd. Thus, G has at most 2n/(2 + ) vertices of degree
at least (2 + )d. This means that the n/(2 + ) vertices of smallest degree in G
each have degree at most (2 + )d. Therefore, every element of set S created in
line 3 has at most (2+)d neighbors in (the remaining graph) G. When we add S
to L in line 4, we keep the property that every element of L has at most (2 + )d
neighbors in G that are placed behind it in L. Furthermore, note that, after we
remove vertices in S (and their incident edges) from G in line 5, G is still at most
d-degenerate (every subgraph of a d-degenerate graph is at most d-degenerate);
(cid:117)(cid:116)
hence, an inductive argument applies to the remainder of the algorithm.
Note that, after (cid:100)log(2+)/2(dn)(cid:101) = O(lg n) iterations, we must have processed
all of G and placed all its vertices on L, which is a (2 + )d-degeneracy ordering
for G and that this property holds even though the algorithm does not take the
value of d into account.
The following lemma is proved in the Appendix.
Lemma 2. An iteration of the while loop (lines 3-5) of Algorithm 1 can be
implemented in O(sort(dn)) I/O's in the external-memory model, where n is
the number of vertices in G at the beginning of the iteration.
Thus, we have the following.
Theorem 1. We can compute a (2 + )d-degeneracy ordering of a d-degenerate
graph, G, in O(sort(dn)) I/O's in the external-memory model, without knowing
the value of d in advance.
4
Proof. Since the number of vertices of G decreases by a factor of 2/(2+) in each
iteration, and each iteration uses O(sort(dn)) I/O's, where n is the number of
vertices in G at the beginning of the iteration (by Lemma 2), the total number
of I/O's, I(G), is bounded by
(cid:18)
(cid:16)(cid:0)2/(2 + )(cid:1)dn
(cid:17)
(cid:17)2
(cid:16) 2
2
+
2 +
2 +
(cid:16)(cid:0)2/(2 + )(cid:1)2
(cid:19)(cid:19)
(cid:17)
+ ···(cid:17)
dn
+ sort
+ ···
(cid:16)
(cid:18)
I(G) = O
sort(dn) + sort
= O
sort(dn)
1 +
= O(sort(dn)).
(cid:117)(cid:116)
This theorem hints at the possibility of effectively using a (2+)d-degeneracy
ordering in place of a d-degeneracy ordering in external-memory algorithms for
naturally sparse graphs. As we show in the remainder of this paper, achieving
this goal is indeed possible, albeit with some additional alterations from previous
internal-memory algorithms.
3 Short Paths and Cycles
In this section, we present external-memory algorithms for finding short cycles in
directed or undirected graphs. Our approach is an external-memory adaptation
of internal-memory algorithms by Alon et al. [3]. We begin with the definition
and an example of a representative due to Monien [22]. A p-set is a set of size p.
Definition 1 (representative). Let F be a collection of p-sets. A sub-collection
(cid:98)F ⊆ F is a q-representative for F, if for every q-set B, there exists a set A ∈ F
such that A ∩ B = ∅ if and only if there exists a set (cid:98)A ∈ (cid:98)F with this property.
Every collection of p-sets F has a q-representative (cid:98)F of size at most (cid:0)p+q
(cid:1)
Monien [22] gives a construction of representatives of size at most O((cid:80)q
(from Bollob´as [7]). An optimal representative, however, seems difficult to find.
i=1 pi).
It uses a p-ary tree of height ≤ q with the following properties.
-- Each node is labeled with either a set A ∈ F or a special symbol λ.
-- If a node is labeled with a set A and its depth is less than q, it has exactly p
children, edges to which are labeled with elements from A (one element per
edge, every element of A is used to label exactly one edge).
p
-- If a node is labeled with λ or has depth q, it has no children.
-- Let E(v) denote the set of all edge labels on the way from the vertex v to
the root of the tree. Then, for every v:
-- if v is labeled with A, then A ∩ E(v) = ∅
-- if v is labeled with λ, then there are no A ∈ F s.t. A ∩ E(v) = ∅.
Monien shows that if a tree T fulfills the above conditions, defining (cid:98)F to be
{4, 7},{4, 8}}. Fig. 1 presents (cid:98)F, a 3-representative of F in the tree form.
the set of all labels of the tree's nodes yields a q-representative for F. As an exam-
ple, consider a collection of 2-sets, F = {{2, 4},{1, 5},{1, 6},{1, 7},{3, 6},{3, 8},
5
{1, 6}
1
6
{3, 8}
8
3
{4, 7}
4
7
{2, 4}
{4, 8}
{1, 7}
{1, 5}
4
{3, 6}
2
{4, 7}
8
{2, 4}
4
λ
7
{1, 5}
1
{3, 8}
1
{2, 4}
5
{3, 8}
Fig. 1. Tree representation of (cid:98)F
The main benefit of using representatives in the tree form stems from the fact
that their sizes are bounded by a function of only p and q (i.e., maximum size of
a representative does not depend on F). It gives a way of storing paths of given
length between two vertices of a graph in a space-efficient way (see Appendix
for details).
The algorithm for finding a cycle of given length has two stages. In the first
stage, vertices of high degree are processed to determine if any of them belongs
to a cycle. This is realized using algorithm cycleThrough from Lemma 5. Since
there are not many vertices of high degree, this can be realized efficiently.
In the second stage, we remove vertices of high degree from the graph. Then,
we group all simple paths that are half the cycle length long by their endpoints
and compute representatives for every such set (see Lemma 3). For each pair of
vertices (u, v), we determine (using findDisjoint from Lemma 4) if there are
two paths: p from u to v and p(cid:48) from v to u, such that p and p(cid:48) do not share any
internal vertices. If this is the case, C = p ∪ p(cid:48) is a cycle of required length.
The following representatives-related lemmas are proved in the Appendix.
Lemma 3. We can compute a q-representative (cid:98)F for a collection of p-sets F,
of size (cid:98)F ≤(cid:80)q
i=1 pi, in O
(cid:16)(cid:0)(cid:80)q+1
i=1 pi(cid:1) · scan(cid:0)F(cid:1)(cid:17)
(cid:16)(cid:0)(cid:80)q+3
i=1 pi +(cid:80)p+3
I/O's.
Lemma 4. For a collection of p-sets, F, and a collection of q-sets, G, there is
an external-memory method, findDisjoint(F,G), that returns a pair of sets
(A, B) (A ∈ F, B ∈ G) s.t. A ∩ B = ∅ or returns if there are no such pairs of
sets. findDisjoint uses O
i=1 qi(cid:1) · scan(cid:0)F + G(cid:1)(cid:17)
I/O's.
Lemma 5. Let G = (V, E). A cycle of length exactly k that passes through
arbitrary v ∈ V , if it exists, can be found by an external-memory algorithm
cycleThrough(G, k, v) in O(cid:0)(k − 1)! · scan(m)(cid:1) I/O's, where m = E, via the
use of representatives.
6
Before we present our result for naturally sparse graphs, we first give an
external-memory method for general graphs.
c ∈ {2k − 1, 2k}, and finds such cycle if it exists, that takes O(cid:0)k · sort(m2− 1
Theorem 2. Let G = (V, E) be a directed or an undirected graph. There is an
external-memory algorithm that decides if G contains a cycle of length exactly
k ) +
(2k − 1)! · scan(m2− 1
k )(cid:1) I/O's.
Proof. Algorithm 2 handles the case of general graphs (which are not necessarily
naturally sparse), and cycles of length c = 2k (the case of c = 2k−1 is analogous).
1
k
C ← cycleThrough(G, 2k, v)
if C (cid:54)= then
return C
1: ∆ ← m
2: for all v -- vertex of degree ≥ ∆ do
3:
4:
5:
6:
end if
7: end for
8: remove vertices of degree ≥ ∆ from G
9: generate all directed paths of length k in G
10: sort the paths lexicographically, according to their endpoints
11: group all paths u (cid:32) v into collection of (k − 1)-sets Fuv
12: for all pairs (Fuv,Fvu) do
13:
14:
15:
16:
end if
17: end for
18: return
P ← findDisjoint(Fuv,Fvu)
if P = (A, B) then
return C = A ∪ B
Algorithm 2: Short cycles in general graphs
scan(m2− 1
k vertices of degree at least ∆, and each
k )(cid:1) I/O's.
Since there are at most m/∆ = m1− 1
call to cycleThrough requires O(cid:0)(2k − 1)! · scan(m)(cid:1) I/O's (by Lemma 5), the
first for loop (lines 2-7) takes O(cid:0)m1− 1
k · (2k − 1)! · scan(m)(cid:1) = O(cid:0)(2k − 1)! ·
Algorithm 1, in O(cid:0)sort(m)(cid:1) I/O's. There are at most m∆k−1 = m2− 1
be generated in line 9. It can be done in O(cid:0)k·sort(m2− 1
k )(cid:1) I/O's (see Appendix).
Sorting the paths (line 10) takes O(cid:0)sort(m2− 1
k )(cid:1) I/O's. After that, creating Fuv's
(line 11) requires O(cid:0)scan(m2− 1
k )(cid:1) I/O's.
Removing vertices of high degree in line 8 is realized just like line 5 of
k paths to
The groupF procedure groups Fuv and Fvu together. Assume we store Fuv's
as tuples (u, v, S), for S ∈ Fuv, in a list F . By u ≺ v we denote that u precedes
v in an arbitrary ordering of V . For u ≺ v, tuples (u, v, 1, S) from line 3 mean
that S ∈ Fuv, while tuples (u, v, 2, S) from line 5 mean that S ∈ Fvu. The for
7
loop (lines 1-7) clearly takes O(cid:0)scan(m2− 1
O(cid:0)sort(m2− 1
from Fvu, allowing us to execute line 9 in O(cid:0)scan(m2− 1
k )(cid:1) I/O's. After sorting F (line 8) in
k )(cid:1) I/O's, tuples for sets from Fuv directly precede those for sets
k )(cid:1) I/O's.
else
if u ≺ v then
write (u, v, 1, S) back to F
proc groupF
1: for all (u, v, S) in F do
2:
3:
4:
5:
end if
6:
7: end for
8: sort F lexicographically
9: scan F to determine pairs (Fuv,Fuv)
write (v, u, 2, S) back to F
O
u,v
Based on Lemma 4, the total number of I/O's in calls to findDisjoint in
Algorithm 2, line 13 is
(cid:16)(cid:80)
i=1 (k − 1)i · scan(Fuv + Fvu)(cid:1)(cid:17)
(cid:0)(cid:80)k+2
(cid:16)(cid:0)(cid:80)k+2
u,v scan(cid:0)Fuv + Fvu(cid:1)(cid:17)
i=1 (k − 1)i(cid:1) ·(cid:80)
k )(cid:1)
= O(cid:0)(2k − 1)! · scan(m2− 1
i=1 (k − 1)i = O(cid:0)(k − 1)k+3(cid:1) = O(cid:0)(2k − 1)!(cid:1).
as we set p = q = k − 1 and(cid:80)k+2
Putting it all together, we get that Algorithm 2 runs in O(cid:0)sort(m2− 1
k )(cid:1) total I/O's.
(2k − 1)! · scan(m2− 1
k ) +
(cid:117)(cid:116)
= O
Theorem 3. Let G = (V, E) be a directed or an undirected graph. There is an
external-memory algorithm that, given L -- a δ-degeneracy ordering of G (for
δ = (2 + )d), finds a cycle of length exactly c, or concludes that it does not exist:
k ) + (4k)! · scan(m2− 1
k ) + (4k)! · scan(m2− 1
I/O's if c = 4k − 2
I/O's if c = 4k − 1 or
k ) + (4k)! · scan(m2− 1
I/O's if c = 4k + 1
δ1− 1
(cid:16)
k ·(cid:0)k · sort(m2− 1
(cid:16)
δ ·(cid:0)k · sort(m2− 1
(cid:16)
k ·(cid:0)k · sort(m2− 1
δ1+ 1
(i) in O
(ii) in O
c = 4k
(iii) in O
k )(cid:1)(cid:17)
k )(cid:1)(cid:17)
k )(cid:1)(cid:17)
Proof. We describe the algorithm for the case of directed G, with c = 4k + 1, as
other cases are similar (and a little easier). We assume that δ < m
2k+1 , which
is obviously the case for naturally sparse graphs. Otherwise, running Algorithm
2 on G achieves the advertised complexity.
1
Algorithm 3 is remarkably similar to Algorithm 2 and so is its analysis.
Differences lie in the value of ∆ and in line 9, when only some paths of length
2k and 2k + 1 are generated. As explained in [3], it suffices to only consider
8
1
k
k /δ1+ 1
1: ∆ ← m
2: for all v -- vertex of degree ≥ ∆ do
C ← cycleThrough(G, 4k + 1, v)
3:
if C (cid:54)= then
4:
5:
return C
end if
6:
7: end for
8: remove vertices of degree ≥ ∆ from G
9: generate directed paths of length 2k and 2k + 1 in G
10: sort the paths lexicographically, according to their endpoints
11: group all paths u (cid:32) v of length 2k into collection of (2k − 1)-sets Fuv
12: group all paths u (cid:32) v of length 2k + 1 into collection of (2k)-sets Guv
13: for all pairs (Fuv,Guv) do
14:
15:
16:
17:
end if
18: end for
19: return
P ← findDisjoint(Fuv,Gvu)
if P = (A, B) then
return C = A ∪ B
Algorithm 3: Short cycles in degenerate graphs
all (2k + 1)-paths that start with two backward-oriented (in L) edges and all
2k-paths that start with a backward-oriented (in L) edge. The number of these
paths is O(m2− 1
k ). Since we can generate them in O
I/O's (see Appendix), and there are at most O(m1− 1
degree ≥ ∆, the theorem follows.
k ) vertices in G of
(cid:117)(cid:116)
k · sort(m2− 1
k )
k δ1+ 1
k δ1+ 1
(cid:17)
(cid:16)
kδ1+ 1
4 All Maximal Cliques
The Bron-Kerbosch algorithm [8] is often the choice when one needs to list all
maximal cliques of an undirected graph G = (V, E). It was initially improved
by Tomita et al. [26]. We present this improvement as the BronKerboschPivot
procedure (Γ (v) denotes the set of neighbors of vertex v).
output R
//maximal clique
proc BronKerboschPivot(P , R, X)
1: if P ∪ X = ∅ then
2:
3: end if
4: u ← vertex from P ∪ X that maximizes P ∩ Γ (u)
5: for all v ∈ P \ Γ (v) do
BronKerboschPivot(P ∩ Γ (v), R ∪ {v}, X ∩ Γ (v))
6:
P ← P \ {v}
7:
8: X ← X ∪ {v}
9: end for
9
The meaning of the arguments to BronKerboschPivot: R is a (possibly non-
maximal) clique, P and X are a division of the set of vertices that are neighbors
of all vertices in R, s.t. vertices in P are to be considered for adding to R while
vertices in X are restricted from the inclusion.
Whereas Tomita et al. run the algorithm as BronKerboschPivot(V ,∅,∅),
Eppstein et al. [12] improved it even further for the case of a d-degenerate G
by utilizing its d-degeneracy ordering L = {v1, v2, . . . , vn} and by performing n
independent calls to BronKerboschPivot. Algorithm 4 presents their version. It
runs in time O(dn3d/3) in the RAM model.
1: for i ← 1 . . . n do
P ← Γ (vi) ∩ {vj : j > i}
2:
3: X ← Γ (vi) ∩ {vj : j < i}
BronKerboschPivot(P ,{vi},X)
4:
5: end for
Algorithm 4: Maximal cliques in degenerate graph
The idea behind Algorithm 4 is to limit the depth of recursive calls to P ≤ d
and then apply the analysis of Tomita et al. [26].
We show how to efficiently implement Algorithm 4 in the external memory
model using a (2 + )d-degeneracy ordering of G. Following [12], we define
subgraphs HP,X of G.
Definition 2 (Graphs HP,X ). Subgraph HP,X = (VP,X , EP,X ) of G = (V, E)
is defined as follows:
VP,X = P ∪ X
EP,X = {(u, v) : (u, v) ∈ E ∧ (u ∈ P ∨ v ∈ P )}
That is, HP,X contains all edges in G whose endpoints are from P ∪ X,
and at least one of them lies in P . To ensure efficiency, HP,X is passed as an
additional argument to every call to BronKerboschPivot with P and X. It is
used in determining u at line 4 of BronKerboschPivot (we simply choose a
vertex of highest degree in HP,X ).
The following two lemmas regarding construction of HP,X 's are proved in
the Appendix.
Lemma 6. Given a δ-degeneracy ordering L of an undirected d-degenerate graph
G (δ = (2 + )d), all initial sets P , X, and graphs HP,X that are passed to
BronKerboschPivot in line 4 of Algorithm 4 can be generated in O(sort(δ2n))
I/O's.
Lemma 7. Given a δ-degeneracy ordering L of an undirected d-degenerate graph
G (δ = (2 + )d), in a call to BronKerboschPivot that was given HP,X , with
P = p and X = x, all graphs HP∩Γ (v),X∩Γ (v) that have to be passed to
recursive calls in line 6, can be formed in O(sort(δp2(p + x))) I/O's.
10
Theorem 4. Given a δ-degeneracy ordering L of an undirected d-degenerate
graph G (δ = (2 + )d), we can list all its maximal cliques in O(3δ/3sort(δn))
I/O's.
Proof. Consider a call to BronKerboschPivot(Pv, {v}, Xv), with Pv = p
and Xv = x. Define (cid:98)D(p, x) to be the maximum number of I/O's in this call.
Based on Lemma 7, (cid:98)D(p, x) satisfies the following recurrence relation:
(cid:98)D(p, x) ≤
(cid:40)
(cid:26) maxk{k(cid:98)D(p − k, x)} + O(cid:0)sort(δp2(p + x))(cid:1)
maxk{k(cid:98)D(p − k, x)} + c · δp2(p+x)
for constant e greater than zero, which can be rewritten as
(cid:98)D(p, x) ≤
if p > 0
if p = 0
logM/B(δ3n) = O(logM/B n) for δ = O(1). Thus, the relation for (cid:98)D(p, x):
for a constant c > 0. Since p ≤ δ and p+x ≤ n, we have logM/B(δp2(p+x)) ≤
(cid:98)D(p, x) ≤
maxk{k(cid:98)D(p − k, x)} + δp2(p + x) · c(cid:48) logM/B n
logM/B(δp2(p + x))
(cid:40)
if p > 0
if p = 0
DB
DB
e
e
if p > 0
if p = 0
e
where c(cid:48) and e are constants greater than zero. Note that this is the relation for
and c2 = e). Since
D(p, x) of Eppstein et. al [12] (we set d = δ, c1 =
the solution for D(p, x) was D(p, x) = O((d + x)3p/3), the solution for (cid:98)D(p, x) is
c(cid:48) logM/B n
DB
(cid:17)
(cid:98)D(p, x) = O
(cid:16)
(δ + x)3p/3 · c(cid:48) logM/B n
DB
= O
3p/3 logM/B n
(cid:17)
(cid:16) δ+x
DB
(cid:16) δ+Xv
The total size of all sets Xv passed to initial calls to BronKerboschPivot is
O(δn), and every set P has at most δ vertices. It follows that the total number
of I/O's in recursive calls is
(cid:88)
= O(cid:0)3δ/3sort(δn)(cid:1)
Algorithm 4 takes O(cid:0)sort(δ2n) + 3δ/3sort(δn)(cid:1) = O(cid:0)3δ/3sort(δn)(cid:1) I/O's.
Combining this with Lemma 6, we get that our external memory version of
(cid:117)(cid:116)
3δ/3 logM/B n
logM/B n
3δ/3 δn
(cid:17)
(cid:16)
(cid:17)
= O
DB
DB
O
v
References
1. N. Alon and S. Gutner. Linear time algorithms for finding a dominating set of
fixed size in degenerated graphs. Algorithmica, 54(4):544 -- 556, 2009.
2. N. Alon, J. Kahn, and P. D. Seymour. Large induced degenerate subgraphs. Graphs
and Combinatorics, 3:203 -- 211, 1987.
3. N. Alon, R. Yuster, and U. Zwick. Finding and counting given length cycles.
Algorithmica, 17(3):209 -- 223, 1997.
11
4. A.-L. Barab´asi and R. Albert. Emergence of scaling in random networks. Science,
286(5439):509 -- 512, 1999.
5. V. Batagelj and M. Zaversnik. An O(m) algorithm for cores decomposition of
networks, 2003. http://arxiv.org/abs/cs.DS/0310049.
6. R. Bauer, M. Krug, and D. Wagner. Enumerating and generating labeled k-
degenerate graphs. In 7th Workshop on Analytic Algorithmics and Combinatorics
(ANALCO), pages 90 -- 98. SIAM, 2010.
7. B. Bollob´as. On generalized graphs. Acta Mathematica Hungarica, 16:447 -- 452,
1965. 10.1007/BF01904851.
8. C. Bron and J. Kerbosch. Algorithm 457: finding all cliques of an undirected graph.
Commun. ACM, 16(9):575 -- 577, 1973.
9. J. Cheng, Y. Ke, S. Chu, and T. Ozsu. Efficient core decomposition in massive
networks. In IEEE Int. Conf. on Data Engineering (ICDE), 2011.
10. M. Chrobak and D. Eppstein. Planar orientations with low out-degree and
compaction of adjacency matrices. Theor. Comput. Sci., 86(2):243 -- 266, 1991.
11. P. Doreian and K. L. Woodard. Defining and locating cores and boundaries of
social networks. Social Networks, 16(4):267 -- 293, 1994.
12. D. Eppstein, M. Loffler, and D. Strash. Listing all maximal cliques in sparse graphs
in near-optimal time. In O. Cheong, K.-Y. Chwa, and K. Park, editors, ISAAC
2010, volume 6506 of LNCS, pages 403 -- 414. Springer-Verlag, 2010.
13. D. Eppstein and D. Strash. Listing all maximal cliques in large sparse real-world
graphs. arXiv eprint, 1103.0318, 2011.
14. P. Erdos and A. Hajnal. On chromatic number of graphs and set-systems. Acta
Mathematica Hungarica, 17(1 -- 2):61 -- 99, 1966.
15. D. Fernholz and V. Ramachandran. The giant k-core of a random graph with a
specified degree sequence. manuscript, 2003.
16. E. C. Freuder. A sufficient condition for backtrack-free search. J. ACM, 29:24 -- 32,
January 1982.
17. P. A. Golovach and Y. Villanger. Parameterized complexity for domination
problems on degenerate graphs. Proc. 34th Int. Worksh. Graph-Theoretic Concepts
in Computer Science (WG 2008), 5344:195 -- 205, 2008.
18. S. Irani. Coloring inductive graphs on-line. Algorithmica, 11:53 -- 72, 1994.
19. L. M. Kirousis and D. M. Thilikos. The linkage of a graph. SIAM Journal on
Computing, 25(3):626 -- 647, 1996.
20. J. Kleinberg. The small-world phenomenon: an algorithm perspective. In 32nd
ACM Symp. on Theory of Computing (STOC), pages 163 -- 170, 2000.
21. D. R. Lick and A. T. White. k-degenerate graphs. Canadian Journal of Mathe-
matics, 22:1082 -- 1096, 1970.
22. B. Monien. How to find long paths efficiently. Annals of Discrete Mathematics,
25:239 -- 254, 1985.
23. B. Pittel, J. Spencer, and N. Wormald. Sudden emergence of a giant k-core in a
random graph. Journal of Combinatorial Theory, Series B, 67(1):111 -- 151, 1996.
24. O. Riordan. The k-core and branching processes. Probability And Computing,
17:111, 2008.
25. S. B. Seidman. Network structure and minimum degree. Social Networks, 5(3):269 --
287, 1983.
26. E. Tomita, A. Tanaka, and H. Takahashi. The worst-case time complexity for
generating all maximal cliques and computational experiments. Theor. Comput.
Sci., 363(1):28 -- 42, 2006.
27. J. S. Vitter. External memory algorithms and data structures: dealing with massive
data. ACM Comput. Surv., 33:209 -- 271, June 2001.
12
A Appendix
A.1 Proof of Lemma 2
The input is a d-degenerate graph G = (V, E). Let V = (1, . . . , n). We store E
as set of edges (u, v). We assume an order on V that we utilize during sorting.
Finding vertices of smallest degree (Algorithm 1, line 3) is realized by the
smallVertices procedure.
d(v) ← degree of v
store pair (d(v), v) in F
proc smallVertices
1: for all v -- vertex of G do
2:
3:
4: end for
5: sort F lexicographically
6: S ← first n/(2 + ) vertices in F
// of smallest degree
Computing degrees of vertices in the for loop (lines 1-4) is easily realized in
O(sort(dn)) I/O's. First, E is sorted lexicographically in O(sort(dn)) I/O's.
After that, edges of E form blocks ordered by their starting vertex, so a simple
scan taking O(scan(dn)) I/O's is enough to determine degrees of the vertices.
Sorting F (line 5) is done in O(sort(n)) I/O's. After that, S is just n/(2 + )
first items of F and its construction (line 6) takes O(scan(n/(2 + ))) I/O's.
Likewise, appending S to L (Algorithm 1, line 4) takes O(scan(n/(2 + )))
I/O's.
Finally, removing edges adjacent to S from G (Algorithm 1, line 5) is realized
as follows.
add tuples (u, v, "-") and (v, u, "-") to E
if u ∈ S then
1: sort E lexicographically
2: for all (u, v) -- edge in E do
3:
4:
end if
5:
6: end for
7: sort E lexicographically
8: for all p, q -- consecutive tuples in E do
9:
10:
11:
12:
13:
14:
end if
15:
16: end for
if p = (u, v) and q = (u, v, "-") then
do not write p back to E
else
//p = (u, v, "-")
do not write p back to E
else if p = (u, v) then
write p back to E
13
Sorting E in line 1 takes O(sort(dn)) I/O's. The first for loop (lines 2-6) takes
O(scan(dn)) I/O's (vertices in S are stored according to the order on V , in the
same relative order as the origins of edges in E, so it is realized by a single
synchronized scan going through E and S at the same time). Each edge in E
causes at most 2 tuples to be added to E in line 4, so clearly the size of E is
O(dn) after line 6. Sorting E (line 7) obviously takes time O(sort(dn)).
The last for loop (lines 8-16) is easily realized by a single scan of E, in
O(scan(dn)) I/O's. Correctness follows from two facts. First, for each edge
(u, v), if u ∈ S, the tuple (u, v, "-") (meaning that (u, v) does not belong to
G after this iteration) is added to E in line 4. Second, if (u, v, "-") is in E, its
direct predecessor is (u, v) (or another copy of (u, v, "-")) after E was sorted
lexicographically in line 7. Therefore, the edges that are no longer in G are
rejected in line 10. Also, no tuples (u, v, "-") are further stored in E (line 14).
Altogether, lines 3-5 of Algorithm 1 are implemented in O(sort(dn)) I/O's.(cid:117)(cid:116)
A.2 Representatives
Monien [22] gave a simple algorithm (which we call repQuery) that operates on
representatives in the tree form described in Sec. 3. Given (cid:98)F -- a representative
for F, repQuery((cid:98)F , B) decides for a q-set B whether there exists a set A ∈ F
s.t. A ∩ B = ∅, and returns such A if it exists. The running time of repQuery is
O(pq) in the RAM model. We use repQuery in our algorithms "as is", i.e., we
allow it to take O(pq) I/O's.
i=1 pi
(number of nodes in a p-ary tree of height q). We can afford to build the tree one
Proof of Lemma 3 The size of the resulting tree is is bounded by (cid:80)q
node at a time, spending O(cid:0)scan(pF)(cid:1) I/O's on each node. Procedure repLabel
labels vertex v in the representative tree for F in O(cid:0)scan(pF)(cid:1) I/O's.
1: (cid:98)A ← set A in F s.t. A ∩ E(v) = ∅
2: if (cid:98)A (cid:54)= then
label v with (cid:98)A
label edges from v to its children with elements from (cid:98)A
proc repLabel(F , v)
create children of v
3:
4:
5:
6: else
7:
8: end if
label v with λ
Set (cid:98)A in line 1 can simply be found by scanning F in O(cid:0)scan(pF)(cid:1) I/O's (pF
is the total size of all p-sets in F). Creating children of v (line 4) and labeling
their edges (line 5) takes O(p) I/O's.
14
O
= O
Proof of Lemma 4 The proof is essentially the same as that of Lemma 3.2
Therefore, we compute a q-representative for F in O
i=1 pi+1(cid:1) · scan(cid:0)F(cid:1)(cid:17)
(cid:16)(cid:0)(cid:80)q
(cid:16)(cid:0)(cid:80)q
i=1 pi(cid:1)·scan(cid:0)pF(cid:1)(cid:17)
i=1 pi(cid:1) · scan(cid:0)F(cid:1)(cid:17)
(cid:16)(cid:0)(cid:80)q+1
(cid:16)(cid:0)(cid:80)q+1
(cid:1) · scan(cid:0)F(cid:1)(cid:17)
in [3]: first we compute a q-representative (cid:98)F of F, in O
(cid:1) · scan(cid:0)G(cid:1)(cid:17)
(cid:16)(cid:0)(cid:80)p+1
I/O's, and a p-representative (cid:98)G of G, in O
The sizes of (cid:98)F and (cid:98)G are bounded by (cid:80)q
i=1 pi and (cid:80)p
Assuming p ≥ q (w.l.o.g.), determining whether (cid:98)F and (cid:98)G contain two disjoint
i=1 pi(cid:1)
i=1 pi+2(cid:1) = O(cid:0)(cid:80)q+3
i=1 pi · pq(cid:1) = O(cid:0)(cid:80)q+1
sets can be easily done in O(cid:0)(cid:80)q+1
I/O's, by querying (cid:98)G (via repQuery) with all sets from (cid:98)F.
i=1 qi, respectively.
I/O's.
i=1
I/O's.
i=1
=
(cid:117)(cid:116)
(cid:117)(cid:116)
Proof of Lemma 5 The (very) big picture of the cycleThrough(G, k, v)
algorithm is as follows:
if there exists simple path p: u (cid:32) v of length exactly k − 1 then
proc cycleThrough(G, k, v)
1: for all u s.t. (v, u) ∈ E do
2:
3:
end if
4:
5: end for
6: return
return p ∪ (v, u)
Let P p
Obviously, main difficulty lies in checking the condition in line 2. To show how
we answer that query, let us first explain how [22] handles the paths.
uv denote the set of all simple paths from u to v of length exactly
(p + 1) (so that these paths have p inner vertices). First, all paths from u to v of
length (p + 1) that have the exact same set of inner vertices are represented as a
single set containing these vertices (the vertices are stored as one of the paths; it
provides a representative of the set). Performing this compression on P p
uv yields
F p
uv -- a family of p-sets:
F p
uv = {S : S is a set of inner vertices on some path from u to v of length p + 1}
The condition from line 2 of cycleThrough is therefore equivalent to F k−2
uv being
nonempty. We will now focus on how to test if this is the case.
uv (for all u ∈ V )
The clou of [22] was that having q-representatives for F p
(for all u ∈ V ).
enables efficient computation of (q − 1)-representatives for F p+1
The labels for a (q− 1)-representative tree for F p+1
uv are computed node by node.
The algorithm is based on the following observation (γ is the node whose label
uv
15
we compute, E(γ) is the set of edge labels on the way from γ to root):
∃U ∈ F p+1
uv
s.t. U ∩ E(γ) = ∅
⇐⇒
∃(cid:98)U ∈ F p
∃w ∈ V \ {u, v} s.t. (u, w) ∈ E ∧ w /∈ E(γ) ∧
wv s.t. (cid:98)U ∩ (E(γ) ∪ {u}) = ∅
wv allows us to find (cid:98)U (or determine that it does
Having a q-representative for F p
not exist) via the repQuery algorithm. Determining the label for γ is therefore
realized as follows:
wv, E(γ) ∪ {u})
(cid:98)U ← repQuery(F p
if (cid:98)U (cid:54)= then
label γ with (cid:98)U ∪ {w}
1: for all w s.t. (u, w) ∈ E and w /∈ E(γ) do
2:
3:
4:
5:
end if
6:
7: end for
8: label γ with λ
return
= O
repQuery (querying a representative tree) takes O(pq) I/O's, so our imple-
mentation of labeling γ takes O(cid:0)pq · scan(cid:0)Γ (u)(cid:1)(cid:1) I/O's (where Γ (u) denotes
size bounded by (cid:80)q−1
the number of neighbors of u in G). It simply scans neighbors of u and calls
repQuery accordingly. Because the (q − 1)-representative tree for F p+1
(cid:16)
q(p + 1)q−1 ·(cid:0)pq · scan(cid:0)Γ (u)(cid:1)(cid:1)(cid:17)
(cid:16)
q2(p + 1)q · scan(cid:0)Γ (u)(cid:1)(cid:17)
has
i=1 (p + 1)i ≤ q(p + 1)q+1, labeling all its nodes requires
I/O's. We are
(cid:16)
1)q·scan(cid:0)Γ (u)(cid:1)(cid:1)(cid:17)
u scan(cid:0)Γ (u)(cid:1)(cid:17)
q2(p+1)q·(cid:80)
Our goal is to compute 0-representatives (cid:98)F k−2
by calling repQuery((cid:98)F k−2
(cid:0)q2(p+
= O(cid:0)q2(p+1)q·scan(m)(cid:1)
O
computing (q−1)-representatives for F p+1
either returns (if F k−2
u to v of length k − 1.
uv's. They are built as trees having
only the root vertex, labeled with either ∅ (if (u, v) ∈ E), or λ (otherwise). They
can be obviously constructed in O(scan(m)) I/O's, via scanning E. Based on
We start with (k−2)-representatives for F 0
(for all u ∈ V ). Then,
is nonempty, as it
uv , representing a path from
uv , ∅), we determine whether F k−2
is empty), or a set A ∈ F k−2
uv 's for all u's, so it takes O
the discussion above, computing (cid:98)F k−2
(cid:16)(cid:80)
for F k−2
uv
I/O's in total.
uv
u
= O
uv
uv
uv
uv 's (for all u ∈ V ) takes
p=0(k − 2 − p)(p + 1)k−2−p · scan(m)(cid:1)
O(cid:0)(cid:80)k−3
p=0(p + 1)k−2−p(cid:1)
= O(cid:0)(k − 2) · scan(m) ·(cid:80)k−3
= O(cid:0)(k − 2) · scan(m) · (k − 2)!(cid:1)
= O(cid:0)(k − 1)! · scan(m)(cid:1)
16
I/O's, as it can be easily shown by induction that(cid:80)k−3
for k ≥ 6.
p=0(p + 1)k−2−p ≤ (k − 2)!
A.3 Path Generation
General Graphs (Algorithm 2) Recall that for Algorithm 2 we need to
generate all directed paths of length k in a graph G = (V, E), where maxi-
mum degree of each vertex is bounded by ∆. As shown in [3], there are at
most O(m∆k−1) = O(m2− 1
k ) such paths. They are generated by the pathGen
procedure.
generate sequences es, for all e ∈ E
proc pathGen
1: generate all sequences of length k − 1, with elements from {1, . . . , ∆}
2: for all s -- sequence ∈ {1, . . . , ∆}k−1 do
3:
4: end for
5: for all s -- sequence ∈ E × {1, . . . , ∆}k−1 do
6:
7:
8:
9:
10:
end if
11:
12: end for
decode s into s(cid:48) ∈ Ek
if s(cid:48) (cid:54)= then
output s(cid:48)
if s(cid:48) is a simple path then
end if
k )(cid:1) I/O's.
k )(cid:1) I/O's. Adding
Line 1 is simply realized in O(cid:0)scan(∆k−1)(cid:1) = O(cid:0)scan(m1− 1
O(cid:0)m · scan(∆k−1)(cid:1) = O(cid:0)scan(m2− 1
an edge at the beginning of each sequence in the first for loop (lines 2-4) takes
Decoding a sequence s ∈ E × {1, . . . , ∆}k−1 into a path s(cid:48) ∈ Ek (pathGen,
line 6) is conceptually straightforward. e is the first edge in the path. Then,
each consecutive number i determines next vertex on the path -- ith neighbor of
the previously decoded one (if it has less than i neighbors, the path is dropped
as invalid). The decodePaths procedure handles decoding of S -- the set of
sequences. s[i] is the ith element of tuple s, s[i].from is the origin, and
s[i].to is the destination vertex of edge at s[i].
decodes the next vertex of s and again cyclically shifts s by one position. Sorting
The first for loop (lines 1-3) decodes the starting edge of the sequence s into
two vertices, and then cyclically shifts the resulting tuple by one position. It
k )(cid:1) I/O's. Each iteration of the second for loop (lines 4-12)
takes O(cid:0)scan(m2− 1
k )(cid:1) I/O's. After that, the inner loop (lines 6-11)
S in line 5 takes O(cid:0)sort(m2− 1
k )(cid:1) I/O's (it takes one synchronized scan of S and
requires only O(cid:0)scan(m2− 1
outer for loop runs for k − 1 iterations, decodePaths uses O(cid:0)k · sort(m2− 1
k )(cid:1)
E). Invalid paths that do not meet the condition at line 8 are dropped. Since the
I/O's.
17
write tuple (s[1].to, s[2], s[3], . . . , s[k], s[1].from) to S
// S contains tuples V × {1, . . . , ∆}k−i × V i
proc decodePaths(S)
1: for all s -- sequence ∈ E × {1, . . . , ∆}k−1 do
2:
3: end for
4: for all i ← 1, . . . , k − 1 do
sort S lexicographically
5:
for all s -- tuple in S do
6:
7:
8:
9:
10:
11:
12: end for
u ← s[2]th neighbor of s[1] in V
if u (cid:54)= then
end if
end for
write tuple (u, s[3], s[4], . . . , s[k+1], s[1]) back to S
does not contain repeated vertices. Thus, pathGen takes O(cid:0)sort(m2− 1
Verifying that a path is simple (pathGen, line 8) is done by checking that it
k )(cid:1) I/O's.
Degenerate Graphs (Algorithm 3) For Algorithm 3, we need to generate
all paths of length 2k + 1 that start with two backward-oriented (in L) edges. As
shown in [3], there are at most O(cid:0)m(cid:80)k
(cid:1)∆iδ2k−i(cid:1) = O(22km∆kδk) paths
(cid:0)2k
of length 2k + 1 in G. It follows from the fact that for each path p, either p or
pR (the reverse of p) has at most k edges with opposite directions than in L.
i=0
i
The procedure pathGenForward generates paths p that have at most k edges
with opposite directions than in L.
for all s -- sequence ∈ V 2 × ({"L", "E"} × {1, . . . , ∆})i−1 do
proc pathGenForward
1: generate sequences s = uv, for all (u, v) ∈ E
2: for i ← 1 . . . 2k do
3:
4:
5:
6:
7:
8:
9: end for
generate all sequences s("L", j), for j ∈ {1, . . . , δ}
if s has < k pairs of the type ("E", j) then
10: for all s -- sequence ∈ V 2 ×(cid:0)(cid:8)"L", "E"(cid:9) ×(cid:8)1, . . . , max{δ, ∆}(cid:9)(cid:1)2k
generate all sequences s("E", j), for j ∈ {1, . . . , ∆}
end if
end for
decode s into s(cid:48) ∈ E2k+1
if s(cid:48) (cid:54)= then
output s(cid:48)
if s(cid:48) is a simple path then
11:
12:
13:
14:
15:
end if
16:
17: end for
end if
do
18
The encoding of the sequences works as follows: it starts with two vertices, u
and v, that represent the starting edge of the path. v is followed by 2k pairs of
the format ("L", i) (with i ∈ {1, . . . , δ}) or ("E", i) (with i ∈ {1, . . . , ∆}). ("L", i)
means that the next vertex is the ith neighbor of the current vertex in degeneracy
ordering L, and ("E", i) means that the next vertex is the ith neighbor of the
current vertex in E. Any sequence s has at most k pairs of the type ("E", j),
and only edges represented by them may have the opposite direction than in L.
The number of sequences generated by the first for loop (lines 2-9) is clearly
O(cid:0)m(cid:80)k
(cid:1)∆iδ2k−i(cid:1) = O(22km∆kδk), and the whole generation process
takes O(cid:0)scan(22km∆kδk)(cid:1) I/O's. The sequences are decoded into paths in the
second for loop (lines 10-17) in a manner similar to decodePaths, using O(cid:0)(2k+
(cid:0)2k
i=0
i
1) · sort(22km∆kδk) I/O's.
Paths of length 2k + 1, that have at most k edges in the opposite direction
than in L when they are read backwards, can be generated by an analogous
procedure pathGenBackward (using ER -- reversed edges instead of E), with the
same I/O complexity.
Procedure pathGen2 generates all paths of length 2k + 1 that start with two
backward-oriented (in L) edges.
generate all sequences ijs, for (i, j) ∈ {1, . . . , δ}2
proc pathGen2
1: generate paths of length 2k − 1, via pathGenForward
2: generate paths of length 2k − 1, via pathGenBackward
3: for all s -- generated path of length 2k − 1 do
4:
5: end for
6: for all s -- sequence ∈ {1, . . . , d}2 × E2k−1 do
7:
8:
9:
10:
11:
end if
12:
13: end for
decode s into s(cid:48) ∈ E2k+1
if s(cid:48) (cid:54)= then
output s(cid:48)
if s(cid:48) is a simple path then
end if
Paths generated in lines 1-2 are the tails of the resulting paths. The two
numbers in sequences added to these paths in line 4 denote the edges in L that
are to be taken to determine first two vertices on the final path, starting at the
tail's first vertex. This assures that these edges are backward-oriented in L.
The number of paths generated by pathGen2 is O(δ2 · 22k−2m∆k−1δk−1) =
k ), and its I/O complexity is O(cid:0)k ·
k δ1+ 1
O(22k−2m∆k−1δk+1) = O(22k−2m2− 1
sort(22k−2m2− 1
k δ1+ 1
k )(cid:1).
Since generating all paths of length 2k that begin with a backward-oriented
(in L) edge is essentially the same as pathGen2, path generation in this case
requires O(cid:0)k · sort(22k−2m2− 1
k δ1+ 1
k )(cid:1) I/O's.
19
A.4 Graph Reordering
Assume we are given a graph G = (V, E), with V = (1, 2, . . . , n), and d-
degeneracy ordering L = (v1, v2, . . . , vn) of V . Our goal is to reorder G according
to L, i.e., substitute each edge (vi, vj) ∈ E with an edge (i, j).
for all vi -- ith vertex in L do
append tuple (vi, "i") to E
// precedes p
do not write (v, i) back to E
end for
sort E lexicographically
for all p -- tuple in E do
if p = (u, v) then
q ← tuple (u, "i")
write (v, i) back to E
proc reorderG(G, L)
1: for all k ← 1, 2 do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: end for
end if
end for
else
A single iteration of the outer for loop (lines 1-14) first renames origins of edges
in E and then reverts them (thus, after 2 iterations edges have their original
directions). First, it adds vertices along with their positions in L to E in lines
2-4. This takes O(scan(n)) I/O's. Then it sorts E (line 5) in O(sort(dn)) I/O's.
The next for loop (lines 6-13) scans E, renames origins of edges to their positions
in L and outputs their opposite versions. The tuples q obtained in line 8 can
clearly be found in the same single scan (tuple (u, "i") directly precedes all edges
(u, v) after E was sorted), so this part is done in O(scan(dn)) I/O's.
Therefore, reorderG runs in O(sort(dn)) I/O's altogether.
A.5 Proof of Lemma 6
First, observe that the total size of all sets P and X passed to initial calls to
BronKerboschPivot is O(m) = O(δn). To see this, note that every edge (vi, vj)
(with vi preceding vj in L = (v1, v2, . . . , vn)) puts vi into initial X for R = {vj}
and vj into initial P for R = {vi}. Since the size of each P is at most δ, the total
number of edges in all HP,X 's is O(δ2n).
Our approach is to generate and store all P 's, X's and HP,X 's at the very
beginning of the algorithm, and then just pass appropriate P , X, and HP,X to
each initial call to BronKerboschPivot.
Procedure genPX generates all initial P 's and X's in O(sort(δn)) I/O's. We
assume that for each vertex v ∈ V we know its position in L (i.e., we know i for
v = vi). Also, for each edge (u, v) ∈ E, we know positions in L of its endpoints
20
(i.e., we know both i and j for (u, v) = (vi, vj)). We can easily achieve this in
L≺ v, we denote that u precedes v in L.
O(sort(δn)) I/O's (see Sec. A.4). By u
The for loop (lines 1-5) clearly takes O(scan(δn)) I/O's.
proc genPX(G, L)
1: for all (vi, vj) -- edge in E do
L≺ vj then
if vi
2:
3:
end if
4:
5: end for
6: sort F lexicographically
7: scan F to create sets P and X
output tuples (vi, "P", vj), and (vj, "X", vi) to set F
The meaning of (vi, "P", vj) is "add vj to Pvi", and of (vj, "X", vi): "add vi
to Xvj ". The above discussion explains why this information allows to generate
all P 's and X's. Therefore, after set F is sorted in O(sort(δn)) I/O's in line 6,
generation of P 's and X's in line 7 takes O(scan(δn)) I/O's.
Procedure genH generates all initial HP,X 's in O(sort(δ2n)) I/O's. Each ver-
tex has at most δ neighbors that succeed it in L, so there are O(δ2n) tuples added
to E in the first for loop (lines 1-5). They can be generated in O(scan(δ2n))
I/O's, as the edges in E are sorted lexicographically.
L≺ v
L≺ w do
end for
append tuple (v, w, ?u?) to E
for all v, w -- neighbors of u s.t. u
q ← tuple of form (vi, vj) immediately preceding p in E
if q = (v, w) then
proc genH(G, L)
1: for all u -- vertex in V do
2:
3:
4:
5: end for
6: sort E lexicographically
7: for all p -- tuple in E do
if p = (v, w, ?u?) then
8:
9:
10:
11:
12:
13:
14:
15:
end if
16:
17: end for
18: sort H lexicographically
19: scan H to create sets Hv
20: scan edges of Hv's and mark the endpoints that belong to Pv's
end if
do not write p back to E
output tuples (Hv, u, w) and (Hv, w, u) to set H
else
//p = (v, w)
write p back to E
21
To understand the meaning of tuples from line 3, refer to Fig. 2.
u
v
w
Fig. 2. v and w are neighbors of u, and u
in genH, line 3. If (v, w) ∈ E, the edge (u, w) is in Hv.
L≺ v
L≺ w, so the tuple (v, w, ?u?) is output
After E is sorted in line 6 in O(sort(δ2n)) I/O's, the next for loop (lines
7-17) identifies edges that belong to sets Hv. In E, each edge (v, w) is followed
by zero or more tuples of the form (v, w, ?u?). This makes it easy to determine
q in line 9, and therefore, the loop takes O(scan(δ2n)) I/O's.
As explained in Fig. 2, (v, w) followed by (v, w, ?u?) means that the edge
(u, w) has to be added to Hv. It is denoted by the tuples (Hv, u, w) and (Hv, w, u)
output in line 11. Sorting H in line 18 takes O(sort(δ2n)) I/O's, and after that,
HP,X 's are generated in line 19 in O(scan(δ2n)) I/O's. The marking of endpoints
(line 20) also takes O(scan(δ2n)) I/O's.
Therefore, total complexity of generating initial P 's, X's, and HP,X 's is
(cid:117)(cid:116)
O(sort(δ2n)) I/O's.
A.6 Proof of Lemma 7
if e = (u, v) or e = (v, w) then
unmark v in e
for all e -- edge in HP,X do
end if
end for
for all e -- edge in HP,X do
proc updateH(v)
1: for all HP,X -- candidate do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: end for
write e back to HP,X
else
end if
end for
if at least one vertex of e is marked then
do not write e back to HP,X
22
Recall that P = p and X = x. Our idea in computing HP,X 's is to first
generate candidates for HP,X 's. Candidates are defined as HP,X 's as they would
be if there were no lines 7 and 8 in BronKerboschPivot (i.e., as if P and X
did not change). We then update the candidates according to lines 7 and 8 of
BronKerboschPivot.
Generation of the candidates for HP,X 's is almost the same as in genH, only
v and w in line 2 are now taken from P , so it uses O(sort(p2(p + x))) I/O's.
Updating HP,X 's (moving v from P to X) is realized by the procedure updateH.
Both inner for loops (lines 2-6 and 7-13) clearly take O(scan(HP,X)) I/O's.
Since the total size of all candidates is O(p2(p + x)), a single call to updateH
takes O(scan(p2(p + x))) I/O's. There are at most p such calls, so generating
HP,X 's that are passed to recursive calls in line 6 of BronKerboschPivot takes
O(sort(p2(p + x)) + p · scan(p2(p + x))) = O(sort(δp2(p + x))) I/O's.
(cid:117)(cid:116)
23
|
1803.09520 | 3 | 1803 | 2018-09-06T08:08:17 | Universal Compressed Text Indexing | [
"cs.DS"
] | The rise of repetitive datasets has lately generated a lot of interest in compressed self-indexes based on dictionary compression, a rich and heterogeneous family that exploits text repetitions in different ways. For each such compression scheme, several different indexing solutions have been proposed in the last two decades. To date, the fastest indexes for repetitive texts are based on the run-length compressed Burrows-Wheeler transform and on the Compact Directed Acyclic Word Graph. The most space-efficient indexes, on the other hand, are based on the Lempel-Ziv parsing and on grammar compression. Indexes for more universal schemes such as collage systems and macro schemes have not yet been proposed. Very recently, Kempa and Prezza [STOC 2018] showed that all dictionary compressors can be interpreted as approximation algorithms for the smallest string attractor, that is, a set of text positions capturing all distinct substrings. Starting from this observation, in this paper we develop the first universal compressed self-index, that is, the first indexing data structure based on string attractors, which can therefore be built on top of any dictionary-compressed text representation. Let $\gamma$ be the size of a string attractor for a text of length $n$. Our index takes $O(\gamma\log(n/\gamma))$ words of space and supports locating the $occ$ occurrences of any pattern of length $m$ in $O(m\log n + occ\log^{\epsilon}n)$ time, for any constant $\epsilon>0$. This is, in particular, the first index for general macro schemes and collage systems. Our result shows that the relation between indexing and compression is much deeper than what was previously thought: the simple property standing at the core of all dictionary compressors is sufficient to support fast indexed queries. | cs.DS | cs |
Universal Compressed Text Indexing 1
Gonzalo Navarro 2
Center for Biotechnology and Bioengineering (CeBiB),
Department of Computer Science, University of Chile.
[email protected]
Nicola Prezza 3
Department of Computer Science, University of Pisa, Italy.
[email protected]
Abstract
The rise of repetitive datasets has lately generated a lot of interest in compressed
self-indexes based on dictionary compression, a rich and heterogeneous family of
techniques that exploits text repetitions in different ways. For each such compres-
sion scheme, several different indexing solutions have been proposed in the last
two decades. To date, the fastest indexes for repetitive texts are based on the
run-length compressed Burrows -- Wheeler transform (BWT) and on the Compact
Directed Acyclic Word Graph (CDAWG). The most space-efficient indexes, on the
other hand, are based on the Lempel -- Ziv parsing and on grammar compression.
Indexes for more universal schemes such as collage systems and macro schemes
have not yet been proposed. Very recently, Kempa and Prezza [STOC 2018] showed
that all dictionary compressors can be interpreted as approximation algorithms for
the smallest string attractor, that is, a set of text positions capturing all distinct
substrings. Starting from this observation, in this paper we develop the first univer-
sal compressed self-index, that is, the first indexing data structure based on string
attractors, which can therefore be built on top of any dictionary-compressed text
representation. Let γ be the size of a string attractor for a text of length n. From
known reductions, γ can be chosen to be asymptotically equal to any repetitiveness
measure: number of runs in the BWT, size of the CDAWG, number of Lempel -- Ziv
phrases, number of rules in a grammar or collage system, size of a macro scheme.
Our index takes O(γ lg(n/γ)) words of space and supports locating the occ occur-
rences of any pattern of length m in O(m lg n + occ lg n) time, for any constant
> 0. This is, in particular, the first index for general macro schemes and collage
systems. Our result shows that the relation between indexing and compression is
much deeper than what was previously thought: the simple property standing at
the core of all dictionary compressors is sufficient to support fast indexed queries.
Key words: Repetitive sequences; Compressed indexes; String attractors
Preprint submitted to Elsevier Preprint
7 September 2018
1
Introduction
Efficiently indexing repetitive text collections is becoming of great importance
due to the accelerating rate at which repetitive datasets are being produced in
domains such as biology (where the number of sequenced individual genomes
is increasing at an accelerating pace) and the web (with databases such as
Wikipedia and GitHub being updated daily by thousands of users). A self-
index on a string S is a data structure that offers direct access to any substring
of S (and thus it replaces S), and at the same time supports indexed queries
such as counting and locating pattern occurrences in S. Unfortunately, classic
self-indexes -- for example, the FM-index [20] -- that work extremely well on
standard datasets fail on repetitive collections in the sense that their compres-
sion rate does not reflect the input's information content. This phenomenon
can be explained in theory with the fact that entropy compression is not able
to take advantage of repetitions longer than (roughly) the logarithm of the
input's length [22]. For this reason, research in the last two decades focused on
self-indexing based on dictionary compressors such as the Lempel -- Ziv 1977
factorization (LZ77) [39], the run-length encoded Burrows -- Wheeler transform
(RLBWT) [11] and context-free grammars (CFGs) [37], just to name the most
popular ones. The idea underlying these compression techniques is to break
the text into phrases coming from a dictionary (hence the name dictionary
compressors), and to represent each phrase using limited information (typi-
cally, a pointer to other text locations or to an external set of strings). This
scheme allows taking full advantage of long repetitions; as a result, dictionary-
compressed self-indexes can be orders of magnitude more space-efficient than
entropy-compressed ones on repetitive datasets.
The landscape of indexes for repetitive collections reflects that of dictionary
compression strategies, with specific indexes developed for each compression
strategy. Yet, a few main techniques stand at the core of most of the indexes.
To date, the fastest indexes are based on the RLBWT and on the Compact Di-
rected Acyclic Word Graph (CDAWG) [10,18]. These indexes achieve optimal-
time queries (i.e., asymptotically equal to those of suffix trees [50]) at the price
of a space consumption higher than that of other compressed indexes. Namely,
the former index [26] requires O(r lg(n/r)) words of space, r being the number
of equal-letter runs in the BWT of S, while the latter [3] uses O(e) words, e
being the size of the CDAWG of S. These two measures (especially e) have
1 This work started during Shonan Meeting 126 "Computation over Compressed
Structured Data".
2 Partially funded by Fondecyt Grant 1-170048 and Basal Funds FB0001, Chile.
3 This work was partially done while the author was holding a post-doc position
at the Technical University of Denmark (DTU). Partially funded by the project
MIUR-SIR CMACBioSeq ("Combinatorial methods for analysis and compression
of biological sequences") grant n. RBSI146R5L.
2
been experimentally confirmed to be not as small as others -- such as the size
of LZ77 -- on repetitive collections [4].
Better measures of repetitiveness include the size z of the LZ77 factorization
of S, the minimum size g of a CFG (i.e., sum of the lengths of the right-
hand sides of the rules) generating S, or the minimum size grl of a run-length
CFG [44] generating S. Indexes using O(z) or O(g) space do exist, but optimal-
time queries have not yet been achieved within this space. Kreft and Navarro
[38] introduced a self-index based on LZ77 compression, which proved to be
extremely space-efficient on highly repetitive text collections [15]. Their self-
index uses O(z) space and finds all the occ occurrences of a pattern of length m
in time O((m2h+(m+occ) lg z) lg(n/z)), where h ≤ z is the maximum number
of times a symbol is successively copied along the LZ77 parsing. A string of
length (cid:96) is extracted in O(h(cid:96)) time. Similarly, self-indexes of size O(g) building
on grammar compression [16,17] can locate all occ occurrences of a pattern in
O(m2 lg lg n + m lg z + occ lg z) time. Within this space, a string of length (cid:96)
can be extracted in time O(lg n + (cid:96)/ lgσ n) [7]. Alternative strategies based on
Block Trees (BTs) [5] appeared recently. A BT on S uses O(z lg(n/z)) space,
which is also the best asymptotic space obtained with grammar compressors
[13,48,29,30,47]. In exchange for using more space than LZ77 compression,
the BT offers fast extraction of substrings: O((1 + (cid:96)/ lgσ n) lg(n/z)) time. A
self-index based on BTs has recently been described by Navarro [42]. Various
indexes based on combinations of LZ77, CFGs, and RLBWTs have also been
proposed [23,24,4,43,8,14]. Some of their best results are O(z lg(n/z)+z lg lg z)
space with O(m + occ(lg lg n + lg z)) query time [14], and O(z lg(n/z)) space
with either O(m lg m + occ lg lg n) [24] or O(m + lg z + occ(lg lg n + lg z)) [14]
query time. Gagie et al. [26] give a more detailed survey.
The above-discussed compression schemes are the most popular, but not the
most space-efficient. More powerful compressors (NP-complete to optimize)
include macro schemes [49] and collage systems [36]. Not much work exists in
this direction, and no indexes are known for these particular compressors.
1.1 String attractors
As seen in the previous paragraphs, the landscape of self-indexes based on dic-
tionary compression -- as well as that of dictionary compressors themselves --
is extremely fragmented, with several techniques being developed for each dis-
tinct compression strategy. Very recently, Kempa and Prezza [35] gathered all
dictionary compression techniques under a common theory: they showed that
these algorithms are approximations to the smallest string attractor, that is,
a set of text positions "capturing" all distinct substrings of S.
3
Definition 1 (String attractor [35]). A string attractor of a string S[1..n] is
a set of γ positions Γ = {j1, . . . , jγ} such that every substring S[i..j] has an
occurrence S[i(cid:48)..j(cid:48)] = S[i..j] with jk ∈ [i(cid:48), j(cid:48)], for some jk ∈ Γ.
Their main result is a set of reductions from dictionary compressors to string
attractors of asymptotically the same size.
Theorem 1 ([35]). Let S be a string and let α be any of these measures:
(1) the size g of a CFG for S,
(2) the size grl of a run-length CFG for S,
(3) the size c of a collage system for S,
(4) the size z of the LZ77 parse of S,
(5) the size b of a macro scheme for S.
Then, S has a string attractor of size γ = O(α). In all cases, the correspond-
ing attractor can be computed in O(S) time and space from the compressed
representation.
Importantly, this implies that any data structure based on string attractors is
universal : given any dictionary-compressed text representation, we can induce
a string attractor and build the data structure on top of it. Indeed, the au-
thors exploit this observation and provide the first universal data structure for
random access, of size O(γ lg(n/γ)). Their extraction time within this space is
O(lg(n/γ) + (cid:96)/ lgσ n). By using slightly more space, O(γ lg(n/γ) lg n) for any
constant > 0, they obtain time O(lg(n/γ)/ lg lg n + (cid:96)/ lgσ n), which is the
optimal that can be reached using any space in O(γ polylog n) [35]. This sug-
gests that compressed computation can be performed independently from the
compression method used while at the same time matching the lower bounds
of individual compressors (at least for some queries such as random access).
1.2 Our Contributions
In this paper we exploit the above observation and describe the first universal
self-index based on string attractors, that is, the first indexing strategy not
depending on the underlying compression scheme. Since string attractors stand
at the core of the notion of compression, our result shows that the relation
between compression and indexing is much deeper than what was previously
thought: the simple string attractor property introduced in Definition 1 is
sufficient to support indexed pattern searches.
Theorem 2. Let a string S[1..n] have an attractor of size γ. Then, for any
constant > 0, there exists a data structure of size O(γ lg(n/γ)) that, given
a pattern string P [1..m], outputs all the occ occurrences of P in S in time
4
O(m lg n + occ(lg γ + lg lg(n/γ))) = O(m lg n + occ lg n). It can be built in
√
O(n + γ lg(n/γ)
lg γ) worst-case time and O(γ lg(n/γ)) space with a Monte
Carlo method returning a correct result with high probability. A guaranteed
construction, using Las Vegas randomization, takes O(n lg n) expected time
(this time also holds w.h.p.) and O(n) space.
We remark that no representation offering random access within o(γ lg(n/γ))
space is known. The performance of our index is close to that of the fastest
self-indexes built on other repetitiveness masures, and it is the first one that
can be built for macro schemes and collage systems.
To obtain our results, we adapt the block tree index of Navarro [42], which is
designed for block trees on the LZ77 parse, to operate on string attractors. The
result is also different from the block-tree-like structure Kempa and Prezza
use for extraction [35], because that one is aligned with the attractors and
this turns out to be unsuitable for indexing. Instead, we use a block-tree-like
structure, which we dub Γ-tree, which partitions the text in a regular form. We
moreover introduce recent techniques [24] to remove the quadratic dependency
on the pattern length in query times.
1.3 Notation
We denote by S[1..n] = S[1]··· S[n] a string of length n over an alphabet of
size σ = O(n). Substrings of S are denoted S[i..j] = S[i]··· S[j], and they are
called prefixes of S if i = 1 and suffixes of S if j = n. The concatenation of
strings S and S(cid:48) is denoted S · S(cid:48). We assume the RAM model of computation
with a computer word of ω = Ω(lg n) bits. By lg we denote the logarithm
function, to the base 2 when this matters.
The term with high probability (w.h.p. abbreviated) indicates with probability
at least 1 − n−c for an arbitrarily large constant c, where n is the input size
(in our case, the input string length).
print φ of a string S = S[1]··· S[n] ∈ [1..σ]n is defined as φ(S) =(cid:80)n−1
In our results, we make use of a modern variant of Karp -- Rabin fingerprinting
[33] (more common nowadays than the original version), defined as follows. Let
q ≥ σ be a prime number, and r be a uniform number in [1..q− 1]. The finger-
i=0 S[n −
i] · ri mod q. The extended fingerprint of S is the triple φ(S) = (cid:104) φ(S), rS
mod q, r−S mod q(cid:105). We say that S (cid:54)= S(cid:48), with S, S(cid:48) ∈ [1..σ]n, collide through
φ (for our purposes, it will be sufficient to consider equal-length strings) if
φ(S) = φ(S(cid:48)), that is, φ(S − S(cid:48)) = 0, where S(cid:48)(cid:48) = S − S(cid:48) is the string defined
as S(cid:48)(cid:48)[i] = S[i] − S(cid:48)[i] mod q. Since φ(S(cid:48)(cid:48)) is a polynomial (in the variable r)
of degree at most n − 1 in the field Zq, it has at most n − 1 roots. As a con-
sequence, the probability of having a collision between two strings is bounded
5
by O(n/q) when r is uniformly chosen in [1..q − 1]. By choosing q ∈ Θ(nc+2)
for an arbitrarily large constant c, one obtains that such a hash function is
collision-free among all equal-length substrings of a given string S of length
n w.h.p. To conclude, we will exploit the (easily provable) folklore fact that
two extended fingerprints φ(U ) and φ(V ) can be combined in constant time to
obtain the extended fingerprint φ(U V ). Similarly, φ(U V ) and φ(U ) (respec-
tively, φ(V )) can be combined in constant time to obtain φ(V ) (respectively,
φ(U )). From now on, we will by default use the term "fingerprint" to mean
extended fingerprint.
2 Γ-Trees
Given a string S[1..n] over an alphabet [1..σ] with an attractor Γ of size γ, we
define a Γ-tree on S as follows. At the top level, numbered l = 0, we split S
into γ substrings (which we call blocks) of length b0 = n/γ. Each block is then
recursively split into two so that if bl is the length of the blocks at level l, then
it holds that bl+1 = bl/2, until reaching blocks of one symbol after lg(n/γ)
levels (that is, bl = n/(γ · 2l)). 4 At each level l, every block that is at distance
< bl from a position j ∈ Γ is marked (the distance between j and a block
S[i..i(cid:48)] is i− j if i > j, j − i(cid:48) if i(cid:48) < j, and 0 otherwise). Blocks S[i..i(cid:48)] that are
not marked are replaced by a pointer (cid:104)ptr1, ptr2, δ(cid:105) to an occurrence S[j(cid:48)..j(cid:48)(cid:48)]
of S[i..i(cid:48)] that includes a position j ∈ Γ, j(cid:48) ≤ j ≤ j(cid:48)(cid:48). Such an occurrence
exists by Definition 1. Moreover, it must be covered by 1 or 2 consecutive
marked blocks of the same level due to our marking mechanism, because all
the positions in S[j(cid:48)..j(cid:48)(cid:48)] are at distance < bl from j. Those 1 or 2 nodes of the
Γ-tree are ptr1 and ptr2, and δ is the offset of j(cid:48) within ptr1 (δ = 0 if j(cid:48) is the
first symbol inside ptr1).
In level l + 1 we explicitly store only the children of the blocks that were
marked in level l. The blocks stored in the Γ-tree (i.e., all blocks at level 0
and those having a marked parent) are called explicit. In the last level, the
marked blocks store their corresponding single symbol from S.
See Figure 1 for an example of a Γ-tree. We can regard the Γ-tree as a binary
tree (with the first lg γ levels chopped out), where the internal nodes are
marked nodes and have two children, whereas the leaves either are marked
and represent just one symbol, or are unmarked and represent a pointer to 1
or 2 marked nodes of the same level. If we call w the number of leaves, then
there are w − γ (marked) internal nodes. From the leaves, γ of them represent
single symbols in the last level, while the other w − γ leaves are unmarked
blocks replaced by pointers. Thus, there are in total 2w − γ nodes in the tree,
4 For simplicity of description, we assume that n/γ is a power of 2.
6
Fig. 1. Example of a Γ-tree built on a text of length n = 24 with γ = 3 attractor po-
sitions (black letters). Marked blocks are colored in gray. Each non-marked block (in
white) is associated with an occurrence (underlined) crossing an attractor position,
and therefore overlapping only marked blocks. Only explicit blocks are shown.
of which w − γ are internal nodes, w − γ are pointers, and γ store explicit
symbols. Alternatively, w nodes are marked (internal nodes plus leaves) and
w − γ are unmarked (leaves only).
The Γ-tree then uses O(w) space. To obtain a bound in terms of n and γ,
note that, at each level, each j ∈ Γ may mark up to 3 blocks; thus there are
w ≤ 3γ lg(n/γ) marked blocks in total and the Γ-tree uses O(γ lg(n/γ)) space.
We now describe two operations on Γ-trees that are fundamental to support
efficient indexed searches. The former is also necessary for a self-index, as
it allows us extracting arbitrary substrings of S efficiently. We remind that
this procedure is not the same described on the original structure of string
attractors [35], because the structures are also different.
2.1 Extraction
To extract a single symbol S[i], we first map it to a local offset 1 ≤ i(cid:48) ≤ b0 in its
corresponding level-0 block. In general, given the local offset i(cid:48) of a character
in the current block at level l, we first see if the current block is marked. If so,
we map i(cid:48) to a position in the next level l + 1, where the current block is split
into two blocks of half the length: if i(cid:48) ≤ bl+1, then we continue on the left
child with the same offset; otherwise, we subtract bl+1 from i(cid:48) and continue
on the right child. If, instead, i is not in a marked block, we take the pointer
(cid:104)ptr1, ptr2, δ(cid:105) stored for that block and add δ to i(cid:48). If the result is i(cid:48) ≤ bl, then
7
b b a b a b b a b a b b a b b a b a b b a b b aa b b ab a b b a b b ab a b b a b b a b b a b a b b a b a b b a b b a b a b b a b a b b b a b b b b b b a we continue in the node ptr1 with offset i(cid:48); otherwise, we continue in ptr2 with
the offset i(cid:48) − bl. In both cases, the new node is marked, so we proceed as on
marked blocks in order to move to the next level in constant time. The total
time to extract a symbol is then O(lg(n/γ)).
A substring of length (cid:96) can thus be extracted in time O((cid:96) lg(n/γ)), which will
be sufficient to obtain the search time complexity of Theorem 2. It is possible
to augment Γ-trees to match the complexity O(lg(n/γ) + (cid:96)/ lgσ n) obtained by
Kempa and Prezza [35] on string attractors as well, though this would have
no impact on our results.
2.2 Fingerprinting
We now show that the Γ-tree can be augmented to compute the Karp -- Rabin
fingerprint of any text substring in logarithmic time.
Lemma 1. Let S[1..n] have an attractor of size γ and φ a Karp -- Rabin finger-
print function. Then we can store a data structure of size O(γ lg(n/γ)) words
supporting the computation of φ on any substring of S in O(lg(n/γ)) time.
Proof. We augment our Γ-tree of S as follows. At level 0, we store the Karp --
Rabin fingerprints of all the text prefixes ending at positions i · n/γ, for i =
1, . . . , γ. At levels l > 0, we store the fingerprints of all explicit blocks.
We first show that we can reduce the problem to that of computing the fin-
gerprints of two prefixes of explicit blocks. Then, we show how to solve the
sub-problem of computing fingerprints of prefixes of explicit blocks.
Let S[i..j] be the substring of which we wish to compute the fingerprint
φ(S[i..j]). Note that φ(S[i..j]) can be computed in constant time from φ(S[1..i−
1]) and φ(S[1..j]) so we can assume, without loss of generality, that i = 1 (i.e.,
the substring is a prefix of S). Then, at level 0 the substring spans a sequence
B1 ··· Bt of blocks followed by a prefix C of block Bt+1 (the sequence of blocks
or C could be empty). The fingerprint of B1 ··· Bt is explicitly stored, so the
problem reduces to that of computing the fingerprint of C.
We now show how to compute the fingerprint of a prefix of an explicit block
(at any level) in O(lg(n/γ)) time. We distinguish two cases.
(A) We wish to compute the fingerprint of B[1..k], for some k ≤ bl, and B
is a marked block at level l. Let Bleft and Bright be the children of B at level
l + 1. Then, the problem reduces to either (i) computing the fingerprint of
Bleft [1..k] if k ≤ bl+1, or combining the fingerprints of Bleft (which is stored)
8
and Bright [1..k − bl+1]. In both sub-cases, the problem reduces to that of com-
puting the fingerprint of the prefix of a block at level l + 1, which is explicit
since B is marked.
(B) We wish to compute the fingerprint of B[1..k], for some k ≤ bl, but B is
an unmarked explicit block. Then, B is linked (through a Γ-tree pointer) to an
occurrence in the same level spanning at most two blocks, both of which are
marked. If the occurrence of B spans only one marked block B(cid:48) at level l, then
B[1..bl] = B(cid:48)[1..bl] and we are back in case (A). Otherwise, the occurrence of
B spans two marked blocks B(cid:48) and B(cid:48)(cid:48) at level l: B[1..bl] = B(cid:48)[i..bl]B(cid:48)(cid:48)[1..i −
1], with i ≤ bl. For each pointer of this kind in the Γ-tree, we store the
fingerprint of B(cid:48)[i..bl]. We consider two sub-cases. (B.1) If k ≥ bl − i + 1,
then B[1..k] = B(cid:48)[i..bl]B(cid:48)(cid:48)[1..k − (bl − i + 1)]. Since we store the fingerprint of
B(cid:48)[i..bl], the problem reduces again to that of computing the fingerprint of the
prefix B(cid:48)(cid:48)[1..k− (bl − i + 1)] of a marked (explicit) block. (B.2) If k < bl − i + 1,
then B[1..k] = B(cid:48)[i..i + k − 1]. Although this is not a prefix nor a suffix of a
block, note that B[1..k]B(cid:48)[i + k..bl] = B(cid:48)[i..i + k − 1]B(cid:48)[i + k..bl] = B(cid:48)[i..bl]. It
follows that we can retrieve the fingerprint of B[1..k] in constant time using
the fingerprints of B(cid:48)[i+k..bl] and B(cid:48)[i..bl]. The latter value is explicitly stored.
The former is the fingerprint of the suffix of an explicit (marked) block. In
this case, note that the fingerprint of a block's suffix can be retrieved from the
fingerprint of the block and the fingerprint of a block's prefix, so we are back
to the problem of computing the fingerprint of an explicit block's prefix.
To sum up, computing a prefix of an explicit block at level l reduces to the
problem of computing a prefix of an explicit block at level l + 1 (plus a con-
stant number of arithmetic operations to combine values). In the worst case,
we navigate down to the leaves, where fingerprints of single characters can
be computed in constant time. Combining this procedure into our main algo-
rithm, we obtain the claimed running time of O(lg(n/γ)).
3 A Universal Self-Index
Our self-index structure builds on the Γ-tree of S. It is formed by two main
components: the first finds all the pattern occurrences that cross explicit block
boundaries, whereas the second finds the occurrences that are completely in-
side unmarked blocks.
Lemma 2. Any substring S[i..j] of length at least 2 either overlaps two con-
secutive explicit blocks or is completely inside an unmarked block.
Proof. The leaves of the Γ-tree partition S into a sequence of explicit blocks:
γ of those are attractor positions and the other w − γ are unmarked blocks.
9
Clearly, if S[i..j] is not completely inside an unmarked block, it must cross a
boundary between two explicit blocks.
We exploit the lemma in the following way. We will define an occurrence of
P as primary if it overlaps two consecutive explicit blocks. The occurrences
that are completely contained in an unmarked block are secondary. By the
lemma, every occurrence of P is either primary or secondary. We will use a
data structure to find the primary occurrences and another one to detect the
secondary ones. The primary occurrences are found by exploiting the fact that
a prefix of P matches at the end of an explicit block and the remaining suffix
of P matches the text that follows. Secondary occurrences, instead, are found
by detecting primary or other secondary occurrences within the area where
an unmarked block points.
We note that this idea is a variant of the classical one [32] used in all in-
dexes based on LZ77 and CFGs. Now we show that the principle can indeed
be applied on attractors, which is the general concept underlying all those
compression methods (and others where no indexes exist yet), and therefore
unifies all those particular techniques.
3.1 Primary Occurrences
We describe the data structures and algorithms used to find the primary occur-
rences. Overall, they require O(w) space and find the occp primary occurrences
in time O(m lg(mn/γ) + occp lg w), for any constant > 0.
Data Structures. The leaves of the Γ-tree partition S into explicit blocks.
The partition is given by the starting positions 1 = p1 < . . . < pw ≤ n of
the leaves. By Lemma 2, every primary occurrence contains some substring
S[pi − 1..pi] for 1 < i ≤ w.
If we find the occurrences considering only their leftmost covered position
of the form pi − 1, we also ensure that each primary occurrence is found
once. Thus, we will find primary occurrences as a prefix of P appearing at
the end of some S[pi−1..pi − 1] followed by the corresponding suffix of P ap-
pearing at the beginning of S[pi..n]. For this sake, we define the set of pairs
B = {(cid:104)S[pi..n], S[pi−1..pi − 1]rev(cid:105), 1 < i ≤ w}, where S[pi−1..pi − 1]rev means
S[pi−1..pi − 1] read backwards, and form multisets X and Y with the left and
right components of B, respectively.
We then lexicographically sort X and Y, to obtain the strings X1, X2, . . . and
Y1, Y2, . . .. All the occurrences ending with a certain prefix of P will form a
10
contiguous range in the sorted multiset Y, whereas all those starting with a
certain suffix of P will form a contiguous range in the sorted multiset X . Each
primary occurrence of P will then correspond to a pair (cid:104)Xx, Yy(cid:105) ∈ B where
both Xx and Yy belong to their range.
Our structure to find the primary occurrences is a two-dimensional discrete
grid G storing one point (x, y) for each pair (cid:104)Xx, Yy(cid:105) ∈ B. The grid G is of size
(w − 1) × (w − 1). We represent G using a two-dimensional range search data
structure requiring O(w) space [12] that reports the t points lying inside any
rectangle of the grid in time O((t + 1) lg w), for any constant > 0. We also
store an array T [1..w− 1] that, for each point (x, y) in G, where Xx = S[pi..n],
stores T [y] = pi, that is, where Xx starts in S.
Queries. To search for a pattern P [1..m], we first find its primary oc-
currences using G as follows. For each partition P< = P [1..k] and P> =
< and X for P>. For each
P [k + 1..m], for 1 ≤ k < m, we search Y for P rev
identified range [x1, x2] × [y1, y2], we extract all the t corresponding primary
occurrences (x, y) in time O((t + 1) lg w) with our range search data struc-
ture. Then we report a primary occurrence starting at T [y] − k for each such
point (x, y). Over the m intervals, this adds up to O((m + occp) lg w) =
O((m + occp) lg(γ lg(n/γ)).
We obtain the m− 1 ranges in the multiset X in overall time O(m lg(mn/γ)),
by using the fingerprint-based technique of Gagie et al. [24] applied to the z-
fast trie of Belazzougui et al. [2] (we use a lemma from Gagie et al. [26] where
the overall result is stated in cleaner form). We use an analogous structure to
obtain the ranges in Y of the suffixes of the reversed pattern.
Lemma 3 (adapted from [26, Lem. 5.2]). Let S[1..n] be a string on alphabet
[1..σ], X be a sorted set of suffixes of S, and φ a Karp -- Rabin fingerprint
function. If one can extract a substring of length (cid:96) from S in time fe((cid:96)) and
compute φ on it in time fh((cid:96)), then one can build a data structure of size
O(X) that obtains the lexicographic ranges in X of the m − 1 suffixes of a
given pattern P in worst-case time O(m lg(σ)/ω+m(fh(m)+lg m)+fe(m)) --
provided that φ is collision-free among substrings of S whose lengths are powers
of two.
In Sections 2.1 and 2.2 we show how to extract in time fe((cid:96)) = O((cid:96) lg(n/γ))
and how to compute a fingerprint in time fh((cid:96)) = O(lg(n/γ)), respectively.
In Section 4.2 we show that a Karp -- Rabin function that is collision-free
among substrings whose lengths are powers of two can be efficiently found.
Together, these results show that we can find all the ranges in X and Y in
time O(m lg(σ)/ω + m(lg(n/γ) + lg m) + m lg(n/γ)) = O(m lg(mn/γ)).
11
Patterns P of length m = 1 can be handled as P [1]∗, where ∗ stands for any
character. Thus, we take [x1, x2] = [1, w] and carry out the search as a normal
pattern of length m = 2. To make this work also for the last position in S,
we assume as usual that S is terminated by a special symbol $ that cannot
appear in search patterns P . Alternatively, we can store a list of the marked
leaves where each alphabet symbol appears, and take those as the primary
occurrences. A simple variant of Lemma 2 shows that occurrences of length 1
are either attractor positions or are inside an unmarked block, so the normal
mechanism to find secondary occurrences from this set works correctly. Since
there are in total γ such leaves, the space for these lists is O(σ + γ) = O(γ).
3.2 Secondary Occurrences
We now describe the data structures and algorithms to find the secondary
occurrences. They require O(w) space and find the occs secondary occurrences
in time O((occp + occs) lg lgω(n/γ)).
Data structures. To track the secondary occurrences, let us call target and
source the text areas S[i..i(cid:48)] and S[j(cid:48)..j(cid:48)(cid:48)], respectively, of an unmarked block
and its pointer, so that there is some j ∈ Γ contained in S[j(cid:48)..j(cid:48)(cid:48)] (if the blocks
are at level l, then i(cid:48) = i+bl−1 and j(cid:48)(cid:48) = j(cid:48) +bl−1). Let S[pos..pos+m−1] be
an occurrence we have already found (using the grid G, initially). Our aim is to
find all the sources that contain S[pos..pos + m− 1], since their corresponding
targets then contain other occurrences of P .
To this aim, we store the sources of all levels in an array R[1..w − γ], with
fields j(cid:48) and j(cid:48)(cid:48), ordered by starting positions R[k].j(cid:48). We build a predecessor
search structure on the fields R[k].j(cid:48), and a range maximum query (RMQ)
data structure on the fields R[k].j(cid:48)(cid:48), able to find the maximum endpoint j(cid:48)(cid:48) in
any given range of R. While a predecessor search using O(w) space requires
O(lg lgω(n/w)) time on an ω-bit-word machine [45, Sec. 1.3.2], the RMQ struc-
ture operates in constant time using just O(w) bits [21].
Queries. Let S[pos..pos+m−1] be a primary occurrence found. A predeces-
sor search for pos gives us the rightmost position r where the sources start at
R[r].j(cid:48) ≤ pos. An RMQ on R[1..r] then finds the position k of the source with
the rightmost endpoint R[k].j(cid:48)(cid:48) in R[1..r]. If even R[k].j(cid:48)(cid:48) < pos+m−1, then no
source covers the occurrence and we finish. If, instead, R[k].j(cid:48)(cid:48) ≥ pos + m − 1,
then the source R[k] covers the occurrence and we process its correspond-
ing target as a secondary occurrence; in this case we also recurse on the
ranges R[1..k − 1] and R[k + 1..r] that are nonempty. It is easy to see that
12
each valid secondary occurrence is identified in O(1) time (see Muthukrish-
nan [41] for an analogous process). In addition, such secondary occurrences,
S[pos(cid:48)..pos(cid:48)+m−1], must be recursively processed for further secondary occur-
rences. A similar procedure is described for tracking the secondary occurrences
in the LZ77-index [38].
The cost per secondary occurrence reported then amortizes to a predecessor
search, O(lg lgω(n/γ)) time. This cost is also paid for each primary occurrence,
which might not yield any secondary occurrence to amortize it. We now prove
that this process is sufficient to find all the secondary occurrences.
Lemma 4. The described algorithm reports every secondary occurrence exactly
once.
Proof. We use induction on the level l of the unmarked block that contains
the secondary occurrence. All the secondary occurrences of the last level are
correctly reported once, since there are none. Now consider an occurrence
S[pos..pos + m − 1] inside an unmarked block S[i..i(cid:48)] of level l. This block is
the target of a source S[j(cid:48)..j(cid:48)(cid:48)] that spans 1 or 2 consecutive marked blocks
of level l. Then there is another occurrence S[pos(cid:48)..pos(cid:48) + m − 1], with pos(cid:48) =
pos − i + j(cid:48). Note that the algorithm can report S[pos..pos + m − 1] only
as a copy of S[pos(cid:48)..pos(cid:48) + m − 1]. If S[pos(cid:48)..pos(cid:48) + m − 1] is primary, then
S[pos..pos + m − 1] will be reported right after S[pos(cid:48)..pos(cid:48) + m − 1], because
[pos(cid:48)..pos(cid:48) + m − 1] ⊆ [j(cid:48)..j(cid:48)(cid:48)], and the algorithm will map [j(cid:48)..j(cid:48)(cid:48)] to [i..i(cid:48)] to
discover S[pos..pos + m − 1]. Otherwise, S[pos(cid:48)..pos(cid:48) + m − 1] is secondary,
and thus it is within one of the marked blocks at level l that overlap the
source. Moreover, it is within one of the blocks of level l + 1 into which those
marked blocks are split. Thus, S[pos(cid:48)..pos(cid:48) + m − 1] is within an unmarked
block of level > l, which by the inductive hypothesis will be reported exactly
once. When S[pos(cid:48)..pos(cid:48) + m− 1] is reported, the algorithm will also note that
[pos(cid:48)..pos(cid:48) + m − 1] ⊆ [j(cid:48)..j(cid:48)(cid:48)] and will find S[pos..pos + m − 1] once as well.
If we handle patterns of length m = 1 by taking the corresponding attractor
positions as the primary occurrences, then there may be secondary occurrences
in the last level, but those point directly to primary occurrences (i.e., attractor
positions), and therefore the base case of the induction holds too.
The total search cost with occ primary and secondary occurrences is there-
fore O(m(lg(γ lg(n/γ)) + lg(mn/γ)) + occ(lg(γ lg(n/γ)) + lg lgω(n/γ))) =
O(m lg n + occ(lg γ + lg lg(n/γ))) = O(m lg n + occ lg n), for any constant
> 0 defined at indexing time (the choice of affects the constant that ac-
companies the size O(γ lg(n/γ)) of the structure G).
13
4 Construction
If we allow the index construction to be correct with high probability only,
then we can build it in O(n + w lg(n/γ) + w
lg w) time and O(w) space (plus
read-only access to S), using a Monte Carlo method. Since w = O(γ lg(n/γ))
is the number of leaves in the Γ-tree and γ(lg(n/γ))O(1) ⊆ O(n), the time can
be written as O(n + w
lg γ). In order to ensure a correct index, a Las Vegas
method requires O(n lg n) time in expectation (and w.h.p.) and O(n) space.
√
√
4.1 Building the Γ-tree
Given the attractor Γ, we can build the index data structure as follows. At
each level l, we create an Aho -- Corasick automaton [1] on the unmarked blocks
at this level (i.e., those at distance ≥ bl from any attractor), and use it to scan
the areas S[j − bl + 1..j + bl − 1] around all the attractor elements j ∈ Γ in
order to find a proper pointer for each of those unmarked blocks. This takes
O(n) time per level. Since the areas around each of the γ attractor positions
are scanned at each level but they have exponentially decreasing lengths, the
scanning time adds up to O(γ n
4γ + ··· ) = O(n).
2γ + γ n
γ + γ n
As for preprocessing time, each unmarked block is preprocessed only once,
and they add up to O(n) symbols. The preprocessing can thus be done in
time O(n) [19]. To be able to scan in linear time, we can build deterministic
dictionaries on the edges outgoing from each node, in time O(n(lg lg σ)2) [46].
In total, we can build the Γ-tree in O(n(lg lg σ)2) deterministic time and O(n)
space.
To reduce the space to O(w), instead of inserting the unmarked blocks into
an Aho -- Corasick automaton, we compute their Karp -- Rabin fingerprints, store
them in a hash table, and scan the areas S[j−bl +1..j +bl−1] around attractor
elements j. This finds the correct sources for all the unmarked blocks w.h.p.
Indeed, if we verify the potential collisions, the result is always correct within
O(n) expected time (further, this time holds w.h.p.).
4.2 Building the Fingerprints
Building the structures for Lemma 1 requires (i) computing the fingerprint of
every text prefix ending at block boundaries (O(n) time and O(w) space in
addition to S), (ii) computing the fingerprint of every explicit block (O(w)
time and O(w) space starting from the leaves and combining results up to the
14
root), (iii) for each unmarked explicit block B, computing the fingerprint of
a string of length at most B (i.e., the fingerprint of B(cid:48)[i..bl]; see case (B) of
Lemma 1). Since unmarked blocks do not have children, each text character
is seen at most once while computing these fingerprints, which implies that
these values can also be computed in O(n) time and O(w) space in addition
to S.
This process, however, does not include finding a collision-free Karp -- Rabin
hash function. As a result, the fingerprinting is correct w.h.p. only. We can use
the de-randomization procedure of Bille et al. [9], which guarantees to find --
in O(n lg n) expected time 5 and O(n) words of space -- a Karp -- Rabin hash
function that is collision-free among substrings of S whose lengths are powers
of two. This is sufficient to deterministically check the equality of substrings 6
in the z-fast trie used in the technique [26, Lem. 5.2] that we use to quickly
find ranges of pattern suffixes/prefixes (in our Section 3.1).
4.3 Building the Multisets X and Y
To build the multisets X and Y for the primary occurrences, we can build the
suffix arrays [40] of S and its reverse, Srev. This requires O(n) deterministic
time and space [31]. Then we can scan those suffix arrays to enumerate X and
Y in the lexicographic order.
To sort X and Y within O(w) space, we can build instead a sparse suffix tree
on the w positions of X or Y in S. This can be done in expected time (and
√
lg w) and O(w) space [27]. If we aim to build the suffix array
w.h.p.) O(n
correctly w.h.p. only, then the time drops to O(n).
We must then build the z-fast trie [2, Thm. 5] on the sets X and Y. Since we
can use any space in O(w), we opt for a simpler variant described by Kempa
and Kosolobov [34, Lem. 5], which is easier to build. Theirs is a compact
trie that stores, for each node v representing the string v.str and with parent
node v.par: (i) the length v.str, (ii) a dictionary mapping each character
c to the child node v(cid:48) of v such that v(cid:48).str[v.str + 1] = c (if such a child
exists), and (iii) the (non-extended) fingerprint of v.str[1..kv], where kv is the
two-fattest number in the range [v.par.str + 1..v.str], that is, the number
in that range whose binary representation has the largest number of trailing
zeros. The trie also requires a global "navigating" hash table that maps the
O(w) pairs (kv, φ(v.str[1..kv])) to their corresponding node v.
5 The time is also O(n lg n) w.h.p., not only in expectation.
6 If the length (cid:96) of the two substrings is not a power of two, then we compare their
prefixes and suffixes whose length is the largest power of two smaller than (cid:96).
15
If p prefixes some string in X (resp. Y) and the fingerprint function φ is
collision-free among equal-length text substrings, then their so-called fat bi-
nary search procedure finds -- in time O(lg p) -- the highest node v such
that v.str is prefixed by a search pattern p (assuming constant-time compu-
tation of the fingerprint of any substring of p, which can be achieved after a
linear-time preprocessing of the pattern). The range of strings of X (resp. Y)
is then associated with the node v.
Their search procedure, however, has an anomaly [34, Lem. 5]: in some cases it
might return a child of v instead of v. We can fix this problem (and always re-
turn the correct v) by also storing, for every internal trie node v, the fingerprint
of v.str and the exit character v.str[v.par.str + 1]. Then, we know that the
procedure returned v if and only if v.par.str < p, the fingerprint of v.par.str
matches that of p[1..v.par.str] and v.str[v.par.str + 1] = p[v.par.str + 1].
If these conditions do not hold, we know that the procedure returned a child
of v and we fix the answer by moving to the parent of the returned node. 7
If p does not prefix any string in X (resp. Y), or if φ is not collision-free, the
search (with our without our fix) returns an arbitrary range. We recall that, to
cope with the first condition (which gives the z-fast trie search the name weak
prefix search), Lemma 3 adds a deterministic check that allows discarding
the incorrect ranges returned by the procedure. The second condition, on the
other hand, w.h.p. does not fail. Still, we recall that we can ensure that our
function is collision-free by checking, at construction time, for collisions among
substrings whose lengths are powers of two, as described in Section 4.2.
The construction of this z-fast trie starts from the sparse suffix tree of X
(or Y), since the topology is the same. The dictionaries giving constant-time
access to the nodes' children can be built correctly w.h.p. in time O(w),
or be made perfect in expected time (and w.h.p.) O(w) [51]. Alternatively,
one can use deterministic dictionaries, which can be built in worst-case time
O(w(lg lg σ)2) [46]. In all cases, the construction space is O(w). Similarly, the
navigating hash can be built correctly w.h.p. in time O(w), or be made perfect
in expected time (and w.h.p.) O(w). We could also represent the navigating
hash using deterministic dictionaries built in worst-case time O(w(lg lg w)2)
[46] and space O(w). The O(w) fingerprints φ(v.str) and the hashes of the
pairs (kv, φ(v.str[1..kv])) can be computed in total time O(w lg(n/γ)) by using
Lemma 1 to compute the Karp -- Rabin fingerprints.
7 This fix works in general. In their particular procedure, we do not need to check
that v.str[v.par.str + 1] = p[v.par.str + 1] (nor to store the exit character). Alter-
natively, we could directly fix their procedure by similarly storing the fingerprint of
v.str (hash(v.str), in their article) and changing line 6 of their extended version [34]
to "if v.len < pat and hash(v.str) = hash(pat[1..v.len]) and v.map(pat[v.len +
1]) (cid:54)= nil then v ← v.map(pat[v.len + 1]);".
16
4.4 Structures to Track Occurrences
√
A variant of the grid data structure G [6], with the same space and time
performance, can be built in O(w
lg w) deterministic time, and O(w) space.
The arrays T and R can be built in O(w) space and O(w lg lg w) deterministic
time [28].
The RMQ structure on R requires O(w) deterministic construction time and
O(w) bits [21]. The predecessor data structure [45], however, requires perfect
hashing. This can be built in O(w) space and O(w(lg lg w)2) deterministic
time [46], or in O(w) expected time (also holding w.h.p.).
5 Conclusions
We have introduced the first universal self-index for repetitive text collections.
The index is based on a recent result [35] that unifies a large number of
dictionary compression methods into the single concept of string attractor. For
each compression method based on Lempel -- Ziv, grammars, run-compressed
BWT, collage systems, macro schemes, etc., it is easy to identify an attractor
set of the same asymptotic size obtained by the compression method, say
γ. Thus, our construction automatically yields a self-index for each of those
compression methods, within O(γ lg(n/γ)) space. No structure is known of
size o(γ lg(n/γ)) able to efficiently extract a substring from the compressed
text (say, within O(polylog n) time per symbol), and thus O(γ lg(n/γ)) is the
minimum space we could hope for an efficient self-index with the current state
of the art.
Indeed, no self-index of size o(γ lg(n/γ)) is known: the smallest ones use
O(z) [38] and O(r) [26] space, respectively, yet there are text families where
z, r = Ω(b∗ lg n) = Ω(γ∗ lg n) [25] (where b∗ and γ∗ denote the smallest macro
scheme and attractor sizes, respectively). The most time-efficient self-indexes
use Θ(r lg(n/r)) and Ω(z lg(n/z)) space, which is asymptotically equal or
larger than ours. The search time of our index, O(m lg n + occ lg n) for any
constant > 0, is close to that of the fastest of those self-indexes, which were
developed for a specific compression format (see [26, Table 1]). Moreover, our
construction provides a self-index for compression methods for which no such
structure existed before, such as collage systems and macro schemes. Those
can provide smaller attractor sets than the ones derived from the more popular
compression methods.
We can improve the search time of our index by using slightly more space.
Our current bottleneck in the per-occurrence query time is the grid data struc-
17
ture G, which uses O(w) space and returns each occurrence in time O(lg w).
Instead, a grid structure [12] using O(w lg lg w) = O(γ lg(n/γ) lg lg n) space
obtains the occp primary occurrences in time O((occp +1) lg lg w). This slightly
larger version of our index can then search in time O(m lg n + occ lg lg n). This
complexity is close to that of some larger indexes in the literature for repetitive
string collections (see [26, Table 1]).
A number of avenues for future work are open, including supporting more
complex pattern matching, handling dynamic collections of texts, supporting
document retrieval, and implementing a practical version of the index. Any
advance in this direction will then translate into all of the existing indexes for
repetitive text collections.
Acknowledgements
We thank the reviewers for their outstanding job in improving our presentation
(and even our bounds).
References
[1] A. V. Aho and M. J. Corasick. Efficient string matching: an aid to bibliographic
search. Communications of the ACM, 18(6):333 -- 340, 1975.
[2] D. Belazzougui, P. Boldi, R. Pagh, and S. Vigna. Fast prefix search in
little space, with applications. In Proc. 18th Annual European Symposium on
Algorithms (ESA), pages 427 -- 438, 2010.
[3] D. Belazzougui and F. Cunial. Fast label extraction in the CDAWG. In Proc.
24th International Symposium on String Processing and Information Retrieval
(SPIRE), LNCS 10508, pages 161 -- 175, 2017.
[4] D. Belazzougui, F. Cunial, T. Gagie, N. Prezza, and M. Raffinot. Composite
In Proc. 26th Annual Symposium on
repetition-aware data structures.
Combinatorial Pattern Matching (CPM), LNCS 9133, pages 26 -- 39, 2015.
[5] D. Belazzougui, T. Gagie, P. Gawrychowski, J. Karkkainen, A. Ord´onez, S. J.
Puglisi, and Y. Tabei. Queries on LZ-bounded encodings. In Proc. 25th Data
Compression Conference (DCC), pages 83 -- 92, 2015.
[6] D. Belazzougui and S. J. Puglisi. Range predecessor and Lempel-Ziv parsing.
In Proc. 27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 2053 -- 2071, 2016.
18
[7] D. Belazzougui, S. J. Puglisi, and Y. Tabei. Access, rank, select in grammar-
compressed strings. In Proc. 23rd Annual European Symposium on Algorithms
(ESA), LNCS 9294, pages 142 -- 154, 2015.
[8] P. Bille, M. B. Ettienne, I. L. Gørtz, and H. W. Vildhøj. Time-space trade-offs
for Lempel-Ziv compressed indexing. Theoretical Computer Science, 713:66 -- 77,
2018.
[9] P. Bille, I. L. Gørtz, M. B. T. Knudsen, M. Lewenstein, and H. W. Vildhøj.
Longest common extensions in sublinear space.
In Proc. 26th Annual
Symposium on Combinatorial Pattern Matching (CPM), LNCS 9133, pages 65 --
76, 2015.
[10] A. Blumer, J. Blumer, D. Haussler, R. McConnell, and A. Ehrenfeucht.
Complete inverted files for efficient text retrieval and analysis. Journal of the
ACM, 34(3):578 -- 595, 1987.
[11] M. Burrows and D. Wheeler. A block sorting lossless data compression
algorithm. Technical Report 124, Digital Equipment Corporation, 1994.
[12] T. M. Chan, K. G. Larsen, and M. Patra¸scu. Orthogonal range searching on the
RAM, revisited. In Proc. 27th ACM Symposium on Computational Geometry
(SoCG), pages 1 -- 10, 2011.
[13] M. Charikar, E. Lehman, D. Liu, R. Panigrahy, M. Prabhakaran, A. Sahai, and
A. Shelat. The smallest grammar problem. IEEE Transactions on Information
Theory, 51(7):2554 -- 2576, 2005.
[14] A. R. Christiansen and M. B. Ettienne.
Compressed indexing with
signature grammars. In Proc. 13th Latin American Symposium on Theoretical
Informatics (LATIN), LNCS 10807, pages 331 -- 345, 2018.
[15] F. Claude, A. Farina, M. Mart´ınez-Prieto, and G. Navarro. Universal indexes
for highly repetitive document collections. Information Systems, 61:1 -- 23, 2016.
[16] F. Claude and G. Navarro.
Self-indexed grammar-based compression.
Fundamenta Informaticae, 111(3):313 -- 337, 2010.
[17] F. Claude and G. Navarro.
Improved grammar-based compressed indexes.
In Proc. 19th International Symposium on String Processing and Information
Retrieval (SPIRE), LNCS 7608, pages 180 -- 192, 2012.
[18] M. Crochemore and R. V´erin. Direct construction of compact directed acyclic
In Proc. 8th Annual Symposium on Combinatorial Pattern
word graphs.
Matching (CPM), LNCS 1264, pages 116 -- 129. Springer, 1997.
[19] S. Dori and G. M. Landau. Construction of Aho Corasick automaton in linear
time for integer alphabets. Information Processing Letters, 98(2):66 -- 72, 2006.
[20] P. Ferragina and G. Manzini. Indexing compressed texts. Journal of the ACM,
52(4):552 -- 581, 2005.
19
[21] J. Fischer and V. Heun.
Space-efficient preprocessing schemes for range
minimum queries on static arrays. SIAM Journal on Computing, 40(2):465 --
492, 2011.
[22] T. Gagie. Large alphabets and incompressibility.
Information Processing
Letters, 99(6):246 -- 251, 2006.
[23] T. Gagie, P. Gawrychowski, J. Karkkainen, Y. Nekrich, and S. J. Puglisi. A
In Proc. 6th International Conference on
faster grammar-based self-index.
Language and Automata Theory and Applications (LATA), LNCS 7183, pages
240 -- 251, 2012.
[24] T. Gagie, P. Gawrychowski, J. Karkkainen, Y. Nekrich, and S. J. Puglisi. LZ77-
based self-indexing with faster pattern matching. In Proc. 11th Latin American
Symposium on Theoretical Informatics (LATIN), LNCS 8392, pages 731 -- 742,
2014.
[25] T. Gagie, G. Navarro, and N. Prezza. On the approximation ratio of Lempel-Ziv
parsing. In Proc. 13th Latin American Symposium on Theoretical Informatics
(LATIN), LNCS 10807, pages 490 -- 503, 2018.
[26] T. Gagie, G. Navarro, and N. Prezza. Optimal-time text indexing in BWT-
runs bounded space. In Proc. 29th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 1459 -- 1477, 2018.
[27] P. Gawrychowski and T. Kociumaka. Sparse suffix tree construction in optimal
In Proc. 28th Annual ACM-SIAM Symposium on Discrete
time and space.
Algorithms (SODA), pages 425 -- 439, 2017.
[28] Y. Han. Deterministic sorting in O(n lg lg n) time and linear space. Journal of
Algorithms, 50(1):96 -- 105, 2004.
[29] A. Jez. Approximation of grammar-based compression via recompression.
Theoretical Computer Science, 592:115 -- 134, 2015.
[30] A. Jez. A really simple approximation of smallest grammar. Theoretical
Computer Science, 616:141 -- 150, 2016.
[31] J. Karkkainen, P. Sanders, and S. Burkhardt.
Linear work suffix array
construction. Journal of the ACM, 53(6):918 -- 936, 2006.
[32] J. Karkkainen and E. Ukkonen. Lempel-Ziv parsing and sublinear-size index
In Proc. 3rd South American Workshop on
structures for string matching.
String Processing (WSP), pages 141 -- 155, 1996.
[33] R. M. Karp and M. O. Rabin.
Efficient randomized pattern-matching
algorithms. IBM Journal of Research and Development, 31(2):249 -- 260, 1987.
[34] D. Kempa and D. Kosolobov. LZ-End parsing in compressed space. In Proc.
27th Data Compression Conference (DCC), pages 350 -- 359, 2017. Extended
version in https://arxiv.org/pdf/1611.01769.pdf.
20
[35] D. Kempa and N. Prezza. At the roots of dictionary compression: String
In Proc. 50th Annual ACM SIGACT Symposium on Theory of
attractors.
Computing (STOC), pages 827 -- 840, 2018.
[36] T. Kida, T. Matsumoto, Y. Shibata, M. Takeda, A. Shinohara, and S. Arikawa.
Collage system: A unifying framework for compressed pattern matching.
Theoretical Computer Science, 298(1):253 -- 272, 2003.
[37] J. C. Kieffer and E. Yang. Grammar-based codes: A new class of universal
lossless source codes. IEEE Transactions on Information Theory, 46(3):737 --
754, 2000.
[38] S. Kreft and G. Navarro. On compressing and indexing repetitive sequences.
Theoretical Computer Science, 483:115 -- 133, 2013.
[39] A. Lempel and J. Ziv. On the complexity of finite sequences.
IEEE
Transactions on Information Theory, 22(1):75 -- 81, 1976.
[40] U. Manber and G. Myers. Suffix arrays: a new method for on-line string
searches. SIAM Journal on Computing, 22(5):935 -- 948, 1993.
[41] S. Muthukrishnan. Efficient algorithms for document retrieval problems.
In
Proc. 13th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 657 -- 666, 2002.
[42] G. Navarro. A self-index on Block Trees.
In Proc. 24th International
Symposium on String Processing and Information Retrieval (SPIRE), LNCS
10508, pages 278 -- 289, 2017.
[43] T. Nishimoto, T. I, S. Inenaga, H. Bannai, and M. Takeda. Dynamic index and
LZ factorization in compressed space. In Proc. Prague Stringology Conference
(PSC), pages 158 -- 170, 2016.
[44] T. Nishimoto, T. I, S. Inenaga, H. Bannai, and M. Takeda. Fully dynamic data
structure for LCE queries in compressed space.
In Proc. 41st International
Symposium on Mathematical Foundations of Computer Science (MFCS), pages
72:1 -- 72:15, 2016.
[45] M. Patra¸scu and M. Thorup. Time-space trade-offs for predecessor search. In
Proc. 38th Annual ACM Symposium on Theory of Computing (STOC), pages
232 -- 240, 2006.
[46] M. Ruzi´c. Constructing efficient dictionaries in close to sorting time. In Proc.
35th International Colloquium on Automata, Languages and Programming
(ICALP), pages 84 -- 95, 2008.
[47] W. Rytter. Application of Lempel-Ziv factorization to the approximation of
grammar-based compression. Theoretical Computer Science, 302(1-3):211 -- 222,
2003.
[48] H. Sakamoto. A fully linear-time approximation algorithm for grammar-based
compression. Journal of Discrete Algorithms, 3(24):416 -- 430, 2005.
21
[49] J. A. Storer and T. G. Szymanski. Data compression via textual substitution.
Journal of the ACM, 29(4):928 -- 951, 1982.
[50] P. Weiner.
Linear pattern matching algorithms.
In Proc. 14th Annual
Symposium on Switching and Automata Theory, pages 1 -- 11, 1973.
[51] D. E. Willard. Examining computational geometry, van Emde Boas trees, and
hashing from the perspective of the fusion tree. SIAM Journal on Computing,
29(3):1030 -- 1049, 2000.
22
|
1512.09002 | 1 | 1512 | 2015-12-30T16:36:17 | The Bounded Edge Coloring Problem and Offline Crossbar Scheduling | [
"cs.DS"
] | This paper introduces a variant of the classical edge coloring problem in graphs that can be applied to an offline scheduling problem for crossbar switches. We show that the problem is NP-complete, develop three lower bounds bounds on the optimal solution value and evaluate the performance of several approximation algorithms, both analytically and experimentally. We show how to approximate an optimal solution with a worst-case performance ratio of $3/2$ and our experimental results demonstrate that the best algorithms produce results that very closely track a lower bound. | cs.DS | cs | The Bounded Edge Coloring Problem
and Offline Crossbar Scheduling
Jonathan Turner
WUCSE-2015-07
Abstract
This paper introduces a variant of the classical edge coloring prob-
lem in graphs that can be applied to an offline scheduling problem for
crossbar switches. We show that the problem is NP-complete, develop
three lower bounds bounds on the optimal solution value and evaluate
the performance of several approximation algorithms, both analytically
and experimentally. We show how to approximate an optimal solution
with a worst-case performance ratio of 3/2 and our experimental results
demonstrate that the best algorithms produce results that very closely
track a lower bound.
1
v
2
0
0
9
0
.
2
1
5
1
:
v
i
X
r
a
1
Introduction
An instance of the bounded edge coloring problem is an undirected graph
G = (V, E) with a positive integer bound b(e) for each edge e. A bounded
edge coloring is a function c, from the edges to the positive integers, with
c(e) ≥ b(e) for all edges e and c(e1) (cid:54)= c(e2) for all edge pairs e1 and e2
that have an endpoint in common. The objective of the problem is to find
a coloring in which the largest color is as small as possible. An example is
shown in Figure 1; the first number labelling each edge is its bound, while
the second is a valid color.
In this paper, we focus on a restricted version of the problem in which
the graph is bipartite, with the vertices divided between inputs and outputs.
1
Figure 1: Example of bounded edge coloring
We also are usually interested in graphs where the bounds on the edges
incident to each input are unique; we refer to this as the unique input bounds
condition. An example of a graph that satisfies the unique input bounds
condition is shown in Figure 2. The right side of the figure shows a tabular
representation of the graph, where each row corresponds to an input, each
column corresponds to an output and each integer denotes an edge and its
bound. Note that this graph requires colors [1 . . . 5].
The bounded edge coloring problem is an abstraction of an offline version
of the crossbar scheduling problem. In this application, the graph's vertices
represent the inputs and outputs of a crossbar switch, while the edges rep-
resent packets to be transferred from inputs to outputs. The edge bounds
represent the arrival times of the packets and the colors represent the times
at which packets are transferred from inputs to outputs. Since an input can
only receive one packet at each time step, the inputs naturally satisfy the
unique input bounds condition. The objective of the problem is to transfer
all packets across the crossbar in the smallest possible amount of time.
The crossbar scheduling problem has been studied extensively in the on-
line context, using several distinct performance criteria. For so-called input
queued switches, the focus has been on ensuring bounded waiting times in
systems subjected to random input traffic. Many scheduling algorithms have
been shown to meet this objective [6, 7, 8]. More compelling, worst-case re-
sults have been shown for combined input and output queued switches, in
which the crossbar is capable of transferring packets somewhat faster than
2
a c e b d 2,3 2,2 1,1 1,2 2,4 1,1 3,3 Figure 2: Bipartite instance that satisfies the unique input bounds condition
they can arrive at the inputs, or be transmitted from the outputs. Some
scheduling algorithms can match the performance of an idealized output-
queued switch when the crossbar is twice as fast as the inputs and out-
puts [1, 4, 5].
In section 2, we show that the bounded edge coloring problem is NP-
complete.
In section 3 we derive several lower bounds on the number of
colors required, and introduce a class of graphs which require substantially
more colors than implied by the weaker of our lower bounds. In section 4,
we describe several algorithms and establish worst-case performance ratios
for some. Experimental performance results appear in section 5.
2 Complexity of Bounded Edge Group Coloring
For the ordinary edge coloring problem, we can color any bipartite graph with
colors [1 . . . ∆], where ∆ is the maximum vertex degree [2]. Unfortunately,
the bounded edge coloring problem is NP-complete. We can show this, using
a reduction from the partial edge coloring completion problem for bipartite
graphs.
In this problem, we are given a bipartite graph with some of its
edges colored, and are asked to complete the coloring using no color larger
than a given integer k. This is a generalization of the partial latin squares
completion problem shown to be NP-complete in [3].
3
w x y z a 1 3 4 b 3 2 c 4 3 2 1 d 1 2 a c w b d 1 3 4 1 x y z 3 4 1 2 2 2 3 Figure 3: Reduction from edge coloring completion to bounded edge coloring
Let G = (V, E) be a bipartite graph, with partial coloring c, using colors
1, . . . , k. We construct a second bipartite graph H = (W, F ) where W in-
cludes all vertices in V , plus some additional vertices to be specified shortly.
Each uncolored edge in E is included in F and is assigned a lower bound of
1. Each edge in E that is colored k is included in F and assigned a lower
bound of k. For each edge in E that is colored k − 1, we include a chain
of three edges, with the inner edge assigned a lower bound of k, while the
outer two edges are assigned lower bounds of k − 1. For each edge in E
that is colored k − i, for i > 1, we include a similar component with two
additional vertices, two "outer" edges and i parallel edges joining the two
added vertices. The outer edges are assigned lower bounds of k− i, while the
inner edges are assigned lower bounds of k− i + 1, . . . , k. The construction of
the chains guarantees that in any valid k-coloring of H, the outer edges are
assigned the same color that the corresponding edge was assigned in G. The
uncolored edges in G are free to use any color in 1, . . . , k. This construction
is illustrated in Figure 3 for the case of k = 3. It's straightforward to show
that the partial coloring of G can be completed using colors 1, . . . , k if and
only if H can be colored using colors 1, . . . , k.
Unfortunately, while this reduction does show that bounded edge color-
ing is NP-complete, the graph H does not satisfy the unique input bounds
condition. To show that instances that satisfy this condition are also hard
to solve, we need to use a more elaborate reduction. The new reduction
4
a c x b 1 y 3 z 2 a c x b 1 y 3 z 2 2 3 1 3 2 1 1 1 1 starts with the graph H constructed above, and then adds 2k to the lower
bounds specified earlier for the components constructed to handle the pre-
colored edges in the original partial edge coloring instance. For edges that
are uncolored in G and incident to an input u, the corresponding edges in
H are assigned distinct lower bounds in the range k + 1, . . . , 2k. We then
attach an additional component to u that has the effect of forcing these edges
to have colors larger than 2k. This construction is illustrated in Figure 4.
This component has two groups of parallel edges joining a pair of vertices
Figure 4: Additional component for each input
to the original input u; these edge groups are each assigned lower bounds of
1, . . . , k. There are two other groups joining additional leaf vertices to the
first two vertices. One group of "leaf edges" has lower bounds of k +1, . . . , 3k
and must be assigned colors equal to their bounds in any valid coloring with
a maximum color of 3k. The other has bounds 2k + 1, . . . , 3k and must also
be assigned colors equal to their bounds in any valid coloring with a maxi-
mum color of 3k. This means that the k parallel edges incident to u must
be assigned colors k + 1, . . . , 2k and this is in turn forces the edges that were
uncolored in G to have colors larger than 2k in the modified version of H.
Figure 5 shows the final version of H, for the example in Figure 3. Observe
that all input vertices in the final version of H do in fact have unique lower
bounds. It's straightforward to show that the final version of H can be col-
ored using colors 1, . . . , 3k if and only if G's coloring can be completed using
colors 1, . . . , k.
5
3k u k+1 1 k 2k+1...3k k+1 ... ... 1 k ... k+i ... u 1 1 ... i unconstrained edges Figure 5: Extension of example in Figure 3
3 Lower Bounds
In this section, we present several methods to compute lower bounds on
the maximum edge color used by an instance of the bounded edge-coloring
problem. The first is referred to as the degree bound. Before describing it,
we need a few definitions. For any vertex u in a graph G, let δG(u) denote the
number of edges incident to u (the vertex degree) and let ∆G = maxu δG(u).
If G is an instance of the bounded edge coloring problem, we let Gk be
the subgraph of G containing edges with bounds ≥ k and we let Gk be the
subgraph containing edges with bounds ≤ k.
Observe that if a vertex u has d edges with bounds ≥ c, then some edge
incident to u must be assigned a color ≥ c + d − 1. Consequently, for any k,
the number of colors needed to color Gk is at least k + ∆Gk − 1. The degree
bound for G is denoted D(G) and defined by D(G) = maxk k + ∆Gk − 1.
For the graph in Figure 2, D = 5 and this graph can be colored using colors
1, . . . , 5. Some graphs require more than D colors. Figure 6 shows a graph
6
a c x b 7 y 9 z 8 8 9 7 8 4 5 4 4 9 1 3 2 1 3 2 4 6 5 7 9 8 7 9 8 1 3 2 1 3 2 4 6 5 7 9 8 7 9 8 1 3 2 1 3 2 4 6 5 7 9 8 7 9 8 Figure 6: D colors are not always enough (top table shows edge bounds,
bottom shows colors in an optimum coloring)
(in the tablular format) with D = 4 that requires five colors (an optimal
coloring is shown in the second table).
The graph in Figure 6 is actually a special case of a class of graphs that
require substantially more than D colors. The graph Bn has inputs u1, . . . , un
and outputs v1, . . . , v2n−1. For 1 ≤ i ≤ n, 1 ≤ j ≤ i, there is an edge (ui, vj)
with bound j, and for each 1 ≤ i < n, i < j ≤ n, there is an edge (ui, vn+i)
with bound j. Figure 7 shows the case of B7, along with a coloring using
nine colors. Note that in general, DBn = n.
Our second lower bound is obtained by computing a sequence of match-
ings. Observe that for any integer k, the edges of G that are colored k must
form a matching in Gk. If we let mk be the number of edges in a maximum
size matching of Gk, then m1 + m2 + ··· + mk is an upper bound on the
number of edges that can be assigned colors in 1, . . . , k. If this sum is less
than the number of edges in G, then G requires more than k colors. So,
7
e f g h i j k a 1 2,3,4 b 1 2 3,4 c 1 2 3 4 d 1 2 3 4 e f g h i j k a 5 2,3,4 b 1 5 3,4 c 2 3 5 4 d 4 2 3 5 Figure 7: Graph B7 and a valid edge coloring using colors 1, . . . , 9
if we let M (G) be the smallest integer k for which m1 + m2 + ··· + mk is
greater than or equal to the number of edges in G, then M (G) is a lower
bound on the number of colors required to color G. We refer to M as the
matching bound. For the graph B4 in Figure 6, the sequence of matching
sizes is 1, 3, 4, 4, 4 and consequently, M = 5. For the graph B7 in Figure 7,
the sequence of matching sizes is 1, 3, 5, 7, 7, 7, 7, 7, 7 and M = 9. It's not
difficult to show that in general, the matching bound for Bn is (cid:100)5n/4(cid:101).
Now, we turn to our third lower bound. Let G = (V, E) be a bipartite
graph with edge bounds b(e) and let c be a valid coloring of G using colors
1, . . . , C. Let Hk be the subgraph of G that is colored using colors 1, . . . , k
8
A B C D E F G H I J K L M a 1 2-7 b 1 2 3-7 c 1 2 3 4-7 d 1 2 3 4 5-7 e 1 2 3 4 5 6-7 f 1 2 3 4 5 6 7 g 1 2 3 4 5 6 7 A B C D E F G H I J K L M a 9 2-7 b 8 9 3-7 c 2 8 9 4-7 d 1 2 8 9 5-7 e 3 4 5 8 9 6-7 f 6 3 4 5 8 9 7 g 7 6 3 4 5 8 9 and let Jk be the subgraph of G that is colored using colors k + 1, . . . , C.
Note that Hk is a subgraph of Gk, DHk ≤ k and ∆Jk ≤ C − k. So, for any
k, it is possible to split G into two subgraphs that have these properties.
Now, let C(cid:48) be an integer smaller than C and note that if there is some k,
for which we cannot split G into subgraphs Hk and Jk with Hk a subgraph
of Gk, DHk ≤ k and ∆Jk ≤ C(cid:48) − k, then G cannot be colored using only
colors 1, . . . , C(cid:48). This leads to a lower bound on the number of colors needed
to color a given graph.
To make this bound useful, we need an efficient way to partition G into
subgraphs Hk and Jk for given integers k and C. This can be done by
solving a network flow problem. Let Fk,C be a flow graph that includes a
source vertex s, a sink vertex t and a chain of vertices for each vertex in
Gk. Specifically, for each input u in Gk, Fk,C contains a chain consisting of
vertices ui for each i ∈ {1, . . . , k} and edges (ui, ui+1) with capacity k − i.
There is also an edge from s to u1 with capacity k. For each output v in
Gk, Fk,C contains a chain consisting of vertices vi for each i ∈ {1, . . . , k} and
edges (vi+1, vi) with capacity k − i. There is also an edge from v1 to t with
capacity k. For each edge (u, v) in Gk with bound i, F contains an edge
(ui, vi) of capacity 1. We refer to this last set of edges as the core edges of
Fk,C. Observe that the subset of the core edges that have positive capacity
in any integer flow on Fk,C correspond to a subgraph of Gk that has a degree
bound D that is no larger than k. To complete the construction of Fk,C,
we specify minimum flow requirements for the edges incident to s and t. In
particular, for input u of G, the edge (s, u1) is assigned a minimum flow of
min{0, δG(u) − (C − k)}. Similarly, for output v of G, the edge (v1, t) is
assigned a minimum flow of min{0, δG(v) − (C − k)}. Given an integer flow
on Fk that satisfies the minimum flow requirements, we define Hk to consist
of those edges in G that correspond to core edges that have positive flow.
We define Jk to include the remaining edges in G. It is straightforward to
show that Hk is a subgraph of Gk, DHk ≤ k and ∆Jk ≤ C − k. If there is
no integer flow on Fk that satisfies the minimum flow requirements, then G
cannot be colored using only colors 1, . . . , C.
Figure 8 shows the graph B4 and the corresponding flow graph F2,5.
There is an integer flow for F2,5 that uses the edges that are emphasized in
9
Figure 8: Graph B4 and corresponding flow graph F2,5
"bold" and satisfies all the minimum flow requirements. The edges of B4
that correspond to the bold core edges are also emphasized in bold. These
edges define the graph H2, while the remaining edges define J2. Note that
DH2 = 2 and ∆J2 = 3.
The flow bound for G is denoted by PG and is defined as the smallest value
of C for which Fk,C has a flow that satisfies the minimum flow requirements,
for all values of k ∈ [1, bmax], where bmax is the largest edge bound in G.
The flow bound for B8 is 11, while the matching bound is 10. This gap
increases for larger graphs. For example, B64 has a matching bound of 80
and a flow bound of 83, while B256 has a matching bound of 320 and a flow
bound of 331.
We close this section by describing a general method for coloring graphs
Bn using colors 1, . . . , n + (cid:100)(n − 1)/3(cid:101). For B64 the largest color is 85, for
B256 the largest color is 341. Figure 9 illustrates the method for coloring
Bn. The top part of the figure shows the coloring given earlier for B7, with
several regions of the table highlighted. Note that each of the four highlighted
regions uses a distinct set of colors and each color used is repeated along a
10
c A b d 1 3 B C D 2 4 1 E F G 1 1 2 2 3 3 2 4 a a1 c1 A1 b1 d1 B1 1,0 E1 1,0 1,0 1,0 a2 c2 A2 b2 d2 B2 E2 s t 1,0 2,1 2,1 2,1 2,1 2,1 2,0 2,0 1,0 1,0 1,0 1,0 1,0 1,0 1,0 1,0 1,0 1,0 capacity, min flow 4 3 4 Figure 9: General method for coloring Bn
diagonal within the region. The edges incident to the outputs H, . . . , M
are assigned colors equal to their bounds. The bottom part of the figure
shows how the colors of the edges incident to the first n outputs are assigned
in the general case, using a parameter x = (cid:100)(n − 1)/3(cid:101). Again, we divide
the table definining the edges into four regions and assign disjoint sets of
colors to those regions. Within each region, colors are used repeatedly along
diagonals. The edges incident to the last n − 1 outputs are assigned colors
equal to their bounds. The choice of x ensures that the assignment of colors
to edges yields a legal coloring of the graph that respects all the edge bounds.
11
n+1,n+x 1,x x+1,2x+1 2x+2,n A B C D E F G H I J K L M a 9 2-7 b 8 9 3-7 c 2 8 9 4-7 d 1 2 8 9 5-7 e 3 4 5 8 9 6-7 f 6 3 4 5 8 9 7 g 7 6 3 4 5 8 9 4 Approximate algorithms
In this section, we describe several approximate algorithms for the bounded
edge coloring problem and analyze their worst-case performance. Let G =
(V, E) be a graph with edge bounds b(e) and let C∗(G) be the largest color
value used by an optimal solution.
We start with a very simple method that produces solutions with a max-
imum color ≤ 2C∗. Let bmax be the largest color bound used by an instance
of the bounded edge coloring problem and let ∆ be the maximum vertex de-
gree. Using any method for the ordinary edge coloring problem, we can color
the edges using colors bmax, . . . , bmax+∆−1 and since C∗ ≥ max{bmax, ∆},
the largest color used by this solution is < 2C∗.
We can get a better approximation by first spliting G into two subgraphs
and coloring each of the subgraphs separately. Consider an optimal coloring
of G and let Hk be the subgraph defined by the edges with colors ≤ k.
Note that Hk is a subgraph of Gk and that ∆Hk ≤ k. If we let Jk be the
subgraph of G defined by the edges with colors > k, then ∆Jk ≤ C∗ − k. So
an optimal coloring must be divisible into a pair of subgraphs that satisfy
these inequalities. Note that we can color Hk using colors k, . . . , 2k− 1 using
any algorithm for the ordinary edge coloring problem. We can also color Jk
using colors bmax, . . . , bmax + C∗ − (k + 1). So long as 2k − 1 < bmax, these
two sets of colors do not overlap meaning that we can color the entire graph
using colors k, . . . , bmax + C∗ − (k + 1). If we let k = (cid:98)bmax/2(cid:99), the largest
color is ≤ (bmax/2) + C∗ ≤ (3/2)C∗.
In order to construct an approximation algorithm based on this obser-
vation, we need a way to split G into subgraphs Hk and Jk. This can be
done by solving the same flow problem that was used in the flow lower bound
computation. Alternatively, we can simplify the flow problem by collapsing
each of the input-side chains u1, . . . , uk into a single vertex u and each of the
output-side chains v1, . . . , vk into a single vertex v. Thus, we can efficiently
color any graph G using colors ≤ (3/2)C∗. We'll refer to this as the splitting
method.
Our next algorithm is a simple greedy algorithm that repeats the follow-
ing step until all edges are colored.
12
Select an edge e = (u, v) and let c be the smallest color that is at least
as large as b(e) and is not yet in use at both u and v.
We select edges that are incident to a vertex of maximum degree in the
uncolored subgraph. While we have no worst-case-performance bound for
this algorithm, in pactice it out-performs the splitting method, as we will see
in the next section.
Now, we consider an algorithm based on the classical augmenting path
algorithm for the ordinary edge coloring problem. For the bounded edge
coloring problem, a path p is an ij-augmenting path if its edges alternate
in color between i and j, it cannot be extended any further at either of its
endpoints and every edge in the path has a bound that is ≤ min{i, j}. The
augmenting path algorithm for the bounded edge coloring problem colors the
edges by repeatingly selecting an edge e = (u, v) and then applying the first
case from the following list that applies.
• If there is some eligible color that is unused at both endpoints, color e
with one such color.
• If there are eligible colors i and j, where i is available at u and j is
available at v, and there is an ij-augmenting path starting at v, then
reverse the colors of the edges on the path and and let c(e) = i.
• If there are eligible colors i and j, where i is available at u and j is
available at v, and there is a ji-augmenting path starting at u, then
reverse the colors of the edges on the path and let c(e) = j.
• Allocate a new eligible color and use it to color e.
Initially, colors 1, . . . , bmax are eligible. New colors are allocated sequentially,
as needed. We select edges that are incident to a vertex of maximum degree
in the uncolored subgraph. When selecting colors, we give preference to
colors with smaller values. In the first case, we select the smallest eligible
color that is unused at both endpoints. In the second and third cases, we
select color pairs i and j that minimize max{i, j}.
Since the augmenting path algorithm for the ordinary edge-coloring prob-
lem uses ∆ colors, this version colors uses no color larger than bmax + ∆− 1.
13
We can also use it to color the subgraphs in the splitting method to obtain
an algorithm that uses no color larger (bmax/2) + C∗.
Next, we consider algorithms based on constructing a series of match-
ings. The first such algorithm starts by initializing k = 1 then repeating the
following step so long as there are uncolored edges.
Find a maximum size matching on the uncolored edges in Gk, assign
color k to all edges in the matching, then increment k.
We refer to this as the maximum size matching algorithm. We can improve
it by selecting matchings that maximize the number of matched vertices that
have maximum degree in the uncolored subgraph. This can be done using an
algorithm described in [10]. This version of the matching algorithm is called
the maximum degree matching algorithm. Reference [10] also shows how to
find maximum size matchings that maximize a general priority score based
on arbitrary integer priorities assigned to the vertices. Our third matching
algorithm uses this method. Priorities are assigned in decreasing order of
vertex degree in the uncolored subgraph. So, vertices of maximum degree are
assigned priority 1, those with the next largest degree are assigned priority
2 and so forth. We refer to this version of the matching algorithm as the
priority matching algorithm. The last two of the matching algorithms use no
color larger bmax + ∆ − 1. Also, like the augmenting path algorithm, they
can be used with the splitting method to obtain an algorithm that uses no
color larger (bmax/2) + C∗.
5 Experimental Evaluation
In this section, we evaluate the performance of the algorithms introduced
in the last section, experimentally. We start, by evaluating the performance
using random graphs that are generated using the following procedure.
• Generate a random regular bipartite graph, with n inputs, n outputs
and all vertices having degree ∆.
• At each input, assign the incident edges a unique random color in
1, . . . , bmax, where bmax ≥ ∆
14
Figure 10: Performance results for random instances
This produces a random problem instance that satisfies the unique input
bounds condition. Figure 10 shows results for graphs on 100 vertices where
the vertex degree is varied from 5 to 50 and bmax = ∆+3. The y-axis displays
the difference between the maximum color used and the vertex degree. Each
data point shows the average from ten random problem instances. Error bars
have been omitted for clarity, but the relative error was generally less than
2%.
The lowest curve shows the results for the flow lower bound and the
priority matching algorithm (for these graphs, the priority matching algo-
rithm always produced results equal to the bound). The maximum degree
matching algorithm was almost identical to the priority matching algorithm,
but did occasionally exhibit small differences. The maximum size matching
algorithm was not quite as good as the other two. The greedy and augment-
ing path algorithms also produced perfectly respectable results, generally
exceeding the largest color used by the priority matching algorithm by less
than 5%.
Figure 11 shows performance results for the graphs Bn. Here the max
degree and priority match algorithms generally use just 2 or 3 more colors
than the number given by the flow lower bound. The maximum size match-
15
split priority match and flow lower bound max degree match max size match greedy augmenting path random n=100 Figure 11: Performance results for Bn
ing algorithm performs less well in this case, actually under-performing the
augmenting path algorithm.
The algorithms described here are all available as part of an open-source
library of graph algorithms and data structures [9].
6 Closing Remarks
The bounded edge-coloring problem is a natural generalization of the ordi-
nary edge-coloring problem and interesting in its own right, independent of
its application to the crossbar scheduling problem. There is a considerable
gap between the best worst-case performance ratio achieved by our algo-
rithms and the experimental performance measurements. Closing that gap
is the main open problem to be addressed.
It's worth noting that most of our algorithms can be applied to general
graphs, as well as bipartite graphs. It would be interesting to understand how
they perform in this context. Unfortunately, the flow lower bound cannot be
applied to general graphs, although the degree bound and matching bound
can be.
While our results do not apply directly to the online version of the cross-
16
split flow lower bound max degree match and priority match max size match greedy augmenting path bad case bar scheduling problem, there is potential for extending them to make them
more applicable. For example, one can model systems in which the crossbars
operate at faster speeds than the inputs and outputs by restricting the values
allowed as edge bounds. This can be used to derive a lower bound on the
"speedup ratio" needed to match the performance of an ideal output-queued
switch.
References
[1] Attiya, H., D. Hay and I. Keslassy. "Packet-Mode Emulation of Output-
Queued Switches," Proc. of ACM SPAA, 2006.
[2] Bondy, J. A. and U. S. R. Murty. Graph Theory and Applications,
Elsevier North Holland, 1976.
[3] Colburn, Charles. "The complexity of completing partial latin squares,"
Discrete Applied Mathematics 8 (1984), pp. 25-30.
[4] Chuang, S.-T. A. Goel, N. McKeown, B. Prabhakar "Matching output
queueing with a combined input output queued switch," IEEE J. on
Selected Areas in Communications, 12/1999.
[5] Krishna, P., N. Patel, A. Charny and R. Simcoe. "On the speedup
required for work-conserving crossbar switches," IEEE J. Selected Areas
of Communications, 6/1999.
[6] Leonardi, E., M. Mellia, F. Neri, and M.A. Marsan, "On the stability
of input-queued switches with speed-up," IEEE/ACM Transactions on
Networking, Vol. 9, No. 1, pp. 104 -- 118, 2/2001.
[7] McKeown, N. "iSLIP: a scheduling algorithm for
input-queued
switches," IEEE Trans. on Networking, 4/1999.
[8] McKeown, N., A. Mekkittikul, V. Anantharam, and J. Walrand.
"Achieving 100% Throughput in an Input-Queued Switch," IEEE Trans.
on Communications, Vol. 47, No. 8, 8/1999.
17
[9] Turner, Jonathan. "Grafalgo - a library of graph algorithms and sup-
porting data structures," Washington University Department of Com-
puter Science and Engineering, wucse-2015-01, 1/2015.
[10] Turner, Jonathan. "Maximum priority matchings," Washington Univer-
sity Department of Computer Science and Engineering, wucse-2015-
06, 11/2015.
18
|
1711.00599 | 1 | 1711 | 2017-11-02T02:46:56 | Optimal Parametric Search for Path and Tree Partitioning | [
"cs.DS"
] | We present linear-time algorithms for partitioning a path or a tree with weights on the vertices by removing $k$ edges to maximize the minimum-weight component. We also use the same framework to partition a path with weight on the vertices, removing $k$ edges to minimize the maximum-weight component. The algorithms use the parametric search paradigm, testing candidate values until an optimum is found while simultaneously reducing the running time needed for each test. For path-partitioning, the algorithm employs a synthetic weighting scheme that results in a constant fraction reduction in running time after each test. For tree-partitioning, our dual-pronged strategy makes progress no matter what the specific structure of our tree is. | cs.DS | cs | Optimal Parametric Search for Path and Tree
Partitioning
Greg N. Frederickson∗
Samson Zhou†
November 3, 2017
Abstract. We present linear-time algorithms for partitioning a path or a tree with weights
on the vertices by removing k edges to maximize the minimum-weight component. We also
use the same framework to partition a path with weight on the vertices, removing k edges
to minimize the maximum-weight component. The algorithms use the parametric search
paradigm, testing candidate values until an optimum is found while simultaneously reducing
the running time needed for each test. For path-partitioning, the algorithm employs a syn-
thetic weighting scheme that results in a constant fraction reduction in running time after
each test. For tree-partitioning, our dual-pronged strategy makes progress no matter what
the specific structure of our tree is.
Key words and phrases. Adaptive algorithms, data structures, max-min, min-max, para-
metric search, partial order, path partitioning, prune-and-search, sorted matrices, tree par-
titioning.
7
1
0
2
v
o
N
2
]
S
D
.
s
c
[
1
v
9
9
5
0
0
.
1
1
7
1
:
v
i
X
r
a
∗Supported in part by the National Science Foundation under grants CCR-86202271 and CCR-9001241,
†Research supported by NSF CCF-1649515. Email: [email protected]
and by the Office of Naval Research under contract N00014-86-K-0689.
1
1
Introduction
Parametric search is a powerful technique for solving various optimization problems [1, 4, 5,
6, 9, 10, 11, 13, 15, 16, 17]. To solve a given problem by parametric search, one must find
a threshold value, namely a largest (or smallest) value that passes a certain feasibility test.
In general, we must identify explicitly or implicitly a set of candidate values, and search
the set, choosing one value at a time upon which to test feasibility. As a result of the test,
we discard from further consideration various values in the set. Eventually, the search will
converge on the optimal value.
In [13], Megiddo emphasized the role that parallel algorithms can play in the implemen-
tation of such a technique. In [5], Cole improved the time bounds for a number of problems
addressed in [13], abstracting two techniques upon which parametric search can be based:
sorting and performing independent binary searches. Despite the clever ideas elaborated in
these papers, none of the algorithms presented in [13, 5] is known to be optimal. In fact,
rarely have algorithms for parametric search problems been shown to be optimal. (For one
that is optimal, see [6].)
In at least some cases, the time required for feasibility testing
matches known lower bounds for the parametric search problem. But in the worst case,
Ω(log n) values must be tested for feasibility, where n is the size of the input. Is this extra
factor of at least log n necessary? For several parametric search problems on paths and
trees, we show that a polylogarithmic penalty in the running time can be avoided, and give
linear-time (and hence optimal) algorithms for these problems.
We consider the max-min tree k-partitioning problem [14]. Let T be a tree with n
vertices and a nonnegative weight associated with each vertex. Let k < n be a given positive
integer. The problem is to delete k edges in the tree so as to maximize the weight of the
lightest of the resulting subtrees. Perl and Schach introduced an algorithm that runs in
O(k2rd(T ) + kn) time, where rd(T ) is the radius of the tree [14]. Megiddo and Cole first
1
considered the special case in which the tree is a path, and gave O(n(log n)2)-time and
O(n log n)-time algorithms, resp., for this case.1 (Megiddo noted that an O(n log n)-time
algorithm is possible, using ideas from [9].) For the problem on a tree, Megiddo presented
an O(n(log n)3)-time algorithm [13], and Cole presented an O(n(log n)2)-time algorithm [5].
The algorithm we present here runs in O(n) time, which is clearly optimal.
A closely related problem is the min-max path k-partitioning problem [2], for which we
must delete k edges in the path so as to minimize the weight of the heaviest of the resulting
subpaths. Our techniques yield linear-time algorithms for both versions of path-partitioning
and for max-min tree k-partitioning problem. Our approach is less convoluted than that
in [7, 8], and thus its analysis should be more amenable to independent confirmation and
publication. We believe that this paper provides for the full acknowledgement of linear time
algorithms for these particular problems.
Our results are a contribution to parametric search in several ways beyond merely pro-
ducing optimal algorithms for path and tree partitioning. In contrast to [8], we introduce
synthetic weights on candidate values or groups of candidate values from a certain implicit
matrix. We use weighted selection to resolve the values representing many shorter paths
before the values for longer paths, enabling future feasibility tests to run faster. We also use
unweighted selection to reduce the size of the set of candidate values so that the time for
selecting test values does not become an obstacle to achieving linear time.
Our parametric search on trees reduces feasibility test time across subpaths, while simul-
taneously also pruning paths that have no paths that branch off of them. To demonstrate
progress, we identify an effective measure of current problem size. Surprisingly, the num-
ber of vertices remaining in the tree seems not to be so helpful, as is also the case for the
number of candidate values. Instead, we show that the time to perform a feasibility test
actually is effective. Our analysis for tree partitioning captures the total progress from the
1All logarithms are to the base 2.
2
beginning of the algorithm, not necessarily between consecutive rounds, as was the case for
the path-partitioning problem. Thus it seems necessary to have a dual-pronged strategy to
simultaneously compress the search structure of paths and also prune paths that end with a
leaf.
We organize our paper as follows. In Section 2 we discuss features of parametric search,
and lay the foundation for the optimal algorithms that we describe in subsequent sections.
In Section 3 we present our approach for solving both the min-max and the max-min par-
titioning problem on a path. In Section 4 we build on the results in Section 3 and present
our approach for optimally solving the max-min partitioning problem on a tree.
2 Paradigms for parametric search
In this section we discuss features of parametric search, and lay the foundation for the
optimal algorithms that we give in subsequent sections. We first review straightforward
feasibility tests to test the feasibility of a search value in a tree. We then review how to
represent all possible search values of a path within a "sorted matrix". We next present a
general approach for search that uses values from a collection of sorted matrices. Finally,
we describe straightforward approaches for the path and the tree that are as good as any
algorithms in [13, 5].
We first describe straightforward feasibility tests for cutting edges in a tree so as to max-
imize the minimum weight of any resulting component (max-min problem). The feasibility
test takes a test value λ, and determines if at least k cuts can be placed in the tree such
that no component has weight less than λ. We take λ∗ to be the largest value that passes
the test.
A straightforward test for the max-min problem in a tree is given in [13], and the feasibility
test for the min-max problem is similar [12]. We focus first on max-min, using an algorithm
3
FTEST0, which starts by rooting the tree at a vertex of degree 1 and initializing the number
of cuts, numcut, to 0. It next calls explore(root) to explore the tree starting at the root.
When the exploration is complete, if numcut > k, then λ is a lower bound on λ∗, and
otherwise λ is an upper bound on λ∗. The procedure explore is:
proc explore (vertex v)
accum wgt(v) ← weight of v
for each child w of v do explore(w ) endfor
Adjust accum wgt(v) and numcut.
For the max-min problem, we adjust accum wgt(v) and numcut by:
for each child w of v do Add accum wgt(w) to accum wgt(v). endfor
if accum wgt(v) ≥ λ
then numcut ← numcut + 1
accum wgt(v) ← 0
endif
In all but one case, we cut an edge in the tree whenever we increment numcut by 1.
For the max-min problem, we generally cut the edge from the current vertex v to its parent
except for the last increment of numcut in FTEST0. This is because either v will be the
root and thus does not have a parent, or the fragment of the tree above v will have total
weight less than λ. By not adjusting numcut to reflect the actual number of cuts in this
case, we are able to state the threshold test for both versions of FTEST0 in precisely the
same form. The feasibility test takes constant time per vertex, and thus uses O(n) time.
In a parametric search problem, the search can be restricted to considering values from a
finite set of values. The desired value is the largest (in the case of a max-min problem) or the
smallest (in the case of a min-max problem) that passes the feasibility test. Let each value
in the finite set be called a candidate value. We next discuss a data structure that contains
all candidate values for problems on just a path. This structure is based on ideas in [9] and
[10]. Let a matrix be called a sorted matrix if for every row the values are in nondecreasing
4
order, and for every column the values are in nonincreasing order. (Note that for notational
convenience, this definition varies slightly from the definition in [9].) Let the vertices on the
path P be indexed from 1 to n.
The set of values that we need to search is the set of sums of weights of vertices from
i to j, for all pairs i ≤ j. We succinctly maintain this set of values, plus others, in a data
structure, the succinct description. For i = 0, 1,··· , n, let Ai be the sum of the weights of
vertices 1 to i. Note that for any pair i, j with i ≤ j, the sum of the weights from vertex
i to j is Aj − Ai−1. Let X1 be the sequence of sums A1, A2,··· , An, and let X2 be the
sequence of sums A0, A1,··· , An−1. Then sorted matrix M (P ) is the n× n Cartesian matrix
X1 − X2, where the ij-th entry is Aj − Ai−1. In determining the above, we can use proper
subtraction, in which, a − b gives max{a − b, 0}. Clearly, the values in any row of M (P ) are
in nondecreasing order, and the values in any column of M (P ) are in nonincreasing order.
Representing M (P ) explicitly would require Θ(n2) time and space. Thus, our data structure
succinctly represents M (P ) (and hence, P ) in O(n) space by the two vectors X1 and X2.
In general, our algorithm also needs to inspect specific subpaths of P . However, repeat-
edly copying subvectors of X1 and X2 can take more than linear time in total. On the
other hand, for a subpath Q of P , the corresponding matrix M (Q) is a submatrix of M (P ),
which we can recover from the succinct representation of P . Thus, our algorithm succinctly
represents M (Q) by the start and end indices of M (Q) within M (P ). In this way, we can
generate the values of M (Q) from the vectors X1 and X2 of M (P ), as well as the location
of the submatrix as given by the succinct representation of Q. Therefore, our algorithm
avoids needlessly recopying vectors and instead generates in O(n) total time the succinct
representations of all subpaths that it may inspect.
As an example, we show a vertex-weighted path P in Fig. 1, and its associated matrix
M (P ). We list the sequence X1 horizontally above M (P ), and the sequence X2 vertically to
the left of M (P ), in such a way that the ij-th element of M (P ) is beneath the j-th element
5
of X1 and to the right of the i-th element of X2.
We next describe the general form PARAM SEARCH of all of our searching algorithms.
It is related to, and uses some of the ideas in, algorithms found in [5], [9], [10], and [13]. By
specifying the specific subroutines INIT MAT, TEST VAL, UPDATE MAT, we will be able
to give four versions of PARAM SEARCH, namely PATH0, TREE0, PATH1, and TREE1.
In all four versions, PARAM SEARCH takes as its arguments an integer k and either a
vertex-weighted path or a tree. It initializes searching bounds λ1 and λ2, where λ1 < λ2.
The algorithm progressively narrows these bounds until they satisfy the following conditions
by the end of the algorithm. For a max-min problem, λ1 is the largest value that is feasible,
and λ2 is the smallest value that is not feasible. (For a min-max problem, λ1 is the largest
value that is not feasible, and λ2 is the smallest value that is feasible.)
PARAM SEARCH will use INIT MAT to initialize M, a collection of succinctly rep-
resented square sorted matrices, the union of whose values is the set of values that we
consider. PARAM SEARCH then performs a series of iterations. On each iteration, it will
use TEST VAL to identify and test a small number of values λ drawn from matrices in M.
Each value λ will be either the largest or the smallest element in some matrix in M. As a
result of the feasibility tests, UPDATE MAT updates M by deleting certain matrices from
M, dividing certain matrices into 4 submatrices, and inserting certain matrices into M.
Note that in initializing λ2 to ∞, we take ∞ to be any value greater than the total weight
of all vertices in the path or the tree.
Algorithm PARAM SEARCH
λ1 ← 0
λ2 ← ∞
INIT MAT
while M is not empty do
TEST VAL
UPDATE MAT
endwhile
6
Output λ1 and λ2.
/* For max-min, λ∗ will be the final λ1 and for min-max, λ∗ will be the final λ2. */
In the remainder of this section we describe simple algorithms PATH0 and TREE0 for
partitioning a path and a tree, resp. These algorithms match the time of the algorithms in
[5], and set the stage for the improved algorithms that we present in the next two sections,
in which we introduce data structures that enable faster feasibility tests and strategies that
prune the tree quickly. We first describe a simple approach to the max-min problem on a path
P . Below are the three routines PATH0 init mat, PATH0 test val, and PATH0 update mat.
For a max-min problem, if λ > λ1 and λ is feasible, then we reset λ1 to λ. Otherwise, if
λ < λ2 and λ is not feasible, then we reset λ2 to λ. Thus at termination, λ∗ = λ1. (For the
min-max problem, we reset λ2 to λ if λ is feasible, and λ1 to λ if λ is not feasible. Thus, at
termination, λ∗ = λ2.) On every iteration we split matrices of size greater than 1 × 1 into
four smaller submatrices. We assume that the dimension of each sorted matrix is a power of
2. If this is not the case, then we pad out the matrix logically with zeroes. In the following,
the notation (λ1, λ2) denotes the open interval of values between λ1 and λ2. We also use
R(P ) to denote the representatives whose values are within the interval (λ1, λ2).
PATH0 init mat:
Implicitly split sorted matrix M (P ) for the path P into four square submatrices.
Initialize M to be the set containing these four square submatrices.
PATH0 test val:
if each submatrix in M contains just 1 element
then Let R be the multiset of values in the submatrices.
else Let R be the multiset consisting of the smallest and the largest
element from each matrix in M.
endif
for two times do:
Let R(cid:48) be the subset of R that contains only values in the interval (λ1, λ2).
if R(cid:48) is not empty then
Select the median element λ in R(cid:48).
if FTEST0(P, k, λ) = "lower(cid:48)(cid:48) (i.e., k < numcuts)
then λ1 ← λ
7
else λ2 ← λ
endif
endif
endfor
PATH0 update mat:
Discard from M any matrix with no values in (λ1, λ2).
if each submatrix in M contains more than 1 element
then Split each submatrix M in M into four square submatrices,
discarding any resulting submatrix with no values in (λ1, λ2).
endif
The following lemma is similar in spirit to Lemma 5 in [9] and Theorem 2 in [10].
Lemma 2.1 Let P be a path of n > 2 vertices. The number of iterations needed by PATH0
is O(log n), and the total time of PATH0 exclusive of feasibility tests is O(n).
Proof We call the multiset of smallest and largest values from each submatrix in M
the representatives of M, and we call the subset of representatives that are in (λ1, λ2) the
unresolved representatives of M. For iteration i = 1, 2, . . . , log n − 1, let S(i) be the number
of submatrices in M, and U (i) be the number of unresolved representatives of M. We first
show that S(i) ≤ 7 ∗ 2i+1 − 4i − 8, and U (i) ≤ 3 ∗ 2i+3 − 8i − 14. We prove this by induction
on i. The basis is for i = 1. At the beginning of iteration 1, there are 4 submatrices in
M and 8 unresolved representatives of M. The first feasibility test resolves at least 4 of
these representatives, and the second feasibility test leaves at most 2 unresolved. At most
all 4 submatrices remain after discarding. Splitting the submatrices at the end of iteration
1 gives at most 16 submatrices, Note that for i = 1, S(i) ≤ 16 = 7 ∗ 21+1 − 4 − 8 and 32
representatives, at most 32−6 = 26 of which are unresolved. and U (i) ≤ 26 = 3∗16−8−14.
Thus the basis is proved.
For the induction step, i > 1. By the induction hypothesis S(i − 1) ≤ 7 ∗ 2i − 4i − 4 and
U (i−1) ≤ 3∗2i+2−8i−6. Let R(i−1) be the set of representatives of M at the end of iteration
8
i−1. Note that these elements fall on at most 2i+1−1 diagonals of M (P ), since each iteration
can only split the existing submatrices into at most four smaller submatrices. The feasibility
tests on iteration i will leave u ≤ (cid:98)U (i − 1)/4(cid:99) ≤ 3 ∗ 2i − 2i − 2 representatives unresolved.
Let dj be the number of elements of R(i−1) on the j-th diagonal that are unresolved, so that
(cid:80) dj = u. Except for possibly the submatrices with the largest and smallest representatives,
all other submatrices on the j-th diagonal have two representatives, each in the range (λ1, λ2).
Thus, there will be at most (cid:98)(dj + 2)/2(cid:99) submatrices whose representatives are on the j-th
2i+1−1 diagonals of M (P ), there are S(i)/4 ≤(cid:80)
diagonal and are not both at most λ1 and not both at least λ2. Then summing over at most
j(cid:98)(dj + 2)/2(cid:99) ≤ (u+2i+2−2)/2 submatrices
that cannot be discarded at the end of iteration i. There will be 2S(i)/4 − u ≤ 2i+2 − 2
representatives of these submatrices that are resolved. After quartering, there are S(i) ≤
4∗(u+2i+2−2)/2 ≤ 3∗2i+1−4i−4+2i+3−4 submatrices at the end of iteration i. Simplifying,
we have S(i) ≤ 7∗ 2i+1 − 4i− 8. After quartering the submatrices, the number of unresolved
representatives of submatrices in M will be U (i) = 2S(i) − (2S(i)/4 − u) = (3/2)S(i) + u
≤ (3/2) ∗ (7 ∗ 2i+1 − 4i − 8) + 3 ∗ 2i − 2i − 2. Simplifying, we have U (i) ≤ 3 ∗ 2i+3 − 8i − 14.
This concludes the proof by induction.
There will be at most log n − 1 iterations until all submatrices consist of single values.
At that point we will have S(log n− 1) ≤ 7∗ 2log n − 4 log n− 8. On each remaining iteration,
the number of elements in M will be at least quartered. Thus the total number of iterations
is O(log n). The work on iteration i, for i = 1, 2, . . . , log n − 1, will be O(S(i)) = O(2i).
Thus the total time on these iterations, exclusive of feasibility testing, will be O((cid:80)log n−1
2i)
= O(n). The total time on the remaining iterations will be O(2(7 ∗ 2log n − 4 log n − 8))
i=1
= O(n).
We illustrate PATH0 using path P as in Fig. 2, with k = 3. First we set λ1 to 0 and
λ2 to ∞. We initialize the set M to the set consisting of the four 4 × 4 submatrices of
the matrix M (P ). On the first iteration of the while-loop, R = {59, 3, 28, 0, 31, 0, 0, 0} and
9
R(cid:48) = {59, 3, 28, 31}. The median of this set is (28 + 31)/2 = 29.5. For λ = 29.5, no cuts
are required, so we reset λ2 to 29.5. Then we recompute R(cid:48) to be {3, 28}, whose median is
(3 + 28)/2 = 15.5. For λ = 15.5, 1 cut is required, so we reset λ2 to 15.5. We discard the
submatrix with all values less than or equal to λ1, leaving three submatrices. We quarter
these submatrices into 12 submatrices as shown in Fig. 2. Of these 12 submatrices, five have
all values too large, and two have all values too small. We discard them, leaving the five
submatrices pictured in Fig. 3.
On iteration 2, R = {17, 0, 27, 3, 11, 0, 16, 0, 15, 0}, and R(cid:48) = {3, 11, 15}. The median of
this set is 11. For λ = 11, 3 cuts are required, so we reset λ1 to 11. Then we recompute
R(cid:48) to be {15}, whose median is 15. For λ = 15, 2 cuts are required, so we reset λ2 to 15.
There are no submatrices with all values at least 15.5, and one submatrix with all values at
most 11, and we discard the latter. We quarter the remaining four submatrices, giving 16
submatrices of dimension 1 × 1, of which all but the one containing 12 are either too large
or too small. On iteration 3, R = {12}, R(cid:48) = {12}, and the median is 12. For λ = 12, 3
cuts are required, so we reset λ1 to 12. At this point, all values are discarded, so that the
revised R(cid:48) is empty, and a second selection is not performed on iteration 3. All submatrices
will be discarded from M, and PATH0 will terminate with λ1 = 12 and λ2 = 15, and output
λ∗ = 12.
Theorem 2.2 Algorithm PATH0 finds a max-min partition of a path of n weighted vertices
in O(n log n) time.
Proof Correctness follows from the correctness of FTEST0, from the fact that all possible
candidates for λ∗ are included in M (P ), and from the fact that each value discarded is either
at most λ1 or at least λ2.
By Lemma 2.1, PATH0 will take O(n) total time, exclusive of the feasibility tests, and
will produce a sequence of O(log n) values to be tested. It follows that the total time for all
10
feasibility tests is O(n log n).
The time for PATH0 corresponds to the times achieved by Megiddo and Cole for the
path problem. We show how to do better in Section 3.
We next describe a simple approach TREE0 to the max-min problem on a tree. We first
define an edge-path-partition of a tree rooted at a vertex of degree 1. Partition the edges of
the tree into paths, where a vertex is an endpoint of a path if and only if it is of degree not
equal to 2 with respect to the tree. Call any path in an edge-path-partition that contains a
leaf in the tree a leaf-path. As an example, consider the vertex-weighted tree in Fig. 4(a).
Fig. 4(b) shows the edge-path-partition for T . There are 7 paths in the partition, as shown.
Four of these paths are leaf-paths.
We now proceed with the approach for the tree. Below are the three routines TREE0 init mat,
TREE0 test val, and TREE0 update mat. The basic idea is to perform the search first on
the leaf-paths, and thus determine which edges in the leaf-paths should be cut. When no
search value on a leaf-path is contained in the open interval (λ1, λ2), we prune the tree and
repeat the process. We use the straightforward feasibility test described earlier.
The determination of cuts and pruning of the tree proceeds as follows. If T contains more
than one leaf, do the following. For each leaf-path Pj, infer the cuts in Pj such that each
component in turn going up in Pj, except the highest component on Pj, has total weight
as small as possible but greater than λ1. Delete all vertices beneath the top vertex of each
leaf-path Pj, and add to the weight of the top vertex in Pj the weight of the other vertices
in the highest component of Pj. This leaves a smaller tree in which all leaf-paths in the
original tree have been deleted. The smaller tree has at most half of the number of leaves of
T . Reset T to be the smaller tree, and k to be the number of cuts remaining to be made in
this tree.
TREE0 init mat:
Initialize T to be the tree rooted at a vertex of degree 1.
11
Concatenate the leaf-paths of T together, yielding path P (cid:48).
Split sorted matrix M (P (cid:48)) for the path P (cid:48) into 4 square submatrices.
Let M be the set containing these four square submatrices.
TREE0 test val:
(Identical to PATH0 ident val,
except that FTEST0 is called with argument T rather than P ).
TREE0 update mat:
Call PATH0 update mat on M.
if M is empty and T is not a path
then
for each leaf-path of T do
Determine cuts on the leaf-path, and decrease k accordingly.
Add to the weight of the top vertex the total weight of the
other vertices in the highest component of the leaf-path.
Delete from T all vertices in the leaf-path except the top vertex.
endfor
Concatenate the leaf-paths of T together, yielding new path P (cid:48).
Split sorted matrix M (P (cid:48)) for path P (cid:48) into 4 square submatrices.
Let M be the set containing these four square submatrices.
endif
As an example we consider the max-min problem on the tree shown in Fig. 4(a), with k =
3. In the initialization, four leaf-paths are identified and concatenated together, giving the
path P (cid:48) = 4, 5, 4, 5, 15, 6, 2, 1, 1, 3, 2. Once all search values associated with P (cid:48) are resolved,
λ1 = 10 and λ2 = 13. We place one cut between the vertices of weight 15 and 6, and reset k
to 2. We add the weights of the two leaves of weight 4 to the weight of their parent, giving
it weight 13. We add the weights of all descendants of the vertex with weight 2, except
the weight of 15, to the vertex of weight 2, giving it a weight also of 13. Then we delete
all edges in the leaf-paths. Fig. 5(a) shows the edge-path-partition of the resulting tree.
There are two leaf-paths in this partition, with vertex weights 13, 4, and 3 on one, and 13
and 3 on the other. We form the path P (cid:48) = 13, 4, 3, 13, 3. Once we have resolved all search
values associated with P (cid:48), we still have λ1 = 10 and λ2 = 13. We can then place two cuts,
above each of the vertices of weight 13, and reset k to 0. We add the weight of the vertex
12
of weight 4 to its parent, giving it weight 7. Then we delete all edges in the leaf-paths. The
edge-path-partition of the resulting tree is shown in Fig. 5(b). There is just one leaf-path
in this partition, with vertex weights 5 and 7. Once all search values associated with P (cid:48) are
resolved, we have λ1 = 12 and λ2 = 13. Since the tree consists of a single path, the algorithm
then terminates with λ∗ = 12.
Theorem 2.3 Algorithm TREE0 finds a max-min partition of a tree of n weighted vertices
in O(n(log n)2) time.
Proof For correctness of the tree algorithm, note that λ1 always corresponds to a feasible
value, that all possible values resulting from leaf paths are represented in M (P (cid:48)), and that
the cuts are inferred on leaf-paths assuming (correctly) that any subsequent value λ to be
tested will have λ > λ1.
We analyze the time as follows. By Lemma 2.1, resolving the path P (cid:48) will use O(log n)
feasibility tests, and time exclusive of feasibility tests of O(n). Since each feasibility test
takes O(n) time, the feasibility tests will use O(n log n) time. Since resolving the path P (cid:48)
will at least halve the number of leaves in the tree, the number of such paths until the
final version of P (cid:48) is O(log n). Thus the total time to partition the tree by this method is
O(n(log n)2).
The time for TREE0 beats the time of O(n(log n)3) for Megiddo's algorithm, and matches
the time of O(n(log n)2) for Cole's algorithm. We show how to do better for the max-min
problem in Section 4.
3 Partitioning a Path
In this section we present an optimal algorithm to perform parametric search on a graph
that is a path of n vertices. The algorithm follows the paradigm of parametric search by
13
repeatedly performing feasibility tests over potential optimal values, as it gathers information
so that subsequent feasibility tests can be performed increasingly faster. The improvement
in speed of the subsequent feasibility tests is enough so that the entire running time is linear.
Our discussion focuses on the max-min problem; at the end of the section we identify the
changes necessary for the min-max problem. Throughout this section we assume that the
vertices of any path or subpath are indexed in increasing order from the start to the end of
the path.
We first consider the running time of PATH0, to determine how the approach might be
accelerated. All activities except for feasibility testing use a total of O(n) time. Feasibility
testing can use a total of Θ(n log n) time in the worst case, since there will be Θ(log n) values
to be tested in the worst case, and each feasibility test takes Θ(n) time. It seems unlikely
that one can reduce the number of tests that need to be made, so then to design a linear-time
algorithm, it seems necessary to design a feasibility test that will quickly begin to take o(n)
time. We show how to realize such an approach.
We shall represent the path by a partition into subpaths, each of which can be searched in
time proportional to the logarithm of its length. Each such subpath will possess a property
that makes feasibility testing easier. Either the subpath will be singular or resolved. A
subpath P (cid:48) is singular if it consists of one vertex, and it is resolved if no value in M (P (cid:48)) falls
in the interval (λ1, λ2). If a subpath is resolved, then the position of any one cut determines
the positions of all other cuts in the subpath, irrespective of what value of λ we choose from
within (λ1, λ2).
If we have arranged suitable data structures when the subpath becomes
resolved, then we do not need a linear scan of it.
To make the representation simple, we restrict the subpaths in the partition to have
lengths that are powers of 2. Each subpath will consist of vertices whose indices are (j −
1)2i + 1, . . . , j 2i for integers j > 0 and i ≥ 0.
Initially the partition will consist of n
singular subpaths. Each nonsingular subpath will have i > 0. When introduced into the
14
partition, each such subpath will replace its two constituent subpaths, the first with indices
(2j − 2)2i−1 + 1, . . . , (2j − 1)2i−1, and the second with indices (2j − 1)2i−1 + 1, . . . , (2j)2i−1.
We represent the partition of path P into the subpaths with three arrays last[1..n],
ncut[1..n] and next[1..n]. Consider any subpath in the partition, with first vertex vf and
last vertex vt. Let vl be an arbitrary vertex in the subpath. The array last identifies the
end of a subpath, given the first vertex of a subpath. Thus last(l) = t if l = f and is
arbitrary otherwise. Given a cut on the subpath, the array next identifies, in constant time,
a cut further on in that subpath. The array ncut identifies the number of cuts skipped in
moving to that further cut. Let w(l, t) be the sum of the weights of vertices vl through vt. If
w(l, t) < λ2, then next(l) = null and ncut(l) = 0. Otherwise, next(l) ≥ l is the index of the
last vertex before a cut, given that l = 1 or vl is the first vertex after a cut. Then ncut(l) is
the number of cuts after vl up to and including the one following vnext(l). We assume that
the last cut on a subpath will leave a (possibly empty) subset of vertices of total weight less
than λ2. (Note that the last cut on the path as a whole must then be ignored.)
Given the partition into subpaths, we describe feasibility test FTEST1. Let λ be the
value to be tested, with λ1 < λ < λ2. For each subpath, we use binary search to find the
first cut, and then, follow next pointers and add ncut values to identify the number of cuts
on the subpath. When we follow a path of next pointers, we will compress this path. This
turns out to be a key operation as we consider subpath merging and its effect on feasibility
testing. (Note that the path compression makes FTEST1 a function with side effects.)
func FTEST1 (path P , integer k, real λ)
f ← 1
numcut ← −1; remainder ← 0
while f ≤ n do /* search the next subpath: */
t ← last(f )
if remainder + w(f, t) < λ
then remainder ← remainder + w(f, t)
else
15
numcut ← numcut + 1
Binary search for a smallest r so that w(f, r) + remainder ≥ λ.
if r < t
then
(s, sumcut) ← search next path (r, t)
numcut ← numcut + sumcut
compress next path (r, s, t, sumcut)
endif
remainder ← w(s + 1, t)
endif
f ← t + 1
endwhile
if numcut ≥ k then return("lower") else return("upper") endif
endfunc
search next path (vertex index l, t)
sumcut ← 0
while l < t and next(l + 1) (cid:54)= null
sumcut ← sumcut + ncut(l + 1)
l ← next(l + 1)
endwhile
return(l, sumcut)
compress next path (vertex index l, s, t, integer sumcut)
while l < t and next(l + 1) (cid:54)= null
sumcut ← sumcut − ncut(l + 1)
ncut(l + 1) ← ncut(l + 1) + sumcut
temp ← next(l + 1)
next(l + 1) ← s
l ← temp
endwhile
Use of this feasibility test by itself is not enough to guarantee a quick reduction in the
time for feasibility testing. This is because there is no assurance that the interval (λ1, λ2) will
be narrowed in a manner that allows longer subpaths to quickly replace shorter subpaths in
the partition of path P . To achieve this effect, we reorganize the positive values from M (P )
into submatrices that correspond in a natural way to subpaths. Furthermore, we associate
synthetic weights with these submatrices and use these weights in selecting the weighted
median for testing. The synthetic weights place a premium on resolving first the values from
16
submatrices corresponding to short subpaths. However, using synthetic weights could mean
that we consider many values of small weight repeatedly, causing the total time for selection
to exceed Θ(n). To offset this effect, we also find the unweighted median, and test this value.
This approach guarantees that we discard at least half of the submatrices' representatives
on each iteration, so that each submatrix inserted into M need be charged only a constant
for its share of the total work in selecting values to test.
We now proceed to a broad description of PATH1. The basic structure follows that of
PARAM SEARCH, in that there will be the routines PATH1 init mat, PATH1 test val, and
PATH1 update mat. However, we will slip the set-up and manipulation of the data structures
for the subpaths into the routines PATH1 init mat and PATH1 update mat. Let large(M )
be the largest element in submatrix M , and let small(M ) be the smallest element in M .
PATH1 init mat:
Initialize M to be empty.
Call mats for path(P, 1, n) to insert submatrices of M (P ) into M.
for l ← 1 to n do last(l) ← l; next(l) ← 0; ncut(l) ← 0 endfor
PATH1 test val:
R ← ∅
for each M in M do
if large(M ) < λ2
then Insert large(M ) into R with synthetic weight w(M )/4. endif
if small(M ) > λ1
then Insert small(M ) into R with synthetic weight w(M )/4. endif
endfor
Select the (synthetic) weighted median element λ in R.
if FTEST1(P, k, λ) = "lower" then λ1 ← λ else λ2 ← λ endif
Remove from R any values no longer in (λ1, λ2).
if R is not empty
then
Select the unweighted median element λ(cid:48) in R.
if FTEST1(P, k, λ(cid:48)) = "lower" then λ1 ← λ(cid:48) else λ2 ← λ(cid:48) endif
endif
PATH1 update mat:
while there is an M in M such that small(M ) ≥ λ2 or large(M ) ≤ λ1
17
or small(M ) ≤ λ1 ≤ λ2 ≤ large(M ) do
if small(M ) ≥ λ2 or large(M ) ≤ λ1
then
Delete M from M.
if this is the last submatrix remaining for a subpath P (cid:48)
then glue paths(P (cid:48))
endif
endif
if small(M ) ≤ λ1 and large(M ) ≥ λ2
then Split M into four square submatrices, each of synthetic weight w(M )/8.
endif
endwhile
Below is the procedure mats for path, which inserts submatrices of the appropriate previ-
ously discussed synthetic weight into M. A call with arguments f and t will generate subma-
trices for all required subpaths of a path containing the vertices with indices f, f + 1, . . . , t.
(In this section, we assume that f = 1 and t = n, but we state the procedure in this form so
that it can be used in the next section too.) The submatrices for the matrix M (P ) in Fig. 1
are shown in Fig. 6. Every subpath P (cid:48) that will appear in some partition of P is initialized
with cleaned(P (cid:48)) and glued(P (cid:48)) to false, where cleaned(P (cid:48)) indicates whether or not all
values in the submatrix associated with P (cid:48) are outside the interval (λ1, λ2), and glued(P (cid:48))
indicates whether or not all values in M (P (cid:48)) are outside the interval (λ1, λ2).
proc mats for path (path P , integer f, t)
size ← 1
w ← 4n4 /* synthetic weight for 1 × 1 submatrices */
while f ≤ t do
for i ← f to t by size do
(i+(cid:100)size/2(cid:101)).. (i+size−1)] for M (P ) into M with synthetic weight w.
i, . . . , i + size − 1.
Insert the succinct description of submatrix [i .. (i+(cid:100)size/2(cid:101)−1),
Let P (cid:48) be the subpath of P whose vertices have indices
cleaned(P (cid:48)) ← false; glued(P (cid:48)) ← false
endfor
size ← size ∗ 2
w ← w/2 /* smaller synthetic weight for increased size of submatrix */
f ← size ∗ (cid:100)(f − 1)/size(cid:101) + 1
18
t ← size ∗ (cid:98)t/size(cid:99)
endwhile
endproc
We next describe the procedure glue paths, which checks to see if two constituent sub-
paths can be combined together. Let P (cid:48) be any subpath of weight at least λ2, and let f(cid:48) and
t(cid:48) be the indices of the first and last vertices in P (cid:48), resp. Let the λ-prefix of P (cid:48), designated
λpref (P (cid:48)), be vertices vf(cid:48),··· , vl in P (cid:48) where l is the largest index such that w(f(cid:48), l) < λ2.
Let the λ-suffix of P (cid:48), designated λsuff (P (cid:48)), be vertices vl,··· , vt(cid:48) in P (cid:48) where l is the small-
est index such that w(l, t(cid:48)) < λ2. To glue two subpaths P2 and P3 together into a subpath
P1, we must have (glued(P2) and glued(P3) and cleaned(P1)). Procedure glue paths sets
the next pointers from vertices in λsuff (P2) to vertices in λpref (P3).
proc glue paths (path P1)
cleaned(P1) ← true
if P1 has length 1
then
glued(P1) ← true
Let l be the index of the vertex in P1.
if w(l, l) ≥ λ2 then next(l) ← l; ncut(l) ← 1 endif
Reset P1 to be subpath of which P1 is now a constituent subpath.
endif
Let P2 and P3 be the constituent subpaths of P1.
while glued(P2) and glued(P3) and cleaned(P1) and P1 (cid:54)= P do
glued(P1) ← true
Let f2 and t2 be resp. the indices of the first and last vertices in P2.
Let f3 and t3 be resp. the indices of the first and last vertices in P3.
last(f2) ← t3
if w(f2, t3) ≥ λ2
then
if w(f2, t2) < λ2
then f ← f2
else /* initialize f to the front of λsuff (P2) */
f ← t2
while w(f − 1, t2) < λ2 do f ← f − 1 endwhile
endif
t ← f3
19
while f ≤ t2 and w(f, t3) ≥ λ2 do
/* set next pointers for λsuff (P2) */
while w(f, t) < λ2 do t ← t + 1 endwhile
ncut(f ) ← 1
next(f ) ← t
f ← f + 1
endwhile
endif
Reset P1 to be subpath of which P1 is now a constituent subpath.
Let P2 and P3 be the constituent subpaths of P1.
endwhile
endproc
It would be nice if procedure glue paths, after setting pointers in λsuff (P2), would perform
pointer jumping, so that next pointers for vertices in λpref (P1) would point to vertices
in λsuff (P1). Unfortunately, it is not apparent how to incorporate pointer jumping into
glue paths without using O(n) time over all calls in the worst case. We thus opt for having
glue paths do no pointer jumping, and instead we do pointer jumping under the guise of
path compression in FTEST1. We use an argument based on amortization to show that this
works well.
Lemma 3.1 Let P be a path of n vertices. All calls to glue-paths will take amortized time
of O(n), and FTEST1 will search each subpath that is glued in amortized time proportional
to the logarithm of its length.
Proof
Suppose two subpaths P2 and P3 are glued together to give P1, where P1 has weight
at least λ2. We consider two cases. First suppose either P2 or P3 has weight at most λ1.
Let Pm represent this subpath. The time for glue paths is in worst case proportional to the
length of Pm. Also, we leave a glue-credit on each vertex in the λpref (P1) and λsuff (P1), and
a jump-credit on each vertex in the λpref (P1). The number of credits will be proportional to
the length of Pm. We charge this work and the credits to the vertices of Pm, at a constant
charge per vertex. It is clear that any vertex in P is charged at most once, so that the total
20
charge to vertices over all calls to glue paths is O(n). The second case is when both P2 and
P3 have weight at least λ2. In this case, the time for glue paths is proportional to the sum of
the lengths of λsuff (P2) and λpref (P3), and we use the glue-credits of P2 and P3 to pay for
this. The jump-credits are used by FTEST1 rather than by glue-paths, and in fact, would
present a problem if one tries to use them in glue-paths, as we discuss below.
Suppose both P2 and P3 have weight at least λ2. Then we might have wanted to have
glue-paths jump the next pointers so that the next pointer for a vertex in λpref (P2) would
be reset to point to λsuff (P3). The time to reset such pointers would be proportional to the
length of λpref (P2). The jump-credits of λpref (P3) could pay for this, as long as the length
of λpref (P2) is at most some constant (say 2) times the length of the λpref (P3). When the
length of the λpref (P2) is at most twice the length of the λpref (P3), we would not want to
jump the pointers.
In general we could view a subpath as containing a sequence of λ-regions, where each
λ-region was once the λ-prefix of some subpath, and the length of a λ-region is less than
twice the length of the preceding λ-region. Note that the number of λ-regions in the sequence
could be at most the logarithm of the length of the subpath. When gluing two subpaths
together, we could concatenate their sequences of λ-regions together, and jump pointers
over any λ-region that gets a predecessor whose length is not more than twice its length. Its
jump-credits could then be used for this pointer jumping. Searching a subpath in FTEST1
would then take time at most proportional to its length.
Of course glue-paths does not jump pointers. However, a lazy form of pointer-jumping
is found in the path-compression of FTEST1, and the same sort of analysis can be seen to
apply. Suppose that FTEST1 follows a pointer to a vertex that was once in the λ-prefix
of some subpath. Consider the situation if glue-paths had jumped pointers. If that pointer
would have been present in the sequence of λ-regions, then charge the operation of following
that pointer to FTEST1. Otherwise, charge the operation of following that pointer to the
21
jump-credits of the λ-prefix containing the vertex. It follows that the number of pointers
followed during a search of a subpath that are not covered by jump-credits is at most the
logarithm of the length of the subpath. Also, the binary search to find the position of the
first cut on a subpath uses time at most proportional to the logarithm of the length of the
subpath.
Lemma 3.2 Let P be a path of n vertices. On the i-th iteration of the while-loop of PATH1,
the amortized time used by feasibility test FTEST1 will be O(i(5/6)i/5 n).
Proof We first consider the synthetic weights assigned to submatrices in M. Correspond-
ing to subpaths of lengths 1, 2, 4, . . . , n, procedure mats for path creates n submatrices of size
1× 1, n/2 submatrices of size 1× 1, n/4 submatrices of size 2× 2, and so on, up through one
submatrix of size n/2 × n/2. The total synthetic weight of the submatrices corresponding
to each path length is 4n5, n5, n5/4, . . . , 4n3, resp. It follows that the total synthetic weight
for submatrices of all path lengths is less than (4/3) ∗ 4n5.
Let wgt(M ) be the synthetic weight assigned to submatrix M . Define the effective
weight, denoted eff wgt(M ), of a submatrix M in M to be wgt(M ) if both its smallest value
and largest value are contained in the interval (λ1, λ2) and (3/4)wgt(M ) if only one of its
smallest value and largest value is contained in the interval (λ1, λ2). Let eff wgt(M) be the
total effective weight of all submatrices in M. When a feasibility test renders a value (the
smallest or largest) from M no longer in the interval (λ1, λ2), we argue that eff wgt(M ) is
reduced by at least wgt(M )/4 because of that value.
If both values were contained in the interval (λ1, λ2), and one no longer is, then clearly
our claim is true. If only one value was contained in the interval (λ1, λ2), and it no longer
is, then M is replaced by four submatrices of effective weights wgt(M )/8 + wgt(M )/8 +
(3/4)wgt(M )/8 + (3/4)wgt(M )/8 < wgt(M )/2, so that there is a reduction in effective
weight by greater than wgt(M )/4. If both values were contained in the interval (λ1, λ2), and
22
both are no longer in, then we consider first one and then the other. Thus every element that
was in (λ1, λ2) but no longer is causes a decrease in eff wgt(M) by an amount at least equal
to its synthetic weight in R. Values with at least half of the total weight in R find themselves
no longer in (λ1, λ2). Furthermore, the total weight of R is at least 1/3 of eff wgt(M). This
follows since a submatrix M has either two values in R at a total of 2w(M )/4 = (1/2)w(M )
or one value at a weight of w(M )/4 = (1/3)(3w(M )/4). Thus eff wgt(M) decreases by a
factor of at least (1/2)(1/3) = 1/6 per iteration.
Corresponding to the subpaths of length 2j, there are n/2j submatrices created by
mats for path. If such a matrix is quartered repeatedly until 1 × 1 submatrices result, each
such submatrix will have weight n4/24j−2. Thus when as little as (n/2j) ∗ (n4/24j−2) syn-
thetic weight remains, all subpaths of length 2j can still be unresolved. This can be as late as
iteration i, where i satisfies (4/3) ∗ 4n5 ∗ (5/6)i = n5/25j−2, or 25j = (3/4) ∗ (6/5)i. While all
subpaths of length 2j can still be unresolved on this iteration, at most (1/2 ∗ 1/24)k = 1/25k
of the subpaths of length 2j−k can be unresolved for k = 1, . . . , j. Also, by Lemma 3.1,
each subpath can be searched by FTEST1 in amortized time proportional to the logarithm
of its length. Thus the time to search path P on iteration i is at worst proportional to
(j/2j)n(1 + 1/25 + 1/210 + ··· ), which is O((j/2j)n). From the relationship of i and j,
2j = (3/4)1/5(6/5)i/5, and j = (1/5) log(3/4) + (i/5) log(6/5). The lemma then follows.
Lemma 3.3 The total time for handling M and performing selection over all iterations of
PATH1 is O(n).
Proof Using the algorithm of [3], the time to perform selection in a given iteration is
proportional to the size of R. Define the effective count, denoted eff cnt(M ), of a submatrix
M in M to be 2 if both its smallest value and largest value are contained in the interval
(λ1, λ2) and 1 if only one of its smallest value and largest value is contained in the interval
(λ1, λ2). Let eff cnt(M) be the total effective count of all submatrices in M. Then the size
23
of R equals eff cnt(M). On any iteration, the result of the feasibility test of λ(cid:48) is to resolve
at least half of the values in R.
The time for inserting values into R and performing the selections is O(n). We show this
by using an accounting argument, charging 2 credits for each value inserted into R. As R
changes, we maintain the invariant that the number of credits is twice the size of R. When
R has k elements, a selection takes O(k) time, paid for by k credits, leaving k/2 elements
covered by k credits. Since n elements are inserted into R during the whole of PATH1, the
time for forming R and performing selections is O(n).
It remains to count the number of submatrices inserted into M. Initially 2n − 1 subma-
trices are inserted into M. For j = 1, 2, . . . , log n − 1, consider all submatrices of size 2j × 2j
that are at some point in M. A matrix that is split must have its smallest value at most
λ1 and its largest value at least λ2. However, Mi,j > Mi−k,j+k for k > 0, since the path
represented by Mi+k,j−k is a subpath of the path represented by Mi,j. Hence, for any 2j × 2j
submatrix that is split, at most one submatrix can be split in each diagonal going from lower
left to upper right. There are fewer than 2n diagonals, so there will be fewer than 2(n/2j)
submatrices that are split. Thus the number resulting from quartering is less than 8(n/2j).
Summing over all j gives O(n) submatrices in M resulting from quartering.
We illustrate PATH1 on path P in Fig. 1, with k = 3. (This is the same example that
we discussed near the end of the previous section.) The initial submatrices for M are shown
in Fig. 6. The submatrix of size 4 × 4 has synthetic weight 211, the two submatrices of size
2× 2 have synthetic weight 212, four submatrices of size 1× 1 have synthetic weight 213, and
the remaining eight submatrices of size 1 × 1 have synthetic weight 214. Initially λ1 = 0 and
λ2 = ∞. The last, next and ncut arrays are initialized as stated.
On the first iteration, R contains two copies each of 6, 11, 9, 2, 1, 15, 7, 8 of synthetic
weight 212, two copies each of 17, 11, 16, 15 of synthetic weight 211, 20, 28, 22, 31 of synthetic
weight 210, and 3, 59 of synthetic weight 29. The weighted median of R is 11. For λ =
24
11, three cuts are required, so we reset λ1 to 11. The revised version of R will then be
{15, 15, 17, 17, 16, 16, 15, 15, 20, 28, 22, 31, 59}. The median of this is 17. For λ(cid:48) = 17, one
cut is required, so we reset λ2 to 17. The set M is changed as follows. We discard all
submatrices of size 1× 1 except the one of synthetic weight 214 containing 15, and the two of
synthetic weight 213 containing 15 in one and 16 in the other. We discard all submatrices of
size 2 × 2. We quarter the submatrix of size 4 × 4, giving four submatrices each of synthetic
weight 28. We discard three of these submatrices because their values are all too large. We
quarter the remaining submatrix (containing 27, 12, 18, 3), giving four submatrices each of
synthetic weight 25. We discard all submatrices but the submatrix of size 1 × 1 containing
12.
As a result of submatrix discarding, we render a number of subpaths cleaned and glued.
We clean and glue every subpath of length 1 except v6, the subpaths v1, v2 and v3, v4, the
subpath v1, v2, v3, v4, and we clean but do not glue the subpath v5, v6, v7, v8. When we glue
subpath v6, we set next(6) to 6, and ncut(6) to 1. When we glue subpath v1, v2, we set
last(1) to 2, next(1) to 2, and ncut(1) to 1. When we glue subpath v3, v4, we set last(3) to
4, but we change no next or ncut values. When we glue subpath v1, v2, v3, v4, we set last(1)
to 4, next(2) to 3, and ncut(2) to 1. This completes all activity on the first iteration.
On the second iteration, R contains two copies of 15 of synthetic weight 212, two copies
each of 16, 15 of synthetic weight 211, and two copies of 12 of synthetic weight 23. The
weighted median of R is 15. For λ = 15, 2 cuts are required, so we reset λ2 to 15. The
revised version of R will then be {12, 12}. The median of this is 12. For λ(cid:48) = 12, 3 cuts are
required, so we reset λ1 to 12.
All remaining submatrices are discarded from M. When this happens, we clean and glue
all remaining subpaths. When we glue subpath v5, v6, we set last(5) to 6, next(5) to 6, and
ncut(5) to 1. When we glue subpath v7, v8, we set last(7) to 8, next(7) to 8, and ncut(7) to
1. When we glue subpath v5, v6, v7, v8, we set last(5) to 8, but we change no next or ncut
25
values. When we glue subpath v1, v2, v3, v4, v5, v6, v7, v8, we set last(1) to 8, next(3) and
next(4) to 6, and ncut(3) and ncut(4) to 1. Since M is empty, PATH1 will then terminate
with λ1 = 12, λ2 = 15, and output λ∗ = 12.
Note that we performed no compression of search paths on the second iteration. Suppose
for the sake of example that we performed a subsequent search with λ = 13. The initial cut
would come after v2, and next(3) and next(7) would be followed to arrive at v8. Then we
would reset ncut(3) to 2, and next(3) to 8.
Theorem 3.4 Algorithm PATH1 solves the max-min k-partitioning problem on a path of n
vertices in O(n) time.
Proof
The use of arrays next and ncut, indexed by vertices on the path, is the key
difference between PATH0 and PATH1. Feasibility test FTEST1 requires next(a) to point
to a further vertex, b, in the subpath, such that ncut(a) is the number of cuts between a
and b. Enforcing this requirement, we update arrays next and ncut by two subroutines of
PATH1.
The first is glue paths. When small(M ) ≥ λ2 or large(M ) ≤ λ1, certain vertices on
the subpaths need no longer be inspected, so glue paths combines two adjacent subpaths of
equal length by updating the next and ncut values on the subpaths. We repeat the gluing
and updating until no longer possible. The second subroutine is compress next path, which
we call after search next path(a, b) in FTEST1. Procedure search next path(a, b) determines
(v, sumcut), where v is the location of the final cut before b and sumcut is the number of
cuts between a and v. We then update the values of next and ncut for vertices between a
and v.
The correctness of FTEST1 follows, as we increment numcut once for each subpath
whose weight exceeds λ∗, or by sumcut for each compressed subpath with the corresponding
number of cuts. Correctness of PATH1 follows from the correctness of FTEST1, from the
26
fact that all possible candidates for λ∗ are included in M, and from the fact that each value
discarded is either at most λ1 or at least λ2.
The time to initialize M is clearly O(n). By Lemma 3.3, the total time to select all values
to test for feasibility is O(n). By Lemma 3.2, the amortized time to perform feasibility test on
iteration i is O(i(5/6)i/5n). Summed over all iterations, this quantity is O(n). By Lemma 3.3,
the total time to handle submatrices in M is O(n). By Lemma 3.1, the time to manipulate
data structures for the subpaths is O(n). The time bound then follows.
We briefly survey the differences needed to solve the min-max path-partitioning prob-
lem. Procedure FTEST1 is similar except that r is the largest index such that w(f, r) +
remainder ≤ λ, we subtract 1 from numcut after completion of the while-loop if remainder =
0, and use numcut ≥ k rather than numcut > k. Upon termination, FTEST1 outputs
λ∗ = λ2. Procedure glue paths is similar, except that we replace the statement
while w(f, t) < λ2 do t ← t + 1 endwhile
by the statement
while w(f, t + 1) < λ2 do t ← t + 1 endwhile
Note that w(l, l) ≤ λ2 always holds for the min-max problem.
Theorem 3.5 The min-max k-partitioning problem can be solved on a path of n vertices in
O(n) time.
Proof The above changes will not affect the asymptotic running time of algorithm PATH1.
27
4 Partitioning for Max-Min on a Tree
In this section, we present an optimal algorithm to perform parametric search for the max-
min problem on a graph that is a tree. The algorithm differs from that in Section 3 as either
long paths or many leaves can overwhelm the running time, so we must simultaneously
compress long paths and delete leaves. The situation is further complicated by the challenge
of handling problematic vertices, which are vertices of degree greater than two. Thus, we
pursue a dual-pronged strategy that identifies both paths to compress and paths to delete.
4.1 Our Tree Algorithm
Our algorithm proceeds through rounds, running an increasing number of selection and
feasibility tests on each round. As we shall show in Section 4.3, each round halves the
feasibility test time, and the overall time for selection is linear. Thus, our algorithm runs in
linear time.
For the purposes of discussion and analysis, we initially classify paths in the tree as either
pending paths or processed paths, based on whether the actual weights associated with the
path have already been resolved. Let an internal path be a subpath between the root and a
problematic vertex, or between two problematic vertices such that all intermediate vertices
are of degree 2. Let a leaf path be a subpath with either the root or a problematic vertex as
one endpoint and the other endpoint being a leaf. Furthermore, let a processed path be an
internal path that is completely cleaned and glued and let a pending path be a path, either
an interval or a leaf path, that contains some value in the interval (λ1, λ2). Moreover, we
define and classify paths or subpaths as light paths, middleweight paths, or heavy paths,
based on how the actual weights in the path compare to λ1 and λ2. A subpath is light if its
total weight is at most λ1, and a subpath is heavy if its total weight is at least λ2. Otherwise,
a subpath is middleweight. Note that a middleweight path may consist of a sequence of both
light and/or middleweight subpaths.
28
For each round, our algorithm runs an increasing number of selection and feasibility tests,
while maintaining three selection sets, H, U, and V, consisting of candidates for the max-
min value. The first selection set is for heavy subpaths, the second is for middleweight paths
formed as sequences of light or middleweight subpaths after the resolution of a problematic
vertex, and the third is for handling problematic vertices whose leaf paths have been resolved.
We assign synthetic weights to values inserted into H or U as a function of the lengths of
the corresponding paths, as defined in the procedure mats for path in Section 3, assuming
that we round up the length to a power of 2.
Our algorithm invokes a separate procedure to address each of these sets specifically.
Procedure handle middleweight paths inserts the total actual weight of each middleweight
path as elements into U. Our algorithm then performs two weighted selections on the
elements in U, one using the length of each path as the weight, and the other using the
synthetic weight of each path. For each of the weighted selections, our algorithm then tests
the selected value for feasibility, and adjusts λ1 and λ2 accordingly. For any middleweight
path that becomes heavy, our algorithm succinctly identifies the corresponding submatrix
and inserts the representatives (of the submatrix) whose values are within (λ1, λ2) into H.
For any path that becomes light, our algorithm cleans and glues the path, so that future
feasibility tests require only polylogarithmic time to search the path.
The second procedure, handle pending paths, selects the weighted and unweighted medi-
ans from H separately, tests each for feasibility, and adjusts λ1 and λ2 accordingly. After the
feasibility test, our algorithm cleans and glues adjacent subpaths that it has just resolved.
When a leaf path is completely resolved, our algorithm represents the leaf path by a single
vertex with the remaining accumulated weight, accum wgt(v), as described in Section 2.
Our algorithm defers until the next iteration of the for-loop the insertion into H of the
representatives of any subpaths that have become heavy.
The third procedure is handle leaves. For any problematic vertex with a resolved leaf
29
path hanging off it, the procedure inserts into V the sum of the weight of that vertex plus
the accumulated remaining weight left over from the resolved leaf path. The procedure then
selects the median from V, performs the feasibility test, and adjusts λ1 and λ2 accordingly.
Since we want to find the max-min, if the number of cuts is at least k, then for each vertex
whose weight plus the accumulated remaining weight from the resolved leaf path is at most
the median, our algorithm merges the vertex with the accumulated remaining weight from
the resolved leaf path. On the other hand, if the number of cuts is less than k, then for each
vertex whose weight plus the accumulated remaining weight from the resolved leaf path is at
least the median, our algorithm cuts below the parent vertex, because any future feasibility
tests would do the same. Furthermore, our algorithm assigns, to any middleweight path
created by the resolution of a problematic vertex, a synthetic weight that is a function of
the length (rounded up to a power of two) as defined in the procedure mats for path in
Section 3. Our algorithm then inserts the weight of that path into U at the beginning of
the next invocation, along with the corresponding synthetic weight. Note that we merge
middleweight paths only when we know whether they will become light or heavy. For the
representative of any subpath that becomes heavy during this procedure, we wait until the
next iteration of the for-loop to insert the representative into H.
We now give the top level of our algorithm. We defer our discussion of the corresponding
data structures until Section 4.2.
TREE1:
Initialize the data structures for the algorithm and set round r ← 1.
while H ∪ U ∪ V (cid:54)= ∅ do
while feasibility test time is more than n/2r do
Call procedure handle middleweight paths.
repeat Call procedure handle pending paths
until the feasibility test time on heavy paths has been reduced by 50%
and the feasibility test time on leaf paths has been reduced by 25%
Call procedure handle leaves.
endwhile
r ← r + 1
30
endwhile
4.2 Data Structures
As our algorithm progresses, it prunes the tree to delete some of the paths in the edge-
path-partition and it glues together other paths, while updating the values of λ1 and λ2. In
this section, we discuss the necessary data structures so that our algorithm can efficiently
perform feasibility testing as these updates occur.
Our algorithm represents a path in an edge-path-partition in consecutive locations of an
array. To achieve that, our algorithm initially stores the actual weights of the vertices of the
tree in an array using the following order. It orders the children of each vertex from left to
right by nondecreasing height in the tree. Using this organization of the tree, our algorithm
then lists the vertices of the tree in postorder. Our algorithm can find the heights of all
vertices in linear time. It can also order pairs containing parent and height lexicographically
in linear time. Our algorithm does this process only during the initialization phase. It uses
arrays that are parallel to the actual weight array to store pointer values, as well as the
accumulated weight (see procedure explore in Section 2), which it then uses to compute the
actual weight of specified subpaths, exactly the same as the last, ncut, and next arrays in
Section 3. We discuss this additional information in due course.
When our algorithm removes paths from the tree, it reorganizes the storage of the tree
within the array as follows. Let P be a leaf-path in the edge-path-partition of the current
tree, and let t be the top vertex of P . To remove P − t, we proceed as follows: Let P (cid:48) be
the path whose bottom vertex is t. If t will have only one child remaining after removal of
P − t, and this child was originally not the rightmost child of t, do the following. Let P (cid:48)(cid:48)
be the other path whose top vertex is t. Copy the vertices of P (cid:48)(cid:48) − t in order, so that the
top vertex of this path is in the location preceding the location of t in the array. Modify
the actual weight of t by adding the remainder left from removing P − t. Also, copy pointer
31
values for all copied vertices into the new locations in their arrays. That is, update last,
ncut, and next pointers in P (cid:48). In particular, the last pointer for the first vertex in P (cid:48) should
now point to the last vertex in P (cid:48)(cid:48). We should also copy and modify the accumulated weight,
as we discuss shortly.
Note that the bottom vertex of P (cid:48)(cid:48) may have been the parent of several vertices. When
P (cid:48)(cid:48) − t is moved, we do not copy the children of its bottom vertex (and subtrees rooted
at them). It is simple to store in a location formerly assigned to the bottom vertex of P (cid:48)(cid:48)
the current location of this vertex, and to also store a pointer back. If we copy the path
containing P (cid:48)(cid:48), then we reset this pointer easily. When only one child of the bottom vertex of
P (cid:48)(cid:48) remains, we copy the corresponding path to in front of P (cid:48)(cid:48). We claim that the total time
to perform all rearrangements will be O(n): The time to copy each vertex and to copy and
adjust its accumulated weight is constant. Because of the way in which the tree is stored in
the array, at most one vertex will be copied from any array location.
We next discuss the representation of a heavy path. Each heavy path P is represented
as a sequence of overlapping subpaths, each of actual weight at least λ2. Each vertex of P is
in at most two overlapping subpaths. Each overlapping subpath, except the first, overlaps
the previous overlapping subpath on vertices of total actual weight at least λ2, and each
overlapping subpath, except the last, overlaps the following overlapping subpath on vertices
of total actual weight at least λ2. Thus any sequence of vertices of weight at most λ2 that is
contained in the path is contained in one of its overlapping subpaths. For each overlapping
subpath, we shall maintain a succinct version of the corresponding sorted matrix M (P ) (as
described in Section 2), where P is the overlapping subpath excluding the top vertex of the
overlapping subpath.
Initially, no path is heavy, since initially λ2 = ∞. Our algorithm recognizes a path P as
heavy in one of two ways.
1. Path P was middleweight until a feasibility test reduced the value of λ2.
32
2. Path P results from the concatenation of two or more paths following the resolution
of a problematic vertex.
If a heavy path P arises in the first way, then represent P by two overlapping subpaths
that are both copies of P , arbitrarily designating one as the first overlapping subpath and
the other as the second. For each overlapping subpath, create the corresponding succinct
description of a sorted matrix. If heavy path P is the concatenation of paths, all of which
were light, then do the same thing.
Otherwise, path P results from the concatenation of paths, with at least one of them
being heavy. Do the following to generate the representation for P . While there is a light or
middleweight path P (cid:48) to be concatenated to a heavy path P (cid:48)(cid:48), combine the two as follows.
If P (cid:48) precedes P (cid:48)(cid:48), then extend the first overlapping subpath of P (cid:48)(cid:48) to incorporate P (cid:48). If P (cid:48)
follows P (cid:48)(cid:48), then extend the last overlapping subpath of P (cid:48)(cid:48). This completes the description
of the concatenation of light or middleweight paths with heavy paths. While there is more
than one heavy path, concatenate adjacent heavy paths P (cid:48) and P (cid:48)(cid:48) as follows. Assume that
P (cid:48) precedes P (cid:48)(cid:48). Combine the last overlapping subpath of P (cid:48) with the first of P (cid:48)(cid:48). Note that
any vertex can be changed at most twice before it is in overlapping subpaths that are neither
the first nor the last.
We now discuss how to perform efficient feasibility testing, given our representation of
paths. To efficiently search along paths, we maintain a second representation of paths as
red-black balanced search trees. In each tree, there will be a node x for every vertex v in
the subpath. Node x contains two fields, wt(x) and ct(x). Field wt(x) contains the sum of
the actual weights of all vertices whose nodes are in the subtree rooted at x, and field ct(x)
will equal the number of nodes in the subtree rooted at x. With this tree it is easy to search
for a vertex v in subpath P such that v is the first vertex in P so that the sum of the actual
weights to v is at least a certain value, and to determine at the same time the position of v
in P . This search takes time proportional to the logarithm of the length of P . When two
33
paths P (cid:48) and P (cid:48)(cid:48) need to be concatenated together, we merge the corresponding search trees
in time proportional to the logarithm of the combined path length.
Moreover, a problematic vertex also needs access to the accumulated actual weight of
each of its subtrees. Thus, we also maintain a linked list child, which points to all children
of a vertex, if they exist. When leaf paths become resolved, remove the pointer from the
parent vertex, but since we do not move the locations of the other descending paths until
only one remains, we do not need to change the pointers for the other children. Thus, when
we delete a leaf path, we update the subpaths and pointers accordingly, so that each subpath
in the edge-path-partition remains in a contiguous block of memory.
4.3 Analysis of our Tree Algorithm
The analysis of our algorithm is not obvious because techniques in Section 3 force the syn-
thetic weight of data structure M to decrease monotonically as our algorithm progresses, but
neither U nor H individually have this property. Moreover, neither U nor H alone provides
an accurate count of the number of resolved paths.
In a feasibility test, our algorithm must explore problematic vertices as well as both
processed and pending paths. Specifically, the feasibility test time is the number of vertices
our algorithm lands on in pending paths, plus the number of vertices our algorithm lands on
in processed paths, plus the number of problematic vertices. Since leaf paths are pending
paths, and the number of problematic vertices is less than the number of leaf paths, the
feasibility test time is upper bounded by double the number of vertices our algorithm lands
on in both pending and processed paths. Thus, we show an upper bound on the time spent
by the feasibility test after each round. Finally, we show that our algorithm needs at most
66(r2 + 9) iterations to cut in half the feasibility test time.
We remark that cleaning and gluing will be the same as in Section 3 since our algorithm
performs cleaning and gluing only along subpaths that were initially created as part of
34
the edge-path-partition or along heavy paths, for which we would have a corresponding
submatrix.
To analyze the performance of our algorithm, we consider, in an auxiliary data structure
P, the representatives of heavy and middleweight subpaths of the tree, along with their
corresponding synthetic weights as a function of their lengths (as defined in Section 2).
When we resolve a problematic vertex, the handle leaves routine produces a newly formed
path that is either a light path, a heavy path, or a middleweight path. Recall that if a newly
formed subpath is middleweight, we insert a representative consisting of actual weight of that
subpath into U. If a newly formed subpath is heavy, we generate our succinct representation
of its corresponding submatrix, and we insert the representatives of the submatrix into H.
In each case, we also insert the representative(s) into P with their same synthetic weights, as
described earlier. Furthermore, at the beginning of each iteration, P includes exactly H, U,
and the representatives of middleweight subpaths of paths represented in U. For convenience
in analysis, let n be the smallest power of two greater than the number of vertices in the
graph. We now consider P to show results analogous to Lemma 3.2 and Lemma 3.3.
Lemma 4.1 At any point in our algorithm, the total synthetic weight of the values in P is
less than (4/3) ∗ 4n5.
Proof Consider the synthetic weights assigned to the representative values in P. Even
if all representatives in P represent heavy paths, procedure mats for path creates at most n
submatrices for subpaths of length 1, at most n/2 more submatrices for subpaths of length
1, at most n/4 submatrices for subpaths of length 2, and so on, up to at most one submatrix
for a subpath of length n/2, corresponding to subpaths of lengths 1, 2, 4, . . . , n. The total
synthetic weight of the representative values corresponding to each heavy path length is at
most 4n5, n5, n5/4, . . . , 4n3, resp. If a path is middleweight, no submatrix is generated for
the path, but by construction, its representative value in P has the same synthetic weight as
35
it would if the path were heavy. Thus, the total synthetic weight in P is less than (4/3)∗4n5.
In Section 4 we defined functions wgt and eff wgt on matrices to analyze the progress of
our algorithm. We use a similar analysis in this section, but we do not have submatrices
for middleweight paths. To emphasize the parallels, we overload the definitions of wgt
and eff wgt, as defined below. Let wgt(P ) be the synthetic weight assigned to subpath P ,
as a function of its length, as defined in Section 2. Define the effective weight, denoted
eff wgt(P ), of a subpath P containing a representative value in P to be wgt(P ) if both its
smallest value and largest value are contained in the interval (λ1, λ2) and (3/4)wgt(P ) if only
one of its smallest value and largest value is contained in the interval (λ1, λ2). We analyze
our algorithm using this notion of eff wgt(P ) without ever the need to explicitly calculate it.
Let eff wgt(P) be the total effective weight of all subpaths with representative values in P.
Lemma 4.2 A weighted and unweighted selection, first both from U and then both from H,
resolves at least 1/24 of eff wgt(P)
Proof When a feasibility test renders a value (the smallest or largest) from a heavy
subpath P no longer in the interval (λ1, λ2), we argue that eff wgt(P ) is reduced by at least
wgt(P )/4 because of that value.
If both values were contained in the interval (λ1, λ2), but one no longer is, then clearly
eff wgt(P ) is reduced by at least wgt(P )/4. If only one value was contained in the interval
(λ1, λ2), and it no longer is, then we consider whether λ1 increased or λ2 decreased. If λ1
increased and P has no value within (λ1, λ2), then any subpath will also have no value within
(λ1, λ2). Thus, eff wgt(P ) is clearly reduced by at least wgt(P )/4. On the other hand, if λ2
decreased so that P is now a heavy path, then we generate M (P ). We then replace M (P )
with the succinct description of four submatrices of effective weights wgt(P )/8 + wgt(P )/8 +
(3/4)wgt(P )/8+(3/4)wgt(P )/8 < wgt(P )/2, so that there is a reduction in weight by greater
36
than wgt(P )/4. If both values from the submatrix were contained in the interval (λ1, λ2),
and both no longer are, then we consider first one and then the other.
Thus every representative that was in (λ1, λ2) but no longer is causes a decrease in
eff wgt(P) by an amount at least equal to its synthetic weight in P. Representatives with
more than half of the total weight in P find themselves no longer in (λ1, λ2).
Recall that we include values representing a subpath P in selection set H if and only if P
is a heavy subpath, and the corresponding values are contained within (λ1, λ2). Of a subpath
P that does have representative values contained in H, then P has either two values in P at a
total of 2wgt(P )/4 = (1/2)wgt(P ) or one value at a weight of wgt(P )/4 = (1/3)(3wgt(P )/4).
For each selection and feasibility test from H and U, at least half of the weight of P is
contained in U or contained in H.
Suppose at least 1/2 of P is contained in U. Then following a selection and feasibility
test from U, at least (1/2)(1/2) = 1/4 of the weight of P will be removed from U, either
determined to be smaller than an updated λ1, or larger than an updated λ2 and inserted
into H as a part of a heavy subpath. Then, following a round of selection and feasibility
testing from H, the weight of P is reduced by at least (1/2)(1/4) = 1/8.
On the other hand, if there is more weight of P in H than in U, then a round of selection
and feasibility testing from H decreases the weight of P by at least (1/2)(1/2) = 1/4.
Thus, in the worse of the two cases, the weight of P decreases by at least (1/8)(1/3) =
1/24 per iteration.
Lemma 4.3 Following i iterations of weighted and unweighted selection and feasibility test-
ing for H, where 25j = (3/4) ∗ (24/23)i, at most 1/25k of the subpaths, of length 2j−k,
represented by values in H can be pending. (Recall that a pending path is defined to contain
some value that is not resolved.)
We omit the proof of Lemma 4.3, as it is essentially contained in the proof of Lemma 3.2,
37
with 24/23 replacing 6/5.
Lemma 4.4 The total time to generate and maintain the overlapping subpaths and thus,
the representatives of the corresponding submatrices for all the heavy subpaths is O(n).
Proof Recall that each vertex can be in at most two overlapping subpaths, so the total time
for generating the representatives of the corresponding submatrices for all heavy subpaths
is at most 2n.
Theorem 4.5 The time for inserting, selecting, and deleting representatives of H and U
across all iterations is O(n).
First, we count the number of representatives inserted into H. Initially, at most
Proof
2n − 1 subpaths have representatives in H. For j = 1, 2, . . . , log n − 1, consider all subpaths
of length 2j that at some point have representatives in H. A path that can be split must
have its smallest value at most λ2 and its largest value at least λ1. Note that light paths
are not split further, while heavy paths are represented succinctly by submatrices. However,
Mi,j > Mi−k,j+k for k > 0, since the heavy path represented by Mi−k,j+k is a subpath of the
path represented by Mi,j. Hence, for any submatrix of size 2j × 2j which is split, at most
one submatrix can be split in each diagonal extending upwards from left to right. There are
fewer than 2n diagonals, so there will be fewer than 2(n/2j) submatrices that are split. Thus
the number of submatrices resulting from quartering is less than 8(n/2j). Summing over all
j gives O(n) insertions into H overall.
Next, we count the number of representatives inserted into U. There are at most n
subpaths of length 1, n/2 subpaths of length 2, and so forth, up to at most 1 path of length
n. Since we insert a representative of each subpath at most once into U, the number of
representatives inserted into U is O(n).
38
Finally, we show that the total selection time from H and U is linear in n. We give an
accounting argument similar to that used in the proof of Lemma 3.3. Charge 2 credits for
each value inserted into H or U. As H and U change, we maintain the invariant that the
number of credits is twice the size of H or U, respectively. The rest of the proof is analogous
to that of Lemma 3.3. Thus, we conclude that the time for performing selection is O(n).
We make the following observations: The time spent by the feasibility test on a pending
path is at least as much time as spent by the feasibility test as if the pending path were a
processed path. Resolving a processed path does not increase the feasibility test time spent
on pending paths. Based on these observations, we note that the feasibility test time spent
on pending paths cannot increase and once all pending paths are resolved, the time spent
on processed paths cannot increase. We analyze our algorithm in each of the following three
exhaustive cases:
1. Lemma 4.6: The feasibility test spends more time on heavy paths than middleweight
paths and at least as much time on pending paths as on processed paths.
2. Lemma 4.10: The feasibility test spends more time on heavy paths than middleweight
paths and more time on processed paths than pending paths.
3. Lemma 4.11: The feasibility test spends at least as much time on middleweight paths
as on heavy paths.
Lemma 4.6 Suppose at the beginning of round r, that the feasibility test lands on at most
n/2r−1 vertices. Suppose further that the feasibility test lands on more vertices in heavy
paths than middleweight paths. If the feasibility test also lands on at least as many vertices
in pending paths as vertices in processed paths, then following at most 163r + 170 iterations
of selection and feasibility testing from H, the number of vertices in pending paths that the
feasibility test lands on is either halved or at most n/2r+2.
39
Proof By assumption, the feasibility test lands on at most n/2r−1 vertices and lands on
more vertices in heavy paths than middleweight paths, and at least as many vertices in
pending paths as vertices in processed paths. If the number of vertices in pending paths
that the feasibility test lands on is at most n/2r+2, then the result follows. Thus, we assume
the feasibility test lands on more than n/2r+2 vertices in pending paths. By Lemma 4.3,
in iteration i, at most 1/25k of the subpaths of length 2j−k can be pending, where 25j =
(3/4) ∗ (24/23)i. Then for j = 2r + 2 and k = r, at most 1/25r of the subpaths of length
2(r+2) can be pending, so there are at most n/25r vertices remaining in pending paths. Thus,
it takes at most i = (5(2r + 2) − log(3/4))/ log(24/23) < 163r + 170 iterations of selection
and feasibility testing from H to reduce the amount of time spent on the pending paths by
at least half. Since each iteration of feasibility testing lands on at most n/2r−1 vertices, the
total number of vertices checked is at most (n/2r−1) (163r + 170).
Before we can show an analogous result for processed paths, we introduce three preliminary
lemmas.
Lemma 4.7 If a feasible lands on at least t vertices in a processed path P , then P has length
√
at least 2
t/2.
Proof
√
t/2. From the edge-path-partition, P can contain
Suppose P has length less than 2
√
t/2−1. However, if P previously contained a prob-
disjoint subpaths of lengths 1, 2, 22, . . . , 2
lematic vertex, then it may have two disjoint subpaths of each length. Thus, the feasibility
test lands on at most 2(1 + 2 + . . . +(cid:112)t/2 − 1) = ((cid:112)t/2 − 1)(cid:112)t/2 < t vertices in total in
P , which is a contradiction.
Lemma 4.8 Suppose a feasibility test lands on at most n/2r vertices but more than n/2r+2
vertices. If the feasibility test spends more time on processed paths than pending paths, then
the median length of leaf paths is at most max(24r2, 2400).
40
Proof
Suppose, by way of contradiction, the median length of a leaf path is more than
24r2 and r > 10. Let x be the mean number of vertices the feasibility test lands on, across all
leaf paths, so that 4r2 < x. Let t be the mean number of vertices the feasibility test lands
on, across all processed paths. Recall that all leaf paths are pending paths. By assumption,
the feasibility test spends more time on processed paths than pending paths. Moreover, the
number of leaf paths is more than the number of internal paths, pending or processed, so
that t > x. By Lemma 4.7, each processed path in which the feasibility test lands on at
t/2. Thus, if x ≤ 2r, then by considering the pending
least t vertices has length is at least 2
√
paths, the ratio of the time spent by the feasibility test to the total number of vertices is at
x
24r2 < 1
most
2r+2 , which contradicts the assumption that the feasibility test lands on more
than n/2r+2 vertices. On the other hand, if x > 2r, then by considering the processed paths,
the ratio of the time spent by the feasibility test to the total number of vertices is at most
√
t/2 < 2t
2
2r+2 for r ≥ 10, since t > x > 2r. This again contradicts the assumption
√
x+t
24r2 +2
that the feasibility test lands on more than n/2r+2 vertices. Thus, the median length of a
t/2 < 1
leaf path is at most max(24r2, 2400).
For the remainder of the section, we analyze r ≥ 10, noting that for r < 10, the median
length of a leaf path is at most 2400 and can be handled in a constant number of feasibility
tests.
Lemma 4.9 Suppose the feasibility test lands on at most n/2r vertices but more than n/2r+1
vertices. Suppose further that the feasibility test lands on more vertices in heavy paths than
middleweight paths. If the feasibility test spends more time on processed paths than pending
paths, then the median length of processed paths is at most 2r2+9. Hence, the number of
vertices the feasibility test lands on in a median length processed path is at most 2(r2 + 9)2.
Proof
Suppose, by way of contradiction, the median length of processed paths is more than
2r2+9. Then by Lemma 4.7, the number of vertices in processed paths that the feasibility
41
test lands on is at most
(cid:16)
2(r2 + 9)2/2r2+9(cid:17)
n. By assumption, the feasibility test spends
more time on heavy paths than middleweight paths, and more time on processed paths than
pending paths, so the number of vertices in processed paths that the feasibility test lands
on is at least n/2r+2. But for all positive integers i, it holds that 1/2i+2 > 2(i2 + 9)/2i2+9.
(cid:16)
2(r2 + 9)2/2r2+9(cid:17)
Thus, n/2r+2 >
n, which contradicts the assumption that the feasibility
test lands on more than n/2r+1 vertices. Hence, the median length of a processed path is at
most 2r2+9.
We are now ready to show the reduction in processed paths from repeated instances of
feasibility testing.
Lemma 4.10 Suppose at the beginning of round r, that the feasibility test lands on at most
n/2r−1 vertices. Suppose further that the feasibility test lands on more vertices in heavy paths
than middleweight paths. If the feasibility test lands on more vertices in processed paths than
vertices in pending paths, then following at most 6(r2 + 9)(408r2 + 815r + 415) iterations of
selection and feasibility testing from H, which takes O(nr4/2r) time, the number of vertices
in processed paths that the feasibility test lands on is either halved or at most n/2r+2.
Proof
By assumption, the feasibility test lands on at most n/2r−1 vertices and lands
on more vertices in heavy paths than middleweight paths, and more vertices in processed
paths than vertices in pending paths. If the number of vertices in processed paths that the
feasibility test lands on is at most n/2r+2, then the lemma immediately follows. Thus, we
assume the feasibility test lands on more than n/2r+2 vertices in processed paths.
By Lemma 4.8, the median length of a leaf path is at most 24(r−1)2. By Lemma 4.3,
in iteration i, at most 1/25k of the subpaths of length 2j−k can be pending, where 25j =
(3/4)(24/23)i. Then for j = 5(r + 1)2 and k = (r + 1)2, at most 1/25(r+1)2 of the subpaths
of length 24(r+1)2 > 24(r−1)2 can be pending, so at least half of the leaf paths are resolved.
It takes at most i = (25(r + 1)2 − log(3/4))/ log(24/23) < 408r2 + 815r + 414 iterations to
42
resolve half of the leaf paths, so that the appropriate values can be inserted into V. One
more iteration of feasibility testing is run using the selected median from V. Thus, running
at most 408r2 + 815r + 414 iterations of handle pending paths, followed by an iteration of
handle processed paths (for a total of at most 408r2 + 815r + 415 iterations) reduces the
number of leaf paths by a factor of 1/4, or equivalently, reduces the total number of paths
by a factor of 1/8.
Since (7/8)6 < 1/2, then by repeating at most 6(r2 + 9) times, the total number of paths
is reduced by a factor of at least 1/2r2+9. If the average length of the remaining processed
paths is more than 2r2+9, then by Lemma 4.9, the feasibility test lands on at most n/2r+1
vertices, which is a reduction of 1/2 > 1/4 in the number of vertices checked by the feasibility
test. Otherwise, if the average length of the remaining processed paths is less than 2r2+9,
then by reducing the total number of paths by factor of at least 1/2r2+9, the time spent
by the feasibility test on vertices in processed paths is at least halved. We require at most
6(r2 + 9) cycles, each with 408r2 + 815r + 415 iterations of feasibility testing. Thus, at most
6(r2 + 9)(408r2 + 815r + 415) = O(r4) iterations are needed to reduce the feasibility test by
at least half, each checking at most n/2r vertices, for a total of O(nr4/2r) time.
Lemma 4.11 For each round, the feasibility test time is reduced by at least 1/2. Thus at
the beginning of round r, the feasibility test lands on at most n/2r−1 vertices.
Proof We note that prior to the first round, the feasibility test lands on exactly n vertices,
and we proceed via induction. Suppose at the beginning of round r, the feasibility test lands
on at most n/2r−1 vertices.
If the feasibility test in fact lands on at most n/2r vertices,
then the induction already holds. Recall that we have three cases, as we claimed just before
Lemma 4.6. The feasibility test can spend at least as much time on middleweight paths
as on heavy paths. Otherwise, the feasibility test spends more time on heavy paths than
middleweight paths, but can spend more time on either pending paths or processed paths.
43
If the feasibility test spends at least as much time on pending paths as on processed paths,
then by Lemma 4.6, no more than 163r + 170, which is certainly less than 408r2 + 815r + 415,
iterations of selection and feasibility from H are needed to reduce the portion spent by the
feasibility test on pending paths by at least half. Hence, the overall feasibility test time is
reduced by at least 1/8.
Otherwise, if the feasibility test spends more time on processed paths than pending
paths, then by Lemma 4.10, at most (6r2 + 9)(408r2 + 815r + 415) iterations of selection and
feasibility from H are needed to reduce the portion spent by the feasibility test on processed
paths by at least half. Hence, the overall feasibility test time is reduced by at least 1/8.
If the feasibility test spends at least as much time on middleweight paths as on heavy
paths, then following a selection in U that is weighted by path length, and then a feasibility
test, at least half the vertices in middleweight paths will be determined to be either in light
paths or in heavy paths. If the vertices are determined to be in light paths, then the portion
spent by the feasibility test on these paths is reduced by at least half, since newly formed
light paths are cleaned and glued in handle middleweight paths, without increasing the time
spent by the feasibility test on heavy paths. Hence, if λ1 is increased by a feasibility test
from U, the overall feasibility test time is reduced by at least 1/8.
However, if the vertices are determined to be in heavy paths, our algorithm will insert
the representatives of the corresponding submatrices into H. Our algorithm thus reduces
the portion spent by the feasibility test on heavy paths by at least 1/8. Hence, taking the
reductions of both types into account, the overall feasibility test time is reduced by at least
1/16. Thus, we can reduce the overall number of vertices that the feasibility test lands on by
1/16 in O(nr4/2r) time, for each r. Since (15/16)11 < 1/2, then at most eleven repetitions
suffice to halve the overall number of vertices checked by the feasibility test. Indeed, each
round requires at most 11[6(r2 + 9)] = 66(r2 + 9) repetitions, and so at the beginning of
round r + 1, the runtime is at most n/2r.
44
Corollary 4.12 Round r has O(r4) calls to the feasibility test.
Proof
Since round r requires at most 66(r2 +9) repetitions of the inner loop, and the inner
loop uses at most (408r2 + 815r + 415) feasibility tests, then the total number of feasibility
tests in round r is O(r4).
Now, we claim the main result of paper.
Theorem 4.13 The runtime of our algorithm is O(n).
Proof By Lemma 4.11, the feasibility test lands on at most n/2r−1 vertices at the beginning
O(n/2r−1) time each. Hence, there are at most log n rounds. Note that (cid:80)log n
of round r. By Corollary 4.12, round r has O(r4) calls to the feasibility test, which take
r=0 nr4/2r−1 is
O(n). By Theorem 4.5, the total time for handling H and U and performing selection over
all iterations is O(n), while the same result clearly holds for V. Therefore, the total time
required by our algorithm is O(n).
5 Conclusion
Our algorithms solve the max-min and min-max path k-partition problems, as well as the
max-min tree k-partition problem in asymptotically optimal time. Consequently, they use
the paradigm of parametric search without incurring any log factors. We avoid these log
factors by searching within collections of candidate values that are implicitly encoded as
matrices or as single values. We assign synthetic weights for these values based on their
corresponding path lengths so that weighted selections favor the early resolution of shorter
paths, which will speed up subsequent feasibility tests. In fact, our analysis relies on demon-
strating a constant fraction reduction in the feasibility test time as the algorithm progresses.
Simultaneously, unweighted selections quickly reduce the overall size of the set of candidate
values so that selecting test values is also achieved in linear time.
45
Furthermore, we have successfully addressed the challenge of developing a meaningful
quantity to track progress as the algorithm proceeds. We have proved that the time to
perform a feasibility test describes in a natural way the overall progress from the beginning
of the algorithm. Unfortunately, even these observations alone are not enough to overcome
the challenges of tree partitioning. In particular, without both compressing long paths and
pruning leaf paths quickly, the feasibility test time might improve too slowly for the overall
algorithm to take linear time. Our dual-pronged strategy addresses both issues simultane-
ously, demonstrating that parallel algorithms are not essential or even helpful in designing
optimal algorithms for certain parametric search problems.
References
[1] E. M. Arkin and C. H. Papadimitriou. On the complexity of circulations. J. Algorithms,
7:134–145, 1986.
[2] R. I. Becker, Y. Perl, and S. R. Schach. A shifting algorithm for min-max tree parti-
tioning. J. ACM, 29:58–67, 1982.
[3] M. Blum, R. W. Floyd, V. R. Pratt, R. L. Rivest, and R. E. Tarjan. Time bounds for
selection. J. Comput. Syst. Sci., 7:448–461, 1972.
[4] R. Cole. Partitioning point sets in arbitrary dimensions. Theor. Comput. Sci., 49:239–
265, 1987.
[5] R. Cole. Slowing down sorting networks to obtain faster sorting algorithms. J. ACM,
34:200–208, 1987.
[6] R. Cole, J. S. Salowe, W. L. Steiger, and E. Szemeredi. An optimal-time algorithm for
slope selection. SIAM J. Comput., 18:792–810, 1989.
[7] G. N. Frederickson. Optimal algorithms for partitioning trees and locating p centers in
trees. Technical Report CSD-TR-1029, Purdue University, 1990.
[8] G. N. Frederickson. Optimal algorithms for tree partitioning. In Proc. 2nd ACM-SIAM
Symp. on Discrete Algorithms, pages 168–177, San Francisco, January 1991.
46
[9] G. N. Frederickson and D. B. Johnson. Finding kth paths and p-centers by generating
and searching good data structures. J. Algorithms, 4:61–80, 1983.
[10] G. N. Frederickson and D. B. Johnson. Generalized selection and ranking: sorted
matrices. SIAM J. on Computing, 13:14–30, 1984.
[11] D. Gusfield. Parametric combinatorial computing and a problem of program module
distribution. J.ACM, 30:551–563, 1983.
[12] S. Kundu and J. Misra. A linear tree partitioning algorithm. SIAM J. on Computing,
6:151–154, 1977.
[13] N. Megiddo. Applying parallel computation algorithms in the design of serial algorithms.
J. ACM, 30:852–865, 1983.
[14] Y. Perl and S. R. Schach. Max-min tree partitioning. J. ACM, 28:5–15, 1981.
[15] M. Reichling. On the detection of a common intersection of k-convex polyhedra. Lect.
Notes Comput. Sci., 333:180–187, 1988.
[16] J. S. Salowe. L-infinity interdistance selection by parametric search. Inf. Proc. Lett.,
30:9–14, 1989.
[17] E. Zemel. A linear time randomizing algorithm for searching ranking functions. Algo-
rithmica, 2:81–90, 1987.
47
Figure 1: Path P , vectors X1 and X2, and explicit illustration of matrix M (P )
48
Figure 2: Explicit illustration of submatrices, just after quartering on the first iteration of PATH0
49
Figure 3: Explicit illustration of submatrices at the end of the first iteration PATH0
50
Figure 4: An edge-path-partition for T
51
Figure 5: The edge-path-partition of the resulting tree after all leaf-paths are deleted
52
Figure 6: Explicit representation of initial submatrices for M (P ) in PATH1
53
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.