MIT Thesis Jsoto

Contributions on Secretary Problems, Independent
Sets of Rectangles and Related Problems

by
José Antonio Soto
Submitted to the Department of Mathematics
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2011

c 2011 José Antonio Soto. All rights reserved.
The author hereby grants to MIT permission to reproduce and to
distribute publicly paper and electronic copies of this thesis document
in whole or in part in any medium now known or hereafter created.
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Mathematics
May 18, 2011
Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Michel X. Goemans
Leighton Family Professor of Applied Mathematics
Thesis Supervisor
Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Michel X. Goemans
Chairman, Department Committee on Graduate Theses
2
Contributions on Secretary Problems, Independent Sets of
Rectangles and Related Problems
by
José Antonio Soto
Submitted to the Department of Mathematics

on May 18, 2011, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy
Abstract
We study three problems arising from different areas of combinatorial optimization.
We first study the matroid secretary problem, which is a generalization proposed
by Babaioff, Immorlica and Kleinberg of the classical secretary problem. In this
problem, the elements of a given matroid are revealed one by one. When an element
is revealed, we learn information about its weight and decide to accept it or not,
while keeping the accepted set independent in the matroid. The goal is to maximize
the expected weight of our solution. We study different variants for this problem
depending on how the elements are presented and on how the weights are assigned
to the elements. Our main result is the first constant competitive algorithm for the
random-assignment random-order model. In this model, a list of hidden nonnegative
weights is randomly assigned to the elements of the matroid, which are later presented
to us in uniform random order, independent of the assignment.
The second problem studied is the jump number problem. Consider a linear ex-
tension L of a poset P . A jump is a pair of consecutive elements in L that are not
comparable in P . Finding a linear extension minimizing the number of jumps is
NP-hard even for chordal bipartite posets. For the class of posets having two direc-
tional orthogonal ray comparability graphs, we show that this problem is equivalent
to finding a maximum independent set of a well-behaved family of rectangles. Using
this, we devise combinatorial and LP-based algorithms for the jump number problem,
extending the class of bipartite posets for which this problem is polynomially solvable
and improving on the running time of existing algorithms for certain subclasses.
The last problem studied is the one of finding nonempty minimizers of a symmet-
ric submodular function over any family of sets closed under inclusion. We give an
efficient O(n3 )-time algorithm for this task, based on Queyranne’s pendant pair tech-
nique for minimizing unconstrained symmetric submodular functions. We extend this
algorithm to report all inclusion-wise nonempty minimal minimizers under hereditary
constraints of slightly more general functions.
Thesis Supervisor: Michel X. Goemans

Title: Leighton Family Professor of Applied Mathematics
3
4
Acknowledgments
This thesis is the result of many experiences that I have encountered over the last
five years at MIT, and the support of many wonderful people I have met here. There
are always more people to acknowledge than one ever has the space and memory for.
First and foremost, I want to thank my advisor Michel Goemans. His perpetual
enthusiasm for research, as well as his patience, encouragement and guidance beyond
academic activities have been invaluable for me. Even on a very busy schedule, Michel
always managed to dedicate a nontrivial amount of time to work together on different
problems.
I also want to acknowledge all the patience and feedback provided by my thesis
committee, Jonathan Kelner and Andreas Schulz. Special thanks go to Andreas not
only for attending my thesis defense remotely from Germany, but also for inviting me
to many mittag seminare and for proposing many interesting problems.
I gratefully acknowledge the support of the National Science Foundation (NSF)
under contract CCF-0829878 and the Office of Naval Research (ONR) under grant
number N00014-11-1-0053.
I extend my gratitude to all the teachers and professors who have motivated
me to study Mathematics, specially to my undergraduate advisor Marcos Kiwi for
introducing me to the world of scientific research and for encouraging me to come to
MIT.
I do not want to forget to mention all the people who contributed to make my
time at MIT enjoyable: my officemates, colleagues and friends; students, faculty,
staff, visitors and postdocs; not only inside the math department, but also in CSAIL,
and ORC. Many thanks to the chilean community in Boston and to my neighbors
at Tang and Westgate, many of whom I am lucky to call friends. Special thanks go
to Claudio Telha who not only has been a very good friend and neighbor but also a
great coauthor: without him a big part of this thesis would have not been written.
I thank my parents, my brother and sister for their support, care and love. I would
have never made it here without them. I also thank all my friends back in Chile for
somehow always being present in my life.
Finally, but not least, I give my deepest love to my wife Giannina for her company,
infinite love and understanding. I thank her for tolerating all the time that this thesis
has stolen away from us. My last year at MIT has been the happiest by far and that
is all because of her.
5
6
Contents
Introduction 11
I Matroid Secretary Problem 21

1 Secretary Problems and Matroids 23
1.1 Introduction to Secretary Problems . . . . . . . . . . . . . . . . . . . 23
1.1.1 Comparison and Value Based Algorithms . . . . . . . . . . . . 24
1.2 Classical Secretary Problem . . . . . . . . . . . . . . . . . . . . . . . 24
1.2.1 Previous Results for the Classical Secretary Problem . . . . . 25
1.3 Multiple Choice Secretary Problem . . . . . . . . . . . . . . . . . . . 29
1.3.1 Previous Results for the Multiple Choice Secretary Problem . 29
1.4 Generalized Secretary Problems . . . . . . . . . . . . . . . . . . . . . 32
1.4.1 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.4.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.4.3 Performance Tool: Competitive Analysis . . . . . . . . . . . . 33
1.4.4 Standard Assumptions . . . . . . . . . . . . . . . . . . . . . . 33
1.4.5 Special Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.6 Lower Bound for Generalized Secretary Problems . . . . . . . 35
1.5 Matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.5.1 Operations on Matroids . . . . . . . . . . . . . . . . . . . . . 37
1.5.2 Matroid Examples . . . . . . . . . . . . . . . . . . . . . . . . 38
1.5.3 Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.5.4 Matroid Secretary Problems . . . . . . . . . . . . . . . . . . . 40
1.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2 New Results for Matroid Secretary Problems 45

2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2 Divide and Conquer . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3 Uniformly Dense Matroids . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3.1 Random-Assignment Random-Order Model . . . . . . . . . . 49
2.3.2 Random-Assignment Adversarial-Order Model . . . . . . . . . 55
2.4 Principal Partition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4.2 Matroids Related to the Principal Partition . . . . . . . . . . 63
7
2.5 General Matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.5.1 Random-Assignment Random-Order Model . . . . . . . . . . 68
2.5.2 Random-Assignment Adversarial-Order Model . . . . . . . . . 70
2.6 New Results for the Adversarial-Assignment Random-Order Model . 70
2.6.1 General O(log r)-Competitive Algorithm . . . . . . . . . . . . 71
2.6.2 Column-Sparse Linear Matroids . . . . . . . . . . . . . . . . . 74
2.6.3 Low Density Matroids . . . . . . . . . . . . . . . . . . . . . . 75
2.6.4 Cographic Matroids . . . . . . . . . . . . . . . . . . . . . . . . 77
2.6.5 Matroids with Small Cocircuits . . . . . . . . . . . . . . . . . 79
2.7 Summary and Open Problems . . . . . . . . . . . . . . . . . . . . . . 80
2.7.1 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 81
II Jump Number of Two Directional Orthogonal Ray

Graphs and Independent Sets of Rectangles 83
3 Posets and Perfect Graphs 85
3.1 Basic Notions of Posets . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2 Chains and Antichains . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3 Extensions and Poset Dimension . . . . . . . . . . . . . . . . . . . . . 87
3.4 Survey on Comparability Graph Classes . . . . . . . . . . . . . . . . 89
3.4.1 Geometric Representation of Posets in the Plane . . . . . . . . 90
3.4.2 Permutation Graphs . . . . . . . . . . . . . . . . . . . . . . . 91
3.4.3 Chordal Bipartite Graphs . . . . . . . . . . . . . . . . . . . . 93
3.4.4 Two Directional Orthogonal Ray Graphs (2dorgs) . . . . . . 93
3.4.5 Interval Bigraphs . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.4.6 Convex Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.4.7 Biconvex Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4.8 Bipartite Permutation Graphs . . . . . . . . . . . . . . . . . . 100
3.4.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.5 Perfect Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4 A Primer on the Jump Number Problem 105

4.1 Jump Number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.1.1 Complexity of the Jump Number Problem . . . . . . . . . . . 108
4.2 Cross-Free Matchings and Biclique Covers . . . . . . . . . . . . . . . 110
4.3 Related Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.1 Matrices: Boolean rank, Antiblocks and Block Covers . . . . . 111
4.3.2 Geometry: Antirectangles and Rectangle Covers . . . . . . . . 114
4.3.3 Interval Combinatorics: Bases and Irredundancy . . . . . . . . 115
4.3.4 Survey on the Complexity of the Presented Problems . . . . . 117
4.4 Summary of Presented Problems and Results . . . . . . . . . . . . . . 119
8
5 Jump Number of 2DORGs 121
5.1 Maximum Independent Sets and Minimum Hitting Sets of Rectangles 121
5.2 Geometric Interpretation for 2dorgs . . . . . . . . . . . . . . . . . . 128
5.3 Linear Programming Formulation . . . . . . . . . . . . . . . . . . . . 130
5.4 Combinatorial Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.4.1 Running Time Improvement . . . . . . . . . . . . . . . . . . . 138
5.4.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.4.3 Data Structure . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.4.4 Admissible Flips . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.4.5 Refined Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 145
5.4.6 Bounds for c.f.i. Families . . . . . . . . . . . . . . . . . . . . . 145
5.4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.5 Relation to Frank and Jordan’s Set-Pair Covering Problem . . . . . . 148
5.6 Summary of Results and Open Problems . . . . . . . . . . . . . . . . 149
6 Weighted Cross-free Matching 151

6.1 NP-Hardness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.2 Polynomial Time Algorithms . . . . . . . . . . . . . . . . . . . . . . . 151
6.2.1 Bipartite Permutation . . . . . . . . . . . . . . . . . . . . . . 151
6.2.2 Convex Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.3 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
III Constrained Set Function Minimization 157

7 Set Function Minimization under Hereditary Constraints 159
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.2 Unconstrained Minimization . . . . . . . . . . . . . . . . . . . . . . . 162
7.3 Constrained Minimization . . . . . . . . . . . . . . . . . . . . . . . . 164
7.4 Set and Bi-set Functions . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.4.1 Fusions and Minors . . . . . . . . . . . . . . . . . . . . . . . . 170
7.4.2 Submodular Functions . . . . . . . . . . . . . . . . . . . . . . 172
7.4.3 Posimodular Functions . . . . . . . . . . . . . . . . . . . . . . 174
7.4.4 Rizzi Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.4.5 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.5 Nagamochi’s Flat Pair Based Algorithm . . . . . . . . . . . . . . . . 181
7.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Bibliography 198
9
10
Introduction
In this thesis we focus on three problems in combinatorial optimization arising from

different areas: the matroid secretary problem, the jump number problem and hered-
itarily constrained minimization of symmetric submodular functions.
The first problem belongs to the realm of online selection algorithms. We wish to
select a set of elements satisfying certain properties from a stream revealed in random
order, where the decision of selecting an element or not must be made at the moment
the element arrives. We study different variants of this problem, exhibiting simple
constant competitive algorithms for some of them.
The second is a scheduling problem coming from order theory. We show a sur-
prising connection between the jump number problem of certain partially ordered
sets and the geometric problem of finding a maximum collection of disjoint rectangles
from a family of axis-parallel rectangles in the plane. Using this geometric interpre-
tation we devise efficient algorithms for this problem and explore nontrivial relations
to other problems in the literature.
The third problem is a set function minimization problem. We contribute a simple
and efficient algorithm that solves it exactly and give some extensions.
In this chapter we describe each of the aforementioned problems, some motivation
and historical background on them, and finally our results and the organization of
the corresponding chapters. The rest of the thesis is divided into three parts, each
one dedicated to one of these problems.
Basic notation
The common notation used in this thesis is fairly standard. We use Z and R to
denote the sets of integer and real numbers respectively. We also use Z+ and R+
to denote the sets of nonnegative integer and real numbers respectively. For a finite
number n, we use [n] to denote the set {1, . . . , n}. The collection of all subsets of a
set V is denoted as 2V .
For vectors v ∈ Rd , we use subindices to denote its coordinates v = (v1 , . . . , vd ).
For the specific case of d = 2 or d = 3, we also use vx , vy and vz to denote the
first, second and third coordinates. The notation RE denotes the Euclidean space of
dimension |E| over R, where the coordinates are indexed by elements of E. For every
set F ⊆ E, we use χF ∈ {0, 1}E to denote its characteristic vector, where χF (e) = 1
if e ∈ F and 0 otherwise.
We use G = (V, E) to denote a graph G having vertex set V and edge set E. On
occasions, we use G = (V, E, w) to denote a graph (V, E) along with a real weight
11
function w : E → R on its edges. To describe the asymptotic behavior of a function
when its value goes to infinity, we use the standard O(·), Ω(·), o(·), and ω(·) notations.
Finally, we use Pr(A) to denote the probability of a given event A and E[X] to
denote the expected value of a random variable X. We use subindices on Pr(·) and
E[·] if we want to be specific about the probability space over which the probability
or expectation is taken.
In each part of the thesis, we introduce specific notation for each of the studied
problems.
Matroid Secretary Problem

The optimal stopping problem known as the classical secretary problem can be re-
garded as follows. An employer wishes to hire the best secretary among n candidates
about which she has no prior information. She interviews them in random order. Af-
ter each interview, she is able to compare the candidates she has already interviewed
and must decide immediately whether to hire the current one or continue. Candidates
that are declined can not be recalled, and the employer is satisfied with nothing but
the best.
This problem has a very elegant solution. Lindley [101] and Dynkin [48] have
independently shown that the best strategy consists in observing roughly n/e of the
candidates and then selecting the first one that is better than all the observed ones.
This strategy returns the best candidate with probability 1/e. Since then, this simple
problem has been generalized and extended in many different directions, rapidly be-
coming a field in its own. People have studied different objective functions, the case
where the number of candidates is unknown, selecting multiple candidates, allowing
full or partial recall of candidates, and adding extra cost for selecting later candi-
dates, among other variations. To learn more about the early history of this problem
and some of its variants we recommend the entertaining article of Ferguson [55] and
Freeman’s survey [65]. We also mention different variants in Section 1.6.
In this thesis, we follow the line of work initiated by Babaioff, Immorlica and
Kleinberg [8]. Their work is motivated by online auctions: An online algorithm can
be regarded as an auctioneer having one or many identical items, and the secretaries
as agents arriving at random times, each one having a different valuation for the
items. The goal of the algorithm is to assign the items to the agents as they arrive
while maximizing the total social welfare, subject to some combinatorial restrictions.
In many cases, these restrictions can be modeled by matroid constraints.
More precisely, in the matroid secretary problem, the elements of the ground set
of a given matroid arrive one by one, each one having a hidden weight. Upon arrival,
an algorithm gets to know the weight and is given the choice to take the element
or reject it, where this decision is irreversible. The only restriction is that the set
of accepted elements must be independent in the matroid. The task is to select
an independent set having total sum of weights as large as possible. It is usually
assumed that the elements arrive in random order independent of their weights, but
this is not a requirement. Another model assumes that a set of adversarially selected
12
weights is randomly assigned to the elements of the matroid before the execution of
the algorithm.
Matroids are very rich combinatorial structures. They generalize certain aspects
of linear independence in vector spaces and some aspects of graph theory. As indepen-
dently shown by Rado [142], Gale [67] and Edmonds [50], matroids characterize those
hereditary set systems for which optimizing a linear objective function is attained via
a natural greedy algorithm. This property makes them suitable for online scenarios.
Because of this, Babaioff et al. [8] have conjectured that it is possible to devise a
constant competitive algorithm for the previous problem, as long as the order or the
weight assignment is randomly selected.
There has been a significant amount of work on this conjecture on the model where
the weights are adversarially selected but the order is random. Constant competitive
algorithms are known for partition matroids [101, 48, 95, 6], transversal matroids [44,
97, 162], graphic matroids [8, 5, 97] and laminar matroids [87]. For general matroids,
the best algorithm known so far, due to Babaioff et al. [8], is O(log r)-competitive,
where r is the rank of the matroid.
Our results
Part of the work of this thesis has already been presented in a recent SODA pa-
per [152]. In this thesis we partially answer Babaioff et al.’s conjecture by ex-
hibiting constant competitive algorithms for the random-assignment random-order
model. This is the model for which both the arrival order is selected at random
and the weights are randomly assigned. In [152], the author has already presented
a 2e/(1 − 1/e) ≈ 8.6005 algorithm for this model. In this work, we present a new
algorithm achieving a competitive ratio of at most 5.7187.
On a very high level our algorithm is based on a simple divide and conquer ap-
proach: replace the matroid by a collection of matroids of a simpler class for which
we can easily derive a constant-competitive algorithm, and then return the union of
the answers. The simpler matroids we use are known as uniformly dense matroids.
Uniformly dense matroids are those for which the density of a set, that is, the
ratio of its cardinality to its rank, is maximized on the entire ground set. The sim-
plest examples are precisely the uniform matroids. We show that uniformly dense
matroids and uniform matroids of the same rank behave similarly, in the sense that
the distribution of the rank of a random set is similar for both matroids. We use this
fact to devise constant competitive algorithms for uniformly dense matroids in this
model.
In order to extend the above algorithms to general matroids we exploit some
notions coming from the theory of principal partitions of a matroid, particularly its
principal sequence. Roughly speaking, the principal sequence of a matroid M is
a decomposition of its ground set into a sequence of parts, each of which is the
underlying set of a uniformly dense minor of M. Furthermore, if we select one
independent set in each of these minors, their union is guaranteed to be independent
in M. By employing separately the previous algorithms in each of these minors, we
obtain an algorithm that returns an independent set of M, while only increasing an
13
extra factor of e/(e − 1) on its competitive ratio. By comparing the weight of our
solution to the optimum of certain randomly defined partition matroids, we give a
tighter analysis for the competitive ratio of our algorithms.
As first noticed by Oveis Gharan and Vondrák [136], it is possible to also ap-
ply the methods in [152] to obtain constant competitive algorithms for the stronger
model in which the weights are randomly assigned but the order in which the elements
are presented is prescribed by an adversary (the random-assignment adversarial-order
model). In this thesis, we present alternative algorithms for this second model achiev-
ing a competitive ratio of 16/(1 − 1/e) ≈ 25.311.
Babaioff et al.’s conjecture is still open for the “standard” adversarial-assignment
random-order model. For this model, we present simple algorithms for various ma-
troid classes. We show a ke-competitive algorithm for the case in which the matroid
is representable by a matrix where each column has at most k non-zero entries.
This result generalizes the 2e-competitive algorithm for graphic matroids of Korula
and Pál [97]. We also give algorithms for general matroids having competitive ratio
proportional to the density of the matroid. Using this, we obtain a 3e-competitive
algorithm for cographic matroids, and a k-competitive algorithm for matroids where
each element is in a cocircuit of size at most k.
For general matroids, we give a new O(log r)-competitive algorithm. Unlike the
previous algorithm of Babaioff et al. [8], our algorithm does not use the numerical
value of the weights. It only needs the ability to make comparisons among seen
elements. This is a desirable property since the features revealed by the elements
may be of qualitative type (for example the qualifications of a person applying for
a job), but the actual value or profit may be an unknown increasing function of
the features revealed. In fact, all the algorithms proposed in this thesis have the
mentioned desirable property.
Organization
This part of the thesis is presented in Chapters 1 and 2.
In Chapter 1 we formally define secretary problems and describe two of the sim-
plest cases: the classical and the multiple choice secretary problems, presenting pre-
vious results for both of them. Later, we describe the generalized secretary problem,
introduced by Babaioff et al. [8], in which the sets of elements that can be simultane-
ously selected obey arbitrary hereditary constraints. We pay special attention to the
different models that arise depending on how the candidates are presented and how
the weights are assigned. Afterwards, we turn our attention to the matroid secretary
problem. Before tackling this problem, we present some background on matroids and
their relation to the greedy algorithm. We conclude the chapter by giving a brief
survey of related results.
Chapter 2 contains our new results. We start by introducing notation and the
divide and conquer idea on which our algorithms for random-assignment models are
based. Then, we study properties of uniformly dense matroids and give different
algorithms for them on both the random-assignment random-order model and the
random-assignment adversarial-order model. Afterwards, we describe the sequence of
14
uniformly dense minors arising from the principal partition of a loopless matroid and
give some background on this construction. We show how to use these matroids to
give constant competitive algorithms for general matroids in both random-assignment
models.
Later in that chapter, we focus on algorithms for the adversarial-assignment
random-order model. We present a new O(log r)-competitive algorithm for matroids
of rank r which only uses the relative order of the weights seen and not their actual
values. We also present an algorithm for linear matroids represented by a matrix A
having competitive ratio proportional to the maximum number of nonzero entries in
a column of A, and an algorithm for general matroids having competitive ratio pro-
portional to the density of the matroid. We conclude that chapter with a summary
of results and a description of open problems in the area.
Jump Number of Two Directional Orthogonal Ray

Graphs and Independent Sets of Rectangles
Although the jump number problem is a purely combinatorial problem arising from
order theory, it admits a natural scheduling interpretation: schedule a collection of
jobs on a single machine obeying the precedences induced by a partially ordered set
(poset), in such a way that the total setup cost is minimized. Here, the setup cost of
a job is 0 if the job immediately before it in the schedule is constrained to precede it,
or 1 otherwise; we say in the latter case that there is a jump in the schedule.
It is known that this problem is NP-hard even for bipartite posets [139], and
that the jump number itself, this is the minimum number of jumps that a given
poset admits over all its linear extensions, depends only on the poset’s comparability
graph [82]. The literature of this problem is vast, specially in what respects determin-
ing natural classes of comparability graphs for which the problem is polynomial time
solvable. For instance, there are polynomial time algorithms to compute this num-
ber on bipartite permutation graphs [156, 53, 16], biconvex graphs [16] and convex
graphs [40].
An important open question in this area is to determine the complexity of the jump
number problem as a function of the poset dimension. The dimension of a partial
order P is the minimum number k for which P can be expressed as the intersection
of k linear orders. Alternatively, this can be defined as the minimum k for which the
poset P can be embedded in the k-dimensional Euclidean space Rk endowed with
the natural coordinate-wise partial order relation ≤Rk . Many problems that are hard
for general posets become polynomially solvable for two-dimensional posets. This has
led Bouchitté and Habib [15] to conjecture that the jump number problem is also
polynomial time solvable in this class. This conjecture remains open today.
From the description above, two-dimensional posets can be defined geometrically
as follows. The elements are a collection of points in the plane and the precedences are
given by the natural partial order on the plane: a point precedes another if the first
is below and to the left of the second. In this thesis we consider a natural variant of
15
these posets in which the points have an associated color and relations between points
of the same colors are ignored. More precisely bicolored 2D-poset can be defined as
follows. Given a sets of red points R and blue points B in the plane, a red element r
precedes a blue element b if r is located below and to the left of b. These are the
only precedences included in the poset. The comparability graphs of these posets
correspond exactly to the two directional orthogonal ray graphs considered recently
by many authors [135, 151].
Our results
A big part of the material presented in this part of the thesis is based on joint work
with Claudio Telha [153].
In this thesis we show that solving the jump number problem on bicolored 2D-
posets defined as above, is equivalent to finding a maximum cardinality family of
disjoint axis-parallel rectangles having a red point as bottom-left corner and a blue
point as top-right corner. Furthermore, we show that this problem can be solved in
polynomial time using either a linear programming based algorithm or a combinatorial
algorithm. The combinatorial algorithm presented runs in time O(n e ω ), where 2 ≤
ω ≤ 2.376 is the exponent for the matrix multiplication problem.
As a biproduct of our work we show a min-max relation between two problems on
rectangles: for the families of rectangles previously described, the maximum size of a
disjoint subfamily of rectangles equals the minimum number of points needed to hit
every rectangle in the family. This relation between the maximum independent set
and the minimum hitting set of rectangles is not true for general rectangle families,
although it is very simple to argue that the first quantity is always at most the
second. A big conjecture in the area is whether or not the ratio between the sizes of
the minimum hitting set and the maximum independent set of any rectangle family
is bounded by a constant. Our result can be seen as a nontrivial step to answer this
conjecture. While studying the aforementioned conjecture, we observe that a simple
application of recent approximation algorithms for the maximum independent set
and the minimum hitting set problems gives a bound of O(log2 log α) for this ratio,
where α is the size of the maximum independent set, improving the existing bounds
of O(log α).
The presented min-max relation translates immediately to min-max relations be-
tween other problems related to the jump number problem of two directional or-
thogonal ray graphs: the maximum cross-free matching problem and the minimum
biclique cover problem. The minimum biclique cover problem of a bipartite graph
is also equivalent to the problem of computing the boolean rank of its biadjacency
matrix A: this is finding the minimum value k for which the s × t matrix A can be
written as the boolean product of an s × k matrix P and a k × t matrix Q. The
boolean rank is used, for example, to find lower bounds for communication complex-
ity [99]. As a corollary, we also expand the class of matrices for which the boolean
rank can be computed exactly in polynomial time.
Furthermore, we relate the previous min-max relations to other relations arising
from apparently unrelated problems in combinatorial optimization: the minimum
16
rectangle cover and the maximum antirectangle of an orthogonal biconvex board,
studied by Chaiken et al. [25], the minimum base and the maximum irredundant
subfamily of an interval family, studied by Györi [80]; and the minimum edge-cover
and the maximum half-disjoint family of set-pairs, studied by Frank and Jordán [63].
Our min-max relations can be seen in a certain way as both a generalization of Györi’s
result and as a non-trivial application of Frank and Jordán’s result. By relating our
problem to Györi’s result we also give an efficient O(n2 )-time algorithm to compute
the jump number of convex graphs, significantly improving on the previous O(n9 )-
algorithm of Dahlhaus [40].
To conclude, we also study a weighted version of the maximum cross-free matching
problem. We show that this problem is NP-hard on two directional orthogonal ray
graphs, and give combinatorial algorithms for certain subclasses where the problem
is polynomial time solvable.
Organization
This part of the thesis is presented in Chapters 3 through 6.
In Chapter 3 we introduce partially ordered sets and state some properties. Later,
we give a small survey on different comparability graph classes, focusing on subclasses
of two directional orthogonal ray graphs and discuss some useful geometrical char-
acterizations. Afterwards, we briefly discuss perfect graphs and recall some of their
properties, since we need them for some of our results.
Chapter 4 focuses on the jump number problem of a comparability graph and
surveys some previous results. In this chapter we also define two related problems:
the maximum cross-free matching and the minimum biclique cover of a graph. The
former problem is known to be equivalent to the jump number problem for chordal
bipartite graphs. Later, we review apparently unrelated problems on 0-1 matrices,
planar geometry and interval combinatorics. The reviewed problems come in pairs
of a minimization and a maximization problem; for some cases a min-max relation
between them is known. We show that all these problems are either equivalent or
special cases of the maximum cross-free matching and the minimum biclique cover
problem.
In Chapter 5 we turn our attention to computing the jump number of two di-
rectional orthogonal ray graphs. For these graphs, we show an equivalence between
the maximum cross-free matchings and minimum biclique covers and the maximum
independent set and minimum hitting sets of certain families of rectangles. Because
of this equivalence, we dedicate a section to review the two mentioned rectangle prob-
lems, paying special attention to existing linear programming relaxations. Later in
the chapter, we use these relaxations to obtain a polynomial time algorithm for the
maximum cross-free matching problem on two directional orthogonal ray graphs. Af-
terwards, we describe a combinatorial algorithm to compute both a cross-free match-
ing and a biclique cover of a given two directional orthogonal ray graph having the
same cardinality. This shows in particular that a min-max relation holds. Later, we
explore the relation between our results with previous works. In particular, we show
how to obtain our min-max relation as a particular case of a strong result by Frank
17
and Jordán [63]. Finally, we summarize our work and present open problems.
In Chapter 6 we describe the weighted version of the maximum cross-free matching
problem. We show NP-hardness for the case of two directional orthogonal ray graphs
and give some subclasses for which we can solve this problem in polynomial time.
Constrained Set Function Minimization

A real valued function f : 2V → R is called submodular if it satisfies the submodular
inequality f (A ∪ B) + f (A ∩ B) ≤ f (A) + f (B) for every pair of subsets A and B of V .
Submodularity is one of the most useful properties in combinatorial optimization (see,
e.g. Lóvasz’s article [104]).
Many important problems reduce to minimizing submodular functions. This is
solved in polynomial time by Grötschel, Lovász and Schrijver [75, 76] using the el-
lipsoid method. Currently it is possible to minimize a general submodular function
given by a value oracle in O(n6 )-time and using O(n5 ) oracle calls, where n is the size
of the ground set, by a combinatorial algorithm due to Orlin [134].
When the submodular function has more structure, it is possible to solve the
minimization problem faster. This is true, for instance for cut functions of weighted
graphs. These functions are not only submodular, but also symmetric. In this case,
the problem is equivalent to the minimum cut problem, which can be solved efficiently
using network flow based algorithms.
Nagamochi and Ibaraki [125, 126] have proposed a combinatorial algorithm to
solve the minimum cut problem on weighted graphs without relying on network flows.
After some improvements and simplifications, Queyranne [141] obtains a purely com-
binatorial algorithm that finds a nontrivial minimizer of a symmetric submodular
function running in O(n3 )-time and using only O(n3 ) function oracle calls. Here
nontrivial means a set that is neither empty or the entire ground set.
We are interested in the problem of minimizing symmetric submodular functions
over hereditary subfamilies of 2V , this is, over families that are closed under inclusion.
Many simple constraints can be expressed in this way. For example, we can add
a maximum cardinality constraint of k to our problem by considering the family
F = {X : |X| ≤ k}. It is important to notice that the problem of minimizing a
general submodular functionq over the family F just defined is already NP-hard to
approximate within an o n/ log n factor, as shown by Svitkina and Fleischer [159].
Our results
In this thesis, we show how to extend Queyranne’s algorithm to return a nontrivial

minimizer of a symmetric submodular function over any hereditary family. This
hereditary minimization problem includes, for example, the problem of finding a
planar induced subgraph in an undirected graph minimizing the number of edges
having precisely one endpoint in the vertex set of the induce subgraph. Our algorithm
runs in O(n3 )-time and uses O(n3 ) oracle calls.
18
For the unrestricted problem, Nagamochi and Ibaraki [127] have presented a mod-
ification of Queyranne’s algorithm that finds all inclusion-wise minimal minimizers
of a symmetric submodular function still using a cubic number of oracle calls. Using
similar ideas, we can also list all minimal solutions of an hereditary minimization
problem using only O(n3 ) oracle calls. As these minimal solutions can be shown to
be disjoint, there are at most n of them.
We also give some general conditions for other function classes for which our
methods can still be applied. For instance, we can find all the minimal minimizers
of a function f under hereditary constraints when f is a restriction of a symmetric
submodular function (also known as submodular-posimodular functions) or when f (S)
is defined as d(S, V \ S) for a monotone and consistent symmetric by-set map d in
the sense of Rizzi [145].
Organization
This part of the thesis is presented in Chapter 7 which is organized as follows.
We first give a brief introduction to the problem and some background on previous
work. Later, we explore Queyranne’s [141] technique to solve the unconstrained
minimization of symmetric submodular functions. This technique is based on the
concept of pendant pairs which we also describe in detail. We explain what are the
conditions that a general set function must satisfy in order for this algorithm to work.
Afterwards, we move to the problem of minimizing set functions under hereditary
constraints. We show how to modify the pendant pair technique in order to find a
minimizer of the constrained problem. We further modify our algorithm so that it
outputs all inclusion-wise minimal minimizers.
We later focus on the function class for which the previous techniques work. This
class includes symmetric submodular functions, their restrictions and more general
functions that we denote as weak Rizzi functions.
After the completion of this work, we were informed of an independently discovered
algorithm of Nagamochi [124] which is able to perform a slightly stronger feat for
certain function class that includes symmetric submodular functions. In the last part
of this chapter we compare our results.
19
20
Part I
Matroid Secretary Problem
21
22
Chapter 1
Secretary Problems and Matroids
In this chapter we define secretary problems and describe two of the simplest cases:
the classical secretary problem and the multiple choice secretary problem. Later, we
describe a common generalization known as the matroid secretary problem introduced
by Babaioff et al. [8]. We conclude by a brief survey of known results in this area.
1.1 Introduction to Secretary Problems

We call secretary problem any process in which we select one or more elements
from a stream, under the condition that every element must be selected or rejected
at the moment it is revealed.
Secretary problems arise as an idealization of real world sequential decision-making
tasks. Examples of tasks modeled by secretary problems include selling an item to the
highest bidder and hiring the most qualified person applying for a job – for example
a secretary position (this is the reason why this problem got its, in my opinion,
unfortunate name). Even though online selection problems like the ones mentioned
seem natural, it was not until the second half of the twentieth century that serious
attempts to formalize and study these problems were undertaken.
Many variations of the secretary problem have been introduced under different
names or in the form of recreational games. Usually, these variants are obtained by
changing some parameters of the problem, for example:
• Do we know the number of elements a priori?
• In what order are the elements presented?
• Can we select more than one element? If so, what sets of elements can we select
simultaneously?
• What information we receive at the moment we examine an element?
• What information we know a priori about the elements?
• What is our objective function?
23
The rather large number of variants makes it very tedious to try to study each of
them in detail. Since this is not our objective, we only focus on a few ones. In the
rest of this section, we formulate the exact variants we attempt to study and explore
some previous work on them.
1.1.1 Comparison and Value Based Algorithms

Consider the two following toy problems.
In the beauty contest, a collection of n contestants is presented one by one and
in random order to a judge. After evaluating each contestant, the judge must make
a choice before seeing the next one. He can either declare that contestant to be the
winner, at which point the contest ends; or keep looking, disqualifying the current
contestant. If the judge reaches the last contestant, she is declared the winner by
default. What is the judge’s best strategy to select the best contestant?
In the game of googol, Alice is given n slips of paper. In each paper she must
write a positive real number. After that, the slips are put face down on a table and
shuffled. Bob starts turning the slips face up and his aim is to stop turning when
he sees a number which he guesses is the largest of the series. What is Bob’s best
strategy?
Both problems are very similar: the objective is to select the best of a sequence
of randomly sorted elements. However, they differ in one small detail: in the second
problem, Bob can actually see the values that Alice has written, and can use that
information for his advantage. For example, if he knew that Alice is prone to select
her numbers uniformly from a fixed range, he can probably devise a better strategy.
In the first problem, this information is of no use, since the judge can not see values:
he can only compare previously presented candidates.
We say that a decision maker or algorithm is comparison-based if it is able
only to ‘rank’ or ‘compare’ previously seen elements, but it is not able to ‘see’ their
actual values. Otherwise, we say it is a value-based algorithm. In this work we
focus mostly on comparison-based algorithms (as in the case of the beauty contest
problem). However, for reasons that will become apparent later, we won’t restrict
ourselves only to cases where the elements are revealed in a uniform random order.
We also assume that the decision maker is able to break ties in a random but
consistent way using a uniformly random permutation τ : [n] → [n] as follows: if
two of the revealed elements have the same value, the one having larger τ -value is
considered larger.
1.2 Classical Secretary Problem

The setting for the classical secretary problem is as follows.
Consider a set of n nonnegative hidden weights w1 ≥ w2 ≥ · · · ≥ wn ≥ 0.
These weights are presented to an online algorithm in some order. When a weight is
presented, the algorithm is able to compare it to previously seen ones, but can not see
its numerical value (i.e. it is a comparison-based algorithm). Using this information,
24
the algorithm must decide whether to select the element or to ignore it and continue.
Its objective is to select an element as heavy as possible.
We focus on two different models:
1. Full-adversarial model: An adversary selects both the hidden weights and

the order in which the elements are presented to the algorithm.
2. Random model: An adversary selects the weights, but they are presented to
the algorithm in uniform random order.
Every adversarial selection is taken before the execution of the algorithm. In

other words, the adversary is oblivious to the decisions and random coins tosses of
the algorithm.
Most of the literature regarding the classical secretary problem focus on design-
ing algorithms that maximize the probability of accepting the largest element in the
sequence. As we are interested in generalizations of this problem we instead take the
route of competitive analysis.
Let ALG be the set returned by an algorithm A for the classical secretary problem.
The set ALG can be either a singleton or the empty set. The payoff of A is the
P
weight of the set ALG: w(ALG) = x∈ALG w(x). Usually, the competitive ratio
of A is defined as the ratio between the maximum possible payoff (in this case,
the maximum weight) and the algorithm’s payoff. As there is randomness involved,
expected values are used. However, since the payoff could be zero, the previous ratio
is not always well defined. We fix this by using the following definition.
We say that a randomized algorithm is α-competitive if for any adversarial
selection that the model allows,
E[αw(ALG) − w∗] ≥ 0, (1.1)
where the expectation is over all the random choices given by the model and the
random choices performed by the algorithm, and w∗ is the maximum payoff of a pos-
sible answer under the current realization of the weights (in this case, the maximum
weight w1 of the stream). The competitive ratio of an algorithm is the minimum
value of α for which this algorithm is α-competitive.
1.2.1 Previous Results for the Classical Secretary Problem

Consider the naive algorithm for any model of the classical secretary problem that
chooses a random element of the stream. Since this algorithm recovers at least 1/n-
fraction of the weight of the maximum element, it is n-competitive.
We show that the full-adversarial model is, as expected, really hard and that for
this model the naive algorithm above is optimal.
Lemma 1.2.1 (Folklore). No algorithm for the full-adversarial classic secretary prob-
lem admits a competitive ratio smaller than n.
25
Proof. Let ε > 0 and fix any (possibly randomized) algorithm A for the prob-
lem. Consider the adversarial input given by the increasing sequence of weights
ε, 2ε, 3ε, . . . , nε. Let pi be the probability that A selects the i-th element on this
sequence. Since the expected number of elements that A selects is at most one:
n
X
pi ≤ 1. (1.2)
i=1
In particular, there is an index j for which pj ≤ 1/n. Consider the sequence of

weights given by ε, 2ε, . . . , (j − 1)ε, n2 , 0, . . . , 0. Since the information available for
the algorithm on this sequence up and including the revelation of the j-th element is
equal to the information available for the totally increasing sequence (as the algorithm
can only make comparisons), the probability that A selects the j-th element on this
second sequence is still pj . Hence, for the set ALG returned by the algorithm on the
second sequence:
E[w(ALG)] ≤ pj n2 + (j − 1)ε ≤ n(1 + ε). (1.3)
Since n2 is the maximum payoff of the second sequence, the competitive ratio of
algorithm A is at least n/(1 + ε) ≥ n(1 − ε). Because ε is arbitrary, we conclude the
proof.
Unlike the full-adversarial model, it is very easy to get a constant competitive
algorithm for the random model. Before doing that, we give the following observation.
Lemma 1.2.2. Any algorithm guaranteeing a probability p of selecting the largest

element in the random model of the classical secretary problem is 1/p-competitive and
vice versa.
Proof. For any algorithm selecting the largest element with probability at least p in
the random model,
E[1/p · w(ALG) − w∗] ≥ 1/p · pw∗ − w∗ ≥ 0. (1.4)
Conversely, let A be a 1/p-competitive algorithm for the random model. Let 0 <
ε < 1, L = 1/ε and consider running algorithm A on the instances induced by the n!
permutations of the set of weights W = {L, L2 , . . . , Ln }. Let q be the probability that
A selects the largest element of the set. This probability is inherent to the algorithm
and does not depend on the set W . Since the algorithm is 1/p-competitive for the
random model,
0 ≤ E[1/p · w(ALG) − w∗ ] ≤ 1/p · (qLn + Ln−1 ) − Ln = Ln /p · (q + ε − p).
This implies that q ≥ p − ε. Using that ε is arbitrary, we conclude the proof.

Consider the following simple algorithm for the random model. Observe and reject
the first bn/2c elements and then accept the first element (if any) that is better than
all element seen so far. We claim that this algorithm selects the best element in the
sequence with probability at least 1/4 and thus, it is a 4-competitive algorithm. We
26
only show this for even values of n (as we will show a much more general bound later).
Indeed, it is easy to see that if the second-best weight appears in the first n/2 positions
and the top-weight appears in the last n/2-positions, then this algorithm returns the
top-weight. The claim follows since the event just described has probability 1/4.
In 1961, Lindley [101] gave a dynamic programming argument to show that the
optimal algorithm for this problem is very similar to the one above: for every value
of n, there is a number r(n) such that the algorithm maximizing the probability
of selecting the best element has the following form. Observe and reject the first
r(n) − 1 elements and then accept the first element (if any) that is better than all
element seen so far. In 1963, Dynkin [48] gave an alternative proof of the previous
fact using Markov chain arguments.
Lemma 1.2.3 (Lindley [101], Dynkin [48]). Consider the algorithm for the random
classical secretary problem obtained by observing and rejecting the first r − 1 elements
and then accepting the first element (if any) that is better than all elements seen so
far. The probability p(r) of selecting the best element under this algorithm is
1 r − 1 n−1
X 1
p(1) = and p(r) = , for 1 < r ≤ n.
n n i=r−1 i
The optimum value r(n), for n ≥ 2 is obtained by picking the only r such that
n−1 n−1
X 1 X 1
≤1< . (1.5)
i=r i i=r−1 i
For a nice exposition and proof of the result above, we refer to the survey article
of Gilbert and Mosteller [70]. Two properties of the above lemma are the following.
1. The optimum value r(n) is such that limn→∞ r(n)/n = 1/e.
2. The optimal probability p(r(n)) is decreasing with n and limn→∞ p(r(n)) = 1/e.
In particular, in order to obtain an almost optimal e-competitive algorithm it is

enough to set r(n) to be roughly n/e and proceed as above.
Lindley–Dynkin algorithm is deterministic. Interestingly, we can obtain a ran-
domized e-competitive algorithm with a somewhat simpler analysis by choosing the
number of elements to “observe and reject” as a random variable with expectation
n/e. This algorithm, denoted as Algorithm 1 below, is also easier to generalize to
other versions of the secretary problem.
Algorithm 1 For the random-order classical secretary problem.

1: Choose m from the binomial distribution Bin(n, p).
2: Observe and reject the first m elements – call this set the sample.
3: Accept the first element (if any) that is better than all the sampled ones.
27
Lemma 1.2.4.
Let wi be the i-th top weight of the stream. Algorithm 1 returns an empty set with
probability p, and returns the singleton {wi } with probability at least
Z 1
(1 − t)i−1
p dt.
p t
In particular, it returns w1 with probability at least (−p ln p). Therefore, by setting

p = 1/e, we obtain an e-competitive algorithm.
Proof. Consider the following offline simulation. Let w1 ≥ w2 ≥ · · · ≥ wn ≥ 0 be

the weights in non-increasing order. Each weight wi selects an arrival time ti in
the interval (0, 1) uniformly and independently. The offline algorithm processes the
weights in order of arrival, considering the ones arriving before time p as the sample.
When a weight arriving after time p is processed, the algorithm checks if this weight
is larger than every sampled weight. In that case, the algorithm selects the currently
processed weight. Otherwise, it continues.
Since the cardinality of the sampled set has binomial distribution with parameters
n and p, the set returned by this simulation has the same distribution as the one
returned by Algorithm 1. Hence, we analyze the simulation.
Consider the i-th top weight wi of the stream. Condition on the time ti of its
arrival being at least p (as otherwise the algorithm will not select it). Condition also
on the event Ei that all (i − 1) weights higher than wi arrive after time ti .
Let A(ti ) = {wj : tj < ti } be the set of weights seen before wi arrives. If this set
is empty then wi is selected. If not, let wA be the top weight of A(ti ). It is easy to
see that under the previous conditioning, wi is selected if and only if tA < p.
Conditioned on ti , Ei and on the set A(ti ) being nonempty, the variable tA is a
uniform random variable in (0, ti ). Hence,
Pr(wi is selected | ti ≥ p, Ei )
= Pr(A(ti ) = ∅ | ti ≥ p, Ei ) + Pr(A(ti ) 6= ∅, tA < p | ti ≥ p, Ei )
≥ 1 · Pr(A(ti ) = ∅ | ti ≥ p, Ei ) + Pr(A(ti ) 6= ∅ | ti ≥ p, Ei )p/ti
≥ p/ti .
As event Ei has probability (1 − ti )i−1 , we have

Z 1
p
Pr(wi is selected) ≥ (1 − ti )i−1 dti .
p ti
In particular the probability that w1 is selected is p1 pt dt = −p ln p. This is maximized

R
by setting p = 1/e. For that value, the probability above also equals 1/e.
To conclude the proof, note that the only way that this algorithm returns an empty
set is when the top weight w1 arrives before p. This happens with probability p.
28
1.3 Multiple Choice Secretary Problem
A natural generalization of the classical secretary problem is the multiple choice
secretary problem, also known as the r-choice secretary problem. In this
problem, the algorithm is allowed to select a set ALG consisting on up to r elements
of the stream, where r is a fixed parameter. The algorithm must follow the same
rules as in the classical secretary problem: it has to select or reject elements as they
appear. The full-adversarial model and the random model are defined analogously
to the classical case. The payoff of the algorithm is, as before, the total weight of
P
the set selected: w(ALG) = x∈ALG w(x).
The r-choice secretary problem has a nice interpretation as an online auction. We
can regard the algorithm as an auctioneer having r identical items, and the secretaries
(the weights) as agents arriving at random times, each one having a different valuation
for the item. The goal of the algorithm is to assign the items to the agents as they
arrive while maximizing the total social welfare.
1.3.1 Previous Results for the Multiple Choice Secretary

Problem
Lemma 1.3.1 (Folklore). No algorithm for the full-adversarial r-choice secretary
problem admits a competitive ratio smaller than n/r.
Proof. We offer a proof very similar to the proof of Lemma 1.2.1. Let 0 < ε < 1 and
fix any (possibly randomized) algorithm A for the problem with 1 ≤ r ≤ n. Consider
the adversarial input given by the increasing sequence of weights ε, 2ε, 3ε, . . . , nε. Let
pi be the probability that A selects the i-th element on this sequence. Since the
expected number of elements that A selects is at most r:
n
X
pi ≤ r. (1.6)
i=1
In particular, there is an index j for which pj ≤ r/n. Consider the sequence of

weights given by ε, 2ε, . . . , (j −1)ε, n3 , 0, . . . , 0. Since the information available for the
algorithm on this sequence up and including the revelation of the j-th element is equal
than the information available for the totally increasing sequence, the probability that
A selects the j-th element on this second sequence is still pj . This means that, for
the second sequence:
E[w(ALG)] ≤ pj n3 + r(j − 1)ε ≤ n2(r + ε). (1.7)
As the maximum payoff of the second sequence is bounded below by n3 , the compet-
itive ratio of algorithm A is at least w(OPT)/E[w(ALG)] ≥ n/(r + ε) ≥ (1 − ε)n/r.
Since ε is arbitrary, we conclude the proof.
A natural optimum algorithm for the full-adversarial model would be simply to

select a random set of r elements from the stream. Since each one of the top r weights
29
(say w1 ≥ w2 ≥ · · · ≥ wr ) is selected with probability at least r/n we obtain that
E[w(ALG)] ≥ (Pri=1 wi)r/n, implying this algorithm is n/r-competitive.
We have seen that for the full-adversarial model, the multiple-choice secretary
problem admits algorithms having better competitive ratios than the ones for the
classical secretary problem. A natural question to ask is whether the same happens
for the random model. Kleinberg [95] has shown that this holds asymptotically: the
random multiple-choice secretary problem admits algorithms with competitive ratio
tending to 1 as r goes to infinity.
Lemma 1.3.2 (Kleinberg [95]). Kleinberg’s algorithm for the random multiple-choice
√
secretary problem returns a set having total expected weight at least 1 − 5/ r times
q r largest
the sum of the weights of the elements. In our notation, his algorithm has a
competitive ratio of 1 + O 1/r .
Kleinberg’s algorithm is recursively defined. If r = 1, then simply apply Lindley–

Dynkin algorithm. If r > 1, then let m be a random variable from the binomial
distribution Bin(n, 1/2). On the first m weights of the stream apply recursively
Kleinberg’s algorithm to select br/2c weights. Once they have been selected, let x be
the top br/2c-th weight of the first m arriving ones. Use x as a threshold to select the
remaining weights: accept any weight that is better than x until r weights have been
selected or until we have run out of weights. Kleinberg offers a simple analysis for his
algorithm in [95] which we do not reproduce. Also, he has claimed (but the proof, as
far as the author knows, has not been published anywhere in the literature) that this
algorithm is asymptotically optimal, in the sense that any algorithm for q
this problem

returns a set having weight at least has competitive ratio at least 1 + Ω 1/r .
A much simpler suboptimal constant-competitive algorithm for this problem was
proposed later by Babaioff et al. [6]: maintain the set T consisting on the top r
elements seen so far (initialize this set with r dummy elements that are considered
to be smaller than everyone else). Observe and reject the first m = bn/ec elements
without adding them to the output; refer to this set as the sample. An element
arriving later will be added to the output only if at the moment it is seen, it enters T
and the element that leaves T is either in the sample or a dummy element. Babaioff
et al. have shown that this algorithm returns a set of at most r elements and that
every element of the optimum is in the output of the algorithm with probability at
least 1/e, making this algorithm e-competitive.
The previous algorithm is deterministic. In the same way we modified Lindley–
Dynkin algorithm for the classical secretary problem we can modify Babaioff et al.’s
algorithm to obtain a randomized e-competitive algorithm having simpler analysis.
The modification is depicted as Algorithm 2. The only difference with respect to
the algorithm above is that the number of sampled elements is given by a binomial
distribution Bin(n, p).
Lemma 1.3.3. Algorithm 2 is (−p ln p)−1 -competitive. In particular, by setting p =

1/e, we obtain an e-competitive algorithm for the random model of the multiple choice-
secretary problem.
30
Algorithm 2 For the random r-choice secretary problem.
1: Maintain a set T containing the top r elements seen so far at every moment.
Initialize T with r dummy elements.
3: Observe and reject the first m elements – call this set the sample.
4: for each element x arriving after the first m elements do
5: if x enters T and the element leaving T is in the sample or is a dummy element.
then Add x to the output ALG.
6: end if
7: end for
Proof. The proof of this lemma is very similar to the proof of Lemma 1.2.4.
Consider the following offline simulation algorithm. Let w1 ≥ w2 ≥ · · · ≥ wn ≥ 0
be the weights in non-increasing order. Each weight wi selects an arrival time ti in
the interval (0, 1) uniformly and independently. The offline algorithm maintains a set
T containing at every moment the top r weights seen so far, initializing T with a set
of r dummy weights. The algorithm then processes the incoming weights in order of
arrival, sampling the ones arriving before time p. When a weight arriving after time p
is processed, the algorithm checks if this weight enters T and if the one leaving T is in
the sample or a dummy weight. If this is the case, the algorithm selects it. Otherwise,
it continues.
Since the cardinality of the sampled set has binomial distribution with parameters
n and p, the set of elements returned by this simulation has the same distribution as
the one returned by Algorithm 2. Hence, we analyze the simulation.
As every weight selected by the algorithm pushes out of T either a sampled weight
or a dummy weight, the algorithm can select at most r elements. We estimate the
probability that each one of the top r weights appears in the output set. Focus on
one such weight wi , i ≤ r and condition on the event that it is not sampled, i.e. that
ti ≥ p.
Let A(ti ) = {wj : tj < ti } be the set of weights seen before wi arrives. If this
set has less than r elements then the element leaving T at time ti will be a dummy
weight. Consider the case where A(ti ) has cardinality at least r and let wA be the
r-th top element of this set. Note that this element wA and the time of its arrival tA
are independent random variables. In particular, conditioned on ti , tA is a uniform
random variable in (0, ti ). Hence,
Pr(wi is selected | ti ≥ p) = Pr(|A(ti )| < r | ti ≥ p) + Pr(|A(ti )| ≥ r, tA < p | ti ≥ p)

≥ 1 · Pr(|A(ti )| < r | ti ≥ p) + Pr(|A(ti )| ≥ r | ti ≥ p)p/ti
≥ p/ti .
31
Then,
Z 1
p
Pr(wi is selected) ≥ dti = −p ln p.
p ti
Therefore,
r r
E[w(ALG)] =
X X
wi Pr(wi is selected) ≥ (−p ln p) wi .
i=1 i=1
By setting p = 1/e, we obtain a competitive ratio of e.

In Section 2.3.1 we propose yet another simple algorithm, Algorithm 4, having a
slightly worse competitive ratio, but having better behavior for some extensions of
the r-choice secretary problem.
1.4 Generalized Secretary Problems

Babaioff et al. [8, 7] have proposed a framework that generalizes both the classical
and multiple-choice secretary problems. In what follows, we describe the setting,
models, standard assumptions and the tools we use to measure the performance of
an algorithm on this framework.
1.4.1 Setting
In the generalized secretary problem there is a ground set S of so-called ele-
ments and a fixed collection of subsets I ⊆ 2S closed under inclusion that describes
the sets of elements that can be simultaneously accepted. These sets are called the
feasible sets, and I is known as the domain of the problem. Each element x ∈ S
has a hidden nonnegative weight w(x). The elements are presented to an online
algorithm in a certain order x1 , . . . , xn .
As each element is presented the algorithm must decide whether to select it or
reject it subject to the following constraints.
1. The decision to accept or reject an element is irreversible.
2. This decision must be made before the next element is presented.
3. The set of selected elements, denoted as ALG, must belong to I at all times.
The payoff of the algorithm is the total sum of the weights of the selected ele-
ments: X
w(ALG) = w(x).
x∈ALG
1.4.2 Models
We classify generalized secretary problems based on how the weights and the order
of the elements are chosen.
32
Assignment of weights. Consider the following two variants.
1. Adversarial-Assignment: The weights are assigned to the elements of S by

an adversary. This information is hidden to the algorithm.
2. Random-Assignment: An adversary selects a list W = {w1 , . . . , wn } of

weights, hidden from the algorithm. This list is assigned to the elements of S
via a uniform random bijection.
Order of the elements. Consider the following two variants.
1. Adversarial-Order: An adversary selects the ordering of the elements, this

order is unknown to the algorithm.
2. Random-Order: The elements are presented in uniform random order.
There are four models for this problem arising from combining the variants above.
We assume that the random choices are performed after the adversarial choices. For
example, in the adversarial-assignment random-order model, the adversary selecting
the weights does not know the order in which the elements are going to be presented. If
randomness is involved for both parameters, we assume the choices to be independent
random variables.
1.4.3 Performance Tool: Competitive Analysis

We use the same performance tool as in the classical and multiple-choice secretary
problems: the competitive ratio.
For any given model, we say that a (randomized) algorithm is α-competitive if
for every adversarial selection that the model allows,
E[αw(ALG) − w∗] ≥ 0, (1.8)
where the expectation is over all the random choices given by the model and the
random choices performed by the algorithm and w∗ is the maximum payoff of a
feasible set under the current assignment of weights. The competitive ratio of an
algorithm is the minimum value of α for which this algorithm is α-competitive.
1.4.4 Standard Assumptions

Unless explicitly stated otherwise, we assume that the algorithm has full access to
the ground set S and the family I at all moments, even if I is exponentially larger
than S (It is often the case though, that a compact representation of I is available.)
When an element x is revealed, the algorithm learns its role in S (that is, it is able
to distinguish x from any other element in S).
We make this caveat here since one could consider a scenario in which we replace
the above ability by the following one: the algorithm can only test if a given collection
of elements already revealed is feasible or not. For example, consider the case where
33
the feasible sets are all the sets of size at most 2, except for a single pair X = {x, y}.
In this setting we can not know if the element we are accepting is x or y until both
elements have already been presented.
We also assume that at every moment, the algorithm is able to compare the values
of two previously presented elements, but it is not able to see their numerical values.
In other words, we look for comparison-based algorithms.
Regarding computational power, we assume that between the arrival of two con-
secutive elements, the algorithm is able to compute any function on the data it already
knows. The polynomiality of the algorithm is not the real issue: unbounded com-
putational time and space, or even the ability to solve the halting problem is not
enough, in some cases, to achieve small competitive ratios. Having a polynomial-time
algorithm is seen as a desirable feature, but it is not a requirement.
1.4.5 Special Cases

The two simplest problems fitting the settings for the generalized secretary problem
are the classical and multiple-choice secretary problems. In these problems, the do-
mains are the collection of all sets having cardinality at most one (classical case) or at
most a fixed number r (multiple-choice case). We denote these domains as uniform
domains.
Since uniform domains are totally symmetric, some of the models we have defined
in Section 1.4.2 coincide. The adversarial-assignment adversarial-order model on
uniform domains is equivalent with what we called the full-adversarial model for the
classical and multiple-choice secretary problems. The other three models (random-
assignment random-order model; random-assignment adversarial-order model; and
adversarial-assignment random-order model) are equivalent to the random model for
the classical and multiple-choice secretary problem.
Generalized secretary problems on different domains have many applications, spe-
cially for online auctions. Consider the following examples.
1. A popular website has space for advertising. This space is used to display a
(probably) different banner ad to every visitor. The owner wishes to sell the ad
slot to different advertisers. Each one arrives with a request for some number
of impression to be displayed on a given day, and the price he is willing to pay.
The owner accepts or rejects each request provided that the total number of
impressions sold does not exceed the estimated number of different impressions
the site gets that day. In the formalism described before, each element x of the
set S of advertisers is assigned a length fx , defined as the fraction of the total
number of impressions he wants, and the domain I consists on those sets whose
total length is at most 1. The domain for this problem is called a knapsack
domain.
2. An auctioneer has a set M of non-identical items. Each agent x ∈ S wants

to buy a specific item but the auctioneer is not allowed to sell more than k ≤
|M | items. The domain I of feasible sets that the auctioneer can choose in
34
the associated generalized secretary problem is known as a truncated partition
matroid.
3. Certain internet provider has a collection of new high speed servers in a town, in-
terconnected via routers forming a network. In order to have access to these new
servers, clients must buy private routes. The sets of clients that this provider
can serve are exactly those sets which can be connected to the servers using
vertex disjoint paths. The domain I of feasible sets of clients is known as a
gammoid.
Besides the examples above, there are other domains of interests: For instance
matching domains (set of matchings of a graph or hypergraph), matroid domains
(independent sets of a given matroid) and matroid intersection domains. Matroid
domains appear often in many applications, for that reason there has been a long line
of work dedicated to them. In Section 1.5 we focus on the matroid secretary problem
and give some background on previous work.
1.4.6 Lower Bound for Generalized Secretary Problems

There is not much to say about the adversarial-assignment adversarial-order model
for general domains. Since this problem includes uniform domains of unit size we
know, using Lemma 1.2.1, that no algorithm achieves a competitive ratio smaller
than n for general domains. For this reason, we only focus on the other three random
models: adversarial-assignment random-order, random-assignment adversarial-order
and random-assignment random-order.
Uniform domains admit constant-competitive algorithms on these three models
(as they are equivalent on these domains). However, the following example, given by
Babaioff et al. [8] shows that general domains do not have this feature.
Let S = [n] be the ground set and let r = bln nc. Partition S into m = dn/re
subsets S1 , . . . , Sm each having size r or r − 1. Consider the domain
I = {I ⊆ [n] : I ⊆ Si for some i}. (1.9)
Lemma 1.4.1 (Babaioff et al. [8]). For any model of the generalized secretary prob-
lem, no algorithm has competitive ratio smaller than o(log n/ log log n) for the do-
main I defined above.
Proof. Babaioff et al.’s argument is to show that under certain independent assign-
ment of weights, the expected weight of the set selected by any online algorithm is less
than 2 while the expected value of the maximum weight in I is Ω(log n/ log log n).
We reproduce their argument here for completeness.
For every element x ∈ S, assign a weight w(x) at random with the following
probabilities:
1 1
Pr(w(x) = 1) = and Pr(w(x) = 0) = 1 − . (1.10)
r r
Consider any randomized algorithm A for this problem. Suppose that the first element
that A selects in one execution is x ∈ Si . Then all the elements selected later must
35
also be in Si . The weights of those elements are independent from the information
observed so far. Since the weight of x is at most 1 and there are at most r additional
elements that can be selected, each with expected value 1/r, the expected value of
the set selected by the algorithm is at most 2. Note that this is true regardless of the
order in which the elements are presented: Even if the algorithm could choose the
order in which the elements are presented, the expected value returned would still be
at most 2.
For the rest of the proof, let j = b 2 lnr r c ≤ 2 lnlnlnn n and let Ei be the event that at
least j elements of Si have value 1. The probability of this event is
  ln n
j
1 1 1

X 2 ln ln n
Pr  w(x) ≥ j  ≥ ≥ =√ . (1.11)
x∈Si r ln n n
Since the events E1 ,. . . , Em are independent, the probability that none of them occurs
m
is at most 1 − √1n = o(1). If one of the events Ei occurs, then the corresponding
set Si ∈ I has weight at least j. The previous implies that the expected value of the
log n
maximum weight set in I is at least j = Ω log log n .
The previous lower bound for the optimal competitive ratio is probably not tight.
Currently, there are no known upper-bounds for generalized secretary problems apart
from the trivial bound of n.
1.5 Matroids
In view of Lemma 1.4.1 in the previous section, a natural question is to determine
other domains (apart from uniform) admitting constant-competitive algorithms for
some of the three random models. Matroid domains are interesting candidates to
consider. In this section, we define these domains and give some important properties
and examples.
A matroid is a pair M = (S, I), where S is denoted as the ground set and
I ⊆ 2S is the family of independent sets, satisfying the following properties.
1. (Nonempty) ∅ ∈ I.
2. (Hereditary) If A ∈ I and B ⊆ A, then B ∈ I.
3. (Augmentation) If A, B ∈ I and |A| < |B|, then there is x ∈ B \ A such that

A ∪ {x} ∈ I.
Matroids satisfy many properties. For a nice exposition of matroids we refer the
reader to Oxley’s [137] and Schrijver’s [147] books.
In this thesis, we use the following matroid related definitions. A basis of a
matroid M is an inclusion-wise maximal independent set. The augmentation property
of matroids guarantees that all the bases of M have the same cardinality. The size
of a base is known as the total rank of M. The rank of a set A in M, denoted as
36
rkM (A) is the size of any maximal independent set contained in A. A circuit is a
minimal non-independent set, and a loop is a singleton circuit. The span of a set A,
denoted span(A) is the set of all elements x in S, such that rkM (A) = rkM (A ∪ {x}).
An important property, that we often use, is that the rank function rkM : S → Z+
is submodular, this is,
rkM (A ∪ B) + rkM (A ∩ B) ≤ rkM (A) + rkM (B), for all A, B ⊆ S. (1.12)
In the following sections we define some matroids and explore some properties.
1.5.1 Operations on Matroids

There are many ways to construct matroids, starting from existing ones. Here we
describe some of the operations that allow us to do that.
1. Truncation: Given M = (S, I) and an integer k ≥ 0, the truncation of M to

size k is the matroid M(k) = (S, I 0 ) having the same ground set, where a set
is independent in the new matroid if it is also independent in the original one
and has cardinality at most k.
2. Restriction: Given M = (S, I) and a subset E ⊆ S of elements, the restriction

of M to E is the matroid M|E = (E, I 0 ) having E as ground set and where a
subset of E is independent if it is independent in the original matroid.
3. Contraction: Given M = (S, I) and a subset E ⊆ S, the matroid obtained

by contracting E is M/E = (S \ E, I 0 ) having S \ E as ground set, where the
independent sets are defined as follows. Consider any maximal independent set
X inside E, a set I is independent in M/E if and only if I ∪ X is independent
in M.
4. Duality: Given M = (S, I), the dual matroid M∗ = (S, I ∗ ) is the matroid
having the same ground set as M where the bases are the complement of the
bases of M. Equivalently, a set I is independent in M∗ if S = spanM (S \ I).
5. Direct Sum: Given two matroids M1 = (S1 , I1 ) and M2 = (S2 , I2 ), the direct
sum M1 ⊕ M2 is the matroid having as ground set two disjoint copies of S1
and S2 (i.e. even if S1 and S2 intersect in an element x, there are two copies of
this element in the new ground set), where a set I is independent if and only
if the set of elements of I belonging to the copy of Si is independent in Mi for
both i = 1 and i = 2.
6. Union: Given two matroids M1 = (S1 , I1 ) and M2 = (S2 , I2 ) where the

ground sets possibly intersect, the union M1 ∪M2 is the matroid having S1 ∪S2
as ground set and where a set I is independent if it can be written as I1 ∪ I2 ,
where I1 ∈ I1 and I2 ∈ I2 .
37
1.5.2 Matroid Examples
1. Free matroids.
The free matroid on S is the matroid having ground set S where every set is
independent.
2. Uniform matroids.
For r ≥ 0, the uniform matroid U(S, r) is the matroid having S as ground set
and where a set I is independent if it has cardinality at most r. In other words,
it is the truncation of the free matroid on S to size r.
3. Partition matroids.
A partition matroid is the direct sum of uniform matroids.
4. Transversal matroids. A matroid M = (S, I) is transversal if there is a

family of subsets C1 , . . . , Cr ⊆ S such that a set I is independent if and only
if we can assign each element of I to a set Ci in the family containing it,
in an injective way. In other words, a transversal matroid is the union of
rank 1 uniform matroids. Equivalently, one can define transversal matroids via
a bipartite graph: the ground set are the vertices on one part of the bipartition,
say A, and the independent sets are those subsets of A that can be covered by
a matching.
5. Laminar matroids. A matroid M = (S, I) is a laminar matroid if there is

a laminar family1 L of subsets of S and a collection of nonnegative integers
{kL }L∈L such that I is independent if and only if |I ∩ L| ≤ kL for all L ∈ L.
Partition matroids and their truncations are examples of laminar matroids.
6. Gammoids. Consider a graph G = (V, E) and two special sets of vertices

S, T ⊆ V . The gammoid M = (S, I) induced by G, S and T is the matroid
having S as ground set and where a set I ⊆ S is independent if there is a
collection of vertex-disjoint paths connecting each element of I to an element
of T .
7. Graphic matroids. Consider a graph G = (V, E). The graphic matroid

M(G) = (E, I) is the matroid having the edges of G as ground set and where
a set of edges is independent if it does not contain a cycle.
8. Linear matroids. Given a collection of vectors V ⊆ Fk in a finite-dimensional

vector space over a field F, the associated linear matroid M = (V, I) is the one
having V as ground set and where a set of vectors is independent if and only if
they are linearly independent in Fk .
1
A family is laminar if every time two sets A and B in the family have nonempty intersection,
they are disjoint or one is contained in the other.
38
1.5.3 Greedy Algorithm
Consider a matroid M = (S, I) and an ordering
σ(1), σ(2), . . . , σ(n)
of its elements, where σ : [n] → S is a bijective function. The greedy procedure on

the ordered sequence above is the following. Initialize I as the empty set and then for
all i from 1 to n, check if I ∪ {σ(i)} is independent. If so, add σ(i) to I. Otherwise,
continue.
The set I = {x1 , . . . , xr } returned by this procedure is not only independent, but
also a base of the matroid. This base is called the lexicographic first base of M
according to the ordering σ. The name comes from the following lemma.
Lemma 1.5.1. Let I = {x1 , . . . , xr } be the base of M obtained by the greedy procedure
on the sequence σ(1), . . . , σ(n), sorted in increasing order of σ −1 . Let J = {y1 , . . . , yr }
be a different base of the matroid, sorted again according to σ −1 and let j be the
maximum index such that
{x1 , . . . , xj } = {y1 , . . . , yj }.
Then
σ −1 (xj+1 ) < σ −1 (yj+1 ).
Furthermore, if v : S → R+ is a nonnegative weight function on the ground set such
that
v(σ(1)) ≥ v(σ(2)) ≥ · · · ≥ v(σ(n)), (1.13)
then the lexicographic first base I according to σ is the base having maximum total
weight.
Proof. Suppose by contradiction that σ −1 (xj+1 ) > σ −1 (yj+1 ). Then, the greedy pro-
cedure processes yj+1 before xj+1 and, since {x1 , . . . , xj , yj+1 } ⊆ J is independent,
this algorithm would have added yj+1 to I, contradicting the hypothesis.
To prove the second claim, suppose that J = {y1 , . . . , yr } is a base having strictly
bigger weight than the lexicographic first base I = {x1 , . . . , xr }, where the elements
are sorted according to their σ −1 value. Let k be the first index for which v(yk ) >
v(xk ). Consider the sets A = {x1 , . . . , xk−1 } ⊆ I and B = {y1 , . . . , yk−1 , yk } ⊆ J.
Since |A| < |B| and both sets are independent, there is an element yi ∈ B \ A so
that A ∪ {yi } is independent. But then, as σ −1 (yi ) ≤ σ −1 (yk ) < σ −1 (xk ), the greedy
algorithm would have included yi into I before processing xk , which is a contradiction.
This means that v(yk ) ≤ v(xk ) for all k, in particular, v(J) ≤ v(I).
The previous lemma states that in order to find a maximum weight independent
set of a nonnegatively weighted matroid, it is enough to sort the elements in decreasing
order of weight and apply the greedy procedure to the sequence obtained. In fact, as
Gale [67] and Edmonds [50] have shown, the previous property characterizes matroids
as the only hereditary families for which the previous algorithm outputs a maximum
weight set in the family.
39
Interestingly enough, the lexicographic first base according to σ can be obtained
also by a different greedy-like algorithm.
Consider an arbitrary ordering of the elements of the matroid (not necessarily
equal to σ): π(1), . . . , π(n). Perform the following procedure. Start with an empty
set I. For all i from 1 to n, check if I ∪ {π(i)} is independent. If so, add π(i) to I.
Otherwise, find the element x ∈ I with σ −1 (x) as large as possible such that the
set obtained from I by swapping x and π(i) is independent. If σ −1 (x) > σ −1 (π(i))
perform the swap and continue. Otherwise, ignore element π(i) and continue.
Lemma 1.5.2. The greedy-like algorithm described above returns the lexicographic
first base of M according to σ.
Proof. Let B be the lexicographic first base of M. Consider an element b of this base
and let I be the independent set considered by the algorithm above just before b is
considered. If b does not enter I at that moment, there must be a circuit C ⊆ (I ∪{b})
containing b. Let x be the element of C \ {b} having largest σ −1 (x). We must have
σ −1 (x) > σ −1 (b) as otherwise b would be in the span of elements having strictly
smaller value of σ −1 , contradicting the fact that b is in the lexicographic first base.
Therefore, x and b are swapped and b does enter the independent set.
Suppose now that at some moment b is swapped out by another element y arriving
later. This means that there is a circuit C 0 containing y and b where b is the element
having largest σ −1 (·)-value. Again, this is a contradiction as b would be spanned by
elements of smaller value of σ −1 .
During the rest of this work we use many matroid properties that we do not
describe in detail. Whenever we do so, we give a pointer to a reference where the
property can be found.
1.5.4 Matroid Secretary Problems

Suppose that we relax the first condition in the setting of the generalized secretary
problem, by allowing to revoke elements from the selected set. Consider the following
greedy-like algorithm for a given domain I: Whenever a new element can be added to
the current set, do it. If this is not the case, check if it is possible to improve the total
weight of the current selected set by discarding a subset of previous elements and
selecting the new one while keeping feasibility2 . If this is possible, perform the most
beneficial swap. As noted in the previous section, this greedy algorithm is guaranteed
to select the maximum weight feasible set when I is the family of independent sets of a
matroid. In fact, this property characterizes matroids: Gale’s [67] and Edmonds’s [50]
proofs can be modified to show that the previous greedy-like algorithm is guaranteed
to return a maximum weight feasible set for every assignment of weights if and only
if I defines a matroid.
2
Checking the weight improvement of swapping the new element for an arbitrary subsets of
previous elements requires the ability to see the values of the elements. However, if we only allow
to discard singletons we can do it with comparisons.
40
The intuition in the last paragraph motivated Babaioff et al. [8, 7] to propose the
following conjecture.
Conjecture 1.5.3 (Babaioff et al. [8, 7]). The matroid secretary problem admits a
constant competitive algorithm (where the constant is perhaps even e), for every one
of the random models proposed.
The conjecture above states that the obligation to honor past commitments is not
too costly for matroids, as it only reduces the expected selected value by a constant.
In Chapter 2 we partially answer the above conjecture by showing an affirmative
answer for two of the three random models proposed.
For the random-assignment random-order model, the author has presented a
constant-competitive algorithm in a recent SODA paper [152], which we improve
in this work. As first noticed by Oveis Gharan and Vondrák [136], it is possible to
apply the methods in [152] to also obtain constant-competitive algorithms for the
random-assignment adversarial-order model. In this thesis, we present alternative
algorithms achieving the same task.
The adversarial-assignment random-order model is still open, even for the case in
which we allow to use value-based algorithms. For this last particular case, Babaioff
et al. [8] have given an O(log r)-competitive algorithm, where r is the total rank of
the matroid.
Lemma 1.5.4 (Babaioff et al.[8]). There is an O(log r)-competitive algorithm for the
adversarial-assignment random-order model of the matroid secretary problem, under
the assumption that we can see the numerical values of the elements as they arrive.
In Section 2.6.1 we prove that it is indeed possible to achieve a competitive ratio
of O(log r) using a comparison-based algorithm.
There is a long line of work dedicated to devise constant competitive algorithms for
specific matroid classes on the adversarial-assignment random-order model. Before
describing these works we make the following important observations.
Lemma 1.5.5 (Folklore).
1. If M admits an α-competitive algorithm for the adversarial-assignment random-

order matroid secretary problem, then so do all the restrictions of M.
2. If M1 and M2 admit an α-competitive algorithm for the adversarial-assignment
random-order matroid secretary problem then so does the direct sum M1 ⊕ M2 .
Proof. For the first item, let M0 be a restriction of a matroid M and A be an α-

competitive algorithm for M. Consider the modified algorithm A0 that virtually
extends M0 to M by adding a set of zero weight dummy elements. Then, this
algorithm simulates the augmented input in such a way that the dummy elements
arrive uniformly at random similarly to the real ones. The resulting algorithm will
also have a competitive ratio of α.
For the second point, run the corresponding α-competitive algorithms in paral-
lel both in M1 and M2 . In other words, when an element is processed, use the
41
corresponding algorithm and accept or reject it accordingly. This algorithm is α-
competitive in the direct sum matroid.
Using Karger’s matroid sampling theorem [93], Babaioff et al. [8] have shown the
following result: If an α-competitive algorithm is known for a given matroid M then
they can devise a max(13α, 400)-competitive algorithm for any truncation of M.
The previous paragraphs imply that the class of matroids admitting a constant-
competitive algorithm for the adversarial-assignment random-order matroid secretary
problem is closed under truncation, restriction and direct sum. It is still open if this
class is also closed for duality, contraction and union.
In what follows we mention some of the previous results for the adversarial-
assignment random-order matroid secretary problem.
Uniform and Partition Matroids. The uniform case corresponds exactly to the
classical and multiple choice secretary problems described in Sections 1.2 and 1.3,
for which e-constant competitive algorithms exist (e.g. Algorithms 1 and 2). Since
partition matroids are direct sum of uniform matroids, they also admit e-competitive
algorithms.
Transversal Matroids. Recall that transversal matroids can be defined using a

bipartite graph, where the ground sets are the vertices on the left part, and the
independent sets are those subsets of the left part that can be covered by a matching.
Babaioff et al. [8] have given a 4d-competitive algorithm for these matroids, where d is
the maximum left-degree of the bipartite graph defining the matroid. In a later article,
Dimitrov and Plaxton [44] give a 16-competitive algorithm for arbitrary transversal
matroids. Korula and Pál [97] improve their analysis to show that their algorithm
is in fact 8-competitive. Later, Thain [162], using a result by Goel and Mehta [72],
shows that Dimitrov and Plaxton’s algorithm is 4/(1 − 1/e) ≈ 6.33-competitive.
Graphic Matroids. Babaioff et al. [8] have shown a 16-competitive algorithm for
graphic matroids, by modifying slightly their 4d-competitive algorithm for transversal
matroids. Later, Korula and Pál [97] give an improved 2e-competitive algorithm for
this class. Around the same time, Babaioff et al. [5] give an alternative 3e-competitive
algorithm.
Laminar Matroids. Very recently, Im and Wang [87] have shown that laminar
matroids also admit constant competitive algorithms.
1.6 Related Work

As discussed in the introductory section of this chapter, the secretary problem admits
many variants and extensions. In this section we present a very limited discussion
about other studied variants.
42
An old line of work involve variants of the classical secretary problem where the
algorithm is allowed to see values, and there is some information available a priori on
the distribution of the incoming weights.
Gilbert and Mosteller [70] study the full-information game where we want to
maximize the probability of selecting the top value of a sequence of weights drawn
from a known (continuous) distribution. By applying a monotonic transformation,
they argue that it is enough to focus on the uniform distribution on the interval (0, 1)
and give a strategy that selects the top element with probability at least 0.58016 . . . ,
which is much larger than what we can get using comparison-based algorithms.
An intermediate problem between the full-information game and the classical sec-
retary problem using only comparisons is called the partial information game in
which weights are drawn from a distribution of known form but containing one or
more unknown parameters; for example, a uniform distribution over an interval of
unknown extremes. Many authors have studied this problem (see, e.g. Ferguson’s
survey [55]). Interestingly enough there are families of distributions for which the
best strategy, in the limit when the number of elements tends to infinity, achieves a
probability of at most 1/e [157]; for these distributions, value-based algorithms have
no advantage over comparison-based ones.
Other variants of the classical secretary problem involve considering objective
functions different from the total weight recovered or the probability of selecting the
top weight. For example, Lindley [101] himself give a recurrence to solve the problem
of minimizing the expected rank of the selected element, i.e. its ordinal position on the
sorted list. Chow et al. [31] solve this problem completely, by showing that as n goes to
1/(j+1)
infinity, the optimal strategy recovers an element of expected rank ∞ j+2
≈
Q
j=1 j
3.8695 · · · . A related problem, still open today, is Robbins’ secretary problem in which
we want to minimize the expected rank using a value-based algorithm and knowing
that the weights are selected uniformly and independently in (0, 1). It is know that
the optimal achievable rank, as n goes to infinity tends to some value between 1.908
and 2.329, but the exact value is not yet known (see [20]).
Gusein-Zade [78] consider the problem of accepting any of the best r candidates.
This corresponds to find the optimal stopping polity for an utility function that is
equal to 1 if the selected element is one of the top r candidates and 0 otherwise.
Mucci [119] was the first person to study general utility functions depending only on
the ordinal position of each element in the sorted list.
Another possibility is to consider an objective function where the gain of selecting
an element equals the product of its weight and a discount factor depending on the
time we select it. Discounted versions of the secretary problem have been studied by
Rasmussen and Pliska [143] and Babaioff et al. [5].
The r-choice secretary problem also admits different variants by changing the ob-
jective function. Ajtai et al. [1] consider the problem of minimizing the expected sum
of the ranks of the selected objects. Babaioff et al. [5] consider the following weighted
secretary problem in which up to r elements can be selected: at the moment an ele-
ment is selected, it must be assigned to one of r possible positions, each one having
a known weight. The benefit of assigning an element to a position is the product
43
of the element weight and the position weight, and the objective is to maximize the
total benefit. Buchbinder et al. [21] study the J-choice K-best secretary problem,
in which one can select at most J elements receiving profit for every element select
from the top K of the stream. A different approach was taken by Bateni et al. [10],
who consider the case where the gain of selecting a set is a nonnegative submodular
objective function, which can be accessed only as the element arrive.
There has also been some work for the adversarial-assignment random-order gen-
eralized secretary problem on non-matroidal domains. Babaioff et al. [6] show a
10e-competitive algorithms for knapsack domains even in the case where both the
weights and lengths are revealed online. Korula and Pál [97] give constant compet-
itive algorithms for some cases of intersection of partition matroids, specifically for
matchings in graphs and hypergraphs where the edges have constant size. Im and
Wang [87] have also studied an interval scheduling secretary problem that generalizes
knapsack domains.
44
Chapter 2
New Results for Matroid Secretary

Problems
2.1 Preliminaries
In this chapter we give new algorithms for matroid secretary problems. Here we
briefly review the notation used.
Consider a matroid M = (S, I) with ground set S of size n. An adversary selects
a set W of n nonnegative weights w1 ≥ · · · ≥ wn ≥ 0, which are assigned to the
elements of the matroid using an ordering of S,
σ(1), σ(2), σ(3), . . . , σ(n);
defined by a bijective map σ : [n] → S. In other words, the weight assignment is given
by
w(σ(i)) = wi .
The elements are then presented to an online algorithm in the order
π(1), π(2), π(3), . . . , π(n);
defined by a certain bijective function π : [n] → S.

Depending on how the assignment σ and the ordering π are selected we recover
the four models discussed in the previous chapter. Each one can be selected in an
adversarial way or uniformly at random.
When an element is presented, the algorithm must decide whether to add it or
not to the current solution set, denoted as ALG, under the condition that this set is
independent (ALG ∈ I) at all times. The objective is to output a solution set whose
P
payoff w(ALG) = e∈ALG w(e) is as high as possible.
We focus on comparison-based algorithms: We assume that when the i-th element
of the stream, π(i), is presented, the algorithm only learns the relative order of the
weight with respect to the previously seen ones. This is, it can compare w(π(j)) with
w(π(k)) for all j, k ≤ i, but it can not use the actual numerical values of the weights.
Without loss of generality, we assume that there are no ties in W , because otherwise
45
we break them using a new random permutation τ (independent of W, σ and π); if
the comparison between two elements seen gives a tie, then we consider heavier the
one having larger τ -value.
As usual, we say that an algorithm A returning an independent set ALG is α-
competitive, if for any adversarial selection the model permits E[αw(ALG) − w∗ ] ≥ 0,
where the expectation is taken over all the random choices given by the model and
the random choices performed by the algorithm, and w∗ is the maximum possible
payoff of an independent set in I.
It is important to remark that for the matroid secretary problem, w∗ depends on
both W and σ, however the set achieving this maximum payoff depends only on σ.
Indeed, let OPTM (σ) be the the lexicographic first base of M under ordering σ.
In other words, OPTM (σ) is the set obtained by applying the greedy procedure that
selects an element if it can be added to the previously selected ones while preserving
independence in M, on the sequence σ(1), σ(2), . . . , σ(n).
Lemma 1.5.1 states that OPTM (σ) is a maximum weight independent set with
respect to any weight function v for which v(σ(1)) ≥ · · · ≥ v(σ(n)) ≥ 0. In particular,
this is true for the weight function w defined before and E[w∗ ] = Eσ [w(OPTM (σ))].
We drop the subindex M in OPTM (σ) whenever there is no possible confusion.
2.2 Divide and Conquer

A natural divide and conquer approach for any model of the matroid secretary prob-
lems is the following.
Given a matroid M = (S, I), find matroids {Mi = (Ei , Ii )}i , with S = ∪i Ei
satisfying three properties:
Mi is independent in M.
L
(i) Any independent set of i
(ii) For each matroid Mi , we have access to an α-competitive algorithm Ai for the
same model.
Mi is at least β times the maximum
L
(iii) The maximum (expected) weight in i
(expected) weight in M.
If we can find such family of matroids, then we claim that the following algorithm is
(at most) α/β-competitive: Run in parallel the α-competitive algorithm in every Mi
and return the union of the answers. Call this algorithm A.
To prove this claim, note first that the sub-algorithm Ai controlling the elements
of Mi effectively receives an instance of the same model as the one we are trying to
solve in the full matroid: If the ordering of the entire ground set was adversarial or
uniformly at random selected, then so is the ordering of the elements presented to
Ai , and the same hold for the weight assignment.
Also, by conditions (i) and (ii), the set that algorithm A returns is independent
in M. By the second condition, the expected weight of the returned set is at least
1/α times the expected maximum weight of i Mi . The claim follows immediately
L
now from the third condition.
46
We use the divide and conquer approach above to devise constant-competitive
algorithms for the random-assignment models of the matroid secretary problem.
In Section 2.3 we show a natural class of matroids admitting simple constant-
competitive algorithms for random-assignment models: the class of uniformly dense
matroids. In Section 2.4 we show how to use the powerful notion of principal partition
to obtain a collection of uniformly dense matroids satisfying condition (i) above. We
also show that a condition slightly weaker than (iii) is satisfied for the considered
matroids. In Section 2.5 we study the obtained algorithms in detail.
2.3 Uniformly Dense Matroids

In order to give constant competitive algorithms for random-assignment models we
start by defining a class of matroids for which this task is relatively simple to attain.
Define the density γ(M) of a loopless matroid 1 M = (S, I) with rank function
rk : S → Z+ to be the maximum over all nonempty sets X of the quantity |X|/rk(X).
The matroid M is uniformly dense if γ(M) is attained by the entire ground
set; that is, if
|X| |S|
≤ , (2.1)
rk(X) rk(S)
for every nonempty X ⊆ S.
Examples of uniformly dense matroids include uniform matroids, the graphic ma-
troid of a complete graph, and all the linear matroids arising from projective geome-
tries P G(r − 1, F).
The following properties are important for our analysis.
Lemma 2.3.1. Let (x1 , . . . , xj ) be a sequence of different elements of a uniformly

dense matroid chosen uniformly at random. The probability that element xj is selected
by the greedy procedure on that sequence is at least 1 − (j − 1)/r, where r = rk(S).
Proof. An element is selected by the greedy procedure only if it is outside the span
of the previous elements. Denote by Ai = {x1 , . . . , xi } the set of the first i elements
of the sequence, and let n be the number of elements of the matroid, then:
n − |span(Aj−1 )| n − rk(Aj−1 )n/r

Pr[xj is selected] = ≥
n − |Aj−1 | n − (j − 1)
n j−1 j−1

≥ 1− ≥1− ,
n − (j − 1) r r
where the first inequality holds since the matroid is uniformly dense and the second
holds because the rank of a set is always at most its cardinality.
A simple application of the previous lemma is the following. Consider a uniform
random set X of j elements. The rank of X = {x1 , . . . , xj } is equal to the cardinality
1
A loop is an element x such that {x} is not independent. A loopless matroid is a matroid
having no loops.
47
of the set returned by the greedy procedure on any ordering of its elements. By
Lemma 2.3.1,
j min{j,r}
E[rk(X)] ≥ i−1 i−1
X X
max 1− ,0 = 1− . (2.2)
i=1 r i=1 r
Note that if j ≤ r, the right hand side is j − j(j−1)

2r
≥ j/2, and if j ≥ r, the right hand
r(r−1)
side is r − 2r ≥ r/2. In any case the expected rank of X is at least min{j, r}/2.
This shows that the rank of a random set in a uniformly dense matroid is close to
what it would have been if the matroid was uniform. The following lemma tightens
this bound.
Lemma 2.3.2. Let X be a set of a fixed cardinality j whose elements are chosen
uniformly at random in a uniformly dense matroid. Then,
j !
E[rk(X)] ≥ r 1 − 1 − 1r

≥ r 1 − e−j/r . (2.3)
In particular,
E[rk(X)] ≥ min{j, r} 1 − 1e .

(2.4)
Proof. Let (x1 , x2 , . . . , xn ) be a random ordering of the elements of S, and let Xj be

the set {x1 , . . . , xj }. As j increases, the rank of Xj increases by one unit every time
an element is outside the span of the previous elements. Then, for all 1 ≤ j ≤ n,
" #
E[rk(Xj )] − E[rk(Xj−1)] = Pr(xj 6∈ span(Xj−1)) = E n −n|span(X j−1 )|
− |Xj−1 |
E[rk(Xj−1)] .
" #
≥E
n − rk(Xj−1 )n/r
≥1−
n − (j − 1) r
Therefore, the sequence Zj = E[rk(Xj )], for j = 0, . . . , n satisfies:
Z0 = 0,
1

Zj ≥ 1 + Zj−1 1− , for j ≥ 1.
r
Solving the previous recurrence yields
j−1 i j !
X 1 1

Zj ≥ 1− =r 1− 1− ≥ r 1 − e−j/r . (2.5)
i=0 r r
This completes the first part of the proof.

The function (1 − e−x ) is increasing, thus if j ≥ r,

Zj ≥ r 1 − e−j/r ≥ r 1 − e−1 .
48
In addition, the function (1 − e−x )/x is decreasing, thus if j ≤ r,

1 − e−j/r
Zj ≥ j ≥ j 1 − e−1 .
j/r
2.3.1 Random-Assignment Random-Order Model

We have seen that uniformly dense matroids behave similarly to uniform matroids
on the same ground set and with the same total rank. A natural approach to obtain
a constant-competitive algorithm for uniformly dense matroids is to adapt constant-
competitive algorithms for uniform matroids.
Recall the e-competitive algorithm (Algorithm 2) for uniform matroids of rank r
we presented in Section 1.3. We show that a slight modification of this algorithm is at
most 2e-competitive for uniformly dense matroids in the random-assignment random-
order model. The only difference with respect to Algorithm 2 is that before adding
an element to the output, we have to test if its addition maintains independence in
the matroid. We depict the entire procedure as Algorithm 3 below.
Algorithm 3 for uniformly dense matroids of n elements and rank r in the random-
assignment random-order model.
1: Maintain a set T containing the heaviest r elements seen so far at every moment
(initialize T with r dummy elements).
2: ALG ← ∅.
4: Observe the first m elements and denote this set as the sample.
6: if x enters T and the element leaving T is in the sample or is a dummy element
then check if ALG ∪ {x} is independent. If so, add x to ALG.
7: end if
8: end for
9: Return the set ALG.
Theorem 2.3.3. Let ALG be the set returned by Algorithm 3 when applied to a
uniformly dense matroid M of rank r. Then
r
Eσ,π [w(ALG(σ))] ≥ (−p2 ln p) wi ≥ (−p2 ln p)Eσ [w(OPTM (σ))].
X
i=1
In particular, by setting p = e−1/2 , we obtain a 2e-competitive algorithm for uni-

formly dense matroids in the random-assignment random-order model of the matroid
secretary problem.
Proof. Similar to the analysis of Algorithm 2, we use an offline simulation, but now
we make a clear distinction between the elements of the matroid and their weights.
49
In the first step of the simulation, each weight wi in the adversarial list W selects
an arrival time ti in the interval (0, 1) uniformly and independently. The simulation
keeps a set T containing the top r weights seen at every moment (initially containing r
dummy weights of negative value) and processes the weights as they arrive, sampling
the ones arriving before time p. When a weight arriving after time p is processed, the
algorithm marks it as a candidate if, first, that weight enters T , and second, the one
leaving T is either in the sample or a dummy weight.
In the second step of the simulation, the algorithm assigns to each weight marked
as a candidate a different element of the matroid’s ground set uniformly at random.
Then, it runs the greedy procedure on the sequence of candidates in the order they
arrived and returns its answer.
Using that the cardinality of the sampled set has binomial distribution with pa-
rameters n and p, it is not hard to check that the sets of elements and weights returned
by this simulation have the same distribution as the ones returned by Algorithm 3.
For this reason, we focus on the simulation.
We estimate the probability that each one of the top r weights appears in the
output set. Focus on one such weight wi with i ≤ r. Condition on the event that wi
arrives after time p, and let ` be the random variable counting the number of weights
in {w1 , . . . , wr } \ {wi } arriving in the time interval (0, p). Each of these high weights
enters T during the sample period and never leaves T . Since every weight marked as
a candidate pushes out either a dummy or a sampled weight of T at the moment it
is marked, the previous statement implies that the number of weights marked as a
candidate by the simulation algorithm is at most r − `.
Since wi is one of the top r weights, it enters the set T at the time it is considered.
Thus, it will be marked as a candidate if and only if the weight leaving T at that
time is either a dummy or a sampled weight.
Let A(ti ) = {wj : tj < ti } be the set of weights seen before wi arrives. If this set
has less than r elements then the element leaving T at ti will be a dummy weight.
Consider the case where A(ti ) has cardinality at least r and let wA be the r-th top
element of this set. Since wA is not one of the top r elements in the full adversarial
list, its arrival time tA is independent of `. In particular, conditioned on {ti , `}, tA is
a uniform random variable in (0, ti ). Hence,
Pr(wi is as a candidate | `, ti )
= 1 · Pr(|A(ti )| < r | `, ti ) + Pr(|A(ti )| ≥ r, tA < p | `, ti )
= Pr(|A(ti )| < r | `, ti ) + Pr(|A(ti )| ≥ r | `, ti )p/ti
≥ p/ti .
The elements of the matroid assigned to the weights marked as candidates form
a random set. Conditioned on the value ` and on wi being a candidate, Lemma 2.3.1
implies that no matter what position wi takes in the list of at most r − ` candidates,
the probability that the corresponding element gets added to the output is at least
1 − (r − ` − 1)/r = (` + 1)/r; therefore, the probability that wi appears in the output
50
is at least
E[`] + 1 = (−p ln p)((r − 1)p + 1) ≥ −p2 ln p.
" #
E `+1Z 1 p
r p ti
dti = −p ln p
r r
Theorem 2.3.3 follows easily from here.
We can obtain an alternative constant-competitive algorithm for uniformly dense

matroids based on Algorithm 1 for the classical secretary problem. The idea is to
perform a random partition of the matroid’s ground set into r consecutive groups.
For each of them, apply Algorithm 1 to find its top-valued element and try to add it
to the output set if this is possible. The procedure is depicted as Algorithm 4 below.
assignment random-order model.
1: ALG ← ∅.
2: Select n values v1 , . . . , vn uniformly at random from {1, . . . , r} and let Ni be the
number of times value i was selected.
3: Denote as S1 the set of the first N1 elements arriving, S2 the set of the next N2
elements, and so on.
4: For each i in [r], run Algorithm 1 with parameter p (not necessarily p = 1/e) on
the sequence Si . Mark the elements that are selected.
5: Whenever an element x is marked, check if ALG ∪ {x} is independent. If so, add
x to ALG.
Before stating the analysis of Algorithm 4, it is convenient to define the following

auxiliary construction.
For any matroid M = (S, I) of total rank r, the associated random partition
matroid P(M) = (S, I 0 ) is obtained as follows. Partition the set S into r classes,
where each element selects its own class in a uniform random way. A set of elements
in S is independent in P(M) if it contains at most one element of each class. The
following lemma states that the weight of the optimum base of P(M) is at least a
constant fraction of the weight of the optimum in M.
Lemma 2.3.4. Let w1 ≥ w2 ≥ · · · ≥ wn ≥ 0 be the set of weights determined by the

adversary, then:
r
r X
Eσ [w(OPTP(M)(σ))] ≥ 1
wi ≥ (1 − 1/e)Eσ [w(OPTM (σ))]. (2.6)

1− 1−
r i=1
To prove this lemma, we need Chebyshev’s sum inequality (see, e.g. [112]) which
states that if
a1 ≥ a2 ≥ · · · ≥ ar and b1 ≥ b2 ≥ · · · ≥ br
51
then
r r r
! !
X 1 X X
ai b i ≥ ai bi .
i=1 r i=1 i=1
Proof of Lemma 2.3.4. For i ∈ [r], wi is in OPTP(M) (σ) if and only if the weights
w1 , . . . , wi−1 are assigned to elements in a different class of P(M) than the one the
element of wi is assigned. Then Pr(wi ∈ OPTP(M) (σ)) = (1 − 1/r)i−1 . Note that both
((1 − 1/r)i−1 )i=1,...,r and (wi )i=1,...,r are non-increasing sequences. Using Chebyshev’s
sum inequality we have
r i−1 r r
i−1 ! X !
E[w(OPTP(M)(σ))] ≥ 1 1X 1
X
1− wi ≥ 1− wi
i=1 r r i=1 r i=1
r
r ! X
1

= 1− 1− wi .
r i=1
We give three different analysis for Algorithm 4.
uniformly dense matroid M of total rank r.
(i) By setting p = 1/e, we have:
Eσ,π [w(ALG(σ))] ≥ (1 − 1/e)(1/e)EP(M),σ [w(OPTP(M)(σ))].

By Lemma 2.3.4, this algorithm is (e/(1 − 1/e)2 ) ≈ 6.80291-competitive.
(ii) For p ∈ (0, 1), we have:
Eσ,π [w(ALG(σ))] ≥ γ1(p) · EP(M),σ [w(OPTP(M)(σ))],

1
where γ1 (p) = (1 − e−(1−p) )(−p ln p).
(1 − p)
This is optimized by setting p = p1 ≈ 0.433509. In that case, we get γ1 (p1 ) ≈

0.276632. By Lemma 2.3.4, this algorithm is ((1 − 1/e)γ1 (p1 ))−1 ≈ 5.7187-
competitive.
(iii) For p ∈ (0, 1), we have:

r
Eσ,π [w(ALG(σ))] ≥ γ2(p) ·
X
wi ,
i=1
−(1−p)
p(1 − e )Z 1 1 − e−t
where γ2 (p) = dt.
1−p p t2
This is optimized by setting p = p2 ≈ 0.384374. In that case, we get γ2 (p2 ) ≈

0.20322. Thus, the competitive ratio of this algorithm is at most 1/(γ2 (p2 )) ≈
4.92078.
52
Proof. As usual, we analyze the algorithm via a simulation. On the first step of the
simulation every weight wi of the adversarial list selects a random color c(wi ) in [r].
Let Wj be the collection of weights that are assigned color j. For every j, we present
the weights of Wj in uniform random order to Algorithm 1 with parameter p. This
algorithm returns either a singleton weight or an empty set. Let T be the set of
weights selected by this algorithm.
On the second step of the simulation, we randomly assign to each weight in T
an element in S to obtain a set X(T ). Finally, we apply the greedy procedure on
the elements X(T ) with weights in T , in uniform random order. Let ALG be the set
obtained and W (ALG) the corresponding set of weights.
It is straightforward to check that the distribution of ALG and W (ALG) is the
same both if we apply the greedy procedure on X(T ) in random order or if we apply it
on increasing ordering of their colors (as the offline simulation does). We use random
order because it makes our analysis simpler. From here, it is easy to see that the sets
of elements and weights returned by the simulation have the same distribution as the
ones returned by Algorithm 4.
Because of the way it was constructed, W (ALG) is a uniform random subset of T
of size rk(X(T )). Then, for any weight w of the list of the adversary,
Pr[w ∈ W (ALG) | w ∈ T, t = |T |] =
E[rk(X(T ))] ≥ (1 − 1/e), (2.7)
t
where the last inequality comes from (2.4) in Lemma 2.3.2 and the fact that X(T ) is
a random set of size t = |T |. Thus, we can remove the conditioning on t.
Recall that Algorithm 1 for the classical secretary problem is e-competitive when
we set p = 1/e. Hence, for each color class j ∈ [r], the expected value in Wj ∩ T is at
least 1/e-fraction of the maximum weight wj∗ in Wj (set wj∗ = 0 for the pathological
case where Wj is empty). Therefore:
r r
E[ΣW (ALG)] ≥ (1 − 1/e)E[Σ(T ∩ Wj )] ≥ (1 − 1/e)(1/e) E[wj∗],
X X
(2.8)
j=1 j=1
where ΣV denotes the total sum of the weights inside V . We conclude the proof
of part (i) by noting that rj=1 E[wj∗ ] is the expected weight of the optimum in the
P
random partition matroid P(M) induced by the coloring.

For part (ii), we refine the previous analysis. By using (2.3) of Lemma 2.3.2, we
get an improved version of (2.7):
Pr[w ∈ W (ALG) | w ∈ T, t = |T |] =
E[rk(X(T ))] ≥ r
1− 1−
1
t !
. (2.9)
t t r
In order to remove the conditioning on t, we need to compute the expected value of

the right hand side. First note that t equals 1 (because of w) plus the number of colors
different to the color of w, for which the classical secretary algorithm did not return an
empty set. For each color, the algorithm returns a nonempty set with probability at
53
most2 (1 − p) (see Lemma 1.2.4). Therefore, the variable t is stochastically dominated
by one plus the sum of r − 1 Bernoulli random variables with parameter (1 − p). As
the right hand side of (2.9) is decreasing in t, we can effectively replace t to obtain:
r−1 1+k ! !
r 1 r−1

(1 − p)k pr−1−k
X
Pr(w ∈ W (ALG) | w ∈ T ) ≥ 1− 1−
k=0 1 + k r k
 
r r j
! !
1 X r r 1
(1 − p)j pr−j − (1 − p)j pr−j 
X
= 1−
1 − p j=1 j j=1 j r
1
= 1 − pr − ((1 − 1/r)(1 − p) + p)r − pr
1−p
1 − (1 − (1 − p)/r)r 1 − e−(1−p)
= ≥ . (2.10)
1−p 1−p
Finally, using that Algorithm 1 returns the best weight of each class with probability
−p ln p (see Lemma 1.2.4), we obtain that
r r
1 − e−(1−p)
E[ΣW (ALG)] ≥ E E[wj∗],
X X
[Σ(T ∩ Wj )] ≥ γ1 (p)
j=1 1 − p j=1
−(1−p)
where γ1 (p) = 1−e1−p (−p ln p). This concludes the proof of part (ii).
The analysis for part (iii) is, again, a refinement of the previous one. For this one,
we compute directly the probability that the simulation selects each one of the top r
weights in the adversarial list. Let wi , for i ≤ r, be the i-th top weight. Let Eij be
the event that wi is the j-th top weight of its own color class. We have
 j−1 i−j
 i−1 1
1− 1 , if j ≤ i;
Pr(Eij ) =  j−1 r r (2.11)
0, otherwise.
Using Lemma 1.2.4, we conclude that the probability that wi is marked is

i
X
Pr(wi ∈ T ) = Pr(wi ∈ T | Eij ) Pr(Eij )
j=1
i Z 1 j−1 i−j
(1 − t)j−1 p
!
X i−1 1 1
= dt 1−
j=1 p t j−1 r r
i j−1 i−j
!
Z 1
pX i−1 1−t 1
= 1− dt
p t j=1 j − 1 r r
Z 1
((1 − t)/r + (1 − 1/r))i−1 Z 1
(1 − t/r)i−1
=p dt = p dt.
p t p t
The sequence (Pr(wi ∈ T ))i=1,...,r is non-increasing. Using (2.10) and Chebyshev’s
2
The reason why this is not exactly (1 − p) is the pathological case where the color class is empty.
In this case the algorithm will always return an empty set.
54
sum inequality we conclude that
−(1−p) r
E[ΣW (ALG)] ≥ 1 −1e− p
X
wi Pr(wi ∈ T )
i=1
−(1−p) X r r
(1 − t/r)i−1
! Z
1−e 1 1X
≥ wi p dt
1−p i=1 p r i=1 t
−(1−p) r
1 − (1 − t/r)r
! Z
1−e X 1
= wi p dt
1−p i=1 p t2
−(1−p) r
1 − e−t
! Z
1−e X 1
≥ wi p dt.
1−p i=1 p t2
This concludes the proof.

The third part of the above analysis for Algorithm 4 shows that this algorithm has
a better competitive ratio (≈ 4.92078) than Algorithm 3 (2e ≈ 5.43656). The second
part, as we will see in Section 2.5, is also useful for the case of general matroids.
2.3.2 Random-Assignment Adversarial-Order Model

After the first publication of some of the results of this thesis in [152], Oveis Gha-
ran and Vondrák [136] have devised a 40-competitive algorithm for uniformly dense
matroids on the random-assignment adversarial-order model.
In what follows we give an alternative algorithm for this task that is very similar
to Algorithm 4. In this new algorithm we also partition the ground set of the matroid
into r groups, in a way that is independent of the order. Afterwards, we use the
threshold algorithm for the classical secretary in each group to find its top-valued
element, but in a coupled way: the sample consists of all the elements of that group
seen in the first half of the stream. After that, unlike Algorithm 4, we only try to
add the elements found to the output set with certain fixed probability (otherwise,
we discard it). The procedure is depicted as Algorithm 5 below.
As in the analysis of Algorithm 4, it is convenient to compare to the random
partition matroid P(M) = (S, I 0 ), associated to matroid M.
uniformly dense matroid M of total rank r, in the random-assignment adversarial-
order model. Then
Eσ [w(ALG(σ))] ≥ 1/16 · EP(M),σ [w(OPTP(M)(σ))].

By Lemma 2.3.4, Algorithm 5 is 16/(1 − 1/e) ≈ 25.3116-competitive.
Proof. As usual, we use a two-step offline simulation algorithm. On the first step, the
simulation assigns to each weight wi in the adversarial list W a color c(wi ) in [r] and
an arrival time ti ∈ (0, 1). After that, the simulation samples all the weights arriving
before time p = 1/2 and, for each color, marks the first weight (if any) arriving after
p = 1/2 that is larger than all the weights of that color seen so far.
55
assignment adversarial-order model.
1: ALG ← ∅.
2: Assign to each element of the matroid a color i ∈ [r] uniformly at random.
3: Choose m from the binomial distribution Bin(n, 1/2).
4: Observe and reject the first m elements and denote this set as the sample.
5: for each element x arriving after that do
6: if the color i of x has not been marked and x is the heaviest element seen so
far with color i then
7: Mark color i.
8: With probability 1/2 ignore x and continue.
9: Otherwise, check if ALG ∪ {x} is independent. If so, add x to ALG.
10: end if
11: end for
On the second step, the simulation assigns to the weight arriving in the k-th posi-
tion the corresponding k-th element of the adversarially sorted list. Then, it unmarks
each marked weight with probability q = 1/2, and applies the greedy procedure on
the remaining marked elements in their order of arrival, returning its answer.
It is easy to see that the sets of elements and weights this simulation returns have
the same distribution as the ones returned by Algorithm 5.
We say that color j ∈ [r] is successful if either the top weight having color j is
part of the set returned by the algorithm or if no weight receives color j.
Claim 1. Every color j is successful with probability at least p(1 − p)q(1 − q).
If this claim holds then the total weight returned by the algorithm is at least
p(1 − p)q(1 − q) = 1/16 times the weight of the optimum of the partition matroid
induced by the coloring, where the independent sets are those containing at most one
element of each color. As this matroid has the same distribution as P(M), the claim
implies the lemma.
Let us prove the claim. Fix a nonempty color class j and let w0 and w00 be the
two top weights of color class j. In the case this class has only one element, let w00
be an arbitrary different weight of the list. Color j is successful if the weight w0 is
selected in the output of the algorithm. In what follows, we estimate the probability
that this event occurs.
Let A be the set of weights arriving before time p and B be the set of weights
arriving after time p. Note that if w0 ∈ B and w00 ∈ A then no matter what relative
position they take inside A and B, weight w0 will be marked in the first step of the
simulation. In order for w0 to be part of the output of the greedy procedure, it is
enough that this weight is not unmarked in the second step, and that the element x0
in B that receives weight w0 is outside the span of the set X of other marked elements.
In order to estimate |X|, we need another definition. We say that a color is feasible
if one of the two following events occur. Either exactly one weight of that color is
56
marked in the first step of the simulation and it is not unmarked later, or no weight at
all receives that color and two independent biased coins fall head, one with probability
1 − p and another one with probability 1 − q.
Let s be the random variable counting the number of feasible colors different
than j. Note that just before applying the greedy algorithm on the second part of the
simulation, the number |X| of remaining marked elements of color different than j,
is always at most s. Then, we have
Pr(w0 is selected by the algorithm | w0 ∈ B, w00 ∈ A, s)

= (1 − q) Pr(x0 6∈ span(X) | w0 ∈ B, w00 ∈ A, s)
 
= (1 − q) E 
|B \ span(X)| 0

00
w ∈ B, w ∈ A, s . (2.12)

|B|
As the matroid is uniformly dense,
|B \ span(X)| |span(X)| n |X| ns 1

≥1− ≥1− ≥1− . (2.13)
|B| |B| r |B| r |B|
We conclude that (2.12) is at least

" #!
ns
E 1 0

(1 − q) 1 − w ∈ B, w00 ∈ A . (2.14)
r |B|
Conditioned on the fact that weight w0 is already assigned to A and weight w00 is
already assigned to B, the variable |B| behaves like a binomial random variable with
parameters n − 2 and 1 − p shifted up by one unit. Therefore,
n−2
" # !
E 1 0 n−2
X 1

00
w ∈ B, w ∈ A = (1 − p)j pn−2−j
|B| j=0 1 + j j
n−2
!
1 X n−1
= (1 − p)j+1 pn−1−(j+1)
(1 − p)(n − 1) j=0 j + 1
1 − pn−1
= . (2.15)
(1 − p)(n − 1)
By removing the conditioning on w0 ∈ B and w00 ∈ A in (2.12), we obtain
ns 1 − pn−1
!
Pr(Color j is successful | s) ≥ p(1 − p)(1 − q) 1 − . (2.16)
r (1 − p)(n − 1)
By Lemma 1.2.4 the probability that a given color is feasible is at most (1 − p)(1 − q),
and since all these events are independent, the variable s is stochastically dominated
by a binomial random variable with parameters r − 1 and (1 − p)(1 − q). As the right
57
hand side of (2.16) is decreasing in s, we can effectively replace s to obtain:
Pr(Color j is successful)
nE[Bin(r − 1, (1 − p)(1 − q))] 1 − pn−1
!
≥ p(1 − p)(1 − q) 1 −
r (1 − p)(n − 1)
n−1
!
n(r − 1)(1 − q)(1 − p )
≥ p(1 − p)(1 − q) 1 −
r(n − 1)
≥ p(1 − p)(1 − q)(1 − (1 − q)(1 − pn−1 )) ≥ p(1 − p)(1 − q)q,
where in the third inequality we have used that r ≤ n. This concludes the proof of
the claim and the lemma.
2.4 Principal Partition

Recall the divide and conquer strategy of Section 2.2. In this section we describe a
technique which lets us find a collection of uniformly dense matroid whose direct sum
approximates any given matroid.
Consider a loopless matroid M = (S, I) that is not uniformly dense, and let E1
be a maximum cardinality set achieving the density of M. This is
|X| |E1 |
γ(M) = max = , (2.17)
∅6=X⊆S rk(X) rk(E1 )
and |E1 | ≥ |X| for any set X achieving the same density.
We claim that the matroid M1 = M|E1 restricted to the set E1 is uniformly dense
with density λ1 = γ(M). Indeed, since the rank function of M1 is equal to the one
of M, every subset of E1 has the same density in both M and M1 , making E1 the
densest set in M1 , with density |E1 |/rk(E1 ) = γ(M) = λ1 .
On the other hand, consider the matroid M01 = M/E1 obtained by contracting E1 .
We show that this matroid is loopless and has strictly smaller density than M1 .
Indeed, recall that the rank function of the contracted matroid (see, e.g. [137]) is
rkM01 (X) = rkM (E1 ∪ X) − rkM (E1 ), for all X ⊆ S \ E1 .
Hence, if x ∈ S \ E1 is a loop of M01 , then rkM (E1 ∪ {x}) = rkM (E1 ), and so
|E1 ∪ {x}| |E1 | + 1 |E1 |

= > ,
rkM (E1 ∪ {x}) rkM (E1 ) rkM (E1 )
contradicting the definition of E1 . Therefore, M01 is loopless. By maximality of E1 ,

every set X with ∅ =
6 X ⊆ S \ E1 satisfies
|E1 ∪ X| |E1 |
< .
rkM (E1 ∪ X) rkM (E1 )
58
Hence,
|E1 |
|X| |E1 ∪ X| − |E1 | rkM (E1 )
(rkM (E1 ∪ X) − rkM (E1 ))
= <
rkM01 (X) rkM (E1 ∪ X) − rkM (E1 ) rkM (E1 ∪ X) − rkM (E1 )
|E1 |
= ,
rkM (E1 )
implying that γ(M01 ) < γ(M). Then, we have the following lemma.
Lemma 2.4.1. Let M = (S, I) be a non-uniformly dense loopless matroid and E1

be the unique maximum cardinality set with γ(M) = |E1 |/rkM (E1 ), then the matroid
M1 = M|E1 is uniformly dense with density λ1 = γ(M) and the matroid M01 =
M/E1 is loopless with density strictly smaller than the one of γ(M).
Proof. The only missing step to prove is that E1 is unique. Indeed suppose that
there are two sets E1 and E10 of the same cardinality achieving the density λ1 of M,
in particular they also have the same rank, rkM (E1 ) = rkM (E10 ). By submodularity
of the rank function:
|E1 ∪ E10 | |E1 | + |E10 | − |E1 ∩ E10 |
≥ ,
rkM (E1 ∪ E10 ) rkM (E1 ) + rkM (E10 ) − rkM (E1 ∩ E10 )
Using that λ1 = |E1 |/rkM (E1 ) = |E10 |/rkM (E10 ) ≥ |E1 ∩ E10 |/rkM (E1 ∩ E10 ), we have
|E1 ∪ E10 | λ1 (rkM (E1 ) + rkM (E10 ) − rkM (E1 ∩ E10 ))

≥ = λ1 .
rkM (E1 ∪ E10 ) rkM (E1 ) + rkM (E10 ) − rkM (E1 ∩ E10 )
Thus, E1 ∪ E10 is a set strictly larger than E1 with the same density. This contradicts
the choice of E1 .
If the loopless matroid M01 is not uniformly dense, we can use the above lemma
in this matroid to find a second uniformly dense matroid M2 = M01 |E2 with density
|E2 | |E2 |
λ2 = γ(M01 ) = = < λ1 ,
rkM01 (E2 ) rkM (E1 ∪ E2 ) − rkM (E1 )
and a loopless matroid M02 = M01 /E2 of strictly smaller density. Here, E2 is the
maximum cardinality set achieving M01 ’s the density. By repeating this process we
obtain a sequence of sets (E1 , . . . , Ek ) partitioning S and a sequence of values λ1 >
λ2 > · · · > λk ≥ 0 for which the following holds.
Theorem 2.4.2 (Principal Partition). Let M = (S, I) be a loopless matroid and let
(Ei )ki=1 and (λi )ki=1 be the partition and sequence of values just described. Let also
F0 = ∅ and Fi = ij=1 Ej . Then, for every 1 ≤ i ≤ k, the matroid
S
Mi = (M/Fi−1 )|Ei
59
is uniformly dense with density
|Ei |
λi = .
rkM (Fi ) − rkM (Fi−1 )
Sk
Furthermore, if for every i, Ii is an independent set of Mi , then the set i=1 Ii is
independent in M.
Proof. We know from properties of matroids that contractions and restrictions com-
mute (see [137]). Therefore, the matroids (Mi )ki=1 coincide with the uniformly dense
matroids constructed iteratively using Lemma 2.4.1. The density condition follows
since
|Ei | |Ei |
γ(Mi ) = = .
rkMi (Ei ) rkM (Ei ∪ Fi−1 ) − rkM (Fi−1 )
To finish, let Ii be an independent set of matroid Mi for every i ∈ [k]. We prove by

induction that ji=1 Ii is independent in M|Fj . The claim holds for j = 1trivially.
S

For
2 ≤ j ≤ k, the fact that Ij is independent in Mj = (M/Fj−1 ) |Ej = M|Fj /Fj−1
implies that Ij ∪ J is independent in M|Fj for any J independent in M|Fj−1 . By
setting J = j−1
S
i=1 Ii we conclude the proof.
The unique sequence of sets ∅ = F0 ⊂ F1 ⊂ · · · ⊂ Fk = E is called the principal

sequence of the matroid M; λ1 > · · · > λk ≥ 1 is the associated sequence of critical
values and M1 , . . . , Mk are the principal minors of M. These sequences have been
extensively studied in the past under different names, becoming part of what today
is known as the theory of principal partitions of matroids and submodular systems.
2.4.1 Background
In this section we briefly introduce some background on principal partitions. For
more detailed explanations, we refer the reader to the Fujishige’s survey [66] and
Narayanan’s monograph [131].
The ideas of this area started with the following principal partition of a graph,
studied by Kishi and Kajitani [94]. Suppose we are given a connected graph G =
(V, E). Define the distance d(T1 , T2 ) between two spanning trees T1 and T2 as the
value |T1 \ T2 | = |T2 \ T1 |. A pair of trees is a maximally distant pair if this distance
is maximized. Kishi and Kajitani have shown the following.
Theorem 2.4.3 (Kishi and Kajitani [94]). The edge set of any connected graph
G = (V, E) can be partitioned into three sets (F − , F 0 , F + ) satisfying the following
conditions.
1. For any e ∈ F − , there is a maximally distant pair of spanning trees T1 and T2
such that e 6∈ T1 ∪ T2 .
2. For any pair of maximally distant spanning trees T1 and T2 , every element
e ∈ F 0 belongs to exactly one of T1 or T2 .
60
3. For any e ∈ F + , there is a maximally distant pair of spanning trees T1 and T2
such that e ∈ T1 ∩ T2 .
The above partition has many applications. For example, it can be used to find
the topological degree of freedom of an electrical network. This is defined as the
minimum number of current and voltage variables whose value uniquely determine
all the currents and voltages of the network using Kirchhoff’s law (see, e.g. [94] and
[131, Chapter 14]). It can also be used to solve Shannon’s switching game (see,
e.g. [19]).
Before continuing it is useful to observe the following (see, e.g. [129, Chapter 9]).
Let S be a finite set and f : 2S → R be a submodular function, this is
f (A ∪ B) + f (A ∩ B) ≤ f (A) + f (B), for all A, B ∈ S. (2.18)
Then the family of minimizers of f form a distributive lattice, that is, a collection
closed for set union and set intersection. In particular this family has a unique
minimal and a unique maximal set.
Kishi and Kajitani’s principal partition can be characterized by the next theorem.
Theorem 2.4.4. Consider the set DG of all the minimizers of the submodular func-
tion
f (X) = 2 rk(X) + |E \ X|,
for X ∈ 2E , where rk(·) is the rank function of the graphic matroid of G. The unique
minimal element of DG is given by F − and the unique maximal element of DG by
F − ∪ F 0 = E \ F + , where (F − , F 0 , F + ) is the Kishi–Kajitani partition of E for
G = (V, E).
Recall also the following min-max theorem for the union of matroids (see, e.g. [147,
Chapter 42]).
Theorem 2.4.5. For a matroid M = (S, I) with rank function rk(·), and any p ≥ 1,
( p )
[
max Ii : Ii
∈I = min {p rk(X) + |S \ X| : X ⊆ S} .
i=1
In particular, the value minX⊆E f (X), with f (X) as in Theorem 2.4.4, corresponds
exactly to the maximum size of the union of 2 trees in G. Bruno and Weinberg [19]
have noticed the matroidal structure of Kishi and Kajitani’s result and have extended
their partition to unions of p copies of a given matroid. Their result can be stated as
follows.
Theorem 2.4.6 (Bruno and Weiberg [19]). For any matroid M = (S, I), and p ≥ 1,
let Dp be the distributive lattice of the minimizers of the submodular function
fp (X) = p rk(X) + |S \ X|.
Let Fp− and Fp+ be the unique minimum and maximum elements of Dp respectively.
Suppose that the collection of distinct Dp is given by Dp1 , . . . , Dpk with p1 > · · · > pk ,
61
then we have
Fp−1 ⊆ Fp+1 ⊆ Fp−2 ⊆ Fp+2 ⊆ · · · ⊆ Fp−k ⊆ Fp+k .
In particular, if the difference set Fp+i \ Fp−i is nonempty, the minor (M|Fp+ )/Fp−i is a
i
matroid having pi disjoint bases partitioning Fp+i \ Fp−i .
Tomizawa [163] and Narayanan [128] independently generalized the decomposition

above by considering rational numbers instead of integers p. For a rational number
p/q with p and q positive integers, they find a minor of M that has p bases uniformly
covering each element of the underlying set q times. More precisely, consider the
following extended min-max theorem for the union of matroids.
Theorem 2.4.7. For a matroid M = (S, I) with rank function rk(·), and p, q ≥ 1,
( p )
X
max |Ii | : Ii ∈ I, |{i : i ∈ [p], s ∈ Ii }| ≤ q, ∀s ∈ S
i=1
= min {p rk(X) + q |S \ X| : X ⊆ S} .
The above result follows from applying Theorem 2.4.5 to the q-parallel extension
of M, which is the matroid obtained by replacing replacing each element of M with
q parallel elements3 . Tomizawa and Narayanan show the following.
Theorem 2.4.8 (Tomizawa [163], Narayanan [128]). For any matroid M = (S, I),
and any λ = p/q for positive integers p, q ≥ 1, let Dλ be the distributive lattice of the
minimizers of the submodular function
fλ (X) = p rk(X) + q |S \ X|.
We call the value λ critical if Dλ contains more than one element. Then, there is a
finite sequence of critical values λ1 > · · · > λk ≥ 0. For each i ∈ [k], let Fλ−i and Fλ+i
be the minimum and maximum elements of Dλi , respectively. Then we have:
∅ = Fλ−1 ⊂ Fλ+1 = Fλ−2 ⊂ Fλ+2 = · · · ⊂ Fλ+k−1 = Fλ−k ⊂ Fλ+k = S.
And the minor Mλi = (M|F + )/Fλ−i , for λi = p/q is a matroid having p bases covering
λi
each element of Fλ+i \ Fλ−i exactly q times.
Note that we could have replaced the function fλ (X) in the previous theorem by
fλ (X) = λ rk(X) + |S \ X| or by fλ (X) = λ rk(X) − |X| and we would have obtained
the same result.
The sequence of distinct sets ∅ = Fλ−1 ⊂ Fλ+1 ⊂ Fλ+2 ⊂ · · · ⊂ Fλ+k = S is denoted as
the principal sequence of M, the values (λi )ki=1 are the associated critical values and
3
To be precise, let S 0 be a set having q copies of each element of S. The q-parallel extension of
M is the matroid having S 0 as ground set and where a set I 0 is independent if it has at most one
copy of each element of S and the respective set I of original elements is independent in M.
62
the minors (Mλi )ki=1 are the principal minors of M. Furthermore, the union of all
the distributed lattices (Dλi )ki=1 is usually known as the principal partition of M.
It is important to remark that when M is loopless, the construction above coin-
cides with the one given in Theorem 2.4.2. (See also [131] and [22].)
To conclude this small survey, we remark that the constructions above for the
rank function of a matroid can be extended also to polymatroid ranks and to even
more general functions (see Fujishige’s survey [66]). An interesting extension is the
following (see [131, Chapter 10]).
Theorem 2.4.9. Let f : 2S → R be a submodular function and g : 2S → R be a

strictly increasing polymatroid rank function (this is, g(∅) = 0 and for all X ⊂ Y ,
g(X) < g(Y )). For every λ ≥ 0, let Dλ be the distributive lattice of the minimizers
of the submodular function
hλ (X) = λf (X) + g(S \ X).
We call the value λ critical if Dλ contains more than one element. Then, there is
a finite sequence of critical values λ1 > · · · > λk ≥ 0. For each i ∈ [k], let Fλ−i and
Fλ+i be the minimum and maximum elements of Dλi , respectively. Then we have
∅ = Fλ−1 ⊂ Fλ+1 = Fλ−2 ⊂ Fλ+2 = · · · ⊂ Fλ+k−1 = Fλ−k ⊂ Fλ+k = S.
The sequence of distinct sets ∅ = Fλ−1 ⊂ Fλ+1 ⊂ Fλ+2 ⊂ · · · ⊂ Fλ+k = S is called the
principal sequence of (f (·), g(·)).
As shown, for example in [131, Chapter 10], for the particular case where f is
submodular and g is a weight function (such as in the case of the principal partition
of matroids) there are polynomial time algorithms to find the principal sequence of
(f (·), g(·)) and the associated critical values.
2.4.2 Matroids Related to the Principal Partition

Consider a general (not necessarily loopless) matroid M = (S, I) and let L be its set
of loops. Let (Fi )ki=0 , (λi )ki=1 and (Mi )ki=1 be the principal sequence, critical values
and principal minors of the loopless matroid M \ L. Also, for every i ∈ [k], let
Ei = Fi \ Fi−1 and ri denote the ground set and the total rank of matroid Mi . Recall
also that the density γ(Mi ) of the matroid Mi is equal to λi = |Ei |/ri .
The family {L, E1 , . . . , Ek } is a partition of the ground set S. Define two new
matroids M0 and P 0 with ground set S and independent sets as follows:
( k )
0
[
I(M ) = Ii : Ii ∈ I(Mi ) , and
i=1
( k )
0
[
I(P ) = Ii : Ii ⊆ Ei , |Ii | ≤ ri .
i=1
63
In other words, if U(X, r) is the uniform matroid on set X having rank r then
k
M0 = U(L, 0) ⊕
M
Mi . (2.19)
i=1
k
P 0 = U(L, 0) ⊕
M
U(Ei , ri ). (2.20)
i=1
Our main task now is to show that M0 and P 0 are, in a well-defined manner,
good approximations of M. Before doing that, let us define a third random matroid
related to the previous ones.
Recall the definition of the random partition matroid P(Mi ) associated to Mi .
In P(Mi ), each element of Ei receives a color in [ri ] uniformly at random. Let Bij be
the set of elements in Ei that are assigned color j. The independent sets of P(Mi )
are those subsets of Ei having at most one element in each part Bij .
Consider the random matroid Q0 obtained by replacing each summand Mi of M0
by the matroid P(Mi ). This is,
k ri
k M
Q0 = U(L, 0) ⊕
M M
P(Mi ) = U(L, 0) ⊕ U(Bij , 1). (2.21)
i=1 i=1 j=1
Now we show some properties of the previous matroids.
Lemma 2.4.10. Any independent set of M0 is independent in M.

Proof. It follows directly from the definition of each Mi = (M/Fi−1 ) .
Ei
Lemma 2.4.11. Any independent set of Q0 is independent in P 0 .
Proof. As any independent set of Q0 contains at most one element in each Bij , it also
contains at most ri element in each Ei .
The following theorem is the main result of this section. It states that random
independent sets in Q0 are likely to have large rank in the original matroid M.
Theorem 2.4.12. Let 1 ≤ ` ≤ n, and X` be a random subset of ` elements of S.

Then
EX`,Q0 [rkQ0 (X`)] ≥ 1 − e EX` [rkM(X`)].
1

(2.22)
In order to prove Theorem 2.4.12, we need two technical lemmas.
Lemma 2.4.13. For every 1 ≤ ` ≤ n, and every 1 ≤ i ≤ k,
EX ,Q [rkQ (X` ∩ Ei)] ≥ ri

`
0 0 1 − exp(−λi `/n) .
Proof. The value on the left hand side depends only on X` and on the subpartition
{Bij }j∈[ri ] of Ei . Since these two objects are chosen independently at random, we can
assume they are constructed as follows.
64
First select X` ⊆ S uniformly at random. Assign then to each element of X`
a color in {1, . . . , ri }. Let X`,j be the set of elements in X` having color j. Use
those colors to assign the partition of Ei : Every element in Ei ∩ X`,j is assigned to
the set Bij . Finally, each of the elements in Ei \ X` selects a value j ∈ {1, . . . , ri }
uniformly, and is assigned then to the corresponding Bij .
Using the definition of Q0 ,
ri ri
EX ,Q [rkQ (X` ∩ Ei)] =
X X
`
0 0 Pr(X`,j ∩ Ei 6= ∅) = (1 − Pr(X`,j ∩ Ei = ∅)) . (2.23)
j=1 j=1
Focus on the j-th term of the sum above and condition on the size t of X`,j . Under
this assumption, X`,j is a random subset of S of size t. From here,

n−|Ei | t−1
! !t

t Y |Ei | |Ei |
Pr X`,j ∩ Ei = ∅ |X`,j | = t = = 1− ≤ 1− .

n n−`

`=0 n
t
By removing the conditioning and using that t is a binomial random variable with
parameters ` and 1/ri ,
`
!t ! t `−t
X |Ei | ` 1 1
Pr(X`,j ∩ Ei = ∅) ≤ 1− · 1−
t=0 n t ri ri
! !`
1 |Ei | 1

= 1− + 1−
ri n ri
!`
|Ei |
= 1− .
nri
Replacing this in (2.23), and using that λi = |Ei |/ri we have

 !` 
EX`,Q0 [rkQ0 (X` ∩ Ei)] ≥ ri 1 − 1 − |λni|

 ≥ ri 1 − exp(−λi `/n) .
Consider a random set X` of size ` in S. The rank of X` in Q0 |Ei = P(Mi ) is

simply the number of subparts in {Bi1 , . . . , Biri } this set intersect. If Ei has high
density (say λi ≥ n/`), then we expect Ei to contain |Ei |(`/n) ≥ ri elements of X` .
As they are roughly equally distributed among the subparts of Ei , we expect the rank
of X` ∩ Ei to be close to ri . On the other hand, if the set Ei has low density then we
expect it to contain less than ri elements of X` , and so we expect the rank of X` ∩ Ei
to be closer to its expected cardinality. The following lemma formalizes this intuition.
65
Lemma 2.4.14. For every 1 ≤ ` ≤ n, and every 1 ≤ i ≤ k,
EX ,Q [rkQ (X` ∩ Ei)] ≥ 1

min EX` [|X` ∩ Ei |], ri
n o
`
0 0 1−
 e
(1 − 1/e) r , if λi ≥ n/`;
i
=
(1 − 1/e)|Ei |`/n, if λi ≤ n/`.
Proof. First note that EX` [|X` ∩ Ei |] = |Ei |(`/n). This quantity is larger than or
equal to ri if and only if λi ≥ n/`. Suppose that this is the case. Using Lemma 2.4.13
and that the function (1 − e−x ) is increasing, we obtain
EX ,Q [rkQ (X` ∩ Ei)] ≥

`
0 0 1 − exp(−λi `/n) ri
min EX` [|X` ∩ Ei |], ri .

1 1
n o
≥ 1− ri = 1 −
e e
Suppose now that λi ≤ n/`. Since the function (1 − e−x )/x is decreasing, we obtain

1 − exp(−λi `/n)
EX ,Q [rkQ (X` ∩ Ei)] ≥
`
0 0
λi `/n
ri λi `/n
min EX` [|X` ∩ Ei |], ri .

1 1
n
o
≥ 1− |Ei |`/n = 1 −
e e
Now we are ready to prove Theorem 2.4.12
Proof of Theorem 2.4.12. Since the densities (λi )ki=1 form a decreasing sequence, there
S∗
is an index i∗ such that λi ≥ n/` if and only if 1 ≤ i ≤ i∗ . The set ii=1 Ei is equal to
the set Fi∗ in the principal sequence of the matroid M \ L. Let F = L ∪ Fi∗ .
Every set in the principal sequence has the same rank in both M and Q0 . Using
this fact and properties of the rank function we get:
EX [rkM(X`)] ≤ EX [rkM(X` ∩ F ) + rkM(X` ∩ (S \ F ))]

` `
≤ rkM (Fi ) + EX [|X` ∩ (S \ F )|]

∗
`
i∗
X k
X
= ri + |Ei |(j/n).
i=1 i=i∗ +1
≤
Pk
E
i=1 X` ,Q0 [rkQ0 (X` ∩ Ei )]
(1 − 1/e)
=
EX` [rkQ0 (X`)] ,
,Q0
(1 − 1/e)
where the last inequality follows from Lemma 2.4.14. This concludes the proof.
To end this section, we prove a lemma that translates the result of Theorem 2.4.12
to a more useful setting for the matroid secretary problem.
66
Lemma 2.4.15. For two matroids M1 and M2 on the same ground set S of size
n (but possibly having randomly defined independent set families), and a constant
α ≥ 0, the following two statements are equivalent:
(i) For all 1 ≤ ` ≤ n,
EX ,M [rkM (X`)] ≥ α EX ,M [rkM (X`)],

` 1 1 ` 2 2 (2.24)
where X` is a cardinality ` subset of S selected uniformly at random, indepen-

dently from any random choice defining the matroids.
(ii) For every adversarial list of weights w1 ≥ w2 ≥ · · · ≥ wn ≥ 0,
Eσ,M [w(OPTM (σ))] ≥ αEσ,M [w(OPTM (σ))],

1 1 2 2 (2.25)
where σ : [n] → S is a bijective map selected uniformly at random, independently

from any random choice defining the matroids.
Proof. We start by rewriting Eσ,M [w(OPTM (σ))] in a more useful way.

Let X`σ = {σ(1), σ(2), . . . , σ(`)} be the (random) set of elements in S receiving
the top ` weights of the adversarial list of weights. Note that
Pr(σ(`) ∈ OPTM (σ)) = Pr(rkM (X`σ ) − rkM (X`−1

σ
) = 1)
= Eσ,M [rkM (X` )] − Eσ,M [rkM (X`−1
σ σ
)]
= EX` ,M [rkM (X` )] − EX`−1 ,M [rkM (X`−1 )],
where for the last line we used that X`σ is a uniform random set of ` elements. Then,
n
Eσ [w(OPTM(σ))] =
X
w` Pr(σ(`) ∈ OPTM (σ))
`=1
n
EX ,M[rkM(X`)] − EX
X
= w` ` `−1 ,M
[rkM (X`−1 )]
`=1
n−1
= wn EXn ,M [rkM (Xn )] + (w` − w`+1 ) EX` ,M [rkM (X` )]. (2.26)
X
`=1
Assume that condition (i) holds, then each term for M1 in the above sum is at least
α times the corresponding term for M2 , implying that condition (ii) holds.
On the other hand, if condition (i) does not hold, then then there is an index j,
for which
EX`,M1 [rkM1 (X`)] < αEX`,M2 [rkM2 (X`)].
Consider the sequence of weights given by w1 = w2 = · · · = wj = 1 and wj+1 =
· · · = wn = 0. For this sequence
Eσ [w(OPTM (σ))] = EX ,M [rkM (X`)]

1 ` 1 1
< αEX ,M [rkM (X` )] = αEσ [w(OPTM (σ))].

` 2 2 2
67
As a corollary, we obtain the following result for the partition matroids P 0 , Q0
associated to M.
Lemma 2.4.16. For every adversarial list of weights w1 ≥ w2 ≥ · · · ≥ wn ≥ 0,
Eσ [w(OPTP 0 (σ))] ≥ Eσ,Q0 [w(OPTQ0 (σ))] ≥ 1 − 1e Eσ [w(OPTM(σ))].

Proof. The first inequality follows by Lemma 2.4.11, and the second from Theo-
rem 2.4.12 and Lemma 2.4.15.
2.5 General Matroids

In this section we apply the divide and conquer idea of Section 2.2 to obtain constant
competitive algorithms for general matroids.
2.5.1 Random-Assignment Random-Order Model

We use the same notation as in the previous section. Consider Algorithm 6 depicted
below.
Algorithm 6 for general matroids in the random-assignment random-order model.

1: Compute the principal minors (Mi )ki=1 of the matroid obtained by removing the
loops of M.
2: Run Algorithm 4 (with parameter p) for uniformly dense matroids in parallel on
each Mi and return the union of the answers.
As the set returned is independent in the matroid M0 = U(L, 0) ⊕ i∈[k] Mi , it is

L
also independent in M. Therefore, the algorithm is correct. In order to estimate its

competitive ratio, we use the properties of Algorithm 4 for uniformly dense matroids.
Theorem 2.3.5 states that for every uniformly dense matroid N of total rank r, Al-
gorithm 4 (with parameter p) returns a set of expected weight at least γ2 (p) times the
weight of the top r weights in the matroid (this is stronger than just being (γ2 (p))−1 -
competitive). In other words, Algorithm 4 returns a set of expected weight at least
γ2 (p) times the one of the optimum independent set in the uniform matroid of total
rank r having the same ground set as N . As every matroid Mi = (Ei , Ii ) is uniformly
dense, the previous argument suggests that Algorithm 4 recovers, in expectation γ2 (p)-
fraction of the optimum weight of the partition matroid P 0 = U(L, 0)⊕ i∈[k] U(Ei , ri )
L
defined in the previous section. We now formalize this result.

Lemma 2.5.1. Let ALG be the set returned by Algorithm 6. Then,
Eσ,π [w(ALG)] ≥ γ2(p) Eσ [w(OPTP (σ))].

0 (2.27)
Proof. The only ingredient left is to argue that Algorithm 4 is effectively being ap-
plied over an instance of the random-assignment random-order model. Observe that
68
the random bijection σ : [n] → W that is used to assign the weights in W to the
elements of the matroid can be viewed as the composition of a random partition of
[n] and W into blocks of sizes (|L|, |E1 |, |E2 |, . . . , |Ek |), and a collection of random
bijections between the corresponding blocks. Conditioned on the random partition,
each block receives a hidden list of weights which are assigned uniformly at random
to the elements of the block. To complete the proof we only need to observe that
the random order in which the elements of each block are presented to the algorithm
processing that block is also uniform.
Now we can a give a first bound for the competitive ratio of Algorithm 6.
Theorem 2.5.2. For p = p2 ≈ 0.384374, Algorithm 6 is 1/(γ2 (p2 )(1 − 1/e)) ≈

7.78455-competitive for the random-assignment random-order model.
Proof. Direct from Lemmas 2.5.1 and 2.4.16.
We note here that this competitive ratio is already an improvement over the
2e/(1 − 1/e) ≈ 8.60052 algorithm presented by the author in the SODA paper [152],
which used Algorithm 3 instead of Algorithm 4 as a subroutine.
We can further improve the estimation of this competitive ratio by using another
property of Algorithm 4. Theorem 2.3.5 states that, when applied on a uniformly
dense matroid N , Algorithm 4 (with parameter p) returns a set of expected weight
at least γ1 (p) times the expected weight of the optimum in the random partition
matroid P(N ). Recall now that the random partition matroid Q0 defined from a
general matroid M contains, for every uniformly dense matroid Mi , a summand
P(Mi ). Since every summand is treated independently in Algorithm 6, we conclude
that this algorithm recovers, in expectation, γ1 (p)-fraction of the optimum weight of
the partition matroid Q0 .
Lemma 2.5.3. Let ALG be the set returned by Algorithm 6. Then,
Eσ,π [w(ALG)] ≥ γ1(p)Eσ,Q [w(OPTQ (σ))].

0 0 (2.28)
Proof. The proof is analogous to the one of Lemma 2.5.1.
Now we can give the current tightest bound for the competitive ratio of Algo-
rithm 6.
Theorem 2.5.4. For p = p1 ≈ 0.433509, Algorithm 6 is 1/(γ1 (p1 )(1−1/e)) ≈ 5.7187-

competitive for the random-assignment random-order model of the matroid secretary
problem.
Proof. Direct from Lemmas 2.5.3 and 2.4.16.
69
2.5.2 Random-Assignment Adversarial-Order Model
Oveis Gharan and Vondrák [136] have noticed that by combining their 40-competitive
algorithm for uniformly dense matroids on the random-assignment adversarial-order
model with our Lemma 2.4.16 one can get a 40/(1 − 1/e)-competitive algorithm for
general matroids. We improve this result by using Algorithm 5 instead.
Consider the procedure depicted as Algorithm 7 below.
Algorithm 7 for general matroids in the random-assignment adversarial-order

model.
1: Compute the principal minors (Mi )ki=1 of the matroid obtained by removing the
loops of M.
2: Run Algorithm 5 for uniformly dense matroids in parallel on each Mi and return
the union of the answers.
Theorem 2.3.6 states that for every uniformly dense matroid N of total rank r,
Algorithm 5 returns a set of expected weight at least 1/16 times the expected weight
of the optimum in the random partition matroid P(N ). In particular, since each
matroid Mi is uniformly dense, and the elements of each matroid receive a random
permutation of certain hidden list of weights (see the proof of Lemma 2.5.1), Al-
gorithm 7 recovers in expectation 1/16 times the optimum weight of the partition
matroid Q0 defined in Section 2.4.2. Therefore, we have the following theorem.
Theorem 2.5.5. Let ALG be the set returned by Algorithm 7 when applied on a
uniformly dense matroid. Then,
Eσ [w(ALG)] ≥ 161 Eσ,Q [w(OPTQ (σ))].

0 0 (2.29)
By Lemma 2.4.16, Algorithm 7 is 16/(1 − 1/e) ≈ 25.3116-competitive for the random-

assignment adversarial-order model of the matroid secretary problem.
2.6 New Results for the Adversarial-Assignment

Random-Order Model
Constant competitive algorithms for the adversarial-assignment random-order model
of the matroid secretary problem remain elusive.
Unfortunately, the strategy used for random-assignment models is not very useful
in this setting. Lai [100] has shown that every matroid M is a restriction of a
uniformly dense matroid M0 . This means that any algorithm for uniformly dense
matroids can be transformed immediately into one for general matroids having the
same competitive ratio (see Lemma 1.5.5).
The best algorithm for the adversarial-assignment random-order model so far is
an O(log r)-competitive due to Babaioff et al. [8] (see Lemma 1.5.4). This algorithm
has many features, including the fact that it does not need to know the matroid be-
forehand; it only needs to know the number of elements and have access to an oracle
70
that test independence only on subsets of elements it has already seen. Nevertheless,
this algorithm makes use of the actual values of the weights being revealed (in other
words, it is not a comparison-based algorithm, as our definition for the matroid sec-
retary problem requires). In this section we present a new algorithm having the same
features but that only uses the relative order of weights seen and not their numerical
values.
Later in this section, we show new constant-competitive algorithms in this model
for certain matroid classes.
We preserve the same notation used in Section 2.1. The only difference with the
previous settings is that the assignment σ : [n] → S is adversarial. In particular,
w1 ≥ w2 ≥ · · · ≥ wn ≥ 0 denotes the collection of element weights in decreasing order
(without loss of generality we assume there are no ties as otherwise the algorithm
can break them using a random permutation τ ). Also, we use σ(1), σ(2), . . . , σ(n) to
denote the elements of S sorted from largest to smallest weight.
We also use indistinctly OPTM or OPTM (σ) to denote the optimum independent
set {x1 , . . . , xr } of M under this ordering, with σ −1 (x1 ) < σ −1 (x2 ) < · · · < σ −1 (xr )
(or equivalently w(x1 ) > w(x2 ) > · · · > w(xr )). We drop the subindex M whenever
there is no possibility for confusion.
2.6.1 General O(log r)-Competitive Algorithm

The following algorithm returns an independent set of the matroid M. With prob-
ability 1/2, run the classical secretary algorithm (say, Algorithm 1) on the set of
non-loops of the matroid. This returns the heaviest non-loop of the stream with
probability at least 1/e. Otherwise, observe the first m elements of the stream,
where m is chosen from the binomial distribution Bin(n, 1/2) (as usual, denote this
set of elements as the sample) and compute the optimum base A = {a1 , . . . , ak }
(with w(a1 ) ≥ · · · ≥ w(ak )) of the sampled elements. Afterwards, select a number
` ∈ {1, 3, 9, . . . , 3t } with t = blog3 rc uniformly at random, run the greedy procedure
on the set of non-sampled elements having weight at least the one of a` as they ar-
rive and return its answer (if ` > k, run the greedy procedure over the entire set of
non-sampled elements).
It is possible to implement this algorithm without knowing the matroid before-
hand; we only need to know the number of elements n and have access to an oracle
to test independence on subsets of already seen elements. For that we need to make
two changes to the algorithm above.
First we require to make a slight modification to the algorithm for the classical
secretary problem we use (Algorithm 1) so that it only consider the non-loops of
the stream without knowing a priori the number of non-loops. The modification
samples the first N elements of the stream, where N is distributed as Bin(n, 1/e),
and then returns the first non-loop having weight larger than any sampled non-loop
(if any). Observe that if n0 is the number of non-loops of the matroid, then the
number of non-loops sampled has distribution Bin(n0 , 1/e). This observation implies
this algorithms does exactly what Algorithm 1 would do if it knew the number of
non-loops beforehand.
71
The second change deals with the number t = blog3 rc in the algorithm above. As
we do not know the rank of the matroid a priori, we can not use this value. Instead,
we use the rank k of the sampled set (which we can compute) to estimate it: We select
t ∈ {blog3 kc, blog3 kc + 1} uniformly at random and use this value in the previous
algorithm.
The full description of this algorithm is depicted as Algorithm 8.
Algorithm 8 for general matroids.

1: With probability 1/2 run an e-competitive algorithm for the classical secretary
problem on the set of non-loops and return its answer.
2: Otherwise, set ALG ← ∅.
3: Choose m from the binomial distribution Bin(n, 1/2).
4: Observe the first m elements and denote this set as the sample.
5: Compute the optimum base A for the sample. Let a1 , . . . , ak be the elements of
A in decreasing order of weight.
6: If the total rank r of the matroid is known, set t = blog3 rc, otherwise, select
t ∈ {blog3 kc, blog3 kc + 1} uniformly at random.
7: Select ` uniformly at random from the set {1, 3, 9, . . . , 3t }.
9: If ALG ∪ {x} is independent and w(x) ≥ w(a` ) (where w(a` ) = 0 if ` > k),
add x to ALG.
10: end for
To analyze this algorithm, note the following. For every ` the algorithm can
choose (provided ` ≤ k), the sample contains an independent set of size ` containing
only elements of weight at least the one of a` (namely the set {a1 , . . . , a` } itself).
Since the sampled set behaves similarly to the non-sampled one, we expect that the
same happens outside the sample. In particular, the greedy procedure should recover
a weight of roughly `w(a` ). By taking the expectation over the choices of ` it is
not hard to check that the expected weight returned by the algorithm is at least
Ω(E[w(A)/ log3 (r)]) = Ω(E[w(OPT)/ log3 (r)]). We give the formal proof below.
Theorem 2.6.1. Algorithm 8 is O(log r)-competitive for any matroid of rank r.
Proof. Assume first that the rank r of the matroid is known. Let OPT = {x1 , . . . , xr }
with w(x1 ) > · · · > w(xr ) be the maximum independent set of the matroid, T the set
of sampled elements and T 0 the set of non-sampled ones.
Let A = {a1 , . . . , ak } be the optimum set in T , with w(a1 ) > · · · > w(ak ) (in-
dependent of whether the algorithm computes A or not). This set can be obtained
by sorting T in decreasing order of weights and applying the greedy procedure. This
fact implies that if an element xi of the optimum is sampled, then xi appears in the
set A.
Every element of the matroid is sampled independently with probability 1/2,
including the elements of the optimum. Therefore, by the previous paragraph,
72
E[w(A)] ≥ w(OPT)
2
. (2.30)
To simplify our analysis, in the following we assume that for i > k, ai is a dummy
element with w(ai ) = 0. Given the number ` chosen by the algorithm (if the algorithm
reaches that state), the weight of the set returned will be at least w(a` ) times the
number of elements the greedy procedure selects; therefore, E[w(ALG)] is at least
blog3 rc
w(x1 ) 1
E
h i
w(a` ) · |ALG| ` = 3j was selected .
X
+ (2.31)

2e 2(1 + blog3 rc) j=0
Let H(a` ) be the collection of non-sampled elements that are at least as heavy
as a` . If the algorithm chooses the number `, it will then execute the greedy procedure
on H(a` ) and return a set of cardinality equal to the rank of H(a` ). Note that
for every `, w(x` ) ≥ w(a` ); therefore, the rank of H(a` ) is at least the number of
nonsampled elements in {x1 , . . . , x` }.
Using a Chernoff bound (see, e.g. [118]),

0
Pr |{x1 , . . . , x` } ∩ T | ≤ `/4 ≤ exp(−`/8).
In particular, if ` ≥ 9,
E[w(a`) · |ALG| ` ] ≥ E[w(a`)](1 − exp(−`/8))`/4 ≥ E[w(a`)]`/6.

Replacing this in (2.31), and dropping the values of j for which ` < 9, we get
blog3 rc
E[w(ALG)] ≥ w(x 1) 1
E[w(a3 )]3j
X
+ j
2e 12(1 + blog 3 rc) j=2
 
blog3 rc
≥ E
w({a1 , . . . , a8 }) 1 X
+ w({a3j , . . . , a3j+1 −1 })
16e 24(1 + blog3 rc) j=2
≥
E[w(A)] .
16e(1 + blog3 rc)
Using Inequality (2.30), we obtain
E[w(ALG)] ≥ 32e(1w(OPT)
+ blog rc)
,
3
which implies the algorithm is O(log r)-competitive.

Suppose now that the rank r is unknown. If r is small, say r ≤ 12, then with
probability 1/(2e) the algorithm will run the standard secretary algorithm and return
element x1 . This element has weight at least 1/12 fraction of the optimum; therefore
the algorithm is 24e-competitive for this case.
For the case where r > 12 we use a different analysis. The random variable k
73
denoting the rank of the sampled set could be strictly smaller than r. However, the
probability that k ≤ r/3 is small. Indeed, for that event to happen we require that at
most 1/3 of the elements of OPT are in the sample. By Chernoff bound, this happens
with probability

Pr |{x1 , . . . , xr } ∩ T | ≤ r/3 ≤ exp(−r/18) ≤ exp(−13/18) ≤ 1/2.
Noting that r/3 ≤ k ≤ r implies that blog3 rc ∈ {blog3 kc, blog3 kc + 1}, we deduce
that with probability at least 1/4 our algorithm guesses t = blog3 rc right; therefore,
the competitive ratio of this algorithm is at most 4 times worse than the one that
knows the rank beforehand.
2.6.2 Column-Sparse Linear Matroids

Let M = (V, I) be a linear matroid represented by a matrix A. Consider the following
algorithm for M (depicted as Algorithm 9): Randomly permute the rows of A and
define for every row i, the sets Ci = {v ∈ V : vi 6= 0} and Bi = Ci \ j<i Cj , where
S
vi denotes the i-th coordinate of column v in the permuted matrix. Next, run any
e-competitive secretary algorithm for the partition matroid that accepts at most one
element from each Bi .
Algorithm 9 for a matroid M = (V, I) represented by a matrix A.

1: Permute the rows of A at random to obtain a matrix A0 . Index the rows of A0 as
1, 2, . . . , `.
2: Let Ci = {v ∈ V : vi 6= 0} and Bi = Ci \ j<i Cj where vi is the i-th coordinate
S
of the column associated to v in matrix A0

3: Let P be the partition matroid on V whose independent sets contain at most one
element from each Bi .
4: Run any e-competitive secretary algorithm for P on the stream of elements and
return its answer.
Theorem 2.6.2. Algorithm 9 returns an independent set of M. Furthermore, if

the matrix A representing M is such that every column contains at most k non-zero
entries, then this algorithm is ke-competitive (assuming a random-order model).
Proof. We show first that the returned set is independent. If this was not the case
there would be a circuit C inside the output. Let v ∈ C be the element belonging
to the set Bi of smallest index i. By definition of v, the elements of C \ v are not
in Ci ; therefore, C and Ci intersects only in v. This is a contradiction since Ci is in
the cocircuit space of the matroid. (Use, e.g. [137, Proposition 2.1.11].)
To show that the algorithm is ke-competitive, construct the bipartite graph G
having parts the rows and columns of A, where there is an edge (i, v) from row i to
column v if the corresponding entry of A is non-zero. Assign to every edge incident
to column v a weight equal to the weight of the matroid element v.
74
Consider the following simulation algorithm: Randomly permute the vertices in
the row part of the graph. Delete all the edges, except the ones going from a column
vertex to its lowest neighbor (this is, to the row having smallest index in the random
permutation). Finally, run any e-competitive secretary algorithm for the partition
matroid that accepts for each row vertex, at most one edge incident to it. This
returns a matching in G with the same weight as the set of elements the original
algorithm returns.
If X is a set of columns independent in M then the row-rank of the submatrix
of A induced by X is equal to its cardinality. In particular, the number of row
vertices that X dominates in G is at least |X|. Using Hall’s Theorem we conclude
that there is a matching covering each independent set of columns. In particular, the
weight of the maximum weight matching M ∗ of G is at least the one of the optimum
independent set of M. On the other hand, M ∗ has weight at most the one of the edge
set {(i, v ∗ (i)) : i ∈ rows(A)}, where v ∗ (i) is the maximum weight neighbor of i in G.
Since each edge (i, v ∗ (i)) is not deleted with probability 1/k and, given that it is not
deleted, the simulation selects it with probability 1/e, we conclude that Algorithm 9
is ke-competitive.
By applying this algorithm to graphic matroids, which are representable by matri-

ces having only 2 ones per column, we recover the 2e-competitive algorithm of Korula
and Pál [97].
2.6.3 Low Density Matroids

The matroid polytope PM ⊆ RS of a matroid M = (S, I) is the convex hull of
the indicator vectors of its independent sets. This polytope can be characterized (see,
e.g. [147, Chapter 40]) as
PM = {y ∈ RS : y ≥ 0 and y(U ) ≤ rk(U ), for all U ⊆ S}.
Let M = (S, I) be a loopless matroid of density
|U |
γ(M) = max ,
∅6=U ⊆S rk(U )
and τ M ∈ RS be the vector having all its coordinates equal to 1/γ(M), then we have
the following property.
Lemma 2.6.3. For every loopless matroid M, the corresponding vector τ M is in the
matroid polytope PM .
Proof. This follows since for every U ⊆ S,
|U |
τ M (U ) = τ M (u) =
X
≤ rk(U ).
u∈U γ(M)
75
The previous lemma implies that τ M admits a decomposition as convex combina-
tion of independent sets of M:
τM =
X X
λI χI , with λI = 1.
I∈I I∈I
This decomposition can be found in polynomial time given access to an independent

oracle of M (see [147, Chapter 40]). Consider the following algorithm (depicted as
Algorithm 10) for a loopless matroid M = (S, I).
Algorithm 10 for loopless matroid M = (S, I).

Compute the decomposition τ M = I∈I λI χI .
P
1:
2: Select and return a set I ∈ I according to the probability distribution (λI )I∈I .
Lemma 2.6.4. For every loopless matroid M, Algorithm 10 is γ(M)-competitive in

any model of the matroid secretary problem (even adversarial-assignment adversarial-
order)
Proof. Every element u ∈ S is in the output with probability
λI = τ M (u) = 1/γ(M).
X
I∈I : u∈I
Therefore, the expected weight returned is at least 1/γ(M) times the collective total
weight of all the elements in the matroid.
If a loopless matroid M contains parallel elements4 we can potentially get a better

competitive ratio by using its simple version M0 . This matroid M0 = (S 0 , I 0 ) is
obtained by deleting all but one element in each parallel class of M = (S, I). In
particular, for every element u in M there is a unique element in M0 representing
u’s parallel class.
For every independent set I 0 ∈ M0 , let Q(I 0 ) be the partition matroid in S induced
by I 0 . In other words, the independent sets of Q(I 0 ) are those subsets of S containing
at most one element from each parallel class represented in I 0 . In particular, every
independent set in Q(I 0 ) is independent in the original matroid M.
Consider Algorithm 11 depicted below.
Lemma 2.6.5. For every loopless matroid M, Algorithm 11 is eγ(M)-competitive

in the adversarial-assigment random-order model of the matroid secretary problem.
Proof. The set returned by this algorithm is independent in the original matroid M,
hence the algorithm is correct. Note that every element u of the optimum base
4
Two elements u and v are parallel in M if {u, v} is a minimal dependent set. Being parallel is
an equivalence relation (considering that every element is parallel to itself). The parallel classes of
M are the equivalence classes of this relation. A matroid is called simple if it has no loops and no
pair of parallel elements.
76
Algorithm 11 for loopless matroid M = (S, I).
1: Construct the simple matroid M0 = (S 0 , I 0 ) and compute the decomposition
0
τ M = I 0 ∈I 0 λI 0 χI 0 .
P
2: Select a set I 0 ∈ I 0 according to the probability distribution (λ0I )I 0 ∈I 0 .

3: Run any e-competitive algorithm for partition matroids on the matroid Q(I 0 ) and
return its answer.
of M is the heaviest of its own parallel class. Provided that the parallel class of u is
represented in the set I 0 selected in line 2, Algorithm 11 returns u with probability
at least 1/e. We conclude the proof by noting that every parallel class (i.e. every
element of S 0 ) is selected with probability γ(M0 ).
It is possible to modify the previous algorithms to work on matroids having loops:

Simply run them on the matroid obtained by removing the loops. The competitive
ratio of the resulting algorithm depends on the density of the resulting matroid.
The competitive ratios of the described algorithms are linear in the density of
the matroid (or in the density of its simple version). In particular for matroids of
constant density, these algorithms are constant-competitive.
In the next subsections we describe two interesting classes of matroids having this
property: cographic matroids and small cocircuit matroids.
2.6.4 Cographic Matroids

The cographic matroid M∗ (G) of a graph G is the dual of its graphic matroid M(G).
The independent sets in M∗ (G) are those sets of edges that, when removed do not
increase the number of connected components of G. The bases in M∗ (G) are the
complements of the maximum forests of G. The circuits (minimal dependent sets) of
M∗ (G) are exactly the minimal edge-cuts of the connected components of G. This
means that the loops of M∗ (G) are the bridges of G.
A well-known result of graph-theory states that when G is 3-edge-connected then
we can find three spanning trees T1 , T2 and T3 , such that the union of their com-
plements covers E(G) (This follows from e.g. Edmonds’ Matroid Partitioning Theo-
rem [49]). In particular, we have the following.
Lemma 2.6.6. If every connected component of G = (V, E) is 3-edge-connected, then

γ(M∗ (G)) ≤ 3.
Proof. The above result implies that there are 3 forests F1 , F2 and F3 whose comple-
ments cover all the edges of G. Let Bi = E \ Fi . For every set of edges X ⊆ E,
3
X 3
X
|X| ≤ |X ∩ Bi | = rkM∗ (G) (X ∩ Bi ) ≤ 3 rkM∗ (G) (X),
i=1 i=1
where the middle equality follows since X ∩ Bi ⊆ Bi is independent in M∗ (G).
77
This result implies that Algorithm 10 is 3-competitive for cographic matroids of
graphs having only 3-edge-connected components. An alternative algorithm in the
same spirit is depicted as Algorithm 12.
Algorithm 12 for the cographic matroid of a graph G = (V, E) having only 3-edge-
connected components.
1: Find three forests F1 , F2 and F3 whose complement cover E.
2: Select i ∈ {1, 2, 3} uniformly at random and return the set Bi = E \ Fi .
Lemma 2.6.7. Algorithm 12 is 3-competitive for cographic matroids of graphs with

3-edge-connected components. This holds for any model of the matroid secretary prob-
lem.
Proof. This follows since every edge of G is selected with probability at least 1/3.
We can not extend the previous results to arbitrary graphs: the cographic matroid
of a bridgeless graph G can have arbitrarily high density. To see this consider the
cycle Cn on n vertices. For this graph
|E(Cn )|
= n. (2.32)
rkM∗ (Cn )
Nevertheless we can still show the following result.

Lemma 2.6.8. For any bridgeless graph G, the simple version of its cographic matroid
has density at most 3.
Proof. Let {C1 , . . . , Ck } be the collection of 2-edge-connected components5 of the
bridgeless graph G. Consider the graph H obtained by taking the disjoint union of
the graphs Ci (using copies of the vertices that are in two or more of the Ci ’s). The
graph H has the same graphic and cographic matroids as G, so we use H instead.
Let P1 , . . . , P` be the cographic parallel classes of M∗ (H). The simple version M0
of M∗ (H) is the matroid obtained from M∗ (H) by deleting (in a matroid-sense) all
but one element in each Pj . Since deleting an element of a matroid corresponds to
contracting the same element in its dual, we conclude that the matroid M0 is the
cographic matroid of the graph H 0 obtained by contracting (in the graph-sense) all
but one edge in each of the Pj .
Since M0 has no pair of parallel elements, the components of H 0 have no edge-
cut of size 2. Therefore, all the components of H 0 are 3-edge-connected. Using
Lemma 2.6.6, we conclude that M0 has density at most 3.
The result above implies that the following algorithm is 3e-competitive-algorithm
for cographic matroids of an arbitrary graph G: Remove the bridges of G and then
apply Algorithm 11. An alternative algorithm in the same spirit is depicted as Algo-
rithm 13.
5
A 2-edge-connected component is a maximal 2-edge-connected subgraph. The 2-edge-connected
components of a bridgeless graph provide a partition of the edges of the graph.
78
Algorithm 13 for the cographic matroid of a graph G = (V, E).
1: Remove the bridges of G.
2: Construct the associated graph H 0 described in the proof of Lemma 2.6.8.
3: Find three forests F1 , F2 and F3 whose complements cover H 0 .
4: Define the partition matroids Qi = Q(E(H 0 ) \ Fi ) having E(G) as ground set (see
Section 2.6.3).
5: Select i ∈ {1, 2, 3} uniformly and run any e-competitive algorithm for partition
matroids on Qi , returning its answer.
Lemma 2.6.9. Algorithm 13 is 3e-competitive for cographic matroids of general

graphs in the adversarial-assignment random-order model.
Proof. Analogous to the proof of Lemma 2.6.5.
2.6.5 Matroids with Small Cocircuits

For each element u of a loopless matroid M = (S, I), let c∗ (u) be the size of the
smallest cocircuit (i.e. circuits of the dual matroid) containing u, and let
c∗ (M) = max c∗ (u). (2.33)

u∈S
Consider the algorithm that greedily constructs an independent set of M selecting

elements as they appear without looking at their weights.
Theorem 2.6.10. The algorithm described above is c∗ (M)-competitive (in random-

order models).
Proof. To see this, fix an element u ∈ S and let C ∗ be a cocircuit of minimum

size containing it. If u appears before all the other elements of C ∗ in the random
order then it has to be selected by the algorithm. Otherwise, there would be a
circuit C that intersects C ∗ only in element u, which is a contradiction. (See, e.g.
[137, Proposition 2.1.11].) From here, we conclude that u is selected with probability
at least 1/c∗ (u) ≥ 1/c∗ (M).
We also have the following property.
Lemma 2.6.11. For every loopless matroid M = (S, I), γ(M) ≤ c∗ (M).
Proof. Let n be the size of S. An element u is selected by the algorithm above if and
only if u is in the lexicographic first base OPT(π) of the ordering π : [n] → S in which
the elements are presented. Consider the vector ρ ∈ RS having each coordinate equal
to 1/c∗ (M). Using the proof of Theorem 2.6.10, we conclude that
1 X
ρ(u) ≤ Prπ (u is selected by the algorithm) = χOPT(π) (u).
n! π
79
In particular, for every set U ⊆ S we have:
|U | 1 X 1 X
= ρ(U ) ≤ |U ∩ OPT(π)| ≤ rk(U ) = rk(U ),
c∗ (M) n! π n! π
where we used the fact that U ∩ OPT(π) is an independent subset of U .
In particular, the algorithm presented is no better than the Algorithm 10 for low-
density matroids at least in terms of its competitive ratio. Nevertheless, the algorithm
above is simpler and does need require to know the matroid beforehand.
2.7 Summary and Open Problems

In this chapter we have given new algorithms for the matroid secretary problem.
Some of these results have already been presented by the author in a recent SODA
paper [152].
Regarding the random-assignment random-order model, our main contribution is
Algorithm 6. This algorithm is the best constant-competitive algorithm known so far
for this model, achieving a competitive ratio of at most 5.7187. (See Theorem 2.5.4.)
A slightly weaker algorithm achieving a competitive ratio of 2e/(1 − 1/e) ≈ 8.60052
was already presented by the author in [152]. The existence of constant-competitive
algorithm solves an open question of Babaioff et al. [8] (see Conjecture 1.5.3).
On a very high level Algorithm 6 is an application of a simple divide and conquer
idea: replace the matroid by a collection of disjoint simpler matroids for which we
can easily derive a constant-competitive algorithm, and then return the union of the
answers. The proposed simpler matroids are the uniformly dense matroids.
In Section 2.3 we have studied some properties of the class of uniformly dense
matroids and shown how to modify existent algorithms for uniform matroids to work
on this class. The best algorithm presented, Algorithm 4 has competitive ratio of
at most 4.92078, improving on the 2e ≈ 5.43653 competitive algorithm previously
presented by the author [152] (see Algorithm 3).
In Section 2.4 we have studied the notion of principal partition of a matroid
and how to use it to construct a collection of uniformly dense minors. These ma-
troids are used for the development of Algorithm 6 in Section 2.5. Roughly speaking,
Lemma 2.4.16 states that every algorithm for uniformly dense matroids can be trans-
formed into an algorithm for general matroids, increasing an extra factor of at most
(e/(e − 1)) on its competitive ratio.
As first noticed by Oveis Gharan and Vondrák [136] after the first publication
of some of the results above, it is possible to apply our methods in the random-
assignment adversarial-order model. Oveis Gharan and Vondrák have devised a 40-
competitive algorithm for uniformly dense matroids on this stronger model, which by
our results, can be transformed in a 40/(1 − 1/e) ≈ 63.279-competitive algorithm for
general matroids.
In this thesis we have improved those algorithm showing a 16/(1 − 1/e) ≈ 25.311
for both uniformly dense matroids (Algorithm 5) and general matroids (Algorithm 7).
80
The last sections of this chapter deal with the “standard” matroid secretary prob-
lem: the arbitrary-assignment random-order model.
We have presented the first O(log r)-competitive comparison-based algorithm for
general matroids. The best previous result for this problem was an valued-based
algorithm of Babaioff et al. [8] achieving the same asymptotic competitive ratio.
Afterwards, we have given a ke-competitive algorithm for linear matroids rep-
resentable by k-column sparse matrices. This result contains as a special case the
2e-competitive for graphic matroids of Korula and Pál [97].
We have also given algorithms for general matroids having competitive ratio pro-
portional to the density of the matroid. Two special cases studied are the cographic
matroids, for which we give 3e-competitive algorithms, and matroids where each ele-
ment is in a cocircuit of size at most k, for which we give a k-competitive algorithm.
2.7.1 Open Problems

Besides the obvious problem of finding a constant-competitive algorithm for general
matroids in the adversarial-assignment random-order model there are several other
problems to work on. Here is a list of some of them.
Consider the following setting. A weighted matroid with arbitrary hidden weights
is presented to an algorithm. At every step, the algorithm can pick any element,
ask for its weight, and depending on the answer add it to the solution set or reject
it, provided that the solution set is always an independent set. Does this ability of
choosing the order help to attain a constant-competitive algorithm for adversarial-
value models of the matroid secretary problem?
Another interesting question is to determine if there is an approximate notion of
principal partition for domains extending matroids, in particular for matroid intersec-
tion domains. An affirmative answer to this question could allow us to use the divide
and conquer approach presented for matroids to give low competitive algorithms for
the random-assignment random-order model of the corresponding generalized secre-
tary problem.
Babaioff et al. [8] (see Lemma 1.4.1) have shown that for general domains, it
is impossible to achieve a competitive ratio of o(log n/ log log n). The proof of this
lemma shows that this is true even for cases where (i) each element receives their
weights from a known distribution, (ii) the algorithm is value-based and (iii) the
algorithm can choose the order in which the elements are revealed. Currently, there
are no known algorithms achieving an o(n)-competitive ratio for general domains even
if the three conditions above hold. It would be interesting to close the gap between
the upper and lower bounds on the competitive ratio for this problem.
Finally, it is open to determine if there are cases of the adversarial-assignment
model of the generalized secretary problem in which value-based algorithms out-
perform comparison-based ones. This seems to be the case since comparison-based
algorithms are not allowed to compare the weight of sets of elements, but currently
no examples are available.
81
82
Part II
Jump Number of Two Directional

Orthogonal Ray Graphs and
Independent Sets of Rectangles.
83
84
Chapter 3
Posets and Perfect Graphs
In this chapter, we define partially ordered sets (posets), comparability graphs and
related notions; and state some properties. We then give a survey on different classes
of comparability graphs, focusing on subclasses of two directional orthogonal ray
graphs and discuss some geometrical characterizations. Finally, we discuss perfect
graphs and recall some of their properties.
3.1 Basic Notions of Posets

We mostly follow the notation of Trotter [164].
A partial order relation over a set V is a reflexive, antisymmetric and transitive
binary relation. A pair P = (V, ≤P ) where ≤P is a partial order over V is called a
partially ordered set or poset.
Two elements u and v in V are comparable if u ≤P v or v ≤P u. Otherwise they
are incomparable. For u and v satisfying u ≤P v, the interval [u, v]P is the set of
elements x such that u ≤P x ≤P v. A linear order, also known as a total order is
a poset in which every pair of elements is comparable. The poset ([n], ≤) where ≤ is
the natural order relation is an example of total order.
An element v ∈ V covers u ∈ V if u <P v and there is no w ∈ V such that
u <P v <P w. Two important undirected graphs associated to the poset P are
defined as follows.
The graph GP having V as vertex set such that uv is an edge if u and v are
comparable and u 6= v is the comparability graph of the poset P . A graph G is a
comparability graph if there is a poset P such that G is isomorphic to GP .
The graph CP having V as vertex set such that uv is an edge if u covers v or if
v covers u is the covering graph of the poset P . A graph G is a covering graph if
there is a poset P such that G is isomorphic to CP .
Different posets can share the same comparability graph or the same covering
graph. Covering graphs are usually represented in the plane with a structured drawing
denoted as a Hasse diagram in such a way that if v covers u, then v is above u in
the plane. Since the problems considered in this part of the thesis deal with bipartite
graphs, the following remark is relevant.
85
Remark 3.1.1. Every bipartite graph is both a comparability and a covering graph.
Given two posets P = (V, ≤P ) and Q = (V, ≤Q ) on the same set V , the inter-
section P ∩ Q is defined as the poset (V, ) where u v if and only if u ≤P v and
u ≤Q v. The intersection of a family of posets is defined analogously.
Given two posets P = (U, ≤P ) and Q = (V, ≤Q ) on possibly different sets, the
product P × Q is defined as the poset (U × V, ), where (u, v) (u0 , v 0 ) if u ≤P u0
and v ≤Q v 0 . The product of a sequence of posets is defined analogously. For d ≥ 1,
we use P d to denote the iterative product of d copies of P .
A poset P = (V, ≤P ) can be embedded in a poset Q = (W, ≤P ) if there is a map
ϕ : V → W for which u ≤P v if and only if ϕ(u) ≤Q ϕ(v). Note that if P can be
embedded in Q and Q can be embedded in R, then the first poset can be embedded
in the third one by just composing the corresponding maps.
Unless specifically stated, we assume the set V to be finite and use n to denote its
cardinality. The only infinite posets we consider are the usual orders of the integers
(Z, ≤), the reals (R, ≤), and their finite powers (Zd , ≤Zd ) and (Rd , ≤Rd ). We reserve
the symbol ≤, without subindex, to denote the total order between numbers.
3.2 Chains and Antichains

Let P = (V, ≤P ) be a poset. A subset {u1 , u2 , . . . , uk } of V satisfying u1 ≤P u2 ≤P
· · · ≤P uk is called a chain. A set of elements that are mutually incomparable is
called an antichain. The height and width of P are the maximum sizes of a chain
and an antichain respectively. Two important results relating these concepts are the
following.
Theorem 3.2.1 (Mirsky’s antichain partitioning [116]). The height of a poset, that
is the size of a maximum chain, is equal to the minimum number of antichains into
which the poset can be partitioned.
Theorem 3.2.2 (Dilworth’s chain partitioning [43]). The width of a poset, that is the
size of a maximum antichain, is equal to the minimum number of chains into which
the poset can be partitioned.
Mirsky’s theorem is very simple to prove: for a given element u ∈ V , let h(u) be
its height, that is, the size of a maximum chain in P having u as its largest element.
The set of elements having the same height forms an antichain, and these antichains
partition P into a number of antichains equal to the poset height. This means that we
can efficiently find a maximum chain and a minimum antichain partition by solving
a longest path problem in the directed acyclic graph obtained from P by appending
one element u0 that is considered smaller than all the others. As finding longest
paths in directed acyclic graphs can be done in linear time (see, e.g. [38]), we have
the following result.
Lemma 3.2.3 (Algorithm for antichain partitioning). Let P = (V, ≤P ) be a poset and
let n and m be the number of vertices and edges of the comparability graph of P . We
can find a maximum chain and a minimum antichain partition of P in O(n+m)-time.
86
Dilworth’s theorem is equivalent to König’s theorem on bipartite graph matchings
and many other related theorems including Hall’s marriage theorem. An important
algorithmic consequence of this is that we can find a minimum partition of a poset into
chains by solving a maximum bipartite matching problem. The following reduction
can be found in [59].
Given a poset P = (V, ≤P ), construct a bipartite graph G having two copies of V as
bipartition, where (u, v) is an edge of G if u ≤P v. For any partition of P into chains,
we can construct a matching by including all the edges between pairs of elements
that are consecutive in some chain. The converse also hold: for any matching M we
can construct a collection of disjoint chains by including u and v in the same chain
whenever (u, v) ∈ M . Observe that the trivial partition of P into n = |V | singleton
chains corresponds to the empty matching, and that adding an edge to a matching
has the effect of merging two of the chains in the collection. The previous observation
implies that this construction maps chain partitions of size k into matchings of size
n − k; therefore, finding a minimum chain partition of P corresponds to finding a
maximum cardinality matching in G.
Computing maximum cardinality matchings in a bipartite graph G with n vertices
and m edges is a well-studied problem. The current best algorithms for this task√are
the deterministic algorithm of Hopcroft and Karp [86] that runs in time O(n + m n),
and the randomized algorithm of Mucha and Sankowski [120] that runs in time O(nω ),
where ω is the exponent for matrix multiplication. That is, ω is the smallest exponent
for which there is an algorithm running in time O(nω ) to multiply two n × n matrices.
Currently it is known that 2 ≤ ω ≤ 2.376 (the upper bound is due to the fast matrix
multiplication algorithm of Coppersmith and Winograd [37]). Using this, we have
Lemma 3.2.4 (Algorithms for chain partitioning). Let P = (V, ≤P ) be a poset and
let n and m be the number of vertices and edges of the comparability graph of P √
. We
can find a maximum antichain and a minimum chain partition of P in O(n + m n)-
time using the deterministic algorithm of Hopcroft and Karp or in O(nω )-time using
the randomized algorithm of Mucha and Sankowski.
3.3 Extensions and Poset Dimension

Chains and linear orders correspond to our intuition of “sorted lists” of elements.
We can represent the linear order L = {u1 , . . . , un }, {(ui , uj ) : 1 ≤ i ≤ j ≤ n} ,
compactly as L = (u1 , . . . , un ). Using this notation, we say that ui is the element in
the i-th position on the linear order L.
Any poset can be sorted into a linear order maintaining all preexistent relations.
We formalize this observation as follows. Given two partial orders P = (V, ≤P ) and
Q = (V, ≤Q ) on the same set V , we say that Q is an extension of P if for any two
elements u and v in V , u ≤P v implies that u ≤Q v. A linear extension is an
extension that is also a total order.
87
Theorem 3.3.1 (Szpilrajn [161]). The collection L of linear extensions of a poset P
T
is nonempty and P = L∈L L.
Theorem 3.3.1 implies that every poset is fully characterized by its family of
linear extensions. An interesting parameter of a poset is the minimum number of
linear extensions needed to characterize it.
A family {L1 , . . . , Ld } of linear orders in V is called a realizer of P if their
intersection is P . The dimension of a poset P is the smallest cardinality of a realizer
for P . A d-dimensional poset is a poset of dimension at most d.
Even though the above is the standard way to define it, the usual idea of dimen-
sion is something based on geometry, not combinatorics. The following alternative
definition is implicit in a book by Ore [132]. The product dimension of a poset P
is the smallest number of total orders C1 , . . . , Cd , with Ci = (Vi , ≤i ) such that P can
be embedded in (Rd , ≤Rd ).
Theorem 3.3.2 (Ore [132]). The dimension and the product dimension of a poset
are the same.
Proof. For completeness, we include a proof of this theorem. Let {L1 , . . . , Ld } be a

realizer of a d-dimensional partial order P = (V, ≤P ). Consider the map ϕ : V → Rd
where the i-th coordinate of ϕ(u) corresponds to the position of u in Li . By definition
of a realizer, we have u ≤P v if and only if ϕ(u) ≤Rd ϕ(v). This embedding shows
that the product dimension of P is at most its dimension.
For the other direction, let ϕ : V → Rd be an embedding of a poset P . Consider
the linear order Li = (V, ≤i ) obtained by setting u <i v if ϕ(u)i < ϕ(v)i or if ϕ(u)i =
ϕ(v)i and ϕ(u)j < ϕ(v)j , where j is the first index in [d] such that ϕ(u)i 6= ϕ(v)j . It
is easy to check that Li is a total order on V .
We prove now that {L1 , . . . , Ld } is a realizer for P . Consider two points u 6= v
such that u <i v for all i ∈ [d], then we have ϕ(u) ≤Rd ϕ(v) for otherwise there would
be a coordinate j such that ϕ(u)j > ϕ(v)j , contradicting that u <j v. Therefore, by
the definition of the embedding, u ≤P v.
Conversely, consider two points u 6= v such that ϕ(u) ≤Rd ϕ(v) and suppose
for contradiction that for some i, u >i v. By definition, this means that either
ϕ(u)i > ϕ(v)i or for some j 6= i, the strict inequality ϕ(u)j > ϕ(v)j holds. Both cases
contradict the hypothesis.
One-dimensional posets correspond simply to linear orders. Two-dimensional
posets are well understood and have many different characterizations. The following
one is worth mentioning.
Theorem 3.3.3 (Dushnik and Miller [47]). A poset P has dimension at most two if
and only if it admits a nonseparating linear extension. This is, a linear extension
L = (v1 , . . . , vn ) of P such that if i < j < k and vi ≤P vk then vj is comparable in P
to at least one of vi and vk .
Two-dimensional posets can be recognized in polynomial time (see Theorem 3.4.6).

However, no simple characterization of d-dimensional posets is known for d ≥ 3.
88
Yannakakis [168] has shown that deciding whether a poset has dimension d is NP-
complete, for d ≥ 3. Hegde and Jain [83] have shown that the problem of computing
the dimension of a poset is hard to approximate within a factor n0.5−ε unless NP =
ZPP.
The following theorem, often attributed to Galai [68] and Hiragushi [84], states
that the dimension of a poset only depends on its comparability graph.
Theorem 3.3.4 (Gallai [68]; Hiragushi [84]). Posets having the same comparability
graph have the same dimension.
Properties of posets depending only on their comparability graphs are called com-
parability invariants. In Chapter 4 we study another comparability invariant: the
jump number.
3.4 Survey on Comparability Graph Classes

To efficiently solve algorithmic problems on graphs, it is important to know the struc-
tural properties of the objects in consideration. Certain graph problems, such as com-
puting a clique of maximum size, are NP-hard for general graphs, but they become
polynomial time solvable when restricted to special classes, such as comparability
graphs. For two graphs classes G1 ⊂ G2 , if a problem is polynomially tractable for
the larger class it will remain easy for the smaller one, and if the problem is NP-hard
in the smaller class it will remain hard in the larger one. Assuming P 6= NP, it is
interesting to refine the threshold classes between P and NP for certain graph prob-
lems. Even for classes that are “easy” for a given problem, it is interesting to find
subclasses where faster algorithms can be provided.
Special graph classes are an interesting topic. For additional results and references
we refer the reader to the excellent survey of Brandstädt, Le and Spinrad [17].
In this section, we survey a particular collection of comparability graphs and their
associated posets which are of relevance for the problems we study in the next chapter.
For some of these classes we present different characterizations and give new proofs
of equivalence between some of them. We start by defining some basic concepts.
Given a collection of objects X for which intersection makes sense, the inter-
section graph I(X) is the graph having X as vertex set, and where two objects
are adjacent if their intersection is nonempty. Similarly, if inclusion makes sense, the
containment graph of X is the graph having X as vertex set and where two objects
are adjacent if one is included in the other.
Given a bipartite graph G = (A ∪ B, E), an ordering (a1 , . . . , as ) of A, and an
ordering (b1 , . . . , bt ) of B where {A, B} is a fixed bipartition of the vertices of G, the
associated biadjacency matrix is the s × t matrix M having Mi,j = 1 if ai bj is an
edge and Mi,j = 0 otherwise. For a matrix M and a collection of matrices N , we say
that M is N -free if M does not contain any matrix of N as an (ordered) submatrix.
If N consists of only one matrix N , we say that a matrix is N -free if it is N -free.
89
3.4.1 Geometric Representation of Posets in the Plane
A two-dimensional comparability graph (or 2D-graph for short) is the compa-
rability graph of a two-dimensional poset. Theorem 3.3.2 states that any such poset P
can be obtained from a subset of points V in R2 , where for two points u and v in V ,
u ≤P v if and only if ux ≤ vx and uy ≤ vy . For a given 2D-graph G, we denote any
collection of points V ⊆ R2 realizing it as above a 2D-representation of G.
In what follows we introduce bipartite versions of the previous concepts. The
following definitions are not standard, but they are deeply related to the class of two
directional orthogonal ray graphs we later introduce (see Section 3.4.4).
Given two multisets1 A and B in R2 , define the bipartite poset P (A, B) on the
disjoint union2 A t B, where a ≤P b if and only if a ∈ A, b ∈ B, ax ≤ bx and ay ≤ by .
The comparability graph of any such poset is denoted as a bicolored 2D-graph.
Any pair of multisets A, B ⊆ R2 realizing a bicolored 2D-graph G as above is called
a bicolored 2D-representation of G.
Let G be a (bicolored) 2D-graph with n vertices. A (bicolored) 2D-representation
of G having every point in the grid [n]2 = {1, . . . , n} × {1, . . . , n}, and where no
two points are in the same horizontal or vertical line is called a (bicolored) rook
representation of G. In particular, every rook representation of a 2D-graph can be
obtained from a collection of points {(i, π(i)) : i ∈ [n]} for some permutation π of [n].
We say in this case that the rook representation is induced by π. We remark that if
(A, B) is a bicolored rook representation of a graph then A and B are disjoint sets
(we do not consider them as multisets anymore).
Lemma 3.4.1. Any (bicolored) 2D-graph admits a (bicolored) rook representation.
Proof. By Theorem 3.3.2, any 2D-representation V ⊆ R2 defines a poset admitting

a realizer {L1 , L2 } of size 2. We recover a rook representation by mapping each
element v of this poset to the point having x-coordinate equal to the position of v in
L1 and y-coordinate equal to the position of v in L2 .
Consider now the bicolored case. The main difficulty is that A and B are not
necessarily disjoint multisets: two points of A t B can share the same position in the
plane. For a bicolored 2D-representation (A, B), consider the map ϕ : A t B → R3
obtained by appending to each point in A (resp. in B) a third z-coordinate equal to
some negative number (resp. positive number), in such a way that no two elements of
AtB have the same z-coordinate. The poset Q defined by u ≤Q v if ϕ(u) ≤R3 ϕ(v) is
an extension of P (A, B) satisfying a ≤P b if and only if a ≤Q b for all (a, b) ∈ A × B.
Using the same construction as in the proof of Theorem 3.3.2, we obtain a realizer
{L1 , L2 , L3 }, Li = (A t B, ≤i ) for Q. Consider the set A0 obtained by moving each
element a ∈ A to the point a0 in [n]2 having x-coordinate equal to the position of v in
L1 and y-coordinate equal to the position of v in L2 . Define B 0 from B analogously.
We claim that (A0 , B 0 ) is a bicolored rook representation of P (A, B).
1
For our purposes a multiset of points in the plane is a collection of objects where even though
two of them may share the same position in the plane, they are considered as different elements.
2
By A t B we mean the multiset containing a copy of A and a copy of B where elements repeated
in both are considered as different.
90
It is clear that in A0 ∪ B 0 no two points are in the same horizontal or vertical line.
We only need to check that a ≤R2 b if and only if a0 ≤R2 b0 , for a ∈ A, and b ∈ B.
One direction is easy: If a ≤R2 b then ϕ(a) ≤R3 ϕ(b) and so, a <i b for i = 1, 2, 3.
This means that a0 ≤R2 b0 . For the converse, suppose that a0 ≤R2 b0 . By construction,
a0 is located strictly to the left and strictly below b0 . This is, a <1 b and a <2 b. Since
we are using the same construction as in the proof of Theorem 3.3.2, the fact that
ϕ(a)3 < 0 < ϕ(b)3 implies that a <3 b also holds. From here, we get ϕ(a) ≤R3 ϕ(b),
implying that a ≤R2 b.
Using the previous lemma we can show the following.
Lemma 3.4.2. Bicolored 2D-graphs are comparability graphs of dimension at most 3.
Proof. Let (A, B) be a bicolored rook representation in [n]2 of a bicolored 2D-graph G.

Define ϕ : A ∪ B → R3 as

x , vy , −vx ) if v = (vx , vy ) ∈ A;
(v
ϕ(v) =
(vx , vy , n + 1 − vx ) if v = (vx , vy ) ∈ B.
Note that if v = (vx , vy ) and w = (wx , wy ) are two different elements in A ∪ B
then, by assumption, vx 6= wx . In particular, if both v and w are in the same color
class (A or B), then the first and third coordinates of (ϕ(v) − ϕ(w)) have different
sign. This means that ϕ(v) and ϕ(w) are incomparable in (R3 , ≤R3 ). On the other
hand, if a ∈ A and b ∈ B then the third coordinate of ϕ(a) is always negative, while
the third coordinate of ϕ(b) is always positive. Therefore, a ≤R2 b if and only if
ϕ(a) ≤R3 ϕ(b).
We have proved that ϕ is an embedding in R3 of a poset having comparability
graph G. This concludes the proof.
Now we are ready to describe the classes of graphs and posets that we focus on
this thesis. We start by defining permutation graphs and then we proceed to define
a chain of bipartite graph classes, each one contained in the previous one.
3.4.2 Permutation Graphs

Permutation graphs have many equivalent characterization. Their standard definition
is the following: A graph G = (V, E) is a permutation graph if there is a labeling
for V = {v1 , . . . , vn } and a permutation σ on [n] such that vi vj is an edge of G if and
only if (i − j)(σ −1 (i) − σ −1 (j)) < 0.
Two equivalent geometric definitions are the following.
Lemma 3.4.3 (Even et al. [52] and Baker er al. [9]). The following are equivalent:
(i) G is a permutation graph.
(ii) G is the intersection graph of a family {S1 , . . . , Sn } of line segments connecting

two parallel lines in the plane.
91
(iii) G is a 2D-graph.
Proof. We give the proofs for completeness.

(i) ⇔ (ii): Consider two parallel (say horizontal) lines in the plane. On the top line
select n points and label them from left to right as 1, 2, . . . , n. On the bottom line
select n points and label them from left to right as σ(1), . . . , σ(n), where σ is the
permutation defining G. Let Si be the line segment connecting the points labeled i
in both the top and bottom line. Then Si intersects Sj if and only if i < j and
σ −1 (i) > σ −1 (j), or if i > j and σ −1 (i) < σ −1 (j). Therefore, G is the intersection
graph of the segments {S1 , . . . , Sn }. Since this construction is reversible we obtain
the equivalence.
(i) ⇔ (iii): Consider a permutation graph given by some permutation σ. Let π be
the related permutation π(i) = n + 1 − σ −1 (i) and define the mapping ϕ : V → R2
given by ϕ(vi ) = (i, π(i)). From here, two elements vi and vj are adjacent in G if
and only if ϕ(vi ) and ϕ(vi ) are comparable by ≤R2 in the plane. Therefore, G is
the 2D-graph having rook representation induced by π. Since this construction is
reversible we obtain the equivalence.
Using Lemma 3.4.3 we can give an alternative geometric proof for yet another
characterization of permutation graphs.
Lemma 3.4.4 (Baker et al. [9]; Dushkin and Miller [47]). A graph G is a permutation
graph if and only if it is the containment graph of a collection of (distinct) closed
intervals.
Proof. Consider a graph G having rook representation induced by a permutation π

on [n]. Let vi be the vertex represented by (i, π(i)). Associate vi to the interval
Ii = [−π(i), i]. Noting that Ii ⊆ Ij if and only if (i, π(i)) ≤R2 (j, π(j)) we conclude
that G is the containment graph of the intervals {Ii : i ∈ [n]}. Conversely, given
the containment graph G of a family of intervals, we map each interval Ii = [ai , bi ]
to the point pi = (−bi , ai ) in the plane. As before, Ii ⊆ Ij if and only if pi ≤R2 pj .
Hence, the points pi are a 2D-representation of the graph G and therefore, G is a
2D-graph. Intuitively, the previous assignment gives a bijection between intervals and
points weakly above the diagonal line y = −x in such a way that their horizontal and
vertical projections onto this line define the corresponding intervals. We illustrate
this construction in Figure 3-1.
We conclude by mentioning a couple of properties of these graphs.
Theorem 3.4.5 (Dushnik and Miller [47]). A graph G is a permutation graph if and
only if both G and its complement G are comparability graphs.
Theorem 3.4.6 (McConell and Spinrad [115] ). Permutation graphs can be recognized
in time O(n + m) where n and m are the number of vertices and edges of the graph.
92
1 2 3 4 5
2 4
2 4 5 3 1
I5 1 3
I4
I3
I2 5
I1
0
Figure 3-1: Different representations of the permutation graph given by σ =

(2, 4, 5, 3, 1) or equivalently, by the rook representation induced by π = (1, 5, 2, 4, 3).
3.4.3 Chordal Bipartite Graphs

A chord of a cycle is an edge connecting two vertices that are not consecutive in the
cycle. A graph is chordal (or triangulated) if every cycle of length at least 4 has at
least one chord. We are interested in the following related class.
A chordal bipartite graph is a bipartite graph where each cycle of length at
least 6 has a chord. We must be careful as this definition can be misleading: chordal
bipartite graphs are not chordal in general, since the cycle on four vertices C4 is
chordal bipartite. In other words, chordal bipartite graphs are not the intersection of
bipartite graphs and chordal graphs.
Chordal bipartite graphs have many important properties. However, we only
mention some theorems that are relevant for our purposes.
Theorem 3.4.7 (Hoffman, Kolen and Sakarovitch [85]). A graph G"is chordal# bipar-
1 1
tite if and only if it admits a Γ-free biadjacency matrix, where Γ = .
1 0
Theorem 3.4.8 (e.g. Lubiw [108]). Chordal bipartite graphs can be recognized in
polynomial time.
3.4.4 Two Directional Orthogonal Ray Graphs (2DORGs)

An orthogonal ray graph is the intersection graph of rays (closed half-lines) in
the plane that are parallel to either the x or the y axis, provided we only consider
intersections between horizontal and vertical rays (i.e. if two horizontal rays intersect,
there is no associated edge in the graph).
A two directional orthogonal ray graph (or 2dorg) is an orthogonal ray
graph where all the horizontal rays go in the same direction (without loss of generality,
they go to the right) and all the vertical rays go in the same direction (without loss
of generality, they go down). Formally, a 2dorg is a bipartite graph on A ∪ B where
93
each vertex v is associated to a point (vx , vy ) ∈ R2 , so that a ∈ A and b ∈ B are
neighbors if and only the rays [ax , ∞)×{ay } and {bx }×(−∞, by ] intersect each other.
Lemma 3.4.9. Two directional orthogonal ray graphs are exactly the bicolored 2D-
graphs.
Proof. This follows directly from the fact that [ax , ∞)×{ay } intersects {bx }×(−∞, by ]
if and only if a ≤R2 b.
This class of graphs can also be characterized by their biadjacency matrices.
Lemma 3.4.10 (Shrestha, Tayu and Ueno [151]). The following are equivalent:
(i) G is a two directional orthogonal ray graph.
(" # " #)
1 1 0 1
(ii) G admits a biadjacency matrix that is γ-free, where γ = , .
1 0 1 0
(iii) G admits a biadjacency matrix such that if Mij is a 0-entry, then at least one
of the following holds: every entry above it is a 0-entry or every entry to its left
is a 0-entry.
Proof. It is easy to check the equivalence between the second and third conditions.
The proof of equivalence between the first and second ones we present is similar to
the one given in [151]. Let G be a 2dorg with rook representation (A, B). Sort
A from top to bottom as a1 , . . . , as and B from left to right as b1 , . . . , bt . We claim
this ordering induces a γ-free biadjacency matrix M . Indeed, suppose that i ≤ i0
and j ≤ j 0 are such that the submatrix M{i,i0 },{j,j 0 } is in γ. As Mi0 ,j = Mi,j 0 = 1, we
have (ai0 )x ≤ (bj )x ≤ (bj 0 )x and (ai0 )y ≤ (ai )y ≤ (bj 0 )y , contradicting the fact that
Mi0 ,j 0 = 0. Figure 3-2 illustrates this construction.
Conversely, let M be a γ-free s × t biadjacency matrix of G. For every i ∈ [s],
let L(i) be the leftmost index j in the matrix such that Mij = 1 (set L(i) = 0 if no
such entry exists). Similarly, for every j ∈ [t], let T (j) be the topmost index i in the
matrix such that Mij = 1 (set T (j) = 0 if no such entry exists). Map each vertex
ai ∈ A to the position (L(i), s − i + 1) and vertex bj ∈ B to position (j, s − T (j) + 1).
The corresponding pair (A0 , B 0 ) of obtained multisets in R2 is a 2D-representation
of G. Indeed, since M is γ-free, Mij = 1 if and only if L(i) ≤ j and T (j) ≤ i. This is
equivalent to (L(i), s − i + 1) ≤R2 (j, s − T (j) + 1).
b2
b1 b1 b2 b3 b4
a1 b4
 
a1 0 1 0 0
a2 b2 a2  1 1 0 1 
 
a3 a3 0 1 1 1
Figure 3-2: A 2dorg and its γ-free biadjacency matrix.
Using the previous lemma we can show the following.
94
Lemma 3.4.11. The class of two directional orthogonal ray graphs is strictly con-
tained in the class of chordal bipartite graphs.
Proof. The inclusion follows since γ-free matrices are also Γ-free. To see that the
inclusion is strict, consider the 3-claw G in Figure 3-3.
b1
b1 a1 b4
b2
a0 b2 a2 b5
b3
b3 a3 b6
a0
Figure 3-3: Chordal bipartite graph that is not a 2dorg and a sketch for the impos-
sibility of a bicolored rook representation.
This graph is obtained by replacing each edge in a star with three edges (K1,3 ) by
a path of length 3. Since G has no cycles, it is chordal bipartite. Shrestha et al. [150]
have shown that G is not a 2dorg by using more involved characterizations of these
graphs. Here we give a direct proof. Let a0 be the root, {b1 , b2 , b3 }, {a1 , a2 , a3 } and
{b4 , b5 , b6 } be the vertices at distance 1, 2 and 3 from the root and assume that G
admits a bicolored rook representation (A, B). In this representation, a0 divides the
plane in four quadrants. The points b1 , b2 , b3 must be in the quadrant above and to
the right of a0 . Furthermore, they must form an antichain for ≤R2 since, if for two
of them we have bi ≤R2 bj , the fact that ai ≤R2 bi would contradict that ai bj is not
an edge of G.
Without loss of generality, assume that b1 , b2 , b3 are the elements of this antichain
from left to right as in Figure 3-3. Since a2 is connected to b2 , but not to b1 or b3 , the
point a2 must be in the gray rectangular area shown in the figure, located between
the vertical line passing through b1 , the horizontal line passing through b3 and the
vertex b2 . But then, since a2 is above and to the right of a0 , so must be b5 contradicting
the fact that a0 and b5 are incomparable.
Finally, we mention the following result.
Theorem 3.4.12 (Shrestha, Tayu and Ueno [151]). Two directional orthogonal ray
graphs can be recognized in polynomial time.
3.4.5 Interval Bigraphs

An interval graph is the intersection graph of a collection of intervals {I1 , . . . , In }
of the real line. An interesting bipartite variant of these graphs is the following.
An interval bigraph is a bipartite graph G = (A ∪ B, E), where each vertex
v ∈ A ∪ B is associated to a real interval Iv so that a ∈ A and b ∈ B are adjacent if
and only if Ia ∩ Ib 6= ∅.
95
Since we are only considering graphs with finite cardinality, it is very simple to
prove that every interval graph (and every interval bigraph) can be represented using
only closed intervals having extreme points in the set {1, 2, . . . , 2n}, where n is the
number of vertices of the graph.
Lemma 3.4.13. The class of interval bigraphs is strictly contained in the class two
directional orthogonal ray graphs.
Proof. We give a simple proof of the inclusion using the bicolored 2D-representation
of two directional orthogonal ray graphs. Let G be an interval bigraph with parts
A and B. For a ∈ A with interval Ia = [s, t] and b ∈ B with interval Ib = [s0 , t0 ],
we identify a with the point (s, −t) ∈ Z2 and b with the point (t0 , −s0 ) ∈ Z2 . By
definition, ab is an edge of G if and only if [s, t] ∩ [s0 , t0 ] 6= ∅, or equivalently if
(s, −t) ≤Z2 (t0 , −s0 ). Intuitively, the previous identification maps A (resp. B) to
points weakly below (resp. weakly above) the diagonal line y = −x in such a way
that their horizontal and vertical projections onto this line define the corresponding
intervals. We illustrate this construction in Figure 3-4. We call this the natural
bicolored 2D-representation of an interval bigraph.
(10, 0)
[0, 10] [2, 7] [5, 7] (7, −2)
b1 b1 b1
0 10
(0, −4) (7, −5)
(2, −6)
a1 a2 a3 a4
[0, 4] [2, 6] [4, 10] [8, 10] (4, −10)
(8, −10)
Figure 3-4: An interval bigraph, its interval representation and a geometric represen-
tation as 2dorg.
We have shown that every interval bigraph is a 2dorg. To see that the inclusion
is strict, consider the graph G in Figure 3-5.
b1
b01 a1 b1 a01 b2
a01
b3
a02
b01
b02 a2 b2 a02 a03
b02
a1
b03
a2
b03 a3 b3 a03
a3
Figure 3-5: A two directional orthogonal ray graph that is not an interval bigraph
and one bicolored rook representation.
Müller [123] calls G an “insect” and gives a proof that this graph is not an interval
bigraph. For completeness, we include a slightly different proof here.
96
Suppose that G admits an interval representation as an interval bigraph. Consider
its natural bicolored 2D-representation (A, B). The set B0 = {b1 , b2 , b3 } must be an
antichain for ≤R2 since if for two of them, bi ≤R2 bj then the fact that a0i ≤R2 bi
would contradict that a0i bj is not an edge. By symmetry, the set A0 = {a1 , a2 , a3 } is
also an antichain. By possibly relabeling the vertices, we may assume that b2 and a2
are the middle points (from left to right) of their respective antichains. Let p be the
bottommost and leftmost point in the plane that is larger than all elements of A0 in
the ≤R2 order and q be the topmost and rightmost point in the plane that is smaller
than all elements of B0 in the ≤R2 order. Since every ai bj is an edge, we must have
p ≤R2 q. On the other hand, since a02 is only comparable to the middle point of B0 ,
q ≤R2 a02 . Symmetrically, b02 ≤R2 p. This implies that b02 ≤R2 a02 . Since the points in
A are weakly below the diagonal and the points in B are weakly above it, we must
have a02 = b02 , which is a contradiction since a02 b02 is not an edge.
We conclude, by mentioning the following results (See Figure 3-6)
Theorem 3.4.14 (Das et al. [41]). A graph G is an interval bigraph if it admits a
biadjacency matrix M with the following zero-partition property: Every 0-entry
in M can be labeled as R or C in such a way that every entry above a C is also a C
and every entry to the left of an R is also an R.
b1 b2 b3 b4 b5 b6

C C
 b1
a1 1 1 1 0 1 0 b2 b5
0R 1 1 1 1 0C b3
 
a2   a1
a2 b4
0C
 
a3 
 1 1 1 1 1 
 a3 a4

a4  0R 0R 1 1 1 0C 
 b6
a5 0R 0R 1 1 1 1 a5
Figure 3-6: A biadjacency matrix with the zero partition property and a bicolored
2D-representation of the same interval graph.
Theorem 3.4.15 (Müller [123]). Interval bigraphs can be recognized in polynomial

time.
3.4.6 Convex Graphs

A bipartite graph with bipartition {A, B} is convex on A if there is a labeling for
A = {a1 , . . . , as } so that the neighborhood of each b ∈ B is a set of elements of A
consecutive in this labeling. A graph is convex (also known as convex bipartite) if it
admits a partition {A, B} of its vertices such that G is bipartite with parts A and B
and convex on one of its parts.
Another, perhaps more usual, characterization of convex graphs is the following.
Given the graph of a finite path P , define A to be the set of edges of the path and B
to be a collection of subpaths of P . It is easy to verify that the intersection graph of
A ∪ B is a convex graph and that every convex graph can be obtained in this way.
97
Lemma 3.4.16. The class of convex graphs is strictly contained in the class of in-
terval bigraphs.
Proof. Consider a convex graph G = (A ∪ B, E) such that G is convex on A under
the labeling A = {a1 , . . . , as }. Map each element ai ∈ A to the (singleton) interval
{i} ⊆ R, and each element b ∈ B having neighborhood {a`(b) , . . . , ar(b) } to the interval
[`(b), r(b)]. Then, G is the interval bigraph associated to the intervals defined in this
way. This proves that every convex graph is an interval bigraph.
To prove that the inclusion is strict, consider the interval bigraph induced by two
equal families of intervals A = B = {{1}; {2}; {3}; [1, 2]; [2, 3]; [1, 3]}. This graph is
depicted in Figure 3-7.
u v w0
u0 v0
r0 s0 t0 w0
r0 s0 t0
r s t
w r s t
u v
u0 v0 w
Figure 3-7: Interval bigraph that is not a convex graph.
We claim that this graph is not convex on A. By symmetry, this implies that G
is not convex on B either. Thus, G is not a convex graph because {A, B} is the only
partition of the graph that makes G bipartite. To prove the claim, denote r = {1},
s = {2}, t = {3}, u = [1, 2], v = [2, 3] and w = [1, 3] to the vertices of A and
r0 , s0 , t0 , u0 , v 0 , w0 the corresponding vertices of B. Assume, for sake of contradiction,
that there is a labeling for A that witness convexity.
The vertex v 0 is adjacent to every vertex of A except for r, similarly the vertex u0
is adjacent to every vertex except for t. This implies that the first and last vertices of
the labeling must be r and t. Assume without loss of generality then that a1 = r and
a6 = t. Since the neighborhood of r0 is {r, u, w}, these three vertices should appear
consecutively in the labeling. Thus, w has label a2 or a3 . On the other hand, the
neighborhood of t0 is {t, v, w}, implying that w must have label a4 or a5 , which is a
contradiction.
By using the natural bicolored 2D-representation for interval bigraphs, we can
represent geometrically a convex graph G = (A ∪ B, E) in the triangular region
{(i, −j) ∈ Z2 : 1 ≤ j ≤ i ≤ |A|} with the points of A lying on the line y = −x.
It is very simple to characterize convex graphs by their biadjacency matrices. A
0-1 matrix is vertically (resp. horizontally) convex if the ones in every column
(resp. in every row) are consecutive.
Theorem 3.4.17. A bipartite graph G = (A ∪ B, E) is convex on A if it admits a
biadjacency matrix that is vertically convex.
98
Proof. Direct from the definitions.
We also have the following result.
Theorem 3.4.18 (Booth and Lueker [14]). Convex bigraphs can be recognized in
polynomial time.
3.4.7 Biconvex Graphs

A biconvex graph is a bipartite graph with bipartition {A, B} that is convex on
both A and B. We say that the graph is biconvex on A and B if this happens.
Lemma 3.4.19. The class of biconvex graphs is strictly contained in the class of
convex graphs.
Proof. The inclusion holds trivially by definition. To prove that the inclusion is strict,
we show that a 2-claw, that is, the tree obtained by replacing each edge in a star with
three edges K1,3 by a path of length 2 is convex but not biconvex. This graph is
depicted in Figure 3-8.
a1 b1
b0 b1 b2 b3
 
a1 1 1 0 0
b0 a2 b2 MG =  1 0 1 0 
 
a2
a3 1 0 0 1
a3 b3
Figure 3-8: Convex graph that is not biconvex, and one of its convex biadjacency
matrices.
Let b0 be the root, {a1 , a2 , a3 } and {b1 , b2 , b3 } be the vertices at distance 1 and 2
from the root. The only bipartition of this graph is A = {a1 , a2 , a3 } and B =
{b0 , b1 , b2 , b3 }. Since the neighborhood of every vertex of B consists of either one
vertex or all the vertices in A, any labeling of A witness convexity on this set. However,
there is no way to order B so that the neighbors of all the vertices of A are consecutive:
we would require b0 to be consecutive to all b1 , b2 and b3 which is impossible.
We can characterize biconvex graphs by their biadjacency matrices. A matrix is

biconvex if it is both horizontally and vertically convex.
Theorem 3.4.20. A bipartite graph G is biconvex if it admits a biconvex biadjacency

matrix.
Finally, we remark that we can recognize biconvex graphs in polynomial time

using the algorithms for convex graph recognition.
99
3.4.8 Bipartite Permutation Graphs
A graph G is a bipartite permutation graph if it is simultaneously bipartite and a
permutation graph. In particular, G admits a bicolored rook representation (A, B)
where both A and B are antichains for ≤R2 .
We can characterize bipartite permutation graphs by their biadjacency matrices.
Theorem 3.4.21 (Spinrad, Brandstädt and Steward [154]; Chen and Yesha [30]).
A bipartite graph G without isolated vertices is a bipartite permutation graph if it
admits a biadjacency matrix M in 2-staircase normal form. This is, M satisfies
the following properties.
1. The matrix M is biconvex.
2. Let [`(i), r(i)] be the (horizontal) interval of indices j for which Mij = 1. Then
`(1) ≤ `(2) ≤ · · · ≤ `(|A|) and r(1) ≤ r(2) ≤ · · · ≤ r(|A|).
(" # " # " #)
0 1 1 1 0 1
Equivalently, M is β-free, where β = , , .
1 1 1 0 1 0
Proof. Consider a rook representation of a bipartite permutation graphs and let A =

{a1 , . . . , as } and B = {b1 , . . . , bt } be the orderings obtained on A and B by sorting
them in increasing order of x-coordinates. Since A and B are antichains with respect
to ≤R2 , in the above orderings the elements are also sorted in decreasing order of their
y-coordinates. We claim these orderings induce a biadjacency matrix M in 2-staircase
normal form.
To prove the biconvexity of M , consider any element ai ∈ A with degree at least
two, and let bj and bj 0 with j ≤ j 0 be two of its neighbors. Suppose that bk ∈ B is
an element with j < k < j 0 . This implies that (ai )x < (bj )x < (bk )x < (bj 0 )x and
(bj )y > (bk )y > (bj 0 )y > (ai )y , and so ai bk must be an edge of the graph. Therefore,
the matrix M is horizontally convex. The proof of vertical convexity is analogous.
Let ai and ai0 be two elements of A with i < i0 . We show that `(i) ≤ `(i0 ).
Indeed, suppose for contradiction that j > j 0 , where j = `(i) and j 0 = `(i0 ), then
we have (ai )y < (bj )y < (bj 0 )y . Since i < i0 , we also have (ai )x < (ai0 )x < (bj 0 )x .
Therefore ai ≤Z2 bj 0 , contradicting the fact that j was the first neighbor of ai . The
proof that r(i) ≤ r(i0 ) is analogous. The fact that β-free matrices are exactly the
ones in 2-staircase normal form follows easily from the definitions.
For an example of a 2-staircase normal form matrix, see Figure 3-9.
Lemma 3.4.22. The class of bipartite permutation graphs is strictly contained in the
class of biconvex graphs.
Proof. The inclusion follows from Theorem 3.4.21 and the fact that both bipartite
permutation graphs and biconvex graphs are closed under adding isolated vertices.
To see that the inclusion is strict consider the graph G in Figure 3-10. The
biconvex matrix MG witness the biconvexity of G.
100
b1
b2 
b1 b2 b3 b4 b5 b6

b3 a1 1 1 0 0 0 0
a1 b4 1 1 1 1 0 0
 
a2  
 
a2 b5
MG = a3 
 0 0 1 1 0 0 

a3 
a4  0 0 1 1 1 0 
b6

a4 a5 0 0 0 1 1 1
a5
Figure 3-9: A rook representation and a 2-staircase normal form matrix associated
to a bipartite permutation graph G.
a1 b2 a3
b0 b1 b2 b3
 
a1 0 1 1 0
b1 a2 b3 MG =  1 1 1 1 
 
a2
a3 0 0 1 1
b0
Figure 3-10: Biconvex graph that is not a permutation graph and one of its biadja-
cency matrices.
Suppose for contradiction that G is a bipartite permutation graph. Then it would

admit a biadjacency matrix in 2-staircase normal form. But it is easy to see that (up
to relabeling of the vertices), the only biconvex biadjacency matrices of G are MG
and the ones obtained by taking the horizontal reflection, vertical reflection or both.
None of them is in 2-staircase normal form.
Finally, we remark that we can recognize bipartite permutation graphs in polyno-
mial time using the same algorithms used for permutation graphs.
3.4.9 Summary
Figure 3-11 depicts all the strict containments among the presented comparability
graph classes. In Table 3.1 we summarize some of the properties of their bicolored
2D-representations and admitted biadjacency matrices.
101
Table 3.1: Geometric and biadjacency properties of comparability graph classes.
Class Bicolored 2D-representation Biadjacency Matrix

Chordal Bipartite None in general. Γ-free.
2dorgs General. γ-free.
Interval bigraphs A : weakly below y = −x, Zero-partition property.
B : weakly above y = −x.
Convex graphs A : on y = −x, Vertically convex.
B : weakly above y = −x.
Biconvex graphs – Biconvex.
Bip. permutation graphs A and B antichains. β-free.
Chordal Bipartite Graphs

Two Directional Orthogonal Ray Graphs
Interval Bigraphs
Convex Graphs
Biconvex Graphs
Bipartite Permutation Graphs
Permutation Graphs
Figure 3-11: Summary of presented comparability graph classes.
3.5 Perfect Graphs

We now recall some important notions about graphs. A clique is a subset of the
vertices of a graph that are mutually adjacent to one another. A stable set is the
graph complement of a clique, this is, a subset of vertices that are mutually non-
adjacent. A clique cover of a graph G is a collection of cliques containing every
vertex of G. A stable set cover of G is a collection of stable sets containing every
vertex of G. A clique partition is a clique cover where all the cliques are mutually
disjoint. Similarly, a stable set partition is a stable set cover where all the stable
sets are mutually disjoint. Since cliques are closed under inclusion, any clique cover
can be transformed into a clique partition of the same cardinality by possibly dropping
some vertices of some cliques in the cover. The same property holds for stable set
covers and stable set partitions.
A proper coloring of a graph G is an assignment of labels, traditionally called
colors, to the vertices of G in such a way that no two adjacent vertices share the
same color. Note that color classes, this is maximal sets of vertices sharing the same
color, are stable sets. Therefore, the collection of color classes of a proper coloring
defines naturally a stable set partition.
The clique number of G, denoted ω(G), is the size of a largest clique in G. The
stability number of G, denoted α(G), is the size of a largest stable set in G. The
102
clique partition number of G, denoted κ(G), is the size of a smallest clique cover
of G. The coloring number of G, denoted χ(G), is the minimum number of colors
needed in a proper coloring of G.
Since cliques in a graph G correspond exactly to stable sets in the graph comple-
ment G, it follows easily that
ω(G) = α(G), α(G) = ω(G), κ(G) = χ(G) and χ(G) = κ(G). (3.1)
Also, since cliques and stable sets intersects in at most one vertex, we easily conclude
that
ω(G) ≤ χ(G) and α(G) ≤ κ(G). (3.2)
An important question in graph theory is to identify classes of graphs for which
equality holds above. Note that by taking the disjoint union of any graph with a very
large clique and a very large stable set we obtain a graph where equality holds for
both inequalities in (3.2). This (pathological) example shows that it might be difficult
to find conditions to characterize the cases where equality holds. This example also
partially explains the study of perfect graphs, which are defined below.
A (vertex) induced subgraph of a graph is a subgraph preserving all the
adjacency relations of the subset of the vertices in which it is defined. Formally
H = (W, F ) is an induced subgraph of G = (V, E) if W is a subset of V and for any
two vertices u, v ∈ W , u, v ∈ F if and only if u, v ∈ E.
A graph G is ω-perfect if for every nonempty induced subgraph H of G, ω(H) =
χ(H). A graph G is α-perfect if for every nonempty induced subgraph H of G,
α(H) = κ(H). A graph is perfect if it is both α-perfect and ω-perfect.
The Weak Perfect Graph Theorem [103] states that a graph is α-perfect if and
only if its complement is α-perfect. By using the complementary relations in (3.1),
we obtain that the three notions of perfectness (perfect, α-perfect and ω-perfect) are
equivalent.
Cycles of odd length are not perfect since their largest clique has size 2 whereas
their chromatic number is 3. By the perfect graph theorem, their complements are
also imperfect graphs. However, any proper induced subgraph of an odd cycle or
of its complement is perfect: they are minimally imperfect. In the nineteen sixties,
Berge [13] conjectured that these two graph classes are the only families of minimally
imperfect graphs. This conjecture attracted much attention during the next forty
years. It was proved in 2002 by Chudnovsky, Robertson, Seymour and Thomas [32]
and it is now known as the Strong Perfect Graph Theorem.
We can restate this theorem as follows. A hole of a graph is an induced chordless
cycle of length at least 4. An antihole is the graph complement of a hole. A hole
(or an antihole) is odd if it contains an odd number of vertices. A graph is a Berge
graph if it contains no odd hole and no odd antihole.
Theorem 3.5.1 (Strong Perfect Graph Theorem [32]). A graph is perfect if and only
if it is a Berge graph.
Some important classes of perfect graphs are the following.
103
• Bipartite graphs and their complements.
• Line graphs of bipartite graphs and their complements.
• Comparability graphs and their complements (co-comparability graphs).
• Interval graphs and their complements.
• Chordal graphs (graphs having no hole of length at least 4).
• Weak chordal graphs (graphs having no hole or antihole of length at least 5).
Perfect graphs can also be characterized polyhedrally. Given a graph G = (V, E),
the stable set polytope STAB(G) and the clique-constrained stable set poly-
tope QSTAB(G) are defined as
STAB(G) = convex-hull({χS ∈ RV : S stable set of G}), and (3.3)

QSTAB(G) = x ∈ RV : x ≥ 0,
X
xv ≤ 1, for all K maximal clique of G . (3.4)
v∈K
The vectors in QSTAB(G) are known as fractional stable sets. Since for every
stable set S, χS ∈ QSTAB(G), we conclude that STAB(G) ⊆ QSTAB(G). However,
the polytope QSTAB(G) might have non-integral vertices associated to fractional
stable sets. Chvátal [33], using results from Lovász [103, 102] characterizes perfect
graphs in terms of these two polytopes.
Theorem 3.5.2 (Chvátal [33]). QSTAB(G) is integral if and only if G is a perfect

graph and in this case QSTAB(G) = STAB(G).
We can find maximum cliques, maximum independent sets, minimum clique parti-
tion and minimum colorings of perfect graph in polynomial time (see e.g. [76, Chapter
9]). For special classes faster algorithms are available. For example, for comparabil-
ity graphs, the four mentioned objects are in correspondence with maximum chains,
maximum antichains, minimum chain partitions and minimum antichain partitions
respectively; therefore, we can compute them using the algorithms described in Lem-
mas 3.2.4 and 3.2.3.
104
Chapter 4
A Primer on the Jump Number

Problem
In this chapter, we define and motivate the jump number problem of a comparability
graph and survey previous results. We also define two related problems, the maximum
cross-free matching and the minimum biclique cover problems. The former problem
is known to be equivalent to the jump number problem for chordal bipartite graphs.
Later, we define and review some apparently unrelated problems on 0-1 matrices,
on planar geometry and on interval combinatorics. Most of the defined problems
come in pairs: a minimization and a maximization problem. For some of these pairs,
a min-max relation is known.
We show that all these problems are either equivalent or special cases of the
maximum cross-free matching and the minimum biclique cover problems. Some of
the connections presented in this chapter are very simple, yet they have not been
explicitly noticed in the literature before our joint work with Telha [153] and this
thesis.
As simple corollaries of the presented equivalences we obtain new min-max re-
lations for the maximum cross-free matching and the minimum biclique cover on
biconvex and convex graphs, and an O(n2 )-time algorithm for the jump number prob-
lem on convex graphs, considerably improving the previous O(n9 )-time algorithm of
Dahlhaus [40].
4.1 Jump Number

Let us revisit for a minute both Dilworth’s chain partitioning theorem (Theorem 3.2.2)
and Szpilrajn’s linear extension theorem (Theorem 3.3.1).
Consider a family C of disjoint chains partitioning P and having minimum cardi-
nality. It is not always the case that we can sort the chains in C = {C1 , . . . , Ck } in
such a way that the obtained sequence induces a linear extension of P . On the other
hand, even though every linear extension L of P can be naturally partitioned into
chains by “breaking” the sequence induced by L whenever an element and the next
one are incomparable in P , this partition is not necessarily of minimum size. As an
105
Figure 4-1: Hasse diagram of K2,2
example, consider a poset having the complete bipartite graph K2,2 as comparability
graph (see Figure 4-1 for its Hasse diagram). This poset can be partitioned into 2
chains, but every linear extension induces a partition into 3 chains.
Suppose that we want to find a linear extension of P for which the natural parti-
tion described above yields the minimum number of chains possible. This problem,
formally defined later in this section, conciliates in a way, both Dilworth’s and Szpil-
rajn’s problems.
Let us motivate this problem a little more. We can reinterpret it as a scheduling
problem. This interpretation is often attributed to an unpublished manuscript of
Pulleyblank [140]. We want to schedule a collection of jobs V on a single machine,
respecting the precedences induced by a poset P = (V, ≤P ). These precedence con-
straints model the fact that certain jobs can only be processed after a collection of
jobs have finished. Moreover, every time a job is scheduled immediately after a task
not constrained to precede it we incur in a unit “setup” cost. We want to find a
schedule that minimizes the total setup cost. Let us define this problem formally
using poset terminology.
Let P = (V, ≤P ) be a poset and L = (u1 , . . . , un ) be a linear extension of P . The
pair (ui , ui+1 ), where 1 ≤ i ≤ n − 1 is a jump (or setup) of P in L if ui 6≤P ui+1 .
Otherwise, it is said to be a bump of P in L.
We use J(P, L) and B(P, L) to denote the number of jumps and bumps of P in L
respectively. Obviously, J(P, L) + B(P, L) = n − 1 where n is the cardinality of V .
The jump number j(P ) of a poset P is defined as
j(P ) = min{J(P, L) : L linear extension of P }. (4.1)
Following the notation of Chaty and Chein [27] we also define the step number1
s(P ) as
s(P ) = max{B(P, L) : L linear extension of P }. (4.2)
From the relation above, we also have that for every poset of cardinality n,
j(P ) + s(P ) = n − 1. (4.3)
The jump number problem is defined as follows.
Problem 1 (Jump Number Problem). Given a finite poset P , find j(P ) and a linear
1
While it would be more natural to use the notation “bump number”, traditionally this term is
reserved to denote the minimum number of bumps in a linear extension of the poset. We use step
number to avoid confusion.
106
extension L of P such that j(P ) = J(P, L).
As commented at the beginning of this section, every linear extension L of P
induces naturally a partition of V into 1 + J(P, L) chains separated by bumps. More-
over, each of these chains is convex, that is, they are also intervals: indeed, if an
element v is such that u <P v <P w for u and w in some chain C of the partition,
then v can not appear before C or after C in the linear order, implying that v must
be in C. A natural question is then: What are necessary and sufficient conditions for
a partition of a poset into convex chains to induce a linear extension?
To answer this question, denote any partition C of a poset P into convex chains
as a chaining. A chaining is proper if there is a linear extension L such that the
elements of each chain are consecutive in L. Since each chain C has |C| − 1 bumps,
it is immediate that
X
B(P, L) = (|C| − 1) (4.4)
C∈C
and so X
s(P ) = min (|C| − 1). (4.5)
C proper chaining of P
C∈C
Theorem 4.1.1 below characterizes the proper chainings of a poset in terms of

alternating cycles. A 2k-alternating cycle (a1 , b1 , a2 , b2 , . . . , ak , bk ) of a set of chains
{C1 , C2 , . . . , Ck } is a collection of elements ai , bi ∈ Ci such that the only comparisons
under ≤P in {a1 , . . . , ak , b1 , . . . , bk } are
ai ≤P bi , i = 1, . . . , k. (4.6)
bi ≥P a(i+1) (mod k) , i = 1, . . . , k. (4.7)
or compactly
a1 ≤P b1 ≥P a2 ≤P b2 ≥P · · · ≤P bk−1 ≥P ak ≤P bk ≥P a1 . (4.8)
In particular note that all the points a1 , . . . , ak , b1 , . . . , bk are distinct.

A chaining C of a poset P is cycle-free if it does not contain a subset of chains
C1 , . . . , Ck and points ai , bi ∈ Ci such that (a1 , b1 , a2 , b2 , . . . , ak , bk ) is a 2k-alternating
cycle. The following theorem is already part of the folklore of the field, and can be
obtained as a consequence of the work of Duffus et al. [45].
Theorem 4.1.1 (Duffus et al. [45]). A chaining C of a poset P is proper if and only
if it is cycle-free.
The previous theorem, together with (4.3) and (4.5), implies that the jump number
problem is equivalent to the problem of finding a cycle-free chaining of maximum
cardinality.
This observation has interesting corollaries for bipartite posets. A poset is
bipartite if its comparability graph is bipartite or equivalently, when every chain con-
tains at most two elements. In particular, for a bipartite poset P with comparability
graph GP = (A ∪ B, E):
107
1. Every nonempty chain in P is convex and corresponds to either a singleton or
an edge in GP .
2. Chainings in P are in bijection with matchings in GP .
3. Proper chainings in P correspond to matchings M in GP such that there are no
edges e1 , . . . , ek in M and f1 , . . . , fk in E \ M where e1 f1 e2 f2 · · · ek fk is a cycle
in GP . This is, there are no M -alternating cycles in GP .
The last consequence implies the following theorem:

Theorem 4.1.2 (Chaty and Chein [27]). Solving the jump number problem in a bipar-
tite poset is equivalent to finding an alternating-cycle-free matching M of maximum
size in the corresponding comparability graph.
As a corollary, we obtain that the jump number of bipartite posets only depends
on the comparability graph of the poset. Habib [82] has shown that, in fact, this
holds for every poset: the jump number is a comparability invariant.
Theorem 4.1.3 (Habib [82]). Posets having the same comparability graph have the
same jump number.
The previous theorem states that the jump number j(G) of a comparability
graph G is a well-defined parameter.
4.1.1 Complexity of the Jump Number Problem

The literature concerning the jump number is vast. In what follows we mention some
of the most relevant results regarding the complexity of this parameter.
Pulleyblank [139] has shown that determining the jump number of a general poset
is NP-hard. On the other hand, there are efficient algorithms for many interesting
classes of posets. Chein and Martin [29] give a simple algorithm to determine the
jump number in forests. Habib [81] presents new algorithms for forests, for graphs
having a unique cycle, and for graphs having only disjoint cycles. Cogis and Habib [35]
introduce the notion of greedy linear extension. They prove that every such exten-
sion is optimal for the jump number in series-parallel posets, obtaining an algorithm
for this poset class. Rival [144] extends this result to N -free posets. Later, Gierz
and Poguntke [69] show that the jump number is polynomial for cycle-series-parallel
posets.
Duffus et al. [45] have shown that for alternating-cycle-free posets, the jump num-
ber of a poset is always one unit smaller than its width (see Theorem 4.1.1). Chein
and Habib [28] give algorithms for posets of width two and later Colbourn and Pul-
leyblank [36] provide a dynamic program algorithm for posets of bounded width.
Steiner [155] also gives polynomial time algorithms for posets admitting a particular
bounded decomposition width.
Since the jump number is polynomial for posets of bounded width, it is natural
to ask the same question for posets of bounded height. On the negative side, Pul-
leyblank’s proof [139] of NP-hardness shows that the jump number is NP-hard even
108
when restricted to posets of height 2 or equivalently, to bipartite graphs. Müller [121]
has extended this negative result to chordal bipartite graphs. On the positive side,
there are efficient algorithms to compute the jump number of bipartite permutation
graphs: an O(n + m)-time algorithm by Steiner and Stewart [156] and O(n)-time
algorithms independently discovered by Fauck [53] and Brandstädt [16], where n and
m represent the numbers of vertices and edges of the graph respectively. The last
two algorithms require a succinct representation of the graphs (e.g. a permutation).
There are also an O(n2 )-time algorithm for biconvex graphs by Brandstädt [16], an
O(m2 )-time and an O(nm)-time algorithm for bipartite distance hereditary graphs
developed by Müller [121] and Amilhastre et al. [2] respectively, and an O(n9 )-time
dynamic program algorithm by Dahlhaus [40] for convex graphs.
All the results in the previous paragraph are based on the equivalence for bipartite
graphs between the jump number and the maximum alternating-cycle free matching
problem (Theorem 4.1.2). For chordal bipartite graphs, the only induced cycles have
length 4. Therefore, in this class, the jump number problem is equivalent to finding
a maximum matching having no 4-alternating cycle. We study the latter problem in
detail in Section 4.2. The algorithms for bipartite permutation and biconvex graphs
are further inspired by a problem in computational geometry: the minimum rectangle
cover of a biconvex rectangular region. We explore this problem in Section 4.3.1.
An important class of posets for which the complexity of this problem is still not
settled is the class of two-dimensional posets. The tractability of the jump number
problem on permutation graphs (that is, comparability graph of two-dimensional
posets, see Lemma 3.4.3) is still open and conjectured to be polynomially solvable
by Bouchitté and Habib [15]. As mentioned above, the jump number is polynomial
for bipartite permutation graphs. Ceroi [23] extended this result to two-dimensional
posets having bounded height. Interestingly enough, Ceroi [24] also gives a proof of
the NP-hardness of a weighted version of this problem and conjectures, in contrast
to Bouchitté and Habib, that the unweighted case is also NP-hard. Both results of
Ceroi use a clever reduction to the maximum weight independent set of a particular
family of rectangles. A similar construction for other types of graphs is studied in
Section 5.2.
We end this survey by mentioning other results regarding the jump number. Mi-
tas [117] has shown that the jump number problem is NP-hard also for interval posets.
However, 3/2-approximation algorithms have been independently discovered by Mi-
tas [117], Felsner [54] and Syslo [160]. As far as we know, these are the only ap-
proximation results in this area, and it would be interesting to obtain approximation
algorithms for other families of posets. It is also worth mentioning that, as shown
by von Arnim and de la Higuera [3], this parameter is polynomial for the subclass of
semiorders.
There has also been work on posets avoiding certain substructures (for exam-
ple, N -free posets as stated before in this survey). Sharary and Zaguia [149] and
Sharary [148] have shown that the jump number is polynomial for K-free posets and
for Z-free posets respectively. Lozin and Gerber [106] propose some necessary condi-
tions for the polynomial-solvability of the jump number on classes of bipartite graphs
characterized by a finite family of forbidden induced graphs and give polynomial time
109
algorithms for some of these classes. In a different article, Lozin [105] shows that the
jump number problem is also polynomial for E-free bipartite graphs.
Another interesting result is that the jump number is fixed parameter tractable:
El-Zahar and Schmerl [51] have shown that the decision problem asking whether
the jump number is at most a fixed number k is solvable in polynomial time and
McCartin [114] has given an algorithm linear in the number of vertices (but factorial
in k) to solve the same question.
4.2 Cross-Free Matchings and Biclique Covers

In this section we introduce new notation. Given a graph G = (V, E), a biclique
is the edge set of a (not necessarily induced) complete bipartite subgraph of G. A
biclique cover is a family of bicliques whose union is E. Two distinct edges e = ab
and f = a0 b0 cross if and only if ab0 and a0 b are also edges of the graph. A cross-free
matching is a collection of edges that are pairwise non-crossing. We denote the
maximum size of a cross-free matching by α∗ (G) and the minimum size of a biclique
cover2 by κ∗ (G). Two natural problems are the following.
Problem 2 (Maximum Cross-Free Matching Problem). Given a graph G = (V, E),
find a cross-free matching M of maximum cardinality α∗ (G).
Problem 3 (Minimum Biclique Cover Problem). Given a graph G = (V, E), find a
biclique cover B of minimum cardinality κ∗ (G).
Even though the definitions above make sense for any type of graph, let us focus
for the rest of this work on bipartite graphs. An important observation is that two
edges cross if and only if they are incident to the same vertex or are part of a 4-
cycle. Hence, the cross-free matchings of a graph are exactly the matchings having
no 4-alternating cycle. This observation implies the following.
Lemma 4.2.1 (Müller [121]). Let G be a chordal bipartite graph. Then cross-free
matchings and alternating-cycle-free matchings of G coincide3 .
Using this lemma, together with Theorem 4.1.2, (4.3) and (4.5), we obtain the
following theorem.
Theorem 4.2.2 (Implicit in Müller [121]). Computing the jump number of a bipartite
chordal graph G is equivalent to finding a maximum cross-free matching G, and
j(G) + α∗ (G) = n − 1, (4.9)
where n is the number of vertices of G.

2
Different authors have introduced these concepts before under different names. Müller [121, 122]
uses edge-cliques, dependent edges, and alternating C4 -free matchings to denote bicliques, crossing
edges and cross-free matchings respectively. He also uses αe (G) and κe (G) instead of α∗ (G) and
κ∗ (G). Dawande [42] denotes crossing edges that are not incident as cross-adjacent edges and uses
m∗ (G) and β ∗ (G) instead of α∗ (G) and κ∗ (G). Fishburn and Hammer [56] call κ∗ (G) the bipartite
dimension of G.
3
Müller shows this also holds for distance-hereditary graphs and strongly chordal graphs.
110
Observe that we could also have defined crossing edges of bipartite graphs as
follows. Two distinct edges cross if and only if there is a biclique containing them.
This observation implies that for any cross-free matching M and for every biclique
cover B of a bipartite graph, |M | ≤ |B|. Therefore, α∗ (G) ≤ κ∗ (G). It is a very
interesting problem to find classes of graphs for which equality holds.
Define the crossing graph4 X(G) as the graph on the edges of G where the
adjacencies are given by the crossing relation between them, this is
X(G) = (E(G), {ef : e crosses f }). (4.10)
It is immediate from their definition that cross-free matchings of a bipartite

graph G correspond exactly to stable sets of X(G). On the other hand, while ev-
ery biclique of G is a clique of X(G), the opposite is not necessarily true. However,
maximal cliques in X(G) correspond to maximal bicliques of G. Noting that there is
always a minimum cardinality biclique cover of G that uses only maximal bicliques
we conclude that
α∗ (G) = α(X(G)) and κ∗ (G) = κ(X(G)), (4.11)
where α(H) and κ(H) denote the size of a maximum stable set and the size of a
minimum clique cover of H, justifying our choice of notation. An important case
where both quantities can be computed in polynomial time is when the graph X(G)
is perfect. In this case we say that the graph G is cross-perfect. We refer the
reader to Section 3.5 for notions related to perfect graphs. For a survey of known
algorithmic results for the cross-free matchings and biclique covers we refer the reader
to Section 4.3.4.
4.3 Related Problems

4.3.1 Matrices: Boolean rank, Antiblocks and Block Covers
Here we explore some problems involving 0-1 matrices. To define them, we need some
definitions.
The boolean algebra consists of elements 0 and 1, where the sum ‘+’ is replaced
by the logical AND (or the minimum) ‘∧’ and the product ‘·’ is replaced by the logical
OR (or the maximum) ‘∨’. A boolean matrix is a matrix whose entries belong to
the boolean algebra. Addition and multiplication of boolean matrices are defined as
usual, except that we use operations in the boolean algebra. The boolean rank of
an s × t matrix M is the minimum r such that there exists an s × r boolean matrix P
and an r × t boolean matrix Q such that M = P · Q.
Problem 4 (Boolean Rank Problem). Given a boolean matrix M , find its boolean
rank and the corresponding matrices P and Q with M = P · Q.
4
Also denoted as the dependence graph by Müller [122] and as the modified line graph L0 (G) by
Dawande [42].
111
A (combinatorial) block of an s × t matrix M is a submatrix which has only
1-entries. More formally, it is a set S × T , where S ⊆ [s] is a set of rows and T ⊆ [t]
is a set of columns, such that Mij = 1 for all i ∈ S and j ∈ T . If (i, j) ∈ S × T , we
say that the entry (i, j) is covered by the block S × T .
Problem 5 (Block Cover Problem). Given a boolean matrix M , find a minimum

cardinality collection of blocks of M such that every 1-entry of M is covered by at
least one block.
The following lemma shows that the two problems above are equivalent. This
lemma can be found with a slightly different notation in an article by Gregory et
al. [74]. We prove it for completeness.
Lemma 4.3.1 (Gregory et al. [74]). The boolean rank of a matrix M is the mini-
mum number r such that the set of 1-entries of M can be covered by an union of r
combinatorial blocks of M .
Proof. Let M be an s × t matrix whose 1-entries are covered by the union of r

combinatorial blocks Sk × Tk , 1 ≤ k ≤ r. Define P as the s × r matrix having as
columns the characteristic vectors of all Sk and Q as the r × t matrix having as rows
the characteristic vectors of all Tk . This is, Pik = 1 if and only if Sk contains element
i, and Qkj = 1 if and only if Tk contains element j. Note that Mij = 1 if and only
if there is k such that (i, j) ∈ Sk × Tk , or equivalently, if and only if there is k such
that Pik ∧ Qkj = 1. Therefore M = P · Q, and so its boolean rank is at most r.
Conversely, let M be equal to P · Q, where P and Q are s × r and r × t boolean
matrices respectively. Consider the sets Sk = {i : Pik = 1} and Tk = {j : Qjk = 1}.
Each Sk × Tk defines a combinatorial block of M . Indeed, for i ∈ Sk , j ∈ Tk , Mij =
(P · Q)ij = r`=1 Pi` ∧ Q`j ≥ Pik ∧ Qkj = 1, which means that Mij = 1. By definition of
W
boolean product, Mij = 1 if and only if there exists k such that Pik = 1 and Qkj = 1,
or equivalently if and only if (i, j) is a 1-entry of some combinatorial block Sk × Tk .
Therefore the 1-entries of M are covered by the union of at most r combinatorial
blocks of M .
Consider any biadjacency matrix M for a bipartite graph G = (A ∪ B, E). It

is easy to see that the combinatorial blocks of M are in bijection with the bicliques
of G. This observation implies the following result.
Lemma 4.3.2 (Orlin [133], Gregory et al. [74]). The minimum biclique cover prob-
lem on bipartite graphs is equivalent to the boolean rank problem and the block cover
problem.
Proof. The proof follows directly from Lemma 4.3.1.
The problems related by the previous lemma also have applications to communi-
cation complexity as the non-deterministic communication complexity of a boolean
function f : X × Y → {0, 1} is precisely the binary logarithm of the boolean rank of
the input matrix of f (see, e.g. [99]).
112
We have seen how the minimum biclique cover problem of a bipartite graph is
restated in terms of biadjacency matrices. The same can be done for the maximum
cross-free matching problem. For that, define a (combinatorial) antiblock of a
boolean matrix M as a collection of 1-entries of M , no two of them contained in the
same block of M .
Problem 6 (Maximum Antiblock Problem). Given a boolean matrix M , find a max-

imum cardinality antiblock of M .
The next lemma follows immediately by interpreting the cross-free matchings of

a bipartite graph in its corresponding biadjacency matrix.
Lemma 4.3.3. The maximum cross-free matching problem of a bipartite graph is

equivalent to the maximum antiblock problem.
Abusing notation, we use α∗ (M ) and κ∗ (M ) to denote the size of a maximum

antiblock and a minimum block cover of a boolean matrix M . It is direct from their
definitions that for every such matrix α∗ (M ) ≤ κ∗ (M ). As the example in Figure 4-2
shows these quantities do not need to be equal.
 
0 0 1 1 0 0 0
0 0 1 1 1 0 0
 
 
0 1 1 0
 
 
 1 1 0 0 0 0 0 
   1 1 1 0 
M1 = 
1 1 0 1 1 0 0 .

M2 =  .

1 1 1 1
 
   
 0 1 0 1 1 1 0 
  0 0 1 1

 0 0 0 0 1 1 1 

0 0 0 0 0 1 1
Figure 4-2: The matrix on the left (example by Lubiw [109]) is not firm: the maximum
size of an antiblock is 5, but the minimum size of a block cover is 6. The matrix on
the right is firm, but the corresponding graph G is not cross-perfect: the bold entries
(1, 3), (2, 2), (3, 1), (3, 4), (4, 3) form a hole in X(G).
Lubiw [109] has proposed the problem of characterizing firm matrices, that is,
matrices for which the min-max relation κ∗ (M ) = α∗ (M ) holds for M and for all
submatrices M 0 of M . Despite its similarity, this is not the same as characterizing
cross-perfect bipartite graph (see Section 4.2) which are the graphs for which X(G) is
perfect: Given a graph G with biadjacency matrix M , the vertices of X(G) correspond
to 1-entries of M , and two entries are connected in X(G) if there is a block containing
both. In particular, X(G) is perfect if for any matrix M 0 obtained from M by changing
some 1-entries into 0-entries, the min-max relation κ∗ (M 0 ) = α∗ (M 0 ) holds. From
here, we get that cross-perfection implies firmness, however the converse does not
hold as the example in Figure 4-2 shows.
113
4.3.2 Geometry: Antirectangles and Rectangle Covers
We define two geometrical problems related to the problems presented in the previous
section. An orthogonal polygon is a polygon in the plane having only horizontal
and vertical segments.
Problem 7 ((Geometric) Minimum Rectangle Cover Problem). Given an orthogo-

nal polygon P , find a collection of orthogonal rectangles whose union is P having
minimum cardinality.
Problem 8 ((Geometric) Maximum Antirectangle Problem). Given an orthogonal

polygon P , find a collection of points inside P , no two of them contained in an
orthogonal rectangle contained in P .
It is a simple exercise to restate these problems in terms of matrices: A rectangle

of a boolean matrix M is a block whose rows and columns are consecutive in M . An
antirectangle is a collection of 1-entries of M no two of them contained in the same
rectangle of M .
Problem 9 ((Matrix) Minimum Rectangle Cover Problem). Given a boolean ma-

trix M , find a minimum cardinality collection of rectangles of M such that every
1-entry of M is covered by at least one rectangle.
Problem 10 ((Matrix) Maximum Antirectangle Problem). Given a boolean ma-

trix M , find a maximum cardinality antirectangle of M .
Problems 9 and 10 are more general than their geometrical counterparts as the
set of 1-entries of a boolean matrix does not necessarily define a polygon: the region
defined could be disconnected. Just like for the case of block covers and antiblocks,
the size of a maximum antirectangle is always at most the size of a minimum rectangle
cover.
Chvátal once conjectured that both quantities are equal for any polygon. Small
counterexamples were quickly proposed by Chung even for simply connected polygons
(see Figure 4-3). Chaiken et al. [25] have shown that the min-max relation holds for
biconvex polygons (respectively for biconvex matrices). Later, Györi [80] extends this
result to vertically convex polygons (resp. for convex matrices). Györi obtains this
result by studying a different problem, namely computing a minimum base of a family
of intervals. We revisit that problem and related ones, in the next section.
It is important to remark that in the case of biconvex matrices, maximal blocks
are rectangles. This follows directly from the following fact. If two 1-entries (i, j)
and (i0 , j 0 ) are in the same block, then (i, j 0 ) and (i0 , j) are also 1-entries. Therefore,
by biconvexity, all the entries in the rectangle with corners in (i, j) and (i0 , j 0 ) are
also 1-entries. Because in any block cover (resp. rectangle cover) we can assume
every block (resp. rectangle) to be maximal without increasing the cardinality of the
cover, the minimum block cover and the minimum rectangle cover problem coincide
for biconvex matrices. Similarly, since antiblocks and antirectangles can be defined
as collection of 1-entries, no two of them in the same maximal blocks or maximal
114
 
0 0 0 1 1 0 0 0
0 0 1 1 1 0 0 0
 
 
 

 0 1 1 1 0 0 1 1 

M=

 0 0 1 1 1 1 1 1 .

 

 0 1 1 1 1 1 1 0 


 1 1 1 0 1 1 0 0 

1 1 0 0 0 0 0 0
Figure 4-3: Chung’s counterexample (see [25]). The maximum size of an antirectangle
is 7, but the minimum size of a rectangle cover is 8.
antirectangle respectively, the maximum antiblock and the maximum antirectangle

problem coincide for biconvex matrices. In particular, using Lemmas 4.3.2, 4.3.3
and Chaiken et al.’s min-max result [25] for rectangle cover in biconvex polygons, we
obtain the following theorem.
Theorem 4.3.4. For a biconvex graph G, the size α∗ (G) of a maximum cross-free
matching and the size κ∗ (G) of a minimum biclique cover coincide.
Even though the theorem above follows from simple observations, to the author
knowledge it has not been noted in the literature for the minimum biclique cover
problem before our joint work with Telha [153] and this thesis.
4.3.3 Interval Combinatorics: Bases and Irredundancy

To simplify our discussions later, in this section we denote as interval every set of
consecutive elements of a finite linear order.
We say that a sequence (I1 , . . . , Ik ) of intervals is irredundant5 if every interval
contains a point not contained in the union of the preceding intervals. We extend this
definition to unordered families by saying that a family of intervals is irredundant if
can be sorted into a irredundant sequence.
We say that a family F of intervals generates an interval I if the latter can be
obtained as the union of intervals contained in F. A family of intervals is a basis for
another family if every interval of the latter family is generated by the former. Two
natural problems are then the following.
Problem 11 (Maximum Irredundant Subfamily Problem). Given a family F of in-
tervals, find an irredundant subfamily of maximum cardinality.
Problem 12 (Minimum Basis Problem). Given a family F of intervals, find an
interval basis B for F of minimum size, where B is not restricted to be a subfamily
of F.
If B generates an irredundant family of intervals I then we clearly have that
|B| ≥ |I| as every element of the sequence requires a new generator. Frank conjectured
5
S
Also called -increasing by Györi [80].
115
that the maximum size of an irredundant subfamily is always equal to the minimum
size of a basis. This question has been answered affirmatively by Györi [80] in a
very beautiful work. Györi reached this conclusion while studying the rectangle cover
problem defined in the previous section. Indeed, as we describe next, his result
implies as a corollary the min-max relation for convex matrices between the maximum
antirectangle and the minimum rectangle cover.
Consider a vertically convex matrix M and let B = {1, 2, . . . , t} its set of columns.
Since the matrix is vertically convex, every maximal rectangle of 1-entries of M is
defined uniquely by an interval in B.
For each row of M , consider the intervals on B induced by maximal sets of con-
secutive 1-entries. Let F be the full collection of (different) intervals obtained in this
way. Using the previous paragraph, it is easy to see that the collection of maximal
rectangles of M are in bijection with the intervals of F. See Figure 4-4 for an example.
1 2 3 4 5 6 7
1 1 0 0 1 0 0
1 1 0 1 1 0 0
I = { {1, 2}; {5}; {4, 5}; {2, 3, 4, 5, 6};
0 1 1 1 1 1 0
0 1 1 1 0 1 1 {2, 3, 4}; {6, 7}; {3}; {7} }.
0 0 1 0 0 0 1
Figure 4-4: Horizontal intervals associated to a vertically convex matrix.
Consider a minimal base B and a maximum irredundant subfamily I for F and let
k be their common cardinality. The rectangles associated to B form a rectangle cover
of M . We can construct an antirectangle of M of the same cardinality as follows. Let
(I1 , . . . , Ik ) be the irredundant sequence of I. There must be a point bi in each interval
S
Ii that is not contained in j<i Ij . Recall that each Ii was selected as a maximum
interval of consecutive 1-entries for some row ai . Let Pi be the 1-entry Mai bi . The
collection {P1 , . . . , Pk } must form an antirectangle, for suppose there was a rectangle
covering two of these points, say Pi and Pj with i < j. This means that in row ai ,
the entire interval between columns bi and bj is formed by 1-entries, contradicting the
definition of point bj . This proves the min-max relation between rectangle covers and
antirectangles for convex matrices.
Interestingly enough, Györi’s results also allow us to show a min-max relation
between block covers and antiblocks for convex matrices (or equivalently, between
biclique covers and cross-free matchings of convex graphs). We can not use the exact
construction above though since, unlike the biconvex case, for convex matrices max-
imum blocks and maximum rectangles are not equivalent, and neither are antiblocks
and antirectangles.
To do that, it is convenient to introduce some notation. A collection of point-
interval pairs {(pi , Ii ) : i ∈ [k]} is called independent if for every i 6= j, we have
pi 6∈ Ij or pj 6∈ Ii . Consider the following problem.
116
Problem 13. [Maximum Independent Set of Point-Interval Pairs] Given a family of
intervals F, find a maximum cardinality independent collection of point-interval pairs
having the intervals in F.
Györi’s proof of his min-max relation for intervals uses the following fact which
we do not prove.
Theorem 4.3.5 (Györi [80], see also Lubiw [110] and Knuth [96]). Problems 11
and 13 are equivalent: From any irredundant family of intervals we can obtain an
independent family of point-interval pairs and vice versa.
Consider now a family of intervals F over a finite total ordered set X. It is
straightforward to see that the containment bipartite graph G having parts X and F
is convex on X, and that every (finite) convex graph arise in this way.
Lemma 4.3.6. Let X, F and G be as above. The cross-free matchings of G cor-
respond to independent families of point-interval pairs in X × F. Also, from every
biclique cover of G we can obtain a base of F of the same cardinality and vice versa.
Proof. The first part follows since, by definition, two point-interval pairs are inde-
pendent if and only if the corresponding edges do not cross in G. For the second part
consider a biclique cover. We can assume that every biclique of this cover is maximal
without changing the cardinality of the cover. Since G is biconvex, the elements in
X covered by a maximal biclique form an interval. It is easy to see that the family of
all such intervals obtained from the maximal bicliques in this cover is a base for F.
Conversely, consider a base for F. For every interval I in this base, consider the
largest biclique having I as one of its parts. The collection of bicliques constructed
in this way covers all the edges.
From here, we obtain as a corollary the following.
Theorem 4.3.7. For a convex graph G, the size of maximum cross-free matching,
α∗ (G) and the size of a minimum biclique cover, κ∗ (G), coincide.
Proof. Use Györi’s min-max relation for intervals and Lemma 4.3.6.
4.3.4 Survey on the Complexity of the Presented Problems

We have already established that for chordal bipartite graphs, the jump number
problem and the maximum cross-free matching problem are equivalent (see Theo-
rem 4.2.2). For this reason, most of the algorithmic results concerning the maximum
cross-free matching problem were already mentioned in Section 4.1.1.
There are three types of results for the described problems we wish to survey:
hardness results, min-max relations, and algorithmic results. We review each type in
no particular order (the problems presented are so intertwined, that it is somehow
hard to separate them by problem).
We review first the hardness results. In 1977, Orlin [133] shows that the mini-
mum biclique cover problem is NP-hard even for bipartite graphs. This implies that
117
computing the boolean rank or finding a minimum block cover of a matrix is also
NP-hard. Two years later, Masek [113], in an unpublished manuscript (in)famous for
its elusiveness (see Johnson’s article [92]) shows that the geometric problem of cov-
ering orthogonal polygons by a minimum number of rectangles is NP-hard. Masek’s
proof, however, requires the use of polygons that are not simply connected. Several
years later, Culberson and Reckhow [39] show that the rectangle cover problem is
still NP-hard when restricted to simply connected polygons. In 1990, Müller [121]
extends Orlin’s result by showing that the minimum biclique cover is NP-hard even
for chordal bipartite graphs.
Regarding min-max relations, Chaiken et al. [25] show in 1981, that minimum rect-
angle covers and maximum antirectangles have the same size for biconvex polygons.
Three years later, Györi [80] proves that for any collection of intervals, maximum
irredundant subfamilies and minimum bases have the same cardinality. Using this,
he also extends Chaiken et al.’s result to vertically convex polygons. There are also
generalizations of Györi’s min-max result for intervals given by Lubiw [110] and Frank
and Jordán [63]. These results are described in Sections 6.2.2 and 5.5 respectively.
Now we focus on algorithms for the presented problems.
Chaiken et al.’s result is algorithmic but Gyori’s proof of his min-max relation
is not. Franzblau and Kleitman [64] have given an algorithmic proof of Györi’s re-
sult which can find maximum irreducible families and minimum bases of intervals in
O(n2 )-time. This translates immediately to an O(n2 ) algorithm for minimum rect-
angle covers and maximum antirectangles of biconvex and convex polygons. A nice
implementation of this algorithm can be found in an article by Knuth [96]. Lu-
biw [107, 109] also gives polynomial time algorithms for a somewhat larger class of
polygons, called plaid polygons. For a simpler class of biconvex polygons, denoted as
monotone orthogonal polygons in [53], Branstädt [16] and Fauck [53] independently
present O(n)-time algorithms for the minimum rectangle cover problem.
Some of the results for the minimum rectangle cover problem translate immedi-
ately to the minimum biclique cover problem. Franzblau and Kleitman’s O(n2 )-time
algorithm for rectangle covers, when applied to biconvex polygon, becomes an O(n2 )-
time algorithm for minimum biclique covers of biconvex graphs. Similarly, since
monotone orthogonal polygons corresponds exactly to matrices having 2-staircase
normal form (see Theorem 3.4.21), and these matrices are the biadjacency matrices
of bipartite permutation graphs, the minimum biclique cover of these graphs can be
computed in O(n)-time using Branstädt’s or Fauck’s algorithm.
Regarding other algorithms for the biclique cover problem, Müller [121, 122] has
shown that if G is a C4 -free bipartite graph, a strongly chordal graph, a distance
hereditary graph or a bipartite permutation graph, then the associated crossing graph
X(G) is perfect, and so (in view of (4.11)) both the minimum biclique cover and max-
imum cross-free matchings are polynomially computable in those classes. Amilhastre
et al. [2] also give polynomial algorithms for the minimum biclique cover of bipartite
domino-free graphs.
118
JUMP AC-FREE CROSS-FREE BICLIQUE-COVER
on Compar. Graphs on General Graphs on General Graphs on General Graphs
JUMP = AC-FREE CROSS-FREE ANTIRECTANGLE BICLIQUE-COVER RECT. COVER

on Bipart. Graphs on Bipart. Graphs on 0-1 Matrices on Bipart. Graphs on 0-1 Matrices
ANTIBLOCK BLOCK-COVER
on 0-1 Matrices on 0-1 Matrices
JUMP = AC-FREE = CROSS-FREE ANTIRECTANGLE RECT. COVER

on Chordal Bip. Graphs on Polygons on Polygons
ANTIBLOCK
on Biadj. mat. of Chordal Bip. Graphs
IRRED. FAM. of INTERVALS ANTIRECTANGLE MIN. INTERVAL BASE RECT. COVER

on Convex Polygons on Convex Polygons
ANTIRECTANGLE RECT. COVER

on Biconvex Polygons on Biconvex Polygons
Figure 4-5: Relation between problems
4.4 Summary of Presented Problems and Results

In Figure 4-5, we present a visual summary of some of the problems presented in
this section, where a line between one problem and another means that the problem
located below is a particular case of the problem located higher.
In Table 4.1, we summarize some of the known algorithmic results for the minimum
biclique cover and the maximum cross-free matching problems. In this table we
include a result marked as new for the jump number of convex graphs. Even though
the same result for biclique covers has already been hinted by Amilhastre et al. [2],
the possibility of efficiently computing the maximum cross-free matching of convex
graphs has often been overlooked: In the literature concerning the jump number
problem before our work, the best algorithm reported for convex graphs is still the
O(n9 )-time dynamic program algorithm by Dahlhaus [40].
Table 4.1: Table of known results.
Bip. Perm. Biconvex Convex

Max. cross-free matching (Jump number) O(n) [16, 53] O(n2 ) [16] O(n9 ) [40]
(new) - - O(n2 )a
Min. biclique-cover O(n) [16, 53] O(n2 )a O(n2 )a
a
These results follows from Theorem 4.3.7 and the algorithm of Franzblau and Kleitman [64] for
geometric rectangle cover. The possibility of applying this algorithm to compute the biclique
cover of convex graphs has been noticed by Amilhastre et al. [2].
119
120
Chapter 5
Jump Number of 2DORGs
In this chapter we show that the jump number problem is polynomially solvable
for the class of two directional orthogonal ray graphs, using its equivalence to the
maximum cross-free matching problem. A key tool for this result, in a way the most
important contribution of this part of the thesis, is a new geometrical reformulation
of the maximum cross-free matching and the minimum biclique cover problems as
the problem of finding maximum independent set and the minimum hitting set of an
associated collection of rectangles in the plane respectively.
Most of the results of this chapter are joint work with Claudio Telha. This chapter
is organized as follows.
In Section 5.1, we review some results for independent and hitting sets of rectan-
gles. A big conjecture in the area is whether or not the ratio between the size of the
minimum hitting set and the size of the maximum independent set of any family of
axis-aligned rectangles is bounded by a constant. In passing, we observe that a sim-
ple application of recent approximation algorithms for both problems gives a bound
of O(log2 log α) for this ratio, where α is the size of the maximum independent set,
improving the existing bounds of O(log α).
In Section 5.2, we describe our key tool. In Section 5.3 we give a linear pro-
gramming based algorithm for the maximum cross-free matching on 2dorgs. The
e ω )-time combinatorial algorithm to compute, for
main result of Section 5.4 is an O(n
a given 2dorg, both a cross-free matching and a biclique cover of the same size. In
particular, this shows that the min-max relation holds.
We later explore the relation between our results with previous work. In particular,
we show how our min-max relation can be obtained as a particular case of a strong
result by Frank and Jordán [63]. Finally, we present some open problems.
5.1 Maximum Independent Sets and Minimum

Hitting Sets of Rectangles
As preparation for our results, we review two geometric problems: the maximum
independent set of rectangles and the minimum hitting set of rectangles.
We start by defining the geometric notion of a rectangle. For our purposes, a
121
rectangle is the cartesian product of two closed intervals, viewed as a subset of the
plane R2 . In other words, we only consider possibly degenerated rectangles in the
plane having their sides parallel to the x and y axes. Alternatively, we can define
a rectangle by describing its bottom-left corner and its top-right corner. Formally,
given two points a and b in R2 , the (geometric) rectangle Γ(a, b) is defined as:
Γ(a, b) = {p ∈ R2 : ax ≤ px ≤ bx , ay ≤ py ≤ by }. (5.1)
Note that if a 6≤R2 b, the rectangle Γ(a, b) is empty. Given a nonempty rectangle R, we
use bl(R) and tr(R) to denote its bottom-left corner and top-right corner respectively.
Two rectangles are independent if they do not intersect, otherwise they are
intersecting. Given a family C of rectangles, an independent set of C is a family
of pairwise independent rectangles. We say that a point p ∈ R2 hits a rectangle R if
p ∈ R. A set of points H ⊆ R2 is a hitting set of C if every rectangle in C is hit by at
least one point of this set. We denote by mis(C) and mhs(C) the sizes of a maximum
independent set of rectangles in C and a minimum hitting set for C respectively. It is
natural to study the following problems.
Problem 14 (Maximum Independent Set of Rectangles Problem). Given a family C

of rectangles, find an independent set I ⊆ C of maximum cardinality, this is |I| =
mis(C).
Problem 15 (Minimum Hitting Set of Rectangles Problem). Given a family C of

rectangles, find a hitting set H of C of minimum cardinality, this is |H| = mhs(C).
Consider the intersection graph of a family C of rectangles,
I(C) = (C, {RS : R ∩ S 6= ∅}). (5.2)
Naturally, independent sets of C correspond to stable sets in I(C). Therefore,
mis(C) = α(I(C)). (5.3)
It is a well known fact that rectangles have the Helly property. This is, if a
collection of rectangles pairwise intersect, then all of them share a point in the plane.
In fact, their intersection is a nonempty rectangle. In particular, we can assign to
every clique C in I(C) a unique witness point, defined as the bottom-left point of
the rectangle formed by the intersection of all elements in C.
Consider now a hitting set H for C. The set of rectangles hit by a point p in H
is a (not necessarily maximal) clique C(p) of I(C). By replacing p by the witness
point of any maximal clique containing C(p) we obtain a new hitting set of the same
cardinality. From here we conclude that C admits a minimum hitting set consisting
only of witness points of maximal cliques, or equivalently a clique cover of I(C) using
only maximal cliques. From here,
mhs(C) = κ(I(C)). (5.4)
122
Since any point hits at most one rectangle of an independent set, we deduce the
relation
mis(C) ≤ mhs(C). (5.5)
An elementary, but important observation is that for both problems defined above
we can restrict ourselves to the family C↓ of inclusion-wise minimal rectangles in C.
Lemma 5.1.1.
mis(C) = mis(C↓ ) and mhs(C) = mhs(C↓ ).
Proof. We can transform any independent set in C into an independent set in C↓ of the
same cardinality by replacing each non-minimal rectangle by a minimal one contained
in it. Since every independent set in C↓ is also independent in C, we obtain the first
equality. For the second one, use that every hitting set of C↓ is also a hitting set of C
and vice versa.
Given any hitting set H ⊆ R2 containing the witness points of all maximal cliques
in C, define the following polytopes:
( )
PH (C) = x ∈ R : C
X
xR ≤ 1, for all p ∈ H, x ≥ 0 , (5.6)
R: p∈R
( )
DH (C) = y ∈ R : H
X
yp ≥ 1, for all R ∈ C, y ≥ 0 , (5.7)
p: p∈R
and the associated integer and linear programs:

( )
C
X
IPH (C) = max xR : x ∈ PH (C) ∩ {0, 1} , (5.8)
R∈C
( )
X
LPH (C) = max xR : x ∈ PH (C) , (5.9)
R∈C
( )
IP0H (C) H
X
= max yp : y ∈ DH (C) ∩ {0, 1} , (5.10)
p∈H
( )
LP0H (C)
X
= max yp : y ∈ DH (C) . (5.11)
p∈H
Here, abusing notation we use IPH (C), LPH (C), IP0H (C), LP0H (C) to denote both
the program formulations and their optimal values.
The characteristic vector of any independent set of C is inside PH (C). Conversely,
every integer vector x of PH (C) is the characteristic vector of an independent set of C:
If there is a pair of intersecting rectangles in the support of x, then the inequality
associated to any witness point of a maximal clique containing both rectangles is not
satisfied. In particular, (5.8) is an integer program formulation for the maximum
independent set problem and (5.9) is its corresponding linear program relaxation.
123
Note also that every integer vector y of DH (C) is the characteristic vector of a
hitting set of C. However, only the characteristic vectors of hitting sets that are sub-
sets of H are contained in DH (C). Nevertheless, since there is an optimum hitting set
using only witness points, the set DH (C) contains at least one such optimum. There-
fore, (5.10) and (5.11) are integer and linear program relaxations for the minimum
hitting set. It is important to note that the values LPH (C) and LP0H (C) coincide since
they correspond to dual linear programs.
Another important observation is that PH (C) does not depend on the choice of H.
As the following lemma shows, it only depends on the intersection graph I(C).
Lemma 5.1.2. For any hitting set H ⊆ R2 containing the witness points of all
maximal cliques in C,
PH (C) = QSTAB(I(C)).
Proof. It follows directly from the definition of QSTAB(I(C)) (see (3.4) in Section 3.5)
and the fact that inequalities associated to non-witness points are implied by the ones
associated to witness points.
Denote by P(C) and LP(C) the common polytope and associated linear program
value. For every collection of rectangles C we have
mis(C) ≤ LP(C) ≤ mhs(C). (5.12)
It is interesting to find cases where (5.12) holds with equality. This happens, for
instance when the polytope P (C) is integral. We can characterize those cases.
Lemma 5.1.3. Given a family of rectangles C the polytope P(C) is integral if and
only if the intersection graph I(C) is perfect.
Proof. The lemma follows using that P(C) = QSTAB(I(C)) (Lemma 5.1.2) and The-
orem 3.5.2 for perfect graphs.
We give two examples of families where I(C) is perfect.
Interval Families.
Lemma 5.1.4. Let C be a family of rectangles such that there is a single horizontal (or
vertical) line intersecting all of them. Then I(C) is an interval graph, and therefore
a perfect graph.
Proof. Assume without loss of generality, by rotating and translating the collection,
that the line intersecting every rectangle is the x-axis. For each rectangle R, let I(R)
be the interval obtained by projecting R over the x-axis. For every pair of rectangles
R and R0 on the family, R ∩ R0 6= ∅ implies that I(R) ∩ I(R0 ) 6= ∅. Conversely, if
I(R) ∩ I(R0 ) 6= ∅, then both R and R0 contain the line segment I(R) ∩ I(R0 ), and
therefore, they intersect. This means that the intersection graph of C is equal to the
intersection graph of the corresponding intervals, completing the proof.
124
Skew Intersecting Families. Every rectangle R is defined as the cartesian product
of two closed intervals Rx and Ry (where one or both could possibly be a singleton
interval). We say that R and R0 skew-intersect if
Rx ⊆ Rx0 and Ry ⊇ Ry0 (5.13)
or vice versa. We give some examples of pairs of skew-intersecting rectangles in

Figure 5-1.
Figure 5-1: Skew-intersecting rectangles.
Lemma 5.1.5. Let C be a family of rectangles such that every pair of intersecting
rectangles are skew-intersecting. Then I(C) is a comparability graph, and therefore a
perfect graph.
Proof. Consider the following partial order in C:
R R0 if and only if Rx ⊆ Rx0 and Ry ⊇ Ry0 . (5.14)
Note that two rectangles R and R0 skew-intersect if and only if R R0 or R0 R.

Therefore, every pair of intersecting rectangles in C are comparable by . This
means that I(C) is a comparability graph. In every pair of intersecting rectangles of
Figure 5-1, the rectangle in bold precedes the other one in the partial order .
For families C of rectangles having perfect intersection graph I(C), not only we
have equality in (5.12), but we can also compute a maximum independent set by
finding an optimal extreme solution of the linear program LP(C). This can be done in
polynomial time since there is a formulation of the linear program using a polynomial
number of inequalities. For that, use that LP(C) = LPH (C), where
H = {(x, y) : x is the x-coordinate of a corner of a rectangle in R and

y is the y-coordinate of a corner of a rectangle in R}. (5.15)
It is easy to see that H is a hitting set containing the witness points of maximal
cliques of C. Furthermore, |H| ≤ (2m)2 , where m = |C|.
Since there are polynomial time algorithms to find minimum clique-covers of per-
fect graphs (see Section 3.5), we can also find a minimum hitting set of these families
in polynomial time. For the particular case of interval families and skew-intersecting
families, there are more efficient algorithms that exploit the fact that the corre-
sponding intersecting graphs are interval graphs, where a simple greedy algorithm
works [77], or comparability graphs, as observed in Lemma 3.2.3.
Both the maximum independent and the minimum hitting set problems are NP-
hard, as shown by, e.g. Fowler et al. [60]. Hence, the problem of approximating
125
Figure 5-2: The family Cn contains 4n + 4 rectangles: n fat vertical, n thin vertical,
n fat horizontal, n thin horizontal and 4 rectangles on the outside.
these quantities is also of relevance. An important related conjecture attributed to

G. Wegner1 is the following. Is it always the case that
mis(C) ≤ 2 · mhs(C) − 1 ? (5.16)
Note that this conjecture is trivially tight by letting C be a single rectangle, where
mhs(C) = mis(C) = 1. Gyárfás and Lehel [79] relaxed this conjecture to the following.
Is there a constant c such that
mis(C) ≤ c · mhs(C) ? (5.17)
Their conjecture is also still open. An important question is then: how large can the
ratio mhs(C)/mis(C) be?
Gyárfás and Lehel [79] give a simple example for which mhs(C)/mis(C) ≥ 3/2.
Fon-Der-Flaass and Kostochka [58] improve this lower bound to 5/3. In an un-
published manuscript, Cibulka et al. [34] give a family {Cn }n≥3 of examples, where
mis(Cn ) = n + 2 and mhs(Cn ) = 2n + 2, from which
mhs(Cn )
lim = 2. (5.18)
n→∞ mis(Cn )
The examples achieving this bound are depicted in Figure 5-2. In particular, this
example shows that if Gyárfás and Lehel’s conjecture holds, then the constant c in
(5.17) is at least 2.
On the upper bound side, many authors (e.g. Károly and Tardos [98], Fon-Der-
Flaas and Kostochka [58]) have shown that
mhs(C)
≤ O(log(mis(C))). (5.19)
mis(C)
1
In the original text by Wegner [167], he actually asked whether for the case where mis(C) ≥ 2,
mis(C) ≤ 2mhs(C) − 3 or not.
126
We use this opportunity to give an improvement on this bound2 .
Theorem 5.1.6.
mhs(C)
≤ O log2 log(mis(C)) . (5.20)
mis(C)
In order to prove this bound, we need two recent approximation algorithms for
mis(C) and mhs(C).
Chalermsook and Chuzhoy [26] have developed an O(log log m)-approximation al-
gorithm for the maximum independent set of a family of m rectangles. More precisely,
they have shown that for any family of rectangles C having their corners in a grid [t]2 ,
it is possible to find an independent set K with
mis(C) ≤ LP[t]2 (C) ≤ |K|O(log log(t)) (5.21)
On the other hand, Aronov et al. [4] have shown the existence of O( 1ε log log 1ε )-nets
for families of axis-parallel rectangles. This implies, using the approach of Brönni-
mann and Goodrich [18], that for every family C of rectangles, there exist a (polyno-
mial time computable) hitting set H, with
|H| ≤ O(LP(C) log log(LP(C))). (5.22)
To prove Theorem 5.1.6, we require the following lemma.

Lemma 5.1.7. For every family of rectangles C, with α = mis(C), there is another
family C 0 of rectangles with corners in the grid [α]2 such that
mis(C 0 ) ≤ mis(C) ≤ LP(C) ≤ 9LP(C 0 ). (5.23)
Proof. Let Cx (resp. Cy ) be the family of intervals obtained by projecting C on the

x-axis (resp. y-axis). Let Hx (resp. Hy ) be a minimum hitting set for Cx (resp. Cy ).
The intersection graph of Cx is an interval graph, and therefore, a perfect graph.
Hence,
|Hx | = mhs(Cx ) = mis(Cx ) ≤ mis(C) = α, (5.24)
and similarly |Hy | ≤ α.

Consider the grid Hx × Hy of size ≤ α × α. By translating and piece-wise scaling
the plane, we can identify Hx with the set {(i, 0) : 1 ≤ i ≤ |Hx |} and Hy with the
set {(0, j) : 1 ≤ j ≤ |Hy |} without changing the intersection graph associated to C.
Thus, we can identify the grid Hx × Hy with a subgrid of [α] × [α]. Note that this
grid is itself, a hitting set of C.
Furthermore, consider the family Ce = {R ∩ [1, α] × [1, α] : R ∈ C}. This is, Ce is
obtained by trimming the rectangles to the rectangular region [1, α] × [1, α]. It is easy
to see that this operation does not change the intersection graph of the family. So,
for our purposes, we can assume w.l.o.g. that C = C.e
2
Even though this observation is an almost straightforward application of [26] and [4], as far as
we know, it has not been yet published in any medium.
127
Let C 0 be the family of rectangles obtained by “growing” each rectangle of C in
such a way that everycorner of it is on a vertex of the grid. This is, we replace
R = Γ(a, b) by R+ = Γ (bax c, bay c), (dbx e, dby e) .
The first inequality of (5.23) follows since any independent set of C 0 induces an
independent set of C of the same size. The second inequality follows from (5.12). The
only non-trivial inequality is the last one.
Since [α]2 is a hitting set for C and C 0 , LP[α]2 (C) = LP(C) and LP[α]2 (C 0 ) = LP(C).
Consider a fractional optimal solution y 0 for LP0[α]2 (C 0 ) (see (5.11)) and recall that the
support of y 0 is contained in [α]2 . Observe that if p is a point in the support of y
that fractionally hits some grown rectangle R+ , then either p, one of its 4 immediate
neighbors in the grid or one of its 4 diagonal neighbors in the grid will hit the original
rectangle R. Define y as
yq = yq0 + yp0 , for all q ∈ [α]2 .

X
(5.25)
p∈[α]2 : p immediate or diagonal neighbor of q
By the previous observation, y is a fractional feasible solution for the dual of

LP[α]2 (C), and by definition, its value is at most 9 times the value of y 0 .
Now we are ready to prove Theorem 5.1.6.
Proof of Theorem 5.1.6. Let C be a family of rectangles with mis(C) = α and C 0 the
family guaranteed by Lemma 5.1.7. Then, by combining (5.22) and (5.21), we have:
mhs(C) ≤ O(LP(C) log log(LP(C))) ≤ O(LP(C 0 ) log log(LP(C 0 )))

= O(LP[α]2 (C 0 ) log log(LP[α]2 (C 0 )))
≤ O(α log log(α) log log(α log log(α))) = O(α(log log(α))2 ).
5.2 Geometric Interpretation for 2DORGs

In this section we establish a geometric interpretation of maximum cross-free match-
ings and minimum biclique covers of 2dorgs as maximum independent sets and
minimum hitting sets of certain families of rectangles in the plane.
Consider a 2dorg G with bicolored 2D-representation (A, B). For every edge
ab in G, the rectangle Γ(a, b) is nonempty. We denote by R(A, B) the set of all
nonempty rectangles Γ(a, b) with a ∈ A and b ∈ B.
We remark that since A and B are maybe multisets, some of the rectangles in
R(A, B) could be equal. Also, since A and B are not necessarily disjoint, some of the
rectangles in R(A, B) may consist of a single point in the plane. These degeneracies do
not happen if (A, B) is a bicolored rook representation: In this case, all the rectangles
in R(A, B) are distinct, they have corners in a fixed grid and have nonempty interior.
In what follows, we use G(A, B, R) to represent the 2dorg G having bicolored
2D-representation (A, B) and such that R = R(A, B). We do not impose (A, B) to
be a bicolored rook representation.
128
The following theorem characterizes crossing edges of G(A, B, R) in terms of R
(see Figure 5-3.)
Theorem 5.2.1 (Soto and Telha [153]). Two edges ab and a0 b0 of G(A, B, R) cross
if and only if Γ(a, b) and Γ(a0 , b0 ) intersect as rectangles.
Proof. Let R = Γ(a, b) and R0 = Γ(a0 , b0 ) be distinct rectangles in R. The edges ab

and a0 b0 cross if and only if Γ(a, b0 ) and Γ(a0 , b) are also in R(A, B). This is equivalent
to max(ax , a0x ) ≤ min(bx , b0x ) and max(ay , a0y ) ≤ min(by , b0y ). From here, if ab and a0 b0
cross, then the point p = (max(ax , a0x ), max(ay , a0y )) is in the intersection of R and R0
as rectangles. Conversely, if there is a point p ∈ R ∩ R0 , then max(ax , a0x ) ≤ px ≤
min(bx , b0x ) and max(ay , a0y ) ≤ py ≤ min(by , b0y ), implying that ab and a0 b0 cross.
Figure 5-3: For 2dorgs, crossing edges are equivalent to intersecting rectangles
Recall the definition of the crossing graph of a graph G (see (4.10)). This is
the graph X(G) on the edges of G, where two elements are adjacent if they cross.
Theorem 5.2.1 states that for a 2dorg G = G(A, B, R) the intersection graph I(R)
and the crossing graph X(G) are isomorphic.
As a consequence of this, all the properties we explored in Section 5.1 about
independent sets and hitting sets of rectangles translate to X(G). For example, as
shown in Figure 5-4, we can associate to every maximal biclique of G a unique witness
point in the plane hitting all the corresponding rectangles, and to every point in the
plane we can associate a (possibly empty) biclique of G.
Figure 5-4: For 2dorgs, maximal bicliques corresponds to witness points.
More importantly, we obtain immediately the following theorem.
Theorem 5.2.2 (Soto and Telha [153]). For a 2dorg G = G(A, B, R),
α∗ (G) = mis(R) = mis(R↓ ) and κ∗ (G) = mhs(R) = mhs(R↓ ).
129
5.3 Linear Programming Formulation
In this section, we use the linear programming relaxation for the maximum indepen-
dent set of rectangles (see Section 5.1) to obtain a polynomial time algorithm for the
maximum cross-free matching of 2dorgs.
If G = G(A, B, R) is a 2dorg such that I(R) (or I(R↓ )) is perfect, then, by
Theorem 5.2.2 and Lemma 5.1.3,
α∗ (G) = κ∗ (G), (5.26)
and solving the linear program LP(R) (or LP(R↓ )) gives a polynomial time algorithm
for finding a maximum cross-free matching. As observed in Section 5.1, we can find
minimum clique-covers of perfect graphs in polynomial time. Therefore, we can also
obtain a minimum biclique cover of G.
Not every 2dorg G = G(A, B, R) is such that I(R) or I(R↓ ) is a perfect graph.
However, this holds for some subclasses. For a bipartite permutation graph G, the
graph X(G) is known to be perfect since it is both weakly chordal [122] and co-
comparability [16]. Since X(G) corresponds to I(R) for any representation of G =
G(A, B, R) as a 2dorg, we can find using the previous observation a maximum
cross-free matching and minimum biclique cover for bipartite permutation graphs in
polynomial time. As mentioned in Section 4.1.1, linear-time algorithms have been
developed for both problems [156, 16, 53]. These algorithms use other properties of
the structure of bipartite permutation graphs.
Let G = G(A, B, R) be a biconvex graph. As Figure 5-5 shows, the graph X(G) =
I(R) is not necessarily perfect.
b1 b2 b3 b4 b2
b3
0 1 1 0
 
a1 a1
b1
a2

 1 1 1 0 

a2
1 1 1 1
 
a3  
b4
a4 0 0 1 1 a3
a4
(a) biadjacency matrix (b) 5-hole
Figure 5-5: A biadjacency matrix and a bicolored 2D-representation of a biconvex

graph. The bold entries form a 5-hole. This 5-hole is represented as a 5-cycle of
rectangles in the right.
However, for biconvex graphs we can show the following result.
Theorem 5.3.1. For every biconvex graph G, there is a representation as a 2dorg

G = G(A, B, R) for which I(R↓ ) is perfect.
Proof. Consider a graph G with biconvex adjacency matrix M . Let {a1 , . . . , as } and
{b1 , . . . , bt } be the vertices of both parts of G, in the order induced by M . Consider
130
each element bj ∈ B as an interval Ij = [`(j), r(j)] of elements in A, this is `(j) and
r(j) are the minimum and maximum indices i for which Mij = 1.
It is easy to check that there are no four indices j1 , j2 , j3 , j4 for which the left and
right extremes of the first three intervals are strictly increasing,
`(j1 ) < `(j2 ) < `(j3 ), r(j1 ) < r(j2 ) < r(j3 ), (5.27)
and simultaneously, the second and fourth interval satisfy
`(j4 ) < `(j2 ), r(j2 ) < r(j4 ). (5.28)
This follows from the fact that no matter how j1 , j2 , j3 and j4 are sorted in the
labeling of B, there is always an element a ∈ A for which the corresponding ones in
the associated row of M , restricted to the previous four columns, are not consecutive.
We use the previous property to prove the lemma. Recall that the graph G admits
a bicolored representation (A, B), with A on the line y = −x and B weakly above
this line. This representation is obtained by the identification
ai = (i, −i), (5.29)

bi = (ri , −li ). (5.30)
We claim that the family of rectangles R = R(A, B) obtained from the above
representation is such that the intersection graph of R↓ is perfect. We show this
using the Strong Perfect Graph Theorem [32].
Suppose there is a hole H = {R1 , R2 , . . . , Rk } ⊆ R↓ of I(R↓ ) with k ≥ 5, where
R` intersects R`−1 and R`+1 (mod k), and R` = Γ(ai` , bj` ).
Assume that R1 is the rectangle in H having i1 as small as possible. It is easy to
check that i1 6= i2 , since otherwise any rectangle in H intersecting the thinnest3 of R1
and R2 must intersect the other one. By a similar argument, i1 6= ik and ik 6= ik−1 .
Also, since R2 and Rk do not intersect, i2 6= ik . Therefore, without loss of generality
we can assume that i1 < i2 < ik . For the rest of the argument, we use Figure 5-6.
We use λ1 to denote the vertical line x = (aik )x = ik , λ2 to denote the horizontal line
y = (ai1 )y = −i1 , Z1 to denote the triangular zone bounded by λ1 , λ2 and the line
y = −x, Z2 to denote the rectangular area strictly above λ2 and strictly to the left of
λ1 , and Z3 to denote the rectangular area strictly below λ2 and strictly to the right
of λ1 . Note that the point where λ1 and λ2 intersect is in R1 ∩ Rk .
First, we claim that i` < ik , for all ` 6= k. Assume, for contradiction, that ik < i`
for some `, meaning that ai` is outside the triangular zone Z1 in Figure 5-6. Since H
is a hole, it is possible to draw a continuous curve in the plane going from the point
ai2 ∈ R2 to the point ai` ∈ R` without ever touching rectangles R1 and Rk . But this
is impossible since to leave Z1 we must touch either λ1 or λ2 . Here, we have used
that Z1 ∩ λ1 ⊆ Rk and Z1 ∩ λ2 ⊆ R1 .
Recall that R2 intersects R1 and that Rk−1 intersects Rk . From here it is easy
to see that i2 < ik−1 (as in the picture) as otherwise R2 and Rk−1 would intersect,
3
The one having smallest length in the x-direction.
131
λ1
Z2
λ2
ai1
Z1
ai2 Z3
aik−1
aik y = −x
Figure 5-6: Proof for no holes.
contradicting the fact that H is a hole with k ≥ 5.

Now we can deduce where the corners in B are located.
Both bi1 and bik are weakly above λ2 and to the right of λ1 . Since R2 intersects R1
but not Rk , the corner bi2 of R2 must be located strictly to the left of the line λ1 , as
otherwise R2 would intersect Rk . The corner bi2 must also be strictly above the line λ2 :
It can not be strictly below λ2 as otherwise R1 and R2 would not intersect, and it can
not be on the line λ2 since then bi2 ∈ R1 , contradicting the inclusion-wise minimality
of R1 . This means that bi2 is in zone Z2 of Figure 5-6. By an analogous argument,
the corner bik−1 of Rk−1 lies in zone Z3 . Also, since all the points {ai1 , . . . , aik } are in
the zone Z1 and the only rectangles intersecting λ1 or λ2 are Rk−1 , Rk , R1 and R2 ,
we conclude that the points {bi3 , . . . , bik−2 } are also in Z1 .
Note that at least one of the following holds:
1. bi3 ∈ Z1 is located above bik−1 .
2. bik−2 ∈ Z1 is located to the right of bik−1 .

If neither does, it is not hard to show that R2 and Rk−1 intersect. Set j2 equal to
i3 if the first condition holds and equal to ik2 if the second condition holds. Also set
j1 = i2 , j3 = ik−1 and j4 = i1 . Then the sequence (j1 , j2 , j3 , j4 ) satisfies (5.27) and
(5.28), contradicting that G is biconvex.
We have shown that I(R↓ ) contains no odd-hole of size ≥ 5. Now we focus on
antiholes. Suppose that A = {R1 , R2 , . . . , Rk } is an antihole of I(R↓ ) of length at
least 7, where Rj intersects every rectangle except Rj−1 (mod k) , and Rj+1 (mod k) . Let
R1 and Rm be the rectangles in A with minimum and maximum value of its lower
left corner (i1 and im respectively). As in the case of holes, we can check that R1 and
Rj are the only rectangles with corners at ai1 and aim respectively.
We now consider two cases, depending on whether R1 and Rm intersect.
If R1 and Rm have empty intersection, then m = 2 or m = k. For both values
of m, R4 and R5 intersect both R1 and Rm (since the antihole has length k ≥ 7). This
implies that R4 and R5 must contain the point p where the line through the bottom
side of R1 and the line through the left side of Rm intersect (see Figure 5-7), which
contradicts the antihole definition.
The second case is if R1 and Rm intersect (see Figure 5-8). Since the length of the
antihole is k ≥ 7, there is j ∈ [k] such that {1, m} ∩ {j − 1, j, j + 1}mod k = ∅. This
132
R1 p
Rm
y = −x
Figure 5-7: First case of the proof for no antiholes.
means that Rj intersects both R1 and Rm . Consider the rectangles Rj−1 and Rj+1 ,
since they do not intersect Rj , each one must be completely contained in one of the
semi-infinite zones Z1 or Z2 . Furthermore, since Rj−1 and Rj+1 do intersect, they
both lie in the same zone. But this implies that one of R1 and Rm is not intersected
by both Rj−1 and Rj+1 , which is a contradiction with the definition of the antihole.
ai1 Z1
Rj
aij Z2
aim y = −x
Figure 5-8: Second case of the proof for no antiholes.
We have shown that I(R↓ ) contains no hole of size ≥ 5 and no odd antihole of
size ≥ 7 (size 5 is covered since antiholes and holes of size 5 coincide). By the Strong
Perfect Graph Theorem [32], we conclude that I(R↓ ) is perfect.
Let us go back to the general case of 2dorgs. Recall that any such graph admits
a bicolored rook representation (by Lemmas 3.4.1 and 3.4.9). In what follows let
G = G(A, B, R) where (A, B) is a bicolored rook representation. In this case no
two points of A ∪ B share the same position and every rectangle has full-dimensional
interior.
An elementary, but important observation is the following.
Remark 5.3.2. If R = Γ(a, b) is an inclusion-wise minimal rectangle in R↓ , then R
does not contain any point in (A ∪ B) \ {a, b}.
This remark holds since the existence of a point c ∈ R ∩ ((A ∪ B) \ {a, b}) im-
plies that either Γ(a, c) or Γ(c, b) is a rectangle in R completely contained in R,
contradicting its minimality.
By Remark 5.3.2, there are only four ways in which a pair of rectangles of R↓ can
intersect. They are shown in Figure 5-9. We say that two intersecting rectangles have
133
corner-intersection if one rectangle contains a corner of the other in its topological
interior. If both are in R↓ , the only way this can happen is that one rectangle contains
the top-left corner of the other, while the other contains the bottom-right corner of
the first.
If the previous case does not happen, we say that the intersection is corner-free.
A corner-free-intersection (c.f.i.) family is a collection of inclusion-wise minimal
rectangles having no corner-intersections.
b b b b = b0
b0 b0 a
a a b0
a0 a0 a = a0 a0
(a) corner-intersection (b) corner-free intersections
Figure 5-9: The only ways two rectangles in R↓ can intersect each other.
Note that c.f.i. families are skew-intersecting families. Therefore, by Lemma 3.4.9,
if R↓ has no pair of corner-intersecting rectangles, I(R↓ ) is perfect.
Let z ∗ be the optimal value of LP(R↓ ) and for every rectangle R ∈ R↓ let µ(R) > 0
be its area4 . Let x̄ be an optimal extreme point of the following linear program
 
 X 
LP(z ∗ , R↓ ) = min  xR = z ∗ and x ∈ P(R↓ ).
X
µ(R)xR :
R∈R↓ R∈R↓
Note that x̄ is a solution to LP(R↓ ) minimizing the total weighted area covered.
It is interesting to note that x̄ is also an optimal extreme point of the linear
program
 
 X 
max (1 − εµ(R))xr : x ∈ P(R↓ ) ,
 
R∈R↓
for sufficiently small values of ε. In particular, this means we can find x̄ by solving
only one linear program.
Theorem 5.3.3. The point x̄ is an integral point.
Proof. Suppose that x̄ is not integral. Let R = {R ∈ R↓ : x̄R > 0} be the set of
rectangles in the support of x̄. We claim that R is a c.f.i. family. Assume for sake
of contradiction that R = Γ(a, b) and R0 = Γ(a0 , b0 ) are two rectangles in R having
corner-intersection as in Figure 5-9a. We apply the following uncrossing procedure:
let ε = min(x̄R , x̄R0 ) > 0 and consider the rectangles S = Γ(a, b0 ) and S 0 = Γ(a0 , b)
in R↓ . Consider the vector x̂ obtained from x̄ by decreasing xR and xR0 by ε and
increasing by the same amount xS and xS 0 . It is easy to see that x̂ is a feasible
4
For our discussion, the area or a rectangle R = Γ(a, b) is defined as (bx − ax )(by − ay ). However,
our techniques also works if we define the area of a rectangle as the number of grid points it contains.
134
solution of LP(z ∗ , R↓ ) with strictly smaller weighted area than x̄, contradicting its
optimality.
Since R is a c.f.i. family, I(R) is a perfect graph. Therefore, P(R) is an integral
polytope. Consider now the set
( )
F = P(R) ∩ x ∈ RR↓ : xR = z ∗ ,
X X X
µ(R)xR = µ(R)x̄R . (5.31)
R∈R R∈R R∈R
This set is a face of P(R) containing only optimum solutions of LP0 (z ∗ , R↓ ). To

conclude the proof we show that x̄ is a vertex of F , and therefore, a vertex of P(R).
If x̄ is not a vertex of F , we can find two different points in F such that x̄ is a convex
combination of those points. This means that x̄ is a convex combination of different
optimum solutions of LP0 (z ∗ , R↓ ), contradicting the choice of x̄.
We can find a maximum cross-free matching of G = G(A, B, R) where (A, B) is

a bicolored rook representation, using Theorem 5.3.3 as follows: Find the optimal
value z ∗ of the linear program LP[n]2 (R↓ ) = LP(R↓ ), where n = |A ∪ B| and then
find an optimal extreme point of LP(z ∗ , R↓ ).
Using that 2dorgs are chordal-bipartite graphs, for which the maximum cross-free
matching and the jump number problems are equivalent, we conclude the following.
Theorem 5.3.4. The maximum cross-free matching and, equivalently, the jump num-
ber of a 2dorg can be computed in polynomial time.
In this section we have shown not only that the value of the linear program LP(R↓ )
equals mis(R↓ ) but also that an integer optimal solution of LP(R↓ ) can be found by
optimizing an easy to describe linear function over the optimal face of this linear
program. In the next section, we give an algorithmic proof that mis(R↓ ) = mhs(R↓ ),
implying that the dual linear program LP0[n]2 (R↓ ) (see (5.11)) also admits an integral
vertex. We conjecture that there is also a simple direction such that optimizing on
this direction over the optimal face of LP0[n]2 (R↓ ) yields the integral solution.
5.4 Combinatorial Algorithm

In this section we give a combinatorial algorithm that computes simultaneously a
maximum cross-free matching and a minimum biclique cover of a 2dorg. Our proce-
dure shares ideas with the algorithmic proof of Györi’s min-max result of intervals [80]
given by Frank [62]. In order to keep the discussion self-contained, we do not rely on
that result and prove every step.
In Section 5.3 we have shown that a maximum cross-free matching of a 2dorg
G = (A, B, R) can be obtained from a maximum independent set of a c.f.i. family R
(as a matter of fact, we have shown that R itself is an independent set). In what
follows, we show that this result also holds if we replace R by a certain maximal
greedy c.f.i. subfamily of R↓ .
135
Consider a 2dorg G = (A, B, R) where (A, B) is a bicolored rook representation.
In particular, the rectangles are in the grid [n]2 , with n = |A ∪ B|. We say that a
rectangle R ∈ R↓ appears before a rectangle S ∈ R↓ in right-top order if either
1. bl(R)x < bl(S)x , or
2. bl(R)x = bl(S)x and tr(R)y < tr(S)y ,

where, as we recall, bl(·) stands for the bottom left corner (the corner in A) and tr(·)
stands for the top right corner (the corner in B). The right-top order is a total order
on R↓ .
Construct a family K by processing the rectangles in R↓ in right-top order and
adding only those that keep K corner-free.
Since I(K) is the comparability graph of (K, ), where is defined as in (5.14),
the size of a maximum independent set R0 of K equals the size of a minimum hit-
ting set H0 for K. We can find optimal solutions of these problems by computing a
maximum antichain and a minimum chain-cover of the poset (K, ), using any poly-
nomial time algorithm for the Dilworth chain-partitioning problem (see, for example,
Lemma 3.2.3). We modify H0 to obtain a set of points H ∗ of the same size hitting
all R.
An admissible flip of a hitting set H for K is an ordered pair of points p and
q in H with px < qx and py < qy , such that the set H \ {p, q} ∪ {(px , qy ), (qx , py )}
obtained by flipping p and q is still a hitting set for K.
Construct H ∗ from H0 by flipping admissible flips while this is possible. Since
we only use points of the grid [n]2 , flipping two points reduces the potential ψ(H) =
2 3
p∈H px py by at least one unit. Using that ψ is positive and ψ(H0 ) ≤ |H0 |n ≤ n ,
P
∗
we do at most this many flips. Therefore, we can construct H in polynomial time.
To continue, we need some definitions and a technical lemma. We say that a
rectangle R ∈ R↓ \ K blames a rectangle S ∈ K if S is the first rectangle (in
right-top order) of K having corner-intersection with R. We denote this rectangle as
S = blame(R). Also, given a set of points H, let hit(H, R↓ ) be the set of rectangles
in R↓ hit by H.
Lemma 5.4.1. Let H be a hitting set of K but not of R↓ and let R = Γ(a, b) be a
rectangle of R↓ \ K that is not hit by H. Let also S = blame(R) = Γ(c, d) ∈ K. Then:
C1. The rectangle U = Γ(c, b) is in K.
If furthermore, R is the last rectangle of R↓ \ K in right-top order that is not hit by
H then
C2. The rectangle T = Γ(a, d) ∈ R↓ is hit by H.
C3. If p is any point of H hitting U and q is any point of H hitting T then (p, q) is
an admissible flip for H.
C4. Let H 0 = H \ {p, q} ∪ {(px , qy ), (py , qx )} be the set obtained by flipping (p, q) as
above. Then hit(H, R↓ ) ∪ {R} ⊆ hit(H 0 , R↓ ).
136
Proof. Since S = Γ(c, d) and R = Γ(a, b) have corner intersection and S appears
before R in top-right order, we have cx < ax < dx < bx and dy > by > cy > ay (see
Figure 5-10). In particular, the rectangles T = Γ(a, d) and U = Γ(c, b) are in R. They
are also inclusion-wise minimal, otherwise there would be a point of A ∪ B \ {a, b, c, d}
in T ∪ U ⊆ R ∪ S, contradicting the minimality of R or S.
d d
Z2 S S r q Z2 ∩Z3
b0 Z3 b0
Z1 U Z1 Z1 p s U
c c c
T R
a
Figure 5-10: Positions of rectangles and zones defined in the proof of Lemma 5.4.2.
Assume by contradiction that the first claim does not hold. Then, there is a rect-
angle U 0 = Γ(c0 , b0 ) ∈ K appearing before U = Γ(c, b) and having corner-intersection
with U . In particular the corner c0 lies in the zone Z1 = (−∞, cx − 1] × [cy + 1, by − 1],
and the corner b0 lies in the zone Z2 = [cx + 1, bx − 1] × [by + 1, ∞) as shown on the
left of Figure 5-10.
Note that S and U 0 intersect since the top-left corner (cx , b0y ) of U is in both
rectangles. Since both S and U 0 are in K, this intersection must be corner-free.
Using that c0 ∈ Z1 we conclude that b0 is either equal to d or it lies in the zone
Z3 = [dx + 1, ∞) × [cy + 1, dy − 1]. See the center of Figure 5-10.
We have c0 ∈ Z1 and b0 ∈ {d} ∪ (Z2 ∩ Z3 ) (see the right of Figure 5-10). Therefore,
R and U 0 have corner-intersection contradicting the choice of S, since U 0 appears
before S in right-top order. This completes the proof of the first claim. Note also
that as U ∈ K, it must be hit by a point p ∈ H.
In what follows assume that R is the last rectangle in R↓ \ K that is not hit
by H. The second claim of the lemma follows by definition: since T appears after R
in right-top order, it must be hit by a point q ∈ H.
Now we show the third and fourth claims. Since p ∈ U \ R and q ∈ T \ R we have
px < ax ≤ qx and py ≤ by < qy . Let r = (px , qy ), s = (qx , py ) and suppose there is a
rectangle R̂ = Γ(â, b̂) ∈ R↓ hit by H but not by H 0 = (H \ {p, q}) ∪ {r, s}. If R̂ is hit
by p (but not by r or s), then b̂ must be in the region [px , qx − 1] × [py , qy − 1]. In
particular, b̂ ∈ S \ {c, d}. Therefore, Γ(c, b̂) is a rectangle in R that contradicts the
inclusion-wise minimality of S.
On the other hand, if R̂ is hit by q (but not by r or s), then â must be in the
region [px + 1, qx ] × [py + 1, qy ]. As before, this means that â ∈ S \ {c, d}, implying
that Γ(â, d) is a rectangle in R contradicting the inclusion-wise minimality of S.
We have shown that hit(H, R↓ ) ⊆ hit(H 0 , R↓ ). Noting that R is also hit by s ∈ H 0 ,
we conclude the proof of the lemma.
Let’s go back to the definition of the hitting sets H0 and H ∗ for K. As a corollary
of the previous lemma we have the following.
Lemma 5.4.2. H ∗ is a hitting set for R↓ , and therefore, a hitting set for R.
137
Proof. If this is not the case then, by condition C3 in Lemma 5.4.1, H ∗ admits an
admissible flip. This contradicts the construction of H ∗ .
Using Lemma 5.4.2, we can find a maximum cross-free matching and a minimum
biclique cover of a 2dorg G = (A, B, R), where (A, B) is a bicolored rook represen-
tation, using the procedure depicted below as Algorithm 14.
Algorithm 14 to compute a maximum independent set and a minimum hitting set

of R(A, B).
1: Compute R, R↓ and the c.f.i. greedy family K.
2: Use an algorithm for the Dilworth chain-partitioning problem to compute a max-
imum independent set R0 and a minimum hitting set H0 for K.
3: Compute H ∗ from H0 by flipping admissible flips while this is possible.
4: Return R0 and H ∗ .
Since H ∗ is a hitting set for R, and R0 is an independent set of R with |H ∗ | =

|H0 | = |R0 |, we conclude they are both optima; therefore, they induce a maximum
cross-free matching and a minimum biclique cover for G.
Theorem 5.4.3 (Soto and Telha [153]). Algorithm 14 computes a maximum cross-
free matching and a minimum biclique cover of the same size for any 2dorg G in
polynomial time. In particular, this shows that α∗ (G) = κ∗ (G).
It is necessary to remark that the ideas for this algorithm were already present
in Frank’s [62] algorithm for Györi’s [80] problem. In fact, by applying Algorithm 14
to families of rectangles arising from convex graphs, we recover Frank’s algorithm
(with different notation). Benczúr et al. [12] have given an efficient implementa-
tion of Frank’s algorithm. In the following section, we use the general lines of that
implementation to obtain a faster version of Algorithm 14.
5.4.1 Running Time Improvement

Consider in this section a fixed 2dorg G = (A, B, R) where (A, B) is a bicolored
rook representation. We use K to denote the greedy c.f.i. subfamily of R↓ obtained
by processing the rectangles in R↓ in right-top order, where a rectangle is added to K
only if it does not have a corner intersection with previously added rectangles. Let
n = |A ∪ B| be the number of vertices and m = |R| be the number of edges of G. Let
k = |K| and ` be the number of vertices and of edges in the intersection graph I(K).
5.4.2 Overview
Consider the following naive implementation of Algorithm 14.
138
First step. Construct R, R↓ and K: The family R can be constructed from (A, B)
in O(n2 ). We can construct R↓ naively in time O(|R|2 ) = O(m2 ) by checking every
pair of rectangles. In order to add a rectangle to the family K we potentially have
to check corner-intersection with all the previous rectangles. We can do this in time
O(|R↓ ||K|) = O(mk). The running time of this step is dominated by O(m2 ).
Second step. Construct R0 and H0 : Let f (k, `) be the running time of an algorithm
solving the Dilworth chain-partitioning problem on the poset (K, ). By applying
that algorithm, we recover a maximum antichain and a minimum chain partition of
this poset. Obtaining R0 from the maximum antichain returned is direct, but to
return the actual hitting set H0 from the chain partition, we need to select for each
chain, one point common to all the rectangles. Since each chain can have k rectangles,
this can be done in O(|H0 |k)-time. The running time of this step is then dominated
by O(f (k, `) + nk).
Third step. Computing H ∗ : A very naive way to do this is the following. We can
check if a pair of points is a flipping pair in time O(|K|) = O(k). Since there are
O(|H0 |2 ) = O(n2 ) possible flips, and the number of flips we require to do is bounded
by n3 , the total running time of this step is O(n5 k).
By using the trivial bounds m = O(n2 ), we conclude that this implementation

has running time O(f (k, `) + n5 k). The current best values for f (k, `) are based
on the deterministic algorithm of √ Hopcroft and Karp [86] for maximum bipartite
matching, running in time O(k+` k), and the randomized algorithm based on matrix
multiplication of Mucha and Sankowski [120] which runs in randomized time O(k ω )
where 2 ≤ ω ≤ 2.376 is the best exponent for the matrix multiplication problem (See
Lemma 3.2.3).
Using the bounds ` ≤ k 2 and k ≤ m ≤ n2 , we obtain that the total running time
of this naive algorithm is dominated by O(n7 ). We can reduce vastly the running time
by using the following strategy, which we describe in detail afterwards.
For the first step, we show how to construct an efficient data structure which
allows us to construct R, R↓ and K in O(n2 )-time.
For the third step, we make a simple observation. To construct a hitting set
for R↓ starting from H0 we do not need to perform every possible admissible flip:
It is enough to perform flips for which the current set hit(H, R↓ ) of hit rectangles
increases (one such flip is guaranteed by Lemma 5.4.1). As |R↓ | ≤ m ≤ n2 , we only
need O(n2 )-flips. In fact, we can do better. Going over the list of rectangles in R↓
in reverse top-right order, it is possible to find a sequence of at most k flips which
yields a hitting set for R↓ . We modify the proposed data structure in order to find
these flipping pairs efficiently. Finally, we show how to implement the flips, so that
each one of them takes time O(k log log n) by using van Emde Boas trees.
Using this approach, we get an O(n2 + (f (k, `) + nk) + (m + k 2 log log n)) running
time. To conclude, we give better bounds for k and ` in terms of n. Namely, we show
that ` = O(n2 ) and that k = O(n log n). Plugging in these bounds and the current
139
√
best bounds for f (k, `), we get that our algorithm runs in O(n5/2 log n)-deterministic
time, or O((n log n)ω + (n log n)2 log log n)-randomized time.
5.4.3 Data Structure

The following observation motivates a fast way to construct R↓ . For every point
a ∈ A, let List(a) be the collection of points v ∈ A ∪ B \ {a} such that Γ(a, v) is a
nonempty rectangle not containing a point of A ∪ B \ {a, v}. Observe that b ∈ B is
such that Γ(a, b) ∈ R↓ if and only if b ∈ List(a) ∩ B.
The points in List(a) form a staircase-shape as shown in Figure 5-11. They are
therefore easy to construct by visiting all the points in A ∪ B \ {a} from bottom to
top and remembering the leftmost visited one that is to the right of a.
Figure 5-11: Staircase shape of List(a).
We now show a fast way to test if a rectangle in R↓ is added to the greedy

c.f.i. family K. Consider the moment when a rectangle R = Γ(a, b) is processed in
the construction of K. Let L = Γ(a0 , b) be the last rectangle added to K before R is
considered having the same top right corner as R, this is tr(L) = b. This rectangle
may not exist, for instance in the case where R is the first rectangle considered
having b as a corner. If L exists, we call it the left-witness of R, and we denote it
as L = LW(R).
If L = LW(R) = Γ(a0 , b) exists, let T = Γ(a0 , b0 ) be the first rectangle in K added
after L with same bottom left corner as L, this is bl(T ) = a0 . The rectangle T may
not exist, for instance, if L is already the last rectangle with corner a0 . If T exists we
call it the top-witness of R, and we denote it as T = TW(R).
Lemma 5.4.4. The rectangle R = Γ(a, b) ∈ R↓ is not added to K if and only if
both rectangles L = LW(R) and T = TW(R) exist and the rectangle T has corner
intersection with R.
Proof. Sufficiency is direct, since the existence of a rectangle T ∈ K having corner
intersection with R implies the latter rectangle is not added to K.
To prove necessity, suppose that R is not added to K. The rectangle R blames
this to a rectangle blame(R) = Γ(c, d) which is the first rectangle in K having corner
intersection with R. By condition C1 of Lemma 5.4.1, Γ(c, b) is a rectangle with
right-top corner b, that is added to K before R is considered. This means that the
left witness L = LW(R) = Γ(a0 , b) exists (since it is the last rectangle added to K with
the previous property). For the next part of the proof, see Figure 5-12 for reference.
Assume for sake of contradiction that the top witness T = TW(R) does not exist
or that T has no corner intersection with R. This is equivalent to saying that there
140
b0
T b00 b0
R00 T
b b
L L
a0 a0
Z
p
a00 R R
a a
Figure 5-12: On the left, the situation if T has no corner intersection with R. On the
right, the situation imposed by Lemma 5.4.4.
is no rectangle with bottom-left corner a0 having corner intersection with R (if there
is one, then T must also have corner-intersection with R).
Recall that R was not added to K. This means that R must contain a bottom-
right corner of a previous rectangle in K in its topological interior int(R). Among
all the possible such bottom-right corners, let p = (px , py ) be the one having larger
x-coordinate, and, in case of ties, select the one with larger y-coordinate. Let R00 =
Γ(a00 , b00 ) ∈ K be the rectangle having p as its bottom-right corner (i.e. p = (b00x , a00y )).
Note that py = a00y must be strictly smaller than a0y . Indeed, if py > a0y , then p
would be in the interior of L, contradicting the fact that L is in K, and if py = a0y then
R00 and L would have the same bottom-left corner a0 , contradicting our assumption
for T . Also, we must have that a0x < a00x since otherwise the point a0 would be inside
the rectangle R00 , contradicting its minimality. See the left of Figure 5-12.
Let U = Γ(a00 , b) ∈ R. This rectangle is inclusion-wise minimal since otherwise,
it would contain a point in A ∪ B \ {a00 , b}, contradicting the minimality of R or R00 .
The discussion of the previous paragraph implies that U appears after L in right-top
order and has the same top-right corner b. We claim that U is also in K. This would
contradict the definition of L as left-witness and conclude the proof of the lemma.
Suppose that the claim does not hold. Then, U must contain the bottom-right
corner q of a previous rectangle in K in its interior. The point q must be in int(U ) \
(int(L) ∪ int(R00 )) as otherwise, L or R00 would not be in K. This implies that q is
in the region Z = [b00x , bx ) × (a00y , a0y ] (the dark area Z in the left of Figure 5-12). But
then q is a bottom-right corner of a rectangle in K, contained in int(R) and having
either higher x-coordinate than p, or having equal x-coordinate, but strictly larger
y-coordinate. This contradicts the definition of p.
The previous lemma states that in order to test if a rectangle R is added to K

we only need to check if its top witness (provided it exists) has corner-intersection
with R or not. This, together with the observation at the beginning of this section,
motivates Algorithm 15 to construct R↓ and K efficiently. This algorithm returns the
entire lists K and K = R↓ \ K. It also returns for every point v ∈ A ∪ B, two lists
K(v) and K(v) containing the rectangles in K and K respectively, having corner v.
All these lists are given in right-top order. Furthermore, every rectangle in K receives
two pointers to its left and top witnesses. In the description of Algorithm 15, last(·)
is a pointer to the last element of the corresponding list.
141
Algorithm 15 Data structure construction.
Require: Two subsets A and B in [n]2 , in rook representation.
Ensure: The associated collections K(·), K(·) and the function LW and TW.
1: Initialize empty list K and K.
2: Initialize empty lists K(v), K(v), for all v ∈ A ∪ B.
3: for each element a ∈ A, from left to right do
4: List(a) ← ∅
5: for each element v ∈ A ∪ B, with vy > ax , from bottom to top do
6: if (vx < ax ) or (vx > last(List(a))x ) then
7: continue to next element.
8: else Add v at the end of List(a).
9: end if
10: end for
11: for each element b ∈ List(a) in the order they were included do
12: if b ∈ B then
13: Let R = Γ(a, b) and L = Γ(a0 , b) ← last(K(b)).
14: if L is NULL then
15: Add R to K, K(a) and K(b).
16: else
17: Let T = Γ(a0 , b0 ) be the next rectangle in K(a0 ) after L.
18: if T is NULL or T has no corner-intersection with R then
20: else . T has corner intersection with R
22: Set LW(R) ← L, TW(R) ← T .
23: end if
24: end if
25: end if
26: end for
27: end for
Since we only manipulate pointers and perform comparisons between points in

the plane, we get the following corollary.
Lemma 5.4.5. Given two subsets A and B in [n]2 in rook representation, with n =
|A ∪ B|, Algorithm 15 runs in O(n2 )-time. Therefore, the first step of the refinement
of Algorithm 14 runs in O(n2 )-time.
5.4.4 Admissible Flips

We can use the left and top witnesses of rectangles in K as a tool to find admissible
flips. Let H be a hitting set of K. For every rectangle S ∈ K, let lastH (S) be the
point in H hitting S having maximum x-coordinate and, in case of ties, select the
one with larger y-coordinate. We call this point the last point in H hitting S.
142
Let R = Γ(a, b) ∈ K be such that all the rectangles in R↓ appearing after R in
right-top order are hit by H. Let L = LW(R) = Γ(a0 , b), T = TW(R) = Γ(a0 , b0 ),
p = lastH (L) and q = lastH (T ).
Lemma 5.4.6. Assume that p and q do not hit R. If qx < ax , then R is in hit(H, R↓ ).
Otherwise, the pair (p, q) is an admissible flip for H, and the set H 0 = (H \ {p, q}) ∪
{(px , qy ), (py , qx )} is such that hit(H, R↓ )∪{S ∈ K : LW(S) = LW(R)} ⊆ hit(H 0 , R↓ ).
Proof. Assume by contradiction that qx < ax and R is not hit by H. By definition
of q, we conclude that the area [ax , b0x ] × [a0y , b0y ] does not contain any point of H. In
particular, as R is not hit by H, the entire rectangle Γ(a, b0 ) is not hit by H. But this
is a contradiction since Γ(a, b0 ) is a rectangle in R↓ appearing after R in right-top
order, so it must be hit by H.
We have proved the first statement of the lemma. To prove the second one,
suppose that qx ≥ ax . Since p and q do not hit R, we must have that px < ax ≤ qx
and py ≤ by < qy as in Figure 5-13
b0
T r q
b
L p s
a0
S
a00 R
a
Figure 5-13: The admissible flip (p, q) and a rectangle S with LW(S) = LW(R) = L.
The proof that the flip (p, q) is admissible is similar to the proof of Lemma 5.4.2,
but we give it here for clarity.
Let r = (px , qy ), s = (qx , py ) and suppose there is a rectangle R̂ = Γ(â, b̂) ∈
hit(H, R↓ ) that is not hit by H 0 = (H \ {p, q}) ∪ {r, s}. If R̂ is hit by p (but not
by r or s), then b̂ must be in the region [px , qx − 1] × [py , qy − 1]. In particular,
b̂ ∈ T \ {a0 , b0 }, contradicting the inclusion-wise minimality of T .
If on the other hand R̂ is hit by q (but not by r or s), then â must be in the region
[px + 1, qx ] × [py + 1, qy ]. As before, this means that â ∈ T \ {a0 , b0 }, contradicting the
inclusion-wise minimality of T . Therefore (p, q) is an admissible flip.
Finally we show that every rectangle S ∈ R↓ \ K with the same left witness as R
is also hit by H 0 . For that, take one such S = Γ(a00 , b), with a00 6∈ {a, a0 }.
If the bottom-left corner of S is to the right of a, this is if a00x > ax , then S appears
after R in right-top order. By hypothesis, S is hit by H and also by H 0 .
Suppose that the previous does not happen. Since LW(S) = LW(R), we know
that a0x < a00x < ax and a0y > a00y > ay as in Figure 5-13. This means that S contains
the point s ∈ H 0 defined above. Therefore, S is also in hit(H 0 , R↓ ).
Consider the following procedure to find a sequence of flips that transforms any
hitting set H0 of K in a hitting set for R↓ . Go over the list of rectangles of K
in reverse top-right order. Check if the current rectangle R is hit by the current
set H by checking the positions of p = lastH (LW(R)) and q = lastH (TW(R)). If the
143
rectangle is hit, continue. Otherwise, flip p and q. By Lemma 5.4.6, when a flip is
performed every rectangle having the same left witness as the current observed one
is immediately hit by the new set H. Therefore, the number of flips this procedure
performs is at most the number of left witnesses |LW(K)| ≤ |K|. This procedure is
depicted as Algorithm 16
Algorithm 16 Flipping Algorithm.

Require: Lists K and K. Left and top witness functions LW, TW. A hitting set H0
for K.
Ensure: A hitting set H for R↓ .
1: H ← H0 .
2: for each rectangle R = Γ(a, b) ∈ K, in reverse top-right order do
3: Let p = lastH (LW(R)) and q = lastH (TW(R)).
4: if R contains p or q, or if qx < ax then
5: continue to next rectangle.
6: else Update H to (H \ {p, q}) ∪ {(px , qy ), (py , qx )}.
7: end if
8: end for
In order to run Algorithm 16 efficiently, we want to keep in each iteration and for
every rectangle R ∈ K, the collection of all points p in H hitting R. We also want
to have quick access to the last point in H hitting R. We manage this as follows.
Consider a collection of k priority queues, one per rectangle in K, each one accepting
priorities on the set {1, . . . , n2 }. Every one of the n2 possible points of the grid [n]2
is given the priority n(px − 1) + py . Recovering lastH (R) is equivalent to finding the
point p ∈ R ∩ H with the highest priority.
We start by populating the priority queue of R with the points in H contained
in R, and every time we update H, we update all the priority queues by adding
and removing at most two elements in each. There are different data structures for
this task. For example, by using heaps or self-balancing binary trees we can perform
insertions, deletions and maximum-priority search in time O(log |H|) = O(log n). By
using van Emde Boas trees [165], we can perform these operations faster, in time
O(log log n2 ) = O(log log n).
By using the fact that we require O(nk)-operations to initially populate the pri-
ority queues and that in total we perform at most k flips, we obtain as corollary the
following.
Lemma 5.4.7. Using one priority queue for each R ∈ K. Algorithm 16 perform at
most O(nk + k 2 ) insertions, deletions and maximum-priority search operations. It
also requires to do constant work for each rectangle in K. Therefore, the second and
third step of the refinement of Algorithm 14 runs in time
O(f (k, `) + m + (nk + k 2 ) log n),
144
using heaps or self-balancing binary trees. And it runs in time
O(f (k, `) + m + (nk + k 2 ) log log n),
using van Emde Boas trees.
5.4.5 Refined Algorithm

The improved version of Algorithm 14 is depicted below as Algorithm 17.
Algorithm 17 (improved version of Algorithm 14)

1: Run Algorithm 14 to compute R, R↓ and the c.f.i. greedy family K.
2: Use an algorithm for the Dilworth chain-partitioning problem to compute a max-
imum independent set R0 and a minimum hitting set H0 for K.
3: Use Algorithm 16 to compute a flipping set H ∗ for R↓ .
4: Return R0 and H ∗ .
5.4.6 Bounds for c.f.i. Families

Our algorithm is, at the moment, fully described. The last ingredient we require to
analyze its running time is to give better bounds for the number of vertices and edges
of the intersection graph of the c.f.i. family K.
For this section, let R↓ be the set of inclusion-wise minimal rectangles coming from
a 2dorg G = (A, B, R), where (A, B) is a bicolored rook representation. Let also K
be any c.f.i. subfamily of R↓ (not necessarily a greedy one). Define the collections
bl(K) and tr(K) of bottom-left corners of rectangles in K (the ones in A), and top-
right corners of rectangles in K (the ones in B) respectively. Finally, let k and ` be
the number of vertices and edges of the intersection graph I(K).
Lemma 5.4.8. If all the rectangles of K intersect a fixed vertical (or horizontal)
line λ, then
k ≤ |bl(K)| + |tr(K)| − 1. (5.32)
Proof. Project all the rectangles in K onto the line λ to obtain a collection of intervals
in the line. Since K is a c.f.i. family, all the intersections in K are skew-intersections
(see (5.13)). Therefore, the collection of intervals forms a laminar family: if two inter-
vals intersect, then one is contained in the other. Let X be the collection of extreme
points of the intervals. Since by assumption (A, B) is a bicolored rook representation,
|X| = |bl(K)| + |tr(K)| and furthermore, every interval is a non-singleton interval. To
conclude the proof of the lemma we only need to prove the following claim.
Claim: Any laminar family I of non-singleton intervals having extreme points
in X has cardinality at most |X| − 1. The claim follows by induction in |X|. For
|X| = 2, it holds trivially. For the general case, let I = [p, q] be a inclusion-wise
minimal interval in I. Note that only one of p or q can be an extreme point of an
interval in I \{I} as otherwise we would contradict laminarity. Therefore, the laminar
145
family I \ {I} has at most |X| − 1 extreme points. By induction hypothesis we get
that |I \ {I}| ≤ (|X| − 1) − 1, or equivalently, |I| ≤ |X| − 1. This concludes the proof
of the claim and the lemma.
Lemma 5.4.9. Let L = {λ1 , . . . , λr } be a collection of vertical lines sorted from left
to right that intersect all the rectangles in K, then
k ≤ n(1 + blog2 (r)c) − r. (5.33)
Proof. Consider the following collections of vertical lines:
L0 = {λ2k+1 : k ≥ 0} = {λ1 , λ3 , λ5 , λ7 , λ9 . . .}.

L1 = {λ4k+2 : k ≥ 0} = {λ2 , λ6 , λ10 , λ14 , . . .}.
L2 = {λ8k+4 : k ≥ 0} = {λ4 , λ12 , λ20 , λ28 , . . .}.
..
.
Lt = {λ2t (2k+1) : k ≥ 0}.
..
.
These collections of lines are disjoint and L0 , . . . , Lblog2 (r)c is a partition of L. Partition
the family K of rectangles as K0 ∪K1 ∪· · ·∪Kblog2 (r)c , where every rectangle is assigned
to the set with largest index containing a vertical line that intersects it. See Figure 5-
14 for an example illustrating this construction (in the example, K is a collection of
disjoint rectangles).
λ1 λ2 λ3 λ4 λ5 λ6 λ7
Figure 5-14: The dashed lines, thin lines and thick lines correspond to L0 , L1 and L2
respectively. White, light-gray and dark-gray rectangles are assigned to K0 , K1 and
K2 respectively.
Fix 0 ≤ t ≤ blog2 (r)c. Every rectangle in Kt intersects a unique line in Lt (if it

intersects two or more, then it would also intersect a line in Lt+1 ). For a given line
λ ∈ Lt , let Kλ be the subset of rectangles in Kt intersecting λ. By Lemma 5.4.9, the
number of rectangles in Kλ is at most |bl(Kλ )| + |tr(Kλ )| − 1. But it is easy to see that
every point a in A belongs to at most one set {bl(Kλ ) : λ ∈ Lt } (in fact, it belongs
to the first line λ which is on or to the right of a). Therefore, λ∈Lt |bl(Kλ )| ≤ |A|,
P
146
λ∈Lt |tr(Kλ )| ≤ |B|. Putting all together we get
P
and similarly,
X
|Kt | = |Kλ | ≤ |A| + |B| − |Lt | = n − |Lt |. (5.34)
λ∈Lt
Therefore,
blog2 (r)c
X
|K| = |Kλ | ≤ (1 + blog2 (r)c)n − |L|. (5.35)
t=0
This completes the proof.
Since n vertical lines are enough to intersect all rectangles in K, we conclude

that in any case k = O(n log n). Now, let us bound the number ` of edges of the
intersection graph I(K).
Lemma 5.4.10.
` = O(n2 ). (5.36)
Proof. Let Λ(n) be the maximum value that ` can take, for a given value of n.
Consider the vertical line λ = {(x, bn/2c) : x ∈ R}. This line divides the grid in
two roughly equal parts. Count the number of edges in (K, ) as follows. Let E1 be
the edges connecting pairs of rectangles that are totally to the left of λ, E2 be the
edges connecting pairs of rectangles that are totally to the right of λ, and E3 be the
remaining edges. Then, |E1 | ≤ Λ(bn/2c) and |E2 | ≤ Λ(dn/2e). We bound the value
of |E3 | in a different way.
Let K0 be the set of rectangles intersecting the vertical line λ, then E3 is exactly
the collection of edges in I(K) having an endpoint in K0 . By Lemma 5.4.9, |K0 | ≤ n.
Now we bound the degree of each element in K0 in the graph I(K). Consider one
rectangle R = Γ(a, b) ∈ K0 . Every rectangle intersecting R must intersect one of the
four lines defined by its sides. By using again Lemma 5.4.9, we conclude that the
total number of rectangles in K intersecting R is at most 4n. Therefore, Λ(n) satisfies
the recurrence
Λ(n) ≤ Λ(bn/2c) + Λ(dn/2e) + 4n2 . (5.37)
By solving this recurrence we get Λ(n) = O(n2 ), completing the proof.
To finish this section, we note that there are c.f.i. families K for which ` = Ω(n2 ).
See Figure 5-15 for an example. In this example, K was obtained using the greedy
procedure on R↓ . Also, if we let n0 = |A| = |B| = n/2, then k = 2n0 + 2 and
` = (n0 )2 + 4 = Ω(n2 ).
On the other hand, it is not clear if there are examples achieving k = Ω(n log n).
In particular, we give the following conjecture
Conjecture 5.4.11. For any c.f.i. family, k = O(n).
147
Figure 5-15: Example with ` = Ω(n2 ).
5.4.7 Conclusion
Combining Lemmas 5.4.5, 5.4.7, the bounds for k and ` obtained in the previous
subsection and the discussion about the running time of the algorithm of Hopcroft
and Karp and the one of Mucha and Sankowski, we obtain the following theorem.
Theorem 5.4.12. Algorithm 17 finds a maximum independent set and a minimum

hitting set for any collection of rectangles R(A, B) where (A, B) are in rook represen-
tation and n = |A ∪ B|, in
q
2.5
O n log n -deterministic time, or in
O((n log n)ω + (n log n)2 log log n)-randomized time,
using Hopcroft and Karp’s algorithm or the one of Mucha and Sankowski respectively,
where 2 ≤ ω ≤ 2.376 is the best exponent for the matrix multiplication problem. If
Conjecture 5.4.11 is true, these running times improve to
O(n2.5 )-deterministic time, and O(nω + n2 log log n)-randomized time.
In particular, if a 2dorg is given as G = (A, B, R) with (A, B) in rook representation,

then the jump-number of G, a maximum cross-free matching and a minimum biclique
cover of G can be computed in the same running times as above.
5.5 Relation to Frank and Jordan’s Set-Pair Cov-

ering Problem
In a seminal paper, Frank and Jordán [63] extend Györi’s result for independent
point-interval pairs (which, by Theorem 4.3.7 is equivalent to the min-max relation
between biclique covers and cross-free matchings for convex graphs) to set-pairs. We
briefly describe a particular case of their result that concern us.
148
A collection of pairs of sets {(Si , Ti )}i is half-disjoint if for every i 6= j, Si ∩ Sj or
Ti ∩ Tj is empty. A directed-edge (s, t) covers a set-pair (S, T ) if s ∈ S and t ∈ T .
A family S of set-pairs is crossing if whenever (S, T ) and (S 0 , T 0 ) are in S, so are
(S ∩ T, S 0 ∪ T 0 ) and (S ∪ T, S 0 ∩ T 0 ). Frank and Jordán prove that for every crossing
family S, the maximum size of a half-disjoint subfamily is equal to the minimum
size of a collection of directed-edges covering S. They also give a linear programming
based algorithm to compute both optimizers. Later, combinatorial algorithms for this
result were also given (e.g. [11]). See Végh’s Ph.D. thesis [166] for related references.
The min-max relation between biclique covers and cross-free matchings of 2dorgs
stated in Theorem 5.4.3, can be seen as a non-trivial application of Frank and
Jordán’s result. Given a 2dorg G = (A ∪ B, R), consider the family of set-pairs
S = {(Rx , Ry ) : R ∈ R↓ }. It is easy to check that this family is crossing, that half-
disjoint families of S correspond to independent sets in R↓ and that coverings of S
by directed-edges correspond to hitting sets for R↓ . We remark that this reduction
relies heavily on the geometric interpretation of 2dorgs we have presented in this
work and also that the presented proofs are self-contained and simpler than the ones
for Frank and Jordán’s result.
5.6 Summary of Results and Open Problems

In this chapter, we have extended the class of bipartite graphs for which the √ jump
2.5
number problem is efficiently solvable, devising algorithms running in O(n log n)-
deterministic time or O((n log n)ω +(n log n)2 log log n)-randomized time. These algo-
rithms compute not only the jump number of 2dorgs, but also a maximum cross-free
matching and a minimum biclique cover.
Previously, the largest class of graphs where the jump number problem had been
shown to be polynomially solvable was the class of convex graphs. In fact, for convex
graphs we have improved considerably the running time to compute the jump number
as before this work, only an O(n9 )-algorithm existed [40]. By relating our work with
a previous result by Györi [80], we have shown in Chapter 4 the existence of an
alternative O(n2 )-algorithm for convex graphs.
We have also shown a min-max relation between biclique covers and cross-free
matchings of 2dorgs. As the minimum biclique cover problem of a graph is equivalent
to the problem of computing the boolean rank of its adjacency matrix A, we have also
expanded the class of matrices for which the boolean rank can be computed exactly
in polynomial time.
We have also presented a min-max relation between hitting sets and independent
sets for the families of rectangles arising from 2dorgs. This result may be of interest
on its own.
Lastly, we have related the previous min-max relations to other relations arising
from apparently unrelated problems in combinatorial optimization: the minimum
rectangle cover and the maximum antirectangle of an orthogonal biconvex polygon,
studied by Chaiken et al. [25], the minimum base of a family of intervals and the
maximum independent set of point-interval pairs, studied by Györi [80]; and the min-
149
imum edge-cover and the maximum half-disjoint family of set-pairs, studied by Frank
and Jordán [63]. Our min-max relations can be seen in a certain way as both a gen-
eralization of Györi’s result and as a non-trivial application of Frank and Jordán’s
result.
We include Table 5.1 containing the current best running times for the studied
problems on a graph G = (A ∪ B, R), with n = |A ∪ B|. This table extends Table 4.1
presented in Chapter 4.
Table 5.1: Table of algorithmic results.
Bip. Perm. Biconvex Convex 2dorg

Max. cross-free matching O(n) [16, 53] O(n ) [16] O(n9 ) [40]
2
-
O(n2.5 )b
f
(Jump number) (new) - - O(n2 )a f ω )b
O(n
Min. biclique-cover O(n) [16, 53] O(n2 )a O(n2 )a -
O(n2.5 )b
f
(new) - - - f ω )b
O(n
a
These results follows from Theorem 4.3.7 and the algorithm of Franzblau and Kleit-
man [64] for geometric rectangle cover. The possibility of applying this algorithm to
compute the biclique cover of convex graphs has been noticed by Amilhastre et al. [2].
b
By Theorem 5.4.12.
Regarding unsolved questions in this area, the problem of determining the com-
plexity of the jump number problem for permutation graphs (two dimensional graphs)
remains open. Interestingly, as first observed by Ceroi [23], for two-dimensional posets
we can interpret the jump number as the problem of finding a maximum weighted in-
dependent set of an associated family of rectangles, for a given assignment of weights.
The fact that the rectangles are weighted makes solving this problem a very challeng-
ing task. In particular, the natural LP-relaxations, unlike the case of 2dorgs, have
non-unit integrality gap. Bounding this gap is also an interesting question.
150
Chapter 6
Weighted Cross-free Matching
In this chapter we consider the problem of finding the maximum weight cross-free
matching of a graph. We show this problem is NP-hard for 2dorgs, and give some
subclasses for which we can solve this problem in polynomial time. The problem is
defined as follows.
Problem 16 (Maximum Weight Cross-Free Matching Problem). Given a graph G =
(V, E), and a nonnegative weight function w : E → R+ , find a cross-free matching M
of maximum total weight.
For chordal bipartite graphs, this problem is equivalent to the maximum weight
jump number, defined by Ceroi [24] and, in the case where G = (A, B, R) is a 2dorg,
to the maximum weight independent set of rectangles in R.
6.1 NP-Hardness
Here we show the following result.
Theorem 6.1.1. The maximum weight cross-free matching problem is NP-hard for
2dorgs, even if the weights are all in {0, 1}.
Proof. To prove this result, we reduce from the maximum independent set of rect-
angles problem (MISRP), which is NP-hard even if the vertices of the rectangles are
all distinct [60]. Given an instance F of MISRP having the previous property, let A
(resp. B) be the set of bottom-left (resp. top-right) corners of rectangles in F. The
2dorg G = (A, B, R) satisfies F ⊆ R, so we can obtain the maximum independent
set of F by finding a maximum weight cross-free matching in G, where we assign a
weight of one to each R ∈ F, and a weight of zero to every other rectangle in R.
6.2 Polynomial Time Algorithms

6.2.1 Bipartite Permutation
We now provide an efficient algorithm for the maximum weight cross-free matching
of bipartite permutation graphs.
151
Recall that permutation graphs admit a rook representation induced by a permu-
tation π: the vertex i of the graph is mapped to the point (i, π(i)). In particular,
when the graph is bipartite, the previous representation is a bicolored rook represen-
tation (A, B) where all the points of A (resp. of B) form an antichain for ≤R2 . In
what follows assume that G = (A, B, R) is given to us in the way we just described.
We can solve the maximum weight independent set of R in O(n2 ) time using the
fact that the complement of the intersection graph I(R) is a comparability graph.
To see this, let us write R & S if R and S are disjoint and either Rx < Sx or Ry > Sy
holds, where we use the notation Px < Qx as shortcut for px < qx for all p ∈ P, q ∈ Q.
It is not hard to verify that D = (R, &) is a partial order whose comparability
graph is the complement of I(R), and that maximum weight cross-free matchings
in G correspond to maximum weight paths in the digraph D, using w as a weight
function on the vertex set R. Since D has |R| vertices and O(|R|2 ) arcs this optimal
path Q∗ can be found in O(|R|2 ) time (see, e.g. [38]).
We can find the optimal path Q∗ faster by exploiting the structure of D. For
simplicity, assume first that all the weights are different.
Let R & S & T be three consecutive rectangles in the optimal path Q∗ , then we
can extract information about the rectangle S. The following properties are easy to
verify.
Right-Right Case. If Rx < Sx and Sx < Tx , then
(i) Among all the rectangles located completely to the right of R and having
the same top-right corner tr(S), S is the heaviest.
Right-Down Case. If Rx < Sx and Sy > Ty , then
(ii) Among all the rectangles having the same bottom-left corner bl(S), S is the
heaviest.
Down-Right Case. If Ry > Sy and Sx < Tx , then
(iii) Among all the rectangles having the same top-right corner tr(S), S is the
heaviest.
Down-Down Case. If Ry > Sy and Sy > Ty , then
(iv) Among all the rectangles located completely below R and having the same
bottom-left corner bl(S), S is the heaviest.
See Figure 6-1 for examples of each case.
The previous properties motivate the following recursion.
Define T (v) for v ∈ A ∪ B as the rectangle of maximum weight (if any) having a
corner in v. Also, define T (b, b0 ), for b, b0 ∈ B, to be the rectangle (if any) of maximum
weight being both strictly to the right of the vertical line passing through b and having
top-right corner b0 . Similarly, define T (a, a0 ), for a, a0 ∈ A to be the rectangle (if any)
of maximum weight being strictly below the horizontal line passing through a and
having bottom-left corner a0 .
For R ∈ R, let V→ (R) (resp. V↓ (R)) be the maximum weight of a path in D starting
on R and having its second rectangle (if any) strictly to the right of R (resp. strictly
152
R R
S S
T
T
(a) Right-Right (b) Right-Down
R R
S S
T
T
(c) Down-Right (d) Down-Down
Figure 6-1: Three consecutive rectangles in Q∗ .
below R). Using properties (i)–(iv), we conclude that:
V (R) = max{V→ (R), V↓ (R)}

0 0
V→ (Γ(a, b)) = w(Γ(a, b)) + max max V→ (T (b, b )), 0 max V↓ (T (a )) .
b0 : b0x >bx 0 a : ax >bx
( )
0 0
V↓ (Γ(a, b)) = w(Γ(a, b)) + max max V↓ (T (a, a )), 0 max V→ (T (b )) .
a0 : a0y <ay 0 b : by <ay
Claim 2. We can compute the entire set of values above in O(n2 ).
To prove this claim, note that {T (a)}a∈A and {T (b)}b∈B can be computed in time
O(|R|) = O(n2 ). Also, by visiting the points a0 in A from bottom to top, we can
find T (a, a0 ) for each a ∈ A in time O(|A|) = O(n). This means that the entire table
{T (a, a0 )}a,a0 ∈A can be computed in O(n2 )-time and so can the table {T (b, b0 )}b,b0 ∈B .
Using those values, compute the quantities:
S(b) = 0 max
0
V→ (T (b, b0 )). S 0 (b) = 0 max
0
V↓ (T (a0 )).
b : b x>bx a : ax >bx
0 0
S(a) = 0 max
0
V↓ (T (a, a )). S (a) = 0 max
0
V→ (T (b0 )).
a : ay <ay b : b y<ay
Above, there are 2|A ∪ B| = O(n) quantities to compute, and since each one is
153
the maximum of at most n known values, the entire tables S(·) and S 0 (·) can be
computed in O(n2 ) time.
After precomputing all the previous values, it takes constant time to output each
entry of the tables V (·), V→ (·), and V↓ (·). As there are 3|R| = O(n2 ) of these values,
we conclude the proof of the claim.
After computing {V (R)}R∈R , we can easily obtain the weight of the optimal path
∗
Q as the maximum over all these values. It is also very simple to recover the actual
path Q∗ from the tables using the recurrences above.
Theorem 6.2.1. By using the dynamic algorithm described above, we can compute
the maximum weight cross-free matching of a bipartite permutation graph in O(n2 )
time.
6.2.2 Convex Graphs

By the discussion in Section 4.3.3 and especially by the proof of Theorem 4.3.7, cross-
free matchings of convex graphs correspond to independent sets of a certain system of
point-interval pairs. Lubiw [110] gives a polynomial time algorithm for the maximum
weight independent set of a system of point-interval pairs. It is straightforward to
implement her algorithm in O(n3 ) time.
We describe her algorithm in our terminology. Let G be a convex graph on A ∪ B
with labeling A = {a1 , . . . , ak }, where the neighborhood of b ∈ B is {a`(b) , . . . , ar(b) }.
We see G as an interval bigraph using Iai = {i} and Ib = [`(b), r(b)]. Let G =
(A, B, R) where (A, B) is the natural bicolored representation for interval bigraphs
described in Section 3.4.5. This representation puts the points A∪B on the triangular
grid U ≡ {(i, −j) ∈ Z2 : 0 ≤ j ≤ i ≤ |A|} with the points of A lying in the line
y = −x.
Since we are looking for a maximum weight independent set of rectangles, we can
assume that no two rectangles in R are equal geometrically. If this is the case, only
keep the one having maximum weight.
Given two disjoint rectangles R and S in R, let us say that R precedes S if either
Rx < Sx and Ry ∩ Sy 6= ∅ or if Ry < Sy and Rx ∩ Sx 6= ∅. It can be shown that this
relation is acyclic; therefore, the maximum weight independent set R∗ must contain a
rectangle T that no one precedes. The algorithm finds R∗ by guessing this rectangle T
and then recursively finding the maximum weighted independent set of the rectangles
lying strictly below T and the rectangles lying strictly to the left of T .
More precisely, for each point q in the grid U let V (q) be the weight of a maximum
independent set using only rectangles R ∈ R having bottom-left corner a ≤R2 q and
top-right corner b ≤R2 q. It is easy to check that
V (q) = max {w(a, q) + V (ax − 1, qy ) + V (qx , ay − 1)}.

a∈A : a≤R2 q
where w(a, q) is the weight of a maximum weight rectangle T contained in Γ(a, q) or

0 if there is no such rectangle. See Figure 6-2.
154
(ax − 1, qy ) q
T
a (qx , ay − 1)
y = −x
Figure 6-2: Example of Lubiw’s algorithm. V (q) equals the weight of rectangle T
plus the weight of the maximum independent set in both shaded regions.
All values {w(a, q)}a∈A,q∈U can be precomputed in time O(|A| · |U |) = O(n3 ).

Once we know those values, we can also find all the values of V (q) in time O(|A| ·
|U |) = O(n3 ) by dynamic programming. Finally, once {V (q)}q∈U is known, the value
V (|A|, 0) holds the weight of the maximum independent set. As usual, we can recover
the object achieving the maximum from the tables of values defined above.
Theorem 6.2.2 (Based on Lubiw’s algorithm [110]). By using the dynamic algorithm
described above, we can compute the maximum weight cross-free matching of a convex
graph in O(n3 ) time.
6.3 Open Problems

We have shown that the weighted cross-free matching is NP-complete for 2dorgs
and polynomially solvable for convex graphs. The most relevant open problem of this
chapter is the complexity of the weighted cross-free matching of interval bigraphs.
155
156
Part III
Constrained Set Function

Minimization
157
158
Chapter 7
Set Function Minimization under

Hereditary Constraints
In this chapter we present an efficient algorithm to find nonempty minimizers of

certain functions over any family of sets closed under inclusion. The function classes
we consider include symmetric submodular functions. Our algorithm makes O(n3 )
oracle calls to the submodular function where n is the cardinality of the ground set. In
contrast, the problem of minimizing a general submodular function
q under acardinality
constraint is known to be inapproximable within a factor of o n/ log n [159].
The results of this chapter are joint work with Michel Goemans. This chapter is
organized as follows. First, we offer an introduction to the problem and some back-
ground on previous work. Later, we explore the unconstrained problem of finding a
nontrivial minimizer of a set function. For the specific case of symmetric submod-
ular functions, this problem can be efficiently solved by a pendant pair technique
consolidated by Queyranne [141]; and Nagamochi and Ibaraki [127]. This technique
finds a collection of at most n candidate sets, containing a minimizer of the problem.
Later in the chapter, we define other classes of functions for which the pendant pair
technique can also be used.
Afterwards, we move to the problem of minimizing set functions under hereditary

constraints. We show how to modify the pendant pair technique in such a way that
the candidate sets are always inside the hereditary family, and that at least one
minimizer of the problem is declared as a candidate. By using ideas of Nagamochi
and Ibaraki [127] we modify our algorithm so that it outputs all the inclusion-wise
minimal minimizers of the problem.
After the completion of this work, we were informed of an independently discovered

algorithm of Nagamochi [124] which is able to perform a stronger feat for certain
classes of functions, including symmetric submodular functions. In the last part of
this chapter we compare our results.
159
7.1 Introduction
Let V be a finite ground set of size n. A set function on V is a real valued function
defined on all the subsets of V . A pair (V, f ) where f is a set function on V is called
a (set function) system. In this work we assume that the set function f is given
through a value oracle, this is an oracle that given a set S returns f (S).
A submodular function f is a set function on V , satisfying that for every pair of
sets A and B,
f (A ∪ B) + f (A ∩ B) ≤ f (A) + f (B). (7.1)
Examples of submodular functions include weight functions, the cut function of

a nonnegatively weighted graph and the entropy of a set of random variables. These
functions have applications in many areas, including game theory, information the-
ory and graph theory. Many important combinatorial optimization problems can be
formulated as finding a minimizer of a submodular function. For this reason the
following is considered a fundamental problem of the field.
Problem 17 (Submodular Function Minimization). Given a submodular function

f : 2V → R, find a subset X ∗ ⊆ V that minimizes f (X ∗ ).
Grötschel, Lovász and Schrijver [75, 76] show that the submodular function min-
imization problem can be solved using the ellipsoid method in strongly polyno-
mial time and using a polynomial number of oracle calls. Later, a collection of
combinatorial strongly polynomial algorithms have been developed by several au-
thors [57, 89, 88, 134, 146, 91]. The current fastest combinatorial algorithms known,
due to Iwata and Orlin [91] and Orlin [134] make O(n5 log n) and O(n5 ) function
oracle calls respectively, and run in O(n6 log n) and O(n6 ) time respectively, where n
is the size of the ground set.
There are faster algorithms available when the function f has more structure.
The case where f is symmetric is of special interest. In this case, we also require the
minimizer A∗ of f to be a nontrivial subset of V , that is ∅ ⊂ A∗ ⊂ V , otherwise
the problem becomes trivial since, by symmetry and submodularity, f (∅) = 21 (f (∅) +
f (V )) ≤ 12 (f (A) + f (V \ A)) = f (A), for all A ⊆ V .
The canonical example of a symmetric submodular function is the cut function
of a nonnegatively weighted graph. Finding a nontrivial minimizer corresponds in
this case to the minimum cut problem. Nagamochi and Ibaraki [125, 126] give a
combinatorial algorithm to solve this problem without relying on network flows. This
algorithm has been improved and simplified independently by Stoer and Wagner [158]
and Frank [61]. Queyranne [141] generalizes this and obtains a purely combinatorial
algorithm that minimizes a symmetric submodular function using only O(n3 ) function
oracle calls.
It is worth noting that if we impose simple additional constraints, minimizing
general submodular functions becomes intractable. For example, the problem of min-
imizing a submodular function
q over all the sets of cardinality at most k is NP-hard to
approximate within an o n/ log n factor, as shown by Svitkina and Fleischer [159].
160
In this work we show that this is not the case for symmetric submodular func-
tions. In Section 7.3 we extend Queyranne’s algorithm so that it returns a nontrivial
minimizer of any hereditary family, this is, a family that it is closed under inclusion.
In other words, the new algorithm solves the following problem.
Problem 18 (Hereditary Symmetric Submodular Function Minimization). Given a
symmetric submodular function f : 2V → R, and an hereditary family I ⊆ 2V , find
a nonempty subset X ∗ ∈ I that minimizes f (X ∗ ).
Common examples of hereditary families include
• Cardinality families: For k ≥ 0, the family of all subsets with at most k
elements: I = {A ⊆ V : |A| ≤ k}.
• Knapsack families: Given a weight function w : V → R+ , consider the family
of all subsets of weight at most one unit: I = {A ⊆ V : v∈A w(v) ≤ 1}.
P
• Matroid families: Given a matroid M over V , consider the family of inde-

pendent sets of M.
• Hereditary graph families: Given a graph G = (V, E), consider the fam-
ily of sets S of vertices such that the induced subgraph G[S] satisfies certain
hereditary property such as being a clique, being triangle-free, being planar or
exclude a given list of minors.
• Matching families: Given a hypergraph H = (V, E), consider the family of
matchings of H, that is sets of edges that are pairwise disjoint.
Since the intersection of hereditary families is hereditary, the hereditary minimization
problem is very broad. This problem includes the following examples.
1. Find a minimum unbalanced cut in a graph; that is for given k, find among all
nonempty sets of at most k vertices, the one inducing the minimum cut.
2. More generally, given a nonnegatively weighted graph, find a nonempty induced
subgraph satisfying an hereditary graph property (e.g. triangle-free, clique,
stable-set, planar) minimizing the weights of the edges having precisely one
endpoint in the subgraph.
For the unrestricted problem (equivalently for the case where I = 2V \ {V }),
Nagamochi and Ibaraki [127] present a modification of Queyranne’s algorithm that
finds all inclusion-wise minimal minimizers of a symmetric submodular function while
still using a cubic number of oracle calls. Using similar ideas, we can also list all
minimal solutions of an hereditary minimization problem.
The algorithms we present are not restricted to work on symmetric submodular
functions. In Section 7.4 we explore other function classes where our methods can
still be applied. For instance, we can solve the hereditary problem for functions f
that are restrictions of a symmetric submodular function (also known as submodular-
posimodular functions) or when f (A) is defined as d(A, V \ A) for a monotone and
consistent symmetric bi-set map d in the sense of Rizzi [145].
161
Other related work. Constrained submodular function minimization problems,
i.e. the minimization of a submodular function over subfamilies of 2V , have been
studied in different contexts. Padberg and Rao [138] show that the minimum odd cut
problem obtained by restricting the minimization over all odd sets can be solved in
polynomial time. This was generalized to submodular functions over larger families
of sets (satisfying certain axioms) by Grötschel, Lovász and Schrijver [76] and by
Goemans and Ramakrishnan [73]. This covers for example the minimization over
all even sets, or all sets not belonging to a given antichain, or all sets excluding all
minimizers (i.e. to find the second minimum). For the particular case of minimizing a
symmetric submodular function under cardinality constraints the best previous result
is a 2-approximation algorithm by Shaddin Dughmi [46]. Recently, Goel et al. [71]
have studied the minimization of monotone submodular functions constrained to sets
satisfying combinatorial structures on graphs, such as vertex covers, shortest paths,
perfect matchings and spanning trees, giving inapproximability results and almost
matching approximation algorithms for them. Independently, Iwata and Nagano [90]
study both the vertex and the edge covering version of this problem.
The algorithm of Nagamochi and Ibaraki [127] also works with functions satisfy-
ing a less restrictive symmetry condition. Narayanan [130] shows that Queyranne’s
algorithm can be used to minimize a wider class of submodular functions, namely
functions that are contractions or restrictions of symmetric submodular functions.
Rizzi [145] has given further extension for a different class of functions.
Nagamochi [124] has recently given an algorithm to find all extreme sets of a
symmetric submodular function, where a set is called extreme if its function value is
strictly smaller than any one of its nontrivial subsets. As we show in Section 7.5, this
algorithm can also be used to solve the hereditary minimization problem.
7.2 Unconstrained Minimization

In this section we review Queyranne’s algorithm [141] to find nontrivial minimizers
of certain systems (V, f ). To simplify notation, we use the convention that for any
set A ⊆ V , any x ∈ A, and any y ∈ V \ A; A + y and A − x stand for the sets A ∪ {y}
and A \ {x} respectively. Also, for x ∈ V we use f (x) to denote f ({x}).
A useful operation for set function systems is the following. Let Π be a partition
of V . For every collection of parts X ⊆ Π, we use VX to denote the set of elements
contained in the union of the parts in X, this is
[
VX = S ⊆ V. (7.2)
S∈X
The fusion of f relative to Π, denoted by fΠ is the set function on Π given by
fΠ (X) = f (VX ). (7.3)
We say that (V 0 , f 0 ) is a fusion of (V, f ) if V 0 corresponds, up to renaming, to a

partition of V and f 0 is the fusion of f relative to this partition.
162
Given a system (V, f ), we say that we fuse a collection of elements X ⊆ V into a
single element s if we replace (V, f ) by the system ((V \ X) + s, fΠ ), where Π is the
partition that has one part equal to X (which, by abusing notation we denote as s)
and where all the other parts are singletons (which again, by abusing notation, keep
the same name as the unique element they contain).
Queyranne’s technique performs iterative fusions on the original system (V, f ). To
keep our explanation simple, we overload the notation above by saying that for any
fusion (V 0 , f 0 ) of (V, f ) and any a ∈ V 0 , Va is the set of elements in V that have been
fused into a, and for every set A ⊆ V 0 , VA is the union of all sets Va with a ∈ A.
We say that a set X ⊆ V separates two elements t and u of V , if X contains
exactly one of t and u. We extend this notion to fusions by saying that two elements
t and u in V 0 are separated by a set X ⊆ V if Vt ⊆ X and Vu ⊆ V \ X or vice versa.
The following concept is crucial for the development of Queyranne’s technique.
An ordered pair (t, u) of elements of V is called a pendant pair for f in V if {u}
has the minimum f -value among all the subsets of V separating u and t, this is,
f (u) = min{f (U ) : U ⊂ V, |U ∩ {u, v}| = 1}. (7.4)
We say that a system (V, f ) is admissible for pendant pairs (or simply, ad-
missible) if for every fusion (V 0 , f 0 ) with |V 0 | ≥ 2 there exists a pendant pair (t, u)
for f 0 in V 0 .
Suppose that (V, f ) is an admissible system and that (t, u) is a pendant pair for f
in V . Let X ∗ be a nontrivial minimizer of (V, f ). Then we have two cases depending
on whether X ∗ separates t and u or not.
If X ∗ separates t and u, then by definition of pendant pairs, f (u) ≤ f (X ∗ ), and
so {u} is also a nontrivial minimizer.
If this is not the case, consider the set system (V 0 , f 0 ) obtained by fusing u and v
into a single element uv. Any nontrivial minimizer X 0 of this system induces a
nontrivial minimizer VX 0 of (V, f ).
By recursively applying the argument above n − 1 times (as all the fused systems
have pendant pairs) we can find a nontrivial minimizer of (V, f ). The procedure is
described in Algorithm 18 below, which we call Queyranne’s algorithm.
Algorithm 18 (Queyranne’s algorithm).

Require: An admissible system (V, f ).
Ensure: A nontrivial minimizer X ∗ for (V, f ).
1: Let (V 0 , f 0 ) = (V, f ), and C ← ∅. . C is the set of candidates.
0
2: while |V | ≥ 2 do
3: Find any pendant pair (t, u) for f 0 in V 0 .
4: Add Vu to C. . Vu is the set of elements of V that have been fused into u.
0 0
5: Update (V , f ) by fusing {t, u} into a single element tu.
6: end while . |V 0 | = 1
7: Add Vu to C, where u is the only element of V 0 .
8: Return the set X ∗ in C with minimum f -value.
163
The argument above implies that Queyranne’s algorithm is correct. The only non-
trivial step of the algorithm correspond to finding pendant pairs of fusions. Suppose
in what follows that we have access to an algorithm A that computes pendant pairs
for any fusion (V 0 , f 0 ) in O(T (|V 0 |))-time and using O(T (|V 0 |)) calls to some value
oracle, where T (·) is an increasing function.
Lemma 7.2.1. By using A as a subroutine, Queyranne’s algorithm returns a non-
trivial minimizer of (V, f ) in O(nT (n))-time and using the same asymptotic number
of oracle calls.
Queyranne originally devised this algorithm for symmetric submodular functions.
He showed not only that these functions are admissible, a fact originally shown by
Mader [111], but also that there is an efficient algorithm to compute pendant pairs.
Let f be a symmetric submodular function on V . Consider an ordering (v1 , . . . , vn )
of the elements of V , such that
f (Wi−1 + vi ) − f (vi ) ≤ f (Wi−1 + vj ) − f (vj ), for all 2 ≤ i ≤ j ≤ n, (7.5)
where v1 can be chosen arbitrarily and Wi denotes the set {v1 , . . . , vi }. An ordering
satisfying (7.5) is called a maximum adjacency ordering. Queyranne [141] shows
Lemma 7.2.2 (Queyranne [141]). For a symmetric submodular function f on V , and
an arbitrarily chosen element v1 ∈ V , the last two elements (vn−1 , vn ) of a maximum
adjacency ordering of V starting from v1 constitute a pendant pair. Furthermore, this
ordering can be found using O(n2 ) oracle calls and in the same running time.
We do not prove Queyranne’s lemma here as in Section 7.4 we show an extension
to more general functions. By using the lemmas above and the fact that fusions
of symmetric submodular functions are also symmetric submodular, Queyranne has
shown the following.
Theorem 7.2.3 (Queyranne [141]). The problem of finding a nontrivial minimizer
of a symmetric submodular function f on V can be solved in O(n3 ) time and using
O(n3 ) oracle calls to f .
In the next section, we extend Queyranne’s algorithm to the problem of finding
nontrivial minimizers of set functions under hereditary constraints.
7.3 Constrained Minimization

A triple (V, f, I) where f is a set function on V and I is an hereditary family of V , is
called an hereditary system. Since hereditary families can be exponentially large
in the size of the ground sets, we assume that we access I through an oracle that
given a set A, answers whether A is in the family or not.
We say that a set A is a minimal optimal solution for (V, f, I) if A is an
inclusion-wise minimal minimizer of the function f over the nontrivial sets in I.
164
Similarly, we say that A is a minimal optimal solution for (V, f ) if A is minimal
optimal for (V, f, 2V \ {V }).
We extend the notion of fusion to hereditary systems as follows. Given a partition
Π of V , the fusion of I relative to Π, denoted by IΠ is the family
IΠ = {I ⊆ Π : VI ∈ I}.
It is easy to see that if I is hereditary then so is IΠ . The fusions of an hereditary

system (V, f, I) are defined analogously to the fusions of (V, f ). It is worth noting at
this point that if (V, f 0 , I 0 ) is a specific fusion of (V, f, I) then we can test if A ∈ I 0
and evaluate f 0 (A) using only one oracle call to I or f respectively.
A system (V, f ) is strongly admissible for pendant pairs (or simply, strongly
admissible) if for every fusion (V 0 , f 0 ) where V 0 has at least 3 elements, and every
v ∈ V 0 , there is a pendant pair (t, u) of V 0 avoiding v (this is, t 6= v and u 6= v). We
say that (V, f, I) is strongly admissible if (V, f ) is.
Queyranne’s lemma (Lemma 7.2.2) implies that (V, f ) is a strongly admissible
system when f is a symmetric submodular function. In Section 7.4 we give other
examples. The following lemma is of interest on its own.
Lemma 7.3.1. The minimal optimal solutions of any strongly admissible system
(V, f, I) are pairwise disjoint.
Proof. The lemma holds trivially if V has at most 2 elements. So assume |V | ≥ 3.

Suppose that two minimal optimal solutions A and B are not disjoint. Since no one
can include the other, we have two cases, either A ∪ B = V , or V \ (A ∪ B) 6= ∅.
In the first case, consider the system (V 0 = {a, b, c}, f 0 ) obtained from (V, f ) by
fusing A \ B into a single element a, B \ A into b and A ∩ B into c. Since (V, f )
is strongly admissible there is a pendant pair for f 0 in V 0 avoiding c. Without loss
of generality assume this pair is (a, b). Therefore f 0 (b) ≤ f 0 ({a, c}) or equivalently,
f (B \ A) ≤ f (A) = f (B), contradicting the inclusion-wise minimality of B.
For the second case, consider the system (V 0 = {a, b, c, d}, f 0 ) obtained by fusing
A \ B into a, B \ A into b, A ∩ B into c and V \ (A ∪ B) into d. Since (V, f ) is
strongly admissible there is a pendant pair for f 0 in V 0 avoiding d. If the pendant
pair also avoids c, then without loss of generality we can assume this pair is (a, b). As
above, this means that f 0 (b) ≤ f 0 ({a, c}) or equivalently, f (B \ A) ≤ f (A) = f (B),
contradicting the choice of B. Therefore the pendant pair contains c.
If c is the first element of the pair, we can assume without loss of generality that
the pair is (c, b). Hence, f 0 (b) ≤ f 0 ({a, c}) which is a contradiction as shown above.
On the other hand, if c is the second element of the pair, we can assume this pair
is (b, c). Hence, f 0 (c) ≤ f 0 ({a, c}) or equivalently f (A ∩ B) ≤ f (A), contradicting the
minimality of A.
As every case leads to a contradiction, we conclude that the minimal optimal
solutions are disjoint.
In what follows we present two algorithms: one to find a particular minimal

optimal solution of a strongly admissible hereditary system (V, f, I), and another to
165
find all minimal optimal solutions. The previous lemma implies that there are at
most n of them.
Both of the mentioned algorithms are direct extensions of the algorithms presented
by Nagamochi and Ibaraki [127]. In fact by setting I to be the hereditary family of
sets not containing a particular element s, we recover their algorithms.
If we just use Queyranne’s algorithm on f , we could introduce candidates that
are not in the hereditary family. In order to avoid that, we first fuse all the loops1 , if
any, of I into a single loop s, and proceed to find a pendant pair not containing it.
In this way, we ensure that every selected candidate belongs to I. If the hereditary
family has no loops, then we just use Queyranne’s algorithm until a loop s is created.
From that point on we continue as before. The complete procedure is depicted below
as Algorithm 19. We defer the problem of finding pendant pairs to the next section.
For now we assume the existence of an algorithm A that given an element s is able
to find a pendant pair avoiding s.
Algorithm 19 FindMinimalOptimal (V, f, I)

Require: A strongly admissible hereditary system (V, f, I).
Ensure: A minimal optimal set X ∗ for the hereditary minimization problem.
1: Set (V 0 , f 0 , I 0 ) ← (V, f, I), and C ← ∅. . C is the set of candidates.
2: while I 0 has no loops do
3: Find any pendant pair (t, u) of f 0 .
4: Add Vu to C. . Vu is the set of elements of V that have been fused into u.
5: Update (V 0 , f 0 , I 0 ) by fusing {t, u} into a single element tu.
6: end while . I 0 has at least one loop.
0 0 0
7: Let (V + s, f , I ) be the system obtained by fusing all the loops of I into a single
element denoted s (keep s as an element outside V 0 ).
8: while |V 0 | ≥ 2 do
9: Find a pendant pair (t, u) of f 0 not containing s.
10: Add Vu to C.
11: if {t, u} ∈ I 0 then
12: Update (V 0 + s, f 0 , I 0 ) by fusing {t, u} into a single element tu.
13: else
14: Update (V 0 + s, f 0 , I 0 ) by fusing {s, t, u} into a single element called s.
15: end if
16: end while
17: if |V 0 | = 1 (say V 0 = {u}) then
18: Add Vu to C.
19: end if
20: Find the sets in C of minimum f -value. Among those, return the set X ∗ that was
added earlier to C.
Theorem 7.3.2. Algorithm 19 outputs an optimal solution of any strongly admissible

hereditary system (V, f, I).
1
A loop of an hereditary family is a singleton that is not in I.
166
Proof. By induction, we can check that at the beginning of each iteration, either I 0
is loopless or s is its only loop. Every candidate Vu introduced in C is obtained from
a non-loop u. Therefore, the output set X ∗ is in I.
Suppose for contradiction that there is a nonempty set Y ∈ I such that f (Y ) <
f (X ∗ ). Assume first, that this set does not separate any pendant pair found in the
execution of the algorithm. Since the algorithm only fuses pendant pairs and loops
we have that at every iteration and for every element w in the current ground set,
the associated set Vw ⊆ V is either completely inside Y or completely outside Y . In
particular, since Vs 6∈ I, Vs is always completely outside Y . Therefore, at the end of
the algorithm, Y must be equal to the set Vu in line 18 and so it is included in the
set of candidates, contradicting the definition of X ∗ .
Consider then the first pendant pair (t, u) found by the algorithm that is separated
by Y . By the property of pendant pairs, f 0 (u) ≤ f (Y ) for the function f 0 at that
iteration. But then, the set Vu ∈ V of elements that were fused into u is a candidate
considered by the algorithm. Thus, f (X ∗ ) ≤ f (Vu ) = f 0 (u) ≤ f (Y ), which contra-
dicts our assumption. Therefore X ∗ has minimum f -value among the nontrivial sets
of I.
Furthermore, since we choose X ∗ as the set that is introduced first into the family
of candidates C (among the ones of minimum value), then this set X ∗ is also an
inclusion-wise minimal minimizer of (V, f, I). Indeed, if there is a set Y ∈ I such
that f (Y ) = f (X ∗ ), with ∅ 6= Y ⊂ X ∗ , then Y separates two elements of X ∗ .
This means that at some moment before the introduction of X ∗ as a candidate, the
algorithm finds a pendant pair (t, u) separated by the set Y with both t and u in X ∗ .
At this iteration, the candidate Vu is such that f (Vu ) = f (Y ) = f (X ∗ ), which is a
contradiction since Vu is introduced before X ∗ to the set of candidates.
Suppose that a given algorithm A computes pendant pairs of (V 0 , f 0 ) avoiding a
given element in O(T (|V 0 |))-time and using O(T (|V 0 |)) oracle calls, where T (·) is an
increasing function. Then we have the following.
Theorem 7.3.3. By using A as a subroutine, Algorithm 19 returns a minimal optimal
solution of the strongly admissible hereditary system (V, f, I) in O(nT (n))-time and
using the same asymptotic number of oracle calls.
Proof. Direct from the fact that each iteration decreases the cardinality of V 0 by one
or two units.
We can use the fact that the minimal optimal solutions are disjoint to find all of
them. We first compute one particular minimal optimal solution X ∗ of the system
and fuse it into a single element s which we will consider a loop for the new family.
Then we run the algorithm again in such a way that, every time a optimal solution X
is found we fuse X + s into s in order to avoid finding solutions having nonempty
intersection with X. A naive implementation of this procedure requires O(n) calls to
Algorithm 19. A better implementation is described in Algorithm 20.
Theorem 7.3.4. Algorithm 20 outputs all minimal optimal solutions of the strongly
admissible hereditary system (V, f, I).
167
Algorithm 20 FindAllMinimalOptimals (V, f, I)
Require: A strongly admissible hereditary system (V, f, I).
Ensure: The family F of minimal optimal solutions for the hereditary minimization
problem.
1: Using FindMinimalOptimal, compute a minimal optimal solution X ∗ for (V, f, I).
Set λ∗ ← f (X ∗ ).
2: Let (V 0 + s, f 0 , I 0 ) be the system obtained by fusing X ∗ and all the loops of I
into a single element, denoted s. (During the execution of the algorithm, keep s
as an element outside V 0 .)
3: I 0 ← I 0 \ {A ∈ I 0 : s ∈ A}. . If s is not a loop, consider it as one.
∗
4: Let F = {X }.
5: for each v ∈ V with f 0 (v) = λ∗ do
6: Add {v} to F.
7: Update (V 0 + s, f 0 , I 0 ) by fusing {s, v} into s.
8: end for
9: while |V 0 | ≥ 2 do . f 0 (v) > λ∗ for all v ∈ V 0 , and s is the only loop of I 0 .
10: Find a pendant pair (t, u) of f 0 not containing s.
11: if {t, u} ∈ I 0 and f 0 ({t, u}) = λ∗ then
12: Add V{t,u} to F.
13: Update (V 0 + s, f 0 , I 0 ) by fusing {s, t, u} into s.
14: else if {t, u} ∈ I 0 and f 0 ({t, u}) > λ∗ then
15: Update (V 0 + s, f 0 , I 0 ) by fusing {t, u} into tu.
16: else . {t, u} 6∈ I 0 .
0 0 0
17: Update (V + s, f , I ) by fusing {s, t, u} into s.
18: end if
19: end while
20: Return the family F.
Proof. Since all the solutions added to F are disjoint and have minimum f -value, it is
enough to show that every minimal optimal solution is eventually added to F. Sup-
pose that this is not the case, and let Y ∈ I be an inclusion-wise minimal minimizer
of (V, f, I) with Y 6∈ F.
We first claim that at every moment and for every w ∈ V 0 +s, the associated set Vw
is always completely inside or completely outside Y . We prove this by induction. The
claim is true at the beginning of the algorithm and after the fusion of all loops and X ∗ .
Since the minimal optimal solutions are disjoint, this also holds after all the optimal
singletons are added to F and fused into s.
Suppose that the claim holds at the beginning of an iteration in the while-loop and
let (t, u) be the pendant pair found at that moment. Note that Y can not separate t
from u, since in that case we would have f 0 (u) = f (Y ) = λ∗ . But, by construction,
the algorithm ensures that at every iteration the singletons are not optimal, i.e.,
f 0 (v) > λ∗ for every v ∈ V 0 , contradicting the previous sentence. It follows that
Vt and Vu are both either completely inside or both completely outside Y . If all the
elements participating in a fusion at this iteration are completely inside or completely
168
outside Y then the claim still holds at the end of the iteration. The only case left to
consider is when Vt and Vu are inside Y , Vs is outside and we fuse {s, t, u} together
into s. We only do this when {t, u} ∈ I 0 and f 0 ({t, u}) = λ∗ or when {t, u} 6∈ I 0 .
As V{t,u} ⊆ Y ∈ I, we must be in the first case. Then, according to the algorithm,
V{t,u} = Vt ∪ Vu is added to F. By minimality of Y we obtain Y = V{t,u} which
contradicts the fact that Y 6∈ F and proves the claim.
Since Y is never added to F, and at every moment Vs ⊇ X ∗ is completely out-
side Y , the previous claim implies that after the while-loop, the set Y corresponds
to the unique element in V 0 , say Y = Vu , for V 0 = {u}. As we maintained that the
singletons cannot be optimal, we get a contradiction.
Similar to the case of Algorithm 19, if algorithm A is able to compute pendant

pairs of (V 0 , f 0 ) avoiding a given element in O(T (|V 0 |))-time and using O(T (|V 0 |))
calls to some oracle, for some increasing function T (·), we conclude that.
Theorem 7.3.5. By using A as a subroutine, Algorithm 20 returns all minimal

optimal solutions of the strongly admissible hereditary system (V, f, I) in O(nT (n))-
time and using the same asymptotic number of oracle calls.
Proof. Direct from the fact that each iteration decreases the cardinality of V 0 by at
least one unit.
7.4 Set and Bi-set Functions

In this section, we study different families of strongly admissible set functions and
show how to find pendant pairs in all of them. The families presented are not new
and can be found in different articles related to Queyranne’s symmetric submodular
function algorithm [127, 145, 130].
Let V be a finite ground set of size n and Q(V ) be the collection of disjoint pairs
of subsets of V ,
Q(V ) = {(A, B) : A, B ⊆ V, A ∩ B = ∅}.
A bi-set function on V is a function d : Q(V ) → R. Just like the case of set

functions, we assume that bi-set functions are given through a value oracle, that is
an oracle that given a pair of disjoint sets (A, B), returns d(A, B).
A bi-set function d on V is symmetric if for all A ⊆ V , d(A, B) = d(B, A).
Every symmetric bi-set function admits a canonical symmetric set function and
vice versa. Given a symmetric bi-set function d on V , the symmetric set function f (d)
is defined as:
f (d) (A) = d(A, V \ A), for all A ⊆ V .
Given a (not necessarily symmetric) set function f on V , the symmetric bi-set function
169
d(f ) is defined2 as:
1
d(f ) (A, B) = (f (A) + f (B) + f (∅) − f (A ∪ B)) , for all A 6= B.
2
(f ) )
Lemma 7.4.1. If f is a symmetric function on V , then f = f (d .
Proof. Let d = d(f ) , and fˆ = f (d) , then for all A ⊆ V we have

1
fˆ(A) = d(A, V \ A) = (f (A) + f (V \ A) + f (∅) − f (V )) = f (A).
2
(d) )
However, even for symmetric d, the functions d and d(f can be extremely dif-
ferent. For example, consider the non constant function
d(A, B) = |A| + |B|, for all (A, B) ∈ Q(V ).

(d)
then f (d) is a constant function equal to n, and d(f ) is also constant equal to n.
We conclude this section with a natural example of a set and bi-set function. Let
G = (V, E, w) be a weighted graph with w : E → R. For all pairs (A, B) of disjoint
sets, define E(A : B) as the set of edges having one endpoint in A and one endpoint
in B. The cut between sets A and B is defined as
dG (A, B) = w(E(A : B)), for all (A, B) ∈ Q(V );
and the cut function of G (also called the weighted degree function of G) is
fG : V → R defined as
fG (A) = w(E(A : V \ A)), for all A ⊆ V .
It is easy to see that f (dG ) = fG and d(fG ) = dG .
7.4.1 Fusions and Minors

We can extend the notion of fusion to bi-set functions. Let Π be a partition of V
into a collection of sets. The fusion of d relative to Π, denoted by dΠ is the bi-set
function on Π given by
dΠ (X, Y ) = d(VX , VY ), (7.6)

S
where we recall that VX = S∈X S. Note that if d is symmetric then so are all their
fusions.
We focus now on certain operations on set functions. Consider a set function f
on V and let S ⊆ V and S = V \ S. The function obtained from f by deleting S,
2
It is also possible to remove the term f (∅) or change its sign in the definition. But this definition
is better behaved.
170
also known as the restriction of f to S, is defined as
f \ S = f.S : 2S → R, where f.S(A) = f (A), for all A ⊆ S. (7.7)
The function obtained from f by contracting S, also known as the contraction of

f to S, is defined as
f /S = f × S : 2S → R, where f × S(A) = f (A ∪ S) − f (S), for all A ⊆ S. (7.8)
We say that a restriction or contraction is non-trivial if the associated set S above

satisfies ∅ ⊂ S ⊂ V .
It is easy to see that deletion and contraction commute, that is, if S, T are disjoint
subsets of V , then (f /S) \ T = (f \ T )/S. Any function obtained from f by deleting
and contracting subsets is called a minor of f .
Let f be an arbitrary set function on V and s be an element outside V . The
anti-restriction of f with extra element s is the function g on V + s such that

f (A), if s 6∈ A,
g(A) =  (7.9)
f (V \ A), if s ∈ A.
The anti-contraction of f with extra element s is the function h on V + s such

that

f (V \ A), if s 6∈ A,
h(A) = (7.10)
f (A − s), if s ∈ A.
Lemma 7.4.2.
1. Let g be the anti-restriction with extra element s of a set function f on V , then

g \ {s} = f
2. Let h be the anti-contraction with extra element s of a set function f on V with

f (∅) = 0, then h/{s} = f .
Proof. The first item follows since by definition, for all A ⊆ V , g \ {s}(A) = g(A) =
f (A). The second item holds since for all A ⊆ V , h/{s}(A) = h(A + s) − h(s) =
f (A) − f (∅).
An application of anti-restrictions and anti-contractions is the following.
Lemma 7.4.3. Let g and h be the anti-restriction and anti-contraction with extra
element s of a set function f on V and let P be a property of set functions closed by
taking fusions and by adding constant functions. Then,
1. Both functions g and h are symmetric;
2. f is a nontrivial restriction of a symmetric function with property P if and only

if its anti-restriction g has property P ;
171
3. f is a nontrivial contraction of a symmetric function with property P if and
only if f (∅) = 0 and its anti-contraction h has property P .
Proof. The first item follows directly from the definitions of g and h. The sufficiency of
the second and third items are due to Lemma 7.4.2. We only need to check necessity.
For the second item, suppose that f = f 0 \ T for some symmetric set function f 0
with property P . Let fˆ be the function obtained from f 0 by fusing T into a single
element which we denote as s. The function fˆ is symmetric and has property P , and
furthermore it is easy to see that f = fˆ \ {s}. To conclude this part, we show that fˆ
is equal to the anti-restriction g of f . Indeed, for all A ⊆ V ,
g(A) = f (A) = fˆ(A), and

g(A + s) = g(V \ A) = f (V \ A) = fˆ(V \ A) = fˆ(A + s).
For the third item, let f = f 0 /T for some symmetric set function f 0 with prop-
erty P . Similar to above, let fˆ be the function obtained from f 0 by fusing T into
a single element s. Then f is equal to fˆ/{s}. In particular, fˆ is symmetric, has
property P and
f (∅) = fˆ(∅ + s) − fˆ(s) = 0.
Moreover, for all A ⊆ V , we have
h(A) = f (V \ A) = fˆ(V \ A + s) − fˆ(s) = fˆ(A) − fˆ(s), and

h(A + s) = f (A) = fˆ(A + s) − fˆ(s).
Therefore, the anti-contraction h of f is equal to fˆ shifted by the constant fˆ(s). We

conclude the proof by recalling that P is closed under the addition of constants.
We also have the following property.
Lemma 7.4.4. Let g be the anti-restriction with extra element s of a set function
f on V . Given an hereditary system I on V , the solutions for the hereditary min-
imization problem on (V, f, I) coincide with the solutions for the hereditary system
(V + s, g, I).
Proof. This follows since f = g \{s} and s is a loop of I when viewed as an hereditary
system on V + s.
7.4.2 Submodular Functions

Consider two different sets A, B ⊆ V . We say that A and B are intersecting if
A \ B, B \ A, and A ∩ B are all nonempty. We further say that A and B are crossing
if they are intersecting and the set V \ (A ∪ B) is also nonempty.
Consider the following submodular inequality,
f (A ∪ B) + f (A ∩ B) ≤ f (A) + f (B). (7.11)
172
A set function f : 2V → R is called fully (resp., intersecting, crossing) sub-
modular if inequality (7.11) is satisfied for every pair A and B in V (resp., for every
pair of intersecting sets or crossing sets). Fully submodular functions are what we
usually denote as submodular functions. However, from this point on we keep the
adjective “fully” to avoid confusion. Note that every fully submodular function is
intersecting submodular and every intersecting submodular function is crossing sub-
modular.
The function f is called fully supermodular if −f is fully submodular, and
fully modular if it is both fully submodular and fully supermodular. We extend
these definitions to their intersecting and crossing versions.
Consider the following examples.
1. The cut function of a nonnegatively weighted undirected or directed graph is

fully submodular.
2. The function f : {1, 2, 3} → R given by


1, if |A| = 2;
f (A) = 
0, otherwise.
The only possibility for two sets A and B to be intersecting is that both of
them have cardinality 2 and intersect in a single element. In this case 2 =
f (A) + f (B) ≥ f (A ∪ B) − f (A ∩ B) = 0. Hence f is intersecting submodular.
But it is not fully submodular since 0 = f (1) + f (2) ≤ f ({1, 2}) + f (∅) = 1.
3. The function g : {1, 2, 3, 4} → R given by


−1, if A ∈ E1 = {{1, 2}, {3, 4}};


g(A) =

1, if A ∈ E2 = {{1, 3}, {1, 4}, {2, 3}, {2, 4}};


0, otherwise.
The only possibility for two sets A and B to be crossing is that both of them have
cardinality 2, intersect in a single element and leave one element outside their
union. At least one of A or B must then be in E2 , from here we have that g(A)+
g(B) ≥ 0 = g(A ∪ B) + g(A ∩ B). This implies that g is crossing submodular.
However, g is not intersecting submodular since −1 = g({1, 2, 3}) + g({3, 4}) ≤
g({1, 2, 3, 4}) + g(3) = 0.
The following lemma states properties of the fusions, restrictions and contractions
of submodular functions.
Lemma 7.4.5. Let f : 2V → R be a set function.
1. If f is (fully, intersecting, crossing) submodular then so are all its fusions.
2. If f is (fully, intersecting, crossing) submodular then so are all its restrictions

and contractions.
173
3. If f is crossing submodular then its non-trivial restrictions are intersecting sub-
modular.
Proof. The proof of the first item is straightforward from the definitions. For the
second item note that if f is (fully, intersecting, crossing) submodular function on V
and T is a nontrivial subset of V then for all A, B ⊆ V \ T ,
f \ T (A) + f \ T (B) − f \ T (A ∪ B) − f \ T (A ∩ B)
= f (A) + f (B) − f (A ∪ B) − f (A ∩ B), (7.12)
implying that f \ T is also (fully, intersecting, crossing) submodular. On the other

hand,
f /T (A) + f /T (B) − f /T (A ∪ B) − f /T (A ∩ B)
= f (A ∪ T ) + f (B ∪ T ) − f ((A ∪ T ) ∪ (B ∪ T )) − f ((A ∩ T ) ∩ (B ∩ T )). (7.13)
Note that if A and B have nonempty intersection, then they are intersecting
(resp. crossing) in V \ T if and only if A ∪ T and B ∪ T are intersecting (resp. crossing)
in V . Therefore, the function f /T is (fully, intersecting, crossing) submodular.
The third item follows from the fact that if A and B are two intersecting sets in
V \ T such that A ∪ B = V \ T then the sets A and B are crossing in V , therefore by
(7.12) the submodular inequality holds for f \ T .
7.4.3 Posimodular Functions

A set function f : 2V → R is called fully posimodular if for all A, B ⊆ V ,
f (A \ B) + f (B \ A) ≤ f (A) + f (B). (7.14)
We say that f is intersecting posimodular or crossing posimodular if in-

equality (7.14) is satisfied for every pair A and B of intersecting sets or crossing sets
respectively. A function f is (fully, intersecting, crossing) negamodular if −f
is (fully, intersecting, crossing) posimodular.
Lemma 7.4.6. Every symmetric fully submodular function is fully posimodular, and
every symmetric crossing submodular function is intersecting posimodular.
Proof. Let f be a symmetric set function on V . Consider first the case where f is
fully submodular. By symmetry,
f (A) + f (B) = f (V \ A) + f (B) ≥ f ((V \ A) ∪ B) + f ((V \ A) ∩ B)

= f (A \ B) + f (B \ A), (7.15)
for all A, B ⊆ V . This implies fully posimodularity.

Now assume that f is only crossing submodular. Let A and B be two intersecting
sets. We have two cases: either A ∪ B = V or A and B are crossing.
174
In the first case, V \ A = B \ A and V \ B = A \ B, then we have
f (A) + f (B) = f (V \ A) + f (V \ B) = f (B \ A) + f (A \ B).
In the second case, V \ A and B are also crossing, hence (7.15) also holds.
The converse of the previous lemma does not hold. For example, the sum of a
fully modular function and a symmetric fully submodular function is always fully
submodular and fully posimodular but it is not necessarily symmetric. It is also
easy to construct examples of functions that are fully posimodular but not fully
submodular. One such example is the indicator function of the ground set V .
Since (fully, intersecting, crossing) submodularity is preserved under taking fu-
sions and adding constants, Lemma 7.4.3 in the previous section allows us to test
whether a function is a restriction or a contraction of a symmetric (fully, intersecting,
crossing) submodular function by looking at its anti-restriction or anti-contraction.
As a corollary, we have the following.
Lemma 7.4.7 (Implicit in Nagamochi and Ibaraki [127] and Narayanan [130]). The
nontrivial restrictions of symmetric crossing submodular functions are exactly those
functions that are both intersecting submodular and intersecting posimodular.
Proof. Let f be a set function on V and g be its anti-restriction with extra element s.
By Lemma 7.4.3 we need to check that the following two statements are equivalent:
(i) g is symmetric crossing submodular.
(ii) f is intersecting submodular and intersecting posimodular.
Suppose that g is symmetric crossing submodular. Then condition (ii) holds by using
Lemmas 7.4.5 and Lemma 7.4.6.
Suppose then that f satisfies condition (ii). We already know, by definition, that
g is symmetric. Let A and B be two crossing sets in V + s.
If s 6∈ A ∪ B, then we have
g(A) + g(B) − g(A ∪ B) − g(A ∩ B) = f (A) + f (B) − f (A ∪ B) − f (A ∩ B) ≥ 0,
where the last inequality holds by intersecting submodularity of f .

If s ∈ A ∩ B, then let A0 = A − s and B 0 = B − s. We have
g(A) + g(B) = f (V \ A0 ) + f (V \ B 0 )
≥ f ((V \ A0 ) ∪ (V \ B 0 )) + f ((V \ A0 ) ∩ (V \ B 0 ))
= f (V \ (A0 ∩ B 0 )) + f (V \ (A0 ∪ B 0 ))
= g((A0 ∩ B 0 ) + s) + g((A0 ∪ B 0 ) + s)
= g(A ∩ B) + g(A ∪ B),
where the inequality holds by intersecting submodularity of f .
175
For the remaining cases assume that s ∈ A \ B without loss of generality and let
0
A = A \ s. By intersecting posimodularity of f we have
g(A) + g(B) = f (V \ A0 ) + f (B)

≥ f ((V \ A0 ) \ B) + f (B \ (V \ A0 ))
= f (V \ (A0 ∪ B)) + f (A0 ∩ B)
= g(A ∪ B) + g(A ∩ B).
Hence, g is crossing submodular.
7.4.4 Rizzi Functions

Rizzi [145] introduces a nice family of bi-set functions, which we describe below.
We say that a symmetric bi-set function d on V is a Rizzi bi-set function if the
following properties hold.
1. (Consistency) For all A, B, C ⊆ V disjoint:
d(A, B) ≤ d(A, C) implies d(A ∪ C, B) ≤ d(A ∪ B, C).
2. (Monotonicity) For all nonempty disjoint A, B, C ⊆ V :
d(A, B) ≤ d(A, B ∪ C).
The associated Rizzi set function is f = f (d) .
Lemma 7.4.8. For every set function f , the bi-set function d(f ) is symmetric and
consistent.
Proof. The function d(f ) (A, B) = 21 (f (A) + f (B) + f (∅) − f (A ∪ B)) is symmetric
by definition. Consistency holds since
d(f ) (A ∪ C, B) − d(f ) (A ∪ B, C)
1
= (f (A ∪ C) + f (B) − f (A ∪ B) − f (C))
2
= d(f ) (A, B) − d(f ) (A, C).
Rizzi observed the following.
Lemma 7.4.9 (Rizzi [145]). If f is an intersecting submodular function on V , then

the associated bi-set function3 d(f ) is a Rizzi bi-set function. Furthermore, if f is
symmetric, f is the associated Rizzi set function of d(f ) .
3
Rizzi considered the function d(A, B) = f (A) + f (B) − f (A ∪ B) instead, but the function we
(f )
consider satisfies f = f (d ) .
176
Proof. To show monotonicity. Let A,B,C be nonempty disjoint sets. Then

2 d(f ) (A, B ∪ C) − d(f ) (A, B)
= f (B ∪ C) − f (A ∪ B ∪ C) − f (B) + f (A ∪ B). (7.16)
Let D = B ∪ C and E = A ∪ B, so that D ∪ E = A ∪ B ∪ C and D ∩ E = B. Since all

A, B and C are nonempty, the sets D and E are intersecting. Hence, (7.16) above is
nonnegative implying monotonicity.
(f )
We have already seen in Lemma 7.4.1 that for symmetric f , f = f (d ) .
Rizzi [145] mistakenly states that the lemma above holds if we replace intersecting
submodularity by the weaker condition of crossing submodularity. However, for the
symmetric crossing submodular function f : V → R satisfying f (∅) = f (V ) = 1, and
f (X) = 0 for all X 6∈ {∅, V }, where |V | ≥ 3 this does not holds. Indeed, if {A, B, C}
is a partition of V in nonempty parts, then
2d(f ) (A, B ∪ C) = f (A) + f (B ∪ C) + f (∅) − f (A ∪ B ∪ C) = 0, and

2d(f ) (A, B) = f (A) + f (B) + f (∅) − f (A ∪ B) = 1.
Hence, we do not have monotonicity. One way to fix this problem is to define a weaker
version of monotonicity. Consider the following property.
1’. (Weak Monotonicity) d(A, B) ≤ d(A, B ∪ C), for all nonempty disjoint sets
A, B, C ⊆ V with A ∪ B ∪ C 6= V .
We call d a weak Rizzi bi-set function if it is symmetric, weak monotone and

consistent. The associated function f (d) is called a weak Rizzi set function.
Lemma 7.4.10. If f is a crossing submodular function on V , then the associated

bi-set function d(f ) is a weak Rizzi bi-set function. Furthermore if f is symmetric, f
is the associated weak Rizzi set function of d(f ) .
Proof. We only need to show weak monotonicity. Indeed, for A, B and C satisfying
A ∪ B ∪ C 6= ∅, the right hand side of (7.16) is nonnegative by crossing submodularity
of f .
We remark here that recognizing that a function f is a (weak) Rizzi function

without having access to the bi-set Rizzi function d for which f = f (d) is not simple.
(d)
This follows since it is not always the case that d = d(f ) . It is also easy to find Rizzi
functions that are not crossing submodular. The following example can be found in
Rizzi [145].
Given a weighted connected graph G = (V, E, w), with w : E → R+ , define the
maximum separation between two disjoint sets of vertices A and B as
d(A, B) = max w(ab).

ab∈E(A:B)
177
This function is symmetric, monotone and consistent. However, for the complete
graph in {a, b, c, d} where all the edges have weight 1, except ac and bd which have
weight 2. The corresponding function f = f (d) given by f (A) = d(A, V \ A), for
A ⊆ V is not crossing submodular since
3 = f ({a, c}) + f ({a, d}) < f ({a, c, d}) + f (a) = 4.
The function f is not even crossing posimodular since
3 = f ({a, c}) + f ({a, d}) < f (c) + f (d) = 4.
In particular, we have the following result.

Lemma 7.4.11. The class of restrictions of weak Rizzi functions strictly contains
the restrictions of symmetric crossing submodular functions.
Proof. The containment follows from Lemma 7.4.10. The fact that the containment
is strict follows from the example above and Lemma 7.4.5.
A useful property of (weak) Rizzi bi-set functions is the following
Lemma 7.4.12. If d is a (weak) Rizzi bi-set function then so are all its fusions.
Proof. Direct from the definitions.
Rizzi has shown that a version of the maximum adjacency ordering, defined by
Queyranne allows us to find pendant pairs for Rizzi set functions. His proof, which
we describe below, naturally extends to weak Rizzi set functions.
Let d be a weak Rizzi bi-set function. An ordering (v1 , . . . , vn ) of the elements
of V , such that
d(vi , Wi−1 ) ≥ d(vj , Wi−1 ), for all 2 ≤ i ≤ j ≤ n, (7.17)
where v1 can be chosen arbitrarily, and Wi denotes the set {v1 , . . . , vi } is called a
maximum adjacency ordering for d. The origin of the name comes from its
interpretation for the cut function on a graph. If d(A, B) represents the cut between
A and B in a given graph G, then the i-th vertex of a maximum adjacency ordering
of d is exactly the one having maximum number of edges going towards the previous
i − 1 vertices.
Lemma 7.4.13 (Essentially in Rizzi [145]). For a weak Rizzi bi-set function d on V ,
and an arbitrary element v1 ∈ V , the last two elements (vn−1 , vn ) of a maximum adja-
cency ordering of V starting from v1 constitute a pendant pair for f (d) . Furthermore,
this order can be found by using O(n2 ) oracle calls to d and in the same running time.
Proof. It is easy to check the claim regarding the running time since to construct
the i-th element of the ordering we only require to find the maximum of the n − i
different values {d(x, Wi−1 )}x∈V \Wi−1 . We show the rest by induction on the number
of elements.
178
The lemma holds trivially for n = 2 since the only set separating v1 from v2 are
the singletons and the function f (d) is symmetric. For n = 3, the only sets separating
v2 from v3 are {v3 }, {v1 , v3 } and their complements. By definition of the ordering,
d(v2 , v1 ) ≥ d(v3 , v1 ). Consistency implies that
f (d) ({v1 , v3 }) = d({v1 , v3 }, v2 ) ≥ d({v1 , v2 }, v3 ) = f (d) (v3 ).
Consider n ≥ 4 and let S be any set separating vn and vn−1 . We must show
d(S, V \ S) ≥ d(vn , V − vn ). (7.18)
It is easy to see that (v{1,2} , v3 , . . . , vn ) is a maximum adjacency ordering for the

function d1,2 obtained by fusing v1 and v2 into v{1,2} . If S does not separate v1 and
v2 then (7.18) holds since (vn−1 , vn ) is a pendant pair of d1,2 by induction. So assume
that S separates v1 and v2 .
We claim that the ordering (v1 , v{2,3} , . . . , vn ) is a maximum adjacency ordering
for the function d2,3 obtained by fusing v2 and v3 into v{2,3} . Indeed, we only need to
prove that d2,3 (v{2,3} , v1 ) ≥ d2,3 (vj , v1 ) for all j ≥ 4. This follows since, by hypothesis
and weak monotonicity, we have
d(vj , v1 ) ≤ d(v2 , v1 ) ≤ d({v2 , v3 }, v1 ).
If S does not separate v2 and v3 then (7.18) holds by induction, since (vn−1 , vn ) is
a pendant pair for d2,3 .
The only possibility left is that S separates v1 from v2 and v2 from v3 . This
means that S does not separate v1 and v3 . To conclude (7.18) it suffices to show that
(v2 , v{1,3} , . . . , vn ) is a maximum adjacency ordering for the function d1,3 obtained by
fusing v1 and v3 into v{1,3} . Assume that this is not the case, then we must have
d1,3 (v{1,3} , v2 ) < d1,3 (vj , v2 )
for some j ≥ 4. Since (v1 , . . . , vn ) is a maximum adjacency ordering of d, we have

d(v2 , v1 ) ≥ d(v3 , v1 ) and d(v3 , {v1 , v2 }) ≥ d(vj , {v1 , v2 }). By consistency we also have
d({v1 , v3 }, v2 ) ≥ d({v1 , v2 }, v3 ). Combining the inequalities above and using weak
monotonicity we get
d(v3 , {v1 , v2 }) ≥ d(vj , {v1 , v2 }) ≥ d(vj , v2 ) = d1,3 (vj , v2 )

> d1,3 (v{1,3} , v2 ) = d({v1 , v3 }, v2 ) ≥ d(v3 , {v1 , v2 }),
which is a contradiction.
7.4.5 Main Results

By combining Lemmas 7.4.13 and 7.4.12 we conclude that weak Rizzi functions are
strongly admissible for pendant pairs. By Lemma 7.4.10, symmetric crossing sub-
modular functions have the same property.
179
Summarizing, we obtain the following results.
Theorem 7.4.14. We can compute all the minimal optimal solutions of the hereditary
system (V, f, I) in time O(n3 ) and using O(n3 ) oracle calls for the following cases:
1. If f = f (d) , for a weak Rizzi bi-set function d on V , and we have access to an
oracle for d.
2. If f is a symmetric (fully, crossing) submodular function on V and we have

access to an oracle for f .
3. If f is an intersecting submodular and intersecting posimodular function on V

and we have access to an oracle for f .
4. If f is defined as
f (A) = d(A, (V \ A) + s), for all A ⊆ V
for some weak Rizzi bi-set function d on V + s, and we have access to an oracle
for d.
Proof. For all results we use Algorithm 20. The first item follows by Lemma 7.4.13 and
the fact that all the fusions of d can be evaluated using oracle calls to d. Item number
two holds since by Lemma 7.4.10, the bi-set function d(f ) is a weak Rizzi function
(f )
and f = f (d ) . Therefore, we can use the result of the first item to conclude.
For the third item, we can construct a value oracle for the anti-restriction g of f .
This value oracle uses one call to the oracle for f . Let s be the extra element added
by the anti-restriction and consider the system (V + s, g, I). By Lemma 7.4.7, g is
crossing submodular on V + s so we can use the result of the second item to find
all the minimal optimal solutions of (V + s, g, I). By Lemma 7.4.4, these solutions
coincide with the ones of the system (V, f, I).
Finally for item number four, consider the system (V + s, g, I) where g = f (d) is
the weak Rizzi function on V + s obtained from d. We can find the minimal optimal
solutions for this system using the first item in this theorem. We claim that g is the
anti-restriction of f with extra element s. Indeed, for all A ⊆ V ,
g(A) = f (d) (A) = d(A, (V + s) \ A) = f (A), and

g(A + s) = f (d) (A + s) = d(A + s, V \ A) = d(V \ A, (V \ (V \ A)) + s) = f (V \ A).
By Lemma 7.4.4, the solutions found for (V + s, g, I) coincide with the ones of the
system (V, f, I).
It is worth noting at this point that we can also use our methods to find all the
inclusion-wise maximum minimizers of some functions constrained to co-hereditary
families, that is families of sets closed under union. Given a family of sets I on
V , its dual with respect to V is the family I ∗V = {A : V \ A ∈ I}. Then the
co-hereditary families are exactly the duals of hereditary family. The set function
dual f ∗V of f is defined as f ∗V (A) = f (V \ A) for all A ⊆ V . A triple (V, f, J ) is
180
a co-hereditary system if (V, f ∗V , J ∗V ) is an hereditary system. A set X ⊆ V is a
maximal optimal solution for a co-hereditary system (V, f, I) if X is a nontrivial
inclusion-wise maximal set minimizing f over all the sets in I. Note that the maximal
optimal solutions for a co-hereditary system (V, f, J ) are exactly the complements of
the minimal optimal solutions of the hereditary system (V, f ∗V , J ∗V ). As a corollary
of the previous theorem we have the following.
Theorem 7.4.15. We can compute all the maximal optimal solutions of the co-
hereditary system (V, f, J ) in time O(n3 ) and using O(n3 ) oracle calls for the follow-
ing cases:
1. If f = f (d) , for a weak Rizzi bi-set function d on V , and we have access to an
oracle for d.
2. If f is a symmetric (fully, crossing) submodular function on V and we have

access to an oracle for f .
3. If the anti-contraction of f is a symmetric crossing submodular function on V

and we have access to an oracle for f .
4. If f is defined as
f (A) = d(A + s, V \ A), for all A ⊆ V
for some weak Rizzi bi-set function d on V + s, and we have access to an oracle
for d.
Proof. The theorem follows since the co-hereditary systems (V, f, J ) for each case
are exactly the duals of the hereditary systems considered in Theorem 7.4.14
7.5 Nagamochi’s Flat Pair Based Algorithm

In this section we briefly describe some of the ideas in a recent algorithm of Nag-
amochi [124] that can also be used to solve the hereditary constrained minimization
problem. We need the following definitions.
Consider a set function f on V . A nonempty proper subset A of V is called an
extreme set of f if
f (A) < f (B), for all ∅ ⊂ B ⊂ A. (7.19)
Let X (f ) be the collection of all extreme sets of f .
A flat pair of f is an unordered pair of elements {t, u} such that
f (X) ≥ min f (x), for all X ⊆ V separating t from u. (7.20)

x∈X
Call a function f admissible for flat pairs4 if f and all its fusions admit flat
pairs.
4
In [124], this class of functions is not named.
181
As noted by Nagamochi, a non-singleton extreme set A can not separate a flat
pair {t, u}. This fact alone implies that the extreme sets of functions admissible for
flat pairs form a laminar family. Nagamochi shows this in an algorithmic way, we
give a direct proof below.
Lemma 7.5.1. The set X (f ) of extreme sets of a function f admissible for flat pairs
is laminar.
Proof. Suppose by contradiction that two extreme sets A and B are intersecting.
Let (V 0 , f 0 ) be the system obtained by fusing all the elements of A \ B into a, all
the elements of B \ A into b, all the elements of A ∩ B into c, and if V \ (A ∪ B) is
nonempty, all the elements of this set into d. Every pair of elements in V 0 is separated
by either A or B. Hence none of them can be a flat pair.
The above lemma implies that the number of extreme sets of a flat pair admissible
function is O(n). Nagamochi gives an algorithm that outputs all the extreme sets of
X (f ) for any function f admissible for flat pairs provided we have access to an algo-
rithm that finds a flat pair of any fusion of f . The algorithm is similar to Queyranne’s
algorithm in the sense that at every iteration we fuse a flat pair together. In what
follows we give a high level overview of the algorithm.
Initialize a set X with the singletons of V . This set contains at every iteration
all the extreme sets of f that are contained inside the singletons of the current fused
system (in other words, X contains all the extreme sets completely inside Vx for some
x ∈ V 0 , where V 0 is the current ground set). In every iteration, the algorithm finds a
flat pair {t, u}, and fuse them into a single element tu. After that, it tests if the set
Vtu containing all the elements of V that have been fused into the new element tu is
extreme by using the information in X . If Vtu is extreme, the algorithm adds it to X .
Nagamochi gives an efficient implementation of the above algorithm that runs in
O(nT (n))-time, where T (n) is the time needed to find a flat pair of a function over a
ground set of n elements.
We can use Nagamochi’s algorithm to solve the hereditary minimization problem
on (V, f, I) where f is admissible for flat pairs as follows. Any minimal optimal
solution of (V, f, I) is an extreme solution of f . Hence, to find all the minimal optimal
solutions of (V, f, I) we can compute the extreme sets of f , keep the ones contained
in I and find among these the ones having minimum f -value. This algorithm has the
same asymptotic time complexity as the algorithm we have presented for the same
problem (see Theorem 7.3.5).
It is an interesting question to decide when a function is admissible for flat pairs.
Nagamochi has shown that symmetric submodular functions (and also their restric-
tions) admit flat pairs, and that we can find them by using a similar technique to the
one used by Queyranne to find pendant pairs.
More precisely, consider an ordering (v1 , . . . , vn ) of the elements of V , such that
f (vi ) + f (Wi−1 + vi ) ≤ f (vj ) + f (Wi−1 + vj ), for all 1 ≤ i ≤ j ≤ n, (7.21)
where Wi denotes the set {v1 , . . . , vi }. An order satisfying (7.21) is called a minimum
degree ordering. Nagamochi [124] shows the following result:
182
Lemma 7.5.2 (Nagamochi [124]). For a symmetric crossing submodular function f
on V , the last two elements {vn−1 , vn } of a minimum degree ordering of V constitute
a flat pair. Furthermore, this ordering can be found by using O(n2 ) value oracle calls
to f and in the same running time.
He also extends the above lemma to intersecting submodular intersecting posi-
modular functions. Given one such function f , modify it so that f (∅) = f (V ) = −∞.
This does not affect its intersecting submodularity or its intersecting posimodularity.
Consider the anti-restriction g of this modified function with extra element s. This
function is symmetric crossing submodular. As g(s) = f (V ) = −∞, the element s is
always the first element of a minimum degree ordering; therefore, by Lemma 7.5.2,
g has a flat pair {t, u} avoiding s. This pair {t, u} is also a flat pair for the orig-
inal function f since if there was a subset X ⊆ V separating t and u for which
f (X) < minx∈X f (x), we would also have g(X) < minx∈X g(x) contradicting the fact
that {t, u} is a flat pair for g.
In what follows, we show a small modification to the lemma above that allows us
to find flat pairs for a slightly larger family of functions.
Consider a symmetric bi-set function d on V satisfying the following property.
d(A ∪ B, C) ≥ d(A, B ∪ C) implies d(A ∪ B, C ∪ D) ≥ d(A, B ∪ C ∪ D),

for all nonempty and disjoint A, B, C, D ⊆ V . (7.22)
Let (v1 , . . . , vn ) be an ordering of the elements of V , such that
d(vi , (V \ Wi−1 ) − vi ) ≤ d(vj , (V \ Wi−1 ) − vj ), for all 1 ≤ i ≤ j ≤ n, (7.23)
where Wi denotes the set {v1 , . . . , vi }. As before, call an ordering satisfying (7.23) a
minimum degree ordering of d. This name also comes from its interpretation for
the case where d(A, B) represents the cut between A and B. In this case, the i-th
vertex of the ordering is selected as the one having minimum degree on the graph
obtained by removing the first i − 1 vertices.
Lemma 7.5.3. Let d be a symmetric bi-set function on V satisfying (7.22). The last
two elements {vn−1 , vn } of a minimum degree ordering of d constitute a flat pair of
the set function f defined by f (A) = d(A, V \ A).
Proof. The proof is essentially the same as the one given by Nagamochi [124] to prove
Lemma 7.5.2. We include it here for completeness.
Define for every i = n − 2, . . . , 0, the symmetric function fi : (V \ Wi ) → R by
fi (X) = d(X, (V \ Wi ) \ X), for all X ⊆ V \ Wi .
We claim that for all i ∈ {n − 2, n − 3, . . . , 0}, {vn , vn } is a flat pair of fi . This is,
fi (X) ≥ min fi (x), for all X ⊆ V \ Wi separating vn−1 and vn . (7.24)

x∈X
Note that f0 (X) = f (X), hence to prove the lemma we need to show the above claim
holds for i = 0.
183
We prove the claim by induction. The case i = n − 2 is trivial as {vn−1 } and {vn }
are the only sets in V \ Wn−2 separating vn−1 and vn . Suppose that (7.24) holds for
i = j, we prove that it also holds for i = j −1. Let X be subset of V \Wj−1 separating
vn−1 from vn . If |X| = 1 then (7.24) holds trivially, so assume that |X| ≥ 2.
We have two cases
Case 1: vj 6∈ X. By hypothesis, there is an element x∗ ∈ X such that fj (X) ≥ fj (x∗ ).
Set A = {x∗ }, B = X − x∗ , C = ((V \ Wj ) \ X) and D = {vj }. Thus A, B, C, D is a
partition of V \ Wj−1 in nonempty sets.
Note that fj (X) ≥ fj (x∗ ) is equivalent to
d(A ∪ B, C) ≥ d(A, B ∪ C).
By (7.22), this implies that
d(A ∪ B, C ∪ D) ≥ d(A, B ∪ C ∪ D),
or equivalently, fj−1 (X) ≥ fj−1 (x∗ ), completing the proof of this case.
Case 2: vj ∈ X. By choice of vj , fj−1 (vj ) ≤ fj−1 (x) for all x ∈ V \ Wj−1 . Consider
the set Y = (V \ Wj ) \ X not containing vj , which separates vn−1 and vn . By the first
case fj−1 (Y ) ≥ miny∈Y fj−1 (y).
Then we have
fj−1 (X) = fj−1 (Y ) ≥ min fj−1 (y) ≥ fj−1 (vj ),

y∈Y
completing the proof.
We now show that the functions obtained from bi-set functions d satisfying (7.22)
include symmetric crossing submodular functions. Indeed, let f be a symmetric
(f )
crossing submodular function. As f is symmetric, f = f (d ) . We show that d(f )
satisfies (7.22). Indeed suppose that A, B, C, D are disjoint nonempty subsets of V .
Then d(f ) (A ∪ B, C) ≥ d(f ) (A, B ∪ C) is equivalent to
f (A ∪ B) + f (C) ≥ f (A) + f (B ∪ C). (7.25)
Since B ∪ C and C ∪ D are crossing sets, we also have
f (B ∪ C) + f (C ∪ D) ≥ f (B ∪ C ∪ D) + f (C). (7.26)
Summing (7.25) and (7.26) we get
f (A ∪ B) + f (C ∪ D) ≥ f (A) + f (B ∪ C ∪ D) (7.27)
which is equivalent to d(f ) (A ∪ B, C ∪ D) ≥ d(f ) (A, B ∪ C ∪ D).

Recall the maximum separation function d(A, B) = maxe∈E(A:B) w(e) for a given
weighted graph G = (V, E, w). As shown in Section 7.4.4, the function f = f (d) is a
Rizzi function that is not necessarily crossing submodular. Note that f also satisfies
(7.22). This shows that Lemma 7.4.10 is a strict generalization of Lemma 7.5.2. In
184
fact, even though our notion of minimum degree ordering (7.23) finds a flat pair,
Nagamochi’s minimum degree ordering (7.21) does not necessarily find one for the
function f defined in this paragraph.
As an example of the last assertion, consider the complete graph G on 6 vertices
{a, b, c, a0 , b0 , c0 }, assign a weight of one to every edge in the triangles T1 = {a, b, c}
and T2 = {a0 , b0 , c0 } and zero weight to all the other edges. See Figure 7-1.
b b0
a T1 T2 a0
c c0
Figure 7-1: Present edges have unit weight. Non-edges have zero weight.
We have

0 if A ⊆ T1 and B ⊆ T2 or vice versa;
d(A, B) =
1 otherwise.
Therefore, f (A) = 0 if and only if A ∈ {T1 , T2 }.

Under Nagamochi’s notion (7.21) of minimum degree ordering, the i-th vertex of
the ordering is selected as the one minimizing the function gi (x) = f (x)+f (Wi−1 +x),
where Wj stands for the first j selected vertices. Note that gi (x) has value 2 unless,
Wi−1 + x ∈ {T1 , T2 }, where gi (x) has unit value. From here it is easy to see that
(a, a0 , b, b0 , c, c0 ) is a minimum degree ordering of f in the sense of Nagamochi. But
the last two vertices {c, c0 } do not form a flat pair since T1 separates them and
0 = f (T1 ) < minx∈T1 f (x) = 1.
Under our notion (7.23) of minimum degree ordering, the i-th vertex is selected
as the one minimizing hi = d(x, (V \ Wi−1 ) − x). In this situation, (a, a0 , b, b0 , c, c0 ) is
not a minimum degree ordering since for i = 4,
hi (b0 ) = d(b0 , (V \ Wi−1 ) − b0 ) = d(b0 , {c, c0 }) = 1, but

hi (c0 ) = d(c0 , (V \ Wi−1 ) − c0 ) = d(c0 , {b, c}) = 0.
We remark that Nagamochi’s notion (7.21) and our notion (7.23) of minimum de-
gree ordering coincide whenever f is symmetric and d = d(f ) (in particular, this holds
for symmetric crossing submodular functions). In this situation, for the functions gi
and hi defined above
2hi (x) = 2d(x, (V \ Wi−1 ) − x) = f (x) + f ((V \ Wi−1 ) − x) − f (∅) + f (V \ Wi−1 )

= f (x) + f (Wi−1 + x) − f (∅) + f (Wi−1 )
= gi (x) − f (∅) + f (Wi−1 ).
Therefore, the minimizers of hi (x) and gi (x) on V \ Wi−1 coincide.
185
7.6 Discussion
In this chapter we have given polynomial time algorithms to solve the hereditary
minimization problem for a class of functions strictly containing symmetric crossing
submodular functions and their restrictions.
We can compare our results to the ones of Nagamochi [124]. On the one hand,
his algorithm can solve the more general problem of finding extreme sets of sym-
metric crossing submodular functions and their restrictions. On the other hand, our
algorithms (Algorithms 19 and 20) work for a strictly larger class of functions.
We have also presented a slight extension of Nagamochi’s minimum degree order-
ing that allows us to find extreme sets of a more general class of functions. Namely,
the ones arising from bi-set functions satisfying (7.22), and their restrictions. We
have not studied how do these functions compare to weak Rizzi functions and their
restrictions, and it would be very interesting to find a superclass of both for which
we can find extreme sets efficiently.
A related question is the following. We have seen that the extreme sets of functions
admissible for flat pairs form a laminar family (Lemma 7.5.1). The same holds for
functions that are strongly admissible for pendant pairs: the proof of this fact is
the same as the one of Lemma 7.3.1 stating that the minimal optimal solutions are
disjoint. Is it possible to devise an efficient algorithm to compute extreme set of a
function given an oracle to find pendant pairs on every fusion?
Nagamochi [124] gives an example of a cut function of a graph for which the
pendant pairs and the flat pairs are different. An interesting question is to find
natural functions admitting one type of pairs but not the other. Similarly, it is
open to find the relation between functions strongly admissible for pendant pairs and
functions that are admissible for flat pairs.
186
Bibliography
[1] M. Ajtai, N. Megiddo, and O. Waarts. Improved algorithms and analysis for
secretary problems and generalizations. SIAM J. Discret. Math., 14(1):1–27,
2001. [43]
[2] J. Amilhastre, M. C. Vilarem, and P. Janssen. Complexity of minimum biclique

cover and minimum biclique decomposition for bipartite domino-free graphs.
Discrete Applied Mathematics, 86(2-3):125–144, 1998. [109, 118, 119, 150]
[3] A. von Arnim and C. de la Higuera. Computing the jump number on semi-
orders is polynomial. Discrete Applied Mathematics, 51(1–2):219–232, 1994.
[109]
[4] B. Aronov, E. Ezra, and M. Sharir. Small-size ε-nets for axis-parallel rectangles
and boxes. SIAM Journal on Computing, 39(7):3248–3282, 2010. [127]
[5] M. Babaioff, M. Dinitz, A. Gupta, N. Immorlica, and K. Talwar. Secretary

problems: Weights and discounts. In Proceedings of the Twentieth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA’09, pages 1245–1254,
2009. [13, 42, 43]
[6] M. Babaioff, N. Immorlica, D. Kempe, and R. Kleinberg. A knapsack secre-

tary problem with applications. In M. Charikar, K. Jansen, O. Reingold, and
J. D. P. Rolim, editors, APPROX-RANDOM, volume 4627 of Lecture Notes in
Computer Science, pages 16–28. Springer, 2007. [13, 30, 44]
[7] M. Babaioff, N. Immorlica, D. Kempe, and R. Kleinberg. Online auctions and

generalized secretary problems. SIGecom Exchanges, 7(2):1–11, 2008. [32, 41]
[8] M. Babaioff, N. Immorlica, and R. Kleinberg. Matroids, secretary problems,

and online mechanisms. In Proceedings of the Eighteenth Annual ACM-SIAM
Symposium on Discrete Algorithms, SODA’07, pages 434–443, 2007. [12, 13,
14, 23, 32, 35, 41, 42, 70, 80, 81]
[9] K. A. Baker, P. C. Fishburn, and F. S. Roberts. Partial orders of dimension 2.

Networks, 2(1):11–28, 1972. [91, 92]
[10] M. Bateni, M. Hajiaghayi, and M. Zadimoghaddam. Submodular secretary

problem and extensions. In M. J. Serna, R. Shaltiel, K. Jansen, and J. D. P.
187
Rolim, editors, APPROX-RANDOM, volume 6302 of Lecture Notes in Com-
puter Science, pages 39–52. Springer, 2010. [44]
[11] A. A. Benczúr. Pushdown-reduce: an algorithm for connectivity augmentation

and poset covering problems. Discrete Applied Mathematics, 129(2-3):233–262,
2003. [149]
[12] A. A. Benczúr, J. Förster, and Z. Király. Dilworth’s theorem and its application
for path systems of a cycle - implementation and analysis. In J. Nešetřil, edi-
tor, ESA, volume 1643 of Lecture Notes in Computer Science, pages 498–509.
Springer, 1999. [138]
[13] C. Berge. Färbung von Graphen, deren sämtliche bzw. deren ungerade Kreise
starr sind. Wissenchaftliche Zeitschrift der Martin-Luther-Universität Halle-
Wittenberg Mathematisch-Naturwissenschaftliche Reihe, 10:114–115, 1961. [103]
[14] K. S. Booth and G. S. Lueker. Testing for the consecutive ones property, interval
graphs, and graph planarity using PQ-tree algorithms. Journal of Computer
and System Sciences, 13(3):335–379, 1976. [99]
[15] V. Bouchitte and M. Habib. The calculation of invariants of ordered sets. In

I. Rival, editor, Algorithms and Order, volume 255 of NATO Science Series C,
pages 231–279. Kluwer Academic Publishers, 1989. [15, 109]
[16] A. Brandstädt. The jump number problem for biconvex graphs and rectangle
covers of rectangular regions. In J. Csirik, J. Demetrovics, and F. Gécseg,
editors, Fundamentals of Computation Theory, volume 380 of Lecture Notes in
Computer Science, pages 68–77. Springer, 1989. [15, 109, 118, 119, 130, 150]
[17] A. Brandstädt, V. B. Le, and J. P. Spinrad. Graph classes: a survey. Society

for Industrial and Applied Mathematics, Philadelphia, PA, USA, 1999. [89]
[18] H. Brönnimann and M. Goodrich. Almost optimal set covers in finite vc-
dimension. Discrete Computational Geometry, 14(1):463–479, 1995. [127]
[19] J. Bruno and L. Weinberg. The principal minors of a matroid. Linear Algebra
and its Applications, 4(1):17–54, 1971. [61]
[20] F. T. Bruss. What is known about Robbins’ problem? Journal of Applied

Probability, 42(1):108–120, 2005. [43]
[21] N. Buchbinder, K. Jain, and M. Singh. Secretary problems via linear program-
ming. In F. Eisenbrand and F. B. Shepherd, editors, IPCO, volume 6080 of
Lecture Notes in Computer Science, pages 163–176. Springer, 2010. [44]
[22] P. A. Catlin, J. W. Grossman, A. M. Hobbs, and H.-J. Lai. Fractional arboricity,

strength, and principal partitions in graphs and matroids. Discrete Applied
Mathematics, 40(3):285–302, 1992. [63]
188
[23] S. Ceroi. Ordres et Géométrie Plane: Application au Nombre de Sauts. PhD
thesis, Université Montpellier II, 2000. (In French). [109, 150]
[24] S. Ceroi. A weighted version of the jump number problem on two-dimensional

orders is NP-complete. Order, 20(1):1–11, 2003. [109, 151]
[25] S. Chaiken, D. J. Kleitman, M. Saks, and J. Shearer. Covering regions by

rectangles. SIAM Journal on Algebraic and Discrete Methods, 2(4):394–410,
1981. [17, 114, 115, 118, 149]
[26] P. Chalermsook and J. Chuzhoy. Maximum independent set of rectangles. In

Proceedings of the Twentieth Annual ACM-SIAM Symposium on Discrete Al-
gorithms, SODA’09, pages 892–901, 2009. [127]
[27] G. Chaty and M. Chein. Ordered matchings and matchings without alternating
cycles in bipartite graphs. Utilitas Math., 16:183–187, 1979. [106, 108]
[28] M. Chein and M. Habib. Jump number of dags having dilworth number 2.
Discrete Applied Mathematics, 7(3):243 – 250, 1984. [108]
[29] M. Chein and P. Martin. Sur le nombre de sauts d’une forêt. C. R. Acad. Sci.
Paris, 275:159–161, 1972. [108]
[30] L. Chen and Y. Yesha. Efficient parallel algorithms for bipartite permutation
graphs. Networks, 23(1):29–39, 1993. [100]
[31] Y. Chow, S. Moriguti, H. Robbins, and S. Samuels. Optimal selection based

on relative rank (the “secretary problem”). Israel Journal of Mathematics,
2(2):81–90, 1964. [43]
[32] M. Chudnovsky, N. Robertson, P. Seymour, and R. Thomas. The strong perfect

graph theorem. Annals of Mathematics, 164(1):51–229, 2006. [103, 131, 133]
[33] V. Chvátal. On certain polytopes associated with graphs. J. Comb. Theory,

Ser. B, 18(2):138–154, 1975. [104]
[34] J. Cibulka, J. Hladký, A. Kazda, B. Lidický, E. Ondráčková, M. Tancer, and

V. Jelínek. The pinning number of overlapping rectangles. Dimacs REU Project
(unpublished manuscript), 2006. [126]
[35] O. Cogis and M. Habib. Nombre de sauts et graphes série-parallèles. RAIRO,

Informatique théorique, 13(1):3–18, 1979. [108]
[36] C. J. Colbourn and W. R. Pulleyblank. Minimizing setups in ordered sets of

fixed width. Order, 1(3):225–229, 1985. [108]
[37] D. Coppersmith and S. Winograd. Matrix multiplication via arithmetic pro-

gressions. J. Symb. Comput., 9(3):251–280, 1990. [87]
189
[38] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to
Algorithms. MIT Press, third edition, 2009. [86, 152]
[39] J. C. Culberson and R. A. Reckhow. Covering polygons is hard. J. Algorithms,

17(1):2–44, 1994. [118]
[40] E. Dahlhaus. The computation of the jump number of convex graphs. In

V. Bouchitté and M. Morvan, editors, ORDAL, volume 831 of Lecture Notes in
Computer Science, pages 176–185. Springer, 1994. [15, 17, 105, 109, 119, 149,
150]
[41] S. Das, M. Sen, A. Roy, and D. West. Interval digraphs: an analogue of interval
graphs. Journal of Graph Theory, 13(2):189–202, 1989. [97]
[42] M. Dawande. A notion of cross-perfect bipartite graphs. Information Processing

Letters, 88(4):143–147, 2003. [110, 111]
[43] R. P. Dilworth. A decomposition theorem for partially ordered sets. The Annals
of Mathematics, 51(1):161–166, 1950. [86]
[44] N. B. Dimitrov and C. G. Plaxton. Competitive weighted matching in transver-

sal matroids. In L. Aceto, I. Damgård, L. A. Goldberg, M. M. Halldórsson,
A. Ingólfsdóttir, and I. Walukiewicz, editors, ICALP, volume 5125 of Lecture
Notes in Computer Science, pages 397–408. Springer, 2008. [13, 42]
[45] D. Duffus, I. Rival, and P. Winkler. Minimizing setups for cycle-free ordered
sets. Proceedings of the American Mathematical Society, 85(4):509–513, 1982.
[107, 108]
[46] S. Dughmi. Submodular functions: Extensions, distributions, and algorithms a

survey. PhD Qualifying Exam Report, Department of Computer Science, Stan-
ford University. ArXiv version in http://arxiv.org/abs/0912.0322, 2009.
[162]
[47] B. Dushnik and E. W. Miller. Partially ordered sets. American Journal of

Mathematics, 63(3):600–610, 1941. [88, 92]
[48] E. B. Dynkin. The optimum choice of the instant for stopping a markov process.
Soviet Mathematics – Doklady, 4:627–629, 1963. [12, 13, 27]
[49] J. Edmonds. Minimum partition of a matroid into independent subsets. Journal

of research of the National Bureau of Standards. Section B, 69:67–72, 1965. [77]
[50] J. Edmonds. Matroids and the greedy algorithm. Mathematical Programming,

1(1):127–136, 1971. [13, 39, 40]
[51] M. H. El-Zahar and J. H. Schmerl. On the size of jump-critical ordered sets.

Order, 1(1):3–5, 1984. [110]
190
[52] S. Even, A. Pnueli, and A. Lempel. Permutation graphs and transitive graphs.
J. ACM, 19(3):400–410, 1972. [91]
[53] H. Fauck. Covering polygons with rectangles via edge coverings of bipartite
permutation graphs. Elektronische Informationsverarbeitung und Kybernetik,
27(8):391–409, 1991. [15, 109, 118, 119, 130, 150]
[54] S. Felsner. A 3/2-approximation algorithm for the jump number of interval

orders. Order, 6(4):325–334, 1990. [109]
[55] T. S. Ferguson. Who solved the secretary problem? Statistical Science,

4(3):282–289, 1989. [12, 43]
[56] P. C. Fishburn and P. L. Hammer. Bipartite dimensions and bipartite degrees

of graphs. Discrete Mathematics, 160(1–3):127–148, 1996. [110]
[57] L. Fleischer and S. Iwata. A push-relabel framework for submodular function

minimization and applications to parametric optimization. Discrete Applied
Mathematics, 131(2):311–322, 2003. [160]
[58] D. G. Fon-Der-Flaass and A. V. Kostochka. Covering boxes by points. Discrete

Mathematics, 120(1–3):269–275, 1993. [126]
[59] L. R. Ford and D. R. Fulkerson. Flow in networks. Princeton University Press

Princeton, 1962. [87]
[60] R. J. Fowler, M. Paterson, and S. L. Tanimoto. Optimal packing and covering

in the plane are NP-complete. Inf. Process. Lett., 12(3):133–137, 1981. [125,
151]
[61] A. Frank. On the edge-connectivity algorithm of Nagamochi and Ibaraki, 1994.

Laboratoire Artemis, IMAG, Université J. Fourier, Grenoble. [160]
[62] A. Frank. Finding minimum generators of path systems. J. Comb. Theory, Ser.
B, 75(2):237–244, 1999. [135, 138]
[63] A. Frank and T. Jordán. Minimal edge-coverings of pairs of sets. J. Comb.

Theory, Ser. B, 65(1):73–110, 1995. [17, 18, 118, 121, 148, 150]
[64] D. S. Franzblau and D. J. Kleitman. An algorithm for covering polygons with

rectangles. Information and Control, 63(3):164–189, 1984. [118, 119, 150]
[65] P. Freeman. The secretary problem and its extensions: A review. International
Statistical Review/Revue Internationale de Statistique, 51(2):189–206, 1983. [12]
[66] S. Fujishige. Theory of principal partitions revisited. In Research Trends in

Combinatorial Optimization, pages 127–162, 2009. [60, 63]
[67] D. Gale. Optimal assignments in an ordered set: An application of matroid

theory. Journal of Combinatorial Theory, 4(2):176–180, 1968. [13, 39, 40]
191
[68] T. Gallai. Transitiv orientierbare graphen. Acta Mathematica Hungarica,
18(1):25–66, 1967. [89]
[69] G. Gierz and W. Poguntke. Minimizing setups for ordered sets: A linear al-
gebraic approach. SIAM J. Algebraic Discrete Methods, 4(1):132–144, 1983.
[108]
[70] J. P. Gilbert and F. Mosteller. Recognizing the maximum of a sequence. Journal
of the American Statistical Association, 61(313):35–73, 1966. [27, 43]
[71] G. Goel, C. Karande, P. Tripathi, and L. Wang. Approximability of combinato-
rial problems with multi-agent submodular cost functions. In Proceedings of the
50th Annual IEEE Symposium on Foundations of Computer Science, FOCS’09,
pages 755–764, 2009. [162]
[72] G. Goel and A. Mehta. Online budgeted matching in random input models with
applications to adwords. In Proceedings of the Nineteenth Annual ACM-SIAM
Symposium on Discrete algorithms, SODA’08, pages 982–991, 2008. [42]
[73] M. X. Goemans and V. S. Ramakrishnan. Minimizing submodular functions
over families of sets. Combinatorica, 15(4):499–513, Dec. 1995. [162]
[74] D. A. Gregory, N. J. Pullman, K. F. Jones, and J. R. Lundgren. Biclique
coverings of regular bigraphs and minimum semiring ranks of regular matrices.
J. Comb. Theory Ser. B, 51(1):73–89, 1991. [112]
[75] M. Grötschel, L. Lovász, and A. Schrijver. The ellipsoid method and its con-
sequences in combinatorial optimization. Combinatorica, 1(2):169–197, 1981.
[18, 160]
[76] M. Grötschel, L. Lovász, and A. Schrijver. Geometric algorithms and combi-
natorial optimization. Springer-Verlag, second edition, 1993. [18, 104, 160,
162]
[77] U. I. Gupta, D. T. Lee, and J. Y.-T. Leung. Efficient algorithms for interval
graphs and circular-arc graphs. Networks, 12(4):459–467, 1982. [125]
[78] S. M. Gusein-Zade. The problem of choice and the sptimal stopping rule for
a sequence of independent trials. Theory of Probability and its Applications,
11(3):472–476, 1966. [43]
[79] A. Gyárfás and J. Lehel. Covering and coloring problems for relatives of inter-
vals. Discrete Mathematics, 55(2):167–180, 1985. [126]
[80] E. Györi. A minimax theorem on intervals. J. Comb. Theory, Ser. B, 37(1):1–9,
1984. [17, 114, 115, 116, 117, 118, 135, 138, 149]
[81] M. Habib. Partitions en chemins des sommets et sauts dans les graphes sans
circuit. PhD thesis, Université Pierre et Marie Curie, 1975. Thèse de 3e cycle.
[108]
192
[82] M. Habib. Comparability invariants. In M. Pouzet and D. Richard, editors,
Ordres: Description et Rôles, volume 99 of North-Holland Mathematics Studies,
pages 371–385. North-Holland, 1984. [15, 108]
[83] R. Hegde and K. Jain. The hardness of approximating poset dimension. Elec-
tronic Notes in Discrete Mathematics, 29:435–443, 2007. European Conference
on Combinatorics, Graph Theory and Applications. [89]
[84] T. Hiraguchi. On the dimension of partially ordered sets. Sci. Rep. Kanazawa
Univ., 1:77–94, 1951. [89]
[85] A. J. Hoffman, A. W. J. Kolen, and M. Sakarovitch. Totally-balanced and
greedy matrices. SIAM Journal on Algebraic and Discrete Methods, 6(4):721–
730, 1985. [93]
[86] J. E. Hopcroft and R. M. Karp. An n5/2 algorithm for maximum matchings in
bipartite graphs. SIAM Journal on Computing, 2(4):225–231, 1973. [87, 139]
[87] S. Im and Y. Wang. Secretary problems: Laminar matroid and interval schedul-
ing. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA’11, pages 1265–1274, 2011. [13, 42, 44]
[88] S. Iwata. A fully combinatorial algorithm for submodular function minimiza-
tion. J. Comb. Theory, Ser. B, 84(2):203–212, Mar. 2002. [160]
[89] S. Iwata, L. Fleischer, and S. Fujishige. A combinatorial strongly polynomial al-
gorithm for minimizing submodular functions. Journal of the ACM, 48(4):761–
777, 2001. [160]
[90] S. Iwata and K. Nagano. Submodular function minimization under covering
constraints. In Proceedings of the 50th Annual IEEE Symposium on Foundations
of Computer Science, FOCS’09, pages 671–680, 2009. [162]
[91] S. Iwata and J. B. Orlin. A simple combinatorial algorithm for submodular
function minimization. In Proceedings of the Twentieth Annual ACM-SIAM
Symposium on Discrete Algorithms, SODA’09, pages 1230–1237, 2009. [160]
[92] D. S. Johnson. The NP-completeness column: An ongoing guide. Journal of
Algorithms, 8(3):438–448, 1987. [118]
[93] D. R. Karger. Random sampling and greedy sparsification for matroid opti-
mization problems. Mathematical Programming, 82(1):41–81, 1998. [42]
[94] G. Kishi and Y. Kajitani. Maximally distant trees and principal partition of a
linear graph. IEEE Transactions on Circuit Theory, 16(3):323–330, 1969. [60,
61]
[95] R. Kleinberg. A multiple-choice secretary algorithm with applications to online
auctions. In Proceedings of the Sixteenth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA’05, pages 630–631, 2005. [13, 30]
193
[96] D. E. Knuth. Irredundant intervals. ACM Journal of Experimental Algorith-
mics, 1, 1996. Article 1. [117, 118]
[97] N. Korula and M. Pál. Algorithms for secretary problems on graphs and hyper-
graphs. In S. Albers, A. Marchetti-Spaccamela, Y. Matias, S. E. Nikoletseas,
and W. Thomas, editors, ICALP, volume 5556 of Lecture Notes in Computer
Science, pages 508–520. Springer, 2009. [13, 14, 42, 44, 75, 81]
[98] G. Károlyi and G. Tardos. On point covers of multiple intervals and axis-parallel
rectangles. Combinatorica, 16(2):213–222, 1996. [126]
[99] E. Kushilevitz and N. Nisan. Communication complexity. Cambridge University

Press, New York, NY, USA, 1997. [16, 112]
[100] H.-J. Lai and H. Lai. Every matroid is a submatroid of a uniformly dense
matroid. Discrete Applied Mathematics, 63(2):151–160, 1995. [70]
[101] D. V. Lindley. Dynamic programming and decision theory. Journal of the Royal
Statistical Society. Series C (Applied Statistics), 10(1):39–51, 1961. [12, 13, 27,
43]
[102] L. Lovász. A characterization of perfect graphs. J. Comb. Theory, Ser. B,

13(2):95 – 98, 1972. [104]
[103] L. Lovász. Normal hypergraphs and the perfect graph conjecture. Discrete
Mathematics, 2(3):253–267, 1972. [103, 104]
[104] L. Lovász. Submodular functions and convexity. In M. G. A. Bachem and

B. Korte, editors, Mathematical Programming – The State of the Art (Bonn
1982), pages 235–257. Springer Verlag, 1983. [18]
[105] V. V. Lozin. E-free bipartite graphs. Diskretn. Anal. Issled. Oper., Ser. 1,
7(1):49–66, 2000. [110]
[106] V. V. Lozin and M. U. Gerber. On the jump number problem in hereditary

classes of bipartite graphs. Order, 17(4):377–385, 2000. [109]
[107] A. Lubiw. Orderings and some combinatorial optimization problems with geo-
metric applications. PhD thesis, University of Toronto, Toronto, Ont., Canada,
1986. [118]
[108] A. Lubiw. Doubly lexical orderings of matrices. SIAM J. Comput., 16(5):854–

879, 1987. [93]
[109] A. Lubiw. The Boolean Basis Problem and How to Cover Some Polygons by
Rectangles. SIAM Journal on Discrete Mathematics, 3(1):98, 1990. [113, 118]
[110] A. Lubiw. A weighted min-max relation for intervals. J. Comb. Theory, Ser.
B, 53(2):151–172, 1991. [117, 118, 154, 155]
194
[111] W. Mader. Über n-fach zusammenhängende eckenmengen in graphen. J. Comb.
Theory, Ser. B, 25(1):74–93, 1978. [164]
[112] R. B. Manfrino, J. A. G. Ortega, and R. V. Delgado. Inequalities: A Mathe-

matical Olympiad Approach. Birkhäuser, 2009. [51]
[113] W. J. Masek. Some NP-complete set covering problems. (unpublished

manuscript), 1979. [118]
[114] C. McCartin. An improved algorithm for the jump number problem. Informa-
tion Processing Letters, 79(2):87–92, 2001. [110]
[115] R. M. McConnell and J. P. Spinrad. Linear-time transitive orientation. In Pro-

ceedings of the Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA’97, pages 19–25, 1997. [92]
[116] L. Mirsky. A dual of Dilworth’s decomposition theorem. The American Math-

ematical Monthly, 78(8):876–877, 1971. [86]
[117] J. Mitas. Tackling the jump number of interval orders. Order, 8(2):115–132,
1991. [109]
[118] R. Motwani and P. Raghavan. Randomized algorithms. Cambridge University

Press, New York, NY, USA, 1995. [73]
[119] A. G. Mucci. Differential equations and optimal choice problems. The Annals
of Statistics, 1(1):104–113, 1973. [43]
[120] M. Mucha and P. Sankowski. Maximum matchings via gaussian elimination. In

Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer
Science, FOCS’04, pages 248–255, 2004. [87, 139]
[121] H. Müller. Alternating cycle-free matchings. Order, 7(1):11–21, 1990. [109, 110,
118]
[122] H. Müller. On edge perfectness and classes of bipartite graphs. Discrete Math-
ematics, 149(1–3):159–187, 1996. [110, 111, 118, 130]
[123] H. Müller. Recognizing interval digraphs and interval bigraphs in polynomial

time. Discrete Applied Mathematics, 78(1–3):189–205, 1997. [96, 97]
[124] H. Nagamochi. Minimum degree orderings. Algorithmica, 56(1):17–34, 2010.

[19, 159, 162, 181, 182, 183, 186]
[125] H. Nagamochi and T. Ibaraki. Computing Edge-Connectivity in multigraphs

and capacitated graphs. SIAM Journal on Discrete Mathematics, 5(1):54–66,
1992. [18, 160]
195
[126] H. Nagamochi and T. Ibaraki. A linear-time algorithm for finding a sparse k-
connected spanning subgraph of a k-connected graph. Algorithmica, 7(1):583–
596, 1992. [18, 160]
[127] H. Nagamochi and T. Ibaraki. A note on minimizing submodular functions.

Information Processing Letters, 67(5):239–244, 1998. [19, 159, 161, 162, 166,
169, 175]
[128] H. Narayanan. Theory of matroids and network analysis. PhD thesis, Depart-
ment of Electrical Engineering, Indian Institute of Technology, Bombay, 1974.
[62]
[129] H. Narayanan. Submodular Functions and Electrical Networks, volume 54 of

Annals of Discrete Mathematics. Elsevier, 1997. [61]
[130] H. Narayanan. A note on the minimization of symmetric and general submod-

ular functions. Discrete Applied Mathematics, 131(2):513–522, 2003. [162, 169,
175]
[131] H. Narayanan and M. N. Vartak. An elementary approach to the principal

partition of a matroid. Transactions of the Institute of Electronics and Com-
munication Engineers of Japan. Section E, E64(4):227–234, 1981. [60, 61, 63]
[132] O. Ore. Theory of graphs. American Mathematical Society, 1962. [88]
[133] J. Orlin. Contentment in graph theory: Covering graphs with cliques. Indaga-
tiones Mathematicae (Proceedings), 80(5):406–424, 1977. [112, 117]
[134] J. B. Orlin. A faster strongly polynomial time algorithm for submodular func-
tion minimization. Mathematical Programming, 118(2):237–251, 2007. [18, 160]
[135] Y. Otachi, Y. Okamoto, and K. Yamazaki. Relationships between the class

of unit grid intersection graphs and other classes of bipartite graphs. Discrete
Applied Mathematics, 155(17):2383–2390, 2007. [16]
[136] S. Oveis Gharan and J. Vondrák. On variants of the matroid secretary problem.
Manuscript. ArXiv version in http://arxiv.org/abs/1007.2140, 2011. [14,
41, 55, 70, 80]
[137] J. G. Oxley. Matroid theory. Oxford University Press, USA, 2006. [36, 58, 60,
74, 79]
[138] M. W. Padberg and M. R. Rao. Odd minimum cut-sets and b-matchings. Math-
ematics of Operations Research, 7:67–80, 1982. [162]
[139] W. R. Pulleyblank. Alternating cycle free matchings. Technical Report CORR

82-18, University of Waterloo - Dept. of Combinatorics and Optimization, 1982.
[15, 108]
196
[140] W. R. Pulleyblank. On minimizing setups in precedence constrained scheduling.
(unpublished), 1982. [106]
[141] M. Queyranne. Minimizing symmetric submodular functions. Mathematical

Programming, 82(1):3–12, 1998. [18, 19, 159, 160, 162, 164]
[142] R. Rado. Note on independence functions. Proceedings of the London Mathe-

matical Society, 3(1):300–320, 1957. [13]
[143] W. T. Rasmussen and S. R. Pliska. Choosing the maximum from a sequence

with a discount function. Applied Mathematics & Optimization, 2(3):279–289,
1975. [43]
[144] I. Rival. Optimal linear extensions by interchanging chains. Proceedings of the

American Mathematical Society, 89(3):387–394, 1983. [108]
[145] R. Rizzi. On minimizing symmetric set functions. Combinatorica, 20(3):445–

450, 2000. [19, 161, 162, 169, 176, 177, 178]
[146] A. Schrijver. A combinatorial algorithm minimizing submodular functions in

strongly polynomial time. J. Comb. Theory, Ser. B, 80(2):346–355, 2000. [160]
[147] A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer,

2003. [36, 61, 75, 76]
[148] A. Sharary. The jump number of Z-free ordered sets. Order, 8(3):267–273, 1991.
[109]
[149] A. H. Sharary and N. Zaguia. On minimizing jumps for ordered sets. Order,
7(4):353–359, 1990. [109]
[150] A. M. S. Shrestha, S. Tayu, and S. Ueno. On orthogonal ray graphs. Discrete

Applied Mathematics, 158(15):1650–1659, 2010. [95]
[151] A. M. S. Shrestha, S. Tayu, and S. Ueno. On two-directional orthogonal ray

graphs. In Proceedings of 2010 IEEE International Symposium on Circuits and
Systems (ISCAS), pages 1807–1810, 2010. [16, 94, 95]
[152] J. A. Soto. Matroid secretary problem in the random assignment model. In

Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA’11, pages 1275–1284, 2011. ArXiv version in http://arxiv.
org/abs/1007.2152. [13, 14, 41, 55, 69, 80]
[153] J. A. Soto and C. Telha. Jump number of two-directional orthogonal ray graphs.
To appear in the proceedings of the 15th Conference on Integer Programming
and Combinatorial Optimization, 2011. [16, 105, 115, 129, 138]
[154] J. Spinrad, A. Brandstädt, and L. Stewart. Bipartite permutation graphs.

Discrete Applied Mathematics, 18(3):279–292, 1987. [100]
197
[155] G. Steiner. On finding the jump number of a partial orber by substitution
decomposition. Order, 2(1):9–23, 1985. [108]
[156] G. Steiner and L. K. Stewart. A linear time algorithm to find the jump number
of 2-dimensional bipartite partial orders. Order, 3(4):359–367, 1987. [15, 109,
130]
[157] T. J. Stewart. Optimal selection from a random sequence with learning of

the underlying distribution. Journal of the American Statistical Association,
73(364):775–780, 1978. [43]
[158] M. Stoer and F. Wagner. A simple min-cut algorithm. Journal of the ACM,
44(4):585–591, 1997. [160]
[159] Z. Svitkina and L. Fleischer. Submodular approximation: Sampling-based algo-

rithms and lower bounds. In Proceedings of the 49th Annual IEEE Symposium
on Foundations of Computer Science, pages 697–706, 2008. [18, 159, 160]
[160] M. M. Syslo. The jump number problem on interval orders: A approximation

algorithm. Discrete Mathematics, 144(1–3):119–130, 1995. [109]
[161] E. Szpilrajn. Sur l’extension de l’ordre partiel. Fundamenta mathematicae,

16:386–389, 1930. [88]
[162] N. Thain. On the transversal matroid secretary problem. Master’s thesis, McGill
University, Canada, 2008. [13, 42]
[163] N. Tomizawa. Strongly Irreducible Matroids and Principal Partition of a Ma-

troid into Strongly Irreducible Minors. Electronics & Communications In Japan,
59(A):1–10, 1976. [62]
[164] W. T. Trotter. Combinatorics and Partially Ordered Sets: Dimension theory.

Johns Hopkins University Press, 2001. [85]
[165] P. van Emde Boas. Preserving order in a forest in less than logarithmic time. In
16th Annual Symposium on Foundations of Computer Science, FOCS’75, pages
75–84, 1975. [144]
[166] L. A. Végh. Connectivity Augmentation Algorithms. PhD thesis, Eötvös Loránd

University, 2010. [149]
[167] G. Wegner. Über eine kombinatorisch-geometrische frage von hadwiger und

debrunner. Israel Journal of Mathematics, 3(4):187–198, 1965. [126]
[168] M. Yannakakis. The complexity of the partial order dimension problem. SIAM
Journal on Algebraic and Discrete Methods, 3(3):351–358, 1982. [89]
198

MIT Thesis Jsoto

Uploaded by

Copyright:

Available Formats

MIT Thesis Jsoto

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MIT Thesis Jsoto

Uploaded by

Copyright:

Available Formats

Contributions on Secretary Problems, Independent

Sets of Rectangles and Related Problems

Submitted to the Department of Mathematics

Thesis Supervisor: Michel X. Goemans

I Matroid Secretary Problem 21

2 New Results for Matroid Secretary Problems 45

II Jump Number of Two Directional Orthogonal Ray

4 A Primer on the Jump Number Problem 105

6 Weighted Cross-free Matching 151

III Constrained Set Function Minimization 157

In this thesis we focus on three problems in combinatorial optimization arising from

Matroid Secretary Problem

Jump Number of Two Directional Orthogonal Ray

Constrained Set Function Minimization

In this thesis, we show how to extend Queyranne’s algorithm to return a nontrivial

Matroid Secretary Problem

Secretary Problems and Matroids

1.1 Introduction to Secretary Problems

• Do we know the number of elements a priori?

• In what order are the elements presented?

• What information we receive at the moment we examine an element?

• What information we know a priori about the elements?

• What is our objective function?

1.1.1 Comparison and Value Based Algorithms

1.2 Classical Secretary Problem

1. Full-adversarial model: An adversary selects both the hidden weights and

Every adversarial selection is taken before the execution of the algorithm. In

E[αw(ALG) − w∗] ≥ 0, (1.1)

1.2.1 Previous Results for the Classical Secretary Problem

In particular, there is an index j for which pj ≤ 1/n. Consider the sequence of

Lemma 1.2.2. Any algorithm guaranteeing a probability p of selecting the largest

E[1/p · w(ALG) − w∗] ≥ 1/p · pw∗ − w∗ ≥ 0. (1.4)

0 ≤ E[1/p · w(ALG) − w∗ ] ≤ 1/p · (qLn + Ln−1 ) − Ln = Ln /p · (q + ε − p).

This implies that q ≥ p − ε. Using that ε is arbitrary, we conclude the proof.

1. The optimum value r(n) is such that limn→∞ r(n)/n = 1/e.

In particular, in order to obtain an almost optimal e-competitive algorithm it is

Algorithm 1 For the random-order classical secretary problem.

In particular, it returns w1 with probability at least (−p ln p). Therefore, by setting

Proof. Consider the following offline simulation. Let w1 ≥ w2 ≥ · · · ≥ wn ≥ 0 be

As event Ei has probability (1 − ti )i−1 , we have

In particular the probability that w1 is selected is p1 pt dt = −p ln p. This is maximized

1.3.1 Previous Results for the Multiple Choice Secretary

In particular, there is an index j for which pj ≤ r/n. Consider the sequence of

E[w(ALG)] ≤ pj n3 + r(j − 1)ε ≤ n2(r + ε). (1.7)

A natural optimum algorithm for the full-adversarial model would be simply to

Kleinberg’s algorithm is recursively defined. If r = 1, then simply apply Lindley–

Lemma 1.3.3. Algorithm 2 is (−p ln p)−1 -competitive. In particular, by setting p =

Pr(wi is selected | ti ≥ p) = Pr(|A(ti )| < r | ti ≥ p) + Pr(|A(ti )| ≥ r, tA < p | ti ≥ p)

By setting p = 1/e, we obtain a competitive ratio of e.

1.4 Generalized Secretary Problems

1. Adversarial-Assignment: The weights are assigned to the elements of S by

2. Random-Assignment: An adversary selects a list W = {w1 , . . . , wn } of

Order of the elements. Consider the following two variants.

1. Adversarial-Order: An adversary selects the ordering of the elements, this

2. Random-Order: The elements are presented in uniform random order.

1.4.3 Performance Tool: Competitive Analysis

E[αw(ALG) − w∗] ≥ 0, (1.8)

1.4.4 Standard Assumptions

1.4.5 Special Cases