arXiv++ Combinatorics

The three-in-a-tree algorithm of Chudnovsky and Seymour decides in time $O(n^4)$ whether three given vertices of a graph belong to an induced tree. Here, we study four-in-a-tree for triangle-free graphs. We give a structural answer to the following question: what does a triangle-free graph look like if no induced tree covers four given vertices? Our main result says that any such graph must have the "same structure", in a sense to be defined precisely, as a square or a cube. We provide an $O(nm)$-time algorithm that given a triangle-free graph $G$ together with four vertices outputs either an induced tree that contains them or a partition of $V(G)$ certifying that no such tree exists. We prove that the problem of deciding whether there exists a tree $T$ covering the four vertices such that at most one vertex of $T$ has degree at least 3 is NP-complete.

We give a structural description of the class $\cal C$ of graphs that do not contain a cycle with a unique chord as an induced subgraph. Our main theorem states that any connected graph in $\cal C$ is either in some simple basic class or has a decomposition. Basic classes are chordless cycles, cliques, bipartite graphs with one side containing only nodes of degree two and induced subgraphs of the famous Heawood or Petersen graph. Decompositions are node cutsets consisting of one or two nodes and edge cutsets called 1-joins. Our decomposition theorem actually gives a complete structure theorem for ${\cal C}$, i.e. every graph in ${\cal C}$ can be built from basic graphs that can be explicitly constructed, and gluing them together by prescribed composition operations; and all graphs built this way are in ${\cal C}$. This has several consequences: an ${\cal O}(nm)$-time algorithm to decide whether a graph is in $\cal C$, an ${\cal O}(n+m)$-time algorithm that finds a maximum clique of any graph in $\cal C$ and an ${\cal O}(nm)$-time coloring algorithm for graphs in $\cal C$. We prove that every graph in $\cal C$ is either 3-colorable or has a coloring with $ω$ colors where $ω$ is the size of a largest clique. The problem of finding a maximum stable set for a graph in $\cal C$ is known to be NP-hard.

A hole in a graph is an induced cycle on at least four vertices. A graph is Berge if it has no odd hole and if its complement has no odd hole. In 2002, Chudnovsky, Robertson, Seymour and Thomas proved a decomposition theorem for Berge graphs saying that every Berge graph either is in a well understood basic class, or has some kind of decomposition. Then, Chudnovsky proved stronger theorems. One of them restricts the allowed decompositions to 2-joins and balanced skew partitions. We prove that the problem of deciding whether a graph has a balanced skew partition is NP-hard. We give an $O(n^9)$-time algorithm for the same problem restricted to Berge graphs. Our algorithm is not constructive: it only certifies whether a graph has a balanced skew partition or not. It relies on a new decomposition theorem for Berge graphs that is more precise than the previously known theorems. Our theorem also implies that every Berge graph can be decomposed in a first step by using only balanced skew partitions, and in a second step by using only 2-joins. Our proof of this new theorem uses at an essential step one of the theorems of Chudnovsky.

We consider the class of graphs containing no odd hole, no odd antihole, and no configuration consisting of three paths between two nodes such that any two of the paths induce a hole, and at least two of the paths are of length 2. This class generalizes claw-free Berge graphs and square-free Berge graphs. We give a combinatorial algorithm of complexity $O(n^{7})$ to find a clique of maximum weight in such a graph. We also consider several subgraph-detection problems related to this class.

We consider the class ${\cal A}$ of graphs that contain no odd hole, no antihole of length at least 5, and no "prism" (a graph consisting of two disjoint triangles with three disjoint paths between them) and the class ${\cal A}'$ of graphs that contain no odd hole, no antihole of length at least 5 and no odd prism (prism whose three paths are odd). These two classes were introduced by Everett and Reed and are relevant to the study of perfect graphs. We give polynomial-time recognition algorithms for these two classes. We proved previously that every graph $G\in{\cal A}$ is "perfectly contractile", as conjectured by Everett and Reed [see the chapter "Even pairs" in the book {\it Perfect Graphs}, J.L. Ramírez-Alfonsín and B.A. Reed, eds., Wiley Interscience, 2001]. The analogous conjecture concerning graphs in ${\cal A}'$ is still open.

We consider the class ${\cal A}$ of graphs that contain no odd hole, no antihole, and no "prism" (a graph consisting of two disjoint triangles with three disjoint paths between them). We prove that every graph $G\in{\cal A}$ different from a clique has an "even pair" (two vertices that are not joined by a chordless path of odd length), as conjectured by Everett and Reed [see the chapter "Even pairs" in the book {\it Perfect Graphs}, J.L. Ramírez-Alfonsín and B.A. Reed, eds., Wiley Interscience, 2001]. Our proof is a polynomial-time algorithm that produces an even pair with the additional property that the contraction of this pair yields a graph in ${\cal A}$. This entails a polynomial-time algorithm, based on successively contracting even pairs, to color optimally every graph in ${\cal A}$. This generalizes several results concerning some classical families of perfect graphs.

A graph is Berge if it has no induced odd cycle on at least 5 vertices and no complement of induced odd cycle on at least 5 vertices. A graph is perfect if the chromatic number equals the maximum clique number for every induced subgraph. Chudnovsky, Robertson, Seymour and Thomas proved that every Berge graph either falls into some classical family of perfect graphs, or has a structural fault that cannot occur in a minimal imperfect graph. A corollary of this is the strong perfect graph theorem conjectured by Berge: every Berge graph is perfect. An even pair of vertices in a graph is a pair of vertices such that every induced path between them has even length. Meyniel proved that a minimal imperfect graph cannot contain an even pair. So even pairs may be considered as a structural fault. Chudnovsky et al. do not use them, and it is known that some classes of Berge graph have no even pairs. The aim of this work is to investigate an "even-pair-like" notion that could be a structural fault present in every Berge graph. An odd pair of cliques is a pair of cliques $\{K_1, K_2\}$ such that every induced path from $K_1$ to $K_2$ with no interior vertex in $K_1 \cup K_2$ has odd length. We conjecture that for every Berge graph $G$ on at least two vertices, either one of $G, \bar{G}$ has an even pair, or one of $G, \bar{G}$ has an odd pair of cliques. We conjecture that a minimal imperfect graph has no odd pair of maximal cliques. We prove these conjectures in some special cases. We show that adding all edges between any 2 vertices of the cliques of an odd pair of cliques is an operation that preserves perfectness.

This work is the PhD thesis of Nicolas Trotignon, written in 2004 under the supervision of Frédéric Maffray. It is motivated by the desire for a better understanding of perfect graphs. The proof of the Claude Berge's perfect graph conjecture in 2002 by Chudnovsky, Robertson, Seymour and Thomas has shed a new light on this field of combinatorics. But some questions are still unsettled, particulary the existence of a combinatorial algorithm for the coloring of perfect graphs. An even pair of a graph is a pair of vertices such that every path joining them has even length. As proved by Fonlupt and Uhry, the contraction of an even pair preserves the chromatic number, and when applied recursively may lead to an optimal coloring. We prove a conjecture of Everett and Reed saying that this method works for a class of perfect graphs: Artemis graphs. This yields a coloring algorithm for Artemis graphs with complexity $O(n^2m)$. We give an $O(n^9)$ algorithm for the recognition of Artemis graphs. Other recognition algorithms are also given, each of them based on subgraph detection routines for Berge graphs. We show that these subgraph detection problems are NP-complete when extended to general graphs.

This document is the Habilitation Thesis (Habilitation à Diriger des Recherches) of Nicolas Trotignon. It presents his work in structural graph theory, from 2001 to 2009.

A \emph{wheel} is a graph made of a cycle of length at least~4 together with a vertex that has at least three neighbors in the cycle. We prove that the problem whose instance is a graph $G$ and whose question is "does $G$ contains a wheel as an induced subgraph" is NP-complete. We also settle the complexity of several similar problems.

A \emph{self-complementary} graph is a graph isomorphic to its complement. An isomorphism between $G$ and its complement, viewed as a permutation of $V(G)$, is then called an \emph{antimorphism}. A \emph{skew partition} of $G$ is a partition of $V(G)$ into 4 sets $A,B,C,D$ such that there is no edge between $A,B$ and every possible edge between $C,D$. A \emph{symmetric partition} of $G$ is a partition of $V(G)$ into 4 sets $A,B,C,D$ such that there is no edge between $A, D$, no edge between $B, C$, every possible edge between $A,B$ and every possible edge between $C,D$. We give a new proof of a theorem of Gibbs saying that every self-complementary graph on $4k$ vertices has $k$ disjoint paths on 4 vertices as induced subgraph. This new proof gives more structural information than the original one. We conjecture that every self-complementary graph on $4k$ vertices either has an induced cycle on 5 vertices, or a skew partition, or a symmetric partition. The new proof of Gibb's theorem yields a proof of the conjecture for the self-complementary graphs that have an antimorphism that is the product of a two circular permutations, one of them of length 4.

We say that a 0-1 matrix $N$ of size $a\times b$ can be found in a collection of sets $\mathcal{H}$ if we can find sets $H_{1}, H_{2}, \dots, H_{a}$ in $\mathcal{H}$ and elements $e_1, e_2, \dots, e_b$ in $\cup_{H \in \mathcal{H}} H$ such that $N$ is the incidence matrix of the sets $H_{1}, H_{2}, \dots, H_{a}$ over the elements $e_1, e_2, \dots, e_b$. We prove the following Ramsey-type result: for every $n\in \N$, there exists a number S(n) such that in any collection of at least S(n) sets, one can find either the incidence matrix of a collection of $n$ singletons, or its complementary matrix, or the incidence matrix of a collection of $n$ sets completely ordered by inclusion. We give several results of the same extremal set theoretical flavour. For some of these, we give the exact value of the number of sets required.

A class of graphs closed under taking induced subgraphs is $χ$-bounded if there exists a function $f$ such that for all graphs $G$ in the class, $χ(G) \leq f(ω(G))$. We consider the following question initially studied in [A. Gy{á}rf{á}s, Problems from the world surrounding perfect graphs, {\em Zastowania Matematyki Applicationes Mathematicae}, 19:413--441, 1987]. For a $χ$-bounded class $\cal C$, is the class $\bar{C}$ $χ$-bounded (where $\bar{\cal C}$ is the class of graphs formed by the complements of graphs from $\cal C$)? We show that if $\cal C$ is $χ$-bounded by the constant function $f(x)=3$, then $\bar{\cal C}$ is $χ$-bounded by $g(x)=\lfloor\frac{8}{5}x\rfloor$ and this is best possible. We show that for every constant $c>0$, if $\cal C$ is $χ$-bounded by a function $f$ such that $f(x)=x$ for $x \geq c$, then $\bar{\cal C}$ is $χ$-bounded. For every $j$, we construct a class of graphs $χ$-bounded by $f(x)=x+x/\log^j(x)$ whose complement is not $χ$-bounded.

A class $\mathcal{G}$ of graphs is said to be {\em $χ$-bounded} if there is a function $f:\mathbb{N} \rightarrow \mathbb{R}$ such that for all $G \in \mathcal{G}$ and all induced subgraphs $H$ of $G$, $χ(H) \leq f(ω(H))$. In this paper, we show that if $\mathcal{G}$ is a $χ$-bounded class, then so is the closure of $\mathcal{G}$ under any one of the following three operations: substitution, gluing along a clique, and gluing along a bounded number of vertices. Furthermore, if $\mathcal{G}$ is $χ$-bounded by a polynomial (respectively: exponential) function, then the closure of $\mathcal{G}$ under substitution is also $χ$-bounded by some polynomial (respectively: exponential) function. In addition, we show that if $\mathcal{G}$ is a $χ$-bounded class, then the closure of $\mathcal{G}$ under the operations of gluing along a clique and gluing along a bounded number of vertices together is also $χ$-bounded, as is the closure of $\mathcal{G}$ under the operations of substitution and gluing along a clique together.

Perfect graphs were defined by Claude Berge in the 1960s. They are important objects for graph theory, linear programming and combinatorial optimization. Claude Berge made a conjecture about them, that was proved by Chudnovsky, Robertson, Seymour and Thomas in 2002, and is now called the strong perfect graph theorem. This is a survey about perfect graphs, mostly focused on the strong perfect graph theorem.

We provide a general method to prove the existence and compute efficiently elimination orderings in graphs. Our method relies on several tools that were known before, but that were not put together so far: the algorithm LexBFS due to Rose, Tarjan and Lueker, one of its properties discovered by Berry and Bordat, and a local decomposition property of graphs discovered by Maffray, Trotignon and Vu\vsković. We use this method to prove the existence of elimination orderings in several classes of graphs, and to compute them in linear time. Some of the classes have already been studied, namely even-hole-free graphs, square-theta-free Berge graphs, universally signable graphs and wheel-free graphs. Some other classes are new. It turns out that all the classes that we study in this paper can be defined by excluding some of the so-called Truemper configurations. For several classes of graphs, we obtain directly bounds on the chromatic number, or fast algorithms for the maximum clique problem or the coloring problem.

The {\em chromatic gap} is the difference between the chromatic number and the clique number of a graph. Here we investigate $\gap(n)$, the maximum chromatic gap over graphs on $n$ vertices. Can the extremal graphs be explored? While computational problems related to the chromatic gap are hopeless, an interplay between Ramsey theory and matching theory leads to a simple and (almost) exact formula for $\gap(n)$ in terms of Ramsey numbers.